Re: [PATCH 12/16] x86/extable: Adjust extable handling to be shadow stack compatible

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 07/05/2020 14:35, Jan Beulich wrote:
> On 02.05.2020 00:58, Andrew Cooper wrote:
>> --- a/xen/arch/x86/traps.c
>> +++ b/xen/arch/x86/traps.c
>> @@ -778,6 +778,28 @@ static bool exception_fixup(struct cpu_user_regs *regs, 
>> bool print)
>>                 vec_name(regs->entry_vector), regs->error_code,
>>                 _p(regs->rip), _p(regs->rip), _p(fixup));
>>  
>> +    if ( IS_ENABLED(CONFIG_XEN_SHSTK) )
>> +    {
>> +        unsigned long ssp;
>> +
>> +        asm ("rdsspq %0" : "=r" (ssp) : "0" (1) );
>> +        if ( ssp != 1 )
>> +        {
>> +            unsigned long *ptr = _p(ssp);
>> +
>> +            /* Search for %rip in the shadow stack, ... */
>> +            while ( *ptr != regs->rip )
>> +                ptr++;
> Wouldn't it be better to bound the loop, as it shouldn't search past
> (strictly speaking not even to) the next page boundary? Also you
> don't care about the top of the stack (being the to be restored SSP),
> do you? I.e. maybe
>
>             while ( *++ptr != regs->rip )
>                 ;
>
> ?
>
> And then - isn't searching for a specific RIP value alone prone to
> error, in case a it matches an ordinary return address? I.e.
> wouldn't you better search for a matching RIP accompanied by a
> suitable pointer into the shadow stack and a matching CS value?
> Otherwise, ...
>
>> +            ASSERT(ptr[1] == __HYPERVISOR_CS);
> ... also assert that ptr[-1] points into the shadow stack?

So this is the problem I was talking about that the previous contexts
SSP isn't stored anywhere helpful.

What we are in practice doing is looking 2 or 3 words up the shadow
stack (depending on exactly how deep our call graph is), to the shadow
IRET frame matching the real IRET frame which regs is pointing to.

Both IRET frames were pushed in the process of generating the exception,
and we've already matched regs->rip to the exception table record.  We
need to fix up regs->rip and the shadow lip to the fixup address.

As we are always fixing up an exception generated from Xen context, we
know that ptr[1] == __HYPERVISOR_CS, and *ptr[-1] = &ptr[2], as we
haven't switched shadow stack as part of taking this exception. 
However, this second point is fragile to exception handlers moving onto IST.

We can't encounter regs->rip in the shadow stack between the current SSP
and the IRET frame we're looking to fix up, or we would have taken a
recursive fault and not reached exception_fixup() to begin with.

Therefore, the loop is reasonably bounded in all cases.

Sadly, there is no RDSS instruction, so we can't actually use shadow
stack reads to spot if we underflowed the shadow stack, and there is no
useful alternative to panic() if we fail to find the shadow IRET frame.

>> --- a/xen/arch/x86/x86_64/entry.S
>> +++ b/xen/arch/x86/x86_64/entry.S
>> @@ -708,7 +708,16 @@ exception_with_ints_disabled:
>>          call  search_pre_exception_table
>>          testq %rax,%rax                 # no fixup code for faulting EIP?
>>          jz    1b
>> -        movq  %rax,UREGS_rip(%rsp)
>> +        movq  %rax,UREGS_rip(%rsp)      # fixup regular stack
>> +
>> +#ifdef CONFIG_XEN_SHSTK
>> +        mov    $1, %edi
>> +        rdsspq %rdi
>> +        cmp    $1, %edi
>> +        je     .L_exn_shstk_done
>> +        wrssq  %rax, (%rdi)             # fixup shadow stack
>> +.L_exn_shstk_done:
>> +#endif
> Again avoid the conditional jump by using alternatives patching?

Well - that depends on whether we're likely to gain any new content in
the pre exception table.

As it stands, it is only the IRET(s) to userspace so would be safe to
turn this into an unconditional alternative.  Even in the crash case, we
won't be returning to guest context after having started the crash
teardown path.

~Andrew