arch/parisc/mm/init.c

   1 /*
   2  *  linux/arch/parisc/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright 1999 SuSE GmbH
   6  *    changed by Philipp Rumpf
   7  *  Copyright 1999 Philipp Rumpf (prumpf@tux.org)
   8  *  Copyright 2004 Randolph Chung (tausq@debian.org)
   9  *  Copyright 2006-2007 Helge Deller (deller@gmx.de)
  10  *
  11  */
  12
  13
  14 #include <linux/module.h>
  15 #include <linux/mm.h>
  16 #include <linux/bootmem.h>
  17 #include <linux/memblock.h>
  18 #include <linux/gfp.h>
  19 #include <linux/delay.h>
  20 #include <linux/init.h>
  21 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
  22 #include <linux/initrd.h>
  23 #include <linux/swap.h>
  24 #include <linux/unistd.h>
  25 #include <linux/nodemask.h>     /* for node_online_map */
  26 #include <linux/pagemap.h>      /* for release_pages */
  27 #include <linux/compat.h>
  28
  29 #include <asm/pgalloc.h>
  30 #include <asm/pgtable.h>
  31 #include <asm/tlb.h>
  32 #include <asm/pdc_chassis.h>
  33 #include <asm/mmzone.h>
  34 #include <asm/sections.h>
  35 #include <asm/msgbuf.h>
  36
  37 extern int  data_start;
  38 extern void parisc_kernel_start(void);  /* Kernel entry point in head.S */
  39
  40 #if CONFIG_PGTABLE_LEVELS == 3
  41 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
  42  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
  43  * guarantee that global objects will be laid out in memory in the same order
  44  * as the order of declaration, so put these in different sections and use
  45  * the linker script to order them. */
  46 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
  47 #endif
  48
  49 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
  50 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
  51
  52 #ifdef CONFIG_DISCONTIGMEM
  53 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
  54 signed char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
  55 #endif
  56
  57 static struct resource data_resource = {
  58         .name   = "Kernel data",
  59         .flags  = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
  60 };
  61
  62 static struct resource code_resource = {
  63         .name   = "Kernel code",
  64         .flags  = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
  65 };
  66
  67 static struct resource pdcdata_resource = {
  68         .name   = "PDC data (Page Zero)",
  69         .start  = 0,
  70         .end    = 0x9ff,
  71         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  72 };
  73
  74 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
  75
  76 /* The following array is initialized from the firmware specific
  77  * information retrieved in kernel/inventory.c.
  78  */
  79
  80 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
  81 int npmem_ranges __read_mostly;
  82
  83 /*
  84  * get_memblock() allocates pages via memblock.
  85  * We can't use memblock_find_in_range(0, KERNEL_INITIAL_SIZE) here since it
  86  * doesn't allocate from bottom to top which is needed because we only created
  87  * the initial mapping up to KERNEL_INITIAL_SIZE in the assembly bootup code.
  88  */
  89 static void * __init get_memblock(unsigned long size)
  90 {
  91         static phys_addr_t search_addr __initdata;
  92         phys_addr_t phys;
  93
  94         if (!search_addr)
  95                 search_addr = PAGE_ALIGN(__pa((unsigned long) &_end));
  96         search_addr = ALIGN(search_addr, size);
  97         while (!memblock_is_region_memory(search_addr, size) ||
  98                 memblock_is_region_reserved(search_addr, size)) {
  99                 search_addr += size;
 100         }
 101         phys = search_addr;
 102
 103         if (phys)
 104                 memblock_reserve(phys, size);
 105         else
 106                 panic("get_memblock() failed.\n");
 107
 108         memset(__va(phys), 0, size);
 109
 110         return __va(phys);
 111 }
 112
 113 #ifdef CONFIG_64BIT
 114 #define MAX_MEM         (~0UL)
 115 #else /* !CONFIG_64BIT */
 116 #define MAX_MEM         (3584U*1024U*1024U)
 117 #endif /* !CONFIG_64BIT */
 118
 119 static unsigned long mem_limit __read_mostly = MAX_MEM;
 120
 121 static void __init mem_limit_func(void)
 122 {
 123         char *cp, *end;
 124         unsigned long limit;
 125
 126         /* We need this before __setup() functions are called */
 127
 128         limit = MAX_MEM;
 129         for (cp = boot_command_line; *cp; ) {
 130                 if (memcmp(cp, "mem=", 4) == 0) {
 131                         cp += 4;
 132                         limit = memparse(cp, &end);
 133                         if (end != cp)
 134                                 break;
 135                         cp = end;
 136                 } else {
 137                         while (*cp != ' ' && *cp)
 138                                 ++cp;
 139                         while (*cp == ' ')
 140                                 ++cp;
 141                 }
 142         }
 143
 144         if (limit < mem_limit)
 145                 mem_limit = limit;
 146 }
 147
 148 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 149
 150 static void __init setup_bootmem(void)
 151 {
 152         unsigned long mem_max;
 153 #ifndef CONFIG_DISCONTIGMEM
 154         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 155         int npmem_holes;
 156 #endif
 157         int i, sysram_resource_count;
 158
 159         disable_sr_hashing(); /* Turn off space register hashing */
 160
 161         /*
 162          * Sort the ranges. Since the number of ranges is typically
 163          * small, and performance is not an issue here, just do
 164          * a simple insertion sort.
 165          */
 166
 167         for (i = 1; i < npmem_ranges; i++) {
 168                 int j;
 169
 170                 for (j = i; j > 0; j--) {
 171                         unsigned long tmp;
 172
 173                         if (pmem_ranges[j-1].start_pfn <
 174                             pmem_ranges[j].start_pfn) {
 175
 176                                 break;
 177                         }
 178                         tmp = pmem_ranges[j-1].start_pfn;
 179                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 180                         pmem_ranges[j].start_pfn = tmp;
 181                         tmp = pmem_ranges[j-1].pages;
 182                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 183                         pmem_ranges[j].pages = tmp;
 184                 }
 185         }
 186
 187 #ifndef CONFIG_DISCONTIGMEM
 188         /*
 189          * Throw out ranges that are too far apart (controlled by
 190          * MAX_GAP).
 191          */
 192
 193         for (i = 1; i < npmem_ranges; i++) {
 194                 if (pmem_ranges[i].start_pfn -
 195                         (pmem_ranges[i-1].start_pfn +
 196                          pmem_ranges[i-1].pages) > MAX_GAP) {
 197                         npmem_ranges = i;
 198                         printk("Large gap in memory detected (%ld pages). "
 199                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 200                                pmem_ranges[i].start_pfn -
 201                                (pmem_ranges[i-1].start_pfn +
 202                                 pmem_ranges[i-1].pages));
 203                         break;
 204                 }
 205         }
 206 #endif
 207
 208         /* Print the memory ranges */
 209         pr_info("Memory Ranges:\n");
 210
 211         for (i = 0; i < npmem_ranges; i++) {
 212                 struct resource *res = &sysram_resources[i];
 213                 unsigned long start;
 214                 unsigned long size;
 215
 216                 size = (pmem_ranges[i].pages << PAGE_SHIFT);
 217                 start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 218                 pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 219                         i, start, start + (size - 1), size >> 20);
 220
 221                 /* request memory resource */
 222                 res->name = "System RAM";
 223                 res->start = start;
 224                 res->end = start + size - 1;
 225                 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 226                 request_resource(&iomem_resource, res);
 227         }
 228
 229         sysram_resource_count = npmem_ranges;
 230
 231         /*
 232          * For 32 bit kernels we limit the amount of memory we can
 233          * support, in order to preserve enough kernel address space
 234          * for other purposes. For 64 bit kernels we don't normally
 235          * limit the memory, but this mechanism can be used to
 236          * artificially limit the amount of memory (and it is written
 237          * to work with multiple memory ranges).
 238          */
 239
 240         mem_limit_func();       /* check for "mem=" argument */
 241
 242         mem_max = 0;
 243         for (i = 0; i < npmem_ranges; i++) {
 244                 unsigned long rsize;
 245
 246                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 247                 if ((mem_max + rsize) > mem_limit) {
 248                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 249                         if (mem_max == mem_limit)
 250                                 npmem_ranges = i;
 251                         else {
 252                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 253                                                        - (mem_max >> PAGE_SHIFT);
 254                                 npmem_ranges = i + 1;
 255                                 mem_max = mem_limit;
 256                         }
 257                         break;
 258                 }
 259                 mem_max += rsize;
 260         }
 261
 262         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 263
 264 #ifndef CONFIG_DISCONTIGMEM
 265         /* Merge the ranges, keeping track of the holes */
 266
 267         {
 268                 unsigned long end_pfn;
 269                 unsigned long hole_pages;
 270
 271                 npmem_holes = 0;
 272                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 273                 for (i = 1; i < npmem_ranges; i++) {
 274
 275                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 276                         if (hole_pages) {
 277                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 278                                 pmem_holes[npmem_holes++].pages = hole_pages;
 279                                 end_pfn += hole_pages;
 280                         }
 281                         end_pfn += pmem_ranges[i].pages;
 282                 }
 283
 284                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 285                 npmem_ranges = 1;
 286         }
 287 #endif
 288
 289 #ifdef CONFIG_DISCONTIGMEM
 290         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 291                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 292         }
 293         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 294
 295         for (i = 0; i < npmem_ranges; i++) {
 296                 node_set_state(i, N_NORMAL_MEMORY);
 297                 node_set_online(i);
 298         }
 299 #endif
 300
 301         /*
 302          * Initialize and free the full range of memory in each range.
 303          */
 304
 305         max_pfn = 0;
 306         for (i = 0; i < npmem_ranges; i++) {
 307                 unsigned long start_pfn;
 308                 unsigned long npages;
 309                 unsigned long start;
 310                 unsigned long size;
 311
 312                 start_pfn = pmem_ranges[i].start_pfn;
 313                 npages = pmem_ranges[i].pages;
 314
 315                 start = start_pfn << PAGE_SHIFT;
 316                 size = npages << PAGE_SHIFT;
 317
 318                 /* add system RAM memblock */
 319                 memblock_add(start, size);
 320
 321                 if ((start_pfn + npages) > max_pfn)
 322                         max_pfn = start_pfn + npages;
 323         }
 324
 325         /* IOMMU is always used to access "high mem" on those boxes
 326          * that can support enough mem that a PCI device couldn't
 327          * directly DMA to any physical addresses.
 328          * ISA DMA support will need to revisit this.
 329          */
 330         max_low_pfn = max_pfn;
 331
 332         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 333
 334 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 335
 336         memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
 337                                 PDC_CONSOLE_IO_IODC_SIZE));
 338         memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
 339                         (unsigned long)(_end - KERNEL_BINARY_TEXT_START));
 340
 341 #ifndef CONFIG_DISCONTIGMEM
 342
 343         /* reserve the holes */
 344
 345         for (i = 0; i < npmem_holes; i++) {
 346                 memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
 347                                 (pmem_holes[i].pages << PAGE_SHIFT));
 348         }
 349 #endif
 350
 351 #ifdef CONFIG_BLK_DEV_INITRD
 352         if (initrd_start) {
 353                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 354                 if (__pa(initrd_start) < mem_max) {
 355                         unsigned long initrd_reserve;
 356
 357                         if (__pa(initrd_end) > mem_max) {
 358                                 initrd_reserve = mem_max - __pa(initrd_start);
 359                         } else {
 360                                 initrd_reserve = initrd_end - initrd_start;
 361                         }
 362                         initrd_below_start_ok = 1;
 363                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 364
 365                         memblock_reserve(__pa(initrd_start), initrd_reserve);
 366                 }
 367         }
 368 #endif
 369
 370         data_resource.start =  virt_to_phys(&data_start);
 371         data_resource.end = virt_to_phys(_end) - 1;
 372         code_resource.start = virt_to_phys(_text);
 373         code_resource.end = virt_to_phys(&data_start)-1;
 374
 375         /* We don't know which region the kernel will be in, so try
 376          * all of them.
 377          */
 378         for (i = 0; i < sysram_resource_count; i++) {
 379                 struct resource *res = &sysram_resources[i];
 380                 request_resource(res, &code_resource);
 381                 request_resource(res, &data_resource);
 382         }
 383         request_resource(&sysram_resources[0], &pdcdata_resource);
 384
 385         /* Initialize Page Deallocation Table (PDT) and check for bad memory. */
 386         pdc_pdt_init();
 387 }
 388
 389 static int __init parisc_text_address(unsigned long vaddr)
 390 {
 391         static unsigned long head_ptr __initdata;
 392
 393         if (!head_ptr)
 394                 head_ptr = PAGE_MASK & (unsigned long)
 395                         dereference_function_descriptor(&parisc_kernel_start);
 396
 397         return core_kernel_text(vaddr) || vaddr == head_ptr;
 398 }
 399
 400 static void __init map_pages(unsigned long start_vaddr,
 401                              unsigned long start_paddr, unsigned long size,
 402                              pgprot_t pgprot, int force)
 403 {
 404         pgd_t *pg_dir;
 405         pmd_t *pmd;
 406         pte_t *pg_table;
 407         unsigned long end_paddr;
 408         unsigned long start_pmd;
 409         unsigned long start_pte;
 410         unsigned long tmp1;
 411         unsigned long tmp2;
 412         unsigned long address;
 413         unsigned long vaddr;
 414         unsigned long ro_start;
 415         unsigned long ro_end;
 416         unsigned long kernel_end;
 417
 418         ro_start = __pa((unsigned long)_text);
 419         ro_end   = __pa((unsigned long)&data_start);
 420         kernel_end  = __pa((unsigned long)&_end);
 421
 422         end_paddr = start_paddr + size;
 423
 424         pg_dir = pgd_offset_k(start_vaddr);
 425
 426 #if PTRS_PER_PMD == 1
 427         start_pmd = 0;
 428 #else
 429         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 430 #endif
 431         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 432
 433         address = start_paddr;
 434         vaddr = start_vaddr;
 435         while (address < end_paddr) {
 436 #if PTRS_PER_PMD == 1
 437                 pmd = (pmd_t *)__pa(pg_dir);
 438 #else
 439                 pmd = (pmd_t *)pgd_address(*pg_dir);
 440
 441                 /*
 442                  * pmd is physical at this point
 443                  */
 444
 445                 if (!pmd) {
 446                         pmd = (pmd_t *) get_memblock(PAGE_SIZE << PMD_ORDER);
 447                         pmd = (pmd_t *) __pa(pmd);
 448                 }
 449
 450                 pgd_populate(NULL, pg_dir, __va(pmd));
 451 #endif
 452                 pg_dir++;
 453
 454                 /* now change pmd to kernel virtual addresses */
 455
 456                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 457                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
 458
 459                         /*
 460                          * pg_table is physical at this point
 461                          */
 462
 463                         pg_table = (pte_t *)pmd_address(*pmd);
 464                         if (!pg_table) {
 465                                 pg_table = (pte_t *) get_memblock(PAGE_SIZE);
 466                                 pg_table = (pte_t *) __pa(pg_table);
 467                         }
 468
 469                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 470
 471                         /* now change pg_table to kernel virtual addresses */
 472
 473                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 474                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
 475                                 pte_t pte;
 476
 477                                 if (force)
 478                                         pte =  __mk_pte(address, pgprot);
 479                                 else if (parisc_text_address(vaddr)) {
 480                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 481                                         if (address >= ro_start && address < kernel_end)
 482                                                 pte = pte_mkhuge(pte);
 483                                 }
 484                                 else
 485 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
 486                                 if (address >= ro_start && address < ro_end) {
 487                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 488                                         pte = pte_mkhuge(pte);
 489                                 } else
 490 #endif
 491                                 {
 492                                         pte = __mk_pte(address, pgprot);
 493                                         if (address >= ro_start && address < kernel_end)
 494                                                 pte = pte_mkhuge(pte);
 495                                 }
 496
 497                                 if (address >= end_paddr) {
 498                                         if (force)
 499                                                 break;
 500                                         else
 501                                                 pte_val(pte) = 0;
 502                                 }
 503
 504                                 set_pte(pg_table, pte);
 505
 506                                 address += PAGE_SIZE;
 507                                 vaddr += PAGE_SIZE;
 508                         }
 509                         start_pte = 0;
 510
 511                         if (address >= end_paddr)
 512                             break;
 513                 }
 514                 start_pmd = 0;
 515         }
 516 }
 517
 518 void free_initmem(void)
 519 {
 520         unsigned long init_begin = (unsigned long)__init_begin;
 521         unsigned long init_end = (unsigned long)__init_end;
 522
 523         /* The init text pages are marked R-X.  We have to
 524          * flush the icache and mark them RW-
 525          *
 526          * This is tricky, because map_pages is in the init section.
 527          * Do a dummy remap of the data section first (the data
 528          * section is already PAGE_KERNEL) to pull in the TLB entries
 529          * for map_kernel */
 530         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 531                   PAGE_KERNEL_RWX, 1);
 532         /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
 533          * map_pages */
 534         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 535                   PAGE_KERNEL, 1);
 536
 537         /* force the kernel to see the new TLB entries */
 538         __flush_tlb_range(0, init_begin, init_end);
 539
 540         /* finally dump all the instructions which were cached, since the
 541          * pages are no-longer executable */
 542         flush_icache_range(init_begin, init_end);
 543
 544         free_initmem_default(POISON_FREE_INITMEM);
 545
 546         /* set up a new led state on systems shipped LED State panel */
 547         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 548 }
 549
 550
 551 #ifdef CONFIG_STRICT_KERNEL_RWX
 552 void mark_rodata_ro(void)
 553 {
 554         /* rodata memory was already mapped with KERNEL_RO access rights by
 555            pagetable_init() and map_pages(). No need to do additional stuff here */
 556         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 557                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
 558 }
 559 #endif
 560
 561
 562 /*
 563  * Just an arbitrary offset to serve as a "hole" between mapping areas
 564  * (between top of physical memory and a potential pcxl dma mapping
 565  * area, and below the vmalloc mapping area).
 566  *
 567  * The current 32K value just means that there will be a 32K "hole"
 568  * between mapping areas. That means that  any out-of-bounds memory
 569  * accesses will hopefully be caught. The vmalloc() routines leaves
 570  * a hole of 4kB between each vmalloced area for the same reason.
 571  */
 572
 573  /* Leave room for gateway page expansion */
 574 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 575 #error KERNEL_MAP_START is in gateway reserved region
 576 #endif
 577 #define MAP_START (KERNEL_MAP_START)
 578
 579 #define VM_MAP_OFFSET  (32*1024)
 580 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 581                                      & ~(VM_MAP_OFFSET-1)))
 582
 583 void *parisc_vmalloc_start __read_mostly;
 584 EXPORT_SYMBOL(parisc_vmalloc_start);
 585
 586 #ifdef CONFIG_PA11
 587 unsigned long pcxl_dma_start __read_mostly;
 588 #endif
 589
 590 void __init mem_init(void)
 591 {
 592         /* Do sanity checks on IPC (compat) structures */
 593         BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
 594 #ifndef CONFIG_64BIT
 595         BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
 596         BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
 597         BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
 598 #endif
 599 #ifdef CONFIG_COMPAT
 600         BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
 601         BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
 602         BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
 603         BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
 604 #endif
 605
 606         /* Do sanity checks on page table constants */
 607         BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
 608         BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
 609         BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 610         BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 611                         > BITS_PER_LONG);
 612
 613         high_memory = __va((max_pfn << PAGE_SHIFT));
 614         set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
 615         free_all_bootmem();
 616
 617 #ifdef CONFIG_PA11
 618         if (hppa_dma_ops == &pcxl_dma_ops) {
 619                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 620                 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
 621                                                 + PCXL_DMA_MAP_SIZE);
 622         } else {
 623                 pcxl_dma_start = 0;
 624                 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 625         }
 626 #else
 627         parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 628 #endif
 629
 630         mem_init_print_info(NULL);
 631 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
 632         printk("virtual kernel memory layout:\n"
 633                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
 634                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
 635                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
 636                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
 637                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
 638
 639                (void*)VMALLOC_START, (void*)VMALLOC_END,
 640                (VMALLOC_END - VMALLOC_START) >> 20,
 641
 642                __va(0), high_memory,
 643                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 644
 645                __init_begin, __init_end,
 646                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
 647
 648                _etext, _edata,
 649                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
 650
 651                _text, _etext,
 652                ((unsigned long)_etext - (unsigned long)_text) >> 10);
 653 #endif
 654 }
 655
 656 unsigned long *empty_zero_page __read_mostly;
 657 EXPORT_SYMBOL(empty_zero_page);
 658
 659 /*
 660  * pagetable_init() sets up the page tables
 661  *
 662  * Note that gateway_init() places the Linux gateway page at page 0.
 663  * Since gateway pages cannot be dereferenced this has the desirable
 664  * side effect of trapping those pesky NULL-reference errors in the
 665  * kernel.
 666  */
 667 static void __init pagetable_init(void)
 668 {
 669         int range;
 670
 671         /* Map each physical memory range to its kernel vaddr */
 672
 673         for (range = 0; range < npmem_ranges; range++) {
 674                 unsigned long start_paddr;
 675                 unsigned long end_paddr;
 676                 unsigned long size;
 677
 678                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 679                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 680                 end_paddr = start_paddr + size;
 681
 682                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 683                           size, PAGE_KERNEL, 0);
 684         }
 685
 686 #ifdef CONFIG_BLK_DEV_INITRD
 687         if (initrd_end && initrd_end > mem_limit) {
 688                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 689                 map_pages(initrd_start, __pa(initrd_start),
 690                           initrd_end - initrd_start, PAGE_KERNEL, 0);
 691         }
 692 #endif
 693
 694         empty_zero_page = get_memblock(PAGE_SIZE);
 695 }
 696
 697 static void __init gateway_init(void)
 698 {
 699         unsigned long linux_gateway_page_addr;
 700         /* FIXME: This is 'const' in order to trick the compiler
 701            into not treating it as DP-relative data. */
 702         extern void * const linux_gateway_page;
 703
 704         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 705
 706         /*
 707          * Setup Linux Gateway page.
 708          *
 709          * The Linux gateway page will reside in kernel space (on virtual
 710          * page 0), so it doesn't need to be aliased into user space.
 711          */
 712
 713         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 714                   PAGE_SIZE, PAGE_GATEWAY, 1);
 715 }
 716
 717 void __init paging_init(void)
 718 {
 719         int i;
 720
 721         setup_bootmem();
 722         pagetable_init();
 723         gateway_init();
 724         flush_cache_all_local(); /* start with known state */
 725         flush_tlb_all_local(NULL);
 726
 727         for (i = 0; i < npmem_ranges; i++) {
 728                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
 729
 730                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
 731
 732 #ifdef CONFIG_DISCONTIGMEM
 733                 /* Need to initialize the pfnnid_map before we can initialize
 734                    the zone */
 735                 {
 736                     int j;
 737                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 738                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 739                          j++) {
 740                         pfnnid_map[j] = i;
 741                     }
 742                 }
 743 #endif
 744
 745                 free_area_init_node(i, zones_size,
 746                                 pmem_ranges[i].start_pfn, NULL);
 747         }
 748 }
 749
 750 #ifdef CONFIG_PA20
 751
 752 /*
 753  * Currently, all PA20 chips have 18 bit protection IDs, which is the
 754  * limiting factor (space ids are 32 bits).
 755  */
 756
 757 #define NR_SPACE_IDS 262144
 758
 759 #else
 760
 761 /*
 762  * Currently we have a one-to-one relationship between space IDs and
 763  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 764  * support 15 bit protection IDs, so that is the limiting factor.
 765  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
 766  * probably not worth the effort for a special case here.
 767  */
 768
 769 #define NR_SPACE_IDS 32768
 770
 771 #endif  /* !CONFIG_PA20 */
 772
 773 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 774 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 775
 776 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 777 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 778 static unsigned long space_id_index;
 779 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 780 static unsigned long dirty_space_ids = 0;
 781
 782 static DEFINE_SPINLOCK(sid_lock);
 783
 784 unsigned long alloc_sid(void)
 785 {
 786         unsigned long index;
 787
 788         spin_lock(&sid_lock);
 789
 790         if (free_space_ids == 0) {
 791                 if (dirty_space_ids != 0) {
 792                         spin_unlock(&sid_lock);
 793                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 794                         spin_lock(&sid_lock);
 795                 }
 796                 BUG_ON(free_space_ids == 0);
 797         }
 798
 799         free_space_ids--;
 800
 801         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 802         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 803         space_id_index = index;
 804
 805         spin_unlock(&sid_lock);
 806
 807         return index << SPACEID_SHIFT;
 808 }
 809
 810 void free_sid(unsigned long spaceid)
 811 {
 812         unsigned long index = spaceid >> SPACEID_SHIFT;
 813         unsigned long *dirty_space_offset;
 814
 815         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 816         index &= (BITS_PER_LONG - 1);
 817
 818         spin_lock(&sid_lock);
 819
 820         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
 821
 822         *dirty_space_offset |= (1L << index);
 823         dirty_space_ids++;
 824
 825         spin_unlock(&sid_lock);
 826 }
 827
 828
 829 #ifdef CONFIG_SMP
 830 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 831 {
 832         int i;
 833
 834         /* NOTE: sid_lock must be held upon entry */
 835
 836         *ndirtyptr = dirty_space_ids;
 837         if (dirty_space_ids != 0) {
 838             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 839                 dirty_array[i] = dirty_space_id[i];
 840                 dirty_space_id[i] = 0;
 841             }
 842             dirty_space_ids = 0;
 843         }
 844
 845         return;
 846 }
 847
 848 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 849 {
 850         int i;
 851
 852         /* NOTE: sid_lock must be held upon entry */
 853
 854         if (ndirty != 0) {
 855                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 856                         space_id[i] ^= dirty_array[i];
 857                 }
 858
 859                 free_space_ids += ndirty;
 860                 space_id_index = 0;
 861         }
 862 }
 863
 864 #else /* CONFIG_SMP */
 865
 866 static void recycle_sids(void)
 867 {
 868         int i;
 869
 870         /* NOTE: sid_lock must be held upon entry */
 871
 872         if (dirty_space_ids != 0) {
 873                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 874                         space_id[i] ^= dirty_space_id[i];
 875                         dirty_space_id[i] = 0;
 876                 }
 877
 878                 free_space_ids += dirty_space_ids;
 879                 dirty_space_ids = 0;
 880                 space_id_index = 0;
 881         }
 882 }
 883 #endif
 884
 885 /*
 886  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
 887  * purged, we can safely reuse the space ids that were released but
 888  * not flushed from the tlb.
 889  */
 890
 891 #ifdef CONFIG_SMP
 892
 893 static unsigned long recycle_ndirty;
 894 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
 895 static unsigned int recycle_inuse;
 896
 897 void flush_tlb_all(void)
 898 {
 899         int do_recycle;
 900
 901         __inc_irq_stat(irq_tlb_count);
 902         do_recycle = 0;
 903         spin_lock(&sid_lock);
 904         if (dirty_space_ids > RECYCLE_THRESHOLD) {
 905             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
 906             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
 907             recycle_inuse++;
 908             do_recycle++;
 909         }
 910         spin_unlock(&sid_lock);
 911         on_each_cpu(flush_tlb_all_local, NULL, 1);
 912         if (do_recycle) {
 913             spin_lock(&sid_lock);
 914             recycle_sids(recycle_ndirty,recycle_dirty_array);
 915             recycle_inuse = 0;
 916             spin_unlock(&sid_lock);
 917         }
 918 }
 919 #else
 920 void flush_tlb_all(void)
 921 {
 922         __inc_irq_stat(irq_tlb_count);
 923         spin_lock(&sid_lock);
 924         flush_tlb_all_local(NULL);
 925         recycle_sids();
 926         spin_unlock(&sid_lock);
 927 }
 928 #endif
 929
 930 #ifdef CONFIG_BLK_DEV_INITRD
 931 void free_initrd_mem(unsigned long start, unsigned long end)
 932 {
 933         free_reserved_area((void *)start, (void *)end, -1, "initrd");
 934 }
 935 #endif