[PATCH] KVM: PPC: Book3S HV: Treat unrecognized TM instructions as illegal

[Segher Boessenkool]

On Mon, Feb 17, 2020 at 12:07:31PM +1100, Michael Neuling wrote:
> On Thu, 2020-02-13 at 10:15 -0500, Gustavo Romero wrote:
> > On P9 DD2.2 due to a CPU defect some TM instructions need to be emulated by
> > KVM. This is handled at first by the hardware raising a softpatch interrupt
> > when certain TM instructions that need KVM assistance are executed in the
> > guest. Some TM instructions, although not defined in the Power ISA, might
> > raise a softpatch interrupt. For instance, 'tresume.' instruction as
> > defined in the ISA must have bit 31 set (1), but an instruction that
> > matches 'tresume.' OP and XO opcodes but has bit 31 not set (0), like
> > 0x7cfe9ddc, also raises a softpatch interrupt, for example, if a code
> > like the following is executed in the guest it will raise a softpatch
> > interrupt just like a 'tresume.' when the TM facility is enabled:
> > 
> > int main() { asm("tabort. 0; .long 0x7cfe9ddc;"); }

> > and then treats the executed instruction as 'nop' whilst it should actually
> > be treated as an illegal instruction since it's not defined by the ISA.
> 
> The ISA has this: 
> 
>    1.3.3 Reserved Fields, Reserved Values, and Reserved SPRs
> 
>    Reserved fields in instructions are ignored by the pro-
>    cessor.
> 
> Hence the hardware will ignore reserved bits. For example executing your little
> program on P8 just exits normally with 0x7cfe9ddc being executed as a NOP.
> 
> Hence, we should NOP this, not generate an illegal.

It is not a reserved bit.

The IMC entry for it matches op1=011111 op2=1////01110 presumably, which
catches all TM instructions and nothing else (bits 0..5 and bits 21..30).
That does not look at bit 31, the softpatch handler has to deal with this.

Some TM insns have bit 31 as 1 and some have it as /.  All instructions
with a "." in the mnemonic have bit 31 is 1, all other have it reserved.
The tables in appendices D, E, F show tend. and tsr. as having it
reserved, which contradicts the individual instruction description (and
does not make much sense).  (Only tcheck has /, everything else has 1;
everything else has a mnemonic with a dot, and does write CR0 always).


Segher