On 5/1/21 10:12 PM, Guenter Roeck wrote:
> On 4/30/21 9:28 PM, Bin Meng wrote:
>> On Fri, Apr 30, 2021 at 10:41 PM Guenter Roeck <[email protected]> wrote:
>>>
>>> Hi,
>>>
>>> On Wed, Jan 06, 2021 at 02:35:03PM +0800, Bin Meng wrote:
>>>> From: Bin Meng <[email protected]>
>>>>
>>>> At present, when booting U-Boot on QEMU sabrelite, we see:
>>>>
>>>> Net: Board Net Initialization Failed
>>>> No ethernet found.
>>>>
>>>> U-Boot scans PHY at address 4/5/6/7 (see board_eth_init() in the
>>>> U-Boot source: board/boundary/nitrogen6x/nitrogen6x.c). On the real
>>>> board, the Ethernet PHY is at address 6. Adjust this by updating the
>>>> "fec-phy-num" property of the fsl_imx6 SoC object.
>>>>
>>>> With this change, U-Boot sees the PHY but complains MAC address:
>>>>
>>>> Net: using phy at 6
>>>> FEC [PRIME]
>>>> Error: FEC address not set.
>>>>
>>>> This is due to U-Boot tries to read the MAC address from the fuse,
>>>> which QEMU does not have any valid content filled in. However this
>>>> does not prevent the Ethernet from working in QEMU. We just need to
>>>> set up the MAC address later in the U-Boot command shell, by:
>>>>
>>>> => setenv ethaddr 00:11:22:33:44:55
>>>>
>>>
>>> With this patch in place, the standard Ethernet interface no longer works
>>> when
>>> booting sabrelite Linux images directly (without u-boot) using the following
>>> qemu command.
>>> qemu-system-arm -M sabrelite -kernel arch/arm/boot/zImage
>>> ...
>>>
>>> The Ethernet interface still instantiates, but packet transfer to the host
>>> no longer works. Reverting this patch fixes the problem for me.
>>>
>>> Is there a qemu command line parameter that is now necessary to instantiate
>>> the Ethernet interface when booting Linux ?
>>
>> Enabling "guest_errors" shows that Linux kernel fec driver is trying
>> to read PHY at address 0, which is not what we want.
>>
>> [imx.fec.phy]imx_phy_read: Bad phy num 0
>>
>> The device tree blob of the sabrelite does not contain a node for the
>> ethernet phy specifying phy address, so I suspect Linux kernel driver
>> is using default phy address 0 instead.
>>
>> Could you please test on a real hardware to see what happens?
>>
>
> The problem is that qemu returns 0 when the OS tries to read from a
> non-existing PHY. Linux expects it to return 0xffff, and believes that
> a PHY is there if 0 is returned.
Correct.
> This helps:
>
> diff --git a/hw/net/imx_fec.c b/hw/net/imx_fec.c
> index f03450c028..3c90c38e26 100644
> --- a/hw/net/imx_fec.c
> +++ b/hw/net/imx_fec.c
> @@ -285,7 +285,7 @@ static uint32_t imx_phy_read(IMXFECState *s, int reg)
> if (phy != s->phy_num) {
> qemu_log_mask(LOG_GUEST_ERROR, "[%s.phy]%s: Bad phy num %u\n",
> TYPE_IMX_FEC, __func__, phy);
> - return 0;
> + return 0xffff;
> }
>
> Note that this is not really a guest error; any OS can and likely
> will scan the MII bus for connected phy chips.
Correct. This should be a trace event instead.
