Index: sys/arm64/arm64/pmap.c
===================================================================
--- sys/arm64/arm64/pmap.c
+++ sys/arm64/arm64/pmap.c
@@ -508,7 +508,8 @@
 static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va);
 
 static uma_zone_t pmap_bti_ranges_zone;
-static bool pmap_bti_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva);
+static bool pmap_bti_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
+    pt_entry_t *pte);
 static pt_entry_t pmap_pte_bti(pmap_t pmap, vm_offset_t va);
 static void pmap_bti_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva);
 static void *bti_dup_range(void *ctx, void *data);
@@ -4955,21 +4956,22 @@
 #endif /* VM_NRESERVLEVEL > 0 */
 
 static int
-pmap_enter_largepage(pmap_t pmap, vm_offset_t va, pt_entry_t newpte, int flags,
+pmap_enter_largepage(pmap_t pmap, vm_offset_t va, pt_entry_t pte, int flags,
     int psind)
 {
-	pd_entry_t *l0p, *l1p, *l2p, origpte;
+	pd_entry_t *l0p, *l1p, *l2p, newpte, origpte;
 	vm_page_t mp;
 
 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
 	KASSERT(psind > 0 && psind < MAXPAGESIZES,
 	    ("psind %d unexpected", psind));
-	KASSERT((PTE_TO_PHYS(newpte) & (pagesizes[psind] - 1)) == 0,
-	    ("unaligned phys address %#lx newpte %#lx psind %d",
-	    PTE_TO_PHYS(newpte), newpte, psind));
+	KASSERT((PTE_TO_PHYS(pte) & (pagesizes[psind] - 1)) == 0,
+	    ("unaligned phys address %#lx pte %#lx psind %d",
+	    PTE_TO_PHYS(pte), pte, psind));
 
 restart:
-	if (!pmap_bti_same(pmap, va, va + pagesizes[psind]))
+	newpte = pte;
+	if (!pmap_bti_same(pmap, va, va + pagesizes[psind], &newpte))
 		return (KERN_PROTECTION_FAILURE);
 	if (psind == 2) {
 		PMAP_ASSERT_L1_BLOCKS_SUPPORTED;
@@ -5123,9 +5125,6 @@
 
 	lock = NULL;
 	PMAP_LOCK(pmap);
-	/* Wait until we lock the pmap to protect the bti rangeset */
-	new_l3 |= pmap_pte_bti(pmap, va);
-
 	if ((flags & PMAP_ENTER_LARGEPAGE) != 0) {
 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
 		    ("managed largepage va %#lx flags %#x", va, flags));
@@ -5197,6 +5196,7 @@
 	orig_l3 = pmap_load(l3);
 	opa = PTE_TO_PHYS(orig_l3);
 	pv = NULL;
+	new_l3 |= pmap_pte_bti(pmap, va);
 
 	/*
 	 * Is the specified virtual address already mapped?
@@ -5478,7 +5478,7 @@
 	 * and let vm_fault() cope.  Check after l2 allocation, since
 	 * it could sleep.
 	 */
-	if (!pmap_bti_same(pmap, va, va + L2_SIZE)) {
+	if (!pmap_bti_same(pmap, va, va + L2_SIZE, &new_l2)) {
 		KASSERT(l2pg != NULL, ("pmap_enter_l2: missing L2 PTP"));
 		pmap_abort_ptp(pmap, va, l2pg);
 		return (KERN_PROTECTION_FAILURE);
@@ -5733,19 +5733,6 @@
 			}
 		}
 		l3p = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(*ml3p));
-
-have_l3p:
-		/*
-		 * If bti is not the same for the whole L3C range, return
-		 * failure and let vm_fault() cope.  Check after L3 allocation,
-		 * since it could sleep.
-		 */
-		if (!pmap_bti_same(pmap, va, va + L3C_SIZE)) {
-			(*ml3p)->ref_count -= L3C_ENTRIES - 1;
-			pmap_abort_ptp(pmap, va, *ml3p);
-			*ml3p = NULL;
-			return (KERN_PROTECTION_FAILURE);
-		}
 	} else {
 		*ml3p = NULL;
 
@@ -5768,8 +5755,21 @@
 			    pmap_load(pde)));
 		}
 	}
+have_l3p:
 	l3p = &l3p[pmap_l3_index(va)];
 
+	/*
+	 * If bti is not the same for the whole L3C range, return failure
+	 * and let vm_fault() cope.  Check after L3 allocation, since
+	 * it could sleep.
+	 */
+	if (!pmap_bti_same(pmap, va, va + L3C_SIZE, &l3e)) {
+		(*ml3p)->ref_count -= L3C_ENTRIES - 1;
+		pmap_abort_ptp(pmap, va, *ml3p);
+		*ml3p = NULL;
+		return (KERN_PROTECTION_FAILURE);
+	}
+
 	/*
 	 * If there are existing mappings, either abort or remove them.
 	 */
@@ -9271,8 +9271,16 @@
 		rangeset_remove_all(pmap->pm_bti);
 }
 
+/*
+ * Returns true if the BTI setting is the same across the specified address
+ * range, and false otherwise.  When returning true, updates the referenced PTE
+ * to reflect the BTI setting.
+ *
+ * Only stage 1 pmaps support BTI.  The kernel pmap is always a stage 1 pmap
+ * that has the same BTI setting implicitly across its entire address range.
+ */
 static bool
-pmap_bti_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
+pmap_bti_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pt_entry_t *pte)
 {
 	struct rs_el *prev_rs, *rs;
 	vm_offset_t va;
@@ -9283,9 +9291,14 @@
 	KASSERT(ADDR_IS_CANONICAL(eva),
 	    ("%s: End address not in canonical form: %lx", __func__, eva));
 
-	if (pmap->pm_bti == NULL || ADDR_IS_KERNEL(sva))
+	if (pmap == kernel_pmap) {
+		*pte |= ATTR_KERN_GP;
+		return (true);
+	}
+	if (pmap->pm_bti == NULL)
 		return (true);
-	MPASS(!ADDR_IS_KERNEL(eva));
+	PMAP_ASSERT_STAGE1(pmap);
+	prev_rs = NULL;
 	for (va = sva; va < eva; prev_rs = rs) {
 		rs = rangeset_lookup(pmap->pm_bti, va);
 		if (va == sva)
@@ -9298,6 +9311,8 @@
 		}
 		va = rs->re_end;
 	}
+	if (prev_rs != NULL)
+		*pte |= ATTR_S1_GP;
 	return (true);
 }