On a 24 CPUs amd64 machine building a kernel with -j24 always result in
a corrupted "RES" `pmap_kernel()->pm_stats.resident_count' view in top(1).

Modifications to this field are (mostly) serialized by the pmap's mutex.
Sadly pmap_enter() do not grab the mutex for the kernel's pmap.  Diff
below changes that and fix the issue for me.  I believe it is easier and
cheaper than using atomics everywhere.

ok?

Index: arch/amd64/amd64/pmap.c
===================================================================
RCS file: /cvs/src/sys/arch/amd64/amd64/pmap.c,v
diff -u -p -r1.171 pmap.c
--- arch/amd64/amd64/pmap.c	8 Aug 2024 15:57:22 -0000	1.171
+++ arch/amd64/amd64/pmap.c	23 Aug 2024 14:13:46 -0000
@@ -416,10 +416,6 @@ pmap_map_ptes(struct pmap *pmap)
 
 	KASSERT(pmap->pm_type != PMAP_TYPE_EPT);
 
-	/* the kernel's pmap is always accessible */
-	if (pmap == pmap_kernel())
-		return 0;
-
 	/*
 	 * Lock the target map before switching to its page tables to
 	 * guarantee other CPUs have finished changing the tables before
@@ -427,6 +423,10 @@ pmap_map_ptes(struct pmap *pmap)
 	 */
 	mtx_enter(&pmap->pm_mtx);
 
+	/* the kernel's pmap is always accessible */
+	if (pmap == pmap_kernel())
+		return 0;
+
 	cr3 = rcr3();
 	KASSERT((cr3 & CR3_PCID) == PCID_KERN ||
 		(cr3 & CR3_PCID) == PCID_PROC);
@@ -443,8 +443,7 @@ pmap_map_ptes(struct pmap *pmap)
 void
 pmap_unmap_ptes(struct pmap *pmap, paddr_t save_cr3)
 {
-	if (pmap != pmap_kernel())
-		mtx_leave(&pmap->pm_mtx);
+	mtx_leave(&pmap->pm_mtx);
 
 	if (save_cr3 != 0)
 		lcr3(save_cr3);