arch/parisc/mm/init.c

   1 /*
   2  *  linux/arch/parisc/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright 1999 SuSE GmbH
   6  *    changed by Philipp Rumpf
   7  *  Copyright 1999 Philipp Rumpf (prumpf@tux.org)
   8  *  Copyright 2004 Randolph Chung (tausq@debian.org)
   9  *  Copyright 2006-2007 Helge Deller (deller@gmx.de)
  10  *
  11  */
  12
  13
  14 #include <linux/module.h>
  15 #include <linux/mm.h>
  16 #include <linux/bootmem.h>
  17 #include <linux/gfp.h>
  18 #include <linux/delay.h>
  19 #include <linux/init.h>
  20 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
  21 #include <linux/initrd.h>
  22 #include <linux/swap.h>
  23 #include <linux/unistd.h>
  24 #include <linux/nodemask.h>     /* for node_online_map */
  25 #include <linux/pagemap.h>      /* for release_pages and page_cache_release */
  26
  27 #include <asm/pgalloc.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/tlb.h>
  30 #include <asm/pdc_chassis.h>
  31 #include <asm/mmzone.h>
  32 #include <asm/sections.h>
  33
  34 extern int  data_start;
  35
  36 #if PT_NLEVELS == 3
  37 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
  38  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
  39  * guarantee that global objects will be laid out in memory in the same order
  40  * as the order of declaration, so put these in different sections and use
  41  * the linker script to order them. */
  42 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
  43 #endif
  44
  45 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
  46 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
  47
  48 #ifdef CONFIG_DISCONTIGMEM
  49 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
  50 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
  51 #endif
  52
  53 static struct resource data_resource = {
  54         .name   = "Kernel data",
  55         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  56 };
  57
  58 static struct resource code_resource = {
  59         .name   = "Kernel code",
  60         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  61 };
  62
  63 static struct resource pdcdata_resource = {
  64         .name   = "PDC data (Page Zero)",
  65         .start  = 0,
  66         .end    = 0x9ff,
  67         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  68 };
  69
  70 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
  71
  72 /* The following array is initialized from the firmware specific
  73  * information retrieved in kernel/inventory.c.
  74  */
  75
  76 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
  77 int npmem_ranges __read_mostly;
  78
  79 #ifdef CONFIG_64BIT
  80 #define MAX_MEM         (~0UL)
  81 #else /* !CONFIG_64BIT */
  82 #define MAX_MEM         (3584U*1024U*1024U)
  83 #endif /* !CONFIG_64BIT */
  84
  85 static unsigned long mem_limit __read_mostly = MAX_MEM;
  86
  87 static void __init mem_limit_func(void)
  88 {
  89         char *cp, *end;
  90         unsigned long limit;
  91
  92         /* We need this before __setup() functions are called */
  93
  94         limit = MAX_MEM;
  95         for (cp = boot_command_line; *cp; ) {
  96                 if (memcmp(cp, "mem=", 4) == 0) {
  97                         cp += 4;
  98                         limit = memparse(cp, &end);
  99                         if (end != cp)
 100                                 break;
 101                         cp = end;
 102                 } else {
 103                         while (*cp != ' ' && *cp)
 104                                 ++cp;
 105                         while (*cp == ' ')
 106                                 ++cp;
 107                 }
 108         }
 109
 110         if (limit < mem_limit)
 111                 mem_limit = limit;
 112 }
 113
 114 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 115
 116 static void __init setup_bootmem(void)
 117 {
 118         unsigned long bootmap_size;
 119         unsigned long mem_max;
 120         unsigned long bootmap_pages;
 121         unsigned long bootmap_start_pfn;
 122         unsigned long bootmap_pfn;
 123 #ifndef CONFIG_DISCONTIGMEM
 124         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 125         int npmem_holes;
 126 #endif
 127         int i, sysram_resource_count;
 128
 129         disable_sr_hashing(); /* Turn off space register hashing */
 130
 131         /*
 132          * Sort the ranges. Since the number of ranges is typically
 133          * small, and performance is not an issue here, just do
 134          * a simple insertion sort.
 135          */
 136
 137         for (i = 1; i < npmem_ranges; i++) {
 138                 int j;
 139
 140                 for (j = i; j > 0; j--) {
 141                         unsigned long tmp;
 142
 143                         if (pmem_ranges[j-1].start_pfn <
 144                             pmem_ranges[j].start_pfn) {
 145
 146                                 break;
 147                         }
 148                         tmp = pmem_ranges[j-1].start_pfn;
 149                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 150                         pmem_ranges[j].start_pfn = tmp;
 151                         tmp = pmem_ranges[j-1].pages;
 152                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 153                         pmem_ranges[j].pages = tmp;
 154                 }
 155         }
 156
 157 #ifndef CONFIG_DISCONTIGMEM
 158         /*
 159          * Throw out ranges that are too far apart (controlled by
 160          * MAX_GAP).
 161          */
 162
 163         for (i = 1; i < npmem_ranges; i++) {
 164                 if (pmem_ranges[i].start_pfn -
 165                         (pmem_ranges[i-1].start_pfn +
 166                          pmem_ranges[i-1].pages) > MAX_GAP) {
 167                         npmem_ranges = i;
 168                         printk("Large gap in memory detected (%ld pages). "
 169                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 170                                pmem_ranges[i].start_pfn -
 171                                (pmem_ranges[i-1].start_pfn +
 172                                 pmem_ranges[i-1].pages));
 173                         break;
 174                 }
 175         }
 176 #endif
 177
 178         if (npmem_ranges > 1) {
 179
 180                 /* Print the memory ranges */
 181
 182                 printk(KERN_INFO "Memory Ranges:\n");
 183
 184                 for (i = 0; i < npmem_ranges; i++) {
 185                         unsigned long start;
 186                         unsigned long size;
 187
 188                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
 189                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 190                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 191                                 i,start, start + (size - 1), size >> 20);
 192                 }
 193         }
 194
 195         sysram_resource_count = npmem_ranges;
 196         for (i = 0; i < sysram_resource_count; i++) {
 197                 struct resource *res = &sysram_resources[i];
 198                 res->name = "System RAM";
 199                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
 200                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
 201                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 202                 request_resource(&iomem_resource, res);
 203         }
 204
 205         /*
 206          * For 32 bit kernels we limit the amount of memory we can
 207          * support, in order to preserve enough kernel address space
 208          * for other purposes. For 64 bit kernels we don't normally
 209          * limit the memory, but this mechanism can be used to
 210          * artificially limit the amount of memory (and it is written
 211          * to work with multiple memory ranges).
 212          */
 213
 214         mem_limit_func();       /* check for "mem=" argument */
 215
 216         mem_max = 0;
 217         num_physpages = 0;
 218         for (i = 0; i < npmem_ranges; i++) {
 219                 unsigned long rsize;
 220
 221                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 222                 if ((mem_max + rsize) > mem_limit) {
 223                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 224                         if (mem_max == mem_limit)
 225                                 npmem_ranges = i;
 226                         else {
 227                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 228                                                        - (mem_max >> PAGE_SHIFT);
 229                                 npmem_ranges = i + 1;
 230                                 mem_max = mem_limit;
 231                         }
 232                 num_physpages += pmem_ranges[i].pages;
 233                         break;
 234                 }
 235             num_physpages += pmem_ranges[i].pages;
 236                 mem_max += rsize;
 237         }
 238
 239         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 240
 241 #ifndef CONFIG_DISCONTIGMEM
 242         /* Merge the ranges, keeping track of the holes */
 243
 244         {
 245                 unsigned long end_pfn;
 246                 unsigned long hole_pages;
 247
 248                 npmem_holes = 0;
 249                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 250                 for (i = 1; i < npmem_ranges; i++) {
 251
 252                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 253                         if (hole_pages) {
 254                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 255                                 pmem_holes[npmem_holes++].pages = hole_pages;
 256                                 end_pfn += hole_pages;
 257                         }
 258                         end_pfn += pmem_ranges[i].pages;
 259                 }
 260
 261                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 262                 npmem_ranges = 1;
 263         }
 264 #endif
 265
 266         bootmap_pages = 0;
 267         for (i = 0; i < npmem_ranges; i++)
 268                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
 269
 270         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
 271
 272 #ifdef CONFIG_DISCONTIGMEM
 273         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 274                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 275                 NODE_DATA(i)->bdata = &bootmem_node_data[i];
 276         }
 277         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 278
 279         for (i = 0; i < npmem_ranges; i++) {
 280                 node_set_state(i, N_NORMAL_MEMORY);
 281                 node_set_online(i);
 282         }
 283 #endif
 284
 285         /*
 286          * Initialize and free the full range of memory in each range.
 287          * Note that the only writing these routines do are to the bootmap,
 288          * and we've made sure to locate the bootmap properly so that they
 289          * won't be writing over anything important.
 290          */
 291
 292         bootmap_pfn = bootmap_start_pfn;
 293         max_pfn = 0;
 294         for (i = 0; i < npmem_ranges; i++) {
 295                 unsigned long start_pfn;
 296                 unsigned long npages;
 297
 298                 start_pfn = pmem_ranges[i].start_pfn;
 299                 npages = pmem_ranges[i].pages;
 300
 301                 bootmap_size = init_bootmem_node(NODE_DATA(i),
 302                                                 bootmap_pfn,
 303                                                 start_pfn,
 304                                                 (start_pfn + npages) );
 305                 free_bootmem_node(NODE_DATA(i),
 306                                   (start_pfn << PAGE_SHIFT),
 307                                   (npages << PAGE_SHIFT) );
 308                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 309                 if ((start_pfn + npages) > max_pfn)
 310                         max_pfn = start_pfn + npages;
 311         }
 312
 313         /* IOMMU is always used to access "high mem" on those boxes
 314          * that can support enough mem that a PCI device couldn't
 315          * directly DMA to any physical addresses.
 316          * ISA DMA support will need to revisit this.
 317          */
 318         max_low_pfn = max_pfn;
 319
 320         /* bootmap sizing messed up? */
 321         BUG_ON((bootmap_pfn - bootmap_start_pfn) != bootmap_pages);
 322
 323         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 324
 325 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 326
 327         reserve_bootmem_node(NODE_DATA(0), 0UL,
 328                         (unsigned long)(PAGE0->mem_free +
 329                                 PDC_CONSOLE_IO_IODC_SIZE), BOOTMEM_DEFAULT);
 330         reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
 331                         (unsigned long)(_end - _text), BOOTMEM_DEFAULT);
 332         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
 333                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT),
 334                         BOOTMEM_DEFAULT);
 335
 336 #ifndef CONFIG_DISCONTIGMEM
 337
 338         /* reserve the holes */
 339
 340         for (i = 0; i < npmem_holes; i++) {
 341                 reserve_bootmem_node(NODE_DATA(0),
 342                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
 343                                 (pmem_holes[i].pages << PAGE_SHIFT),
 344                                 BOOTMEM_DEFAULT);
 345         }
 346 #endif
 347
 348 #ifdef CONFIG_BLK_DEV_INITRD
 349         if (initrd_start) {
 350                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 351                 if (__pa(initrd_start) < mem_max) {
 352                         unsigned long initrd_reserve;
 353
 354                         if (__pa(initrd_end) > mem_max) {
 355                                 initrd_reserve = mem_max - __pa(initrd_start);
 356                         } else {
 357                                 initrd_reserve = initrd_end - initrd_start;
 358                         }
 359                         initrd_below_start_ok = 1;
 360                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 361
 362                         reserve_bootmem_node(NODE_DATA(0), __pa(initrd_start),
 363                                         initrd_reserve, BOOTMEM_DEFAULT);
 364                 }
 365         }
 366 #endif
 367
 368         data_resource.start =  virt_to_phys(&data_start);
 369         data_resource.end = virt_to_phys(_end) - 1;
 370         code_resource.start = virt_to_phys(_text);
 371         code_resource.end = virt_to_phys(&data_start)-1;
 372
 373         /* We don't know which region the kernel will be in, so try
 374          * all of them.
 375          */
 376         for (i = 0; i < sysram_resource_count; i++) {
 377                 struct resource *res = &sysram_resources[i];
 378                 request_resource(res, &code_resource);
 379                 request_resource(res, &data_resource);
 380         }
 381         request_resource(&sysram_resources[0], &pdcdata_resource);
 382 }
 383
 384 static void __init map_pages(unsigned long start_vaddr,
 385                              unsigned long start_paddr, unsigned long size,
 386                              pgprot_t pgprot, int force)
 387 {
 388         pgd_t *pg_dir;
 389         pmd_t *pmd;
 390         pte_t *pg_table;
 391         unsigned long end_paddr;
 392         unsigned long start_pmd;
 393         unsigned long start_pte;
 394         unsigned long tmp1;
 395         unsigned long tmp2;
 396         unsigned long address;
 397         unsigned long vaddr;
 398         unsigned long ro_start;
 399         unsigned long ro_end;
 400         unsigned long fv_addr;
 401         unsigned long gw_addr;
 402         extern const unsigned long fault_vector_20;
 403         extern void * const linux_gateway_page;
 404
 405         ro_start = __pa((unsigned long)_text);
 406         ro_end   = __pa((unsigned long)&data_start);
 407         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
 408         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
 409
 410         end_paddr = start_paddr + size;
 411
 412         pg_dir = pgd_offset_k(start_vaddr);
 413
 414 #if PTRS_PER_PMD == 1
 415         start_pmd = 0;
 416 #else
 417         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 418 #endif
 419         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 420
 421         address = start_paddr;
 422         vaddr = start_vaddr;
 423         while (address < end_paddr) {
 424 #if PTRS_PER_PMD == 1
 425                 pmd = (pmd_t *)__pa(pg_dir);
 426 #else
 427                 pmd = (pmd_t *)pgd_address(*pg_dir);
 428
 429                 /*
 430                  * pmd is physical at this point
 431                  */
 432
 433                 if (!pmd) {
 434                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0), PAGE_SIZE << PMD_ORDER);
 435                         pmd = (pmd_t *) __pa(pmd);
 436                 }
 437
 438                 pgd_populate(NULL, pg_dir, __va(pmd));
 439 #endif
 440                 pg_dir++;
 441
 442                 /* now change pmd to kernel virtual addresses */
 443
 444                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 445                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
 446
 447                         /*
 448                          * pg_table is physical at this point
 449                          */
 450
 451                         pg_table = (pte_t *)pmd_address(*pmd);
 452                         if (!pg_table) {
 453                                 pg_table = (pte_t *)
 454                                         alloc_bootmem_low_pages_node(NODE_DATA(0), PAGE_SIZE);
 455                                 pg_table = (pte_t *) __pa(pg_table);
 456                         }
 457
 458                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 459
 460                         /* now change pg_table to kernel virtual addresses */
 461
 462                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 463                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
 464                                 pte_t pte;
 465
 466                                 /*
 467                                  * Map the fault vector writable so we can
 468                                  * write the HPMC checksum.
 469                                  */
 470                                 if (force)
 471                                         pte =  __mk_pte(address, pgprot);
 472                                 else if (core_kernel_text(vaddr) &&
 473                                          address != fv_addr)
 474                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 475                                 else
 476 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
 477                                 if (address >= ro_start && address < ro_end
 478                                                         && address != fv_addr
 479                                                         && address != gw_addr)
 480                                         pte = __mk_pte(address, PAGE_KERNEL_RO);
 481                                 else
 482 #endif
 483                                         pte = __mk_pte(address, pgprot);
 484
 485                                 if (address >= end_paddr) {
 486                                         if (force)
 487                                                 break;
 488                                         else
 489                                                 pte_val(pte) = 0;
 490                                 }
 491
 492                                 set_pte(pg_table, pte);
 493
 494                                 address += PAGE_SIZE;
 495                                 vaddr += PAGE_SIZE;
 496                         }
 497                         start_pte = 0;
 498
 499                         if (address >= end_paddr)
 500                             break;
 501                 }
 502                 start_pmd = 0;
 503         }
 504 }
 505
 506 void free_initmem(void)
 507 {
 508         unsigned long addr;
 509         unsigned long init_begin = (unsigned long)__init_begin;
 510         unsigned long init_end = (unsigned long)__init_end;
 511
 512         /* The init text pages are marked R-X.  We have to
 513          * flush the icache and mark them RW-
 514          *
 515          * This is tricky, because map_pages is in the init section.
 516          * Do a dummy remap of the data section first (the data
 517          * section is already PAGE_KERNEL) to pull in the TLB entries
 518          * for map_kernel */
 519         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 520                   PAGE_KERNEL_RWX, 1);
 521         /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
 522          * map_pages */
 523         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 524                   PAGE_KERNEL, 1);
 525
 526         /* force the kernel to see the new TLB entries */
 527         __flush_tlb_range(0, init_begin, init_end);
 528         /* Attempt to catch anyone trying to execute code here
 529          * by filling the page with BRK insns.
 530          */
 531         memset((void *)init_begin, 0x00, init_end - init_begin);
 532         /* finally dump all the instructions which were cached, since the
 533          * pages are no-longer executable */
 534         flush_icache_range(init_begin, init_end);
 535
 536         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
 537                 ClearPageReserved(virt_to_page(addr));
 538                 init_page_count(virt_to_page(addr));
 539                 free_page(addr);
 540                 num_physpages++;
 541                 totalram_pages++;
 542         }
 543
 544         /* set up a new led state on systems shipped LED State panel */
 545         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 546
 547         printk(KERN_INFO "Freeing unused kernel memory: %luk freed\n",
 548                 (init_end - init_begin) >> 10);
 549 }
 550
 551
 552 #ifdef CONFIG_DEBUG_RODATA
 553 void mark_rodata_ro(void)
 554 {
 555         /* rodata memory was already mapped with KERNEL_RO access rights by
 556            pagetable_init() and map_pages(). No need to do additional stuff here */
 557         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 558                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
 559 }
 560 #endif
 561
 562
 563 /*
 564  * Just an arbitrary offset to serve as a "hole" between mapping areas
 565  * (between top of physical memory and a potential pcxl dma mapping
 566  * area, and below the vmalloc mapping area).
 567  *
 568  * The current 32K value just means that there will be a 32K "hole"
 569  * between mapping areas. That means that  any out-of-bounds memory
 570  * accesses will hopefully be caught. The vmalloc() routines leaves
 571  * a hole of 4kB between each vmalloced area for the same reason.
 572  */
 573
 574  /* Leave room for gateway page expansion */
 575 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 576 #error KERNEL_MAP_START is in gateway reserved region
 577 #endif
 578 #define MAP_START (KERNEL_MAP_START)
 579
 580 #define VM_MAP_OFFSET  (32*1024)
 581 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 582                                      & ~(VM_MAP_OFFSET-1)))
 583
 584 void *parisc_vmalloc_start __read_mostly;
 585 EXPORT_SYMBOL(parisc_vmalloc_start);
 586
 587 #ifdef CONFIG_PA11
 588 unsigned long pcxl_dma_start __read_mostly;
 589 #endif
 590
 591 void __init mem_init(void)
 592 {
 593         int codesize, reservedpages, datasize, initsize;
 594
 595         /* Do sanity checks on page table constants */
 596         BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
 597         BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
 598         BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 599         BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 600                         > BITS_PER_LONG);
 601
 602         high_memory = __va((max_pfn << PAGE_SHIFT));
 603
 604 #ifndef CONFIG_DISCONTIGMEM
 605         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
 606         totalram_pages += free_all_bootmem();
 607 #else
 608         {
 609                 int i;
 610
 611                 for (i = 0; i < npmem_ranges; i++)
 612                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
 613         }
 614 #endif
 615
 616         codesize = (unsigned long)_etext - (unsigned long)_text;
 617         datasize = (unsigned long)_edata - (unsigned long)_etext;
 618         initsize = (unsigned long)__init_end - (unsigned long)__init_begin;
 619
 620         reservedpages = 0;
 621 {
 622         unsigned long pfn;
 623 #ifdef CONFIG_DISCONTIGMEM
 624         int i;
 625
 626         for (i = 0; i < npmem_ranges; i++) {
 627                 for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
 628                         if (PageReserved(pfn_to_page(pfn)))
 629                                 reservedpages++;
 630                 }
 631         }
 632 #else /* !CONFIG_DISCONTIGMEM */
 633         for (pfn = 0; pfn < max_pfn; pfn++) {
 634                 /*
 635                  * Only count reserved RAM pages
 636                  */
 637                 if (PageReserved(pfn_to_page(pfn)))
 638                         reservedpages++;
 639         }
 640 #endif
 641 }
 642
 643 #ifdef CONFIG_PA11
 644         if (hppa_dma_ops == &pcxl_dma_ops) {
 645                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 646                 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
 647                                                 + PCXL_DMA_MAP_SIZE);
 648         } else {
 649                 pcxl_dma_start = 0;
 650                 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 651         }
 652 #else
 653         parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 654 #endif
 655
 656         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
 657                 nr_free_pages() << (PAGE_SHIFT-10),
 658                 num_physpages << (PAGE_SHIFT-10),
 659                 codesize >> 10,
 660                 reservedpages << (PAGE_SHIFT-10),
 661                 datasize >> 10,
 662                 initsize >> 10
 663         );
 664
 665 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
 666         printk("virtual kernel memory layout:\n"
 667                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
 668                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
 669                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
 670                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
 671                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
 672
 673                (void*)VMALLOC_START, (void*)VMALLOC_END,
 674                (VMALLOC_END - VMALLOC_START) >> 20,
 675
 676                __va(0), high_memory,
 677                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 678
 679                __init_begin, __init_end,
 680                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
 681
 682                _etext, _edata,
 683                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
 684
 685                _text, _etext,
 686                ((unsigned long)_etext - (unsigned long)_text) >> 10);
 687 #endif
 688 }
 689
 690 unsigned long *empty_zero_page __read_mostly;
 691 EXPORT_SYMBOL(empty_zero_page);
 692
 693 void show_mem(unsigned int filter)
 694 {
 695         int i,free = 0,total = 0,reserved = 0;
 696         int shared = 0, cached = 0;
 697
 698         printk(KERN_INFO "Mem-info:\n");
 699         show_free_areas(filter);
 700         if (filter & SHOW_MEM_FILTER_PAGE_COUNT)
 701                 return;
 702 #ifndef CONFIG_DISCONTIGMEM
 703         i = max_mapnr;
 704         while (i-- > 0) {
 705                 total++;
 706                 if (PageReserved(mem_map+i))
 707                         reserved++;
 708                 else if (PageSwapCache(mem_map+i))
 709                         cached++;
 710                 else if (!page_count(&mem_map[i]))
 711                         free++;
 712                 else
 713                         shared += page_count(&mem_map[i]) - 1;
 714         }
 715 #else
 716         for (i = 0; i < npmem_ranges; i++) {
 717                 int j;
 718
 719                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
 720                         struct page *p;
 721                         unsigned long flags;
 722
 723                         pgdat_resize_lock(NODE_DATA(i), &flags);
 724                         p = nid_page_nr(i, j) - node_start_pfn(i);
 725
 726                         total++;
 727                         if (PageReserved(p))
 728                                 reserved++;
 729                         else if (PageSwapCache(p))
 730                                 cached++;
 731                         else if (!page_count(p))
 732                                 free++;
 733                         else
 734                                 shared += page_count(p) - 1;
 735                         pgdat_resize_unlock(NODE_DATA(i), &flags);
 736                 }
 737         }
 738 #endif
 739         printk(KERN_INFO "%d pages of RAM\n", total);
 740         printk(KERN_INFO "%d reserved pages\n", reserved);
 741         printk(KERN_INFO "%d pages shared\n", shared);
 742         printk(KERN_INFO "%d pages swap cached\n", cached);
 743
 744
 745 #ifdef CONFIG_DISCONTIGMEM
 746         {
 747                 struct zonelist *zl;
 748                 int i, j;
 749
 750                 for (i = 0; i < npmem_ranges; i++) {
 751                         zl = node_zonelist(i, 0);
 752                         for (j = 0; j < MAX_NR_ZONES; j++) {
 753                                 struct zoneref *z;
 754                                 struct zone *zone;
 755
 756                                 printk("Zone list for zone %d on node %d: ", j, i);
 757                                 for_each_zone_zonelist(zone, z, zl, j)
 758                                         printk("[%d/%s] ", zone_to_nid(zone),
 759                                                                 zone->name);
 760                                 printk("\n");
 761                         }
 762                 }
 763         }
 764 #endif
 765 }
 766
 767 /*
 768  * pagetable_init() sets up the page tables
 769  *
 770  * Note that gateway_init() places the Linux gateway page at page 0.
 771  * Since gateway pages cannot be dereferenced this has the desirable
 772  * side effect of trapping those pesky NULL-reference errors in the
 773  * kernel.
 774  */
 775 static void __init pagetable_init(void)
 776 {
 777         int range;
 778
 779         /* Map each physical memory range to its kernel vaddr */
 780
 781         for (range = 0; range < npmem_ranges; range++) {
 782                 unsigned long start_paddr;
 783                 unsigned long end_paddr;
 784                 unsigned long size;
 785
 786                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 787                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
 788                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 789
 790                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 791                           size, PAGE_KERNEL, 0);
 792         }
 793
 794 #ifdef CONFIG_BLK_DEV_INITRD
 795         if (initrd_end && initrd_end > mem_limit) {
 796                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 797                 map_pages(initrd_start, __pa(initrd_start),
 798                           initrd_end - initrd_start, PAGE_KERNEL, 0);
 799         }
 800 #endif
 801
 802         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
 803         memset(empty_zero_page, 0, PAGE_SIZE);
 804 }
 805
 806 static void __init gateway_init(void)
 807 {
 808         unsigned long linux_gateway_page_addr;
 809         /* FIXME: This is 'const' in order to trick the compiler
 810            into not treating it as DP-relative data. */
 811         extern void * const linux_gateway_page;
 812
 813         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 814
 815         /*
 816          * Setup Linux Gateway page.
 817          *
 818          * The Linux gateway page will reside in kernel space (on virtual
 819          * page 0), so it doesn't need to be aliased into user space.
 820          */
 821
 822         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 823                   PAGE_SIZE, PAGE_GATEWAY, 1);
 824 }
 825
 826 #ifdef CONFIG_HPUX
 827 void
 828 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
 829 {
 830         pgd_t *pg_dir;
 831         pmd_t *pmd;
 832         pte_t *pg_table;
 833         unsigned long start_pmd;
 834         unsigned long start_pte;
 835         unsigned long address;
 836         unsigned long hpux_gw_page_addr;
 837         /* FIXME: This is 'const' in order to trick the compiler
 838            into not treating it as DP-relative data. */
 839         extern void * const hpux_gateway_page;
 840
 841         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
 842
 843         /*
 844          * Setup HP-UX Gateway page.
 845          *
 846          * The HP-UX gateway page resides in the user address space,
 847          * so it needs to be aliased into each process.
 848          */
 849
 850         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
 851
 852 #if PTRS_PER_PMD == 1
 853         start_pmd = 0;
 854 #else
 855         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 856 #endif
 857         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 858
 859         address = __pa(&hpux_gateway_page);
 860 #if PTRS_PER_PMD == 1
 861         pmd = (pmd_t *)__pa(pg_dir);
 862 #else
 863         pmd = (pmd_t *) pgd_address(*pg_dir);
 864
 865         /*
 866          * pmd is physical at this point
 867          */
 868
 869         if (!pmd) {
 870                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
 871                 pmd = (pmd_t *) __pa(pmd);
 872         }
 873
 874         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
 875 #endif
 876         /* now change pmd to kernel virtual addresses */
 877
 878         pmd = (pmd_t *)__va(pmd) + start_pmd;
 879
 880         /*
 881          * pg_table is physical at this point
 882          */
 883
 884         pg_table = (pte_t *) pmd_address(*pmd);
 885         if (!pg_table)
 886                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
 887
 888         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
 889
 890         /* now change pg_table to kernel virtual addresses */
 891
 892         pg_table = (pte_t *) __va(pg_table) + start_pte;
 893         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
 894 }
 895 EXPORT_SYMBOL(map_hpux_gateway_page);
 896 #endif
 897
 898 void __init paging_init(void)
 899 {
 900         int i;
 901
 902         setup_bootmem();
 903         pagetable_init();
 904         gateway_init();
 905         flush_cache_all_local(); /* start with known state */
 906         flush_tlb_all_local(NULL);
 907
 908         for (i = 0; i < npmem_ranges; i++) {
 909                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
 910
 911                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
 912
 913 #ifdef CONFIG_DISCONTIGMEM
 914                 /* Need to initialize the pfnnid_map before we can initialize
 915                    the zone */
 916                 {
 917                     int j;
 918                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 919                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 920                          j++) {
 921                         pfnnid_map[j] = i;
 922                     }
 923                 }
 924 #endif
 925
 926                 free_area_init_node(i, zones_size,
 927                                 pmem_ranges[i].start_pfn, NULL);
 928         }
 929 }
 930
 931 #ifdef CONFIG_PA20
 932
 933 /*
 934  * Currently, all PA20 chips have 18 bit protection IDs, which is the
 935  * limiting factor (space ids are 32 bits).
 936  */
 937
 938 #define NR_SPACE_IDS 262144
 939
 940 #else
 941
 942 /*
 943  * Currently we have a one-to-one relationship between space IDs and
 944  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 945  * support 15 bit protection IDs, so that is the limiting factor.
 946  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
 947  * probably not worth the effort for a special case here.
 948  */
 949
 950 #define NR_SPACE_IDS 32768
 951
 952 #endif  /* !CONFIG_PA20 */
 953
 954 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 955 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 956
 957 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 958 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 959 static unsigned long space_id_index;
 960 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 961 static unsigned long dirty_space_ids = 0;
 962
 963 static DEFINE_SPINLOCK(sid_lock);
 964
 965 unsigned long alloc_sid(void)
 966 {
 967         unsigned long index;
 968
 969         spin_lock(&sid_lock);
 970
 971         if (free_space_ids == 0) {
 972                 if (dirty_space_ids != 0) {
 973                         spin_unlock(&sid_lock);
 974                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 975                         spin_lock(&sid_lock);
 976                 }
 977                 BUG_ON(free_space_ids == 0);
 978         }
 979
 980         free_space_ids--;
 981
 982         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 983         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 984         space_id_index = index;
 985
 986         spin_unlock(&sid_lock);
 987
 988         return index << SPACEID_SHIFT;
 989 }
 990
 991 void free_sid(unsigned long spaceid)
 992 {
 993         unsigned long index = spaceid >> SPACEID_SHIFT;
 994         unsigned long *dirty_space_offset;
 995
 996         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 997         index &= (BITS_PER_LONG - 1);
 998
 999         spin_lock(&sid_lock);
1000
1001         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
1002
1003         *dirty_space_offset |= (1L << index);
1004         dirty_space_ids++;
1005
1006         spin_unlock(&sid_lock);
1007 }
1008
1009
1010 #ifdef CONFIG_SMP
1011 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
1012 {
1013         int i;
1014
1015         /* NOTE: sid_lock must be held upon entry */
1016
1017         *ndirtyptr = dirty_space_ids;
1018         if (dirty_space_ids != 0) {
1019             for (i = 0; i < SID_ARRAY_SIZE; i++) {
1020                 dirty_array[i] = dirty_space_id[i];
1021                 dirty_space_id[i] = 0;
1022             }
1023             dirty_space_ids = 0;
1024         }
1025
1026         return;
1027 }
1028
1029 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
1030 {
1031         int i;
1032
1033         /* NOTE: sid_lock must be held upon entry */
1034
1035         if (ndirty != 0) {
1036                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1037                         space_id[i] ^= dirty_array[i];
1038                 }
1039
1040                 free_space_ids += ndirty;
1041                 space_id_index = 0;
1042         }
1043 }
1044
1045 #else /* CONFIG_SMP */
1046
1047 static void recycle_sids(void)
1048 {
1049         int i;
1050
1051         /* NOTE: sid_lock must be held upon entry */
1052
1053         if (dirty_space_ids != 0) {
1054                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1055                         space_id[i] ^= dirty_space_id[i];
1056                         dirty_space_id[i] = 0;
1057                 }
1058
1059                 free_space_ids += dirty_space_ids;
1060                 dirty_space_ids = 0;
1061                 space_id_index = 0;
1062         }
1063 }
1064 #endif
1065
1066 /*
1067  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1068  * purged, we can safely reuse the space ids that were released but
1069  * not flushed from the tlb.
1070  */
1071
1072 #ifdef CONFIG_SMP
1073
1074 static unsigned long recycle_ndirty;
1075 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1076 static unsigned int recycle_inuse;
1077
1078 void flush_tlb_all(void)
1079 {
1080         int do_recycle;
1081
1082         do_recycle = 0;
1083         spin_lock(&sid_lock);
1084         if (dirty_space_ids > RECYCLE_THRESHOLD) {
1085             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1086             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1087             recycle_inuse++;
1088             do_recycle++;
1089         }
1090         spin_unlock(&sid_lock);
1091         on_each_cpu(flush_tlb_all_local, NULL, 1);
1092         if (do_recycle) {
1093             spin_lock(&sid_lock);
1094             recycle_sids(recycle_ndirty,recycle_dirty_array);
1095             recycle_inuse = 0;
1096             spin_unlock(&sid_lock);
1097         }
1098 }
1099 #else
1100 void flush_tlb_all(void)
1101 {
1102         spin_lock(&sid_lock);
1103         flush_tlb_all_local(NULL);
1104         recycle_sids();
1105         spin_unlock(&sid_lock);
1106 }
1107 #endif
1108
1109 #ifdef CONFIG_BLK_DEV_INITRD
1110 void free_initrd_mem(unsigned long start, unsigned long end)
1111 {
1112         if (start >= end)
1113                 return;
1114         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1115         for (; start < end; start += PAGE_SIZE) {
1116                 ClearPageReserved(virt_to_page(start));
1117                 init_page_count(virt_to_page(start));
1118                 free_page(start);
1119                 num_physpages++;
1120                 totalram_pages++;
1121         }
1122 }
1123 #endif