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