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  *
   9  */
  10
  11 #include <linux/config.h>
  12
  13 #include <linux/module.h>
  14 #include <linux/mm.h>
  15 #include <linux/bootmem.h>
  16 #include <linux/delay.h>
  17 #include <linux/init.h>
  18 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
  19 #include <linux/initrd.h>
  20 #include <linux/swap.h>
  21 #include <linux/unistd.h>
  22
  23 #include <asm/pgalloc.h>
  24 #include <asm/tlb.h>
  25 #include <asm/pdc_chassis.h>
  26
  27 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  28
  29 extern char _text;      /* start of kernel code, defined by linker */
  30 extern int  data_start;
  31 extern char _end;       /* end of BSS, defined by linker */
  32 extern char __init_begin, __init_end;
  33
  34 #ifdef CONFIG_DISCONTIGMEM
  35 struct node_map_data node_data[MAX_PHYSMEM_RANGES];
  36 bootmem_data_t bmem_data[MAX_PHYSMEM_RANGES];
  37 unsigned char *chunkmap;
  38 unsigned int maxchunkmap;
  39 #endif
  40
  41 static struct resource data_resource = {
  42         .name   = "Kernel data",
  43         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  44 };
  45
  46 static struct resource code_resource = {
  47         .name   = "Kernel code",
  48         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  49 };
  50
  51 static struct resource pdcdata_resource = {
  52         .name   = "PDC data (Page Zero)",
  53         .start  = 0,
  54         .end    = 0x9ff,
  55         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  56 };
  57
  58 static struct resource sysram_resources[MAX_PHYSMEM_RANGES];
  59
  60 static unsigned long max_pfn;
  61
  62 /* The following array is initialized from the firmware specific
  63  * information retrieved in kernel/inventory.c.
  64  */
  65
  66 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES];
  67 int npmem_ranges;
  68
  69 #ifdef __LP64__
  70 #define MAX_MEM         (~0UL)
  71 #else /* !__LP64__ */
  72 #define MAX_MEM         (3584U*1024U*1024U)
  73 #endif /* !__LP64__ */
  74
  75 static unsigned long mem_limit = MAX_MEM;
  76
  77 static void __init mem_limit_func(void)
  78 {
  79         char *cp, *end;
  80         unsigned long limit;
  81         extern char saved_command_line[];
  82
  83         /* We need this before __setup() functions are called */
  84
  85         limit = MAX_MEM;
  86         for (cp = saved_command_line; *cp; ) {
  87                 if (memcmp(cp, "mem=", 4) == 0) {
  88                         cp += 4;
  89                         limit = memparse(cp, &end);
  90                         if (end != cp)
  91                                 break;
  92                         cp = end;
  93                 } else {
  94                         while (*cp != ' ' && *cp)
  95                                 ++cp;
  96                         while (*cp == ' ')
  97                                 ++cp;
  98                 }
  99         }
 100
 101         if (limit < mem_limit)
 102                 mem_limit = limit;
 103 }
 104
 105 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 106
 107 static void __init setup_bootmem(void)
 108 {
 109         unsigned long bootmap_size;
 110         unsigned long mem_max;
 111         unsigned long bootmap_pages;
 112         unsigned long bootmap_start_pfn;
 113         unsigned long bootmap_pfn;
 114 #ifndef CONFIG_DISCONTIGMEM
 115         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 116         int npmem_holes;
 117 #endif
 118         int i, sysram_resource_count;
 119
 120         disable_sr_hashing(); /* Turn off space register hashing */
 121
 122 #ifdef CONFIG_DISCONTIGMEM
 123         /*
 124          * The below is still true as of 2.4.2. If this is ever fixed,
 125          * we can remove this warning!
 126          */
 127
 128         printk(KERN_WARNING "\n\n");
 129         printk(KERN_WARNING "CONFIG_DISCONTIGMEM is enabled, which is probably a mistake. This\n");
 130         printk(KERN_WARNING "option can lead to heavy swapping, even when there are gigabytes\n");
 131         printk(KERN_WARNING "of free memory.\n\n");
 132 #endif
 133
 134 #ifdef __LP64__
 135
 136 #ifndef CONFIG_DISCONTIGMEM
 137         /*
 138          * Sort the ranges. Since the number of ranges is typically
 139          * small, and performance is not an issue here, just do
 140          * a simple insertion sort.
 141          */
 142
 143         for (i = 1; i < npmem_ranges; i++) {
 144                 int j;
 145
 146                 for (j = i; j > 0; j--) {
 147                         unsigned long tmp;
 148
 149                         if (pmem_ranges[j-1].start_pfn <
 150                             pmem_ranges[j].start_pfn) {
 151
 152                                 break;
 153                         }
 154                         tmp = pmem_ranges[j-1].start_pfn;
 155                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 156                         pmem_ranges[j].start_pfn = tmp;
 157                         tmp = pmem_ranges[j-1].pages;
 158                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 159                         pmem_ranges[j].pages = tmp;
 160                 }
 161         }
 162
 163         /*
 164          * Throw out ranges that are too far apart (controlled by
 165          * MAX_GAP). If CONFIG_DISCONTIGMEM wasn't implemented so
 166          * poorly, we would recommend enabling that option, but,
 167          * until it is fixed, this is the best way to go.
 168          */
 169
 170         for (i = 1; i < npmem_ranges; i++) {
 171                 if (pmem_ranges[i].start_pfn -
 172                         (pmem_ranges[i-1].start_pfn +
 173                          pmem_ranges[i-1].pages) > MAX_GAP) {
 174                         npmem_ranges = i;
 175                         break;
 176                 }
 177         }
 178 #endif
 179
 180         if (npmem_ranges > 1) {
 181
 182                 /* Print the memory ranges */
 183
 184                 printk(KERN_INFO "Memory Ranges:\n");
 185
 186                 for (i = 0; i < npmem_ranges; i++) {
 187                         unsigned long start;
 188                         unsigned long size;
 189
 190                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
 191                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 192                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld Mb\n",
 193                                 i,start, start + (size - 1), size >> 20);
 194                 }
 195         }
 196
 197 #endif /* __LP64__ */
 198
 199         sysram_resource_count = npmem_ranges;
 200         for (i = 0; i < sysram_resource_count; i++) {
 201                 struct resource *res = &sysram_resources[i];
 202                 res->name = "System RAM";
 203                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
 204                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
 205                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 206                 request_resource(&iomem_resource, res);
 207         }
 208
 209         /*
 210          * For 32 bit kernels we limit the amount of memory we can
 211          * support, in order to preserve enough kernel address space
 212          * for other purposes. For 64 bit kernels we don't normally
 213          * limit the memory, but this mechanism can be used to
 214          * artificially limit the amount of memory (and it is written
 215          * to work with multiple memory ranges).
 216          */
 217
 218         mem_limit_func();       /* check for "mem=" argument */
 219
 220         mem_max = 0;
 221         for (i = 0; i < npmem_ranges; i++) {
 222                 unsigned long rsize;
 223
 224                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 225                 if ((mem_max + rsize) > mem_limit) {
 226                         printk(KERN_WARNING "Memory truncated to %ld Mb\n", mem_limit >> 20);
 227                         if (mem_max == mem_limit)
 228                                 npmem_ranges = i;
 229                         else {
 230                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 231                                                        - (mem_max >> PAGE_SHIFT);
 232                                 npmem_ranges = i + 1;
 233                                 mem_max = mem_limit;
 234                         }
 235                         break;
 236                 }
 237                 mem_max += rsize;
 238         }
 239
 240         printk(KERN_INFO "Total Memory: %ld Mb\n",mem_max >> 20);
 241
 242 #ifndef CONFIG_DISCONTIGMEM
 243
 244         /* Merge the ranges, keeping track of the holes */
 245
 246         {
 247                 unsigned long end_pfn;
 248                 unsigned long hole_pages;
 249
 250                 npmem_holes = 0;
 251                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 252                 for (i = 1; i < npmem_ranges; i++) {
 253
 254                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 255                         if (hole_pages) {
 256                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 257                                 pmem_holes[npmem_holes++].pages = hole_pages;
 258                                 end_pfn += hole_pages;
 259                         }
 260                         end_pfn += pmem_ranges[i].pages;
 261                 }
 262
 263                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 264                 npmem_ranges = 1;
 265         }
 266 #endif
 267
 268         bootmap_pages = 0;
 269         for (i = 0; i < npmem_ranges; i++)
 270                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
 271
 272         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
 273
 274 #ifdef CONFIG_DISCONTIGMEM
 275         for (i = 0; i < npmem_ranges; i++)
 276                 node_data[i].pg_data.bdata = &bmem_data[i];
 277 #endif
 278         /*
 279          * Initialize and free the full range of memory in each range.
 280          * Note that the only writing these routines do are to the bootmap,
 281          * and we've made sure to locate the bootmap properly so that they
 282          * won't be writing over anything important.
 283          */
 284
 285         bootmap_pfn = bootmap_start_pfn;
 286         max_pfn = 0;
 287         for (i = 0; i < npmem_ranges; i++) {
 288                 unsigned long start_pfn;
 289                 unsigned long npages;
 290
 291                 start_pfn = pmem_ranges[i].start_pfn;
 292                 npages = pmem_ranges[i].pages;
 293
 294                 bootmap_size = init_bootmem_node(NODE_DATA(i),
 295                                                 bootmap_pfn,
 296                                                 start_pfn,
 297                                                 (start_pfn + npages) );
 298                 free_bootmem_node(NODE_DATA(i),
 299                                   (start_pfn << PAGE_SHIFT),
 300                                   (npages << PAGE_SHIFT) );
 301                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 302                 if ((start_pfn + npages) > max_pfn)
 303                         max_pfn = start_pfn + npages;
 304         }
 305
 306         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
 307                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
 308                 BUG();
 309         }
 310
 311         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 312
 313 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 314
 315         reserve_bootmem_node(NODE_DATA(0), 0UL,
 316                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
 317         reserve_bootmem_node(NODE_DATA(0),__pa((unsigned long)&_text),
 318                         (unsigned long)(&_end - &_text));
 319         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
 320                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
 321
 322 #ifndef CONFIG_DISCONTIGMEM
 323
 324         /* reserve the holes */
 325
 326         for (i = 0; i < npmem_holes; i++) {
 327                 reserve_bootmem_node(NODE_DATA(0),
 328                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
 329                                 (pmem_holes[i].pages << PAGE_SHIFT));
 330         }
 331 #endif
 332
 333 #ifdef CONFIG_BLK_DEV_INITRD
 334         if (initrd_start) {
 335                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 336                 if (__pa(initrd_start) < mem_max) {
 337                         unsigned long initrd_reserve;
 338
 339                         if (__pa(initrd_end) > mem_max) {
 340                                 initrd_reserve = mem_max - __pa(initrd_start);
 341                         } else {
 342                                 initrd_reserve = initrd_end - initrd_start;
 343                         }
 344                         initrd_below_start_ok = 1;
 345                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 346
 347                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
 348                 }
 349         }
 350 #endif
 351
 352         data_resource.start =  virt_to_phys(&data_start);
 353         data_resource.end = virt_to_phys(&_end)-1;
 354         code_resource.start = virt_to_phys(&_text);
 355         code_resource.end = virt_to_phys(&data_start)-1;
 356
 357         /* We don't know which region the kernel will be in, so try
 358          * all of them.
 359          */
 360         for (i = 0; i < sysram_resource_count; i++) {
 361                 struct resource *res = &sysram_resources[i];
 362                 request_resource(res, &code_resource);
 363                 request_resource(res, &data_resource);
 364         }
 365         request_resource(&sysram_resources[0], &pdcdata_resource);
 366 }
 367
 368 void free_initmem(void)
 369 {
 370         /* FIXME: */
 371 #if 0
 372         printk(KERN_INFO "NOT FREEING INITMEM (%dk)\n",
 373                         (&__init_end - &__init_begin) >> 10);
 374         return;
 375 #else
 376         unsigned long addr;
 377
 378         printk(KERN_INFO "Freeing unused kernel memory: ");
 379
 380 #if 1
 381         /* Attempt to catch anyone trying to execute code here
 382          * by filling the page with BRK insns.
 383          *
 384          * If we disable interrupts for all CPUs, then IPI stops working.
 385          * Kinda breaks the global cache flushing.
 386          */
 387         local_irq_disable();
 388
 389         memset(&__init_begin, 0x00,
 390                 (unsigned long)&__init_end - (unsigned long)&__init_begin);
 391
 392         flush_data_cache();
 393         asm volatile("sync" : : );
 394         flush_icache_range((unsigned long)&__init_begin, (unsigned long)&__init_end);
 395         asm volatile("sync" : : );
 396
 397         local_irq_enable();
 398 #endif
 399
 400         addr = (unsigned long)(&__init_begin);
 401         for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) {
 402                 ClearPageReserved(virt_to_page(addr));
 403                 set_page_count(virt_to_page(addr), 1);
 404                 free_page(addr);
 405                 num_physpages++;
 406                 totalram_pages++;
 407         }
 408
 409         /* set up a new led state on systems shipped LED State panel */
 410         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 411
 412         printk("%luk freed\n", (unsigned long)(&__init_end - &__init_begin) >> 10);
 413 #endif
 414 }
 415
 416 /*
 417  * Just an arbitrary offset to serve as a "hole" between mapping areas
 418  * (between top of physical memory and a potential pcxl dma mapping
 419  * area, and below the vmalloc mapping area).
 420  *
 421  * The current 32K value just means that there will be a 32K "hole"
 422  * between mapping areas. That means that  any out-of-bounds memory
 423  * accesses will hopefully be caught. The vmalloc() routines leaves
 424  * a hole of 4kB between each vmalloced area for the same reason.
 425  */
 426
 427  /* Leave room for gateway page expansion */
 428 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 429 #error KERNEL_MAP_START is in gateway reserved region
 430 #endif
 431 #define MAP_START (KERNEL_MAP_START)
 432
 433 #define VM_MAP_OFFSET  (32*1024)
 434 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 435                                      & ~(VM_MAP_OFFSET-1)))
 436
 437 void *vmalloc_start;
 438 EXPORT_SYMBOL(vmalloc_start);
 439
 440 #ifdef CONFIG_PA11
 441 unsigned long pcxl_dma_start;
 442 #endif
 443
 444 void __init mem_init(void)
 445 {
 446         int i;
 447
 448         high_memory = __va((max_pfn << PAGE_SHIFT));
 449         max_mapnr = (virt_to_page(high_memory - 1) - mem_map) + 1;
 450
 451         num_physpages = 0;
 452         mem_map = zone_table[0]->zone_mem_map;
 453         for (i = 0; i < npmem_ranges; i++)
 454                 num_physpages += free_all_bootmem_node(NODE_DATA(i));
 455         totalram_pages = num_physpages;
 456
 457         printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
 458
 459 #ifdef CONFIG_PA11
 460         if (hppa_dma_ops == &pcxl_dma_ops) {
 461                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 462                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
 463         } else {
 464                 pcxl_dma_start = 0;
 465                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
 466         }
 467 #else
 468         vmalloc_start = SET_MAP_OFFSET(MAP_START);
 469 #endif
 470
 471 }
 472
 473 int do_check_pgt_cache(int low, int high)
 474 {
 475         return 0;
 476 }
 477
 478 unsigned long *empty_zero_page;
 479
 480 void show_mem(void)
 481 {
 482         int i,free = 0,total = 0,reserved = 0;
 483         int shared = 0, cached = 0;
 484
 485         printk(KERN_INFO "Mem-info:\n");
 486         show_free_areas();
 487         printk(KERN_INFO "Free swap:     %6ldkB\n",
 488                                 nr_swap_pages<<(PAGE_SHIFT-10));
 489         i = max_mapnr;
 490         while (i-- > 0) {
 491                 total++;
 492                 if (PageReserved(mem_map+i))
 493                         reserved++;
 494                 else if (PageSwapCache(mem_map+i))
 495                         cached++;
 496                 else if (!atomic_read(&mem_map[i].count))
 497                         free++;
 498                 else
 499                         shared += atomic_read(&mem_map[i].count) - 1;
 500         }
 501         printk(KERN_INFO "%d pages of RAM\n", total);
 502         printk(KERN_INFO "%d reserved pages\n", reserved);
 503         printk(KERN_INFO "%d pages shared\n", shared);
 504         printk(KERN_INFO "%d pages swap cached\n", cached);
 505 }
 506
 507
 508 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
 509 {
 510         pgd_t *pg_dir;
 511         pmd_t *pmd;
 512         pte_t *pg_table;
 513         unsigned long end_paddr;
 514         unsigned long start_pmd;
 515         unsigned long start_pte;
 516         unsigned long tmp1;
 517         unsigned long tmp2;
 518         unsigned long address;
 519         unsigned long ro_start;
 520         unsigned long ro_end;
 521         unsigned long fv_addr;
 522         unsigned long gw_addr;
 523         extern const unsigned long fault_vector_20;
 524         extern void * const linux_gateway_page;
 525
 526         ro_start = __pa((unsigned long)&_text);
 527         ro_end   = __pa((unsigned long)&data_start);
 528         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
 529         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
 530
 531         end_paddr = start_paddr + size;
 532
 533         pg_dir = pgd_offset_k(start_vaddr);
 534
 535 #if PTRS_PER_PMD == 1
 536         start_pmd = 0;
 537 #else
 538         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 539 #endif
 540         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 541
 542         address = start_paddr;
 543         while (address < end_paddr) {
 544 #if PTRS_PER_PMD == 1
 545                 pmd = (pmd_t *)__pa(pg_dir);
 546 #else
 547                 pmd = (pmd_t *) (PAGE_MASK & pgd_val(*pg_dir));
 548
 549                 /*
 550                  * pmd is physical at this point
 551                  */
 552
 553                 if (!pmd) {
 554                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
 555                         pmd = (pmd_t *) __pa(pmd);
 556                 }
 557
 558                 pgd_val(*pg_dir) = _PAGE_TABLE | (unsigned long) pmd;
 559 #endif
 560                 pg_dir++;
 561
 562                 /* now change pmd to kernel virtual addresses */
 563
 564                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 565                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
 566
 567                         /*
 568                          * pg_table is physical at this point
 569                          */
 570
 571                         pg_table = (pte_t *) (PAGE_MASK & pmd_val(*pmd));
 572                         if (!pg_table) {
 573                                 pg_table = (pte_t *)
 574                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
 575                                 pg_table = (pte_t *) __pa(pg_table);
 576                         }
 577
 578                         pmd_val(*pmd) = _PAGE_TABLE |
 579                                            (unsigned long) pg_table;
 580
 581                         /* now change pg_table to kernel virtual addresses */
 582
 583                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 584                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
 585                                 pte_t pte;
 586
 587 #if !defined(CONFIG_STI_CONSOLE)
 588 #warning STI console should explicitly allocate executable pages but does not
 589                                 /*
 590                                  * Map the fault vector writable so we can
 591                                  * write the HPMC checksum.
 592                                  */
 593                                 if (address >= ro_start && address < ro_end
 594                                                         && address != fv_addr
 595                                                         && address != gw_addr)
 596                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
 597                                 else
 598 #endif
 599                                     pte = __mk_pte(address, pgprot);
 600
 601                                 if (address >= end_paddr)
 602                                         pte_val(pte) = 0;
 603
 604                                 set_pte(pg_table, pte);
 605
 606                                 address += PAGE_SIZE;
 607                         }
 608                         start_pte = 0;
 609
 610                         if (address >= end_paddr)
 611                             break;
 612                 }
 613                 start_pmd = 0;
 614         }
 615 }
 616
 617 /*
 618  * pagetable_init() sets up the page tables
 619  *
 620  * Note that gateway_init() places the Linux gateway page at page 0.
 621  * Since gateway pages cannot be dereferenced this has the desirable
 622  * side effect of trapping those pesky NULL-reference errors in the
 623  * kernel.
 624  */
 625 static void __init pagetable_init(void)
 626 {
 627         int range;
 628
 629         /* Map each physical memory range to its kernel vaddr */
 630
 631         for (range = 0; range < npmem_ranges; range++) {
 632                 unsigned long start_paddr;
 633                 unsigned long end_paddr;
 634                 unsigned long size;
 635
 636                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 637                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
 638                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 639
 640                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 641                         size, PAGE_KERNEL);
 642         }
 643
 644 #ifdef CONFIG_BLK_DEV_INITRD
 645         if (initrd_end && initrd_end > mem_limit) {
 646                 printk("initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 647                 map_pages(initrd_start, __pa(initrd_start),
 648                         initrd_end - initrd_start, PAGE_KERNEL);
 649         }
 650 #endif
 651
 652         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
 653         memset(empty_zero_page, 0, PAGE_SIZE);
 654 }
 655
 656 static void __init gateway_init(void)
 657 {
 658         unsigned long linux_gateway_page_addr;
 659         /* FIXME: This is 'const' in order to trick the compiler
 660            into not treating it as DP-relative data. */
 661         extern void * const linux_gateway_page;
 662
 663         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 664
 665         /*
 666          * Setup Linux Gateway page.
 667          *
 668          * The Linux gateway page will reside in kernel space (on virtual
 669          * page 0), so it doesn't need to be aliased into user space.
 670          */
 671
 672         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 673                 PAGE_SIZE, PAGE_GATEWAY);
 674 }
 675
 676 #ifdef CONFIG_HPUX
 677 void
 678 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
 679 {
 680         pgd_t *pg_dir;
 681         pmd_t *pmd;
 682         pte_t *pg_table;
 683         unsigned long start_pmd;
 684         unsigned long start_pte;
 685         unsigned long address;
 686         unsigned long hpux_gw_page_addr;
 687         /* FIXME: This is 'const' in order to trick the compiler
 688            into not treating it as DP-relative data. */
 689         extern void * const hpux_gateway_page;
 690
 691         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
 692
 693         /*
 694          * Setup HP-UX Gateway page.
 695          *
 696          * The HP-UX gateway page resides in the user address space,
 697          * so it needs to be aliased into each process.
 698          */
 699
 700         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
 701
 702 #if PTRS_PER_PMD == 1
 703         start_pmd = 0;
 704 #else
 705         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 706 #endif
 707         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 708
 709         address = __pa(&hpux_gateway_page);
 710 #if PTRS_PER_PMD == 1
 711         pmd = (pmd_t *)__pa(pg_dir);
 712 #else
 713         pmd = (pmd_t *) (PAGE_MASK & pgd_val(*pg_dir));
 714
 715         /*
 716          * pmd is physical at this point
 717          */
 718
 719         if (!pmd) {
 720                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
 721                 pmd = (pmd_t *) __pa(pmd);
 722         }
 723
 724         pgd_val(*pg_dir) = _PAGE_TABLE | (unsigned long) pmd;
 725 #endif
 726         /* now change pmd to kernel virtual addresses */
 727
 728         pmd = (pmd_t *)__va(pmd) + start_pmd;
 729
 730         /*
 731          * pg_table is physical at this point
 732          */
 733
 734         pg_table = (pte_t *) (PAGE_MASK & pmd_val(*pmd));
 735         if (!pg_table)
 736                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
 737
 738         pmd_val(*pmd) = _PAGE_TABLE | (unsigned long) pg_table;
 739
 740         /* now change pg_table to kernel virtual addresses */
 741
 742         pg_table = (pte_t *) __va(pg_table) + start_pte;
 743         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
 744 }
 745 EXPORT_SYMBOL(map_hpux_gateway_page);
 746 #endif
 747
 748 extern void flush_tlb_all_local(void);
 749
 750 void __init paging_init(void)
 751 {
 752         int i;
 753
 754         setup_bootmem();
 755         pagetable_init();
 756         gateway_init();
 757         flush_cache_all_local(); /* start with known state */
 758         flush_tlb_all_local();
 759
 760         for (i = 0; i < npmem_ranges; i++) {
 761                 unsigned long zones_size[MAX_NR_ZONES] = { 0, 0, 0, };
 762
 763                 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
 764                 free_area_init_node(i,NODE_DATA(i),NULL,zones_size,
 765                                 (pmem_ranges[i].start_pfn << PAGE_SHIFT),0);
 766         }
 767
 768 #ifdef CONFIG_DISCONTIGMEM
 769         /*
 770          * Initialize support for virt_to_page() macro.
 771          *
 772          * Note that MAX_ADDRESS is the largest virtual address that
 773          * we can map. However, since we map all physical memory into
 774          * the kernel address space, it also has an effect on the maximum
 775          * physical address we can map (MAX_ADDRESS - PAGE_OFFSET).
 776          */
 777
 778         maxchunkmap = MAX_ADDRESS >> CHUNKSHIFT;
 779         chunkmap = (unsigned char *)alloc_bootmem(maxchunkmap);
 780
 781         for (i = 0; i < maxchunkmap; i++)
 782             chunkmap[i] = BADCHUNK;
 783
 784         for (i = 0; i < npmem_ranges; i++) {
 785
 786                 ADJ_NODE_MEM_MAP(i) = NODE_MEM_MAP(i) - pmem_ranges[i].start_pfn;
 787                 {
 788                         unsigned long chunk_paddr;
 789                         unsigned long end_paddr;
 790                         int chunknum;
 791
 792                         chunk_paddr = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 793                         end_paddr = chunk_paddr + (pmem_ranges[i].pages << PAGE_SHIFT);
 794                         chunk_paddr &= CHUNKMASK;
 795
 796                         chunknum = (int)CHUNKNUM(chunk_paddr);
 797                         while (chunk_paddr < end_paddr) {
 798                                 if (chunknum >= maxchunkmap)
 799                                         goto badchunkmap1;
 800                                 if (chunkmap[chunknum] != BADCHUNK)
 801                                         goto badchunkmap2;
 802                                 chunkmap[chunknum] = (unsigned char)i;
 803                                 chunk_paddr += CHUNKSZ;
 804                                 chunknum++;
 805                         }
 806                 }
 807         }
 808
 809         return;
 810
 811 badchunkmap1:
 812         panic("paging_init: Physical address exceeds maximum address space!\n");
 813 badchunkmap2:
 814         panic("paging_init: Collision in chunk map array. CHUNKSZ needs to be smaller\n");
 815 #endif
 816 }
 817
 818 #ifdef CONFIG_PA20
 819
 820 /*
 821  * Currently, all PA20 chips have 18 bit protection id's, which is the
 822  * limiting factor (space ids are 32 bits).
 823  */
 824
 825 #define NR_SPACE_IDS 262144
 826
 827 #else
 828
 829 /*
 830  * Currently we have a one-to-one relationship between space id's and
 831  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 832  * support 15 bit protection id's, so that is the limiting factor.
 833  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
 834  * probably not worth the effort for a special case here.
 835  */
 836
 837 #define NR_SPACE_IDS 32768
 838
 839 #endif  /* !CONFIG_PA20 */
 840
 841 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 842 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 843
 844 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 845 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 846 static unsigned long space_id_index;
 847 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 848 static unsigned long dirty_space_ids = 0;
 849
 850 static spinlock_t sid_lock = SPIN_LOCK_UNLOCKED;
 851
 852 unsigned long alloc_sid(void)
 853 {
 854         unsigned long index;
 855
 856         spin_lock(&sid_lock);
 857
 858         if (free_space_ids == 0) {
 859                 if (dirty_space_ids != 0) {
 860                         spin_unlock(&sid_lock);
 861                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 862                         spin_lock(&sid_lock);
 863                 }
 864                 if (free_space_ids == 0)
 865                         BUG();
 866         }
 867
 868         free_space_ids--;
 869
 870         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 871         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 872         space_id_index = index;
 873
 874         spin_unlock(&sid_lock);
 875
 876         return index << SPACEID_SHIFT;
 877 }
 878
 879 void free_sid(unsigned long spaceid)
 880 {
 881         unsigned long index = spaceid >> SPACEID_SHIFT;
 882         unsigned long *dirty_space_offset;
 883
 884         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 885         index &= (BITS_PER_LONG - 1);
 886
 887         spin_lock(&sid_lock);
 888
 889         if (*dirty_space_offset & (1L << index))
 890             BUG(); /* attempt to free space id twice */
 891
 892         *dirty_space_offset |= (1L << index);
 893         dirty_space_ids++;
 894
 895         spin_unlock(&sid_lock);
 896 }
 897
 898
 899 #ifdef CONFIG_SMP
 900 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 901 {
 902         int i;
 903
 904         /* NOTE: sid_lock must be held upon entry */
 905
 906         *ndirtyptr = dirty_space_ids;
 907         if (dirty_space_ids != 0) {
 908             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 909                 dirty_array[i] = dirty_space_id[i];
 910                 dirty_space_id[i] = 0;
 911             }
 912             dirty_space_ids = 0;
 913         }
 914
 915         return;
 916 }
 917
 918 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 919 {
 920         int i;
 921
 922         /* NOTE: sid_lock must be held upon entry */
 923
 924         if (ndirty != 0) {
 925                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 926                         space_id[i] ^= dirty_array[i];
 927                 }
 928
 929                 free_space_ids += ndirty;
 930                 space_id_index = 0;
 931         }
 932 }
 933
 934 #else /* CONFIG_SMP */
 935
 936 static void recycle_sids(void)
 937 {
 938         int i;
 939
 940         /* NOTE: sid_lock must be held upon entry */
 941
 942         if (dirty_space_ids != 0) {
 943                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 944                         space_id[i] ^= dirty_space_id[i];
 945                         dirty_space_id[i] = 0;
 946                 }
 947
 948                 free_space_ids += dirty_space_ids;
 949                 dirty_space_ids = 0;
 950                 space_id_index = 0;
 951         }
 952 }
 953 #endif
 954
 955 /*
 956  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
 957  * purged, we can safely reuse the space ids that were released but
 958  * not flushed from the tlb.
 959  */
 960
 961 #ifdef CONFIG_SMP
 962
 963 static unsigned long recycle_ndirty;
 964 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
 965 static unsigned int recycle_inuse = 0;
 966
 967 void flush_tlb_all(void)
 968 {
 969         int do_recycle;
 970
 971         do_recycle = 0;
 972         spin_lock(&sid_lock);
 973         if (dirty_space_ids > RECYCLE_THRESHOLD) {
 974             if (recycle_inuse) {
 975                 BUG();  /* FIXME: Use a semaphore/wait queue here */
 976             }
 977             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
 978             recycle_inuse++;
 979             do_recycle++;
 980         }
 981         spin_unlock(&sid_lock);
 982         on_each_cpu((void (*)(void *))flush_tlb_all_local, NULL, 1, 1);
 983         if (do_recycle) {
 984             spin_lock(&sid_lock);
 985             recycle_sids(recycle_ndirty,recycle_dirty_array);
 986             recycle_inuse = 0;
 987             spin_unlock(&sid_lock);
 988         }
 989 }
 990 #else
 991 void flush_tlb_all(void)
 992 {
 993         spin_lock(&sid_lock);
 994         flush_tlb_all_local();
 995         recycle_sids();
 996         spin_unlock(&sid_lock);
 997 }
 998 #endif
 999
1000 #ifdef CONFIG_BLK_DEV_INITRD
1001 void free_initrd_mem(unsigned long start, unsigned long end)
1002 {
1003 #if 0
1004         if (start < end)
1005                 printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1006         for (; start < end; start += PAGE_SIZE) {
1007                 ClearPageReserved(virt_to_page(start));
1008                 set_page_count(virt_to_page(start), 1);
1009                 free_page(start);
1010                 num_physpages++;
1011                 totalram_pages++;
1012         }
1013 #endif
1014 }
1015 #endif