* Matthew Wilcox: > I received a suggestion off-list that we only do the PMD alignment on > 64-bit, which seems quite reasonable to me. After all, I don't care > about performance on 32-bit just as much as I don't care about security > on 32-bit. Perhaps we can we repurpose MAP_DENYWRITE to disable this? For shared objects as loaded by a dynamic linker, the alignment is pointless in many cases even if the original mapping is quite a bit larger than 2 MiB because the individual LOAD segments and their protection settings are smaller than 2 MiB, so hugepages cannot be used in the end after all. The dynamic linker knows the LOAD segments, so it can drop MAP_DENYWRITE if it determines that hugepages could be beneficial. (Current glibc sets MAP_DENYWRITE for historic reasons.) On the other hand, I wouldn't object to more explicit control over mmap pointer alignment, either, for anonymous mappings as well. There are some binutils versions that produce 2 MiB aligned file layout on x86-64, but that change was reverted, presumably because kernel hugepage support for non-anonymous memory wasn't available. But there are likely some iffy details that make these binaries unusable for hugepages in practice, like lack of hugepage alignment at the end of LOAD segments. Unfortunately, BFD ld tends to produce approximate PT_LOAD and PT_GNU_RELRO and relies on the dynamic loader to round things up and down in somewhat questionable ways. Thanks, Florian
