Re: [PATCH 3/5] hw/pci-host/prep: Remove unuseful memory region mapping

[Philippe Mathieu-Daudé]

On 3/22/21 5:09 AM, David Gibson wrote:
> On Fri, Mar 12, 2021 at 03:38:21PM -0500, Peter Xu wrote:
>> On Fri, Mar 12, 2021 at 07:28:49PM +0100, Philippe Mathieu-Daudé wrote:
>>> The pci_io_non_contiguous region is mapped on top of pci_io
>>> with higher priority, but simply dispatch into this region
>>> address space. Simplify by directly registering the former
>>> region in place, and adapt the address space dispatch offsets.
>>>
>>> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
>>> ---
>>>  hw/pci-host/prep.c | 11 ++++-------
>>>  1 file changed, 4 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/hw/pci-host/prep.c b/hw/pci-host/prep.c
>>> index 0a9162fba97..00a28c2d18c 100644
>>> --- a/hw/pci-host/prep.c
>>> +++ b/hw/pci-host/prep.c
>>> @@ -159,8 +159,7 @@ static uint64_t raven_io_read(void *opaque, hwaddr addr,
>>>      uint8_t buf[4];
>>>  
>>>      addr = raven_io_address(s, addr);
>>> -    address_space_read(&s->pci_io_as, addr + 0x80000000,
>>> -                       MEMTXATTRS_UNSPECIFIED, buf, size);
>>> +    address_space_read(&s->pci_io_as, addr, MEMTXATTRS_UNSPECIFIED, buf, 
>>> size);
>>>  
>>>      if (size == 1) {
>>>          return buf[0];
>>> @@ -191,8 +190,7 @@ static void raven_io_write(void *opaque, hwaddr addr,
>>>          g_assert_not_reached();
>>>      }
>>>  
>>> -    address_space_write(&s->pci_io_as, addr + 0x80000000,
>>> -                        MEMTXATTRS_UNSPECIFIED, buf, size);
>>> +    address_space_write(&s->pci_io_as, addr, MEMTXATTRS_UNSPECIFIED, buf, 
>>> size);
>>
>> This changes access to s->pci_io_as, but below didn't change s->pci_io_as
>> layout at all (below is about address_space_mem).  Is this intended?
>>
>>>  }
>>>  
>>>  static const MemoryRegionOps raven_io_ops = {
>>> @@ -294,9 +292,8 @@ static void raven_pcihost_initfn(Object *obj)
>>>      address_space_init(&s->pci_io_as, &s->pci_io, "raven-io");
>>>  
>>>      /* CPU address space */
>>> -    memory_region_add_subregion(address_space_mem, 0x80000000, &s->pci_io);
>>> -    memory_region_add_subregion_overlap(address_space_mem, 0x80000000,
>>> -                                        &s->pci_io_non_contiguous, 1);
>>> +    memory_region_add_subregion(address_space_mem, 0x80000000,
>>> +                                &s->pci_io_non_contiguous);
>>
>> I don't know any of this code at all... but it seems the two memory regions 
>> are
>> not identical in size:
>>
>>     memory_region_init(&s->pci_io, obj, "pci-io", 0x3f800000);
>>     memory_region_init_io(&s->pci_io_non_contiguous, obj, &raven_io_ops, s,
>>                           "pci-io-non-contiguous", 0x00800000);
>>
>> Then it seems the memory access dispatching to (0x00800000, 0x3f800000) would
>> change too, from s->pci_io to nothing.  Raise this up too since I don't know
>> either whether it's intended..
> 
> Right, it seems like this removes the mapping of s->pci_io entirely.

Yes, this is on purpose. The dispatching is done via raven_io_ops:

    address_space_init(&s->pci_io_as, &s->pci_io, "raven-io");

raven_io_ops uses raven_io_address() which returns an address in
"pci-io" range if < 0x800000, else it returns io_mem_unassigned
because the address is outside the pci_io_as AddressSpace.

Note, the flatview is unchanged with this patch.

My take here is 1/ I squashed too much changes and 2/ I didn't
documented enough. I'll respin improved.

Thanks for your reviews!

Phil.