In pmap_advise(), when we encounter a superpage mapping, we first demote the

mapping and then destroy one of the 4 KB page mappings so that there is a
potential trigger for repromotion.  Currently, we destroy the first 4 KB
page mapping that falls within the (current) superpage mapping or the
virtual address range [sva, eva).  However, I have found empirically that
destroying the last 4 KB mapping produces slightly better results,
specifically, more promotions and fewer failed promotion attempts.
Accordingly, this revision changes pmap_advise() to destroy the last 4 KB
page mapping.  It also replaces some nearby uses of boolean_t with bool.

Reviewed by:	kib, markj
Differential Revision:	https://reviews.freebsd.org/D21115

sys/amd64/amd64/pmap.c

@@ -7444,7 +7444,7 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
 	pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V;
 	vm_offset_t va, va_next;
 	vm_page_t m;
-	boolean_t anychanged;
+	bool anychanged;
 
 	if (advice != MADV_DONTNEED && advice != MADV_FREE)
 		return;
@@ -7463,7 +7463,7 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
 	PG_M = pmap_modified_bit(pmap);
 	PG_V = pmap_valid_bit(pmap);
 	PG_RW = pmap_rw_bit(pmap);
-	anychanged = FALSE;
+	anychanged = false;
 	pmap_delayed_invl_start();
 	PMAP_LOCK(pmap);
 	for (; sva < eva; sva = va_next) {
@@ -7505,17 +7505,25 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
 			/*
 			 * Unless the page mappings are wired, remove the
 			 * mapping to a single page so that a subsequent
-			 * access may repromote.  Since the underlying page
+			 * access may repromote.  Choosing the last page
-			 * table page is fully populated, this removal never
+			 * within the address range [sva, min(va_next, eva))
-			 * frees a page table page.
+			 * generally results in more repromotions.  Since the
+			 * underlying page table page is fully populated, this
+			 * removal never frees a page table page.
 			 */
 			if ((oldpde & PG_W) == 0) {
-				pte = pmap_pde_to_pte(pde, sva);
+				va = eva;
+				if (va > va_next)
+					va = va_next;
+				va -= PAGE_SIZE;
+				KASSERT(va >= sva,
+				    ("pmap_advise: no address gap"));
+				pte = pmap_pde_to_pte(pde, va);
 				KASSERT((*pte & PG_V) != 0,
 				    ("pmap_advise: invalid PTE"));
-				pmap_remove_pte(pmap, pte, sva, *pde, NULL,
+				pmap_remove_pte(pmap, pte, va, *pde, NULL,
 				    &lock);
-				anychanged = TRUE;
+				anychanged = true;
 			}
 			if (lock != NULL)
 				rw_wunlock(lock);
@@ -7547,7 +7555,7 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
 				if (va == va_next)
 					va = sva;
 			} else
-				anychanged = TRUE;
+				anychanged = true;
 			continue;
 maybe_invlrng:
 			if (va != va_next) {

sys/arm64/arm64/pmap.c

@@ -4888,15 +4888,23 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
 			/*
 			 * Unless the page mappings are wired, remove the
 			 * mapping to a single page so that a subsequent
-			 * access may repromote.  Since the underlying page
+			 * access may repromote.  Choosing the last page
-			 * table page is fully populated, this removal never
+			 * within the address range [sva, min(va_next, eva))
-			 * frees a page table page.
+			 * generally results in more repromotions.  Since the
+			 * underlying page table page is fully populated, this
+			 * removal never frees a page table page.
 			 */
 			if ((oldl2 & ATTR_SW_WIRED) == 0) {
-				l3 = pmap_l2_to_l3(l2, sva);
+				va = eva;
+				if (va > va_next)
+					va = va_next;
+				va -= PAGE_SIZE;
+				KASSERT(va >= sva,
+				    ("pmap_advise: no address gap"));
+				l3 = pmap_l2_to_l3(l2, va);
 				KASSERT(pmap_load(l3) != 0,
 				    ("pmap_advise: invalid PTE"));
-				pmap_remove_l3(pmap, l3, sva, pmap_load(l2),
+				pmap_remove_l3(pmap, l3, va, pmap_load(l2),
 				    NULL, &lock);
 			}
 			if (lock != NULL)

sys/i386/i386/pmap.c

@@ -5167,19 +5167,19 @@ __CONCAT(PMTYPE, advise)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
 	pt_entry_t *pte;
 	vm_offset_t va, pdnxt;
 	vm_page_t m;
-	boolean_t anychanged, pv_lists_locked;
+	bool anychanged, pv_lists_locked;
 
 	if (advice != MADV_DONTNEED && advice != MADV_FREE)
 		return;
 	if (pmap_is_current(pmap))
-		pv_lists_locked = FALSE;
+		pv_lists_locked = false;
 	else {
-		pv_lists_locked = TRUE;
+		pv_lists_locked = true;
 resume:
 		rw_wlock(&pvh_global_lock);
 		sched_pin();
 	}
-	anychanged = FALSE;
+	anychanged = false;
 	PMAP_LOCK(pmap);
 	for (; sva < eva; sva = pdnxt) {
 		pdnxt = (sva + NBPDR) & ~PDRMASK;
@@ -5193,7 +5193,7 @@ resume:
 			if ((oldpde & PG_MANAGED) == 0)
 				continue;
 			if (!pv_lists_locked) {
-				pv_lists_locked = TRUE;
+				pv_lists_locked = true;
 				if (!rw_try_wlock(&pvh_global_lock)) {
 					if (anychanged)
 						pmap_invalidate_all_int(pmap);
@@ -5212,16 +5212,24 @@ resume:
 			/*
 			 * Unless the page mappings are wired, remove the
 			 * mapping to a single page so that a subsequent
-			 * access may repromote.  Since the underlying page
+			 * access may repromote.  Choosing the last page
-			 * table page is fully populated, this removal never
+			 * within the address range [sva, min(pdnxt, eva))
-			 * frees a page table page.
+			 * generally results in more repromotions.  Since the
+			 * underlying page table page is fully populated, this
+			 * removal never frees a page table page.
 			 */
 			if ((oldpde & PG_W) == 0) {
-				pte = pmap_pte_quick(pmap, sva);
+				va = eva;
+				if (va > pdnxt)
+					va = pdnxt;
+				va -= PAGE_SIZE;
+				KASSERT(va >= sva,
+				    ("pmap_advise: no address gap"));
+				pte = pmap_pte_quick(pmap, va);
 				KASSERT((*pte & PG_V) != 0,
 				    ("pmap_advise: invalid PTE"));
-				pmap_remove_pte(pmap, pte, sva, NULL);
+				pmap_remove_pte(pmap, pte, va, NULL);
-				anychanged = TRUE;
+				anychanged = true;
 			}
 		}
 		if (pdnxt > eva)
@@ -5250,7 +5258,7 @@ resume:
 				if (va == pdnxt)
 					va = sva;
 			} else
-				anychanged = TRUE;
+				anychanged = true;
 			continue;
 maybe_invlrng:
 			if (va != pdnxt) {