Re: [Qemu-devel] RFC: memory API changes

[Paolo Bonzini <pbonzini@redhat.com>]

On 24/03/2015 19:06, Peter Maydell wrote:
> On 24 March 2015 at 17:51, Paolo Bonzini <pbonzini@redhat.com> wrote:
>> On 24/03/2015 17:35, Peter Maydell wrote:
>>> On 24 March 2015 at 16:23, Paolo Bonzini <pbonzini@redhat.com> wrote:
>>>>> On 24 March 2015 at 15:08, Paolo Bonzini <pbonzini@redhat.com> wrote:
>>>> , for those callers
>>>> of ld/st*_phys that use cs->as as the first argument.
>>>
>>> ...but I don't understand this caveat. I want to add arguments
>>> and rename the functions for *all* callers of ld/st*_phys.
>>> I don't want to specialcase the ones which happen to be
>>> operating on cs->as.
>>
>> The ones that operate on cs->as could become (for some CPUs at least)
>> special-cased accessors like the bus ones; for example building the
>> MemTxAttrs according to internal CPU state.
> 
> Sure, individual targets could do something like this if they
> wanted (compare the arm_ldl_code functions), once these renames
> have gone in.
> 
>> ld/st*_phys actually started as CPU-specific accessors, and most uses
>> are still of that kind, so it makes sense to me that we special-case
>> them.  Maybe it limits churn, maybe it doesn't.  But if it doesn't, it's
>> not like anything is lost.
> 
> I think this is where we disagree. I see ld/st*_phys as being
> really generic -- they take an AddressSpace, after all, and
> part of the same family with address_space_read &c. If you
> don't leave them as generic, then you end up having to use
> the really awkward _read/_write for simple accesses and
> then manage the byteswapping yourself. That's why I want
> to rename them into address_space_* : to indicate that they
> are all part of the same family, and you can use
> address_space_read if you want to read an arbitrary byte
> buffer, or address_space_ldl_be if you want to read a
> big endian 32 bit word, and so on.

I agree with that.  I just want to keep ld/st*_phys _in addition_ as the
short forms of address_space_ld/st*, and keep ld/st*_phys instead of
address_space_ld/st* for those uses that have cs->as as the first argument.

The rationale is to evolve ld/st*_phys into CPU-specific accessors
paralleling the bus-specific accessors.

Paolo

> (The only reason they started out CPU specific is because
> we didn't have any concept at all of having more than
> one address space, so there wasn't any need to say which
> one you meant when you were doing a load.)
> 
> To me it makes much more sense that if a DMA controller
> like pl080 wants to do an LE word read from the AS which
> its bus master is connected to, that it can just do
>  word = ldl_le_phys(my_as, addr, ...);
> 
> I'd expect pretty much any bus master to want to do this
> kind of thing, in fact. It just happens that most of the
> bus masters we have in QEMU right now are CPUs...
> 
> -- PMM
> 
>