Re: Using ggc-page with plain malloc

[Wolfram Gloger <Wolfram dot Gloger at dent dot med dot uni-muenchen dot de>]

Hello,

Sorry for the late followup to this.

> >> I've been trying to arrange for ggc-page to work upon plain malloc.
> >> All it needs is some way to ask for a memory area aligned at a page
> >> boundary.  It's a bit tricky, because of wide variations among the
> >> malloc() implementations, but I believe it can made to work well
> >> unless someone finds some platform that breaks the following program
> >> in a way that is too hard to fix.
> 
> > Wouldn't it be simpler to use a system interface for obtaining aligned
> > memory? valloc(3) is part of Single Unix, and it gives you
> > sysconf(_SC_PAGESIZE)-aligned memory. In addition (or alternatively?),
> > memalign gives you arbitrarily-aligned memory.
> 
> We already use valloc when it's available.  But there are systems that
> don't support it, and that's precisely the fallback I'm trying to
> create.

I was initially going to suggest using memalign() instead of valloc(),
but that seems to be rather non-standardized (neither ANSI C nor
Single Unix mention it at all, that surprised me).  Also note that SUS
marks valloc() as `Legacy' and says:

     Applications should avoid using valloc() but should use malloc()
     or mmap() instead. On systems with a large page size, the number
     of successful valloc() operations may be zero.

A memalign() emulation via malloc() is certainly the most portable
solution, however using it a lot (for, e.g. allocating a single page)
is a dead-sure method for fragmenting the heap badly.

I would suggest to use anonymous mmap() or mmap() on /dev/zero
(whereever one of those is available) to obtain pages of memory.

Regards,
Wolfram.