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/gfp.h>
  18 #include <linux/delay.h>
  19 #include <linux/init.h>
  20 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
  21 #include <linux/initrd.h>
  22 #include <linux/swap.h>
  23 #include <linux/unistd.h>
  24 #include <linux/nodemask.h>     /* for node_online_map */
  25 #include <linux/pagemap.h>      /* for release_pages and page_cache_release */
  26 #include <linux/compat.h>
  27
  28 #include <asm/pgalloc.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/tlb.h>
  31 #include <asm/pdc_chassis.h>
  32 #include <asm/mmzone.h>
  33 #include <asm/sections.h>
  34 #include <asm/msgbuf.h>
  35
  36 extern int  data_start;
  37 extern void parisc_kernel_start(void);  /* Kernel entry point in head.S */
  38
  39 #if CONFIG_PGTABLE_LEVELS == 3
  40 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
  41  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
  42  * guarantee that global objects will be laid out in memory in the same order
  43  * as the order of declaration, so put these in different sections and use
  44  * the linker script to order them. */
  45 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
  46 #endif
  47
  48 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
  49 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
  50
  51 #ifdef CONFIG_DISCONTIGMEM
  52 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
  53 signed char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
  54 #endif
  55
  56 static struct resource data_resource = {
  57         .name   = "Kernel data",
  58         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  59 };
  60
  61 static struct resource code_resource = {
  62         .name   = "Kernel code",
  63         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  64 };
  65
  66 static struct resource pdcdata_resource = {
  67         .name   = "PDC data (Page Zero)",
  68         .start  = 0,
  69         .end    = 0x9ff,
  70         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  71 };
  72
  73 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
  74
  75 /* The following array is initialized from the firmware specific
  76  * information retrieved in kernel/inventory.c.
  77  */
  78
  79 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
  80 int npmem_ranges __read_mostly;
  81
  82 #ifdef CONFIG_64BIT
  83 #define MAX_MEM         (~0UL)
  84 #else /* !CONFIG_64BIT */
  85 #define MAX_MEM         (3584U*1024U*1024U)
  86 #endif /* !CONFIG_64BIT */
  87
  88 static unsigned long mem_limit __read_mostly = MAX_MEM;
  89
  90 static void __init mem_limit_func(void)
  91 {
  92         char *cp, *end;
  93         unsigned long limit;
  94
  95         /* We need this before __setup() functions are called */
  96
  97         limit = MAX_MEM;
  98         for (cp = boot_command_line; *cp; ) {
  99                 if (memcmp(cp, "mem=", 4) == 0) {
 100                         cp += 4;
 101                         limit = memparse(cp, &end);
 102                         if (end != cp)
 103                                 break;
 104                         cp = end;
 105                 } else {
 106                         while (*cp != ' ' && *cp)
 107                                 ++cp;
 108                         while (*cp == ' ')
 109                                 ++cp;
 110                 }
 111         }
 112
 113         if (limit < mem_limit)
 114                 mem_limit = limit;
 115 }
 116
 117 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 118
 119 static void __init setup_bootmem(void)
 120 {
 121         unsigned long bootmap_size;
 122         unsigned long mem_max;
 123         unsigned long bootmap_pages;
 124         unsigned long bootmap_start_pfn;
 125         unsigned long bootmap_pfn;
 126 #ifndef CONFIG_DISCONTIGMEM
 127         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 128         int npmem_holes;
 129 #endif
 130         int i, sysram_resource_count;
 131
 132         disable_sr_hashing(); /* Turn off space register hashing */
 133
 134         /*
 135          * Sort the ranges. Since the number of ranges is typically
 136          * small, and performance is not an issue here, just do
 137          * a simple insertion sort.
 138          */
 139
 140         for (i = 1; i < npmem_ranges; i++) {
 141                 int j;
 142
 143                 for (j = i; j > 0; j--) {
 144                         unsigned long tmp;
 145
 146                         if (pmem_ranges[j-1].start_pfn <
 147                             pmem_ranges[j].start_pfn) {
 148
 149                                 break;
 150                         }
 151                         tmp = pmem_ranges[j-1].start_pfn;
 152                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 153                         pmem_ranges[j].start_pfn = tmp;
 154                         tmp = pmem_ranges[j-1].pages;
 155                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 156                         pmem_ranges[j].pages = tmp;
 157                 }
 158         }
 159
 160 #ifndef CONFIG_DISCONTIGMEM
 161         /*
 162          * Throw out ranges that are too far apart (controlled by
 163          * MAX_GAP).
 164          */
 165
 166         for (i = 1; i < npmem_ranges; i++) {
 167                 if (pmem_ranges[i].start_pfn -
 168                         (pmem_ranges[i-1].start_pfn +
 169                          pmem_ranges[i-1].pages) > MAX_GAP) {
 170                         npmem_ranges = i;
 171                         printk("Large gap in memory detected (%ld pages). "
 172                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 173                                pmem_ranges[i].start_pfn -
 174                                (pmem_ranges[i-1].start_pfn +
 175                                 pmem_ranges[i-1].pages));
 176                         break;
 177                 }
 178         }
 179 #endif
 180
 181         if (npmem_ranges > 1) {
 182
 183                 /* Print the memory ranges */
 184
 185                 printk(KERN_INFO "Memory Ranges:\n");
 186
 187                 for (i = 0; i < npmem_ranges; i++) {
 188                         unsigned long start;
 189                         unsigned long size;
 190
 191                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
 192                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 193                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 194                                 i,start, start + (size - 1), size >> 20);
 195                 }
 196         }
 197
 198         sysram_resource_count = npmem_ranges;
 199         for (i = 0; i < sysram_resource_count; i++) {
 200                 struct resource *res = &sysram_resources[i];
 201                 res->name = "System RAM";
 202                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
 203                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
 204                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 205                 request_resource(&iomem_resource, res);
 206         }
 207
 208         /*
 209          * For 32 bit kernels we limit the amount of memory we can
 210          * support, in order to preserve enough kernel address space
 211          * for other purposes. For 64 bit kernels we don't normally
 212          * limit the memory, but this mechanism can be used to
 213          * artificially limit the amount of memory (and it is written
 214          * to work with multiple memory ranges).
 215          */
 216
 217         mem_limit_func();       /* check for "mem=" argument */
 218
 219         mem_max = 0;
 220         for (i = 0; i < npmem_ranges; i++) {
 221                 unsigned long rsize;
 222
 223                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 224                 if ((mem_max + rsize) > mem_limit) {
 225                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 226                         if (mem_max == mem_limit)
 227                                 npmem_ranges = i;
 228                         else {
 229                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 230                                                        - (mem_max >> PAGE_SHIFT);
 231                                 npmem_ranges = i + 1;
 232                                 mem_max = mem_limit;
 233                         }
 234                         break;
 235                 }
 236                 mem_max += rsize;
 237         }
 238
 239         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 240
 241 #ifndef CONFIG_DISCONTIGMEM
 242         /* Merge the ranges, keeping track of the holes */
 243
 244         {
 245                 unsigned long end_pfn;
 246                 unsigned long hole_pages;
 247
 248                 npmem_holes = 0;
 249                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 250                 for (i = 1; i < npmem_ranges; i++) {
 251
 252                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 253                         if (hole_pages) {
 254                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 255                                 pmem_holes[npmem_holes++].pages = hole_pages;
 256                                 end_pfn += hole_pages;
 257                         }
 258                         end_pfn += pmem_ranges[i].pages;
 259                 }
 260
 261                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 262                 npmem_ranges = 1;
 263         }
 264 #endif
 265
 266         bootmap_pages = 0;
 267         for (i = 0; i < npmem_ranges; i++)
 268                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
 269
 270         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
 271
 272 #ifdef CONFIG_DISCONTIGMEM
 273         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 274                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 275                 NODE_DATA(i)->bdata = &bootmem_node_data[i];
 276         }
 277         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 278
 279         for (i = 0; i < npmem_ranges; i++) {
 280                 node_set_state(i, N_NORMAL_MEMORY);
 281                 node_set_online(i);
 282         }
 283 #endif
 284
 285         /*
 286          * Initialize and free the full range of memory in each range.
 287          * Note that the only writing these routines do are to the bootmap,
 288          * and we've made sure to locate the bootmap properly so that they
 289          * won't be writing over anything important.
 290          */
 291
 292         bootmap_pfn = bootmap_start_pfn;
 293         max_pfn = 0;
 294         for (i = 0; i < npmem_ranges; i++) {
 295                 unsigned long start_pfn;
 296                 unsigned long npages;
 297
 298                 start_pfn = pmem_ranges[i].start_pfn;
 299                 npages = pmem_ranges[i].pages;
 300
 301                 bootmap_size = init_bootmem_node(NODE_DATA(i),
 302                                                 bootmap_pfn,
 303                                                 start_pfn,
 304                                                 (start_pfn + npages) );
 305                 free_bootmem_node(NODE_DATA(i),
 306                                   (start_pfn << PAGE_SHIFT),
 307                                   (npages << PAGE_SHIFT) );
 308                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 309                 if ((start_pfn + npages) > max_pfn)
 310                         max_pfn = start_pfn + npages;
 311         }
 312
 313         /* IOMMU is always used to access "high mem" on those boxes
 314          * that can support enough mem that a PCI device couldn't
 315          * directly DMA to any physical addresses.
 316          * ISA DMA support will need to revisit this.
 317          */
 318         max_low_pfn = max_pfn;
 319
 320         /* bootmap sizing messed up? */
 321         BUG_ON((bootmap_pfn - bootmap_start_pfn) != bootmap_pages);
 322
 323         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 324
 325 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 326
 327         reserve_bootmem_node(NODE_DATA(0), 0UL,
 328                         (unsigned long)(PAGE0->mem_free +
 329                                 PDC_CONSOLE_IO_IODC_SIZE), BOOTMEM_DEFAULT);
 330         reserve_bootmem_node(NODE_DATA(0), __pa(KERNEL_BINARY_TEXT_START),
 331                         (unsigned long)(_end - KERNEL_BINARY_TEXT_START),
 332                         BOOTMEM_DEFAULT);
 333         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
 334                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT),
 335                         BOOTMEM_DEFAULT);
 336
 337 #ifndef CONFIG_DISCONTIGMEM
 338
 339         /* reserve the holes */
 340
 341         for (i = 0; i < npmem_holes; i++) {
 342                 reserve_bootmem_node(NODE_DATA(0),
 343                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
 344                                 (pmem_holes[i].pages << PAGE_SHIFT),
 345                                 BOOTMEM_DEFAULT);
 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                         reserve_bootmem_node(NODE_DATA(0), __pa(initrd_start),
 364                                         initrd_reserve, BOOTMEM_DEFAULT);
 365                 }
 366         }
 367 #endif
 368
 369         data_resource.start =  virt_to_phys(&data_start);
 370         data_resource.end = virt_to_phys(_end) - 1;
 371         code_resource.start = virt_to_phys(_text);
 372         code_resource.end = virt_to_phys(&data_start)-1;
 373
 374         /* We don't know which region the kernel will be in, so try
 375          * all of them.
 376          */
 377         for (i = 0; i < sysram_resource_count; i++) {
 378                 struct resource *res = &sysram_resources[i];
 379                 request_resource(res, &code_resource);
 380                 request_resource(res, &data_resource);
 381         }
 382         request_resource(&sysram_resources[0], &pdcdata_resource);
 383 }
 384
 385 static int __init parisc_text_address(unsigned long vaddr)
 386 {
 387         static unsigned long head_ptr __initdata;
 388
 389         if (!head_ptr)
 390                 head_ptr = PAGE_MASK & (unsigned long)
 391                         dereference_function_descriptor(&parisc_kernel_start);
 392
 393         return core_kernel_text(vaddr) || vaddr == head_ptr;
 394 }
 395
 396 static void __init map_pages(unsigned long start_vaddr,
 397                              unsigned long start_paddr, unsigned long size,
 398                              pgprot_t pgprot, int force)
 399 {
 400         pgd_t *pg_dir;
 401         pmd_t *pmd;
 402         pte_t *pg_table;
 403         unsigned long end_paddr;
 404         unsigned long start_pmd;
 405         unsigned long start_pte;
 406         unsigned long tmp1;
 407         unsigned long tmp2;
 408         unsigned long address;
 409         unsigned long vaddr;
 410         unsigned long ro_start;
 411         unsigned long ro_end;
 412         unsigned long fv_addr;
 413         unsigned long gw_addr;
 414         extern const unsigned long fault_vector_20;
 415         extern void * const linux_gateway_page;
 416
 417         ro_start = __pa((unsigned long)_text);
 418         ro_end   = __pa((unsigned long)&data_start);
 419         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
 420         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
 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 *) alloc_bootmem_low_pages_node(NODE_DATA(0), 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 *)
 466                                         alloc_bootmem_low_pages_node(NODE_DATA(0), PAGE_SIZE);
 467                                 pg_table = (pte_t *) __pa(pg_table);
 468                         }
 469
 470                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 471
 472                         /* now change pg_table to kernel virtual addresses */
 473
 474                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 475                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
 476                                 pte_t pte;
 477
 478                                 /*
 479                                  * Map the fault vector writable so we can
 480                                  * write the HPMC checksum.
 481                                  */
 482                                 if (force)
 483                                         pte =  __mk_pte(address, pgprot);
 484                                 else if (parisc_text_address(vaddr) &&
 485                                          address != fv_addr)
 486                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 487                                 else
 488 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
 489                                 if (address >= ro_start && address < ro_end
 490                                                         && address != fv_addr
 491                                                         && address != gw_addr)
 492                                         pte = __mk_pte(address, PAGE_KERNEL_RO);
 493                                 else
 494 #endif
 495                                         pte = __mk_pte(address, pgprot);
 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         /* Attempt to catch anyone trying to execute code here
 540          * by filling the page with BRK insns.
 541          */
 542         memset((void *)init_begin, 0x00, init_end - init_begin);
 543         /* finally dump all the instructions which were cached, since the
 544          * pages are no-longer executable */
 545         flush_icache_range(init_begin, init_end);
 546
 547         free_initmem_default(-1);
 548
 549         /* set up a new led state on systems shipped LED State panel */
 550         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 551 }
 552
 553
 554 #ifdef CONFIG_DEBUG_RODATA
 555 void mark_rodata_ro(void)
 556 {
 557         /* rodata memory was already mapped with KERNEL_RO access rights by
 558            pagetable_init() and map_pages(). No need to do additional stuff here */
 559         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 560                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
 561 }
 562 #endif
 563
 564
 565 /*
 566  * Just an arbitrary offset to serve as a "hole" between mapping areas
 567  * (between top of physical memory and a potential pcxl dma mapping
 568  * area, and below the vmalloc mapping area).
 569  *
 570  * The current 32K value just means that there will be a 32K "hole"
 571  * between mapping areas. That means that  any out-of-bounds memory
 572  * accesses will hopefully be caught. The vmalloc() routines leaves
 573  * a hole of 4kB between each vmalloced area for the same reason.
 574  */
 575
 576  /* Leave room for gateway page expansion */
 577 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 578 #error KERNEL_MAP_START is in gateway reserved region
 579 #endif
 580 #define MAP_START (KERNEL_MAP_START)
 581
 582 #define VM_MAP_OFFSET  (32*1024)
 583 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 584                                      & ~(VM_MAP_OFFSET-1)))
 585
 586 void *parisc_vmalloc_start __read_mostly;
 587 EXPORT_SYMBOL(parisc_vmalloc_start);
 588
 589 #ifdef CONFIG_PA11
 590 unsigned long pcxl_dma_start __read_mostly;
 591 #endif
 592
 593 void __init mem_init(void)
 594 {
 595         /* Do sanity checks on IPC (compat) structures */
 596         BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
 597 #ifndef CONFIG_64BIT
 598         BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
 599         BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
 600         BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
 601 #endif
 602 #ifdef CONFIG_COMPAT
 603         BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
 604         BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
 605         BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
 606         BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
 607 #endif
 608
 609         /* Do sanity checks on page table constants */
 610         BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
 611         BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
 612         BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 613         BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 614                         > BITS_PER_LONG);
 615
 616         high_memory = __va((max_pfn << PAGE_SHIFT));
 617         set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
 618         free_all_bootmem();
 619
 620 #ifdef CONFIG_PA11
 621         if (hppa_dma_ops == &pcxl_dma_ops) {
 622                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 623                 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
 624                                                 + PCXL_DMA_MAP_SIZE);
 625         } else {
 626                 pcxl_dma_start = 0;
 627                 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 628         }
 629 #else
 630         parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 631 #endif
 632
 633         mem_init_print_info(NULL);
 634 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
 635         printk("virtual kernel memory layout:\n"
 636                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
 637                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
 638                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
 639                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
 640                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
 641
 642                (void*)VMALLOC_START, (void*)VMALLOC_END,
 643                (VMALLOC_END - VMALLOC_START) >> 20,
 644
 645                __va(0), high_memory,
 646                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 647
 648                __init_begin, __init_end,
 649                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
 650
 651                _etext, _edata,
 652                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
 653
 654                _text, _etext,
 655                ((unsigned long)_etext - (unsigned long)_text) >> 10);
 656 #endif
 657 }
 658
 659 unsigned long *empty_zero_page __read_mostly;
 660 EXPORT_SYMBOL(empty_zero_page);
 661
 662 void show_mem(unsigned int filter)
 663 {
 664         int total = 0,reserved = 0;
 665         pg_data_t *pgdat;
 666
 667         printk(KERN_INFO "Mem-info:\n");
 668         show_free_areas(filter);
 669
 670         for_each_online_pgdat(pgdat) {
 671                 unsigned long flags;
 672                 int zoneid;
 673
 674                 pgdat_resize_lock(pgdat, &flags);
 675                 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 676                         struct zone *zone = &pgdat->node_zones[zoneid];
 677                         if (!populated_zone(zone))
 678                                 continue;
 679
 680                         total += zone->present_pages;
 681                         reserved = zone->present_pages - zone->managed_pages;
 682                 }
 683                 pgdat_resize_unlock(pgdat, &flags);
 684         }
 685
 686         printk(KERN_INFO "%d pages of RAM\n", total);
 687         printk(KERN_INFO "%d reserved pages\n", reserved);
 688
 689 #ifdef CONFIG_DISCONTIGMEM
 690         {
 691                 struct zonelist *zl;
 692                 int i, j;
 693
 694                 for (i = 0; i < npmem_ranges; i++) {
 695                         zl = node_zonelist(i, 0);
 696                         for (j = 0; j < MAX_NR_ZONES; j++) {
 697                                 struct zoneref *z;
 698                                 struct zone *zone;
 699
 700                                 printk("Zone list for zone %d on node %d: ", j, i);
 701                                 for_each_zone_zonelist(zone, z, zl, j)
 702                                         printk("[%d/%s] ", zone_to_nid(zone),
 703                                                                 zone->name);
 704                                 printk("\n");
 705                         }
 706                 }
 707         }
 708 #endif
 709 }
 710
 711 /*
 712  * pagetable_init() sets up the page tables
 713  *
 714  * Note that gateway_init() places the Linux gateway page at page 0.
 715  * Since gateway pages cannot be dereferenced this has the desirable
 716  * side effect of trapping those pesky NULL-reference errors in the
 717  * kernel.
 718  */
 719 static void __init pagetable_init(void)
 720 {
 721         int range;
 722
 723         /* Map each physical memory range to its kernel vaddr */
 724
 725         for (range = 0; range < npmem_ranges; range++) {
 726                 unsigned long start_paddr;
 727                 unsigned long end_paddr;
 728                 unsigned long size;
 729
 730                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 731                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
 732                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 733
 734                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 735                           size, PAGE_KERNEL, 0);
 736         }
 737
 738 #ifdef CONFIG_BLK_DEV_INITRD
 739         if (initrd_end && initrd_end > mem_limit) {
 740                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 741                 map_pages(initrd_start, __pa(initrd_start),
 742                           initrd_end - initrd_start, PAGE_KERNEL, 0);
 743         }
 744 #endif
 745
 746         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
 747 }
 748
 749 static void __init gateway_init(void)
 750 {
 751         unsigned long linux_gateway_page_addr;
 752         /* FIXME: This is 'const' in order to trick the compiler
 753            into not treating it as DP-relative data. */
 754         extern void * const linux_gateway_page;
 755
 756         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 757
 758         /*
 759          * Setup Linux Gateway page.
 760          *
 761          * The Linux gateway page will reside in kernel space (on virtual
 762          * page 0), so it doesn't need to be aliased into user space.
 763          */
 764
 765         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 766                   PAGE_SIZE, PAGE_GATEWAY, 1);
 767 }
 768
 769 void __init paging_init(void)
 770 {
 771         int i;
 772
 773         setup_bootmem();
 774         pagetable_init();
 775         gateway_init();
 776         flush_cache_all_local(); /* start with known state */
 777         flush_tlb_all_local(NULL);
 778
 779         for (i = 0; i < npmem_ranges; i++) {
 780                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
 781
 782                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
 783
 784 #ifdef CONFIG_DISCONTIGMEM
 785                 /* Need to initialize the pfnnid_map before we can initialize
 786                    the zone */
 787                 {
 788                     int j;
 789                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 790                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 791                          j++) {
 792                         pfnnid_map[j] = i;
 793                     }
 794                 }
 795 #endif
 796
 797                 free_area_init_node(i, zones_size,
 798                                 pmem_ranges[i].start_pfn, NULL);
 799         }
 800 }
 801
 802 #ifdef CONFIG_PA20
 803
 804 /*
 805  * Currently, all PA20 chips have 18 bit protection IDs, which is the
 806  * limiting factor (space ids are 32 bits).
 807  */
 808
 809 #define NR_SPACE_IDS 262144
 810
 811 #else
 812
 813 /*
 814  * Currently we have a one-to-one relationship between space IDs and
 815  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 816  * support 15 bit protection IDs, so that is the limiting factor.
 817  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
 818  * probably not worth the effort for a special case here.
 819  */
 820
 821 #define NR_SPACE_IDS 32768
 822
 823 #endif  /* !CONFIG_PA20 */
 824
 825 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 826 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 827
 828 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 829 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 830 static unsigned long space_id_index;
 831 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 832 static unsigned long dirty_space_ids = 0;
 833
 834 static DEFINE_SPINLOCK(sid_lock);
 835
 836 unsigned long alloc_sid(void)
 837 {
 838         unsigned long index;
 839
 840         spin_lock(&sid_lock);
 841
 842         if (free_space_ids == 0) {
 843                 if (dirty_space_ids != 0) {
 844                         spin_unlock(&sid_lock);
 845                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 846                         spin_lock(&sid_lock);
 847                 }
 848                 BUG_ON(free_space_ids == 0);
 849         }
 850
 851         free_space_ids--;
 852
 853         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 854         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 855         space_id_index = index;
 856
 857         spin_unlock(&sid_lock);
 858
 859         return index << SPACEID_SHIFT;
 860 }
 861
 862 void free_sid(unsigned long spaceid)
 863 {
 864         unsigned long index = spaceid >> SPACEID_SHIFT;
 865         unsigned long *dirty_space_offset;
 866
 867         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 868         index &= (BITS_PER_LONG - 1);
 869
 870         spin_lock(&sid_lock);
 871
 872         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
 873
 874         *dirty_space_offset |= (1L << index);
 875         dirty_space_ids++;
 876
 877         spin_unlock(&sid_lock);
 878 }
 879
 880
 881 #ifdef CONFIG_SMP
 882 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 883 {
 884         int i;
 885
 886         /* NOTE: sid_lock must be held upon entry */
 887
 888         *ndirtyptr = dirty_space_ids;
 889         if (dirty_space_ids != 0) {
 890             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 891                 dirty_array[i] = dirty_space_id[i];
 892                 dirty_space_id[i] = 0;
 893             }
 894             dirty_space_ids = 0;
 895         }
 896
 897         return;
 898 }
 899
 900 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 901 {
 902         int i;
 903
 904         /* NOTE: sid_lock must be held upon entry */
 905
 906         if (ndirty != 0) {
 907                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 908                         space_id[i] ^= dirty_array[i];
 909                 }
 910
 911                 free_space_ids += ndirty;
 912                 space_id_index = 0;
 913         }
 914 }
 915
 916 #else /* CONFIG_SMP */
 917
 918 static void recycle_sids(void)
 919 {
 920         int i;
 921
 922         /* NOTE: sid_lock must be held upon entry */
 923
 924         if (dirty_space_ids != 0) {
 925                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 926                         space_id[i] ^= dirty_space_id[i];
 927                         dirty_space_id[i] = 0;
 928                 }
 929
 930                 free_space_ids += dirty_space_ids;
 931                 dirty_space_ids = 0;
 932                 space_id_index = 0;
 933         }
 934 }
 935 #endif
 936
 937 /*
 938  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
 939  * purged, we can safely reuse the space ids that were released but
 940  * not flushed from the tlb.
 941  */
 942
 943 #ifdef CONFIG_SMP
 944
 945 static unsigned long recycle_ndirty;
 946 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
 947 static unsigned int recycle_inuse;
 948
 949 void flush_tlb_all(void)
 950 {
 951         int do_recycle;
 952
 953         __inc_irq_stat(irq_tlb_count);
 954         do_recycle = 0;
 955         spin_lock(&sid_lock);
 956         if (dirty_space_ids > RECYCLE_THRESHOLD) {
 957             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
 958             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
 959             recycle_inuse++;
 960             do_recycle++;
 961         }
 962         spin_unlock(&sid_lock);
 963         on_each_cpu(flush_tlb_all_local, NULL, 1);
 964         if (do_recycle) {
 965             spin_lock(&sid_lock);
 966             recycle_sids(recycle_ndirty,recycle_dirty_array);
 967             recycle_inuse = 0;
 968             spin_unlock(&sid_lock);
 969         }
 970 }
 971 #else
 972 void flush_tlb_all(void)
 973 {
 974         __inc_irq_stat(irq_tlb_count);
 975         spin_lock(&sid_lock);
 976         flush_tlb_all_local(NULL);
 977         recycle_sids();
 978         spin_unlock(&sid_lock);
 979 }
 980 #endif
 981
 982 #ifdef CONFIG_BLK_DEV_INITRD
 983 void free_initrd_mem(unsigned long start, unsigned long end)
 984 {
 985         free_reserved_area((void *)start, (void *)end, -1, "initrd");
 986 }
 987 #endif