On 26/01/2020 02:28, Matthias Brugger wrote:
On 24/01/2020 01:26, André Przywara wrote:
[ ... ]
Found the culprit, after following a lead started by an over-lunch discussion: Colleagues pointed out the SError (interrupts) early in the kernel could just show because they just got unmasked when dropping into EL1. And indeed in AArch64 U-Boot we keep Aborts masked - we don't clear the A bit in DAIF normally (only for Freescale). So allowing SError exceptions in U-Boot's start.S revealed that the SError interrupt was actually triggered by the writel in write_hwaddr(), I guess because the MAC wasn't reset before. And the SError condition stayed pending all the time, until the kernel announced its interest in being told about fatal errors - then it inherited U-Boot's error.
Thanks for the explanation. I think the situation leaves space for improving. Either should we warn about a pending Abort before leaving U-Boot or we should allow aborts in general.
Definitively we should unmasks SErrors in U-Boot, since they point us to serious problems, with this one here actually being somewhat on the harmless side. Also U-Boot has exception handlers that dump useful information, so we should use them.
But doing so would need to be done for all ARMv8 ports (in start.S), so I am a bit reluctant to post something this late in the merge window without proper testing on multiple platforms.
So for me the issue is fixed after adding the reset routine I sketched in that thread before.
But you mentioned that it still didn't work for you?
I just double checked and everything works fine. Please feel free to send a new version :)
Great, thanks! Did just that.
Cheers, Andre