Re: Regression: memory corruption on Atmel SAMA5D31

[Tudor.Ambarus]

On 6/30/22 12:23, Tudor.Ambarus@xxxxxxxxxxxxx wrote:
> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> 
> On 6/30/22 08:20, Peter Rosin wrote:
>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
>>
>> Hi!
> 
> Hi, Peter!
>>
>> 2022-06-27 at 18:53, Tudor.Ambarus@xxxxxxxxxxxxx wrote:
>>> On 6/27/22 15:26, Tudor.Ambarus@xxxxxxxxxxxxx wrote:
>>>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
>>>>
>>>> On 6/21/22 13:46, Peter Rosin wrote:
>>>>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
>>>>>
>>>>> 2022-06-20 at 16:22, Tudor.Ambarus@xxxxxxxxxxxxx wrote:
>>>>>>
>>>>>>>
>>>>>>> git@xxxxxxxxxx:ambarus/linux-0day.git, branch dma-regression-hdmac-v5.18-rc7-4th-attempt
>>>>>>>
>>>>>>
>>>>>> Hi, Peter,
>>>>>>
>>>>>> I've just forced pushed on this branch, I had a typo somewhere and with that fixed I could
>>>>>> no longer reproduce the bug. Tested for ~20 minutes. Would you please test last 3 patches
>>>>>> and tell me if you can still reproduce the bug?
>>>>>
>>>>> Hi!
>>>>>
>>>>> I rebased your patches onto my current branch which is v5.18.2 plus a few unrelated
>>>>> changes (at least they are unrelated after removing the previous workaround to disable
>>>>> nand-dma entirely).
>>>>>
>>>>> The unrelated patches are two backports so that drivers recognize new compatibles [1][2],
>>>>> which should be completely harmless, plus a couple of proposed fixes that happens to fix
>>>>> eeprom issues with the at91 I2C driver from Codrin Ciubotariu [3].
>>>>>
>>>>> On that kernel, I can still reproduce. It seems a bit harder to reproduce the problem now
>>>>> though. If the system is otherwise idle, the sha256sum test did not reproduce in a run of
>>>>> 150+ attempts, but if I let the "real" application run while I do the test, I get a failure rate
>>>>> of about 10%, see below. The real application burns some CPU (but not all of it) and
>>>>> communicates with HW using I2C, native UARTs and two of the four USB-serial ports
>>>>> (FTDI, with the latency set to 1ms as mentioned earlier), so I guess there is more DMA
>>>>> pressure or something? There is a 100mbps network connection, but it was left "idle"
>>>>> during this test.
>>>>>
>>>>
>>>> Thanks, Peter.
>>>> I got back to the office, I'm rechecking what could go wrong.
>>>>
>>>
>>> Hi, Peter,
>>>
>>> Would you please help me with another round of testing? I have difficulties
>>> in reproducing the bug and maybe you can speed up the process while I copy
>>> your testing setup. I made two more patches on top of the same branch [1].
>>> My assumption is that the last problem that you saw is that a transfer
>>> could be started multiple times. I think these are the last less invasive
>>> changes that I try, I'll have to rewrite the logic anyway.
>>>
>>> Thanks!
>>>
>>> [1] To github.com:ambarus/linux-0day.git
>>>    cbb2ddca4618..79c7784dbcf2  dma-regression-hdmac-v5.18-rc7-4th-attempt -> dma-regression-hdmac-v5.18-rc7-4th-attempt
>>
>> I was out of office, but I managed to get a test running over night and can
>> report that It still fails. This is a longer run of about 500 with a failure
>> rate of 5% compared to the last time when the failure rate was 10%. I tend
> 
> Thanks!
> 
>> to think that the observed difference in failure rate may well be statistical
>> noise, but who knows? Would it be useful with a longer run without the last
>> two patches to see if they make a difference?

I forgot to answer, sorry. No, not needed as it still fails.
> 
> I pushed another patch were I added a write mem barrier to make sure everything
> is in place before starting the transfer. Could you also take the last patch
> and re-test if it's not too complicated? I still can't reproduce it on my side,
> I'm checking what else I can add to stress test the DMA.

I could reproduce the bug even with the wmb(). I'm rechecking what I missed.

Cheers,
ta