git.oblomov.eu Git - linux-2.6/blob - 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  *
  10  */
  11
  12 #include <linux/config.h>
  13
  14 #include <linux/module.h>
  15 #include <linux/mm.h>
  16 #include <linux/bootmem.h>
  17 #include <linux/delay.h>
  18 #include <linux/init.h>
  19 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
  20 #include <linux/initrd.h>
  21 #include <linux/swap.h>
  22 #include <linux/unistd.h>
  23 #include <linux/nodemask.h>     /* for node_online_map */
  24 #include <linux/pagemap.h>      /* for release_pages and page_cache_release */
  25
  26 #include <asm/pgalloc.h>
  27 #include <asm/tlb.h>
  28 #include <asm/pdc_chassis.h>
  29 #include <asm/mmzone.h>
  30
  31 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  32
  33 extern char _text;      /* start of kernel code, defined by linker */
  34 extern int  data_start;
  35 extern char _end;       /* end of BSS, defined by linker */
  36 extern char __init_begin, __init_end;
  37
  38 #ifdef CONFIG_DISCONTIGMEM
  39 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
  40 bootmem_data_t bmem_data[MAX_NUMNODES] __read_mostly;
  41 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
  42 #endif
  43
  44 static struct resource data_resource = {
  45         .name   = "Kernel data",
  46         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  47 };
  48
  49 static struct resource code_resource = {
  50         .name   = "Kernel code",
  51         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  52 };
  53
  54 static struct resource pdcdata_resource = {
  55         .name   = "PDC data (Page Zero)",
  56         .start  = 0,
  57         .end    = 0x9ff,
  58         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  59 };
  60
  61 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
  62
  63 /* The following array is initialized from the firmware specific
  64  * information retrieved in kernel/inventory.c.
  65  */
  66
  67 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
  68 int npmem_ranges __read_mostly;
  69
  70 #ifdef __LP64__
  71 #define MAX_MEM         (~0UL)
  72 #else /* !__LP64__ */
  73 #define MAX_MEM         (3584U*1024U*1024U)
  74 #endif /* !__LP64__ */
  75
  76 static unsigned long mem_limit __read_mostly = MAX_MEM;
  77
  78 static void __init mem_limit_func(void)
  79 {
  80         char *cp, *end;
  81         unsigned long limit;
  82         extern char saved_command_line[];
  83
  84         /* We need this before __setup() functions are called */
  85
  86         limit = MAX_MEM;
  87         for (cp = saved_command_line; *cp; ) {
  88                 if (memcmp(cp, "mem=", 4) == 0) {
  89                         cp += 4;
  90                         limit = memparse(cp, &end);
  91                         if (end != cp)
  92                                 break;
  93                         cp = end;
  94                 } else {
  95                         while (*cp != ' ' && *cp)
  96                                 ++cp;
  97                         while (*cp == ' ')
  98                                 ++cp;
  99                 }
 100         }
 101
 102         if (limit < mem_limit)
 103                 mem_limit = limit;
 104 }
 105
 106 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 107
 108 static void __init setup_bootmem(void)
 109 {
 110         unsigned long bootmap_size;
 111         unsigned long mem_max;
 112         unsigned long bootmap_pages;
 113         unsigned long bootmap_start_pfn;
 114         unsigned long bootmap_pfn;
 115 #ifndef CONFIG_DISCONTIGMEM
 116         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 117         int npmem_holes;
 118 #endif
 119         int i, sysram_resource_count;
 120
 121         disable_sr_hashing(); /* Turn off space register hashing */
 122
 123         /*
 124          * Sort the ranges. Since the number of ranges is typically
 125          * small, and performance is not an issue here, just do
 126          * a simple insertion sort.
 127          */
 128
 129         for (i = 1; i < npmem_ranges; i++) {
 130                 int j;
 131
 132                 for (j = i; j > 0; j--) {
 133                         unsigned long tmp;
 134
 135                         if (pmem_ranges[j-1].start_pfn <
 136                             pmem_ranges[j].start_pfn) {
 137
 138                                 break;
 139                         }
 140                         tmp = pmem_ranges[j-1].start_pfn;
 141                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 142                         pmem_ranges[j].start_pfn = tmp;
 143                         tmp = pmem_ranges[j-1].pages;
 144                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 145                         pmem_ranges[j].pages = tmp;
 146                 }
 147         }
 148
 149 #ifndef CONFIG_DISCONTIGMEM
 150         /*
 151          * Throw out ranges that are too far apart (controlled by
 152          * MAX_GAP).
 153          */
 154
 155         for (i = 1; i < npmem_ranges; i++) {
 156                 if (pmem_ranges[i].start_pfn -
 157                         (pmem_ranges[i-1].start_pfn +
 158                          pmem_ranges[i-1].pages) > MAX_GAP) {
 159                         npmem_ranges = i;
 160                         printk("Large gap in memory detected (%ld pages). "
 161                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 162                                pmem_ranges[i].start_pfn -
 163                                (pmem_ranges[i-1].start_pfn +
 164                                 pmem_ranges[i-1].pages));
 165                         break;
 166                 }
 167         }
 168 #endif
 169
 170         if (npmem_ranges > 1) {
 171
 172                 /* Print the memory ranges */
 173
 174                 printk(KERN_INFO "Memory Ranges:\n");
 175
 176                 for (i = 0; i < npmem_ranges; i++) {
 177                         unsigned long start;
 178                         unsigned long size;
 179
 180                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
 181                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 182                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 183                                 i,start, start + (size - 1), size >> 20);
 184                 }
 185         }
 186
 187         sysram_resource_count = npmem_ranges;
 188         for (i = 0; i < sysram_resource_count; i++) {
 189                 struct resource *res = &sysram_resources[i];
 190                 res->name = "System RAM";
 191                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
 192                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
 193                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 194                 request_resource(&iomem_resource, res);
 195         }
 196
 197         /*
 198          * For 32 bit kernels we limit the amount of memory we can
 199          * support, in order to preserve enough kernel address space
 200          * for other purposes. For 64 bit kernels we don't normally
 201          * limit the memory, but this mechanism can be used to
 202          * artificially limit the amount of memory (and it is written
 203          * to work with multiple memory ranges).
 204          */
 205
 206         mem_limit_func();       /* check for "mem=" argument */
 207
 208         mem_max = 0;
 209         num_physpages = 0;
 210         for (i = 0; i < npmem_ranges; i++) {
 211                 unsigned long rsize;
 212
 213                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 214                 if ((mem_max + rsize) > mem_limit) {
 215                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 216                         if (mem_max == mem_limit)
 217                                 npmem_ranges = i;
 218                         else {
 219                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 220                                                        - (mem_max >> PAGE_SHIFT);
 221                                 npmem_ranges = i + 1;
 222                                 mem_max = mem_limit;
 223                         }
 224                 num_physpages += pmem_ranges[i].pages;
 225                         break;
 226                 }
 227             num_physpages += pmem_ranges[i].pages;
 228                 mem_max += rsize;
 229         }
 230
 231         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 232
 233 #ifndef CONFIG_DISCONTIGMEM
 234         /* Merge the ranges, keeping track of the holes */
 235
 236         {
 237                 unsigned long end_pfn;
 238                 unsigned long hole_pages;
 239
 240                 npmem_holes = 0;
 241                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 242                 for (i = 1; i < npmem_ranges; i++) {
 243
 244                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 245                         if (hole_pages) {
 246                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 247                                 pmem_holes[npmem_holes++].pages = hole_pages;
 248                                 end_pfn += hole_pages;
 249                         }
 250                         end_pfn += pmem_ranges[i].pages;
 251                 }
 252
 253                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 254                 npmem_ranges = 1;
 255         }
 256 #endif
 257
 258         bootmap_pages = 0;
 259         for (i = 0; i < npmem_ranges; i++)
 260                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
 261
 262         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
 263
 264 #ifdef CONFIG_DISCONTIGMEM
 265         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 266                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 267                 NODE_DATA(i)->bdata = &bmem_data[i];
 268         }
 269         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 270
 271         for (i = 0; i < npmem_ranges; i++)
 272                 node_set_online(i);
 273 #endif
 274
 275         /*
 276          * Initialize and free the full range of memory in each range.
 277          * Note that the only writing these routines do are to the bootmap,
 278          * and we've made sure to locate the bootmap properly so that they
 279          * won't be writing over anything important.
 280          */
 281
 282         bootmap_pfn = bootmap_start_pfn;
 283         max_pfn = 0;
 284         for (i = 0; i < npmem_ranges; i++) {
 285                 unsigned long start_pfn;
 286                 unsigned long npages;
 287
 288                 start_pfn = pmem_ranges[i].start_pfn;
 289                 npages = pmem_ranges[i].pages;
 290
 291                 bootmap_size = init_bootmem_node(NODE_DATA(i),
 292                                                 bootmap_pfn,
 293                                                 start_pfn,
 294                                                 (start_pfn + npages) );
 295                 free_bootmem_node(NODE_DATA(i),
 296                                   (start_pfn << PAGE_SHIFT),
 297                                   (npages << PAGE_SHIFT) );
 298                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 299                 if ((start_pfn + npages) > max_pfn)
 300                         max_pfn = start_pfn + npages;
 301         }
 302
 303         /* IOMMU is always used to access "high mem" on those boxes
 304          * that can support enough mem that a PCI device couldn't
 305          * directly DMA to any physical addresses.
 306          * ISA DMA support will need to revisit this.
 307          */
 308         max_low_pfn = max_pfn;
 309
 310         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
 311                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
 312                 BUG();
 313         }
 314
 315         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 316
 317 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 318
 319         reserve_bootmem_node(NODE_DATA(0), 0UL,
 320                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
 321         reserve_bootmem_node(NODE_DATA(0),__pa((unsigned long)&_text),
 322                         (unsigned long)(&_end - &_text));
 323         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
 324                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
 325
 326 #ifndef CONFIG_DISCONTIGMEM
 327
 328         /* reserve the holes */
 329
 330         for (i = 0; i < npmem_holes; i++) {
 331                 reserve_bootmem_node(NODE_DATA(0),
 332                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
 333                                 (pmem_holes[i].pages << PAGE_SHIFT));
 334         }
 335 #endif
 336
 337 #ifdef CONFIG_BLK_DEV_INITRD
 338         if (initrd_start) {
 339                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 340                 if (__pa(initrd_start) < mem_max) {
 341                         unsigned long initrd_reserve;
 342
 343                         if (__pa(initrd_end) > mem_max) {
 344                                 initrd_reserve = mem_max - __pa(initrd_start);
 345                         } else {
 346                                 initrd_reserve = initrd_end - initrd_start;
 347                         }
 348                         initrd_below_start_ok = 1;
 349                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 350
 351                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
 352                 }
 353         }
 354 #endif
 355
 356         data_resource.start =  virt_to_phys(&data_start);
 357         data_resource.end = virt_to_phys(&_end)-1;
 358         code_resource.start = virt_to_phys(&_text);
 359         code_resource.end = virt_to_phys(&data_start)-1;
 360
 361         /* We don't know which region the kernel will be in, so try
 362          * all of them.
 363          */
 364         for (i = 0; i < sysram_resource_count; i++) {
 365                 struct resource *res = &sysram_resources[i];
 366                 request_resource(res, &code_resource);
 367                 request_resource(res, &data_resource);
 368         }
 369         request_resource(&sysram_resources[0], &pdcdata_resource);
 370 }
 371
 372 void free_initmem(void)
 373 {
 374         unsigned long addr;
 375
 376         printk(KERN_INFO "Freeing unused kernel memory: ");
 377
 378 #ifdef CONFIG_DEBUG_KERNEL
 379         /* Attempt to catch anyone trying to execute code here
 380          * by filling the page with BRK insns.
 381          *
 382          * If we disable interrupts for all CPUs, then IPI stops working.
 383          * Kinda breaks the global cache flushing.
 384          */
 385         local_irq_disable();
 386
 387         memset(&__init_begin, 0x00,
 388                 (unsigned long)&__init_end - (unsigned long)&__init_begin);
 389
 390         flush_data_cache();
 391         asm volatile("sync" : : );
 392         flush_icache_range((unsigned long)&__init_begin, (unsigned long)&__init_end);
 393         asm volatile("sync" : : );
 394
 395         local_irq_enable();
 396 #endif
 397
 398         addr = (unsigned long)(&__init_begin);
 399         for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) {
 400                 ClearPageReserved(virt_to_page(addr));
 401                 init_page_count(virt_to_page(addr));
 402                 free_page(addr);
 403                 num_physpages++;
 404                 totalram_pages++;
 405         }
 406
 407         /* set up a new led state on systems shipped LED State panel */
 408         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 409
 410         printk("%luk freed\n", (unsigned long)(&__init_end - &__init_begin) >> 10);
 411 }
 412
 413
 414 #ifdef CONFIG_DEBUG_RODATA
 415 void mark_rodata_ro(void)
 416 {
 417         extern char __start_rodata, __end_rodata;
 418         /* rodata memory was already mapped with KERNEL_RO access rights by
 419            pagetable_init() and map_pages(). No need to do additional stuff here */
 420         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 421                 (unsigned long)(&__end_rodata - &__start_rodata) >> 10);
 422 }
 423 #endif
 424
 425
 426 /*
 427  * Just an arbitrary offset to serve as a "hole" between mapping areas
 428  * (between top of physical memory and a potential pcxl dma mapping
 429  * area, and below the vmalloc mapping area).
 430  *
 431  * The current 32K value just means that there will be a 32K "hole"
 432  * between mapping areas. That means that  any out-of-bounds memory
 433  * accesses will hopefully be caught. The vmalloc() routines leaves
 434  * a hole of 4kB between each vmalloced area for the same reason.
 435  */
 436
 437  /* Leave room for gateway page expansion */
 438 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 439 #error KERNEL_MAP_START is in gateway reserved region
 440 #endif
 441 #define MAP_START (KERNEL_MAP_START)
 442
 443 #define VM_MAP_OFFSET  (32*1024)
 444 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 445                                      & ~(VM_MAP_OFFSET-1)))
 446
 447 void *vmalloc_start __read_mostly;
 448 EXPORT_SYMBOL(vmalloc_start);
 449
 450 #ifdef CONFIG_PA11
 451 unsigned long pcxl_dma_start __read_mostly;
 452 #endif
 453
 454 void __init mem_init(void)
 455 {
 456         high_memory = __va((max_pfn << PAGE_SHIFT));
 457
 458 #ifndef CONFIG_DISCONTIGMEM
 459         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
 460         totalram_pages += free_all_bootmem();
 461 #else
 462         {
 463                 int i;
 464
 465                 for (i = 0; i < npmem_ranges; i++)
 466                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
 467         }
 468 #endif
 469
 470         printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
 471
 472 #ifdef CONFIG_PA11
 473         if (hppa_dma_ops == &pcxl_dma_ops) {
 474                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 475                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
 476         } else {
 477                 pcxl_dma_start = 0;
 478                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
 479         }
 480 #else
 481         vmalloc_start = SET_MAP_OFFSET(MAP_START);
 482 #endif
 483
 484 }
 485
 486 unsigned long *empty_zero_page __read_mostly;
 487
 488 void show_mem(void)
 489 {
 490         int i,free = 0,total = 0,reserved = 0;
 491         int shared = 0, cached = 0;
 492
 493         printk(KERN_INFO "Mem-info:\n");
 494         show_free_areas();
 495         printk(KERN_INFO "Free swap:     %6ldkB\n",
 496                                 nr_swap_pages<<(PAGE_SHIFT-10));
 497 #ifndef CONFIG_DISCONTIGMEM
 498         i = max_mapnr;
 499         while (i-- > 0) {
 500                 total++;
 501                 if (PageReserved(mem_map+i))
 502                         reserved++;
 503                 else if (PageSwapCache(mem_map+i))
 504                         cached++;
 505                 else if (!page_count(&mem_map[i]))
 506                         free++;
 507                 else
 508                         shared += page_count(&mem_map[i]) - 1;
 509         }
 510 #else
 511         for (i = 0; i < npmem_ranges; i++) {
 512                 int j;
 513
 514                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
 515                         struct page *p;
 516                         unsigned long flags;
 517
 518                         pgdat_resize_lock(NODE_DATA(i), &flags);
 519                         p = nid_page_nr(i, j) - node_start_pfn(i);
 520
 521                         total++;
 522                         if (PageReserved(p))
 523                                 reserved++;
 524                         else if (PageSwapCache(p))
 525                                 cached++;
 526                         else if (!page_count(p))
 527                                 free++;
 528                         else
 529                                 shared += page_count(p) - 1;
 530                         pgdat_resize_unlock(NODE_DATA(i), &flags);
 531                 }
 532         }
 533 #endif
 534         printk(KERN_INFO "%d pages of RAM\n", total);
 535         printk(KERN_INFO "%d reserved pages\n", reserved);
 536         printk(KERN_INFO "%d pages shared\n", shared);
 537         printk(KERN_INFO "%d pages swap cached\n", cached);
 538
 539
 540 #ifdef CONFIG_DISCONTIGMEM
 541         {
 542                 struct zonelist *zl;
 543                 int i, j, k;
 544
 545                 for (i = 0; i < npmem_ranges; i++) {
 546                         for (j = 0; j < MAX_NR_ZONES; j++) {
 547                                 zl = NODE_DATA(i)->node_zonelists + j;
 548
 549                                 printk("Zone list for zone %d on node %d: ", j, i);
 550                                 for (k = 0; zl->zones[k] != NULL; k++)
 551                                         printk("[%d/%s] ", zl->zones[k]->zone_pgdat->node_id, zl->zones[k]->name);
 552                                 printk("\n");
 553                         }
 554                 }
 555         }
 556 #endif
 557 }
 558
 559
 560 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
 561 {
 562         pgd_t *pg_dir;
 563         pmd_t *pmd;
 564         pte_t *pg_table;
 565         unsigned long end_paddr;
 566         unsigned long start_pmd;
 567         unsigned long start_pte;
 568         unsigned long tmp1;
 569         unsigned long tmp2;
 570         unsigned long address;
 571         unsigned long ro_start;
 572         unsigned long ro_end;
 573         unsigned long fv_addr;
 574         unsigned long gw_addr;
 575         extern const unsigned long fault_vector_20;
 576         extern void * const linux_gateway_page;
 577
 578         ro_start = __pa((unsigned long)&_text);
 579         ro_end   = __pa((unsigned long)&data_start);
 580         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
 581         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
 582
 583         end_paddr = start_paddr + size;
 584
 585         pg_dir = pgd_offset_k(start_vaddr);
 586
 587 #if PTRS_PER_PMD == 1
 588         start_pmd = 0;
 589 #else
 590         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 591 #endif
 592         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 593
 594         address = start_paddr;
 595         while (address < end_paddr) {
 596 #if PTRS_PER_PMD == 1
 597                 pmd = (pmd_t *)__pa(pg_dir);
 598 #else
 599                 pmd = (pmd_t *)pgd_address(*pg_dir);
 600
 601                 /*
 602                  * pmd is physical at this point
 603                  */
 604
 605                 if (!pmd) {
 606                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
 607                         pmd = (pmd_t *) __pa(pmd);
 608                 }
 609
 610                 pgd_populate(NULL, pg_dir, __va(pmd));
 611 #endif
 612                 pg_dir++;
 613
 614                 /* now change pmd to kernel virtual addresses */
 615
 616                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 617                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
 618
 619                         /*
 620                          * pg_table is physical at this point
 621                          */
 622
 623                         pg_table = (pte_t *)pmd_address(*pmd);
 624                         if (!pg_table) {
 625                                 pg_table = (pte_t *)
 626                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
 627                                 pg_table = (pte_t *) __pa(pg_table);
 628                         }
 629
 630                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 631
 632                         /* now change pg_table to kernel virtual addresses */
 633
 634                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 635                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
 636                                 pte_t pte;
 637
 638                                 /*
 639                                  * Map the fault vector writable so we can
 640                                  * write the HPMC checksum.
 641                                  */
 642                                 if (address >= ro_start && address < ro_end
 643                                                         && address != fv_addr
 644                                                         && address != gw_addr)
 645                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
 646                                 else
 647                                     pte = __mk_pte(address, pgprot);
 648
 649                                 if (address >= end_paddr)
 650                                         pte_val(pte) = 0;
 651
 652                                 set_pte(pg_table, pte);
 653
 654                                 address += PAGE_SIZE;
 655                         }
 656                         start_pte = 0;
 657
 658                         if (address >= end_paddr)
 659                             break;
 660                 }
 661                 start_pmd = 0;
 662         }
 663 }
 664
 665 /*
 666  * pagetable_init() sets up the page tables
 667  *
 668  * Note that gateway_init() places the Linux gateway page at page 0.
 669  * Since gateway pages cannot be dereferenced this has the desirable
 670  * side effect of trapping those pesky NULL-reference errors in the
 671  * kernel.
 672  */
 673 static void __init pagetable_init(void)
 674 {
 675         int range;
 676
 677         /* Map each physical memory range to its kernel vaddr */
 678
 679         for (range = 0; range < npmem_ranges; range++) {
 680                 unsigned long start_paddr;
 681                 unsigned long end_paddr;
 682                 unsigned long size;
 683
 684                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 685                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
 686                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 687
 688                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 689                         size, PAGE_KERNEL);
 690         }
 691
 692 #ifdef CONFIG_BLK_DEV_INITRD
 693         if (initrd_end && initrd_end > mem_limit) {
 694                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 695                 map_pages(initrd_start, __pa(initrd_start),
 696                         initrd_end - initrd_start, PAGE_KERNEL);
 697         }
 698 #endif
 699
 700         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
 701         memset(empty_zero_page, 0, PAGE_SIZE);
 702 }
 703
 704 static void __init gateway_init(void)
 705 {
 706         unsigned long linux_gateway_page_addr;
 707         /* FIXME: This is 'const' in order to trick the compiler
 708            into not treating it as DP-relative data. */
 709         extern void * const linux_gateway_page;
 710
 711         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 712
 713         /*
 714          * Setup Linux Gateway page.
 715          *
 716          * The Linux gateway page will reside in kernel space (on virtual
 717          * page 0), so it doesn't need to be aliased into user space.
 718          */
 719
 720         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 721                 PAGE_SIZE, PAGE_GATEWAY);
 722 }
 723
 724 #ifdef CONFIG_HPUX
 725 void
 726 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
 727 {
 728         pgd_t *pg_dir;
 729         pmd_t *pmd;
 730         pte_t *pg_table;
 731         unsigned long start_pmd;
 732         unsigned long start_pte;
 733         unsigned long address;
 734         unsigned long hpux_gw_page_addr;
 735         /* FIXME: This is 'const' in order to trick the compiler
 736            into not treating it as DP-relative data. */
 737         extern void * const hpux_gateway_page;
 738
 739         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
 740
 741         /*
 742          * Setup HP-UX Gateway page.
 743          *
 744          * The HP-UX gateway page resides in the user address space,
 745          * so it needs to be aliased into each process.
 746          */
 747
 748         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
 749
 750 #if PTRS_PER_PMD == 1
 751         start_pmd = 0;
 752 #else
 753         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 754 #endif
 755         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 756
 757         address = __pa(&hpux_gateway_page);
 758 #if PTRS_PER_PMD == 1
 759         pmd = (pmd_t *)__pa(pg_dir);
 760 #else
 761         pmd = (pmd_t *) pgd_address(*pg_dir);
 762
 763         /*
 764          * pmd is physical at this point
 765          */
 766
 767         if (!pmd) {
 768                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
 769                 pmd = (pmd_t *) __pa(pmd);
 770         }
 771
 772         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
 773 #endif
 774         /* now change pmd to kernel virtual addresses */
 775
 776         pmd = (pmd_t *)__va(pmd) + start_pmd;
 777
 778         /*
 779          * pg_table is physical at this point
 780          */
 781
 782         pg_table = (pte_t *) pmd_address(*pmd);
 783         if (!pg_table)
 784                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
 785
 786         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
 787
 788         /* now change pg_table to kernel virtual addresses */
 789
 790         pg_table = (pte_t *) __va(pg_table) + start_pte;
 791         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
 792 }
 793 EXPORT_SYMBOL(map_hpux_gateway_page);
 794 #endif
 795
 796 void __init paging_init(void)
 797 {
 798         int i;
 799
 800         setup_bootmem();
 801         pagetable_init();
 802         gateway_init();
 803         flush_cache_all_local(); /* start with known state */
 804         flush_tlb_all_local(NULL);
 805
 806         for (i = 0; i < npmem_ranges; i++) {
 807                 unsigned long zones_size[MAX_NR_ZONES] = { 0, 0, 0 };
 808
 809                 /* We have an IOMMU, so all memory can go into a single
 810                    ZONE_DMA zone. */
 811                 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
 812
 813 #ifdef CONFIG_DISCONTIGMEM
 814                 /* Need to initialize the pfnnid_map before we can initialize
 815                    the zone */
 816                 {
 817                     int j;
 818                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 819                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 820                          j++) {
 821                         pfnnid_map[j] = i;
 822                     }
 823                 }
 824 #endif
 825
 826                 free_area_init_node(i, NODE_DATA(i), zones_size,
 827                                 pmem_ranges[i].start_pfn, NULL);
 828         }
 829 }
 830
 831 #ifdef CONFIG_PA20
 832
 833 /*
 834  * Currently, all PA20 chips have 18 bit protection id's, which is the
 835  * limiting factor (space ids are 32 bits).
 836  */
 837
 838 #define NR_SPACE_IDS 262144
 839
 840 #else
 841
 842 /*
 843  * Currently we have a one-to-one relationship between space id's and
 844  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 845  * support 15 bit protection id's, so that is the limiting factor.
 846  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
 847  * probably not worth the effort for a special case here.
 848  */
 849
 850 #define NR_SPACE_IDS 32768
 851
 852 #endif  /* !CONFIG_PA20 */
 853
 854 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 855 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 856
 857 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 858 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 859 static unsigned long space_id_index;
 860 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 861 static unsigned long dirty_space_ids = 0;
 862
 863 static DEFINE_SPINLOCK(sid_lock);
 864
 865 unsigned long alloc_sid(void)
 866 {
 867         unsigned long index;
 868
 869         spin_lock(&sid_lock);
 870
 871         if (free_space_ids == 0) {
 872                 if (dirty_space_ids != 0) {
 873                         spin_unlock(&sid_lock);
 874                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 875                         spin_lock(&sid_lock);
 876                 }
 877                 if (free_space_ids == 0)
 878                         BUG();
 879         }
 880
 881         free_space_ids--;
 882
 883         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 884         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 885         space_id_index = index;
 886
 887         spin_unlock(&sid_lock);
 888
 889         return index << SPACEID_SHIFT;
 890 }
 891
 892 void free_sid(unsigned long spaceid)
 893 {
 894         unsigned long index = spaceid >> SPACEID_SHIFT;
 895         unsigned long *dirty_space_offset;
 896
 897         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 898         index &= (BITS_PER_LONG - 1);
 899
 900         spin_lock(&sid_lock);
 901
 902         if (*dirty_space_offset & (1L << index))
 903             BUG(); /* attempt to free space id twice */
 904
 905         *dirty_space_offset |= (1L << index);
 906         dirty_space_ids++;
 907
 908         spin_unlock(&sid_lock);
 909 }
 910
 911
 912 #ifdef CONFIG_SMP
 913 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 914 {
 915         int i;
 916
 917         /* NOTE: sid_lock must be held upon entry */
 918
 919         *ndirtyptr = dirty_space_ids;
 920         if (dirty_space_ids != 0) {
 921             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 922                 dirty_array[i] = dirty_space_id[i];
 923                 dirty_space_id[i] = 0;
 924             }
 925             dirty_space_ids = 0;
 926         }
 927
 928         return;
 929 }
 930
 931 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 932 {
 933         int i;
 934
 935         /* NOTE: sid_lock must be held upon entry */
 936
 937         if (ndirty != 0) {
 938                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 939                         space_id[i] ^= dirty_array[i];
 940                 }
 941
 942                 free_space_ids += ndirty;
 943                 space_id_index = 0;
 944         }
 945 }
 946
 947 #else /* CONFIG_SMP */
 948
 949 static void recycle_sids(void)
 950 {
 951         int i;
 952
 953         /* NOTE: sid_lock must be held upon entry */
 954
 955         if (dirty_space_ids != 0) {
 956                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 957                         space_id[i] ^= dirty_space_id[i];
 958                         dirty_space_id[i] = 0;
 959                 }
 960
 961                 free_space_ids += dirty_space_ids;
 962                 dirty_space_ids = 0;
 963                 space_id_index = 0;
 964         }
 965 }
 966 #endif
 967
 968 /*
 969  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
 970  * purged, we can safely reuse the space ids that were released but
 971  * not flushed from the tlb.
 972  */
 973
 974 #ifdef CONFIG_SMP
 975
 976 static unsigned long recycle_ndirty;
 977 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
 978 static unsigned int recycle_inuse = 0;
 979
 980 void flush_tlb_all(void)
 981 {
 982         int do_recycle;
 983
 984         do_recycle = 0;
 985         spin_lock(&sid_lock);
 986         if (dirty_space_ids > RECYCLE_THRESHOLD) {
 987             if (recycle_inuse) {
 988                 BUG();  /* FIXME: Use a semaphore/wait queue here */
 989             }
 990             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
 991             recycle_inuse++;
 992             do_recycle++;
 993         }
 994         spin_unlock(&sid_lock);
 995         on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
 996         if (do_recycle) {
 997             spin_lock(&sid_lock);
 998             recycle_sids(recycle_ndirty,recycle_dirty_array);
 999             recycle_inuse = 0;
1000             spin_unlock(&sid_lock);
1001         }
1002 }
1003 #else
1004 void flush_tlb_all(void)
1005 {
1006         spin_lock(&sid_lock);
1007         flush_tlb_all_local(NULL);
1008         recycle_sids();
1009         spin_unlock(&sid_lock);
1010 }
1011 #endif
1012
1013 #ifdef CONFIG_BLK_DEV_INITRD
1014 void free_initrd_mem(unsigned long start, unsigned long end)
1015 {
1016         if (start >= end)
1017                 return;
1018         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1019         for (; start < end; start += PAGE_SIZE) {
1020                 ClearPageReserved(virt_to_page(start));
1021                 init_page_count(virt_to_page(start));
1022                 free_page(start);
1023                 num_physpages++;
1024                 totalram_pages++;
1025         }
1026 }
1027 #endif