Fix some issues with pmap_protect().

[markj]

Summary

Note that pmap_protect() is only ever used to remove permissions from
existing mappings.

• When clearing PTE_W, use an atomic update to the PTE; otherwise we may clobber updates done by the hardware.
• Handle VM_PROT_EXECUTE.
• Clear PTE_D and mark the page dirty when removing write permission from a mapping.

[kib]

sys/riscv/riscv/pmap.c

1918 ↗

(On Diff #52487)

I see that you changed this to atomic clear, so my comment from other review is not relevant.

But what if PTE_D bit set after the pmap_load() but before atomic_clear() ? This seems to be legit and would cause missing dirty set and consequent corruption of the user data. Shouldn't you use atomic_cmpxchg() there and retry on failure ?

[markj]

• Fix the bug that kib pointed out.

[markj]

sys/riscv/riscv/pmap.c

1918 ↗

(On Diff #52487)

Indeed, thanks. :(

BTW, I used _long instead of _64 here because I believe we'll eventually want to support 32-bit kernels. I'll update the macros in D18717 in a separate commit.

[jhb]

sys/riscv/riscv/pmap.c

1874 ↗

(On Diff #52498)

Oops.