[PATCH 3/9] arm64: mm: install SError abort handler

[Marc Zyngier]

On Fri, Mar 24 2017 at 07:02:05 PM, Florian Fainelli <f.fainelli at gmail.com> wrote:
> On 03/24/2017 11:31 AM, Mark Rutland wrote:
>> Hi Florian,
>> 
>> On Fri, Mar 24, 2017 at 10:53:48AM -0700, Florian Fainelli wrote:
>>> On 03/24/2017 10:35 AM, Mark Rutland wrote:
>>>> On Fri, Mar 24, 2017 at 09:48:40AM -0700, Doug Berger wrote:
>>>>> On 03/24/2017 08:16 AM, Mark Rutland wrote:
>>>>>> On Fri, Mar 24, 2017 at 07:46:26AM -0700, Doug Berger wrote:
>>>>
>>>>> If you would consider an alternative implementation where we scrap
>>>>> the SError handler (i.e. maintain the ugliness in our downstream
>>>>> kernel) in favor of a more gentle user mode crash on SError that
>>>>> allows the kernel the opportunity to service the interrupt for
>>>>> diagnostic purposes I could try to repackage that.
>>>>
>>>> If this is just for diagnostic purposes, I believe you can register a
>>>> panic notifier, which can then read from the bus. The panic will occur,
>>>> but you'll have the opportunity to log some information to dmesg.
>>>
>>> And crash the kernel? That sounds awful, FWIW the ARM/Linux kernel is
>>> able to recover just fine from user-space accessing e.g: invalid
>>> physical addresses in the GISB register space, bringing the same level
>>> of functionality to ARM64/Linux sounds reasonable to me.
>> 
>> I disagree, given that:
>> 
>> (a) You cannot determine the (HW) origin of the SError in an
>>     architecturally portable way. i.e. when you take an SError, you have
>>     no way of determining what asynchronous event caused it.
>> 
>> (b) SError is effectively an edge-triggered interrupt for fatal system
>>     errors (e.g. it may be triggered in resonse to ECC errors,
>>     corruption detected in caches, etc). Even if you can determine that
>>     the GISB triggered *an* SError, this does not tell you that this was
>>     the *only* SError.
>
> Correct, which is why Doug's changes allow chaining of handlers.
>
>> 
>>     If you take an SError, something bad has already happened. Your data
>>     may already have been corrupted, and worse, you don't know when or
>>     where specifically this occurred (nor how many times).
>
> Sure, but that still allows you to send the correct signal to a faulting
> application (unless I am missing something here).
>
>>     
>> (c) You cannot determine the (SW) origin of an SError without relying
>>     upon implementation details. This cannot be written in a way that
>>     does not rely on microarchitecture, integration, etc, and would need
>>     to be updated for every future system with this misfeature.
>
> Which is exactly what is being done here, with the help of platform
> specific information (we would not load brcmstb_gisb.c if we were not on
> a platform where it makes sense to use that HW).
>
>> 
>> (d) Even if you can determine the (SW) origin of an SError by relying on
>>     IMPLEMENTATION DEFINED details, your handler needs to be intimately
>>     familiar with the arch in question in order to attempt to recover.
>> 
>>     For example, the existing code tries to skip an ARM instruction in
>>     some cases. For arm64 there are three cases that would need to be
>>     handled (AArch64 A64, AArch32 A32/ARM, AArch32 T32/Thumb).
>> 
>>     Further, it appears to me that the existing code is broken given
>>     that it doesn't handle Thumb, and given that it's skipping an
>>     instruction in response to an asynchronous event -- i.e. some
>>     arbitrary instruction after the one which triggered the abort.
>
> OK, that could presumably be fixed though.
>
>> 
>> For better or worse, SError *must* be treated as fatal.
>
> I disagree here, since this is a platform specific SError exception that
> we can actually handle correctly there is a chance to actually not take
> down the system on something that can be made non fatal and informative
> at the same time.
>
>> 
>> As Doug stated:
>> 
>>     The main benefit is to help debug user mode code that accidentally
>>     maps a bad address since we would never make such an egregious error
>>     in the kernel ;)
>> 
>> This is just one of many ways a userspace application with direct HW
>> access can bring down the system. I see no reason to treat it any
>> differently, especially given the above points.
>
> Partially disagree, in the absence of a way to specifically deal with
> the exception, I would almost agree, but this is not the case here, we
> have a piece of HW that can help us locate the problem, display an
> informative message, and send a SIGBUS to the faulting application.
>
> Anyway, I won't argue much further than that, but I certainly don't
> think taking down an entire system is going to prove itself useful when
> you need to deploy such a kernel to hundreds of people who have no clue
> what so ever what their actual problem is in the first place. Taking a
> SIGBUS and printing a message can at least allow us to say: read more
> carefully, it say exactly what's wrong.

I think there is one point that hasn't been made in this discussion. You
seem to assume that an SError should be handled the same way on an ARMv8
system as you handle it on your ARMv7 platform.

In most cases, Linux on ARMv7 runs (for better or worse) in secure mode,
making it the only software agent capable of handling an Asynchronous
Abort. On v8, Linux runs in non-secure mode, and relies on secure
firmware for a large set of platform specific services (PM, CPU
bring-up...). Crucially, error handling is one of these services.

It is largely expected that an SError should be first taken to EL3
(SCR_EL3.EA being set), handled there by any platform-specific FW able
to triage and log the error, and could even be reported to EL1NS via
some standard mechanism. If the FW decides to reinject this SError to
the non-secure side, then this is bound to be fatal, because even the FW
couldn't handle it.

So my view is that you should move that kind of error handling to the
place where it actually belongs. It will give you the opportunity to
print out your debug messages if necessary, and leave the SError
handling in the kernel the way it should be: fatal.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny.