bug in drivers/edac/mpc85xx_edac.c:mpc85xx_mc_check()

[Kumar Gala]

On Apr 29, 2009, at 2:37 AM, Andrew Morton wrote:

> Let's cc the suitable people.
>
> On Tue, 28 Apr 2009 18:23:42 -0700 "Jeff Haran" <jharan at Brocade.COM>  
> wrote:
>
>> Hi,
>>
>> Recent versions of this function contain the following snippets:
>>
>>    if (err_detect & DDR_EDE_SBE)
>>        edac_mc_handle_ce(mci, pfn, err_addr & PAGE_MASK,
>>                  syndrome, row_index, 0, mci->ctl_name);
>>
>>    if (err_detect & DDR_EDE_MBE)
>>        edac_mc_handle_ue(mci, pfn, err_addr & PAGE_MASK,
>>                  row_index, mci->ctl_name);
>>
>> I am pretty sure the references to PAGE_MASK should be proceeded by a
>> tilda, as in:
>>
>>    if (err_detect & DDR_EDE_SBE)
>>        edac_mc_handle_ce(mci, pfn, err_addr & ~PAGE_MASK,
>>                  syndrome, row_index, 0, mci->ctl_name);
>>
>>    if (err_detect & DDR_EDE_MBE)
>>        edac_mc_handle_ue(mci, pfn, err_addr & ~PAGE_MASK,
>>                  row_index, mci->ctl_name);
>>
>
> Could well be.  PAGE_MASK is very easy to get wrong.  I've _never_
> trusted my own memory of it and I always have to go back to the
> definition when reviewing code :(

This should ~PAGE_MASK to get the offset into the page.

>> Much as I would like to submit a tested patch like the rest of the
>> world, I find myself in the situation where the only Freescale target
>> system I have to test on is running a 3 year old kernel (2.6.14),  
>> which
>> preceeds the introduction of EDAC driver support, at least for
>> Freescale. So the best I can do is borrow from the new EDAC driver  
>> and
>> backport it to the old kernel.
>>
>> But I have learned a few things in this process and can thus share  
>> what
>> I've learned as it may be of help to the EDAC driver developers:
>>
>> 1) Before you read the Freescale 8548 CAPTURE_ADDRESS register, you  
>> want
>> to read CAPTURE_ATTRIBUTES first and make sure the VLD bit (least
>> significant bit in the register) is set or else the data in
>> CAPTURE_ADDRESS may not be yet valid.
>>
>> 2) When you are done scrubbing the memory with the single bit  
>> error, you
>> want to write 0 to CAPTURE_ATTRIBUTES so as to clear VLD and thus  
>> setup
>> the ECC capture logic to capture the next single bit error.

This is a correct description based on how FSL error HW works.

- k