On Tue, 11 Oct 2005, Nick Piggin wrote:
Alan Cox wrote:
32000 processes each with 2G mapped as zero pages appears to allow the
refcount to overflow ?
That's right (though I count only 8192 required with 4K page size) -
close to impossible on 32-bit architectures, though not so the 64-bit
ones, which still use 32-bits for count and mapcount.
It needs 16GB of page table space to get the mapcount to go negative,
or if we refine the atomic_add_negative test, 32GB of page table space
to wrap (I'm assuming an 8-byte PAE page table entries for each unit
of mapcount, since we're well beyond the 4GB non-PAE limit).
Do we actually need to worry about i386 above 32GB in the x86_64 era?
I was a bit worried about this too, but Hugh didn't think it was a
really big a deal - I guess because the real solution for the refcount
overflow on 64-bit is to expand the refcount type.
Yes, and I'm imagining some scheme of sharing _count and _mapcount in
a single atomic64, since we don't want to expand struct page for this.
Implement that shortly, unless we find a way to eliminate _mapcount
instead.
But Alan's overflow issue is not new, it's not brought on refcounting
the ZERO_PAGE: sys_remap_file_pages already allows a file page to be
mapped multiple times in single process, without the constriction of
needing lots of vm_area_struct space. ZERO_PAGE is just more obvious.