[RFC] Tunable elision patch for siddhesh/tunables

[Paul E. Murphy]

On 10/06/2015 04:22 AM, Siddhesh Poyarekar wrote:
> On Monday 05 October 2015 09:45 PM, Paul E. Murphy wrote:
>> In continuing the glibc tunables discussion, I've made some
>> additional patches which I think would help our work on the
>> siddhesh/tunables branch.
>>
>> I've split this into two patches:
>>
>> * The first should enable tunables to initialize without
>> calling malloc(), and prior to __environ being set.
>>
>> * The second adds a common tunables initialization function
>> for elision, and enables it for supported archs.
>>
>> I have tested this on PPC64. I need help testing for
>> supported s390 and x86 platforms.
> 
> I haven't done a thorough review, but the overall approach seems sane.
> The only objection I have there is the new macro
> TUNABLE_REGISTER_WITH_ENV.  You should be using the
> COMPAT_TUNABLE_REGISTER instead and optionally pass it the envp.  I

My interpretation is that COMPAT_TUNABLE_* macros should only be used
for porting existing tunables with an existing env var.

> don't currently have an opinion on whether a new environment variable
> for elision is worth having, but that is a separate point that others
> will likely have a strong opinion about.  I would only like it to be

I have intentionally avoided explicitly creating new env vars. It has
been made clear by Roland this is unacceptable. The consensus is that
any such change must use the tunables abstraction. The debate is over
on this topic AFAIK.

> separate from the core tunable infrastructure so that it does not
> dictate which way the framework goes.
> 

I agree. The intent of these patches is to validate your existing
work with tunables by using something clearly demanded by the glibc
users.

What work remains on this branch before you open it for debate?


BR,
Paul