riscv: Handle four-level page tables in various pmap traversal routines

[jrtc27]

I think pmap_change_attr_locked needs changing otherwise you'll fault rather than return EINVAL for bogus VAs that aren't covered by an L0 entry? The function doesn't currently do anything other than exist for LKPI and do a no-op walk in order to give back the right errors.

[markj]

In D34278#777754, @jrtc27 wrote:

I think pmap_change_attr_locked needs changing otherwise you'll fault rather than return EINVAL for bogus VAs that aren't covered by an L0 entry? The function doesn't currently do anything other than exist for LKPI and do a no-op walk in order to give back the right errors.

Hmm, I don't see why we'd fault - pmap_l1() will return NULL in that case (see D34276) and pmap_change_attr_locked() handles that by returning EINVAL.

[jrtc27]

In D34278#777756, @markj wrote:

In D34278#777754, @jrtc27 wrote:

I think pmap_change_attr_locked needs changing otherwise you'll fault rather than return EINVAL for bogus VAs that aren't covered by an L0 entry? The function doesn't currently do anything other than exist for LKPI and do a no-op walk in order to give back the right errors.

Hmm, I don't see why we'd fault - pmap_l1() will return NULL in that case (see D34276) and pmap_change_attr_locked() handles that by returning EINVAL.

Hm, indeed, misremembered how those work. I guess you only need to do the walk explicitly if you care about distinguishing !V from !RX, i.e. if you expect to have L0 leaves, which we don't.