arch/powerpc/mm/hash_utils_64.c

   1 /*
   2  * PowerPC64 port by Mike Corrigan and Dave Engebretsen
   3  *   {mikejc|engebret}@us.ibm.com
   4  *
   5  *    Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
   6  *
   7  * SMP scalability work:
   8  *    Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
   9  *
  10  *    Module name: htab.c
  11  *
  12  *    Description:
  13  *      PowerPC Hashed Page Table functions
  14  *
  15  * This program is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU General Public License
  17  * as published by the Free Software Foundation; either version
  18  * 2 of the License, or (at your option) any later version.
  19  */
  20
  21 #undef DEBUG
  22 #undef DEBUG_LOW
  23
  24 #include <linux/spinlock.h>
  25 #include <linux/errno.h>
  26 #include <linux/sched.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/stat.h>
  29 #include <linux/sysctl.h>
  30 #include <linux/export.h>
  31 #include <linux/ctype.h>
  32 #include <linux/cache.h>
  33 #include <linux/init.h>
  34 #include <linux/signal.h>
  35 #include <linux/memblock.h>
  36 #include <linux/context_tracking.h>
  37
  38 #include <asm/processor.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/mmu.h>
  41 #include <asm/mmu_context.h>
  42 #include <asm/page.h>
  43 #include <asm/types.h>
  44 #include <asm/uaccess.h>
  45 #include <asm/machdep.h>
  46 #include <asm/prom.h>
  47 #include <asm/tlbflush.h>
  48 #include <asm/io.h>
  49 #include <asm/eeh.h>
  50 #include <asm/tlb.h>
  51 #include <asm/cacheflush.h>
  52 #include <asm/cputable.h>
  53 #include <asm/sections.h>
  54 #include <asm/spu.h>
  55 #include <asm/udbg.h>
  56 #include <asm/code-patching.h>
  57 #include <asm/fadump.h>
  58 #include <asm/firmware.h>
  59 #include <asm/tm.h>
  60
  61 #ifdef DEBUG
  62 #define DBG(fmt...) udbg_printf(fmt)
  63 #else
  64 #define DBG(fmt...)
  65 #endif
  66
  67 #ifdef DEBUG_LOW
  68 #define DBG_LOW(fmt...) udbg_printf(fmt)
  69 #else
  70 #define DBG_LOW(fmt...)
  71 #endif
  72
  73 #define KB (1024)
  74 #define MB (1024*KB)
  75 #define GB (1024L*MB)
  76
  77 /*
  78  * Note:  pte   --> Linux PTE
  79  *        HPTE  --> PowerPC Hashed Page Table Entry
  80  *
  81  * Execution context:
  82  *   htab_initialize is called with the MMU off (of course), but
  83  *   the kernel has been copied down to zero so it can directly
  84  *   reference global data.  At this point it is very difficult
  85  *   to print debug info.
  86  *
  87  */
  88
  89 #ifdef CONFIG_U3_DART
  90 extern unsigned long dart_tablebase;
  91 #endif /* CONFIG_U3_DART */
  92
  93 static unsigned long _SDR1;
  94 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
  95
  96 struct hash_pte *htab_address;
  97 unsigned long htab_size_bytes;
  98 unsigned long htab_hash_mask;
  99 EXPORT_SYMBOL_GPL(htab_hash_mask);
 100 int mmu_linear_psize = MMU_PAGE_4K;
 101 int mmu_virtual_psize = MMU_PAGE_4K;
 102 int mmu_vmalloc_psize = MMU_PAGE_4K;
 103 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 104 int mmu_vmemmap_psize = MMU_PAGE_4K;
 105 #endif
 106 int mmu_io_psize = MMU_PAGE_4K;
 107 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
 108 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
 109 u16 mmu_slb_size = 64;
 110 EXPORT_SYMBOL_GPL(mmu_slb_size);
 111 #ifdef CONFIG_PPC_64K_PAGES
 112 int mmu_ci_restrictions;
 113 #endif
 114 #ifdef CONFIG_DEBUG_PAGEALLOC
 115 static u8 *linear_map_hash_slots;
 116 static unsigned long linear_map_hash_count;
 117 static DEFINE_SPINLOCK(linear_map_hash_lock);
 118 #endif /* CONFIG_DEBUG_PAGEALLOC */
 119
 120 /* There are definitions of page sizes arrays to be used when none
 121  * is provided by the firmware.
 122  */
 123
 124 /* Pre-POWER4 CPUs (4k pages only)
 125  */
 126 static struct mmu_psize_def mmu_psize_defaults_old[] = {
 127         [MMU_PAGE_4K] = {
 128                 .shift  = 12,
 129                 .sllp   = 0,
 130                 .penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
 131                 .avpnm  = 0,
 132                 .tlbiel = 0,
 133         },
 134 };
 135
 136 /* POWER4, GPUL, POWER5
 137  *
 138  * Support for 16Mb large pages
 139  */
 140 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
 141         [MMU_PAGE_4K] = {
 142                 .shift  = 12,
 143                 .sllp   = 0,
 144                 .penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
 145                 .avpnm  = 0,
 146                 .tlbiel = 1,
 147         },
 148         [MMU_PAGE_16M] = {
 149                 .shift  = 24,
 150                 .sllp   = SLB_VSID_L,
 151                 .penc   = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
 152                             [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
 153                 .avpnm  = 0x1UL,
 154                 .tlbiel = 0,
 155         },
 156 };
 157
 158 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
 159 {
 160         unsigned long rflags = pteflags & 0x1fa;
 161
 162         /* _PAGE_EXEC -> NOEXEC */
 163         if ((pteflags & _PAGE_EXEC) == 0)
 164                 rflags |= HPTE_R_N;
 165
 166         /* PP bits. PAGE_USER is already PP bit 0x2, so we only
 167          * need to add in 0x1 if it's a read-only user page
 168          */
 169         if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
 170                                          (pteflags & _PAGE_DIRTY)))
 171                 rflags |= 1;
 172         /*
 173          * Always add "C" bit for perf. Memory coherence is always enabled
 174          */
 175         return rflags | HPTE_R_C | HPTE_R_M;
 176 }
 177
 178 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
 179                       unsigned long pstart, unsigned long prot,
 180                       int psize, int ssize)
 181 {
 182         unsigned long vaddr, paddr;
 183         unsigned int step, shift;
 184         int ret = 0;
 185
 186         shift = mmu_psize_defs[psize].shift;
 187         step = 1 << shift;
 188
 189         prot = htab_convert_pte_flags(prot);
 190
 191         DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
 192             vstart, vend, pstart, prot, psize, ssize);
 193
 194         for (vaddr = vstart, paddr = pstart; vaddr < vend;
 195              vaddr += step, paddr += step) {
 196                 unsigned long hash, hpteg;
 197                 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
 198                 unsigned long vpn  = hpt_vpn(vaddr, vsid, ssize);
 199                 unsigned long tprot = prot;
 200
 201                 /*
 202                  * If we hit a bad address return error.
 203                  */
 204                 if (!vsid)
 205                         return -1;
 206                 /* Make kernel text executable */
 207                 if (overlaps_kernel_text(vaddr, vaddr + step))
 208                         tprot &= ~HPTE_R_N;
 209
 210                 hash = hpt_hash(vpn, shift, ssize);
 211                 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
 212
 213                 BUG_ON(!ppc_md.hpte_insert);
 214                 ret = ppc_md.hpte_insert(hpteg, vpn, paddr, tprot,
 215                                          HPTE_V_BOLTED, psize, psize, ssize);
 216
 217                 if (ret < 0)
 218                         break;
 219 #ifdef CONFIG_DEBUG_PAGEALLOC
 220                 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
 221                         linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
 222 #endif /* CONFIG_DEBUG_PAGEALLOC */
 223         }
 224         return ret < 0 ? ret : 0;
 225 }
 226
 227 #ifdef CONFIG_MEMORY_HOTPLUG
 228 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
 229                       int psize, int ssize)
 230 {
 231         unsigned long vaddr;
 232         unsigned int step, shift;
 233
 234         shift = mmu_psize_defs[psize].shift;
 235         step = 1 << shift;
 236
 237         if (!ppc_md.hpte_removebolted) {
 238                 printk(KERN_WARNING "Platform doesn't implement "
 239                                 "hpte_removebolted\n");
 240                 return -EINVAL;
 241         }
 242
 243         for (vaddr = vstart; vaddr < vend; vaddr += step)
 244                 ppc_md.hpte_removebolted(vaddr, psize, ssize);
 245
 246         return 0;
 247 }
 248 #endif /* CONFIG_MEMORY_HOTPLUG */
 249
 250 static int __init htab_dt_scan_seg_sizes(unsigned long node,
 251                                          const char *uname, int depth,
 252                                          void *data)
 253 {
 254         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 255         __be32 *prop;
 256         unsigned long size = 0;
 257
 258         /* We are scanning "cpu" nodes only */
 259         if (type == NULL || strcmp(type, "cpu") != 0)
 260                 return 0;
 261
 262         prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
 263         if (prop == NULL)
 264                 return 0;
 265         for (; size >= 4; size -= 4, ++prop) {
 266                 if (be32_to_cpu(prop[0]) == 40) {
 267                         DBG("1T segment support detected\n");
 268                         cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
 269                         return 1;
 270                 }
 271         }
 272         cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
 273         return 0;
 274 }
 275
 276 static void __init htab_init_seg_sizes(void)
 277 {
 278         of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
 279 }
 280
 281 static int __init get_idx_from_shift(unsigned int shift)
 282 {
 283         int idx = -1;
 284
 285         switch (shift) {
 286         case 0xc:
 287                 idx = MMU_PAGE_4K;
 288                 break;
 289         case 0x10:
 290                 idx = MMU_PAGE_64K;
 291                 break;
 292         case 0x14:
 293                 idx = MMU_PAGE_1M;
 294                 break;
 295         case 0x18:
 296                 idx = MMU_PAGE_16M;
 297                 break;
 298         case 0x22:
 299                 idx = MMU_PAGE_16G;
 300                 break;
 301         }
 302         return idx;
 303 }
 304
 305 static int __init htab_dt_scan_page_sizes(unsigned long node,
 306                                           const char *uname, int depth,
 307                                           void *data)
 308 {
 309         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 310         __be32 *prop;
 311         unsigned long size = 0;
 312
 313         /* We are scanning "cpu" nodes only */
 314         if (type == NULL || strcmp(type, "cpu") != 0)
 315                 return 0;
 316
 317         prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
 318         if (prop != NULL) {
 319                 pr_info("Page sizes from device-tree:\n");
 320                 size /= 4;
 321                 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
 322                 while(size > 0) {
 323                         unsigned int base_shift = be32_to_cpu(prop[0]);
 324                         unsigned int slbenc = be32_to_cpu(prop[1]);
 325                         unsigned int lpnum = be32_to_cpu(prop[2]);
 326                         struct mmu_psize_def *def;
 327                         int idx, base_idx;
 328
 329                         size -= 3; prop += 3;
 330                         base_idx = get_idx_from_shift(base_shift);
 331                         if (base_idx < 0) {
 332                                 /*
 333                                  * skip the pte encoding also
 334                                  */
 335                                 prop += lpnum * 2; size -= lpnum * 2;
 336                                 continue;
 337                         }
 338                         def = &mmu_psize_defs[base_idx];
 339                         if (base_idx == MMU_PAGE_16M)
 340                                 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
 341
 342                         def->shift = base_shift;
 343                         if (base_shift <= 23)
 344                                 def->avpnm = 0;
 345                         else
 346                                 def->avpnm = (1 << (base_shift - 23)) - 1;
 347                         def->sllp = slbenc;
 348                         /*
 349                          * We don't know for sure what's up with tlbiel, so
 350                          * for now we only set it for 4K and 64K pages
 351                          */
 352                         if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
 353                                 def->tlbiel = 1;
 354                         else
 355                                 def->tlbiel = 0;
 356
 357                         while (size > 0 && lpnum) {
 358                                 unsigned int shift = be32_to_cpu(prop[0]);
 359                                 int penc  = be32_to_cpu(prop[1]);
 360
 361                                 prop += 2; size -= 2;
 362                                 lpnum--;
 363
 364                                 idx = get_idx_from_shift(shift);
 365                                 if (idx < 0)
 366                                         continue;
 367
 368                                 if (penc == -1)
 369                                         pr_err("Invalid penc for base_shift=%d "
 370                                                "shift=%d\n", base_shift, shift);
 371
 372                                 def->penc[idx] = penc;
 373                                 pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
 374                                         " avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
 375                                         base_shift, shift, def->sllp,
 376                                         def->avpnm, def->tlbiel, def->penc[idx]);
 377                         }
 378                 }
 379                 return 1;
 380         }
 381         return 0;
 382 }
 383
 384 #ifdef CONFIG_HUGETLB_PAGE
 385 /* Scan for 16G memory blocks that have been set aside for huge pages
 386  * and reserve those blocks for 16G huge pages.
 387  */
 388 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
 389                                         const char *uname, int depth,
 390                                         void *data) {
 391         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 392         __be64 *addr_prop;
 393         __be32 *page_count_prop;
 394         unsigned int expected_pages;
 395         long unsigned int phys_addr;
 396         long unsigned int block_size;
 397
 398         /* We are scanning "memory" nodes only */
 399         if (type == NULL || strcmp(type, "memory") != 0)
 400                 return 0;
 401
 402         /* This property is the log base 2 of the number of virtual pages that
 403          * will represent this memory block. */
 404         page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
 405         if (page_count_prop == NULL)
 406                 return 0;
 407         expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
 408         addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
 409         if (addr_prop == NULL)
 410                 return 0;
 411         phys_addr = be64_to_cpu(addr_prop[0]);
 412         block_size = be64_to_cpu(addr_prop[1]);
 413         if (block_size != (16 * GB))
 414                 return 0;
 415         printk(KERN_INFO "Huge page(16GB) memory: "
 416                         "addr = 0x%lX size = 0x%lX pages = %d\n",
 417                         phys_addr, block_size, expected_pages);
 418         if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
 419                 memblock_reserve(phys_addr, block_size * expected_pages);
 420                 add_gpage(phys_addr, block_size, expected_pages);
 421         }
 422         return 0;
 423 }
 424 #endif /* CONFIG_HUGETLB_PAGE */
 425
 426 static void mmu_psize_set_default_penc(void)
 427 {
 428         int bpsize, apsize;
 429         for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
 430                 for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
 431                         mmu_psize_defs[bpsize].penc[apsize] = -1;
 432 }
 433
 434 static void __init htab_init_page_sizes(void)
 435 {
 436         int rc;
 437
 438         /* se the invalid penc to -1 */
 439         mmu_psize_set_default_penc();
 440
 441         /* Default to 4K pages only */
 442         memcpy(mmu_psize_defs, mmu_psize_defaults_old,
 443                sizeof(mmu_psize_defaults_old));
 444
 445         /*
 446          * Try to find the available page sizes in the device-tree
 447          */
 448         rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
 449         if (rc != 0)  /* Found */
 450                 goto found;
 451
 452         /*
 453          * Not in the device-tree, let's fallback on known size
 454          * list for 16M capable GP & GR
 455          */
 456         if (mmu_has_feature(MMU_FTR_16M_PAGE))
 457                 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
 458                        sizeof(mmu_psize_defaults_gp));
 459  found:
 460 #ifndef CONFIG_DEBUG_PAGEALLOC
 461         /*
 462          * Pick a size for the linear mapping. Currently, we only support
 463          * 16M, 1M and 4K which is the default
 464          */
 465         if (mmu_psize_defs[MMU_PAGE_16M].shift)
 466                 mmu_linear_psize = MMU_PAGE_16M;
 467         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 468                 mmu_linear_psize = MMU_PAGE_1M;
 469 #endif /* CONFIG_DEBUG_PAGEALLOC */
 470
 471 #ifdef CONFIG_PPC_64K_PAGES
 472         /*
 473          * Pick a size for the ordinary pages. Default is 4K, we support
 474          * 64K for user mappings and vmalloc if supported by the processor.
 475          * We only use 64k for ioremap if the processor
 476          * (and firmware) support cache-inhibited large pages.
 477          * If not, we use 4k and set mmu_ci_restrictions so that
 478          * hash_page knows to switch processes that use cache-inhibited
 479          * mappings to 4k pages.
 480          */
 481         if (mmu_psize_defs[MMU_PAGE_64K].shift) {
 482                 mmu_virtual_psize = MMU_PAGE_64K;
 483                 mmu_vmalloc_psize = MMU_PAGE_64K;
 484                 if (mmu_linear_psize == MMU_PAGE_4K)
 485                         mmu_linear_psize = MMU_PAGE_64K;
 486                 if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
 487                         /*
 488                          * Don't use 64k pages for ioremap on pSeries, since
 489                          * that would stop us accessing the HEA ethernet.
 490                          */
 491                         if (!machine_is(pseries))
 492                                 mmu_io_psize = MMU_PAGE_64K;
 493                 } else
 494                         mmu_ci_restrictions = 1;
 495         }
 496 #endif /* CONFIG_PPC_64K_PAGES */
 497
 498 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 499         /* We try to use 16M pages for vmemmap if that is supported
 500          * and we have at least 1G of RAM at boot
 501          */
 502         if (mmu_psize_defs[MMU_PAGE_16M].shift &&
 503             memblock_phys_mem_size() >= 0x40000000)
 504                 mmu_vmemmap_psize = MMU_PAGE_16M;
 505         else if (mmu_psize_defs[MMU_PAGE_64K].shift)
 506                 mmu_vmemmap_psize = MMU_PAGE_64K;
 507         else
 508                 mmu_vmemmap_psize = MMU_PAGE_4K;
 509 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 510
 511         printk(KERN_DEBUG "Page orders: linear mapping = %d, "
 512                "virtual = %d, io = %d"
 513 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 514                ", vmemmap = %d"
 515 #endif
 516                "\n",
 517                mmu_psize_defs[mmu_linear_psize].shift,
 518                mmu_psize_defs[mmu_virtual_psize].shift,
 519                mmu_psize_defs[mmu_io_psize].shift
 520 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 521                ,mmu_psize_defs[mmu_vmemmap_psize].shift
 522 #endif
 523                );
 524
 525 #ifdef CONFIG_HUGETLB_PAGE
 526         /* Reserve 16G huge page memory sections for huge pages */
 527         of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
 528 #endif /* CONFIG_HUGETLB_PAGE */
 529 }
 530
 531 static int __init htab_dt_scan_pftsize(unsigned long node,
 532                                        const char *uname, int depth,
 533                                        void *data)
 534 {
 535         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 536         __be32 *prop;
 537
 538         /* We are scanning "cpu" nodes only */
 539         if (type == NULL || strcmp(type, "cpu") != 0)
 540                 return 0;
 541
 542         prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
 543         if (prop != NULL) {
 544                 /* pft_size[0] is the NUMA CEC cookie */
 545                 ppc64_pft_size = be32_to_cpu(prop[1]);
 546                 return 1;
 547         }
 548         return 0;
 549 }
 550
 551 static unsigned long __init htab_get_table_size(void)
 552 {
 553         unsigned long mem_size, rnd_mem_size, pteg_count, psize;
 554
 555         /* If hash size isn't already provided by the platform, we try to
 556          * retrieve it from the device-tree. If it's not there neither, we
 557          * calculate it now based on the total RAM size
 558          */
 559         if (ppc64_pft_size == 0)
 560                 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
 561         if (ppc64_pft_size)
 562                 return 1UL << ppc64_pft_size;
 563
 564         /* round mem_size up to next power of 2 */
 565         mem_size = memblock_phys_mem_size();
 566         rnd_mem_size = 1UL << __ilog2(mem_size);
 567         if (rnd_mem_size < mem_size)
 568                 rnd_mem_size <<= 1;
 569
 570         /* # pages / 2 */
 571         psize = mmu_psize_defs[mmu_virtual_psize].shift;
 572         pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
 573
 574         return pteg_count << 7;
 575 }
 576
 577 #ifdef CONFIG_MEMORY_HOTPLUG
 578 int create_section_mapping(unsigned long start, unsigned long end)
 579 {
 580         return htab_bolt_mapping(start, end, __pa(start),
 581                                  pgprot_val(PAGE_KERNEL), mmu_linear_psize,
 582                                  mmu_kernel_ssize);
 583 }
 584
 585 int remove_section_mapping(unsigned long start, unsigned long end)
 586 {
 587         return htab_remove_mapping(start, end, mmu_linear_psize,
 588                         mmu_kernel_ssize);
 589 }
 590 #endif /* CONFIG_MEMORY_HOTPLUG */
 591
 592 #define FUNCTION_TEXT(A)        ((*(unsigned long *)(A)))
 593
 594 static void __init htab_finish_init(void)
 595 {
 596         extern unsigned int *htab_call_hpte_insert1;
 597         extern unsigned int *htab_call_hpte_insert2;
 598         extern unsigned int *htab_call_hpte_remove;
 599         extern unsigned int *htab_call_hpte_updatepp;
 600
 601 #ifdef CONFIG_PPC_HAS_HASH_64K
 602         extern unsigned int *ht64_call_hpte_insert1;
 603         extern unsigned int *ht64_call_hpte_insert2;
 604         extern unsigned int *ht64_call_hpte_remove;
 605         extern unsigned int *ht64_call_hpte_updatepp;
 606
 607         patch_branch(ht64_call_hpte_insert1,
 608                 FUNCTION_TEXT(ppc_md.hpte_insert),
 609                 BRANCH_SET_LINK);
 610         patch_branch(ht64_call_hpte_insert2,
 611                 FUNCTION_TEXT(ppc_md.hpte_insert),
 612                 BRANCH_SET_LINK);
 613         patch_branch(ht64_call_hpte_remove,
 614                 FUNCTION_TEXT(ppc_md.hpte_remove),
 615                 BRANCH_SET_LINK);
 616         patch_branch(ht64_call_hpte_updatepp,
 617                 FUNCTION_TEXT(ppc_md.hpte_updatepp),
 618                 BRANCH_SET_LINK);
 619
 620 #endif /* CONFIG_PPC_HAS_HASH_64K */
 621
 622         patch_branch(htab_call_hpte_insert1,
 623                 FUNCTION_TEXT(ppc_md.hpte_insert),
 624                 BRANCH_SET_LINK);
 625         patch_branch(htab_call_hpte_insert2,
 626                 FUNCTION_TEXT(ppc_md.hpte_insert),
 627                 BRANCH_SET_LINK);
 628         patch_branch(htab_call_hpte_remove,
 629                 FUNCTION_TEXT(ppc_md.hpte_remove),
 630                 BRANCH_SET_LINK);
 631         patch_branch(htab_call_hpte_updatepp,
 632                 FUNCTION_TEXT(ppc_md.hpte_updatepp),
 633                 BRANCH_SET_LINK);
 634 }
 635
 636 static void __init htab_initialize(void)
 637 {
 638         unsigned long table;
 639         unsigned long pteg_count;
 640         unsigned long prot;
 641         unsigned long base = 0, size = 0, limit;
 642         struct memblock_region *reg;
 643
 644         DBG(" -> htab_initialize()\n");
 645
 646         /* Initialize segment sizes */
 647         htab_init_seg_sizes();
 648
 649         /* Initialize page sizes */
 650         htab_init_page_sizes();
 651
 652         if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
 653                 mmu_kernel_ssize = MMU_SEGSIZE_1T;
 654                 mmu_highuser_ssize = MMU_SEGSIZE_1T;
 655                 printk(KERN_INFO "Using 1TB segments\n");
 656         }
 657
 658         /*
 659          * Calculate the required size of the htab.  We want the number of
 660          * PTEGs to equal one half the number of real pages.
 661          */
 662         htab_size_bytes = htab_get_table_size();
 663         pteg_count = htab_size_bytes >> 7;
 664
 665         htab_hash_mask = pteg_count - 1;
 666
 667         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 668                 /* Using a hypervisor which owns the htab */
 669                 htab_address = NULL;
 670                 _SDR1 = 0;
 671 #ifdef CONFIG_FA_DUMP
 672                 /*
 673                  * If firmware assisted dump is active firmware preserves
 674                  * the contents of htab along with entire partition memory.
 675                  * Clear the htab if firmware assisted dump is active so
 676                  * that we dont end up using old mappings.
 677                  */
 678                 if (is_fadump_active() && ppc_md.hpte_clear_all)
 679                         ppc_md.hpte_clear_all();
 680 #endif
 681         } else {
 682                 /* Find storage for the HPT.  Must be contiguous in
 683                  * the absolute address space. On cell we want it to be
 684                  * in the first 2 Gig so we can use it for IOMMU hacks.
 685                  */
 686                 if (machine_is(cell))
 687                         limit = 0x80000000;
 688                 else
 689                         limit = MEMBLOCK_ALLOC_ANYWHERE;
 690
 691                 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
 692
 693                 DBG("Hash table allocated at %lx, size: %lx\n", table,
 694                     htab_size_bytes);
 695
 696                 htab_address = __va(table);
 697
 698                 /* htab absolute addr + encoded htabsize */
 699                 _SDR1 = table + __ilog2(pteg_count) - 11;
 700
 701                 /* Initialize the HPT with no entries */
 702                 memset((void *)table, 0, htab_size_bytes);
 703
 704                 /* Set SDR1 */
 705                 mtspr(SPRN_SDR1, _SDR1);
 706         }
 707
 708         prot = pgprot_val(PAGE_KERNEL);
 709
 710 #ifdef CONFIG_DEBUG_PAGEALLOC
 711         linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
 712         linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
 713                                                     1, ppc64_rma_size));
 714         memset(linear_map_hash_slots, 0, linear_map_hash_count);
 715 #endif /* CONFIG_DEBUG_PAGEALLOC */
 716
 717         /* On U3 based machines, we need to reserve the DART area and
 718          * _NOT_ map it to avoid cache paradoxes as it's remapped non
 719          * cacheable later on
 720          */
 721
 722         /* create bolted the linear mapping in the hash table */
 723         for_each_memblock(memory, reg) {
 724                 base = (unsigned long)__va(reg->base);
 725                 size = reg->size;
 726
 727                 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
 728                     base, size, prot);
 729
 730 #ifdef CONFIG_U3_DART
 731                 /* Do not map the DART space. Fortunately, it will be aligned
 732                  * in such a way that it will not cross two memblock regions and
 733                  * will fit within a single 16Mb page.
 734                  * The DART space is assumed to be a full 16Mb region even if
 735                  * we only use 2Mb of that space. We will use more of it later
 736                  * for AGP GART. We have to use a full 16Mb large page.
 737                  */
 738                 DBG("DART base: %lx\n", dart_tablebase);
 739
 740                 if (dart_tablebase != 0 && dart_tablebase >= base
 741                     && dart_tablebase < (base + size)) {
 742                         unsigned long dart_table_end = dart_tablebase + 16 * MB;
 743                         if (base != dart_tablebase)
 744                                 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
 745                                                         __pa(base), prot,
 746                                                         mmu_linear_psize,
 747                                                         mmu_kernel_ssize));
 748                         if ((base + size) > dart_table_end)
 749                                 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
 750                                                         base + size,
 751                                                         __pa(dart_table_end),
 752                                                          prot,
 753                                                          mmu_linear_psize,
 754                                                          mmu_kernel_ssize));
 755                         continue;
 756                 }
 757 #endif /* CONFIG_U3_DART */
 758                 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
 759                                 prot, mmu_linear_psize, mmu_kernel_ssize));
 760         }
 761         memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
 762
 763         /*
 764          * If we have a memory_limit and we've allocated TCEs then we need to
 765          * explicitly map the TCE area at the top of RAM. We also cope with the
 766          * case that the TCEs start below memory_limit.
 767          * tce_alloc_start/end are 16MB aligned so the mapping should work
 768          * for either 4K or 16MB pages.
 769          */
 770         if (tce_alloc_start) {
 771                 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
 772                 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
 773
 774                 if (base + size >= tce_alloc_start)
 775                         tce_alloc_start = base + size + 1;
 776
 777                 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
 778                                          __pa(tce_alloc_start), prot,
 779                                          mmu_linear_psize, mmu_kernel_ssize));
 780         }
 781
 782         htab_finish_init();
 783
 784         DBG(" <- htab_initialize()\n");
 785 }
 786 #undef KB
 787 #undef MB
 788
 789 void __init early_init_mmu(void)
 790 {
 791         /* Setup initial STAB address in the PACA */
 792         get_paca()->stab_real = __pa((u64)&initial_stab);
 793         get_paca()->stab_addr = (u64)&initial_stab;
 794
 795         /* Initialize the MMU Hash table and create the linear mapping
 796          * of memory. Has to be done before stab/slb initialization as
 797          * this is currently where the page size encoding is obtained
 798          */
 799         htab_initialize();
 800
 801         /* Initialize stab / SLB management */
 802         if (mmu_has_feature(MMU_FTR_SLB))
 803                 slb_initialize();
 804         else
 805                 stab_initialize(get_paca()->stab_real);
 806 }
 807
 808 #ifdef CONFIG_SMP
 809 void early_init_mmu_secondary(void)
 810 {
 811         /* Initialize hash table for that CPU */
 812         if (!firmware_has_feature(FW_FEATURE_LPAR))
 813                 mtspr(SPRN_SDR1, _SDR1);
 814
 815         /* Initialize STAB/SLB. We use a virtual address as it works
 816          * in real mode on pSeries.
 817          */
 818         if (mmu_has_feature(MMU_FTR_SLB))
 819                 slb_initialize();
 820         else
 821                 stab_initialize(get_paca()->stab_addr);
 822 }
 823 #endif /* CONFIG_SMP */
 824
 825 /*
 826  * Called by asm hashtable.S for doing lazy icache flush
 827  */
 828 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
 829 {
 830         struct page *page;
 831
 832         if (!pfn_valid(pte_pfn(pte)))
 833                 return pp;
 834
 835         page = pte_page(pte);
 836
 837         /* page is dirty */
 838         if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
 839                 if (trap == 0x400) {
 840                         flush_dcache_icache_page(page);
 841                         set_bit(PG_arch_1, &page->flags);
 842                 } else
 843                         pp |= HPTE_R_N;
 844         }
 845         return pp;
 846 }
 847
 848 #ifdef CONFIG_PPC_MM_SLICES
 849 unsigned int get_paca_psize(unsigned long addr)
 850 {
 851         u64 lpsizes;
 852         unsigned char *hpsizes;
 853         unsigned long index, mask_index;
 854
 855         if (addr < SLICE_LOW_TOP) {
 856                 lpsizes = get_paca()->context.low_slices_psize;
 857                 index = GET_LOW_SLICE_INDEX(addr);
 858                 return (lpsizes >> (index * 4)) & 0xF;
 859         }
 860         hpsizes = get_paca()->context.high_slices_psize;
 861         index = GET_HIGH_SLICE_INDEX(addr);
 862         mask_index = index & 0x1;
 863         return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
 864 }
 865
 866 #else
 867 unsigned int get_paca_psize(unsigned long addr)
 868 {
 869         return get_paca()->context.user_psize;
 870 }
 871 #endif
 872
 873 /*
 874  * Demote a segment to using 4k pages.
 875  * For now this makes the whole process use 4k pages.
 876  */
 877 #ifdef CONFIG_PPC_64K_PAGES
 878 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
 879 {
 880         if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
 881                 return;
 882         slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
 883 #ifdef CONFIG_SPU_BASE
 884         spu_flush_all_slbs(mm);
 885 #endif
 886         if (get_paca_psize(addr) != MMU_PAGE_4K) {
 887                 get_paca()->context = mm->context;
 888                 slb_flush_and_rebolt();
 889         }
 890 }
 891 #endif /* CONFIG_PPC_64K_PAGES */
 892
 893 #ifdef CONFIG_PPC_SUBPAGE_PROT
 894 /*
 895  * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
 896  * Userspace sets the subpage permissions using the subpage_prot system call.
 897  *
 898  * Result is 0: full permissions, _PAGE_RW: read-only,
 899  * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
 900  */
 901 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
 902 {
 903         struct subpage_prot_table *spt = &mm->context.spt;
 904         u32 spp = 0;
 905         u32 **sbpm, *sbpp;
 906
 907         if (ea >= spt->maxaddr)
 908                 return 0;
 909         if (ea < 0x100000000UL) {
 910                 /* addresses below 4GB use spt->low_prot */
 911                 sbpm = spt->low_prot;
 912         } else {
 913                 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
 914                 if (!sbpm)
 915                         return 0;
 916         }
 917         sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
 918         if (!sbpp)
 919                 return 0;
 920         spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
 921
 922         /* extract 2-bit bitfield for this 4k subpage */
 923         spp >>= 30 - 2 * ((ea >> 12) & 0xf);
 924
 925         /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
 926         spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
 927         return spp;
 928 }
 929
 930 #else /* CONFIG_PPC_SUBPAGE_PROT */
 931 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
 932 {
 933         return 0;
 934 }
 935 #endif
 936
 937 void hash_failure_debug(unsigned long ea, unsigned long access,
 938                         unsigned long vsid, unsigned long trap,
 939                         int ssize, int psize, int lpsize, unsigned long pte)
 940 {
 941         if (!printk_ratelimit())
 942                 return;
 943         pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
 944                 ea, access, current->comm);
 945         pr_info("    trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
 946                 trap, vsid, ssize, psize, lpsize, pte);
 947 }
 948
 949 /* Result code is:
 950  *  0 - handled
 951  *  1 - normal page fault
 952  * -1 - critical hash insertion error
 953  * -2 - access not permitted by subpage protection mechanism
 954  */
 955 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
 956 {
 957         enum ctx_state prev_state = exception_enter();
 958         pgd_t *pgdir;
 959         unsigned long vsid;
 960         struct mm_struct *mm;
 961         pte_t *ptep;
 962         unsigned hugeshift;
 963         const struct cpumask *tmp;
 964         int rc, user_region = 0, local = 0;
 965         int psize, ssize;
 966
 967         DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
 968                 ea, access, trap);
 969
 970         /* Get region & vsid */
 971         switch (REGION_ID(ea)) {
 972         case USER_REGION_ID:
 973                 user_region = 1;
 974                 mm = current->mm;
 975                 if (! mm) {
 976                         DBG_LOW(" user region with no mm !\n");
 977                         rc = 1;
 978                         goto bail;
 979                 }
 980                 psize = get_slice_psize(mm, ea);
 981                 ssize = user_segment_size(ea);
 982                 vsid = get_vsid(mm->context.id, ea, ssize);
 983                 break;
 984         case VMALLOC_REGION_ID:
 985                 mm = &init_mm;
 986                 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
 987                 if (ea < VMALLOC_END)
 988                         psize = mmu_vmalloc_psize;
 989                 else
 990                         psize = mmu_io_psize;
 991                 ssize = mmu_kernel_ssize;
 992                 break;
 993         default:
 994                 /* Not a valid range
 995                  * Send the problem up to do_page_fault
 996                  */
 997                 rc = 1;
 998                 goto bail;
 999         }
1000         DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1001
1002         /* Bad address. */
1003         if (!vsid) {
1004                 DBG_LOW("Bad address!\n");
1005                 rc = 1;
1006                 goto bail;
1007         }
1008         /* Get pgdir */
1009         pgdir = mm->pgd;
1010         if (pgdir == NULL) {
1011                 rc = 1;
1012                 goto bail;
1013         }
1014
1015         /* Check CPU locality */
1016         tmp = cpumask_of(smp_processor_id());
1017         if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
1018                 local = 1;
1019
1020 #ifndef CONFIG_PPC_64K_PAGES
1021         /* If we use 4K pages and our psize is not 4K, then we might
1022          * be hitting a special driver mapping, and need to align the
1023          * address before we fetch the PTE.
1024          *
1025          * It could also be a hugepage mapping, in which case this is
1026          * not necessary, but it's not harmful, either.
1027          */
1028         if (psize != MMU_PAGE_4K)
1029                 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1030 #endif /* CONFIG_PPC_64K_PAGES */
1031
1032         /* Get PTE and page size from page tables */
1033         ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
1034         if (ptep == NULL || !pte_present(*ptep)) {
1035                 DBG_LOW(" no PTE !\n");
1036                 rc = 1;
1037                 goto bail;
1038         }
1039
1040         /* Add _PAGE_PRESENT to the required access perm */
1041         access |= _PAGE_PRESENT;
1042
1043         /* Pre-check access permissions (will be re-checked atomically
1044          * in __hash_page_XX but this pre-check is a fast path
1045          */
1046         if (access & ~pte_val(*ptep)) {
1047                 DBG_LOW(" no access !\n");
1048                 rc = 1;
1049                 goto bail;
1050         }
1051
1052         if (hugeshift) {
1053                 if (pmd_trans_huge(*(pmd_t *)ptep))
1054                         rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1055                                              trap, local, ssize, psize);
1056 #ifdef CONFIG_HUGETLB_PAGE
1057                 else
1058                         rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1059                                               local, ssize, hugeshift, psize);
1060 #else
1061                 else {
1062                         /*
1063                          * if we have hugeshift, and is not transhuge with
1064                          * hugetlb disabled, something is really wrong.
1065                          */
1066                         rc = 1;
1067                         WARN_ON(1);
1068                 }
1069 #endif
1070                 goto bail;
1071         }
1072
1073 #ifndef CONFIG_PPC_64K_PAGES
1074         DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1075 #else
1076         DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1077                 pte_val(*(ptep + PTRS_PER_PTE)));
1078 #endif
1079         /* Do actual hashing */
1080 #ifdef CONFIG_PPC_64K_PAGES
1081         /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
1082         if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1083                 demote_segment_4k(mm, ea);
1084                 psize = MMU_PAGE_4K;
1085         }
1086
1087         /* If this PTE is non-cacheable and we have restrictions on
1088          * using non cacheable large pages, then we switch to 4k
1089          */
1090         if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
1091             (pte_val(*ptep) & _PAGE_NO_CACHE)) {
1092                 if (user_region) {
1093                         demote_segment_4k(mm, ea);
1094                         psize = MMU_PAGE_4K;
1095                 } else if (ea < VMALLOC_END) {
1096                         /*
1097                          * some driver did a non-cacheable mapping
1098                          * in vmalloc space, so switch vmalloc
1099                          * to 4k pages
1100                          */
1101                         printk(KERN_ALERT "Reducing vmalloc segment "
1102                                "to 4kB pages because of "
1103                                "non-cacheable mapping\n");
1104                         psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1105 #ifdef CONFIG_SPU_BASE
1106                         spu_flush_all_slbs(mm);
1107 #endif
1108                 }
1109         }
1110         if (user_region) {
1111                 if (psize != get_paca_psize(ea)) {
1112                         get_paca()->context = mm->context;
1113                         slb_flush_and_rebolt();
1114                 }
1115         } else if (get_paca()->vmalloc_sllp !=
1116                    mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1117                 get_paca()->vmalloc_sllp =
1118                         mmu_psize_defs[mmu_vmalloc_psize].sllp;
1119                 slb_vmalloc_update();
1120         }
1121 #endif /* CONFIG_PPC_64K_PAGES */
1122
1123 #ifdef CONFIG_PPC_HAS_HASH_64K
1124         if (psize == MMU_PAGE_64K)
1125                 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1126         else
1127 #endif /* CONFIG_PPC_HAS_HASH_64K */
1128         {
1129                 int spp = subpage_protection(mm, ea);
1130                 if (access & spp)
1131                         rc = -2;
1132                 else
1133                         rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1134                                             local, ssize, spp);
1135         }
1136
1137         /* Dump some info in case of hash insertion failure, they should
1138          * never happen so it is really useful to know if/when they do
1139          */
1140         if (rc == -1)
1141                 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1142                                    psize, pte_val(*ptep));
1143 #ifndef CONFIG_PPC_64K_PAGES
1144         DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1145 #else
1146         DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1147                 pte_val(*(ptep + PTRS_PER_PTE)));
1148 #endif
1149         DBG_LOW(" -> rc=%d\n", rc);
1150
1151 bail:
1152         exception_exit(prev_state);
1153         return rc;
1154 }
1155 EXPORT_SYMBOL_GPL(hash_page);
1156
1157 void hash_preload(struct mm_struct *mm, unsigned long ea,
1158                   unsigned long access, unsigned long trap)
1159 {
1160         int hugepage_shift;
1161         unsigned long vsid;
1162         pgd_t *pgdir;
1163         pte_t *ptep;
1164         unsigned long flags;
1165         int rc, ssize, local = 0;
1166
1167         BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1168
1169 #ifdef CONFIG_PPC_MM_SLICES
1170         /* We only prefault standard pages for now */
1171         if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1172                 return;
1173 #endif
1174
1175         DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1176                 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1177
1178         /* Get Linux PTE if available */
1179         pgdir = mm->pgd;
1180         if (pgdir == NULL)
1181                 return;
1182
1183         /* Get VSID */
1184         ssize = user_segment_size(ea);
1185         vsid = get_vsid(mm->context.id, ea, ssize);
1186         if (!vsid)
1187                 return;
1188         /*
1189          * Hash doesn't like irqs. Walking linux page table with irq disabled
1190          * saves us from holding multiple locks.
1191          */
1192         local_irq_save(flags);
1193
1194         /*
1195          * THP pages use update_mmu_cache_pmd. We don't do
1196          * hash preload there. Hence can ignore THP here
1197          */
1198         ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugepage_shift);
1199         if (!ptep)
1200                 goto out_exit;
1201
1202         WARN_ON(hugepage_shift);
1203 #ifdef CONFIG_PPC_64K_PAGES
1204         /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1205          * a 64K kernel), then we don't preload, hash_page() will take
1206          * care of it once we actually try to access the page.
1207          * That way we don't have to duplicate all of the logic for segment
1208          * page size demotion here
1209          */
1210         if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1211                 goto out_exit;
1212 #endif /* CONFIG_PPC_64K_PAGES */
1213
1214         /* Is that local to this CPU ? */
1215         if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1216                 local = 1;
1217
1218         /* Hash it in */
1219 #ifdef CONFIG_PPC_HAS_HASH_64K
1220         if (mm->context.user_psize == MMU_PAGE_64K)
1221                 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1222         else
1223 #endif /* CONFIG_PPC_HAS_HASH_64K */
1224                 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1225                                     subpage_protection(mm, ea));
1226
1227         /* Dump some info in case of hash insertion failure, they should
1228          * never happen so it is really useful to know if/when they do
1229          */
1230         if (rc == -1)
1231                 hash_failure_debug(ea, access, vsid, trap, ssize,
1232                                    mm->context.user_psize,
1233                                    mm->context.user_psize,
1234                                    pte_val(*ptep));
1235 out_exit:
1236         local_irq_restore(flags);
1237 }
1238
1239 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1240  *          do not forget to update the assembly call site !
1241  */
1242 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1243                      int local)
1244 {
1245         unsigned long hash, index, shift, hidx, slot;
1246
1247         DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1248         pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1249                 hash = hpt_hash(vpn, shift, ssize);
1250                 hidx = __rpte_to_hidx(pte, index);
1251                 if (hidx & _PTEIDX_SECONDARY)
1252                         hash = ~hash;
1253                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1254                 slot += hidx & _PTEIDX_GROUP_IX;
1255                 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1256                 /*
1257                  * We use same base page size and actual psize, because we don't
1258                  * use these functions for hugepage
1259                  */
1260                 ppc_md.hpte_invalidate(slot, vpn, psize, psize, ssize, local);
1261         } pte_iterate_hashed_end();
1262
1263 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1264         /* Transactions are not aborted by tlbiel, only tlbie.
1265          * Without, syncing a page back to a block device w/ PIO could pick up
1266          * transactional data (bad!) so we force an abort here.  Before the
1267          * sync the page will be made read-only, which will flush_hash_page.
1268          * BIG ISSUE here: if the kernel uses a page from userspace without
1269          * unmapping it first, it may see the speculated version.
1270          */
1271         if (local && cpu_has_feature(CPU_FTR_TM) &&
1272             current->thread.regs &&
1273             MSR_TM_ACTIVE(current->thread.regs->msr)) {
1274                 tm_enable();
1275                 tm_abort(TM_CAUSE_TLBI);
1276         }
1277 #endif
1278 }
1279
1280 void flush_hash_range(unsigned long number, int local)
1281 {
1282         if (ppc_md.flush_hash_range)
1283                 ppc_md.flush_hash_range(number, local);
1284         else {
1285                 int i;
1286                 struct ppc64_tlb_batch *batch =
1287                         &__get_cpu_var(ppc64_tlb_batch);
1288
1289                 for (i = 0; i < number; i++)
1290                         flush_hash_page(batch->vpn[i], batch->pte[i],
1291                                         batch->psize, batch->ssize, local);
1292         }
1293 }
1294
1295 /*
1296  * low_hash_fault is called when we the low level hash code failed
1297  * to instert a PTE due to an hypervisor error
1298  */
1299 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1300 {
1301         enum ctx_state prev_state = exception_enter();
1302
1303         if (user_mode(regs)) {
1304 #ifdef CONFIG_PPC_SUBPAGE_PROT
1305                 if (rc == -2)
1306                         _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1307                 else
1308 #endif
1309                         _exception(SIGBUS, regs, BUS_ADRERR, address);
1310         } else
1311                 bad_page_fault(regs, address, SIGBUS);
1312
1313         exception_exit(prev_state);
1314 }
1315
1316 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1317                            unsigned long pa, unsigned long rflags,
1318                            unsigned long vflags, int psize, int ssize)
1319 {
1320         unsigned long hpte_group;
1321         long slot;
1322
1323 repeat:
1324         hpte_group = ((hash & htab_hash_mask) *
1325                        HPTES_PER_GROUP) & ~0x7UL;
1326
1327         /* Insert into the hash table, primary slot */
1328         slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1329                                   psize, psize, ssize);
1330
1331         /* Primary is full, try the secondary */
1332         if (unlikely(slot == -1)) {
1333                 hpte_group = ((~hash & htab_hash_mask) *
1334                               HPTES_PER_GROUP) & ~0x7UL;
1335                 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags,
1336                                           vflags | HPTE_V_SECONDARY,
1337                                           psize, psize, ssize);
1338                 if (slot == -1) {
1339                         if (mftb() & 0x1)
1340                                 hpte_group = ((hash & htab_hash_mask) *
1341                                               HPTES_PER_GROUP)&~0x7UL;
1342
1343                         ppc_md.hpte_remove(hpte_group);
1344                         goto repeat;
1345                 }
1346         }
1347
1348         return slot;
1349 }
1350
1351 #ifdef CONFIG_DEBUG_PAGEALLOC
1352 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1353 {
1354         unsigned long hash;
1355         unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1356         unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1357         unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1358         long ret;
1359
1360         hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1361
1362         /* Don't create HPTE entries for bad address */
1363         if (!vsid)
1364                 return;
1365
1366         ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1367                                     HPTE_V_BOLTED,
1368                                     mmu_linear_psize, mmu_kernel_ssize);
1369
1370         BUG_ON (ret < 0);
1371         spin_lock(&linear_map_hash_lock);
1372         BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1373         linear_map_hash_slots[lmi] = ret | 0x80;
1374         spin_unlock(&linear_map_hash_lock);
1375 }
1376
1377 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1378 {
1379         unsigned long hash, hidx, slot;
1380         unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1381         unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1382
1383         hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1384         spin_lock(&linear_map_hash_lock);
1385         BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1386         hidx = linear_map_hash_slots[lmi] & 0x7f;
1387         linear_map_hash_slots[lmi] = 0;
1388         spin_unlock(&linear_map_hash_lock);
1389         if (hidx & _PTEIDX_SECONDARY)
1390                 hash = ~hash;
1391         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1392         slot += hidx & _PTEIDX_GROUP_IX;
1393         ppc_md.hpte_invalidate(slot, vpn, mmu_linear_psize, mmu_linear_psize,
1394                                mmu_kernel_ssize, 0);
1395 }
1396
1397 void kernel_map_pages(struct page *page, int numpages, int enable)
1398 {
1399         unsigned long flags, vaddr, lmi;
1400         int i;
1401
1402         local_irq_save(flags);
1403         for (i = 0; i < numpages; i++, page++) {
1404                 vaddr = (unsigned long)page_address(page);
1405                 lmi = __pa(vaddr) >> PAGE_SHIFT;
1406                 if (lmi >= linear_map_hash_count)
1407                         continue;
1408                 if (enable)
1409                         kernel_map_linear_page(vaddr, lmi);
1410                 else
1411                         kernel_unmap_linear_page(vaddr, lmi);
1412         }
1413         local_irq_restore(flags);
1414 }
1415 #endif /* CONFIG_DEBUG_PAGEALLOC */
1416
1417 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
1418                                 phys_addr_t first_memblock_size)
1419 {
1420         /* We don't currently support the first MEMBLOCK not mapping 0
1421          * physical on those processors
1422          */
1423         BUG_ON(first_memblock_base != 0);
1424
1425         /* On LPAR systems, the first entry is our RMA region,
1426          * non-LPAR 64-bit hash MMU systems don't have a limitation
1427          * on real mode access, but using the first entry works well
1428          * enough. We also clamp it to 1G to avoid some funky things
1429          * such as RTAS bugs etc...
1430          */
1431         ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1432
1433         /* Finally limit subsequent allocations */
1434         memblock_set_current_limit(ppc64_rma_size);
1435 }