]> git.saurik.com Git - apple/xnu.git/blobdiff - osfmk/i386/i386_vm_init.c
xnu-4903.221.2.tar.gz
[apple/xnu.git] / osfmk / i386 / i386_vm_init.c
index 989895eb0904cde360ee093e239c97088373f031..9b74e8d449911ea58db3ba8e0e8edfa96ff2f799 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2003-2008 Apple Computer, Inc. All rights reserved.
+ * Copyright (c) 2003-2012 Apple Inc. All rights reserved.
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
@@ -54,8 +54,6 @@
  * the rights to redistribute these changes.
  */
 
-#include <platforms.h>
-#include <mach_kdb.h>
 
 #include <mach/i386/vm_param.h>
 
 #include <mach/thread_status.h>
 #include <pexpert/i386/efi.h>
 #include <i386/i386_lowmem.h>
-#include <i386/lowglobals.h>
+#include <x86_64/lowglobals.h>
 #include <i386/pal_routines.h>
 
 #include <mach-o/loader.h>
 #include <libkern/kernel_mach_header.h>
 
+
 vm_size_t      mem_size = 0; 
 pmap_paddr_t   first_avail = 0;/* first after page tables */
 
 uint64_t       max_mem;        /* Size of physical memory (bytes), adjusted by maxmem */
 uint64_t        mem_actual;
-uint64_t       sane_size = 0;  /* Memory size to use for defaults calculations */
+uint64_t       sane_size = 0;  /* Memory size for defaults calculations */
+
+/*
+ * KASLR parameters
+ */
+ppnum_t                vm_kernel_base_page;
+vm_offset_t    vm_kernel_base;
+vm_offset_t    vm_kernel_top;
+vm_offset_t    vm_kernel_stext;
+vm_offset_t    vm_kernel_etext;
+vm_offset_t    vm_kernel_slide;
+vm_offset_t    vm_kernel_slid_base;
+vm_offset_t    vm_kernel_slid_top;
+vm_offset_t vm_hib_base;
+vm_offset_t    vm_kext_base = VM_MIN_KERNEL_AND_KEXT_ADDRESS;
+vm_offset_t    vm_kext_top = VM_MIN_KERNEL_ADDRESS;
+
+vm_offset_t vm_prelink_stext;
+vm_offset_t vm_prelink_etext;
+vm_offset_t vm_prelink_sinfo;
+vm_offset_t vm_prelink_einfo;
+vm_offset_t vm_slinkedit;
+vm_offset_t vm_elinkedit;
+
+vm_offset_t vm_kernel_builtinkmod_text;
+vm_offset_t vm_kernel_builtinkmod_text_end;
 
 #define MAXLORESERVE   (32 * 1024 * 1024)
 
@@ -99,8 +123,10 @@ ppnum_t             lowest_lo = 0;
 ppnum_t                lowest_hi = 0;
 ppnum_t                highest_hi = 0;
 
+enum {PMAP_MAX_RESERVED_RANGES = 32};
 uint32_t pmap_reserved_pages_allocated = 0;
-uint32_t pmap_last_reserved_range = 0xFFFFFFFF;
+uint32_t pmap_reserved_range_indices[PMAP_MAX_RESERVED_RANGES];
+uint32_t pmap_last_reserved_range_index = 0;
 uint32_t pmap_reserved_ranges = 0;
 
 extern unsigned int bsd_mbuf_cluster_reserve(boolean_t *);
@@ -110,21 +136,22 @@ vm_offset_t       virtual_avail, virtual_end;
 static pmap_paddr_t    avail_remaining;
 vm_offset_t     static_memory_end = 0;
 
-vm_offset_t    sHIB, eHIB, stext, etext, sdata, edata, end;
+vm_offset_t    sHIB, eHIB, stext, etext, sdata, edata, end, sconst, econst;
 
 /*
  * _mh_execute_header is the mach_header for the currently executing kernel
  */
-void *sectTEXTB; unsigned long sectSizeTEXT;
-void *sectDATAB; unsigned long sectSizeDATA;
-void *sectOBJCB; unsigned long sectSizeOBJC;
-void *sectLINKB; unsigned long sectSizeLINK;
-void *sectPRELINKB; unsigned long sectSizePRELINK;
-void *sectHIBB; unsigned long sectSizeHIB;
-void *sectINITPTB; unsigned long sectSizeINITPT;
-
-kernel_segment_command_t *segTEXT;
-kernel_section_t *cursectTEXT, *lastsectTEXT;
+vm_offset_t segTEXTB; unsigned long segSizeTEXT;
+vm_offset_t segDATAB; unsigned long segSizeDATA;
+vm_offset_t segLINKB; unsigned long segSizeLINK;
+vm_offset_t segPRELINKTEXTB; unsigned long segSizePRELINKTEXT;
+vm_offset_t segPRELINKINFOB; unsigned long segSizePRELINKINFO;
+vm_offset_t segHIBB; unsigned long segSizeHIB;
+unsigned long segSizeConst;
+
+static kernel_segment_command_t *segTEXT, *segDATA;
+static kernel_section_t *cursectTEXT, *lastsectTEXT;
+static kernel_segment_command_t *segCONST;
 
 extern uint64_t firmware_Conventional_bytes;
 extern uint64_t firmware_RuntimeServices_bytes;
@@ -136,9 +163,62 @@ extern uint64_t firmware_Unusable_bytes;
 extern uint64_t firmware_other_bytes;
 uint64_t firmware_MMIO_bytes;
 
-#if    DEBUG
-#define        PRINT_PMAP_MEMORY_TABLE
+/*
+ * Linker magic to establish the highest address in the kernel.
+ */
+extern void    *last_kernel_symbol;
+
+boolean_t      memmap = FALSE;
+#if    DEBUG || DEVELOPMENT
+static void
+kprint_memmap(vm_offset_t maddr, unsigned int msize, unsigned int mcount) {
+    unsigned int         i;
+    unsigned int         j;
+    pmap_memory_region_t *p = pmap_memory_regions;
+    EfiMemoryRange       *mptr; 
+    addr64_t             region_start, region_end;
+    addr64_t             efi_start, efi_end;
+
+    for (j = 0; j < pmap_memory_region_count; j++, p++) {
+        kprintf("pmap region %d type %d base 0x%llx alloc_up 0x%llx alloc_down 0x%llx top 0x%llx\n",
+            j, p->type,
+            (addr64_t) p->base  << I386_PGSHIFT,
+            (addr64_t) p->alloc_up << I386_PGSHIFT,
+            (addr64_t) p->alloc_down << I386_PGSHIFT,
+            (addr64_t) p->end   << I386_PGSHIFT);
+        region_start = (addr64_t) p->base << I386_PGSHIFT;
+        region_end = ((addr64_t) p->end << I386_PGSHIFT) - 1;
+        mptr = (EfiMemoryRange *) maddr; 
+        for (i = 0; 
+             i < mcount;
+             i++, mptr = (EfiMemoryRange *)(((vm_offset_t)mptr) + msize)) {
+            if (mptr->Type != kEfiLoaderCode &&
+                mptr->Type != kEfiLoaderData &&
+                mptr->Type != kEfiBootServicesCode &&
+                mptr->Type != kEfiBootServicesData &&
+                mptr->Type != kEfiConventionalMemory) {
+                efi_start = (addr64_t)mptr->PhysicalStart;
+                efi_end = efi_start + ((vm_offset_t)mptr->NumberOfPages << I386_PGSHIFT) - 1;
+                if ((efi_start >= region_start && efi_start <= region_end) ||
+                    (efi_end >= region_start && efi_end <= region_end)) {
+                    kprintf(" *** Overlapping region with EFI runtime region %d\n", i);
+                }
+            }
+        }
+    }
+}
+#define DPRINTF(x...)  do { if (memmap) kprintf(x); } while (0)
+
+#else
+
+static void
+kprint_memmap(vm_offset_t maddr, unsigned int msize, unsigned int mcount) {
+#pragma unused(maddr, msize, mcount)
+}
+
+#define DPRINTF(x...)
 #endif /* DEBUG */
+
 /*
  * Basic VM initialization.
  */
@@ -152,74 +232,137 @@ i386_vm_init(uint64_t    maxmem,
        EfiMemoryRange *mptr;
         unsigned int mcount;
         unsigned int msize;
+       vm_offset_t maddr;
        ppnum_t fap;
        unsigned int i;
-       unsigned int safeboot;
        ppnum_t maxpg = 0;
         uint32_t pmap_type;
        uint32_t maxloreserve;
        uint32_t maxdmaaddr;
        uint32_t  mbuf_reserve = 0;
        boolean_t mbuf_override = FALSE;
+       boolean_t coalescing_permitted;
+       vm_kernel_base_page = i386_btop(args->kaddr);
+       vm_offset_t base_address;
+       vm_offset_t static_base_address;
+    
+       PE_parse_boot_argn("memmap", &memmap, sizeof(memmap));
 
-#if DEBUG
-       kprintf("Boot args revision: %d version: %d",
-               args->Revision, args->Version);
-       kprintf("  commandline: \"");
-       for(i=0; i<BOOT_LINE_LENGTH; i++)
-               kprintf("%c", args->CommandLine[i]);
-       kprintf("\"\n");
-#endif
+       /*
+        * Establish the KASLR parameters.
+        */
+       static_base_address = ml_static_ptovirt(KERNEL_BASE_OFFSET);
+       base_address        = ml_static_ptovirt(args->kaddr);
+       vm_kernel_slide     = base_address - static_base_address;
+       if (args->kslide) {
+               kprintf("KASLR slide: 0x%016lx dynamic\n", vm_kernel_slide);
+               if (vm_kernel_slide != ((vm_offset_t)args->kslide))
+                       panic("Kernel base inconsistent with slide - rebased?");
+       } else {
+               /* No slide relative to on-disk symbols */
+               kprintf("KASLR slide: 0x%016lx static and ignored\n",
+                       vm_kernel_slide);
+               vm_kernel_slide = 0;
+       }
+
+       /*
+        * Zero out local relocations to avoid confusing kxld.
+        * TODO: might be better to move this code to OSKext::initialize
+        */
+       if (_mh_execute_header.flags & MH_PIE) {
+               struct load_command *loadcmd;
+               uint32_t cmd;
+
+               loadcmd = (struct load_command *)((uintptr_t)&_mh_execute_header +
+                                                 sizeof (_mh_execute_header));
+
+               for (cmd = 0; cmd < _mh_execute_header.ncmds; cmd++) {
+                       if (loadcmd->cmd == LC_DYSYMTAB) {
+                               struct dysymtab_command *dysymtab;
+
+                               dysymtab = (struct dysymtab_command *)loadcmd;
+                               dysymtab->nlocrel = 0;
+                               dysymtab->locreloff = 0;
+                               kprintf("Hiding local relocations\n");
+                               break;
+                       }
+                       loadcmd = (struct load_command *)((uintptr_t)loadcmd + loadcmd->cmdsize);
+               }
+       }
 
        /*
         * Now retrieve addresses for end, edata, and etext 
         * from MACH-O headers.
         */
-
-       sectTEXTB = (void *) getsegdatafromheader(
-               &_mh_execute_header, "__TEXT", &sectSizeTEXT);
-       sectDATAB = (void *) getsegdatafromheader(
-               &_mh_execute_header, "__DATA", &sectSizeDATA);
-       sectOBJCB = (void *) getsegdatafromheader(
-               &_mh_execute_header, "__OBJC", &sectSizeOBJC);
-       sectLINKB = (void *) getsegdatafromheader(
-               &_mh_execute_header, "__LINKEDIT", &sectSizeLINK);
-       sectHIBB = (void *)getsegdatafromheader(
-               &_mh_execute_header, "__HIB", &sectSizeHIB);
-       sectINITPTB = (void *)getsegdatafromheader(
-               &_mh_execute_header, "__INITPT", &sectSizeINITPT);
-       sectPRELINKB = (void *) getsegdatafromheader(
-               &_mh_execute_header, "__PRELINK_TEXT", &sectSizePRELINK);
-
-       segTEXT = getsegbynamefromheader(&_mh_execute_header, "__TEXT");
+       segTEXTB = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                       "__TEXT", &segSizeTEXT);
+       segDATAB = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                       "__DATA", &segSizeDATA);
+       segLINKB = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                       "__LINKEDIT", &segSizeLINK);
+       segHIBB  = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                       "__HIB", &segSizeHIB);
+       segPRELINKTEXTB = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                               "__PRELINK_TEXT", &segSizePRELINKTEXT);
+       segPRELINKINFOB = (vm_offset_t) getsegdatafromheader(&_mh_execute_header,
+                                               "__PRELINK_INFO", &segSizePRELINKINFO);
+       segTEXT = getsegbynamefromheader(&_mh_execute_header,
+                                       "__TEXT");
+       segDATA = getsegbynamefromheader(&_mh_execute_header,
+                                       "__DATA");
+       segCONST = getsegbynamefromheader(&_mh_execute_header,
+                                       "__CONST");
        cursectTEXT = lastsectTEXT = firstsect(segTEXT);
        /* Discover the last TEXT section within the TEXT segment */
        while ((cursectTEXT = nextsect(segTEXT, cursectTEXT)) != NULL) {
                lastsectTEXT = cursectTEXT;
        }
 
-       sHIB  = (vm_offset_t) sectHIBB;
-       eHIB  = (vm_offset_t) sectHIBB + sectSizeHIB;
+       sHIB  = segHIBB;
+       eHIB  = segHIBB + segSizeHIB;
+       vm_hib_base = sHIB;
        /* Zero-padded from ehib to stext if text is 2M-aligned */
-       stext = (vm_offset_t) sectTEXTB;
+       stext = segTEXTB;
+       lowGlo.lgStext = stext;
        etext = (vm_offset_t) round_page_64(lastsectTEXT->addr + lastsectTEXT->size);
        /* Zero-padded from etext to sdata if text is 2M-aligned */
-       sdata = (vm_offset_t) sectDATAB;
-       edata = (vm_offset_t) sectDATAB + sectSizeDATA;
+       sdata = segDATAB;
+       edata = segDATAB + segSizeDATA;
 
-#if DEBUG
-       kprintf("sectTEXTB    = %p\n", sectTEXTB);
-       kprintf("sectDATAB    = %p\n", sectDATAB);
-       kprintf("sectOBJCB    = %p\n", sectOBJCB);
-       kprintf("sectLINKB    = %p\n", sectLINKB);
-       kprintf("sectHIBB     = %p\n", sectHIBB);
-       kprintf("sectPRELINKB = %p\n", sectPRELINKB);
-       kprintf("eHIB         = %p\n", (void *) eHIB);
-       kprintf("stext        = %p\n", (void *) stext);
-       kprintf("etext        = %p\n", (void *) etext);
-       kprintf("sdata        = %p\n", (void *) sdata);
-       kprintf("edata        = %p\n", (void *) edata);
-#endif
+       sconst = segCONST->vmaddr;
+       segSizeConst = segCONST->vmsize;
+       econst = sconst + segSizeConst;
+
+       assert(((sconst|econst) & PAGE_MASK) == 0);
+       
+       DPRINTF("segTEXTB    = %p\n", (void *) segTEXTB);
+       DPRINTF("segDATAB    = %p\n", (void *) segDATAB);
+       DPRINTF("segLINKB    = %p\n", (void *) segLINKB);
+       DPRINTF("segHIBB     = %p\n", (void *) segHIBB);
+       DPRINTF("segPRELINKTEXTB = %p\n", (void *) segPRELINKTEXTB);
+       DPRINTF("segPRELINKINFOB = %p\n", (void *) segPRELINKINFOB);
+       DPRINTF("sHIB        = %p\n", (void *) sHIB);
+       DPRINTF("eHIB        = %p\n", (void *) eHIB);
+       DPRINTF("stext       = %p\n", (void *) stext);
+       DPRINTF("etext       = %p\n", (void *) etext);
+       DPRINTF("sdata       = %p\n", (void *) sdata);
+       DPRINTF("edata       = %p\n", (void *) edata);
+       DPRINTF("sconst      = %p\n", (void *) sconst);
+       DPRINTF("econst      = %p\n", (void *) econst);
+       DPRINTF("kernel_top  = %p\n", (void *) &last_kernel_symbol);
+
+       vm_kernel_base  = sHIB;
+       vm_kernel_top   = (vm_offset_t) &last_kernel_symbol;
+       vm_kernel_stext = stext;
+       vm_kernel_etext = etext;
+       vm_prelink_stext = segPRELINKTEXTB;
+       vm_prelink_etext = segPRELINKTEXTB + segSizePRELINKTEXT;
+       vm_prelink_sinfo = segPRELINKINFOB;
+       vm_prelink_einfo = segPRELINKINFOB + segSizePRELINKINFO;
+       vm_slinkedit = segLINKB;
+       vm_elinkedit = segLINKB + segSizeLINK;
+       vm_kernel_slid_base = vm_kext_base + vm_kernel_slide;
+       vm_kernel_slid_top = vm_prelink_einfo;
 
        vm_set_page_size();
 
@@ -227,9 +370,6 @@ i386_vm_init(uint64_t       maxmem,
         * Compute the memory size.
         */
 
-       if ((1 == vm_himemory_mode) || PE_parse_boot_argn("-x", &safeboot, sizeof (safeboot))) {
-               maxpg = 1 << (32 - I386_PGSHIFT);
-       }
        avail_remaining = 0;
        avail_end = 0;
        pmptr = pmap_memory_regions;
@@ -237,7 +377,8 @@ i386_vm_init(uint64_t       maxmem,
        pmap_memory_region_count = pmap_memory_region_current = 0;
        fap = (ppnum_t) i386_btop(first_avail);
 
-       mptr = (EfiMemoryRange *)ml_static_ptovirt((vm_offset_t)args->MemoryMap);
+       maddr = ml_static_ptovirt((vm_offset_t)args->MemoryMap);
+       mptr = (EfiMemoryRange *)maddr;
         if (args->MemoryMapDescriptorSize == 0)
                panic("Invalid memory map descriptor size");
         msize = args->MemoryMapDescriptorSize;
@@ -256,6 +397,38 @@ i386_vm_init(uint64_t      maxmem,
                }
                base = (ppnum_t) (mptr->PhysicalStart >> I386_PGSHIFT);
                top = (ppnum_t) (((mptr->PhysicalStart) >> I386_PGSHIFT) + mptr->NumberOfPages - 1);
+
+               if (base == 0) {
+                       /*
+                        * Avoid having to deal with the edge case of the 
+                        * very first possible physical page and the roll-over
+                        * to -1; just ignore that page.
+                        */
+                       kprintf("WARNING: ignoring first page in [0x%llx:0x%llx]\n", (uint64_t) base, (uint64_t) top);
+                       base++;
+               }
+               if (top + 1 == 0) {
+                       /*
+                        * Avoid having to deal with the edge case of the 
+                        * very last possible physical page and the roll-over
+                        * to 0; just ignore that page.
+                        */
+                       kprintf("WARNING: ignoring last page in [0x%llx:0x%llx]\n", (uint64_t) base, (uint64_t) top);
+                       top--;
+               }
+               if (top < base) {
+                       /*
+                        * That was the only page in that region, so
+                        * ignore the whole region.
+                        */
+                       continue;
+               }
+
+#if    MR_RSV_TEST
+               static uint32_t nmr = 0;
+               if ((base > 0x20000) && (nmr++ < 4))
+                       mptr->Attribute |= EFI_MEMORY_KERN_RESERVED;
+#endif
                region_bytes = (uint64_t)(mptr->NumberOfPages << I386_PGSHIFT);
                pmap_type = mptr->Type;
 
@@ -276,15 +449,10 @@ i386_vm_init(uint64_t     maxmem,
                         * sane_size should reflect the total amount of physical
                         * RAM in the system, not just the amount that is
                         * available for the OS to use.
-                        * FIXME:Consider deriving this value from SMBIOS tables
+                        * We now get this value from SMBIOS tables
                         * rather than reverse engineering the memory map.
-                        * Alternatively, see
-                        * <rdar://problem/4642773> Memory map should
-                        * describe all memory
-                        * Firmware on some systems guarantees that the memory
-                        * map is complete via the "RomReservedMemoryTracked"
-                        * feature field--consult that where possible to
-                        * avoid the "round up to 128M" workaround below.
+                        * But the legacy computation of "sane_size" is kept
+                        * for diagnostic information.
                         */
 
                case kEfiRuntimeServicesCode:
@@ -320,10 +488,10 @@ i386_vm_init(uint64_t     maxmem,
                        break;
                }
 
-#if DEBUG
-               kprintf("EFI region %d: type %u/%d, base 0x%x, top 0x%x\n",
-                       i, mptr->Type, pmap_type, base, top);
-#endif
+               DPRINTF("EFI region %d: type %u/%d, base 0x%x, top 0x%x %s\n",
+                   i, mptr->Type, pmap_type, base, top,
+                   (mptr->Attribute&EFI_MEMORY_KERN_RESERVED)? "RESERVED" :
+                   (mptr->Attribute&EFI_MEMORY_RUNTIME)? "RUNTIME" : "");
 
                if (maxpg) {
                        if (base >= maxpg)
@@ -347,6 +515,19 @@ i386_vm_init(uint64_t      maxmem,
                                prev_pmptr = 0;
                                continue;
                        }
+                       /*
+                        * A range may be marked with with the
+                        * EFI_MEMORY_KERN_RESERVED attribute
+                        * on some systems, to indicate that the range
+                        * must not be made available to devices.
+                        */
+
+                       if (mptr->Attribute & EFI_MEMORY_KERN_RESERVED) {
+                               if (++pmap_reserved_ranges > PMAP_MAX_RESERVED_RANGES) {
+                                       panic("Too many reserved ranges %u\n", pmap_reserved_ranges);
+                               }
+                       }
+
                        if (top < fap) {
                                /*
                                 * entire range below first_avail
@@ -361,29 +542,22 @@ i386_vm_init(uint64_t     maxmem,
 
                                pmptr->end = top;
 
-                               /*
-                                * A range may be marked with with the
-                                * EFI_MEMORY_KERN_RESERVED attribute
-                                * on some systems, to indicate that the range
-                                * must not be made available to devices.
-                                * Simplifying assumptions are made regarding
-                                * the placement of the range.
-                                */
-                               if (mptr->Attribute & EFI_MEMORY_KERN_RESERVED)
-                                       pmap_reserved_ranges++;
 
                                if ((mptr->Attribute & EFI_MEMORY_KERN_RESERVED) &&
-                                   (top < I386_KERNEL_IMAGE_BASE_PAGE)) {
-                                       pmptr->alloc = pmptr->base;
-                                       pmap_last_reserved_range = pmap_memory_region_count;
+                                   (top < vm_kernel_base_page)) {
+                                       pmptr->alloc_up = pmptr->base;
+                                       pmptr->alloc_down = pmptr->end;
+                                       pmap_reserved_range_indices[pmap_last_reserved_range_index++] = pmap_memory_region_count;
                                }
                                else {
                                        /*
                                         * mark as already mapped
                                         */
-                                       pmptr->alloc = top;
+                                       pmptr->alloc_up = top + 1;
+                                       pmptr->alloc_down = top;
                                }
                                pmptr->type = pmap_type;
+                               pmptr->attribute = mptr->Attribute;
                        }
                        else if ( (base < fap) && (top > fap) ) {
                                /*
@@ -392,43 +566,54 @@ i386_vm_init(uint64_t     maxmem,
                                 * mark already allocated
                                 */
                                pmptr->base = base;
-                               pmptr->alloc = pmptr->end = (fap - 1);
+                               pmptr->end = (fap - 1);
+                               pmptr->alloc_up = pmptr->end + 1;
+                               pmptr->alloc_down = pmptr->end;
                                pmptr->type = pmap_type;
+                               pmptr->attribute = mptr->Attribute;
                                /*
                                 * we bump these here inline so the accounting
                                 * below works correctly
                                 */
                                pmptr++;
                                pmap_memory_region_count++;
-                               pmptr->alloc = pmptr->base = fap;
+
+                               pmptr->alloc_up = pmptr->base = fap;
                                pmptr->type = pmap_type;
-                               pmptr->end = top;
-                       }
-                       else {
+                               pmptr->attribute = mptr->Attribute;
+                               pmptr->alloc_down = pmptr->end = top;
+
+                               if (mptr->Attribute & EFI_MEMORY_KERN_RESERVED)
+                                       pmap_reserved_range_indices[pmap_last_reserved_range_index++] = pmap_memory_region_count;
+                       } else {
                                /*
                                 * entire range useable
                                 */
-                               pmptr->alloc = pmptr->base = base;
+                               pmptr->alloc_up = pmptr->base = base;
                                pmptr->type = pmap_type;
-                               pmptr->end = top;
+                               pmptr->attribute = mptr->Attribute;
+                               pmptr->alloc_down = pmptr->end = top;
+                               if (mptr->Attribute & EFI_MEMORY_KERN_RESERVED)
+                                       pmap_reserved_range_indices[pmap_last_reserved_range_index++] = pmap_memory_region_count;
                        }
 
                        if (i386_ptob(pmptr->end) > avail_end )
                                avail_end = i386_ptob(pmptr->end);
 
                        avail_remaining += (pmptr->end - pmptr->base);
-
+                       coalescing_permitted = (prev_pmptr && (pmptr->attribute == prev_pmptr->attribute) && ((pmptr->attribute & EFI_MEMORY_KERN_RESERVED) == 0));
                        /*
                         * Consolidate contiguous memory regions, if possible
                         */
                        if (prev_pmptr &&
-                           pmptr->type == prev_pmptr->type &&
-                           pmptr->base == pmptr->alloc &&
-                               pmptr->base == (prev_pmptr->end + 1))
+                           (pmptr->type == prev_pmptr->type) &&
+                           (coalescing_permitted) &&
+                           (pmptr->base == pmptr->alloc_up) &&
+                           (prev_pmptr->end == prev_pmptr->alloc_down) &&
+                           (pmptr->base == (prev_pmptr->end + 1)))
                        {
-                               if(prev_pmptr->end == prev_pmptr->alloc)
-                                       prev_pmptr->alloc = pmptr->base;
                                prev_pmptr->end = pmptr->end;
+                               prev_pmptr->alloc_down = pmptr->alloc_down;
                        } else {
                                pmap_memory_region_count++;
                                prev_pmptr = pmptr;
@@ -437,48 +622,24 @@ i386_vm_init(uint64_t     maxmem,
                }
        }
 
-#ifdef PRINT_PMAP_MEMORY_TABLE
-       {
-        unsigned int j;
-        pmap_memory_region_t *p = pmap_memory_regions;
-        addr64_t region_start, region_end;
-        addr64_t efi_start, efi_end;
-        for (j=0;j<pmap_memory_region_count;j++, p++) {
-            kprintf("pmap region %d type %d base 0x%llx alloc 0x%llx top 0x%llx\n",
-                   j, p->type,
-                    (addr64_t) p->base  << I386_PGSHIFT,
-                   (addr64_t) p->alloc << I386_PGSHIFT,
-                   (addr64_t) p->end   << I386_PGSHIFT);
-            region_start = (addr64_t) p->base << I386_PGSHIFT;
-            region_end = ((addr64_t) p->end << I386_PGSHIFT) - 1;
-           mptr = (EfiMemoryRange *) ml_static_ptovirt((vm_offset_t)args->MemoryMap);
-            for (i=0; i<mcount; i++, mptr = (EfiMemoryRange *)(((vm_offset_t)mptr) + msize)) {
-                if (mptr->Type != kEfiLoaderCode &&
-                    mptr->Type != kEfiLoaderData &&
-                    mptr->Type != kEfiBootServicesCode &&
-                    mptr->Type != kEfiBootServicesData &&
-                    mptr->Type != kEfiConventionalMemory) {
-                efi_start = (addr64_t)mptr->PhysicalStart;
-                efi_end = efi_start + ((vm_offset_t)mptr->NumberOfPages << I386_PGSHIFT) - 1;
-                if ((efi_start >= region_start && efi_start <= region_end) ||
-                    (efi_end >= region_start && efi_end <= region_end)) {
-                    kprintf(" *** Overlapping region with EFI runtime region %d\n", i);
-                }
-              }
-            }
-          }
+       if (memmap) {
+               kprint_memmap(maddr, msize, mcount);
        }
-#endif
 
        avail_start = first_avail;
-       mem_actual = sane_size;
+       mem_actual = args->PhysicalMemorySize;
 
        /*
-        * For user visible memory size, round up to 128 Mb - accounting for the various stolen memory
-        * not reported by EFI.
+        * For user visible memory size, round up to 128 Mb
+        * - accounting for the various stolen memory not reported by EFI.
+        * This is maintained for historical, comparison purposes but
+        * we now use the memory size reported by EFI/Booter.
         */
-
        sane_size = (sane_size + 128 * MB - 1) & ~((uint64_t)(128 * MB - 1));
+       if (sane_size != mem_actual)
+               printf("mem_actual: 0x%llx\n legacy sane_size: 0x%llx\n",
+                       mem_actual, sane_size);
+       sane_size = mem_actual;
 
        /*
         * We cap at KERNEL_MAXMEM bytes (currently 32GB for K32, 96GB for K64).
@@ -497,7 +658,7 @@ i386_vm_init(uint64_t       maxmem,
        if ( (maxmem > (uint64_t)first_avail) && (maxmem < sane_size)) {
                ppnum_t discarded_pages  = (ppnum_t)((sane_size - maxmem) >> I386_PGSHIFT);
                ppnum_t highest_pn = 0;
-               ppnum_t cur_alloc  = 0;
+               ppnum_t cur_end  = 0;
                uint64_t        pages_to_use;
                unsigned        cur_region = 0;
 
@@ -511,15 +672,17 @@ i386_vm_init(uint64_t     maxmem,
                pages_to_use = avail_remaining;
 
                while (cur_region < pmap_memory_region_count && pages_to_use) {
-                       for (cur_alloc = pmap_memory_regions[cur_region].alloc;
-                            cur_alloc < pmap_memory_regions[cur_region].end && pages_to_use;
-                            cur_alloc++) {
-                               if (cur_alloc > highest_pn)
-                                       highest_pn = cur_alloc;
+                       for (cur_end = pmap_memory_regions[cur_region].base;
+                            cur_end < pmap_memory_regions[cur_region].end && pages_to_use;
+                            cur_end++) {
+                               if (cur_end > highest_pn)
+                                       highest_pn = cur_end;
                                pages_to_use--;
                        }
-                       if (pages_to_use == 0)
-                               pmap_memory_regions[cur_region].end = cur_alloc;
+                       if (pages_to_use == 0) {
+                               pmap_memory_regions[cur_region].end = cur_end;
+                               pmap_memory_regions[cur_region].alloc_down = cur_end;
+                       }
 
                        cur_region++;
                }
@@ -561,7 +724,9 @@ i386_vm_init(uint64_t       maxmem,
                        else
                                maxloreserve = MAXLORESERVE / PAGE_SIZE;
 
+#if SOCKETS
                        mbuf_reserve = bsd_mbuf_cluster_reserve(&mbuf_override) / PAGE_SIZE;
+#endif
                } else
                        maxloreserve = (maxloreserve * (1024 * 1024)) / PAGE_SIZE;
 
@@ -603,13 +768,14 @@ boolean_t pmap_next_page_reserved(ppnum_t *);
  */
 boolean_t
 pmap_next_page_reserved(ppnum_t *pn) {
-       if (pmap_reserved_ranges && pmap_last_reserved_range != 0xFFFFFFFF) {
+       if (pmap_reserved_ranges) {
                uint32_t n;
                pmap_memory_region_t *region;
-               for (n = 0; n <= pmap_last_reserved_range; n++) {
-                       region = &pmap_memory_regions[n];
-                       if (region->alloc < region->end) {
-                               *pn = region->alloc++;
+               for (n = 0; n < pmap_last_reserved_range_index; n++) {
+                       uint32_t reserved_index = pmap_reserved_range_indices[n];
+                       region = &pmap_memory_regions[reserved_index];
+                       if (region->alloc_up <= region->alloc_down) {
+                               *pn = region->alloc_up++;
                                avail_remaining--;
 
                                if (*pn > max_ppnum)
@@ -619,6 +785,11 @@ pmap_next_page_reserved(ppnum_t *pn) {
                                        lowest_lo = *pn;
 
                                pmap_reserved_pages_allocated++;
+#if DEBUG
+                               if (region->alloc_up > region->alloc_down) {
+                                       kprintf("Exhausted reserved range index: %u, base: 0x%x end: 0x%x, type: 0x%x, attribute: 0x%llx\n", reserved_index, region->base, region->end, region->type, region->attribute);
+                               }
+#endif
                                return TRUE;
                        }
                }
@@ -641,8 +812,8 @@ pmap_next_page_hi(
                for (n = pmap_memory_region_count - 1; n >= 0; n--) {
                        region = &pmap_memory_regions[n];
 
-                       if (region->alloc != region->end) {
-                               *pn = region->alloc++;
+                       if (region->alloc_down >= region->alloc_up) {
+                               *pn = region->alloc_down--;
                                avail_remaining--;
 
                                if (*pn > max_ppnum)
@@ -670,12 +841,12 @@ pmap_next_page(
               ppnum_t *pn)
 {
        if (avail_remaining) while (pmap_memory_region_current < pmap_memory_region_count) {
-               if (pmap_memory_regions[pmap_memory_region_current].alloc ==
-                               pmap_memory_regions[pmap_memory_region_current].end) {
+               if (pmap_memory_regions[pmap_memory_region_current].alloc_up >
+                   pmap_memory_regions[pmap_memory_region_current].alloc_down) {
                        pmap_memory_region_current++;
                        continue;
                }
-               *pn = pmap_memory_regions[pmap_memory_region_current].alloc++;
+               *pn = pmap_memory_regions[pmap_memory_region_current].alloc_up++;
                avail_remaining--;
 
                if (*pn > max_ppnum)