On Thu, Dec 08, 2022, Oliver Upton wrote:
> On Thu, Dec 08, 2022 at 12:18:07AM +0000, Sean Christopherson wrote:
>
> [...]
>
> > Together, what about? The #ifdef is a bit gross, especially around "hi_start",
> > but it's less duplicate code. And IMO, having things bundled in the same place
> > makes it a lot easier for newbies (to arm64 or kernel coding in general) to
> > understand what's going on and why arm64 is different.
>
> I'd rather we not go this route. We really shouldn't make any attempt to
> de-dupe something that is inherently architecture specific.
>
> For example:
>
> > + /*
> > + * All architectures supports splitting the virtual address space into
> > + * a high and a low half. Populate both halves, except for arm64 which
> > + * currently uses only TTBR0_EL1 (arbitrary selftests "logic"), i.e.
> > + * only has a valid low half.
> > + */
> > + sparsebit_num_t nr_va_bits = (1ULL << (vm->va_bits - 1)) >> vm->page_shift;
>
> This is still wrong for arm64. When we say the VA space is 48 bits, we
> really do mean that TTBR0 is able to address a full 48 bits. So this
> truncates the MSB for the addressing mode.
Ah, I missed the lack of a "-1" in the arm64 code.
> With the code living in the arm64 side of the shop, I can also tailor
> the comment to directly match the architecture to provide breadcrumbs
> tying it back to the Arm ARM.
The main reason why I don't like splitting the code this way is that it makes it
harder for non-arm64 folks to understand what makes arm64 different. Case in
point, my overlooking of the "-1". I read the changelog and the comment and
still missed that small-but-important detail, largely because I am completely
unfamiliar with how TTBR{0,1}_EL1 works.
Actually, before we do anything, we should get confirmation from the s390 and
RISC-V folks on whether they have a canonical hole like x86, i.e. maybe x86 is
the oddball.
Anyways, assuming one architecture is the oddball (I'm betting it's x86), I have
no objection to bleeding some of the details into the common code, including a
large comment to document the gory details. If every architecture manges to be
different, then yeah, a hook is probably warranted.
That said, I also don't mind shoving a bit of abstraction into arch code if that
avoids some #ifdef ugliness or allows for better documentation, flexibility, etc.
What I don't like is duplicating the logic of turning "VA bits" into the bitmap.
E.g. something like this would also be an option. Readers would obviously need
to track down has_split_va_space, but that should be fairly easy and can come
with a big arch-specific comment, and meanwhile the core logic of how selftests
populate the va bitmaps is common.
Or if arm64 is the only arch without a split, invert the flag and have arm64 set
the vm->has_combined_va_space or whatever.
static void vm_vaddr_populate_bitmap(struct kvm_vm *vm)
{
unsigned int eff_va_bits = vm->va_bits;
sparsebit_num_t nr_bits;
/* blah blah blah */
if (vm->has_split_va_space)
eff_va_bits--;
nr_bits = (1ULL << eff_va_bits) >> vm->page_shift;
sparsebit_set_num(vm->vpages_valid, 0, nr_va_bits);
if (vm->has_split_va_space)
sparsebit_set_num(vm->vpages_valid,
(~((1ULL << eff_va_bits) - 1)) >> vm->page_shift,
nr_bits);
}
