Map the vm_page array into KVA on amd64.

r351198 allows the kernel to use domain-local memory to back the vm_page
array (up to 2MB boundaries) and reserves a separate PML4 entry for that
purpose.  One consequence of that change is that the vm_page array is no
longer present in minidumps, which only adds pages mapped above
VM_MIN_KERNEL_ADDRESS.

To avoid the friction caused by having kernel data structures mapped
below VM_MIN_KERNEL_ADDRESS, map the vm_page array starting at
VM_MIN_KERNEL_ADDRESS instead of using a dedicated PML4 entry.

Reviewed by:	kib
Discussed with:	jeff
Sponsored by:	The FreeBSD Foundation
Differential Revision:	https://reviews.freebsd.org/D21491

sys/amd64/amd64/pmap.c

@@ -384,9 +384,6 @@ static u_int64_t	DMPDphys;	/* phys addr of direct mapped level 2 */
 static u_int64_t	DMPDPphys;	/* phys addr of direct mapped level 3 */
 static int		ndmpdpphys;	/* number of DMPDPphys pages */
 
-static uint64_t		PAPDPphys;	/* phys addr of page array level 3 */
-static int		npapdpphys;	/* number of PAPDPphys pages */
-
 static vm_paddr_t	KERNend;	/* phys addr of end of bootstrap data */
 
 /*
@@ -1431,16 +1428,6 @@ create_pagetables(vm_paddr_t *firstaddr)
 	pml4_entry_t *p4_p;
 	uint64_t DMPDkernphys;
 
-	npapdpphys = howmany(ptoa(Maxmem) / sizeof(struct vm_page), NBPML4);
-	if (npapdpphys > NPAPML4E) {
-		printf("NDMPML4E limits system to %lu GB\n",
-		    (NDMPML4E * 512) * (PAGE_SIZE / sizeof(struct vm_page)));
-		npapdpphys = NPAPML4E;
-		Maxmem = atop(NPAPML4E * NBPML4 *
-		    (PAGE_SIZE / sizeof(struct vm_page)));
-	}
-	PAPDPphys = allocpages(firstaddr, npapdpphys);
-
 	/* Allocate page table pages for the direct map */
 	ndmpdp = howmany(ptoa(Maxmem), NBPDP);
 	if (ndmpdp < 4)		/* Minimum 4GB of dirmap */
@@ -1587,12 +1574,6 @@ create_pagetables(vm_paddr_t *firstaddr)
 		p4_p[KPML4BASE + i] = KPDPphys + ptoa(i);
 		p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V;
 	}
-
-	/* Connect the page array slots up to the pml4. */
-	for (i = 0; i < npapdpphys; i++) {
-		p4_p[PAPML4I + i] = PAPDPphys + ptoa(i);
-		p4_p[PAPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx;
-	}
 }
 
 /*
@@ -3456,11 +3437,6 @@ pmap_pinit_pml4(vm_page_t pml4pg)
 		    X86_PG_V;
 	}
 
-	for (i = 0; i < npapdpphys; i++) {
-		pm_pml4[PAPML4I + i] = (PAPDPphys + ptoa(i)) | X86_PG_RW |
-		    X86_PG_V;
-	}
-
 	/* install self-referential address mapping entry(s) */
 	pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW |
 	    X86_PG_A | X86_PG_M;
@@ -3817,8 +3793,6 @@ pmap_release(pmap_t pmap)
 		pmap->pm_pml4[KPML4BASE + i] = 0;
 	for (i = 0; i < ndmpdpphys; i++)/* Direct Map */
 		pmap->pm_pml4[DMPML4I + i] = 0;
-	for (i = 0; i < npapdpphys; i++)
-		pmap->pm_pml4[PAPML4I + i] = 0;
 	pmap->pm_pml4[PML4PML4I] = 0;	/* Recursive Mapping */
 	for (i = 0; i < lm_ents; i++)	/* Large Map */
 		pmap->pm_pml4[LMSPML4I + i] = 0;
@@ -3856,6 +3830,10 @@ kvm_free(SYSCTL_HANDLER_ARGS)
 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 
     0, 0, kvm_free, "LU", "Amount of KVM free");
 
+/*
+ * Allocate physical memory for the vm_page array and map it into KVA,
+ * attempting to back the vm_pages with domain-local memory.
+ */
 void
 pmap_page_array_startup(long pages)
 {
@@ -3868,14 +3846,15 @@ pmap_page_array_startup(long pages)
 
 	vm_page_array_size = pages;
 
-	start = va = PA_MIN_ADDRESS;
+	start = va = VM_MIN_KERNEL_ADDRESS;
-	end = va + (pages * sizeof(struct vm_page));
+	end = va + pages * sizeof(struct vm_page);
 	while (va < end) {
-		pfn = first_page + ((va - start) / sizeof(struct vm_page));
+		pfn = first_page + (va - start) / sizeof(struct vm_page);
 		domain = _vm_phys_domain(ctob(pfn));
 		pdpe = pmap_pdpe(kernel_pmap, va);
 		if ((*pdpe & X86_PG_V) == 0) {
 			pa = vm_phys_early_alloc(domain, PAGE_SIZE);
+			dump_add_page(pa);
 			bzero((void *)PHYS_TO_DMAP(pa), PAGE_SIZE);
 			*pdpe = (pdp_entry_t)(pa | X86_PG_V | X86_PG_RW |
 			    X86_PG_A | X86_PG_M);
@@ -3892,6 +3871,7 @@ pmap_page_array_startup(long pages)
 		pde_store(pde, newpdir);
 		va += NBPDR;
 	}
+	vm_page_array = (vm_page_t)start;
 }
 
 /*

sys/amd64/include/vmparam.h

@@ -160,13 +160,12 @@
  * 0xffff808000000000 - 0xffff847fffffffff   large map (can be tuned up)
  * 0xffff848000000000 - 0xfffff7ffffffffff   unused (large map extends there)
  * 0xfffff80000000000 - 0xfffffbffffffffff   4TB direct map
- * 0xfffffc0000000000 - 0xfffffcffffffffff   unused
+ * 0xfffffc0000000000 - 0xfffffdffffffffff   unused
- * 0xfffffd0000000000 - 0xfffffd7fffffffff   page array 512GB
- * 0xfffffd8000000000 - 0xfffffdffffffffff   unused
  * 0xfffffe0000000000 - 0xffffffffffffffff   2TB kernel map
  *
  * Within the kernel map:
  *
+ * 0xfffffe0000000000                        vm_page_array
  * 0xffffffff80000000                        KERNBASE
  */
 
@@ -216,10 +215,10 @@
 	(x) & ~DMAP_MIN_ADDRESS; })
 
 /*
- * amd64 statically allocates the page array address so that it can
+ * amd64 maps the page array into KVA so that it can be more easily
- * be more easily allocated on the correct memory domains.
+ * allocated on the correct memory domains.
  */
-#define PMAP_HAS_PAGE_ARRAY	1
+#define	PMAP_HAS_PAGE_ARRAY	1
 
 /*
  * How many physical pages per kmem arena virtual page.

sys/vm/vm_kern.c

@@ -759,7 +759,7 @@ kmem_init(vm_offset_t start, vm_offset_t end)
 	vm_map_lock(m);
 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
 	kernel_map = m;
-	(void) vm_map_insert(m, NULL, (vm_ooffset_t) 0,
+	(void)vm_map_insert(m, NULL, 0,
 #ifdef __amd64__
 	    KERNBASE,
 #else		     
@@ -767,6 +767,18 @@ kmem_init(vm_offset_t start, vm_offset_t end)
 #endif
 	    start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
 	/* ... and ending with the completion of the above `insert' */
+
+#ifdef __amd64__
+	/*
+	 * Mark KVA used for the page array as allocated.  Other platforms
+	 * that handle vm_page_array allocation can simply adjust virtual_avail
+	 * instead.
+	 */
+	(void)vm_map_insert(m, NULL, 0, (vm_offset_t)vm_page_array,
+	    (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
+	    sizeof(struct vm_page)),
+	    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
+#endif
 	vm_map_unlock(m);
 
 	/*

sys/vm/vm_page.c

@@ -135,11 +135,7 @@ static int vm_pageproc_waiters;
  */
 vm_page_t bogus_page;
 
-#ifdef PMAP_HAS_PAGE_ARRAY
-vm_page_t vm_page_array = (vm_page_t)PA_MIN_ADDRESS;
-#else
 vm_page_t vm_page_array;
-#endif
 long vm_page_array_size;
 long first_page;