[PATCH v3] slab_arena: Enhance slab_arena_create to support madvise hints

[Vladimir Davydov]

On Wed, May 15, 2019 at 03:53:04PM +0300, Cyrill Gorcunov wrote:
> On Wed, May 15, 2019 at 03:37:56PM +0300, Vladimir Davydov wrote:
> > 
> > I'd also patch all places (tests, etc) where old flags (MAP_*) are used
> > to switch them to new flags. Sooner or later we should drop this compat
> > layer.
> 
> Maybe we should do this in a separate patch? I would like the tests to sit
> for some time with MAP_ flags to make sure everything works. And I'll prepare
> the patch covering MAP_ next week, sounds OK?

Okay.

> 
> > > -#if !defined(MAP_ANONYMOUS)
> > > -#define MAP_ANONYMOUS MAP_ANON
> > > +static void
> > > +madvise_checked(void *ptr, size_t size, int flags)
> > > +{
> > > +#ifdef TARANTOOL_SMALL_USE_MADVISE
> > 
> > It would be nice to have a warning in case MADV_DONTDUMP is unavailable.
> > It should be printed only once. May be, here, or, even better, in
> > Tarantool itself (via say_warn()).
> 
> You know I thought if we could use small/conf.h inside tarantool.
> For this sake I prefixed symbols with TARANROOL_SMALL_ to not interfere
> with any other symbols.
> 
> Another option -- provide something like
> 
> static const uint64_t small_features = (bitset of features);

I would define HAVE_MADV_DONTNEED in tarantool's config.h.

> 
> bool small_have_feature(uint64_t mask)
> {
> 	return small_features & mask;
> }
> 
> ?
> 
> > >  
> > >  static void
> > >  munmap_checked(void *addr, size_t size)
> > >  {
> > >  	if (munmap(addr, size)) {
> > >  		char buf[64];
> > > -		intptr_t ignore_it = (intptr_t)strerror_r(errno, buf,
> > > -							  sizeof(buf));
> > > +		char *ignore_it = strerror_r(errno, buf, sizeof(buf));
> > 
> > This wouldn't compile on OS X. Please leave it as is. Same's fair for
> > madvise_checked.
> 
> The problem is that if I leave it as is it doesn't compile on linux,
> due to _GNU_SOURCE definition, which is required for MADV_ flags :/

It does compile on my box, which runs Linux. Note the cast: it should be
fine to explicitly cast (char *) to intptr_t.