amd64: Handle 1GB mappings in pmap_enter_quick_locked()

[markj]

Summary

This code path can be triggered by applying MADV_WILLNEED to a 1GB
mapping.

[kib]

sys/amd64/amd64/pmap.c

7556

Remove the extra () if you are fixing style

7568

[alc]

sys/amd64/amd64/pmap.c

7567–7569

In the newly introduced if statement above, we already know whether pde is not NULL, but I'm pretty sure that the compiler can't infer that pmap_pdpe_to_pde(pdpe, va) cannot be NULL. For example, it doesn't know that physical page 0 is not allocatable. So, you could fold this if/else statement into the above and thereby eliminate the pde NULL check. The PHYS_TO_VM_PAGE line might be the only one that crosses 80 characters.

[markj]

Handle review comments

[markj]

sys/amd64/amd64/pmap.c

7567–7569

I rewrote this fragment to avoid needless comparisons, but I can't see a way to do so that doesn't require adding either some code duplication or an ugly goto.

[alc]

Shouldn't the same change be made on arm64?

sys/amd64/amd64/pmap.c

7567–7569

In this case, I would expect the compiler to be able to deduplicate the generated code with no extra branches.

[markj]

In D36674#832879, @alc wrote:

Shouldn't the same change be made on arm64?

Yes, I simply haven't tested my arm64 patch yet.

As a side note, the existing regression tests for posix shared memory would have caught this if not for a bug in the test (largepage_madvise). :(