On 14/06/2019 23:14, Fabiano Rosas wrote:
> Mark Cave-Ayland <mark.cave-ayland@xxxxxxxxxxxx> writes:
>
>> QEMU's kvm_handle_debug() function identifies software breakpoints by checking
>> for a value of 0 in kvm_debug_exit_arch's status field. Since this field isn't
>
> LGTM, but let me start a discussion:
>
> In arch/powerpc/include/uapi/asm/kvm.h I see the following:
>
>
> /*
> * Defines for h/w breakpoint, watchpoint (read, write or both) and
> * software breakpoint.
> * These are used as "type" in KVM_SET_GUEST_DEBUG ioctl and "status"
> * for KVM_DEBUG_EXIT.
> */
> #define KVMPPC_DEBUG_NONE 0x0
> #define KVMPPC_DEBUG_BREAKPOINT (1UL << 1)
> #define KVMPPC_DEBUG_WATCH_WRITE (1UL << 2)
> #define KVMPPC_DEBUG_WATCH_READ (1UL << 3)
>
>
> this seems to be biased towards the BookE implementation which uses
> KVMPPC_DEBUG_BREAKPOINT to indicate a "hardware breakpoint" (i.e. Instruction
> Address Compare - ISA v2 Book III-E, section 10.4.1), and then
> KVMPPC_DEBUG_NONE ends up implicitly meaning "software breakpoint" for
> Book3s PR/HV.
>
>> explicitly set to 0 when the software breakpoint instruction is detected, any
>> previous non-zero value present causes a hang in QEMU as it tries to process
>> the breakpoint instruction incorrectly as a hardware breakpoint.
>
> What QEMU does (again biased towards BookE) is to check the 'status'
> field and treat both h/w breakpoints and watchpoints as hardware
> breakpoints (which is correct in a sense) and then proceed to look for
> s/w breakpoints:
>
> if (arch_info->status) {
> return kvm_handle_hw_breakpoint(cs, arch_info);
> }
>
> if (kvm_find_sw_breakpoint(cs, arch_info->address)) {
> return kvm_handle_sw_breakpoint(cs, arch_info->address);
> }
Right so this feels a bit like BookE was used for the original implementation and
then the Book3S functionality followed on from that.
> So I wonder if it would not be beneficial for future developers if we
> drew the line and made a clear distinction between:
>
> software breakpoints - triggered by a KVMPPC_INST_SW_BREAKPOINT execution
> hardware breakpoints - triggered by some register match
>
> Maybe by turning the first two defines into:
>
> #define KVMPPC_DEBUG_SW_BREAKPOINT 0x0
> #define KVMPPC_DEBUG_HW_BREAKPOINT (1UL << 1)
No objection from me, however I don't really have enough history with KVM PPC to be
able to opine as to whether this fits with the general design of
KVM/PPC/BookE/Book3S. I'm really coming at this from the perspective of toying with a
hobby project on my iMac G4 and fixing up the things I find are broken.
>> Ensure that the kvm_debug_exit_arch status field is set to 0 when the software
>> breakpoint instruction is detected (similar to the existing logic in booke.c
>> and e500_emulate.c) to restore software breakpoint functionality under Book3S
>> PR.
>>
>> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@xxxxxxxxxxxx>
>> ---
>
> Anyway, the proposed patch fixes the issue.
>
> Reviewed-by: Fabiano Rosas <farosas@xxxxxxxxxxxxx>
Great, thank you! Is this enough for the patch to be queued?
>> arch/powerpc/kvm/emulate.c | 1 +
>> 1 file changed, 1 insertion(+)
>>
>> diff --git a/arch/powerpc/kvm/emulate.c b/arch/powerpc/kvm/emulate.c
>> index bb4d09c1ad56..6fca38ca791f 100644
>> --- a/arch/powerpc/kvm/emulate.c
>> +++ b/arch/powerpc/kvm/emulate.c
>> @@ -271,6 +271,7 @@ int kvmppc_emulate_instruction(struct kvm_run *run, struct kvm_vcpu *vcpu)
>> */
>> if (inst == KVMPPC_INST_SW_BREAKPOINT) {
>> run->exit_reason = KVM_EXIT_DEBUG;
>> + run->debug.arch.status = 0;
>> run->debug.arch.address = kvmppc_get_pc(vcpu);
>> emulated = EMULATE_EXIT_USER;
>> advance = 0;
ATB,
Mark.
