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
  27 #include <asm/pgalloc.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/tlb.h>
  30 #include <asm/pdc_chassis.h>
  31 #include <asm/mmzone.h>
  32 #include <asm/sections.h>
  33
  34 extern int  data_start;
  35 extern void parisc_kernel_start(void);  /* Kernel entry point in head.S */
  36
  37 #if PT_NLEVELS == 3
  38 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
  39  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
  40  * guarantee that global objects will be laid out in memory in the same order
  41  * as the order of declaration, so put these in different sections and use
  42  * the linker script to order them. */
  43 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
  44 #endif
  45
  46 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
  47 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
  48
  49 #ifdef CONFIG_DISCONTIGMEM
  50 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
  51 signed char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
  52 #endif
  53
  54 static struct resource data_resource = {
  55         .name   = "Kernel data",
  56         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  57 };
  58
  59 static struct resource code_resource = {
  60         .name   = "Kernel code",
  61         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  62 };
  63
  64 static struct resource pdcdata_resource = {
  65         .name   = "PDC data (Page Zero)",
  66         .start  = 0,
  67         .end    = 0x9ff,
  68         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
  69 };
  70
  71 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
  72
  73 /* The following array is initialized from the firmware specific
  74  * information retrieved in kernel/inventory.c.
  75  */
  76
  77 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
  78 int npmem_ranges __read_mostly;
  79
  80 #ifdef CONFIG_64BIT
  81 #define MAX_MEM         (~0UL)
  82 #else /* !CONFIG_64BIT */
  83 #define MAX_MEM         (3584U*1024U*1024U)
  84 #endif /* !CONFIG_64BIT */
  85
  86 static unsigned long mem_limit __read_mostly = MAX_MEM;
  87
  88 static void __init mem_limit_func(void)
  89 {
  90         char *cp, *end;
  91         unsigned long limit;
  92
  93         /* We need this before __setup() functions are called */
  94
  95         limit = MAX_MEM;
  96         for (cp = boot_command_line; *cp; ) {
  97                 if (memcmp(cp, "mem=", 4) == 0) {
  98                         cp += 4;
  99                         limit = memparse(cp, &end);
 100                         if (end != cp)
 101                                 break;
 102                         cp = end;
 103                 } else {
 104                         while (*cp != ' ' && *cp)
 105                                 ++cp;
 106                         while (*cp == ' ')
 107                                 ++cp;
 108                 }
 109         }
 110
 111         if (limit < mem_limit)
 112                 mem_limit = limit;
 113 }
 114
 115 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
 116
 117 static void __init setup_bootmem(void)
 118 {
 119         unsigned long bootmap_size;
 120         unsigned long mem_max;
 121         unsigned long bootmap_pages;
 122         unsigned long bootmap_start_pfn;
 123         unsigned long bootmap_pfn;
 124 #ifndef CONFIG_DISCONTIGMEM
 125         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
 126         int npmem_holes;
 127 #endif
 128         int i, sysram_resource_count;
 129
 130         disable_sr_hashing(); /* Turn off space register hashing */
 131
 132         /*
 133          * Sort the ranges. Since the number of ranges is typically
 134          * small, and performance is not an issue here, just do
 135          * a simple insertion sort.
 136          */
 137
 138         for (i = 1; i < npmem_ranges; i++) {
 139                 int j;
 140
 141                 for (j = i; j > 0; j--) {
 142                         unsigned long tmp;
 143
 144                         if (pmem_ranges[j-1].start_pfn <
 145                             pmem_ranges[j].start_pfn) {
 146
 147                                 break;
 148                         }
 149                         tmp = pmem_ranges[j-1].start_pfn;
 150                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
 151                         pmem_ranges[j].start_pfn = tmp;
 152                         tmp = pmem_ranges[j-1].pages;
 153                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
 154                         pmem_ranges[j].pages = tmp;
 155                 }
 156         }
 157
 158 #ifndef CONFIG_DISCONTIGMEM
 159         /*
 160          * Throw out ranges that are too far apart (controlled by
 161          * MAX_GAP).
 162          */
 163
 164         for (i = 1; i < npmem_ranges; i++) {
 165                 if (pmem_ranges[i].start_pfn -
 166                         (pmem_ranges[i-1].start_pfn +
 167                          pmem_ranges[i-1].pages) > MAX_GAP) {
 168                         npmem_ranges = i;
 169                         printk("Large gap in memory detected (%ld pages). "
 170                                "Consider turning on CONFIG_DISCONTIGMEM\n",
 171                                pmem_ranges[i].start_pfn -
 172                                (pmem_ranges[i-1].start_pfn +
 173                                 pmem_ranges[i-1].pages));
 174                         break;
 175                 }
 176         }
 177 #endif
 178
 179         if (npmem_ranges > 1) {
 180
 181                 /* Print the memory ranges */
 182
 183                 printk(KERN_INFO "Memory Ranges:\n");
 184
 185                 for (i = 0; i < npmem_ranges; i++) {
 186                         unsigned long start;
 187                         unsigned long size;
 188
 189                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
 190                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
 191                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
 192                                 i,start, start + (size - 1), size >> 20);
 193                 }
 194         }
 195
 196         sysram_resource_count = npmem_ranges;
 197         for (i = 0; i < sysram_resource_count; i++) {
 198                 struct resource *res = &sysram_resources[i];
 199                 res->name = "System RAM";
 200                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
 201                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
 202                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 203                 request_resource(&iomem_resource, res);
 204         }
 205
 206         /*
 207          * For 32 bit kernels we limit the amount of memory we can
 208          * support, in order to preserve enough kernel address space
 209          * for other purposes. For 64 bit kernels we don't normally
 210          * limit the memory, but this mechanism can be used to
 211          * artificially limit the amount of memory (and it is written
 212          * to work with multiple memory ranges).
 213          */
 214
 215         mem_limit_func();       /* check for "mem=" argument */
 216
 217         mem_max = 0;
 218         for (i = 0; i < npmem_ranges; i++) {
 219                 unsigned long rsize;
 220
 221                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
 222                 if ((mem_max + rsize) > mem_limit) {
 223                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
 224                         if (mem_max == mem_limit)
 225                                 npmem_ranges = i;
 226                         else {
 227                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
 228                                                        - (mem_max >> PAGE_SHIFT);
 229                                 npmem_ranges = i + 1;
 230                                 mem_max = mem_limit;
 231                         }
 232                         break;
 233                 }
 234                 mem_max += rsize;
 235         }
 236
 237         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
 238
 239 #ifndef CONFIG_DISCONTIGMEM
 240         /* Merge the ranges, keeping track of the holes */
 241
 242         {
 243                 unsigned long end_pfn;
 244                 unsigned long hole_pages;
 245
 246                 npmem_holes = 0;
 247                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
 248                 for (i = 1; i < npmem_ranges; i++) {
 249
 250                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
 251                         if (hole_pages) {
 252                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
 253                                 pmem_holes[npmem_holes++].pages = hole_pages;
 254                                 end_pfn += hole_pages;
 255                         }
 256                         end_pfn += pmem_ranges[i].pages;
 257                 }
 258
 259                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
 260                 npmem_ranges = 1;
 261         }
 262 #endif
 263
 264         bootmap_pages = 0;
 265         for (i = 0; i < npmem_ranges; i++)
 266                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
 267
 268         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
 269
 270 #ifdef CONFIG_DISCONTIGMEM
 271         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
 272                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
 273                 NODE_DATA(i)->bdata = &bootmem_node_data[i];
 274         }
 275         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
 276
 277         for (i = 0; i < npmem_ranges; i++) {
 278                 node_set_state(i, N_NORMAL_MEMORY);
 279                 node_set_online(i);
 280         }
 281 #endif
 282
 283         /*
 284          * Initialize and free the full range of memory in each range.
 285          * Note that the only writing these routines do are to the bootmap,
 286          * and we've made sure to locate the bootmap properly so that they
 287          * won't be writing over anything important.
 288          */
 289
 290         bootmap_pfn = bootmap_start_pfn;
 291         max_pfn = 0;
 292         for (i = 0; i < npmem_ranges; i++) {
 293                 unsigned long start_pfn;
 294                 unsigned long npages;
 295
 296                 start_pfn = pmem_ranges[i].start_pfn;
 297                 npages = pmem_ranges[i].pages;
 298
 299                 bootmap_size = init_bootmem_node(NODE_DATA(i),
 300                                                 bootmap_pfn,
 301                                                 start_pfn,
 302                                                 (start_pfn + npages) );
 303                 free_bootmem_node(NODE_DATA(i),
 304                                   (start_pfn << PAGE_SHIFT),
 305                                   (npages << PAGE_SHIFT) );
 306                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 307                 if ((start_pfn + npages) > max_pfn)
 308                         max_pfn = start_pfn + npages;
 309         }
 310
 311         /* IOMMU is always used to access "high mem" on those boxes
 312          * that can support enough mem that a PCI device couldn't
 313          * directly DMA to any physical addresses.
 314          * ISA DMA support will need to revisit this.
 315          */
 316         max_low_pfn = max_pfn;
 317
 318         /* bootmap sizing messed up? */
 319         BUG_ON((bootmap_pfn - bootmap_start_pfn) != bootmap_pages);
 320
 321         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
 322
 323 #define PDC_CONSOLE_IO_IODC_SIZE 32768
 324
 325         reserve_bootmem_node(NODE_DATA(0), 0UL,
 326                         (unsigned long)(PAGE0->mem_free +
 327                                 PDC_CONSOLE_IO_IODC_SIZE), BOOTMEM_DEFAULT);
 328         reserve_bootmem_node(NODE_DATA(0), __pa(KERNEL_BINARY_TEXT_START),
 329                         (unsigned long)(_end - KERNEL_BINARY_TEXT_START),
 330                         BOOTMEM_DEFAULT);
 331         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
 332                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT),
 333                         BOOTMEM_DEFAULT);
 334
 335 #ifndef CONFIG_DISCONTIGMEM
 336
 337         /* reserve the holes */
 338
 339         for (i = 0; i < npmem_holes; i++) {
 340                 reserve_bootmem_node(NODE_DATA(0),
 341                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
 342                                 (pmem_holes[i].pages << PAGE_SHIFT),
 343                                 BOOTMEM_DEFAULT);
 344         }
 345 #endif
 346
 347 #ifdef CONFIG_BLK_DEV_INITRD
 348         if (initrd_start) {
 349                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
 350                 if (__pa(initrd_start) < mem_max) {
 351                         unsigned long initrd_reserve;
 352
 353                         if (__pa(initrd_end) > mem_max) {
 354                                 initrd_reserve = mem_max - __pa(initrd_start);
 355                         } else {
 356                                 initrd_reserve = initrd_end - initrd_start;
 357                         }
 358                         initrd_below_start_ok = 1;
 359                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
 360
 361                         reserve_bootmem_node(NODE_DATA(0), __pa(initrd_start),
 362                                         initrd_reserve, BOOTMEM_DEFAULT);
 363                 }
 364         }
 365 #endif
 366
 367         data_resource.start =  virt_to_phys(&data_start);
 368         data_resource.end = virt_to_phys(_end) - 1;
 369         code_resource.start = virt_to_phys(_text);
 370         code_resource.end = virt_to_phys(&data_start)-1;
 371
 372         /* We don't know which region the kernel will be in, so try
 373          * all of them.
 374          */
 375         for (i = 0; i < sysram_resource_count; i++) {
 376                 struct resource *res = &sysram_resources[i];
 377                 request_resource(res, &code_resource);
 378                 request_resource(res, &data_resource);
 379         }
 380         request_resource(&sysram_resources[0], &pdcdata_resource);
 381 }
 382
 383 static int __init parisc_text_address(unsigned long vaddr)
 384 {
 385         static unsigned long head_ptr __initdata;
 386
 387         if (!head_ptr)
 388                 head_ptr = PAGE_MASK & (unsigned long)
 389                         dereference_function_descriptor(&parisc_kernel_start);
 390
 391         return core_kernel_text(vaddr) || vaddr == head_ptr;
 392 }
 393
 394 static void __init map_pages(unsigned long start_vaddr,
 395                              unsigned long start_paddr, unsigned long size,
 396                              pgprot_t pgprot, int force)
 397 {
 398         pgd_t *pg_dir;
 399         pmd_t *pmd;
 400         pte_t *pg_table;
 401         unsigned long end_paddr;
 402         unsigned long start_pmd;
 403         unsigned long start_pte;
 404         unsigned long tmp1;
 405         unsigned long tmp2;
 406         unsigned long address;
 407         unsigned long vaddr;
 408         unsigned long ro_start;
 409         unsigned long ro_end;
 410         unsigned long fv_addr;
 411         unsigned long gw_addr;
 412         extern const unsigned long fault_vector_20;
 413         extern void * const linux_gateway_page;
 414
 415         ro_start = __pa((unsigned long)_text);
 416         ro_end   = __pa((unsigned long)&data_start);
 417         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
 418         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
 419
 420         end_paddr = start_paddr + size;
 421
 422         pg_dir = pgd_offset_k(start_vaddr);
 423
 424 #if PTRS_PER_PMD == 1
 425         start_pmd = 0;
 426 #else
 427         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 428 #endif
 429         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 430
 431         address = start_paddr;
 432         vaddr = start_vaddr;
 433         while (address < end_paddr) {
 434 #if PTRS_PER_PMD == 1
 435                 pmd = (pmd_t *)__pa(pg_dir);
 436 #else
 437                 pmd = (pmd_t *)pgd_address(*pg_dir);
 438
 439                 /*
 440                  * pmd is physical at this point
 441                  */
 442
 443                 if (!pmd) {
 444                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0), PAGE_SIZE << PMD_ORDER);
 445                         pmd = (pmd_t *) __pa(pmd);
 446                 }
 447
 448                 pgd_populate(NULL, pg_dir, __va(pmd));
 449 #endif
 450                 pg_dir++;
 451
 452                 /* now change pmd to kernel virtual addresses */
 453
 454                 pmd = (pmd_t *)__va(pmd) + start_pmd;
 455                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
 456
 457                         /*
 458                          * pg_table is physical at this point
 459                          */
 460
 461                         pg_table = (pte_t *)pmd_address(*pmd);
 462                         if (!pg_table) {
 463                                 pg_table = (pte_t *)
 464                                         alloc_bootmem_low_pages_node(NODE_DATA(0), PAGE_SIZE);
 465                                 pg_table = (pte_t *) __pa(pg_table);
 466                         }
 467
 468                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
 469
 470                         /* now change pg_table to kernel virtual addresses */
 471
 472                         pg_table = (pte_t *) __va(pg_table) + start_pte;
 473                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
 474                                 pte_t pte;
 475
 476                                 /*
 477                                  * Map the fault vector writable so we can
 478                                  * write the HPMC checksum.
 479                                  */
 480                                 if (force)
 481                                         pte =  __mk_pte(address, pgprot);
 482                                 else if (parisc_text_address(vaddr) &&
 483                                          address != fv_addr)
 484                                         pte = __mk_pte(address, PAGE_KERNEL_EXEC);
 485                                 else
 486 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
 487                                 if (address >= ro_start && address < ro_end
 488                                                         && address != fv_addr
 489                                                         && address != gw_addr)
 490                                         pte = __mk_pte(address, PAGE_KERNEL_RO);
 491                                 else
 492 #endif
 493                                         pte = __mk_pte(address, pgprot);
 494
 495                                 if (address >= end_paddr) {
 496                                         if (force)
 497                                                 break;
 498                                         else
 499                                                 pte_val(pte) = 0;
 500                                 }
 501
 502                                 set_pte(pg_table, pte);
 503
 504                                 address += PAGE_SIZE;
 505                                 vaddr += PAGE_SIZE;
 506                         }
 507                         start_pte = 0;
 508
 509                         if (address >= end_paddr)
 510                             break;
 511                 }
 512                 start_pmd = 0;
 513         }
 514 }
 515
 516 void free_initmem(void)
 517 {
 518         unsigned long init_begin = (unsigned long)__init_begin;
 519         unsigned long init_end = (unsigned long)__init_end;
 520
 521         /* The init text pages are marked R-X.  We have to
 522          * flush the icache and mark them RW-
 523          *
 524          * This is tricky, because map_pages is in the init section.
 525          * Do a dummy remap of the data section first (the data
 526          * section is already PAGE_KERNEL) to pull in the TLB entries
 527          * for map_kernel */
 528         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 529                   PAGE_KERNEL_RWX, 1);
 530         /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
 531          * map_pages */
 532         map_pages(init_begin, __pa(init_begin), init_end - init_begin,
 533                   PAGE_KERNEL, 1);
 534
 535         /* force the kernel to see the new TLB entries */
 536         __flush_tlb_range(0, init_begin, init_end);
 537         /* Attempt to catch anyone trying to execute code here
 538          * by filling the page with BRK insns.
 539          */
 540         memset((void *)init_begin, 0x00, init_end - init_begin);
 541         /* finally dump all the instructions which were cached, since the
 542          * pages are no-longer executable */
 543         flush_icache_range(init_begin, init_end);
 544
 545         free_initmem_default(-1);
 546
 547         /* set up a new led state on systems shipped LED State panel */
 548         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
 549 }
 550
 551
 552 #ifdef CONFIG_DEBUG_RODATA
 553 void mark_rodata_ro(void)
 554 {
 555         /* rodata memory was already mapped with KERNEL_RO access rights by
 556            pagetable_init() and map_pages(). No need to do additional stuff here */
 557         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 558                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
 559 }
 560 #endif
 561
 562
 563 /*
 564  * Just an arbitrary offset to serve as a "hole" between mapping areas
 565  * (between top of physical memory and a potential pcxl dma mapping
 566  * area, and below the vmalloc mapping area).
 567  *
 568  * The current 32K value just means that there will be a 32K "hole"
 569  * between mapping areas. That means that  any out-of-bounds memory
 570  * accesses will hopefully be caught. The vmalloc() routines leaves
 571  * a hole of 4kB between each vmalloced area for the same reason.
 572  */
 573
 574  /* Leave room for gateway page expansion */
 575 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
 576 #error KERNEL_MAP_START is in gateway reserved region
 577 #endif
 578 #define MAP_START (KERNEL_MAP_START)
 579
 580 #define VM_MAP_OFFSET  (32*1024)
 581 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
 582                                      & ~(VM_MAP_OFFSET-1)))
 583
 584 void *parisc_vmalloc_start __read_mostly;
 585 EXPORT_SYMBOL(parisc_vmalloc_start);
 586
 587 #ifdef CONFIG_PA11
 588 unsigned long pcxl_dma_start __read_mostly;
 589 #endif
 590
 591 void __init mem_init(void)
 592 {
 593         /* Do sanity checks on page table constants */
 594         BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
 595         BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
 596         BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 597         BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 598                         > BITS_PER_LONG);
 599
 600         high_memory = __va((max_pfn << PAGE_SHIFT));
 601         set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
 602         free_all_bootmem();
 603
 604 #ifdef CONFIG_PA11
 605         if (hppa_dma_ops == &pcxl_dma_ops) {
 606                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
 607                 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
 608                                                 + PCXL_DMA_MAP_SIZE);
 609         } else {
 610                 pcxl_dma_start = 0;
 611                 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 612         }
 613 #else
 614         parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
 615 #endif
 616
 617         mem_init_print_info(NULL);
 618 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
 619         printk("virtual kernel memory layout:\n"
 620                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
 621                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
 622                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
 623                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
 624                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
 625
 626                (void*)VMALLOC_START, (void*)VMALLOC_END,
 627                (VMALLOC_END - VMALLOC_START) >> 20,
 628
 629                __va(0), high_memory,
 630                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 631
 632                __init_begin, __init_end,
 633                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
 634
 635                _etext, _edata,
 636                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
 637
 638                _text, _etext,
 639                ((unsigned long)_etext - (unsigned long)_text) >> 10);
 640 #endif
 641 }
 642
 643 unsigned long *empty_zero_page __read_mostly;
 644 EXPORT_SYMBOL(empty_zero_page);
 645
 646 void show_mem(unsigned int filter)
 647 {
 648         int i,free = 0,total = 0,reserved = 0;
 649         int shared = 0, cached = 0;
 650
 651         printk(KERN_INFO "Mem-info:\n");
 652         show_free_areas(filter);
 653         if (filter & SHOW_MEM_FILTER_PAGE_COUNT)
 654                 return;
 655 #ifndef CONFIG_DISCONTIGMEM
 656         i = max_mapnr;
 657         while (i-- > 0) {
 658                 total++;
 659                 if (PageReserved(mem_map+i))
 660                         reserved++;
 661                 else if (PageSwapCache(mem_map+i))
 662                         cached++;
 663                 else if (!page_count(&mem_map[i]))
 664                         free++;
 665                 else
 666                         shared += page_count(&mem_map[i]) - 1;
 667         }
 668 #else
 669         for (i = 0; i < npmem_ranges; i++) {
 670                 int j;
 671
 672                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
 673                         struct page *p;
 674                         unsigned long flags;
 675
 676                         pgdat_resize_lock(NODE_DATA(i), &flags);
 677                         p = nid_page_nr(i, j) - node_start_pfn(i);
 678
 679                         total++;
 680                         if (PageReserved(p))
 681                                 reserved++;
 682                         else if (PageSwapCache(p))
 683                                 cached++;
 684                         else if (!page_count(p))
 685                                 free++;
 686                         else
 687                                 shared += page_count(p) - 1;
 688                         pgdat_resize_unlock(NODE_DATA(i), &flags);
 689                 }
 690         }
 691 #endif
 692         printk(KERN_INFO "%d pages of RAM\n", total);
 693         printk(KERN_INFO "%d reserved pages\n", reserved);
 694         printk(KERN_INFO "%d pages shared\n", shared);
 695         printk(KERN_INFO "%d pages swap cached\n", cached);
 696
 697
 698 #ifdef CONFIG_DISCONTIGMEM
 699         {
 700                 struct zonelist *zl;
 701                 int i, j;
 702
 703                 for (i = 0; i < npmem_ranges; i++) {
 704                         zl = node_zonelist(i, 0);
 705                         for (j = 0; j < MAX_NR_ZONES; j++) {
 706                                 struct zoneref *z;
 707                                 struct zone *zone;
 708
 709                                 printk("Zone list for zone %d on node %d: ", j, i);
 710                                 for_each_zone_zonelist(zone, z, zl, j)
 711                                         printk("[%d/%s] ", zone_to_nid(zone),
 712                                                                 zone->name);
 713                                 printk("\n");
 714                         }
 715                 }
 716         }
 717 #endif
 718 }
 719
 720 /*
 721  * pagetable_init() sets up the page tables
 722  *
 723  * Note that gateway_init() places the Linux gateway page at page 0.
 724  * Since gateway pages cannot be dereferenced this has the desirable
 725  * side effect of trapping those pesky NULL-reference errors in the
 726  * kernel.
 727  */
 728 static void __init pagetable_init(void)
 729 {
 730         int range;
 731
 732         /* Map each physical memory range to its kernel vaddr */
 733
 734         for (range = 0; range < npmem_ranges; range++) {
 735                 unsigned long start_paddr;
 736                 unsigned long end_paddr;
 737                 unsigned long size;
 738
 739                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
 740                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
 741                 size = pmem_ranges[range].pages << PAGE_SHIFT;
 742
 743                 map_pages((unsigned long)__va(start_paddr), start_paddr,
 744                           size, PAGE_KERNEL, 0);
 745         }
 746
 747 #ifdef CONFIG_BLK_DEV_INITRD
 748         if (initrd_end && initrd_end > mem_limit) {
 749                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
 750                 map_pages(initrd_start, __pa(initrd_start),
 751                           initrd_end - initrd_start, PAGE_KERNEL, 0);
 752         }
 753 #endif
 754
 755         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
 756         memset(empty_zero_page, 0, PAGE_SIZE);
 757 }
 758
 759 static void __init gateway_init(void)
 760 {
 761         unsigned long linux_gateway_page_addr;
 762         /* FIXME: This is 'const' in order to trick the compiler
 763            into not treating it as DP-relative data. */
 764         extern void * const linux_gateway_page;
 765
 766         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
 767
 768         /*
 769          * Setup Linux Gateway page.
 770          *
 771          * The Linux gateway page will reside in kernel space (on virtual
 772          * page 0), so it doesn't need to be aliased into user space.
 773          */
 774
 775         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
 776                   PAGE_SIZE, PAGE_GATEWAY, 1);
 777 }
 778
 779 #ifdef CONFIG_HPUX
 780 void
 781 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
 782 {
 783         pgd_t *pg_dir;
 784         pmd_t *pmd;
 785         pte_t *pg_table;
 786         unsigned long start_pmd;
 787         unsigned long start_pte;
 788         unsigned long address;
 789         unsigned long hpux_gw_page_addr;
 790         /* FIXME: This is 'const' in order to trick the compiler
 791            into not treating it as DP-relative data. */
 792         extern void * const hpux_gateway_page;
 793
 794         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
 795
 796         /*
 797          * Setup HP-UX Gateway page.
 798          *
 799          * The HP-UX gateway page resides in the user address space,
 800          * so it needs to be aliased into each process.
 801          */
 802
 803         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
 804
 805 #if PTRS_PER_PMD == 1
 806         start_pmd = 0;
 807 #else
 808         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 809 #endif
 810         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 811
 812         address = __pa(&hpux_gateway_page);
 813 #if PTRS_PER_PMD == 1
 814         pmd = (pmd_t *)__pa(pg_dir);
 815 #else
 816         pmd = (pmd_t *) pgd_address(*pg_dir);
 817
 818         /*
 819          * pmd is physical at this point
 820          */
 821
 822         if (!pmd) {
 823                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
 824                 pmd = (pmd_t *) __pa(pmd);
 825         }
 826
 827         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
 828 #endif
 829         /* now change pmd to kernel virtual addresses */
 830
 831         pmd = (pmd_t *)__va(pmd) + start_pmd;
 832
 833         /*
 834          * pg_table is physical at this point
 835          */
 836
 837         pg_table = (pte_t *) pmd_address(*pmd);
 838         if (!pg_table)
 839                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
 840
 841         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
 842
 843         /* now change pg_table to kernel virtual addresses */
 844
 845         pg_table = (pte_t *) __va(pg_table) + start_pte;
 846         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
 847 }
 848 EXPORT_SYMBOL(map_hpux_gateway_page);
 849 #endif
 850
 851 void __init paging_init(void)
 852 {
 853         int i;
 854
 855         setup_bootmem();
 856         pagetable_init();
 857         gateway_init();
 858         flush_cache_all_local(); /* start with known state */
 859         flush_tlb_all_local(NULL);
 860
 861         for (i = 0; i < npmem_ranges; i++) {
 862                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
 863
 864                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
 865
 866 #ifdef CONFIG_DISCONTIGMEM
 867                 /* Need to initialize the pfnnid_map before we can initialize
 868                    the zone */
 869                 {
 870                     int j;
 871                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
 872                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
 873                          j++) {
 874                         pfnnid_map[j] = i;
 875                     }
 876                 }
 877 #endif
 878
 879                 free_area_init_node(i, zones_size,
 880                                 pmem_ranges[i].start_pfn, NULL);
 881         }
 882 }
 883
 884 #ifdef CONFIG_PA20
 885
 886 /*
 887  * Currently, all PA20 chips have 18 bit protection IDs, which is the
 888  * limiting factor (space ids are 32 bits).
 889  */
 890
 891 #define NR_SPACE_IDS 262144
 892
 893 #else
 894
 895 /*
 896  * Currently we have a one-to-one relationship between space IDs and
 897  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
 898  * support 15 bit protection IDs, so that is the limiting factor.
 899  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
 900  * probably not worth the effort for a special case here.
 901  */
 902
 903 #define NR_SPACE_IDS 32768
 904
 905 #endif  /* !CONFIG_PA20 */
 906
 907 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
 908 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
 909
 910 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
 911 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
 912 static unsigned long space_id_index;
 913 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
 914 static unsigned long dirty_space_ids = 0;
 915
 916 static DEFINE_SPINLOCK(sid_lock);
 917
 918 unsigned long alloc_sid(void)
 919 {
 920         unsigned long index;
 921
 922         spin_lock(&sid_lock);
 923
 924         if (free_space_ids == 0) {
 925                 if (dirty_space_ids != 0) {
 926                         spin_unlock(&sid_lock);
 927                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
 928                         spin_lock(&sid_lock);
 929                 }
 930                 BUG_ON(free_space_ids == 0);
 931         }
 932
 933         free_space_ids--;
 934
 935         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
 936         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
 937         space_id_index = index;
 938
 939         spin_unlock(&sid_lock);
 940
 941         return index << SPACEID_SHIFT;
 942 }
 943
 944 void free_sid(unsigned long spaceid)
 945 {
 946         unsigned long index = spaceid >> SPACEID_SHIFT;
 947         unsigned long *dirty_space_offset;
 948
 949         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
 950         index &= (BITS_PER_LONG - 1);
 951
 952         spin_lock(&sid_lock);
 953
 954         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
 955
 956         *dirty_space_offset |= (1L << index);
 957         dirty_space_ids++;
 958
 959         spin_unlock(&sid_lock);
 960 }
 961
 962
 963 #ifdef CONFIG_SMP
 964 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
 965 {
 966         int i;
 967
 968         /* NOTE: sid_lock must be held upon entry */
 969
 970         *ndirtyptr = dirty_space_ids;
 971         if (dirty_space_ids != 0) {
 972             for (i = 0; i < SID_ARRAY_SIZE; i++) {
 973                 dirty_array[i] = dirty_space_id[i];
 974                 dirty_space_id[i] = 0;
 975             }
 976             dirty_space_ids = 0;
 977         }
 978
 979         return;
 980 }
 981
 982 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
 983 {
 984         int i;
 985
 986         /* NOTE: sid_lock must be held upon entry */
 987
 988         if (ndirty != 0) {
 989                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
 990                         space_id[i] ^= dirty_array[i];
 991                 }
 992
 993                 free_space_ids += ndirty;
 994                 space_id_index = 0;
 995         }
 996 }
 997
 998 #else /* CONFIG_SMP */
 999
1000 static void recycle_sids(void)
1001 {
1002         int i;
1003
1004         /* NOTE: sid_lock must be held upon entry */
1005
1006         if (dirty_space_ids != 0) {
1007                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1008                         space_id[i] ^= dirty_space_id[i];
1009                         dirty_space_id[i] = 0;
1010                 }
1011
1012                 free_space_ids += dirty_space_ids;
1013                 dirty_space_ids = 0;
1014                 space_id_index = 0;
1015         }
1016 }
1017 #endif
1018
1019 /*
1020  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1021  * purged, we can safely reuse the space ids that were released but
1022  * not flushed from the tlb.
1023  */
1024
1025 #ifdef CONFIG_SMP
1026
1027 static unsigned long recycle_ndirty;
1028 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1029 static unsigned int recycle_inuse;
1030
1031 void flush_tlb_all(void)
1032 {
1033         int do_recycle;
1034
1035         __inc_irq_stat(irq_tlb_count);
1036         do_recycle = 0;
1037         spin_lock(&sid_lock);
1038         if (dirty_space_ids > RECYCLE_THRESHOLD) {
1039             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1040             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1041             recycle_inuse++;
1042             do_recycle++;
1043         }
1044         spin_unlock(&sid_lock);
1045         on_each_cpu(flush_tlb_all_local, NULL, 1);
1046         if (do_recycle) {
1047             spin_lock(&sid_lock);
1048             recycle_sids(recycle_ndirty,recycle_dirty_array);
1049             recycle_inuse = 0;
1050             spin_unlock(&sid_lock);
1051         }
1052 }
1053 #else
1054 void flush_tlb_all(void)
1055 {
1056         __inc_irq_stat(irq_tlb_count);
1057         spin_lock(&sid_lock);
1058         flush_tlb_all_local(NULL);
1059         recycle_sids();
1060         spin_unlock(&sid_lock);
1061 }
1062 #endif
1063
1064 #ifdef CONFIG_BLK_DEV_INITRD
1065 void free_initrd_mem(unsigned long start, unsigned long end)
1066 {
1067         free_reserved_area((void *)start, (void *)end, -1, "initrd");
1068 }
1069 #endif