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         return __va(phys);
 109 }
 110
 111 #ifdef CONFIG_64BIT
 112 #define MAX_MEM         (~0UL)
 113 #else /* !CONFIG_64BIT */
 114 #define MAX_MEM         (3584U*1024U*1024U)
 115 #endif /* !CONFIG_64BIT */
 116
 117 static unsigned long mem_limit __read_mostly = MAX_MEM;
 118
 119 static void __init mem_limit_func(void)
 120 {
 121         char *cp, *end;
 122         unsigned long limit;
 123
 124         /* We need this before __setup() functions are called */
 125
 126         limit = MAX_MEM;
 127         for (cp = boot_command_line; *cp; ) {
 128                 if (memcmp(cp, "mem=", 4) == 0) {
 129                         cp += 4;
 130                         limit = memparse(cp, &end);
 131                         if (end != cp)
 132                                 break;
 133                         cp = end;
 134                 } else {
 135                         while (*cp != ' ' && *cp)
 136                                 ++cp;
 137                         while (*cp == ' ')
 138                                 ++cp;
 139                 }
 140         }
 141
 142         if (limit < mem_limit)
 143                 mem_limit = limit;
 144 }
 145
 146 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 147
 148 static void __init setup_bootmem(void)
 149 {
 150         unsigned long mem_max;
 151 #ifndef CONFIG_DISCONTIGMEM
 152         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 153         int npmem_holes;
 154 #endif
 155         int i, sysram_resource_count;
 156
 157         disable_sr_hashing(); /* Turn off space register hashing */
 158
 159         /*
 160          * Sort the ranges. Since the number of ranges is typically
 161          * small, and performance is not an issue here, just do
 162          * a simple insertion sort.
 163          */
 164
 165         for (i = 1; i < npmem_ranges; i++) {
 166                 int j;
 167
 168                 for (j = i; j > 0; j--) {
 169                         unsigned long tmp;
 170
 171                         if (pmem_ranges[j-1].start_pfn <
 172                             pmem_ranges[j].start_pfn) {
 173
 174                                 break;
 175                         }
 176                         tmp = pmem_ranges[j-1].start_pfn;
 177                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 178                         pmem_ranges[j].start_pfn = tmp;
 179                         tmp = pmem_ranges[j-1].pages;
 180                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 181                         pmem_ranges[j].pages = tmp;
 182                 }
 183         }
 184
 185 #ifndef CONFIG_DISCONTIGMEM
 186         /*
 187          * Throw out ranges that are too far apart (controlled by
 188          * MAX_GAP).
 189          */
 190
 191         for (i = 1; i < npmem_ranges; i++) {
 192                 if (pmem_ranges[i].start_pfn -
 193                         (pmem_ranges[i-1].start_pfn +
 194                          pmem_ranges[i-1].pages) > MAX_GAP) {
 195                         npmem_ranges = i;
 196                         printk("Large gap in memory detected (%ld pages). "
 197                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 198                                pmem_ranges[i].start_pfn -
 199                                (pmem_ranges[i-1].start_pfn +
 200                                 pmem_ranges[i-1].pages));
 201                         break;
 202                 }
 203         }
 204 #endif
 205
 206         /* Print the memory ranges */
 207         pr_info("Memory Ranges:\n");
 208
 209         for (i = 0; i < npmem_ranges; i++) {
 210                 struct resource *res = &sysram_resources[i];
 211                 unsigned long start;
 212                 unsigned long size;
 213
 214                 size = (pmem_ranges[i].pages << PAGE_SHIFT);
 215                 start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 216                 pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 217                         i, start, start + (size - 1), size >> 20);
 218
 219                 /* request memory resource */
 220                 res->name = "System RAM";
 221                 res->start = start;
 222                 res->end = start + size - 1;
 223                 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 224                 request_resource(&iomem_resource, res);
 225         }
 226
 227         sysram_resource_count = npmem_ranges;
 228
 229         /*
 230          * For 32 bit kernels we limit the amount of memory we can
 231          * support, in order to preserve enough kernel address space
 232          * for other purposes. For 64 bit kernels we don't normally
 233          * limit the memory, but this mechanism can be used to
 234          * artificially limit the amount of memory (and it is written
 235          * to work with multiple memory ranges).
 236          */
 237
 238         mem_limit_func();       /* check for "mem=" argument */
 239
 240         mem_max = 0;
 241         for (i = 0; i < npmem_ranges; i++) {
 242                 unsigned long rsize;
 243
 244                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 245                 if ((mem_max + rsize) > mem_limit) {
 246                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 247                         if (mem_max == mem_limit)
 248                                 npmem_ranges = i;
 249                         else {
 250                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 251                                                        - (mem_max >> PAGE_SHIFT);
 252                                 npmem_ranges = i + 1;
 253                                 mem_max = mem_limit;
 254                         }
 255                         break;
 256                 }
 257                 mem_max += rsize;
 258         }
 259
 260         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 261
 262 #ifndef CONFIG_DISCONTIGMEM
 263         /* Merge the ranges, keeping track of the holes */
 264
 265         {
 266                 unsigned long end_pfn;
 267                 unsigned long hole_pages;
 268
 269                 npmem_holes = 0;
 270                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 271                 for (i = 1; i < npmem_ranges; i++) {
 272
 273                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 274                         if (hole_pages) {
 275                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 276                                 pmem_holes[npmem_holes++].pages = hole_pages;
 277                                 end_pfn += hole_pages;
 278                         }
 279                         end_pfn += pmem_ranges[i].pages;
 280                 }
 281
 282                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 283                 npmem_ranges = 1;
 284         }
 285 #endif
 286
 287 #ifdef CONFIG_DISCONTIGMEM
 288         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 289                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 290         }
 291         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 292
 293         for (i = 0; i < npmem_ranges; i++) {
 294                 node_set_state(i, N_NORMAL_MEMORY);
 295                 node_set_online(i);
 296         }
 297 #endif
 298
 299         /*
 300          * Initialize and free the full range of memory in each range.
 301          */
 302
 303         max_pfn = 0;
 304         for (i = 0; i < npmem_ranges; i++) {
 305                 unsigned long start_pfn;
 306                 unsigned long npages;
 307                 unsigned long start;
 308                 unsigned long size;
 309
 310                 start_pfn = pmem_ranges[i].start_pfn;
 311                 npages = pmem_ranges[i].pages;
 312
 313                 start = start_pfn << PAGE_SHIFT;
 314                 size = npages << PAGE_SHIFT;
 315
 316                 /* add system RAM memblock */
 317                 memblock_add(start, size);
 318
 319                 if ((start_pfn + npages) > max_pfn)
 320                         max_pfn = start_pfn + npages;
 321         }
 322
 323         /* IOMMU is always used to access "high mem" on those boxes
 324          * that can support enough mem that a PCI device couldn't
 325          * directly DMA to any physical addresses.
 326          * ISA DMA support will need to revisit this.
 327          */
 328         max_low_pfn = max_pfn;
 329
 330         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 331
 332 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 333
 334         memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
 335                                 PDC_CONSOLE_IO_IODC_SIZE));
 336         memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
 337                         (unsigned long)(_end - KERNEL_BINARY_TEXT_START));
 338
 339 #ifndef CONFIG_DISCONTIGMEM
 340
 341         /* reserve the holes */
 342
 343         for (i = 0; i < npmem_holes; i++) {
 344                 memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
 345                                 (pmem_holes[i].pages << PAGE_SHIFT));
 346         }
 347 #endif
 348
 349 #ifdef CONFIG_BLK_DEV_INITRD
 350         if (initrd_start) {
 351                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 352                 if (__pa(initrd_start) < mem_max) {
 353                         unsigned long initrd_reserve;
 354
 355                         if (__pa(initrd_end) > mem_max) {
 356                                 initrd_reserve = mem_max - __pa(initrd_start);
 357                         } else {
 358                                 initrd_reserve = initrd_end - initrd_start;
 359                         }
 360                         initrd_below_start_ok = 1;
 361                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 362
 363                         memblock_reserve(__pa(initrd_start), initrd_reserve);
 364                 }
 365         }
 366 #endif
 367
 368         data_resource.start =  virt_to_phys(&data_start);
 369         data_resource.end = virt_to_phys(_end) - 1;
 370         code_resource.start = virt_to_phys(_text);
 371         code_resource.end = virt_to_phys(&data_start)-1;
 372
 373         /* We don't know which region the kernel will be in, so try
 374          * all of them.
 375          */
 376         for (i = 0; i < sysram_resource_count; i++) {
 377                 struct resource *res = &sysram_resources[i];
 378                 request_resource(res, &code_resource);
 379                 request_resource(res, &data_resource);
 380         }
 381         request_resource(&sysram_resources[0], &pdcdata_resource);
 382 }
 383
 384 static int __init parisc_text_address(unsigned long vaddr)
 385 {
 386         static unsigned long head_ptr __initdata;
 387
 388         if (!head_ptr)
 389                 head_ptr = PAGE_MASK & (unsigned long)
 390                         dereference_function_descriptor(&parisc_kernel_start);
 391
 392         return core_kernel_text(vaddr) || vaddr == head_ptr;
 393 }
 394
 395 static void __init map_pages(unsigned long start_vaddr,
 396                              unsigned long start_paddr, unsigned long size,
 397                              pgprot_t pgprot, int force)
 398 {
 399         pgd_t *pg_dir;
 400         pmd_t *pmd;
 401         pte_t *pg_table;
 402         unsigned long end_paddr;
 403         unsigned long start_pmd;
 404         unsigned long start_pte;
 405         unsigned long tmp1;
 406         unsigned long tmp2;
 407         unsigned long address;
 408         unsigned long vaddr;
 409         unsigned long ro_start;
 410         unsigned long ro_end;
 411         unsigned long kernel_end;
 412
 413         ro_start = __pa((unsigned long)_text);
 414         ro_end   = __pa((unsigned long)&data_start);
 415         kernel_end  = __pa((unsigned long)&_end);
 416
 417         end_paddr = start_paddr + size;
 418
 419         pg_dir = pgd_offset_k(start_vaddr);
 420
 421 #if PTRS_PER_PMD == 1
 422         start_pmd = 0;
 423 #else
 424         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 425 #endif
 426         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 427
 428         address = start_paddr;
 429         vaddr = start_vaddr;
 430         while (address < end_paddr) {
 431 #if PTRS_PER_PMD == 1
 432                 pmd = (pmd_t *)__pa(pg_dir);
 433 #else
 434                 pmd = (pmd_t *)pgd_address(*pg_dir);
 435
 436                 /*
 437                  * pmd is physical at this point
 438                  */
 439
 440                 if (!pmd) {
 441                         pmd = (pmd_t *) get_memblock(PAGE_SIZE << PMD_ORDER);
 442                         pmd = (pmd_t *) __pa(pmd);
 443                 }
 444
 445                 pgd_populate(NULL, pg_dir, __va(pmd));
 446 #endif
 447                 pg_dir++;
 448
 449                 /* now change pmd to kernel virtual addresses */
 450
 451                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 452                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
 453
 454                         /*
 455                          * pg_table is physical at this point
 456                          */
 457
 458                         pg_table = (pte_t *)pmd_address(*pmd);
 459                         if (!pg_table) {
 460                                 pg_table = (pte_t *) get_memblock(PAGE_SIZE);
 461                                 pg_table = (pte_t *) __pa(pg_table);
 462                         }
 463
 464                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 465
 466                         /* now change pg_table to kernel virtual addresses */
 467
 468                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 469                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
 470                                 pte_t pte;
 471
 472                                 if (force)
 473                                         pte =  __mk_pte(address, pgprot);
 474                                 else if (parisc_text_address(vaddr)) {
 475                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 476                                         if (address >= ro_start && address < kernel_end)
 477                                                 pte = pte_mkhuge(pte);
 478                                 }
 479                                 else
 480 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
 481                                 if (address >= ro_start && address < ro_end) {
 482                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 483                                         pte = pte_mkhuge(pte);
 484                                 } else
 485 #endif
 486                                 {
 487                                         pte = __mk_pte(address, pgprot);
 488                                         if (address >= ro_start && address < kernel_end)
 489                                                 pte = pte_mkhuge(pte);
 490                                 }
 491
 492                                 if (address >= end_paddr) {
 493                                         if (force)
 494                                                 break;
 495                                         else
 496                                                 pte_val(pte) = 0;
 497                                 }
 498
 499                                 set_pte(pg_table, pte);
 500
 501                                 address += PAGE_SIZE;
 502                                 vaddr += PAGE_SIZE;
 503                         }
 504                         start_pte = 0;
 505
 506                         if (address >= end_paddr)
 507                             break;
 508                 }
 509                 start_pmd = 0;
 510         }
 511 }
 512
 513 void free_initmem(void)
 514 {
 515         unsigned long init_begin = (unsigned long)__init_begin;
 516         unsigned long init_end = (unsigned long)__init_end;
 517
 518         /* The init text pages are marked R-X.  We have to
 519          * flush the icache and mark them RW-
 520          *
 521          * This is tricky, because map_pages is in the init section.
 522          * Do a dummy remap of the data section first (the data
 523          * section is already PAGE_KERNEL) to pull in the TLB entries
 524          * for map_kernel */
 525         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 526                   PAGE_KERNEL_RWX, 1);
 527         /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
 528          * map_pages */
 529         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 530                   PAGE_KERNEL, 1);
 531
 532         /* force the kernel to see the new TLB entries */
 533         __flush_tlb_range(0, init_begin, init_end);
 534
 535         /* finally dump all the instructions which were cached, since the
 536          * pages are no-longer executable */
 537         flush_icache_range(init_begin, init_end);
 538
 539         free_initmem_default(POISON_FREE_INITMEM);
 540
 541         /* set up a new led state on systems shipped LED State panel */
 542         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 543 }
 544
 545
 546 #ifdef CONFIG_DEBUG_RODATA
 547 void mark_rodata_ro(void)
 548 {
 549         /* rodata memory was already mapped with KERNEL_RO access rights by
 550            pagetable_init() and map_pages(). No need to do additional stuff here */
 551         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 552                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
 553 }
 554 #endif
 555
 556
 557 /*
 558  * Just an arbitrary offset to serve as a "hole" between mapping areas
 559  * (between top of physical memory and a potential pcxl dma mapping
 560  * area, and below the vmalloc mapping area).
 561  *
 562  * The current 32K value just means that there will be a 32K "hole"
 563  * between mapping areas. That means that  any out-of-bounds memory
 564  * accesses will hopefully be caught. The vmalloc() routines leaves
 565  * a hole of 4kB between each vmalloced area for the same reason.
 566  */
 567
 568  /* Leave room for gateway page expansion */
 569 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 570 #error KERNEL_MAP_START is in gateway reserved region
 571 #endif
 572 #define MAP_START (KERNEL_MAP_START)
 573
 574 #define VM_MAP_OFFSET  (32*1024)
 575 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 576                                      & ~(VM_MAP_OFFSET-1)))
 577
 578 void *parisc_vmalloc_start __read_mostly;
 579 EXPORT_SYMBOL(parisc_vmalloc_start);
 580
 581 #ifdef CONFIG_PA11
 582 unsigned long pcxl_dma_start __read_mostly;
 583 #endif
 584
 585 void __init mem_init(void)
 586 {
 587         /* Do sanity checks on IPC (compat) structures */
 588         BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
 589 #ifndef CONFIG_64BIT
 590         BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
 591         BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
 592         BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
 593 #endif
 594 #ifdef CONFIG_COMPAT
 595         BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
 596         BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
 597         BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
 598         BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
 599 #endif
 600
 601         /* Do sanity checks on page table constants */
 602         BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
 603         BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
 604         BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 605         BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 606                         > BITS_PER_LONG);
 607
 608         high_memory = __va((max_pfn << PAGE_SHIFT));
 609         set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
 610         free_all_bootmem();
 611
 612 #ifdef CONFIG_PA11
 613         if (hppa_dma_ops == &pcxl_dma_ops) {
 614                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 615                 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
 616                                                 + PCXL_DMA_MAP_SIZE);
 617         } else {
 618                 pcxl_dma_start = 0;
 619                 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 620         }
 621 #else
 622         parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 623 #endif
 624
 625         mem_init_print_info(NULL);
 626 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
 627         printk("virtual kernel memory layout:\n"
 628                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
 629                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
 630                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
 631                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
 632                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
 633
 634                (void*)VMALLOC_START, (void*)VMALLOC_END,
 635                (VMALLOC_END - VMALLOC_START) >> 20,
 636
 637                __va(0), high_memory,
 638                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 639
 640                __init_begin, __init_end,
 641                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
 642
 643                _etext, _edata,
 644                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
 645
 646                _text, _etext,
 647                ((unsigned long)_etext - (unsigned long)_text) >> 10);
 648 #endif
 649 }
 650
 651 unsigned long *empty_zero_page __read_mostly;
 652 EXPORT_SYMBOL(empty_zero_page);
 653
 654 void show_mem(unsigned int filter)
 655 {
 656         int total = 0,reserved = 0;
 657         pg_data_t *pgdat;
 658
 659         printk(KERN_INFO "Mem-info:\n");
 660         show_free_areas(filter);
 661
 662         for_each_online_pgdat(pgdat) {
 663                 unsigned long flags;
 664                 int zoneid;
 665
 666                 pgdat_resize_lock(pgdat, &flags);
 667                 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 668                         struct zone *zone = &pgdat->node_zones[zoneid];
 669                         if (!populated_zone(zone))
 670                                 continue;
 671
 672                         total += zone->present_pages;
 673                         reserved = zone->present_pages - zone->managed_pages;
 674                 }
 675                 pgdat_resize_unlock(pgdat, &flags);
 676         }
 677
 678         printk(KERN_INFO "%d pages of RAM\n", total);
 679         printk(KERN_INFO "%d reserved pages\n", reserved);
 680
 681 #ifdef CONFIG_DISCONTIGMEM
 682         {
 683                 struct zonelist *zl;
 684                 int i, j;
 685
 686                 for (i = 0; i < npmem_ranges; i++) {
 687                         zl = node_zonelist(i, 0);
 688                         for (j = 0; j < MAX_NR_ZONES; j++) {
 689                                 struct zoneref *z;
 690                                 struct zone *zone;
 691
 692                                 printk("Zone list for zone %d on node %d: ", j, i);
 693                                 for_each_zone_zonelist(zone, z, zl, j)
 694                                         printk("[%d/%s] ", zone_to_nid(zone),
 695                                                                 zone->name);
 696                                 printk("\n");
 697                         }
 698                 }
 699         }
 700 #endif
 701 }
 702
 703 /*
 704  * pagetable_init() sets up the page tables
 705  *
 706  * Note that gateway_init() places the Linux gateway page at page 0.
 707  * Since gateway pages cannot be dereferenced this has the desirable
 708  * side effect of trapping those pesky NULL-reference errors in the
 709  * kernel.
 710  */
 711 static void __init pagetable_init(void)
 712 {
 713         int range;
 714
 715         /* Map each physical memory range to its kernel vaddr */
 716
 717         for (range = 0; range < npmem_ranges; range++) {
 718                 unsigned long start_paddr;
 719                 unsigned long end_paddr;
 720                 unsigned long size;
 721
 722                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 723                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 724                 end_paddr = start_paddr + size;
 725
 726                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 727                           size, PAGE_KERNEL, 0);
 728         }
 729
 730 #ifdef CONFIG_BLK_DEV_INITRD
 731         if (initrd_end && initrd_end > mem_limit) {
 732                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 733                 map_pages(initrd_start, __pa(initrd_start),
 734                           initrd_end - initrd_start, PAGE_KERNEL, 0);
 735         }
 736 #endif
 737
 738         empty_zero_page = get_memblock(PAGE_SIZE);
 739 }
 740
 741 static void __init gateway_init(void)
 742 {
 743         unsigned long linux_gateway_page_addr;
 744         /* FIXME: This is 'const' in order to trick the compiler
 745            into not treating it as DP-relative data. */
 746         extern void * const linux_gateway_page;
 747
 748         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 749
 750         /*
 751          * Setup Linux Gateway page.
 752          *
 753          * The Linux gateway page will reside in kernel space (on virtual
 754          * page 0), so it doesn't need to be aliased into user space.
 755          */
 756
 757         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 758                   PAGE_SIZE, PAGE_GATEWAY, 1);
 759 }
 760
 761 void __init paging_init(void)
 762 {
 763         int i;
 764
 765         setup_bootmem();
 766         pagetable_init();
 767         gateway_init();
 768         flush_cache_all_local(); /* start with known state */
 769         flush_tlb_all_local(NULL);
 770
 771         for (i = 0; i < npmem_ranges; i++) {
 772                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
 773
 774                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
 775
 776 #ifdef CONFIG_DISCONTIGMEM
 777                 /* Need to initialize the pfnnid_map before we can initialize
 778                    the zone */
 779                 {
 780                     int j;
 781                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 782                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 783                          j++) {
 784                         pfnnid_map[j] = i;
 785                     }
 786                 }
 787 #endif
 788
 789                 free_area_init_node(i, zones_size,
 790                                 pmem_ranges[i].start_pfn, NULL);
 791         }
 792 }
 793
 794 #ifdef CONFIG_PA20
 795
 796 /*
 797  * Currently, all PA20 chips have 18 bit protection IDs, which is the
 798  * limiting factor (space ids are 32 bits).
 799  */
 800
 801 #define NR_SPACE_IDS 262144
 802
 803 #else
 804
 805 /*
 806  * Currently we have a one-to-one relationship between space IDs and
 807  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 808  * support 15 bit protection IDs, so that is the limiting factor.
 809  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
 810  * probably not worth the effort for a special case here.
 811  */
 812
 813 #define NR_SPACE_IDS 32768
 814
 815 #endif  /* !CONFIG_PA20 */
 816
 817 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 818 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 819
 820 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 821 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 822 static unsigned long space_id_index;
 823 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 824 static unsigned long dirty_space_ids = 0;
 825
 826 static DEFINE_SPINLOCK(sid_lock);
 827
 828 unsigned long alloc_sid(void)
 829 {
 830         unsigned long index;
 831
 832         spin_lock(&sid_lock);
 833
 834         if (free_space_ids == 0) {
 835                 if (dirty_space_ids != 0) {
 836                         spin_unlock(&sid_lock);
 837                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 838                         spin_lock(&sid_lock);
 839                 }
 840                 BUG_ON(free_space_ids == 0);
 841         }
 842
 843         free_space_ids--;
 844
 845         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 846         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 847         space_id_index = index;
 848
 849         spin_unlock(&sid_lock);
 850
 851         return index << SPACEID_SHIFT;
 852 }
 853
 854 void free_sid(unsigned long spaceid)
 855 {
 856         unsigned long index = spaceid >> SPACEID_SHIFT;
 857         unsigned long *dirty_space_offset;
 858
 859         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 860         index &= (BITS_PER_LONG - 1);
 861
 862         spin_lock(&sid_lock);
 863
 864         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
 865
 866         *dirty_space_offset |= (1L << index);
 867         dirty_space_ids++;
 868
 869         spin_unlock(&sid_lock);
 870 }
 871
 872
 873 #ifdef CONFIG_SMP
 874 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 875 {
 876         int i;
 877
 878         /* NOTE: sid_lock must be held upon entry */
 879
 880         *ndirtyptr = dirty_space_ids;
 881         if (dirty_space_ids != 0) {
 882             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 883                 dirty_array[i] = dirty_space_id[i];
 884                 dirty_space_id[i] = 0;
 885             }
 886             dirty_space_ids = 0;
 887         }
 888
 889         return;
 890 }
 891
 892 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 893 {
 894         int i;
 895
 896         /* NOTE: sid_lock must be held upon entry */
 897
 898         if (ndirty != 0) {
 899                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 900                         space_id[i] ^= dirty_array[i];
 901                 }
 902
 903                 free_space_ids += ndirty;
 904                 space_id_index = 0;
 905         }
 906 }
 907
 908 #else /* CONFIG_SMP */
 909
 910 static void recycle_sids(void)
 911 {
 912         int i;
 913
 914         /* NOTE: sid_lock must be held upon entry */
 915
 916         if (dirty_space_ids != 0) {
 917                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 918                         space_id[i] ^= dirty_space_id[i];
 919                         dirty_space_id[i] = 0;
 920                 }
 921
 922                 free_space_ids += dirty_space_ids;
 923                 dirty_space_ids = 0;
 924                 space_id_index = 0;
 925         }
 926 }
 927 #endif
 928
 929 /*
 930  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
 931  * purged, we can safely reuse the space ids that were released but
 932  * not flushed from the tlb.
 933  */
 934
 935 #ifdef CONFIG_SMP
 936
 937 static unsigned long recycle_ndirty;
 938 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
 939 static unsigned int recycle_inuse;
 940
 941 void flush_tlb_all(void)
 942 {
 943         int do_recycle;
 944
 945         __inc_irq_stat(irq_tlb_count);
 946         do_recycle = 0;
 947         spin_lock(&sid_lock);
 948         if (dirty_space_ids > RECYCLE_THRESHOLD) {
 949             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
 950             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
 951             recycle_inuse++;
 952             do_recycle++;
 953         }
 954         spin_unlock(&sid_lock);
 955         on_each_cpu(flush_tlb_all_local, NULL, 1);
 956         if (do_recycle) {
 957             spin_lock(&sid_lock);
 958             recycle_sids(recycle_ndirty,recycle_dirty_array);
 959             recycle_inuse = 0;
 960             spin_unlock(&sid_lock);
 961         }
 962 }
 963 #else
 964 void flush_tlb_all(void)
 965 {
 966         __inc_irq_stat(irq_tlb_count);
 967         spin_lock(&sid_lock);
 968         flush_tlb_all_local(NULL);
 969         recycle_sids();
 970         spin_unlock(&sid_lock);
 971 }
 972 #endif
 973
 974 #ifdef CONFIG_BLK_DEV_INITRD
 975 void free_initrd_mem(unsigned long start, unsigned long end)
 976 {
 977         free_reserved_area((void *)start, (void *)end, -1, "initrd");
 978 }
 979 #endif