On Tue, 11 Jun 2024 at 10:48, Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> Fair enough if that's a pain on x86, but we already have them on arm64, and
> hence using them is a smaller change there. We already have a couple of cases
> which uses MOVZ;MOVK;MOVK;MOVK sequence, e.g.
>
> // in __invalidate_icache_max_range()
> asm volatile(ALTERNATIVE_CB("movz %0, #0\n"
> "movk %0, #0, lsl #16\n"
> "movk %0, #0, lsl #32\n"
> "movk %0, #0, lsl #48\n",
> ARM64_ALWAYS_SYSTEM,
> kvm_compute_final_ctr_el0)
> : "=r" (ctr));
>
> ... which is patched via the callback:
>
> void kvm_compute_final_ctr_el0(struct alt_instr *alt,
> __le32 *origptr, __le32 *updptr, int nr_inst)
> {
> generate_mov_q(read_sanitised_ftr_reg(SYS_CTR_EL0),
> origptr, updptr, nr_inst);
> }
>
> ... where the generate_mov_q() helper does the actual instruction generation.
>
> So if we only care about a few specific constants, we could give them their own
> callbacks, like kvm_compute_final_ctr_el0() above.
I'll probably only have another day until my mailbox starts getting
more pull requests (Mon-Tue outside the merge window is typically my
quiet time when I have time to go through old emails and have time for
private projects).
So I'll look at doing this for x86 and see how it works.
I do suspect that even then it's possibly more code with a
site-specific callback for each case, but maybe it would be worth it
just for the flexibility.
Linus
