RISC-V superpage support, part 1/6.

[markj]

Summary

The change adds routines to promote and demote 2MB superpages. It also
adds support for pmap_enter(psind = 1), described in the commit log
message for r321378. A subsequent commit modifies all of the pmap
routines to handle large mappings appropriately.

The changes are largely adopted from amd64. arm64 has more stringent
requirements around superpage mappings in order to avoid TLB conflict
aborts, and these requirements do not apply to RISC-V, which like amd64
permits TLBs to simultaneously cache 4KB and 2MB translations of a given
page. RISC-V only supports 2 software-defined PTE bits, already used
for PTE_SW_MANAGED and PTE_SW_WIRED, so we do not have an analogue to
amd64's PG_PROMOTED; instead, pmap_remove_l2() always invalidates the
entire 2MB range.

Unlike pmap_remove_pde(), RISC-V's pmap_remove_l2() handles TLB
invalidation.

Note that unlike amd64, RISC-V still has the pvh_global_lock.

[kib]

sys/riscv/riscv/pmap.c

2616 ↗

(On Diff #52909)

Is pmap_enter_l2() allowed to fail when PMAP_ENTER_NOSLEEP is not specified ?

I am asking because vm_fault_populate() is not ready for it, and I have a WIP where I needed to add the following diff:

diff --git a/sys/vm/vm_fault.c b/sys/vm/vm_fault.c
index e095ccc69ed..c3246e4a3b7 100644
--- a/sys/vm/vm_fault.c
+++ b/sys/vm/vm_fault.c
@@ -479,8 +479,20 @@ vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 			    fault_flags, true);
 		}
 		VM_OBJECT_WUNLOCK(fs->first_object);
-		pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ?
-		    PMAP_ENTER_WIRED : 0), psind);
+		rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
+		    (wired ? PMAP_ENTER_WIRED : 0), psind);
+#if defined(__amd64__)
+		if (psind > 0 && rv == KERN_FAILURE) {
+			for (i = 0; i < npages; i++) {
+				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
+				    &m[i], prot, fault_type |
+				    (wired ? PMAP_ENTER_WIRED : 0), 0);
+				MPASS(rv == KERN_SUCCESS);
+			}
+		}
+#else
+		MPASS(rv == KERN_SUCCESS);
+#endif
 		VM_OBJECT_WLOCK(fs->first_object);
 		m_mtx = NULL;
 		for (i = 0; i < npages; i++) {

[markj]

sys/riscv/riscv/pmap.c

2616 ↗

(On Diff #52909)

No, it should not fail for any other reason. (The NOREPLACE flag may cause it to fail upon discovering an existing mapping, but this is not used by pmap_enter().) So your diff doesn't need an update.

IMO it is kind of weird for pmap_enter(psind > 0) to have distinct failure modes from pmap_enter(psind = 0). I'd be interested to hear about it if you feel like describing the WIP here.

sys/vm/vm_fault.c

273 ↗

(On Diff #52909)

We really need a proper predicate for this. Note that platforms may define VM_NRESERVLEVEL > 0 without implementing transparent superpages. sparc64 does this. Should we define a new constant in machine/vmparam.h?

[kib]

sys/riscv/riscv/pmap.c

2616 ↗

(On Diff #52909)

D18893

[jhb]

sys/riscv/riscv/pmap.c

1853 ↗

(On Diff #52909)

s/entires/entries/

[markj]

Fix typo.