Re: [PATCH] arm64: errata: Check for --fix-cortex-a53-843419 and --fix-cortex-a53

From: Florian Fainelli
Date: Thu Nov 03 2016 - 13:20:33 EST

On 11/03/2016 07:16 AM, Will Deacon wrote:
> On Wed, Nov 02, 2016 at 02:57:26PM -0700, Florian Fainelli wrote:
>> On 11/02/2016 02:41 PM, Markus Mayer wrote:
>>> On 2 November 2016 at 14:27, Will Deacon <will.deacon@xxxxxxx> wrote:
>>>> On Wed, Nov 02, 2016 at 02:07:17PM -0700, Markus Mayer wrote:
>>>>> The question I am asking is: What do we have to lose by supporting both options?
>>>> We end up passing "--fix-cortex-a53" to the linker, without knowing what it
>>>> might do in the future.
>>> It seems highly unlikely that such a generic option would be added in
>>> the future, both, because the precedent has been set for topic
>>> specific options, and because they know it has been used in the past,
>>> so they wouldn't add a previously used option to do something
>>> completely different. (And if they really did, then that would be a
>>> huge binutils bug.)
>>> So, we have a trade-off between a real world problem that does
>>> currently exist and avoiding a theoretical issue that may never
>>> materialize.
>> Agreed, also the way Markus' patch is designed makes it such that we
>> first try the full and current option name, and if not supported, try
>> the second (and earlier, now obsolete) option name, so I really don't
>> see a lot of room for things to go wrong here...
> It's not beyond the realms of possibility that ld will grow a
> "fix-cortex-a53" option in the future, that enables all of the a53
> workarounds. Since ld is the linker supported by the kernel and gold isn't,
> I don't want to pass this option down.

No it's entirely reasonable to think this may happen, although:

- this has not happened yet, so once this happens, you will need to cook
a patch for this anyway, and you will be able to gate this catch all
linker option by an appropriate version check presumably

- you would supposedly want a fine grained set of linker options that
are specific to workarounds you have to have enabled, instead of a catch
all "enable all Cortex A53" workarounds

> If you can't change toolchain and you want this worked around, why can't you
> either build gold with it enabled by default, or pass the extra flag on the
> command line to the kernel build system?

Because that creates a distribution problem and now we have to document
this for people who want to build a kernel on their own, without
necessarily understanding if this is something they might need, or why
this is needed, and why the kernel is not taking care of that on its
own? So yes, this comes down to who is responsible for what, in that
case the kernel's Makefile is the best place where to put such knowledge
as to which workaround needs to be enabled by the linker and it
simplifies things a lot for people.