Markus Armbruster <arm...@redhat.com> writes:
[...]
> * "allwinner-a10"
>
> "-nodefaults -serial null -device allwinner-a10" doesn't explode,
> which means I can't exclude the possibility that this might actually
> do something useful for someone. I'd say we treat it just like
> "pc87312": leave alone now, document incompatible change.
[...]
> How could this kind of mistake could look like in NIC device models?
> "allwinner-a10"'s instance_init aw_a10_init() provides a clue: it messes
> with qemu_check_nic_model():
>
> if (nd_table[0].used) {
> qemu_check_nic_model(&nd_table[0], TYPE_AW_EMAC);
> qdev_set_nic_properties(DEVICE(&s->emac), &nd_table[0]);
> }
>
> No other device does that. Without -nodefaults, this fails:
>
> -device allwinner-a10: Unsupported NIC model: xgmac
Even better:
$ qemu-system-arm -S -M highbank -monitor stdio
QEMU 2.2.90 monitor - type 'help' for more information
(qemu) device_add allwinner-a10
Unsupported NIC model: xgmac
[Exit 1 ]
exit() in monitor command, big no-no.
Machine type doesn't matter. To get past this, nd_table[0] must have
model=allwinner-emac. With -net nic,model=allwinner-emac, it has, and
we get to the next round of failures:
* cubieboard
Board already creates an allwinner-a10, and creating a second one with
-device or device_add aborts.
* Any other board with an onboard NIC, e.g. highbank
Refuses to start, as board doesn't support model=allwinner-emac:
qemu-system-arm: Unsupported NIC model: allwinner-emac
* Any other board without an onboard NIC, e.g. virt
Warns on startup:
Warning: requested NIC (anonymous, model allwinner-emac) was not created
(not supported by this machine?
-device / device_add allwinner-a10 succeeds, as long as serial_hds[0]
is set. If you suppress it with -nodefaults, -device / device_add
exit()s (no-no again) with
Can't create serial device, empty char device
If the board has an onboard serial (e.g. collie), both the onboard
serial and the allwinner-a10 get are now connected to it. Not going
to work. We normally catch this error, but bypassing qdev properties
also bypasses the check.
Actual use of -device / device_add allwinner-a10 seems vanishingly
unlikely. Thus, setting cannot_instantiate_with_device_add_yet is
unlikely to break anything. Let's do it to mask the "device_add
exit()s" bug.
> I haven't checked all uses of nd_table[].
I have now, this is the only offender.
