[U-Boot] FEC and EFI Simple Network

[Heinrich Schuchardt <xypron.glpk@gmx.de>]

On 2/23/19 3:23 PM, Alexander Graf wrote:
> 
> 
> On 02.12.18 22:42, Alexander Graf wrote:
>>
>>
>> On 29.03.18 06:02, Joe Hershberger wrote:
>>> Hi Patrick,
>>>
>>> On Wed, Mar 28, 2018 at 4:54 PM, Fabio Estevam <festevam@gmail.com> wrote:
>>>> Adding Joe in case he has some ideas.
>>>>
>>>> On Tue, Mar 27, 2018 at 9:12 AM, Patrick Wildt <patrick@blueri.se> wrote:
>>>>> Hi,
>>>>>
>>>>> I have been debugging network issues when running an EFI Application
>>>>> that uses the EFI Simple Network protocol on an i.MX6 machine (FEC).
>>>>>
>>>>> The symptom is that u-boot's (FEC) internal RX ring index is reset to 0,
>>>>> while the controller is still at idx 3 (or something else).  This is
>>>>> caused by the following circumstances:
>>>>>
>>>>> The Simple Network protocol offers methods like Start(), Stop(),
>>>>> Initialize(), Shutdown(), Reset().  Also the protocol has a state.  The
>>>>> states are Stopped, Started, Initialized.  The transition is as follows:
>>>>>
>>>>> Stopped ---Start()---> Started ---Initialize()--> Initialized
>>>>>
>>>>> Start() does some initialization, Initialize() allocates the TX/RX
>>>>> descriptors and actually kicks off the network engine.
>>>>>
>>>>> So far, only Initialize() is implemented in our u-boot interface, and it
>>>>> calls eth_init() which in the end calls fec_init().  Our network state
>>>>> is _always_ Started.  This means that EFI Applications see that that the
>>>>> state is Started and then call Initialize() to start the actual network
>>>>> traffic.  There is no call to Stop()/Shutdown()/Reset() as it's supposed
>>>>> to be in a sane state.
>>>>>
>>>>> In my case the FEC is already initialized since I booted using network
>>>>> and the RX desc index is already non-zero.  Now the EFI Application sees
>>>>> that the state is Started, calls Initialize() which makes u-boot call
>>>>> eth_init() which then calls fec_init().
>>>>>
>>>>> fec_init() does not reset the controller so that the controller-internal
>>>>> RX desc index is not reset to zero.  fec_init() calls fec_open() which
>>>>> then resets the driver-internal RX desc index to zero.  Now they are out
>>>>> of sync, boom.
>>>
>>> Would it be reasonable for fec_init to use a state variable to keep
>>> track of if it (and the HW) is already initialized and not call
>>> fec_open in that case? Also, fec_halt would need to update that state
>>> as well.
>>
>> Ping? Is this still an issue?
> 
> Last Ping.
> 
> 
> Alex
> 

With commit 0c5d2a3dac01 ("efi_loader: completely initialize network")
we changed efi_net_initialize() from only calling eth_init() to execute
the following sequence:

        /* Setup packet buffers */
        net_init();
        /* Disable hardware and put it into the reset state */
        eth_halt();
        /* Set current device according to environment variables */
        eth_set_current();
        /* Get hardware ready for send and receive operations */
        ret = eth_init();

So I am wondering if our additional call to eth_halt() shouldn't have
fixed the problem that Patrick originally reported.

Best regards

Heinrich