Re: [PATCH] x86/traps: 'Fix' safety of read_registers() in #DF path

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 16/10/2020 12:03, Jan Beulich wrote:
> On 16.10.2020 12:58, Andrew Cooper wrote:
>> On 15/10/2020 08:27, Jan Beulich wrote:
>>> On 14.10.2020 20:00, Andrew Cooper wrote:
>>>> On 13/10/2020 16:51, Jan Beulich wrote:
>>>>> On 12.10.2020 15:49, Andrew Cooper wrote:
>>>>>> All interrupts and exceptions pass a struct cpu_user_regs up into C.  This
>>>>>> contains the legacy vm86 fields from 32bit days, which are beyond the
>>>>>> hardware-pushed frame.
>>>>>>
>>>>>> Accessing these fields is generally illegal, as they are logically out of
>>>>>> bounds for anything other than an interrupt/exception hitting ring1/3 code.
>>>>>>
>>>>>> Unfortunately, the #DF handler uses these fields as part of preparing the
>>>>>> state dump, and being IST, accesses the adjacent stack frame.
>>>>>>
>>>>>> This has been broken forever, but c/s 6001660473 "x86/shstk: Rework the stack
>>>>>> layout to support shadow stacks" repositioned the #DF stack to be adjacent to
>>>>>> the guard page, which turns this OoB write into a fatal pagefault:
>>>>>>
>>>>>>   (XEN) *** DOUBLE FAULT ***
>>>>>>   (XEN) ----[ Xen-4.15-unstable  x86_64  debug=y   Tainted:  C   ]----
>>>>>>   (XEN) ----[ Xen-4.15-unstable  x86_64  debug=y   Tainted:  C   ]----
>>>>>>   (XEN) CPU:    4
>>>>>>   (XEN) RIP:    e008:[<ffff82d04031fd4f>] traps.c#read_registers+0x29/0xc1
>>>>>>   (XEN) RFLAGS: 0000000000050086   CONTEXT: hypervisor (d1v0)
>>>>>>   ...
>>>>>>   (XEN) Xen call trace:
>>>>>>   (XEN)    [<ffff82d04031fd4f>] R traps.c#read_registers+0x29/0xc1
>>>>>>   (XEN)    [<ffff82d0403207b3>] F do_double_fault+0x3d/0x7e
>>>>>>   (XEN)    [<ffff82d04039acd7>] F double_fault+0x107/0x110
>>>>>>   (XEN)
>>>>>>   (XEN) Pagetable walk from ffff830236f6d008:
>>>>>>   (XEN)  L4[0x106] = 80000000bfa9b063 ffffffffffffffff
>>>>>>   (XEN)  L3[0x008] = 0000000236ffd063 ffffffffffffffff
>>>>>>   (XEN)  L2[0x1b7] = 0000000236ffc063 ffffffffffffffff
>>>>>>   (XEN)  L1[0x16d] = 8000000236f6d161 ffffffffffffffff
>>>>>>   (XEN)
>>>>>>   (XEN) ****************************************
>>>>>>   (XEN) Panic on CPU 4:
>>>>>>   (XEN) FATAL PAGE FAULT
>>>>>>   (XEN) [error_code=0003]
>>>>>>   (XEN) Faulting linear address: ffff830236f6d008
>>>>>>   (XEN) ****************************************
>>>>>>   (XEN)
>>>>>>
>>>>>> and rendering the main #DF analysis broken.
>>>>>>
>>>>>> The proper fix is to delete cpu_user_regs.es and later, so no
>>>>>> interrupt/exception path can access OoB, but this needs disentangling from the
>>>>>> PV ABI first.
>>>>>>
>>>>>> Not-really-fixes: 6001660473 ("x86/shstk: Rework the stack layout to support shadow stacks")
>>>>>> Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
>>>>> Reviewed-by: Jan Beulich <jbeulich@suse.com>
>>>>>
>>>>> Is it perhaps worth also saying explicitly that the other IST
>>>>> stacks don't suffer the same problem because show_registers()
>>>>> makes an local copy of the passed in struct? (Doing so also
>>>>> for #DF would apparently be an alternative solution.)
>>>> They're not safe.  They merely don't explode.
>>>>
>>>> https://lore.kernel.org/xen-devel/1532546157-5974-1-git-send-email-andrew.cooper3@citrix.com/
>>>> was "fixed" by CET-SS turning the guard page from not present to
>>>> read-only, but the same CET-SS series swapped #DB for #DF when it comes
>>>> to the single OoB write problem case.
>>> I see. While indeed I didn't pay attention to the OoB read aspect,
>>> me saying "the other IST stacks don't suffer the same problem" was
>>> still correct, wasn't it? Anyway - not a big deal.
>> I've tweaked the commit message to make this more clear.
>>
>> --8<---
>> Accessing these fields is generally illegal, as they are logically out of
>> bounds for anything other than an interrupt/exception hitting ring1/3 code.
>>
>> show_registers() unconditionally reads these fields, but the content is
>> discarded before use.  This is benign right now, as all parts of the
>> stack are
>> readable, including the guard pages.
>>
>> However, read_registers() in the #DF handler writes to these fields as
>> part of
>> preparing the state dump, and being IST, hits the adjacent stack frame.
>> --8<--
> Thanks, lgtm.

On a tangent, what are your views WRT backport beyond 4.14?

Back then, it was #DB which was adjacent to the guard frame (which was
not present), but it doesn't use show_registers() by default, so I think
the problem is mostly hidden.

~Andrew