Re: [Qemu-devel] [PATCH v4 1/8] fw_cfg: max access size and region size are the same for MMIO data reg

[Paolo Bonzini <pbonzini@redhat.com>]

On 16/12/2014 21:06, Laszlo Ersek wrote:
> On 12/16/14 20:49, Paolo Bonzini wrote:
>> fw_cfg_read (and
>> thus fw_cfg_data_mem_read) is not idempotent.  The split/compose stuff
>> accesses the bytes at offsets 8,9,10,11,12,13,14,15 and composes them
>> according to the endianness.
>>
>> In the case of fw_cfg it just retrieves 8 bytes, but in the case of host
>> big endian it reads them in the "wrong" order for some reason (sorry, I
>> haven't looked at this thoroughly).
> 
> I can't imagine how that would happen; fw_cfg_data_mem_read() ignores
> both "addr" and "size", and fw_cfg_read() simply advances the
> "cur_offset" member.

Honestly neither can I.  But still the automatic splitting (which is
even tested by tests/endianness-test.c :)) assumes idempotency of the
components and it's not entirely surprising that it somehow/sometimes
breaks if you don't respect that.

>> So the solution is:
>>
>> 1) make fw_cfg_data_mem_ops DEVICE_LITTLE_ENDIAN
>>
>> 2) make fw_cfg_data_mem_read and fw_cfg_data_mem_write call fw_cfg_read
>> and fw_cfg_write SIZE times and build up a value from the lowest byte up.
> 
> Nonetheless, that's a really nice idea! I got so stuck with the
> automatic splitting that I forgot about the possibility to act upon the
> "size" parameter in fw_cfg_data_mem_read(). Thanks!
> 
> ... Another thing that Andrew mentioned but I didn't cover in my other
> email -- what about fw_cfg_ctl_mem_ops? It's currently DEVICE_NATIVE_ENDIAN.
> 
> You flipped the combined ops to LE in commit 6fdf98f2 (and, apparently,
> I reviewed it). Shouldn't we do the same for the standalone selector?

No.  The standalone selector is used as MMIO, and the BE platforms
expect the platform to be big-endian.  The combined ops are only used on
ISA ports, where the firmware expects them to be little-endian (as
mentioned in the commit message).

That said, the standalone selector is used by BE platforms only, so we
know that the standalone selector is always DEVICE_BIG_ENDIAN.

So if you want, you can make the standalone selector and the standalone
datum BE and swap them in the firmware.  If the suggestion doesn't make
you jump up and down, I understand that. :)

Paolo