Re: [Qemu-devel] [PATCH RFC] s390x/kvm: call cpu_synchronize_state() on every kvm_arch_handle_exit()

[Christian Borntraeger]

On 03/26/2018 11:20 AM, David Hildenbrand wrote:
> Manually having to use cpu_synchronize_state() is error prone. I don't
> think that the performance impact is that huge if we simply synchronize
> the state on every kvm_arch_handle_exit() call. This makes the code
> easier to maintain.
> 
> We now also call it (although not neded) for
> - KVM_EXIT_S390_RESET -> s390_reipl_request()
> - KVM_EXIT_DEBUG -> kvm_arch_handle_debug_exit()
> - unmanagable/unimplemented intercepts
> - ICPT_WAITPSW -> s390_handle_wait() -> cpu gets halted
> - ICPT_CPU_STOP -> do_stop_interrupt() -> cpu gets halted
> - Scenarios where we inject an operation exception
> - handle_sigp() on the source CPU
> - handle_stsi()
> 
> I don't think any of these are performance critical. Especially as we
> have all information directly contained in kvm_run, there are no
> additional IOCTLs to issue on modern kernels.

We had other issues in the past in other (common code) places. For example
see

commit 79ca7a1b898eb97c4192f3c78027a0f20485e7b4
Author:     Christian Borntraeger <address@hidden>
AuthorDate: Tue Mar 7 15:19:08 2017 +0100
Commit:     Paolo Bonzini <address@hidden>
CommitDate: Tue Mar 14 13:26:36 2017 +0100

    exec: add cpu_synchronize_state to cpu_memory_rw_debug

so we might consider going even further.....But this will be tricky.


FWIW, I think your patch even fixes a bug:

--- snip ----
static int handle_diag(S390CPU *cpu, struct kvm_run *run, uint32_t ipb)
{
    int r = 0;
    uint16_t func_code;

    /*
     * For any diagnose call we support, bits 48-63 of the resulting
     * address specify the function code; the remainder is ignored.
     */
    func_code = decode_basedisp_rs(&cpu->env, ipb, NULL) & DIAG_KVM_CODE_MASK;

---->
        static inline hwaddr decode_basedisp_s(CPUS390XState *env, uint32_t ipb,
                                               uint8_t *ar)
        {
            hwaddr addr = 0;
            uint8_t reg;
            
            reg = ipb >> 28;
            if (reg > 0) {
                addr = env->regs[reg];

----> we do the sync_regs after this in the diag handler!
So currently we only handle the case with base reg == 0 correctly.
So
diag x,y,0x500(0) 
works


but things like
lghi 1,0x500
diag x,y,0(1)

not unless I miss something.


[...]
> @@ -1778,6 +1760,8 @@ int kvm_arch_handle_exit(CPUState *cs, struct kvm_run 
> *run)
> 
>      qemu_mutex_lock_iothread();
> 
> +    cpu_synchronize_state(CPU(cpu));
> +
>      switch (run->exit_reason) {
>          case KVM_EXIT_S390_SIEIC:
>              ret = handle_intercept(cpu);
> 

Does it make sense to do this hunk NOW for 2.12 (cc stable)
and queue your full patch for 2.13?