On Tue, May 06, 2025 at 09:31:00AM +0200, Peter Zijlstra wrote: > On Sat, May 03, 2025 at 11:28:37AM -0700, Josh Poimboeuf wrote: > > On Sat, May 03, 2025 at 11:50:23AM +0200, Peter Zijlstra wrote: > > > > +++ b/arch/x86/entry/entry_64_fred.S > > > > @@ -116,7 +116,8 @@ SYM_FUNC_START(asm_fred_entry_from_kvm) > > > > movq %rsp, %rdi /* %rdi -> pt_regs */ > > > > call __fred_entry_from_kvm /* Call the C entry point */ > > > > POP_REGS > > > > - ERETS > > > > + > > > > + ALTERNATIVE "", __stringify(ERETS), X86_FEATURE_FRED > > > > 1: > > > > /* > > > > * Objtool doesn't understand what ERETS does, this hint tells it that > > > > @@ -124,7 +125,7 @@ SYM_FUNC_START(asm_fred_entry_from_kvm) > > > > * isn't strictly needed, but it's the simplest form. > > > > */ > > > > UNWIND_HINT_RESTORE > > > > - pop %rbp > > > > + leave > > > > RET > > > > > > So this, while clever, might be a problem with ORC unwinding. Because > > > now the stack is different depending on the alternative, and we can't > > > deal with that. > > > > > > Anyway, I'll go have a poke on Monday (or Tuesday if Monday turns out to > > > be a bank holiday :-). > > > > Can we just adjust the stack in the alternative? > > > > ALTERNATIVE "add $64 %rsp", __stringify(ERETS), X86_FEATURE_FRED > > Yes, that should work. Nope, it needs to be "mov %rbp, %rsp". Because that is the actual rsp value after ERETS-to-self. > But I wanted to have a poke at objtool, so it > will properly complain about the mistake first. So a metric ton of fail here :/ The biggest problem is the UNWIND_HINT_RESTORE right after the alternative. This ensures that objtool thinks all paths through the alternative end up with the same stack. And hence won't actually complain. Second being of course, that in order to get IRET and co correct, we'd need far more of an emulator. Also, it actually chokes on this variant, and I've not yet figured out why. Whatever state should be created by that mov, the restore hint should wipe it all. But still the ORC generation bails with unknown base reg -1.
