Re: [RFC PATCH v2 2/2] target/ppc: make gdb able to translate priviledged addresses

[Fabiano Rosas <farosas@linux.ibm.com>]

Richard Henderson <richard.henderson@linaro.org> writes:

> On 6/15/21 4:32 AM, Bruno Piazera Larsen wrote:
>> On 14/06/2021 19:37, Richard Henderson wrote:
>>> On 6/14/21 12:16 PM, Bruno Larsen (billionai) wrote:
>>>> This patch changes ppc_cpu_get_phys_page_debug so that it is now
>>>> able to translate both, priviledged and real mode addresses
>>>> independently of whether the CPU executing it has those permissions
>>>>
>>>> This was mentioned by Fabiano as something that would be very useful to
>>>> help with debugging, but could possibly constitute a security issue if
>>>> that debug function can be called in some way by prodution code. the
>>>> solution was implemented such that it would be trivial to wrap it around
>>>> ifdefs for building only with --enable-debug, for instance, but we are
>>>> not sure this is the best approach, hence why it is an RFC.
>>>>
>>>> Suggested-by: Fabiano Rosas<farosas@linux.ibm.com>
>>>> Signed-off-by: Bruno Larsen (billionai)<bruno.larsen@eldorado.org.br>
>>>> ---
>>>>   target/ppc/mmu_helper.c | 23 +++++++++++++++++++++++
>>>>   1 file changed, 23 insertions(+)
>>>
>>> I think the first part is unnecessary.  Either the cpu is in supervisor mode or it 
>>> isn't, and gdb should use the correct address space.  If you really want to force 
>>> supervisor lookup from a guest that is paused in usermode, I suppose you could force 
>>> MSR.PR=1 while you're performing the access and set it back afterward.
>> I don't see why GDB should not be able to see supervisor level addresses just because the 
>> CPU can't.
>
> Because then when you are debugging, you then don't know whether the address is actually 
> accessible in the current cpu context.
>

@Bruno, so this is what I referred to somewhere else on the thread,
people expect GDB to have the same access level of the currently
executing code. So implementing my suggestion would break their
workflow.

>>> I think the second part is actively wrong -- real-mode address lookup will (for the most 
>>> part) always succeed.  Moreover, the gdb user will have no idea that you've silently 
>>> changed addressing methods.
>> 
>> I disagree. Real-mode address will mostly fail, since during the boot process Linux 
>> kernels set the MMU to use only virtual addresses, so real mode addresses only work when 
>> debugging the firmware or the early setup of the kernel. After that, GDB can basically 
>> only see virtual addresses.
>
> Exactly.  But you changed that so that any unmapped address will re-try with real-mode, 
> which (outside of hv) simply maps real->physical and returns the input.
>
> One should have to perform some special action to see addresses in a different cpu 
> context.  I don't think that gdb supports such a special action at the moment.  If you 
> want that feature though, that's where you should start.

I think we can just drop this patch. The scenarios where debugging
across MMU contexts happen are quite limited.

My use case was a while back when implementing single-step for KVM
guests; there were some situations where GDB would have issues setting
breakpoints around kernel code that altered MSR_IR/DR. But that is
mostly anecdotal at this point. If I ever run into that again, now I
know where to look.

>
>
> r~