Implement an alternative solution to the amd64 and i386 pmap problem that we

previously addressed in r348246.

This pmap problem also exists on arm64 and riscv.  However, the original
solution developed for amd64 and i386 cannot be used on arm64 and riscv.  In
particular, arm64 and riscv do not define a PG_PROMOTED flag in their level
2 PTEs.  (A PG_PROMOTED flag makes no sense on arm64, where unlike x86 or
riscv we are required to break the old 4KB mappings before making the 2MB
mapping; and on riscv there are no unused bits in the PTE to define a
PG_PROMOTED flag.)

This commit implements an alternative solution that can be used on all four
architectures.  Moreover, this solution has two other advantages.  First, on
older AMD processors that required the Erratum 383 workaround, it is less
costly.  Specifically, it avoids unnecessary calls to pmap_fill_ptp() on a
superpage demotion.  Second, it enables the elimination of some calls to
pagezero() in pmap_kernel_remove_{l2,pde}().

In addition, remove a related stale comment from pmap_enter_{l2,pde}().

Reviewed by:	kib, markj (an earlier version)
MFC after:	1 week
Differential Revision:	https://reviews.freebsd.org/D20538

1 files changed, 26 insertions, 14 deletions

diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c
index f7599bf364cf..d7a35b5c9cfc 100644
--- a/sys/amd64/amd64/pmap.c
+++ b/sys/amd64/amd64/pmap.c
@@ -1053,7 +1053,7 @@ static int	pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde,
 static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va,
     vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp);
 static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte);
-static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte);
+static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted);
 static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva,
     vm_offset_t eva);
 static void pmap_invalidate_cache_range_all(vm_offset_t sva,
@@ -1757,8 +1757,13 @@ pmap_init(void)
 		mpte->pindex = pmap_pde_pindex(KERNBASE) + i;
 		mpte->phys_addr = KPTphys + (i << PAGE_SHIFT);
 		mpte->wire_count = 1;
+
+		/*
+		 * Collect the page table pages that were replaced by a 2MB
+		 * page in create_pagetables().  They are zero filled.
+		 */
 		if (i << PDRSHIFT < KERNend &&
-		    pmap_insert_pt_page(kernel_pmap, mpte))
+		    pmap_insert_pt_page(kernel_pmap, mpte, false))
 			panic("pmap_init: pmap_insert_pt_page failed");
 	}
 	PMAP_UNLOCK(kernel_pmap);
@@ -3129,12 +3134,15 @@ pmap_add_delayed_free_list(vm_page_t m, struct spglist *free,
  * of idle page table pages.  Each of a pmap's page table pages is responsible
  * for mapping a distinct range of virtual addresses.  The pmap's collection is
  * ordered by this virtual address range.
+ *
+ * If "promoted" is false, then the page table page "mpte" must be zero filled.
  */
 static __inline int
-pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte)
+pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted)
 {
 
 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
+	mpte->valid = promoted ? VM_PAGE_BITS_ALL : 0;
 	return (vm_radix_insert(&pmap->pm_root, mpte));
 }
 
@@ -4626,7 +4634,7 @@ pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
 	 * If the page table page is not leftover from an earlier promotion,
 	 * initialize it.
 	 */
-	if ((oldpde & PG_PROMOTED) == 0)
+	if (mpte->valid == 0)
 		pmap_fill_ptp(firstpte, newpte);
 
 	pmap_demote_pde_check(firstpte, newpte);
@@ -4699,9 +4707,11 @@ pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
 	newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V;
 
 	/*
-	 * Initialize the page table page.
+	 * If this page table page was unmapped by a promotion, then it
+	 * contains valid mappings.  Zero it to invalidate those mappings.
 	 */
-	pagezero((void *)PHYS_TO_DMAP(mptepa));
+	if (mpte->valid != 0)
+		pagezero((void *)PHYS_TO_DMAP(mptepa));
 
 	/*
 	 * Demote the mapping.
@@ -4766,6 +4776,8 @@ pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
 	} else {
 		mpte = pmap_remove_pt_page(pmap, sva);
 		if (mpte != NULL) {
+			KASSERT(mpte->valid == VM_PAGE_BITS_ALL,
+			    ("pmap_remove_pde: pte page not promoted"));
 			pmap_resident_count_dec(pmap, 1);
 			KASSERT(mpte->wire_count == NPTEPG,
 			    ("pmap_remove_pde: pte page wire count error"));
@@ -5399,7 +5411,7 @@ setpte:
 	    ("pmap_promote_pde: page table page is out of range"));
 	KASSERT(mpte->pindex == pmap_pde_pindex(va),
 	    ("pmap_promote_pde: page table page's pindex is wrong"));
-	if (pmap_insert_pt_page(pmap, mpte)) {
+	if (pmap_insert_pt_page(pmap, mpte, true)) {
 		atomic_add_long(&pmap_pde_p_failures, 1);
 		CTR2(KTR_PMAP,
 		    "pmap_promote_pde: failure for va %#lx in pmap %p", va,
@@ -5826,15 +5838,13 @@ pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags,
 		}
 		vm_page_free_pages_toq(&free, true);
 		if (va >= VM_MAXUSER_ADDRESS) {
+			/*
+			 * Both pmap_remove_pde() and pmap_remove_ptes() will
+			 * leave the kernel page table page zero filled.
+			 */
 			mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
-			if (pmap_insert_pt_page(pmap, mt)) {
-				/*
-				 * XXX Currently, this can't happen because
-				 * we do not perform pmap_enter(psind == 1)
-				 * on the kernel pmap.
-				 */
+			if (pmap_insert_pt_page(pmap, mt, false))
 				panic("pmap_enter_pde: trie insert failed");
-			}
 		} else
 			KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p",
 			    pde));
@@ -6824,6 +6834,8 @@ pmap_remove_pages(pmap_t pmap)
 					}
 					mpte = pmap_remove_pt_page(pmap, pv->pv_va);
 					if (mpte != NULL) {
+						KASSERT(mpte->valid == VM_PAGE_BITS_ALL,
+						    ("pmap_remove_pages: pte page not promoted"));
 						pmap_resident_count_dec(pmap, 1);
 						KASSERT(mpte->wire_count == NPTEPG,
 						    ("pmap_remove_pages: pte page wire count error"));