On Wed, Oct 19, 2011 at 08:40:08AM +0200, Andi Kleen wrote:
> From: Andi Kleen <a...@linux.intel.com>
>
> This implements the freeing back of large chunks in the ggc madvise path
> Richard Guenther asked for. This way on systems with limited
> address space malloc() and other allocators still have
> a chance to get back at some of the memory ggc freed. The
> fragmented pages are still just given back, but the address space
> stays allocated.
>
> I tried freeing only aligned 2MB areas to optimize for 2MB huge
> pages, but the hit rate was quite low, so I switched to 1MB+
> unaligned areas. The target size is a param now.
If the size to free is smaller than the quirk size, then it has the very
undesirable effect that with using GC only you might run unnecessarily out
of virtual address space, because it allocates pages in 2MB chunks, but
if they are released in 1MB chunks, those released chunks will never be
usable again for GC. Consider on 32-bit address space allocating 3GB
of GC memory, then freeing stuff in every odd 1MB chunk of pages, then
wanting to allocate through GC the 1.5GB back.
IMHO we should munmap immediately in release_pages the > G.pagesize pages,
those are not very likely to be reused anyway (and it had one in between
ggc_collect cycle to be reused anyway), and for the == G.pagesize
(the usual case, the only ones that are allocated in GGC_QUIRK_SIZE sets)
we should note which page was the first one in the GGC_QUIRK_SIZE chunk
and munmap exactly those 2MB starting at the first page only.
Jakub
