Re: [PATCH] ARM: BCM5301X: Add back handler ignoring external imprecise aborts

[Scott Branden <scott.branden@broadcom.com>]

On 16-10-31 02:56 PM, Florian Fainelli wrote:
> On 10/31/2016 02:46 PM, Scott Branden wrote:
>>
>>
>> On 16-10-31 02:01 PM, Florian Fainelli wrote:
>>> On 10/31/2016 01:59 PM, Hauke Mehrtens wrote:
>>>>
>>>>
>>>> On 10/31/2016 07:08 PM, Scott Branden wrote:
>>>>> Hi Rafal,
>>>>>
>>>>>
>>>>> On 16-10-29 04:12 AM, Rafał Miłecki wrote:
>>>>>> From: Rafał Miłecki <rafal@milecki.pl>
>>>>>>
>>>>>> Since early BCM5301X days we got abort handler that was removed by
>>>>>> commit 937b12306ea79 ("ARM: BCM5301X: remove workaround imprecise
>>>>>> abort
>>>>>> fault handler"). It assumed we need to deal only with pending aborts
>>>>>> left by the bootloader. Unfortunately this isn't true for BCM5301X.
>>>>>>
>>>>>> When probing PCI config space (device enumeration) it is expected to
>>>>>> have master aborts on the PCI bus. Most bridges don't forward (or they
>>>>>> allow disabling it) these errors onto the AXI/AMBA bus but not the
>>>>>> Northstar (BCM5301X) one.
>>>>> Should we only add this workaround code if CONFIG_PCI is on then?
>>>>
>>>> I think all the supported northstar devices have a PCIe controller. We
>>>> could add such a CONFIG_PCI check, but I do not see a big advantage.
>>>
>>> Actually, I do see a couple disadvantages if we gate this with
>>> CONFIG_PCI: if this problem shows up irrespective of your kernel
>>> configuration, you want the error handler to clear it, not rely on
>>> CONFIG_PCI to be enabled for the error to go away and also, without an
>>> additional ifdef, additional compiler coverage.
>>>
>> A problem with reintroducing this change is that all imprecise data
>> aborts are ignored, not just PCI.  So if you don't actually use PCI in
>> your system and want to debug other aborts you are unable to.  I don't
>> know if we care about such situation.
>
> Considering that any abort is pretty much fatal, the options are:
>
> - update the freaking bootloader to a version where there are no such
> aborts generated, not an option on consumer devices, unclear which
> version (if any) fixes that
>
> - fixups the aborts externally, via a boot wrapper, which is going to
> take some time to develop, causes additional burden on the distributors
> to provide instructions/build images on how to do it
I think the abort is already fixed in the kernel now on boot so option 1 
and 2 were not needed - and abort handler was removed as detailed by 
Rafal in the commit.  Only outstanding issue is now this new PCI 
enumeration issue.
>
> - fixups the aborts in the kernel, irrespective of where they come from,
> simple and easy
>
> - fixups the aborts in the kernel, look where they come from, by using
> some bit of magic, looking at PCIe registers and whatnot (provided that
> is even possible), not too hard, but can take a while, and is error prone
Option 4 sounds like the proper solution - check the range the abort is 
due to the PCI device enumeration and only ignore those aborts.
>
> I can certainly advocate that option 3 is the one that gives a working
> device, and this is what matters from a distribution perspective like
> LEDE/OpenWrt.
>
Since I don't use BCM5301x option 3 is fine by me.