Use MADV_FREE on linux kernel >= 4.5

[rajatgoel]

Maybe a new configure option to enable/disable this at compile time for linux?

[djwatson]

Last I tested MADV_FREE doesn't work correctly on swapless systems yet (changes still in review -https://lkml.org/lkml/2015/11/24/753). So probably can't be be the default yet and needs to be an option

[jasone]

It looks like the current state of affairs is that MADV_FREE is equivalent to MADV_DONTNEED on swapless systems, so I think it's okay to use, even if currently a minor pessimization since we won't recognized the pages as being zeroed upon reuse.

[madscientist]

Maybe it's just me but I'm leery of configure-time checks for things like this. What happens if I build this on a system where it's available but then run it on a system where it isn't? Maybe this is a spurious concern.

[jasone]

Yeah, cross compiling of any form is in general a dangerous business, and this particular configuration test is a bit more fragile than most because we tend to think of systems with merely different kernel versions as equivalent targets. That said, I want to support MADV_FREE on Linux, but I don't want to saddle us with a compatibility quagmire or explicit configuration control. What we have here is a forward-compatible approach, with an escape hatch (--config autoconf cache file) to enable backward compatibility for people who are actually cross-compiling from newer systems to older ones.

[nh2]

It looks like the current state of affairs is that MADV_FREE is equivalent to MADV_DONTNEED on swapless systems

@jasone @djwatson This is no longer the case, see

https://www.spinics.net/lists/linux-mm/msg167333.html

MADV_FREE now always returns memory to the operating systems "lazily" (on memory pressure).