On 09.03.2026 23:25, Andrew Cooper wrote:
> On 27/02/2026 11:16 pm, Andrew Cooper wrote:
>> diff --git a/xen/arch/x86/traps.c b/xen/arch/x86/traps.c
>> index 2f40f628cbff..e2c35a046e6b 100644
>> --- a/xen/arch/x86/traps.c
>> +++ b/xen/arch/x86/traps.c
>> ...
>> + case 2: /* SYSENTER */
>> + {
>> + /*
>> + * FRED delivery preserves the interrupted state, but previously
>> + * SYSENTER discarded almost everything.
>> + *
>> + * The guest isn't aware of FRED, so recreate the legacy
>> + * behaviour.
>> + *
>> + * When setting the selectors, clear all upper metadata. In
>> + * particular fred_ss.swint becomes pend_DB on ERETx.
>> + *
>> + * When converting to a fault, hardware finally gives us enough
>> + * information to account for prefixes, so provide the more
>> + * correct behaviour rather than assuming the instruction was
>> two
>> + * bytes long.
>> + */
>> + unsigned int len = regs->fred_ss.insnlen;
>> +
>> + regs->ssx = FLAT_USER_SS;
>> + regs->rsp = 0;
>> + regs->eflags &= ~(X86_EFLAGS_VM | X86_EFLAGS_IF);
>> + regs->csx = 3;
>> + regs->rip = 0;
>> +
>> + if ( !curr->arch.pv.sysenter_callback_eip )
>> + {
>> + regs->rip -= len;
>> + pv_inject_hw_exception(X86_EXC_GP, 0);
>> + }
>> + else
>> + pv_inject_callback(CALLBACKTYPE_sysenter);
>> + break;
>
> This isn't actually a correct transformation of the IDT code. When the
> SYENTER entrypoint isn't registered, this delivers a #GP at
> 0003:fffffffffffffffe
>
> The simple fix to get back to IDT behaviour is to simply drop the
> subtraction of len.
>
> In FRED mode, we can finally point the #GP at the SYSENTER instruction,
> rather than delivering at 0. We could even provide the success case
> pointing sensibly too.
>
> The question is should we? Until now, the differences between FRED and
> IDT mode are minimal. This would be major difference, and it's for
> SYSENTER which all but unused. I'm erring on the side of "match IDT".
I agree. Down the road we could introduce an opt-in "better behavior" mode
when running under FRED (also covering other aspects previously discussed).
Jan
