arch/powerpc/mm/hugetlbpage.c

   1 /*
   2  * PPC64 (POWER4) Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2003 David Gibson, IBM Corporation.
   5  *
   6  * Based on the IA-32 version:
   7  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/slab.h>
  16 #include <linux/err.h>
  17 #include <linux/sysctl.h>
  18 #include <asm/mman.h>
  19 #include <asm/pgalloc.h>
  20 #include <asm/tlb.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/mmu_context.h>
  23 #include <asm/machdep.h>
  24 #include <asm/cputable.h>
  25 #include <asm/spu.h>
  26
  27 #define PAGE_SHIFT_64K  16
  28 #define PAGE_SHIFT_16M  24
  29 #define PAGE_SHIFT_16G  34
  30
  31 #define NUM_LOW_AREAS   (0x100000000UL >> SID_SHIFT)
  32 #define NUM_HIGH_AREAS  (PGTABLE_RANGE >> HTLB_AREA_SHIFT)
  33 #define MAX_NUMBER_GPAGES       1024
  34
  35 /* Tracks the 16G pages after the device tree is scanned and before the
  36  * huge_boot_pages list is ready.  */
  37 static unsigned long gpage_freearray[MAX_NUMBER_GPAGES];
  38 static unsigned nr_gpages;
  39
  40 /* Array of valid huge page sizes - non-zero value(hugepte_shift) is
  41  * stored for the huge page sizes that are valid.
  42  */
  43 unsigned int mmu_huge_psizes[MMU_PAGE_COUNT] = { }; /* initialize all to 0 */
  44
  45 #define hugepte_shift                   mmu_huge_psizes
  46 #define PTRS_PER_HUGEPTE(psize)         (1 << hugepte_shift[psize])
  47 #define HUGEPTE_TABLE_SIZE(psize)       (sizeof(pte_t) << hugepte_shift[psize])
  48
  49 #define HUGEPD_SHIFT(psize)             (mmu_psize_to_shift(psize) \
  50                                                 + hugepte_shift[psize])
  51 #define HUGEPD_SIZE(psize)              (1UL << HUGEPD_SHIFT(psize))
  52 #define HUGEPD_MASK(psize)              (~(HUGEPD_SIZE(psize)-1))
  53
  54 /* Subtract one from array size because we don't need a cache for 4K since
  55  * is not a huge page size */
  56 #define huge_pgtable_cache(psize)       (pgtable_cache[HUGEPTE_CACHE_NUM \
  57                                                         + psize-1])
  58 #define HUGEPTE_CACHE_NAME(psize)       (huge_pgtable_cache_name[psize])
  59
  60 static const char *huge_pgtable_cache_name[MMU_PAGE_COUNT] = {
  61         "unused_4K", "hugepte_cache_64K", "unused_64K_AP",
  62         "hugepte_cache_1M", "hugepte_cache_16M", "hugepte_cache_16G"
  63 };
  64
  65 /* Flag to mark huge PD pointers.  This means pmd_bad() and pud_bad()
  66  * will choke on pointers to hugepte tables, which is handy for
  67  * catching screwups early. */
  68 #define HUGEPD_OK       0x1
  69
  70 typedef struct { unsigned long pd; } hugepd_t;
  71
  72 #define hugepd_none(hpd)        ((hpd).pd == 0)
  73
  74 static inline int shift_to_mmu_psize(unsigned int shift)
  75 {
  76         switch (shift) {
  77 #ifndef CONFIG_PPC_64K_PAGES
  78         case PAGE_SHIFT_64K:
  79             return MMU_PAGE_64K;
  80 #endif
  81         case PAGE_SHIFT_16M:
  82             return MMU_PAGE_16M;
  83         case PAGE_SHIFT_16G:
  84             return MMU_PAGE_16G;
  85         }
  86         return -1;
  87 }
  88
  89 static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
  90 {
  91         if (mmu_psize_defs[mmu_psize].shift)
  92                 return mmu_psize_defs[mmu_psize].shift;
  93         BUG();
  94 }
  95
  96 static inline pte_t *hugepd_page(hugepd_t hpd)
  97 {
  98         BUG_ON(!(hpd.pd & HUGEPD_OK));
  99         return (pte_t *)(hpd.pd & ~HUGEPD_OK);
 100 }
 101
 102 static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr,
 103                                     struct hstate *hstate)
 104 {
 105         unsigned int shift = huge_page_shift(hstate);
 106         int psize = shift_to_mmu_psize(shift);
 107         unsigned long idx = ((addr >> shift) & (PTRS_PER_HUGEPTE(psize)-1));
 108         pte_t *dir = hugepd_page(*hpdp);
 109
 110         return dir + idx;
 111 }
 112
 113 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
 114                            unsigned long address, unsigned int psize)
 115 {
 116         pte_t *new = kmem_cache_zalloc(huge_pgtable_cache(psize),
 117                                       GFP_KERNEL|__GFP_REPEAT);
 118
 119         if (! new)
 120                 return -ENOMEM;
 121
 122         spin_lock(&mm->page_table_lock);
 123         if (!hugepd_none(*hpdp))
 124                 kmem_cache_free(huge_pgtable_cache(psize), new);
 125         else
 126                 hpdp->pd = (unsigned long)new | HUGEPD_OK;
 127         spin_unlock(&mm->page_table_lock);
 128         return 0;
 129 }
 130
 131
 132 static pud_t *hpud_offset(pgd_t *pgd, unsigned long addr, struct hstate *hstate)
 133 {
 134         if (huge_page_shift(hstate) < PUD_SHIFT)
 135                 return pud_offset(pgd, addr);
 136         else
 137                 return (pud_t *) pgd;
 138 }
 139 static pud_t *hpud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long addr,
 140                          struct hstate *hstate)
 141 {
 142         if (huge_page_shift(hstate) < PUD_SHIFT)
 143                 return pud_alloc(mm, pgd, addr);
 144         else
 145                 return (pud_t *) pgd;
 146 }
 147 static pmd_t *hpmd_offset(pud_t *pud, unsigned long addr, struct hstate *hstate)
 148 {
 149         if (huge_page_shift(hstate) < PMD_SHIFT)
 150                 return pmd_offset(pud, addr);
 151         else
 152                 return (pmd_t *) pud;
 153 }
 154 static pmd_t *hpmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long addr,
 155                          struct hstate *hstate)
 156 {
 157         if (huge_page_shift(hstate) < PMD_SHIFT)
 158                 return pmd_alloc(mm, pud, addr);
 159         else
 160                 return (pmd_t *) pud;
 161 }
 162
 163 /* Build list of addresses of gigantic pages.  This function is used in early
 164  * boot before the buddy or bootmem allocator is setup.
 165  */
 166 void add_gpage(unsigned long addr, unsigned long page_size,
 167         unsigned long number_of_pages)
 168 {
 169         if (!addr)
 170                 return;
 171         while (number_of_pages > 0) {
 172                 gpage_freearray[nr_gpages] = addr;
 173                 nr_gpages++;
 174                 number_of_pages--;
 175                 addr += page_size;
 176         }
 177 }
 178
 179 /* Moves the gigantic page addresses from the temporary list to the
 180  * huge_boot_pages list.
 181  */
 182 int alloc_bootmem_huge_page(struct hstate *hstate)
 183 {
 184         struct huge_bootmem_page *m;
 185         if (nr_gpages == 0)
 186                 return 0;
 187         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 188         gpage_freearray[nr_gpages] = 0;
 189         list_add(&m->list, &huge_boot_pages);
 190         m->hstate = hstate;
 191         return 1;
 192 }
 193
 194
 195 /* Modelled after find_linux_pte() */
 196 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 197 {
 198         pgd_t *pg;
 199         pud_t *pu;
 200         pmd_t *pm;
 201
 202         unsigned int psize;
 203         unsigned int shift;
 204         unsigned long sz;
 205         struct hstate *hstate;
 206         psize = get_slice_psize(mm, addr);
 207         shift = mmu_psize_to_shift(psize);
 208         sz = ((1UL) << shift);
 209         hstate = size_to_hstate(sz);
 210
 211         addr &= hstate->mask;
 212
 213         pg = pgd_offset(mm, addr);
 214         if (!pgd_none(*pg)) {
 215                 pu = hpud_offset(pg, addr, hstate);
 216                 if (!pud_none(*pu)) {
 217                         pm = hpmd_offset(pu, addr, hstate);
 218                         if (!pmd_none(*pm))
 219                                 return hugepte_offset((hugepd_t *)pm, addr,
 220                                                       hstate);
 221                 }
 222         }
 223
 224         return NULL;
 225 }
 226
 227 pte_t *huge_pte_alloc(struct mm_struct *mm,
 228                         unsigned long addr, unsigned long sz)
 229 {
 230         pgd_t *pg;
 231         pud_t *pu;
 232         pmd_t *pm;
 233         hugepd_t *hpdp = NULL;
 234         struct hstate *hstate;
 235         unsigned int psize;
 236         hstate = size_to_hstate(sz);
 237
 238         psize = get_slice_psize(mm, addr);
 239         BUG_ON(!mmu_huge_psizes[psize]);
 240
 241         addr &= hstate->mask;
 242
 243         pg = pgd_offset(mm, addr);
 244         pu = hpud_alloc(mm, pg, addr, hstate);
 245
 246         if (pu) {
 247                 pm = hpmd_alloc(mm, pu, addr, hstate);
 248                 if (pm)
 249                         hpdp = (hugepd_t *)pm;
 250         }
 251
 252         if (! hpdp)
 253                 return NULL;
 254
 255         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, psize))
 256                 return NULL;
 257
 258         return hugepte_offset(hpdp, addr, hstate);
 259 }
 260
 261 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 262 {
 263         return 0;
 264 }
 265
 266 static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp,
 267                                unsigned int psize)
 268 {
 269         pte_t *hugepte = hugepd_page(*hpdp);
 270
 271         hpdp->pd = 0;
 272         tlb->need_flush = 1;
 273         pgtable_free_tlb(tlb, pgtable_free_cache(hugepte,
 274                                                  HUGEPTE_CACHE_NUM+psize-1,
 275                                                  PGF_CACHENUM_MASK));
 276 }
 277
 278 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 279                                    unsigned long addr, unsigned long end,
 280                                    unsigned long floor, unsigned long ceiling,
 281                                    unsigned int psize)
 282 {
 283         pmd_t *pmd;
 284         unsigned long next;
 285         unsigned long start;
 286
 287         start = addr;
 288         pmd = pmd_offset(pud, addr);
 289         do {
 290                 next = pmd_addr_end(addr, end);
 291                 if (pmd_none(*pmd))
 292                         continue;
 293                 free_hugepte_range(tlb, (hugepd_t *)pmd, psize);
 294         } while (pmd++, addr = next, addr != end);
 295
 296         start &= PUD_MASK;
 297         if (start < floor)
 298                 return;
 299         if (ceiling) {
 300                 ceiling &= PUD_MASK;
 301                 if (!ceiling)
 302                         return;
 303         }
 304         if (end - 1 > ceiling - 1)
 305                 return;
 306
 307         pmd = pmd_offset(pud, start);
 308         pud_clear(pud);
 309         pmd_free_tlb(tlb, pmd);
 310 }
 311
 312 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 313                                    unsigned long addr, unsigned long end,
 314                                    unsigned long floor, unsigned long ceiling)
 315 {
 316         pud_t *pud;
 317         unsigned long next;
 318         unsigned long start;
 319         unsigned int shift;
 320         unsigned int psize = get_slice_psize(tlb->mm, addr);
 321         shift = mmu_psize_to_shift(psize);
 322
 323         start = addr;
 324         pud = pud_offset(pgd, addr);
 325         do {
 326                 next = pud_addr_end(addr, end);
 327                 if (shift < PMD_SHIFT) {
 328                         if (pud_none_or_clear_bad(pud))
 329                                 continue;
 330                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 331                                                ceiling, psize);
 332                 } else {
 333                         if (pud_none(*pud))
 334                                 continue;
 335                         free_hugepte_range(tlb, (hugepd_t *)pud, psize);
 336                 }
 337         } while (pud++, addr = next, addr != end);
 338
 339         start &= PGDIR_MASK;
 340         if (start < floor)
 341                 return;
 342         if (ceiling) {
 343                 ceiling &= PGDIR_MASK;
 344                 if (!ceiling)
 345                         return;
 346         }
 347         if (end - 1 > ceiling - 1)
 348                 return;
 349
 350         pud = pud_offset(pgd, start);
 351         pgd_clear(pgd);
 352         pud_free_tlb(tlb, pud);
 353 }
 354
 355 /*
 356  * This function frees user-level page tables of a process.
 357  *
 358  * Must be called with pagetable lock held.
 359  */
 360 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 361                             unsigned long addr, unsigned long end,
 362                             unsigned long floor, unsigned long ceiling)
 363 {
 364         pgd_t *pgd;
 365         unsigned long next;
 366         unsigned long start;
 367
 368         /*
 369          * Comments below take from the normal free_pgd_range().  They
 370          * apply here too.  The tests against HUGEPD_MASK below are
 371          * essential, because we *don't* test for this at the bottom
 372          * level.  Without them we'll attempt to free a hugepte table
 373          * when we unmap just part of it, even if there are other
 374          * active mappings using it.
 375          *
 376          * The next few lines have given us lots of grief...
 377          *
 378          * Why are we testing HUGEPD* at this top level?  Because
 379          * often there will be no work to do at all, and we'd prefer
 380          * not to go all the way down to the bottom just to discover
 381          * that.
 382          *
 383          * Why all these "- 1"s?  Because 0 represents both the bottom
 384          * of the address space and the top of it (using -1 for the
 385          * top wouldn't help much: the masks would do the wrong thing).
 386          * The rule is that addr 0 and floor 0 refer to the bottom of
 387          * the address space, but end 0 and ceiling 0 refer to the top
 388          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 389          * that end 0 case should be mythical).
 390          *
 391          * Wherever addr is brought up or ceiling brought down, we
 392          * must be careful to reject "the opposite 0" before it
 393          * confuses the subsequent tests.  But what about where end is
 394          * brought down by HUGEPD_SIZE below? no, end can't go down to
 395          * 0 there.
 396          *
 397          * Whereas we round start (addr) and ceiling down, by different
 398          * masks at different levels, in order to test whether a table
 399          * now has no other vmas using it, so can be freed, we don't
 400          * bother to round floor or end up - the tests don't need that.
 401          */
 402         unsigned int psize = get_slice_psize(tlb->mm, addr);
 403
 404         addr &= HUGEPD_MASK(psize);
 405         if (addr < floor) {
 406                 addr += HUGEPD_SIZE(psize);
 407                 if (!addr)
 408                         return;
 409         }
 410         if (ceiling) {
 411                 ceiling &= HUGEPD_MASK(psize);
 412                 if (!ceiling)
 413                         return;
 414         }
 415         if (end - 1 > ceiling - 1)
 416                 end -= HUGEPD_SIZE(psize);
 417         if (addr > end - 1)
 418                 return;
 419
 420         start = addr;
 421         pgd = pgd_offset(tlb->mm, addr);
 422         do {
 423                 psize = get_slice_psize(tlb->mm, addr);
 424                 BUG_ON(!mmu_huge_psizes[psize]);
 425                 next = pgd_addr_end(addr, end);
 426                 if (mmu_psize_to_shift(psize) < PUD_SHIFT) {
 427                         if (pgd_none_or_clear_bad(pgd))
 428                                 continue;
 429                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 430                 } else {
 431                         if (pgd_none(*pgd))
 432                                 continue;
 433                         free_hugepte_range(tlb, (hugepd_t *)pgd, psize);
 434                 }
 435         } while (pgd++, addr = next, addr != end);
 436 }
 437
 438 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 439                      pte_t *ptep, pte_t pte)
 440 {
 441         if (pte_present(*ptep)) {
 442                 /* We open-code pte_clear because we need to pass the right
 443                  * argument to hpte_need_flush (huge / !huge). Might not be
 444                  * necessary anymore if we make hpte_need_flush() get the
 445                  * page size from the slices
 446                  */
 447                 unsigned int psize = get_slice_psize(mm, addr);
 448                 unsigned int shift = mmu_psize_to_shift(psize);
 449                 unsigned long sz = ((1UL) << shift);
 450                 struct hstate *hstate = size_to_hstate(sz);
 451                 pte_update(mm, addr & hstate->mask, ptep, ~0UL, 1);
 452         }
 453         *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
 454 }
 455
 456 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 457                               pte_t *ptep)
 458 {
 459         unsigned long old = pte_update(mm, addr, ptep, ~0UL, 1);
 460         return __pte(old);
 461 }
 462
 463 struct page *
 464 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 465 {
 466         pte_t *ptep;
 467         struct page *page;
 468         unsigned int mmu_psize = get_slice_psize(mm, address);
 469
 470         /* Verify it is a huge page else bail. */
 471         if (!mmu_huge_psizes[mmu_psize])
 472                 return ERR_PTR(-EINVAL);
 473
 474         ptep = huge_pte_offset(mm, address);
 475         page = pte_page(*ptep);
 476         if (page) {
 477                 unsigned int shift = mmu_psize_to_shift(mmu_psize);
 478                 unsigned long sz = ((1UL) << shift);
 479                 page += (address % sz) / PAGE_SIZE;
 480         }
 481
 482         return page;
 483 }
 484
 485 int pmd_huge(pmd_t pmd)
 486 {
 487         return 0;
 488 }
 489
 490 int pud_huge(pud_t pud)
 491 {
 492         return 0;
 493 }
 494
 495 struct page *
 496 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 497                 pmd_t *pmd, int write)
 498 {
 499         BUG();
 500         return NULL;
 501 }
 502
 503
 504 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 505                                         unsigned long len, unsigned long pgoff,
 506                                         unsigned long flags)
 507 {
 508         struct hstate *hstate = hstate_file(file);
 509         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 510         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
 511 }
 512
 513 /*
 514  * Called by asm hashtable.S for doing lazy icache flush
 515  */
 516 static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags,
 517                                         pte_t pte, int trap, unsigned long sz)
 518 {
 519         struct page *page;
 520         int i;
 521
 522         if (!pfn_valid(pte_pfn(pte)))
 523                 return rflags;
 524
 525         page = pte_page(pte);
 526
 527         /* page is dirty */
 528         if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
 529                 if (trap == 0x400) {
 530                         for (i = 0; i < (sz / PAGE_SIZE); i++)
 531                                 __flush_dcache_icache(page_address(page+i));
 532                         set_bit(PG_arch_1, &page->flags);
 533                 } else {
 534                         rflags |= HPTE_R_N;
 535                 }
 536         }
 537         return rflags;
 538 }
 539
 540 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 541                    unsigned long ea, unsigned long vsid, int local,
 542                    unsigned long trap)
 543 {
 544         pte_t *ptep;
 545         unsigned long old_pte, new_pte;
 546         unsigned long va, rflags, pa, sz;
 547         long slot;
 548         int err = 1;
 549         int ssize = user_segment_size(ea);
 550         unsigned int mmu_psize;
 551         int shift;
 552         mmu_psize = get_slice_psize(mm, ea);
 553
 554         if (!mmu_huge_psizes[mmu_psize])
 555                 goto out;
 556         ptep = huge_pte_offset(mm, ea);
 557
 558         /* Search the Linux page table for a match with va */
 559         va = hpt_va(ea, vsid, ssize);
 560
 561         /*
 562          * If no pte found or not present, send the problem up to
 563          * do_page_fault
 564          */
 565         if (unlikely(!ptep || pte_none(*ptep)))
 566                 goto out;
 567
 568         /*
 569          * Check the user's access rights to the page.  If access should be
 570          * prevented then send the problem up to do_page_fault.
 571          */
 572         if (unlikely(access & ~pte_val(*ptep)))
 573                 goto out;
 574         /*
 575          * At this point, we have a pte (old_pte) which can be used to build
 576          * or update an HPTE. There are 2 cases:
 577          *
 578          * 1. There is a valid (present) pte with no associated HPTE (this is
 579          *      the most common case)
 580          * 2. There is a valid (present) pte with an associated HPTE. The
 581          *      current values of the pp bits in the HPTE prevent access
 582          *      because we are doing software DIRTY bit management and the
 583          *      page is currently not DIRTY.
 584          */
 585
 586
 587         do {
 588                 old_pte = pte_val(*ptep);
 589                 if (old_pte & _PAGE_BUSY)
 590                         goto out;
 591                 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED;
 592         } while(old_pte != __cmpxchg_u64((unsigned long *)ptep,
 593                                          old_pte, new_pte));
 594
 595         rflags = 0x2 | (!(new_pte & _PAGE_RW));
 596         /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
 597         rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N);
 598         shift = mmu_psize_to_shift(mmu_psize);
 599         sz = ((1UL) << shift);
 600         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
 601                 /* No CPU has hugepages but lacks no execute, so we
 602                  * don't need to worry about that case */
 603                 rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte),
 604                                                        trap, sz);
 605
 606         /* Check if pte already has an hpte (case 2) */
 607         if (unlikely(old_pte & _PAGE_HASHPTE)) {
 608                 /* There MIGHT be an HPTE for this pte */
 609                 unsigned long hash, slot;
 610
 611                 hash = hpt_hash(va, shift, ssize);
 612                 if (old_pte & _PAGE_F_SECOND)
 613                         hash = ~hash;
 614                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 615                 slot += (old_pte & _PAGE_F_GIX) >> 12;
 616
 617                 if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_psize,
 618                                          ssize, local) == -1)
 619                         old_pte &= ~_PAGE_HPTEFLAGS;
 620         }
 621
 622         if (likely(!(old_pte & _PAGE_HASHPTE))) {
 623                 unsigned long hash = hpt_hash(va, shift, ssize);
 624                 unsigned long hpte_group;
 625
 626                 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
 627
 628 repeat:
 629                 hpte_group = ((hash & htab_hash_mask) *
 630                               HPTES_PER_GROUP) & ~0x7UL;
 631
 632                 /* clear HPTE slot informations in new PTE */
 633 #ifdef CONFIG_PPC_64K_PAGES
 634                 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HPTE_SUB0;
 635 #else
 636                 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
 637 #endif
 638                 /* Add in WIMG bits */
 639                 rflags |= (new_pte & (_PAGE_WRITETHRU | _PAGE_NO_CACHE |
 640                                       _PAGE_COHERENT | _PAGE_GUARDED));
 641
 642                 /* Insert into the hash table, primary slot */
 643                 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0,
 644                                           mmu_psize, ssize);
 645
 646                 /* Primary is full, try the secondary */
 647                 if (unlikely(slot == -1)) {
 648                         hpte_group = ((~hash & htab_hash_mask) *
 649                                       HPTES_PER_GROUP) & ~0x7UL;
 650                         slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags,
 651                                                   HPTE_V_SECONDARY,
 652                                                   mmu_psize, ssize);
 653                         if (slot == -1) {
 654                                 if (mftb() & 0x1)
 655                                         hpte_group = ((hash & htab_hash_mask) *
 656                                                       HPTES_PER_GROUP)&~0x7UL;
 657
 658                                 ppc_md.hpte_remove(hpte_group);
 659                                 goto repeat;
 660                         }
 661                 }
 662
 663                 if (unlikely(slot == -2))
 664                         panic("hash_huge_page: pte_insert failed\n");
 665
 666                 new_pte |= (slot << 12) & (_PAGE_F_SECOND | _PAGE_F_GIX);
 667         }
 668
 669         /*
 670          * No need to use ldarx/stdcx here
 671          */
 672         *ptep = __pte(new_pte & ~_PAGE_BUSY);
 673
 674         err = 0;
 675
 676  out:
 677         return err;
 678 }
 679
 680 static void __init set_huge_psize(int psize)
 681 {
 682         /* Check that it is a page size supported by the hardware and
 683          * that it fits within pagetable limits. */
 684         if (mmu_psize_defs[psize].shift &&
 685                 mmu_psize_defs[psize].shift < SID_SHIFT_1T &&
 686                 (mmu_psize_defs[psize].shift > MIN_HUGEPTE_SHIFT ||
 687                  mmu_psize_defs[psize].shift == PAGE_SHIFT_64K ||
 688                  mmu_psize_defs[psize].shift == PAGE_SHIFT_16G)) {
 689                 /* Return if huge page size has already been setup or is the
 690                  * same as the base page size. */
 691                 if (mmu_huge_psizes[psize] ||
 692                    mmu_psize_defs[psize].shift == PAGE_SHIFT)
 693                         return;
 694                 hugetlb_add_hstate(mmu_psize_defs[psize].shift - PAGE_SHIFT);
 695
 696                 switch (mmu_psize_defs[psize].shift) {
 697                 case PAGE_SHIFT_64K:
 698                     /* We only allow 64k hpages with 4k base page,
 699                      * which was checked above, and always put them
 700                      * at the PMD */
 701                     hugepte_shift[psize] = PMD_SHIFT;
 702                     break;
 703                 case PAGE_SHIFT_16M:
 704                     /* 16M pages can be at two different levels
 705                      * of pagestables based on base page size */
 706                     if (PAGE_SHIFT == PAGE_SHIFT_64K)
 707                             hugepte_shift[psize] = PMD_SHIFT;
 708                     else /* 4k base page */
 709                             hugepte_shift[psize] = PUD_SHIFT;
 710                     break;
 711                 case PAGE_SHIFT_16G:
 712                     /* 16G pages are always at PGD level */
 713                     hugepte_shift[psize] = PGDIR_SHIFT;
 714                     break;
 715                 }
 716                 hugepte_shift[psize] -= mmu_psize_defs[psize].shift;
 717         } else
 718                 hugepte_shift[psize] = 0;
 719 }
 720
 721 static int __init hugepage_setup_sz(char *str)
 722 {
 723         unsigned long long size;
 724         int mmu_psize;
 725         int shift;
 726
 727         size = memparse(str, &str);
 728
 729         shift = __ffs(size);
 730         mmu_psize = shift_to_mmu_psize(shift);
 731         if (mmu_psize >= 0 && mmu_psize_defs[mmu_psize].shift)
 732                 set_huge_psize(mmu_psize);
 733         else
 734                 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
 735
 736         return 1;
 737 }
 738 __setup("hugepagesz=", hugepage_setup_sz);
 739
 740 static int __init hugetlbpage_init(void)
 741 {
 742         unsigned int psize;
 743
 744         if (!cpu_has_feature(CPU_FTR_16M_PAGE))
 745                 return -ENODEV;
 746
 747         /* Add supported huge page sizes.  Need to change HUGE_MAX_HSTATE
 748          * and adjust PTE_NONCACHE_NUM if the number of supported huge page
 749          * sizes changes.
 750          */
 751         set_huge_psize(MMU_PAGE_16M);
 752         set_huge_psize(MMU_PAGE_16G);
 753
 754         /* Temporarily disable support for 64K huge pages when 64K SPU local
 755          * store support is enabled as the current implementation conflicts.
 756          */
 757 #ifndef CONFIG_SPU_FS_64K_LS
 758         set_huge_psize(MMU_PAGE_64K);
 759 #endif
 760
 761         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 762                 if (mmu_huge_psizes[psize]) {
 763                         huge_pgtable_cache(psize) = kmem_cache_create(
 764                                                 HUGEPTE_CACHE_NAME(psize),
 765                                                 HUGEPTE_TABLE_SIZE(psize),
 766                                                 HUGEPTE_TABLE_SIZE(psize),
 767                                                 0,
 768                                                 NULL);
 769                         if (!huge_pgtable_cache(psize))
 770                                 panic("hugetlbpage_init(): could not create %s"\
 771                                       "\n", HUGEPTE_CACHE_NAME(psize));
 772                 }
 773         }
 774
 775         return 0;
 776 }
 777
 778 module_init(hugetlbpage_init);