Re: [RFC PATCH 0/5] Improve PPC instruction emulation

[Alexander Graf <agraf@suse.de>]

On 19.07.14 12:14, Paul Mackerras wrote:
> This series aims to increase the range of instructions that KVM on PPC
> can emulate and reduce code duplication by using the existing
> instruction emulation code from arch/powerpc/lib/sstep.c for KVM.
>
> The ultimate goal is to make PR KVM run faster on the kind of
> instruction sequences that we get in Linux's first-level interrupt
> handlers, where we have privileged instructions such as move to/from
> SPR, mtmsrd, rfid, etc., intermingled with ordinary unprivileged
> loads, stores, arithmetic instructions, etc.  If KVM could emulate
> those ordinary instructions as well as the privileged ones, we could
> emulate these code sequences without incurring the cost to exit and
> re-enter the guest for every single privileged instruction.  That
> would be a speedup provided the guest entry/exit cost was greater than
> the cost of emulating a few ordinary instructions.
>
> This series doesn't get to that ultimate goal but does lay the
> groundwork.  It splits the emulate_step() function into two parts,
> analyse_instr() and emulate_step(), and uses analyse_instr() in
> kvmppc_emulate_instruction().  This means that KVM needs to store its
> vcpu integer register state in a struct pt_regs like the rest of the
> kernel does.  We also need to make kvmppc_handle_load() and
> kvmppc_handle_store() handle loads and stores to ordinary guest memory
> as well as emulated MMIO.

Please take a look at my other patch set that implemented instruction 
emulation. There we split the code paths between MMIO emulation and 
normal instruction emulation.

I really think that approach is a prerequisite to doing full instruction 
emulation in longer code snippets. Obviously the generic load/store 
should then handle MMIO as well as generic memory operations.


Alex