AARCH64 - 48-bit VA
Enable 48bit VA on AARCH64
- Name: Jeremy Linton
- Email: firstname.lastname@example.org
- Release notes owner:
The current aarch64 kernel is using a 42-bit process virtual address (VA) space and due to the way aarch64 paging works this limits the max physical address as well. This is fairly limiting for some applications, but aarch64 processors also have support for 48-bit VA's. This is actually 1-bit more than the "47-bit" VA's currently used on x86_64. This change is a fairly minor kernel configuration but it has caused problems with mozjs and luajit based projects because they have stored tags in their pointers in the 48th address bit. The upstream mozjs project has been fixed, and there are a set of changes outstanding for luajit. Futher there, have been patches applied to rawhide to support it (bug #1242326, bug #1375305. bug #1375547), but the packages dependent on js185 require a rebuild due to an ABI change.
Benefit to Fedora
Larger aarch64 processes won't be constrained by both the virtual as well as physical limitations of the 42-bit VA currently in use. This actually also helps with things like hugetlb's and potentially provides a performace boost. Lastly, it allows fedora to boot on a class of machines that have the majority of their RAM higher in the address space.
- Proposal owners:
The kernel config needs to be changed, and the core mozjs/luajit tagging problems need to be resolved. This can either take the form of patching the remaining mozjs packages that have not yet been patched and rebuilding the affected js185 packages, or it can take the form of moving dependent packages onto mozjs versions that have been patched. Futher lua needs to either be updated to the latest mainline or the critical patches need to be applied to the lua versions currently in use.
- Other developers:
- Release engineering:
Release engineering needs to assure the js185 packages are rebuilt. Peter Robinson has volonteered to assist in this.
- Policies and guidelines:
- Trademark approval:
N/A (not needed for this Change)
It shouldn't have any noticable impact on upgrades, currently its possible to boot machines with both 42 and 48 bit VA changes simply by picking diffrent kernel options in grub.
How To Test
Normal system operation on aarch64 hardware. This include running package unit tests/etc as normal and watching out for regressions. The change is system wide because it basically affects the entire system, but should be fairly transparent.
Its unlikely a desktop user would notice the change, except possibly it may boot on additional hardware. A server user could potentially utilize more RAM for in-memory databases/etc.
mozjs and luajit packages as well as js's dependent packages.
Need js patches
- js-1.8.5 (fixed with js-1.8.5-27.fc27 and js-1.8.5-27.fc26)
- mozjs17 (hopefully unnecessary due to polkit move to mozjs24 https://lists.freedesktop.org/archives/polkit-devel/2016-August/000503.html)
- mozjs31 (hopefully unnecessary due to 0ad move to mozjs38 http://trac.wildfiregames.com/ticket/3708)
Packages needing rebuild after js-1.8.5 ABI patch
- erlang-js-0:1.3.0-7 (possible upgrade patch https://github.com/basho/erlang_js/pull/44#issuecomment-247323892)
- libproxy-mozjs-0:0.4.12 (hopefully moving to mozjs24)
- pacrunner-0:0.7-7(hopefully moving to mozjs24 or mozjs38)
Pull lua forward to latest upstream.
In the worse case, the kernel can be reverted back to 42-bit VA's with a single configuration change. Otherwise, we can patch additional mozjs versions (mosjz31 if 0ad's patches arn't complete) or drop the dependent packages from the aarch64 disto until they can be fixed (really this only applies to 0ad at this point).
- Contingency deadline: Beta freeze
- Blocks release? No, the likely problem is an additional package is found to have a problem. I would expect that should this happen it won't be a major bug/package as 48-bit VA's have been in test for a while.
As part of this, an attempt has been made to reduce the number of mozjs versions in use. The plan is to move dependent packages to either mozjs24 or mozjs38. There remains some work to move some of the original js 1.8.5 users, but that isn't a critical problem as simply rebuilding them solves the problems caused by the additional address bit.
I can add additional information about the hardware or the dependent projects here if its appropriate.
Enable 48-bit address spaces.