posix_memalign() in libvmmalloc.c

[Steve Byan's Lists]

I'm curious as to why libvmmalloc.c reimplements posix_memalign using je_vmem_memalign() and je_vmem_pool_aligned_alloc() rather than simply delegating to je_vmem_posix_memalign()?

The use of je_vmem_memalign() causes problems on platforms that don't include that legacy API. The jemalloc configure logic doesn't export that function unless the platform already provides memalign(). Should I force jemalloc to export je_vmem_memalign() regardless of the platform, or would it be better to recode posix_memalign() to avoid je_vmem_memalign()?

Best regards,
-Steve

--
Steve Byan
stev...@me.com
Littleton, MA

[Krzysztof Czurylo]

Hi Steve,

To be honest, I don't see any reason for not using je_vmem_posix_memalign().  Simply, I didn't realize there could be any problem with je_vmem_memalign().  Perhaps I got inspired by jemalloc internals, as all of memalign(), posix_memalign(), aligned_alloc() and valloc() are implemented by a call to internal function imemalign().

Anyway, I think we should modify posix_memalign() as you described.  I'll make a 

Also, instead of changing jemalloc configure logic and forcing it to export memalign(), we should rather modify the way how VMMALLOC_OVERRIDE_MEMALIGN is defined and make it depend on JEMALLOC_OVERRIDE_MEMALIGN definition.

Thanks,

K.

[Krzysztof Czurylo]

I was going to say I'll make a patch shortly. Here it is:

https://github.com/pmem/nvml/pull/776

K.

[Steve Byan's Lists]

> On Apr 26, 2016, at 3:02 PM, Krzysztof Czurylo <krzyszto...@intel.com> wrote:
>
> I was going to say I'll make a patch shortly. Here it is:
> https://github.com/pmem/nvml/pull/776

Thanks very much! Works great.