Re: [Qemu-devel] [PATCH v2 09/19] exec.c: Use cpu_get_phys_page_attrs_debug

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/01/2016 15:18, Peter Maydell wrote:
> On 11 January 2016 at 13:47, Paolo Bonzini <pbonzini@redhat.com> wrote:
>>
>>
>> On 16/11/2015 15:05, Peter Maydell wrote:
>>> -    hwaddr phys = cpu_get_phys_page_debug(cpu, pc);
>>> +    MemTxAttrs attrs = {};
>>> +    hwaddr phys = cpu_get_phys_page_attrs_debug(cpu, pc, &attrs);
>>> +    int asidx = cpu_asidx_from_attrs(cpu, attrs);
>>>      if (phys != -1) {
>>> -        tb_invalidate_phys_addr(cpu->as,
>>> +        tb_invalidate_phys_addr(cpu->cpu_ases[asidx].as,
>>>                                  phys | (pc & ~TARGET_PAGE_MASK));
>>>      }
>>
>> The only question I have is whether it is right to have empty MemTxAttrs
>> here.  Is there any way to use the MemTxAttrs for the "current state" of
>> the CPU, whatever that is (on x86 I have added cpu_get_mem_attrs to
>> target-i386/cpu.h)?
> 
> That's what the call to cpu_get_phys_page_attrs_debug() does:
> it fills in the MemTxAttrs :secure and :user fields, and then
> cpu_asidx_from_attrs() uses the returned attributes to pick
> the right address space. (cpu_get_phys_page_attrs_debug()
> only fills in attribute fields it knows about, which is why
> we pass it an empty attrs struct to start with.)

Oops, that's not clear from the documentation in patch 4.  But if that
was the case, shouldn't cpu_get_phys_page_attrs_debug leave *attrs
completely alone when using the fallback?

I think it's clearer if you pass uninitialized MemTxAttrs to
cpu_get_phys_page_attrs_debug, assuming you can do it.  I can see why
the current semantics make sense for helper.c's get_phys_addr, but I
think they are confusing for the topmost function call.

Paolo