On 3/11/21 6:27 PM, Peter Xu wrote:
> On Thu, Mar 11, 2021 at 05:21:49PM +0100, Philippe Mathieu-Daudé wrote:
>> +Aspeed team
>>
>> On 3/11/21 1:18 PM, Philippe Mathieu-Daudé wrote:
>>> On 3/10/21 9:29 PM, Peter Xu wrote:
>>
>>>> Yeah no worry - it's just that I feel one memory_region_init_alias() call
>>>> is
>>>> probably missing in your huge series somewhere, so that you'll take that
>>>> alias
>>>> MR as subregion rather than the real MR (which is the root of one AS).
>>>
>>> OK, with your earlier comments start + Mark other comment I start
>>> to understand better.
>>>
>>> So far:
>>>
>>> (1a) AddressSpace is a physical view, its base address must be zero
>>>
>>> (1b) AddressSpace aperture is fixed (depends on hardware design,
>>> not changeable at runtime
>>>
>>> Therefore due to (1a):
>>> (2) AddressSpace root MemoryRegion is a container and must not be
>>> mmio-mapped anywhere (in particular not on SysBus).
>>>
>>> (3) If hardware has a MMIO view of an AddressSpace, it has to be
>>> via a MemoryRegion alias. That way the alias handles paddr offset
>>> adjustment to the zero-based AddressSpace root container MR.
>>> Aliasing allows resizing the alias size without modifying the AS
>>> aperture size (1b).
>>>
>>> I'll start adding assertions for (1a) and (2) in the code base and
>>> see if (3) adjustments are required.
>>
>> So using:
>>
>> -- >8 --
>> diff --git a/softmmu/memory.c b/softmmu/memory.c
>> index 874a8fccdee..8ce2d7f83b9 100644
>> --- a/softmmu/memory.c
>> +++ b/softmmu/memory.c
>> @@ -713,6 +713,12 @@ static MemoryRegion
>> *memory_region_get_flatview_root(MemoryRegion *mr)
>> continue;
>> }
>> }
>> + if (mr && mr->addr) {
>> + error_report("Detected flatview root memory region '%s' with"
>> + " non-zero base address (0x%"HWADDR_PRIx"):
>> aborting",
>> + memory_region_name(mr), mr->addr);
>> + abort();
>> + }
>>
>> return mr;
>> }
>> ---
>
> Maybe it works, but it looks a bit odd to test here. What I meant was
> something like attached.
>
> Maybe it's still legal to make the root mr a subregion of another, so maybe
> I'm
> completely wrong... then the patch attached won't make any sense either.
Maybe it does, and we need to rework some boards code.
With your patch applied:
$ ./qemu-system-ppc -M 40p
qemu-system-ppc: softmmu/memory.c:2445:
memory_region_add_subregion_common: Assertion `!subregion->is_root_mr'
failed.
Aborted (core dumped)
$ ./qemu-system-arm -M ast2600-evb
qemu-system-arm: softmmu/memory.c:2445:
memory_region_add_subregion_common: Assertion `!subregion->is_root_mr'
failed.
Aborted (core dumped)
> It's
> just that in my mind each MR should have a "parent" - for normal MR it's the
> container MR, then for root MR it's easier to see the AS as its "parent".
>
> Maybe Paolo could clarify this..
>
> Thanks,
>
