arch/powerpc/mm/hugetlbpage.c

   1 /*
   2  * PPC Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2003 David Gibson, IBM Corporation.
   5  * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
   6  *
   7  * Based on the IA-32 version:
   8  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/io.h>
  13 #include <linux/slab.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/export.h>
  16 #include <linux/of_fdt.h>
  17 #include <linux/memblock.h>
  18 #include <linux/moduleparam.h>
  19 #include <linux/swap.h>
  20 #include <linux/swapops.h>
  21 #include <linux/kmemleak.h>
  22 #include <asm/pgtable.h>
  23 #include <asm/pgalloc.h>
  24 #include <asm/tlb.h>
  25 #include <asm/setup.h>
  26 #include <asm/hugetlb.h>
  27 #include <asm/pte-walk.h>
  28
  29 bool hugetlb_disabled = false;
  30
  31 #define hugepd_none(hpd)        (hpd_val(hpd) == 0)
  32
  33 #define PTE_T_ORDER     (__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
  34
  35 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  36 {
  37         /*
  38          * Only called for hugetlbfs pages, hence can ignore THP and the
  39          * irq disabled walk.
  40          */
  41         return __find_linux_pte(mm->pgd, addr, NULL, NULL);
  42 }
  43
  44 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  45                            unsigned long address, unsigned int pdshift,
  46                            unsigned int pshift, spinlock_t *ptl)
  47 {
  48         struct kmem_cache *cachep;
  49         pte_t *new;
  50         int i;
  51         int num_hugepd;
  52
  53         if (pshift >= pdshift) {
  54                 cachep = PGT_CACHE(PTE_T_ORDER);
  55                 num_hugepd = 1 << (pshift - pdshift);
  56         } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
  57                 cachep = PGT_CACHE(PTE_INDEX_SIZE);
  58                 num_hugepd = 1;
  59         } else {
  60                 cachep = PGT_CACHE(pdshift - pshift);
  61                 num_hugepd = 1;
  62         }
  63
  64         new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  65
  66         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  67         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  68
  69         if (! new)
  70                 return -ENOMEM;
  71
  72         /*
  73          * Make sure other cpus find the hugepd set only after a
  74          * properly initialized page table is visible to them.
  75          * For more details look for comment in __pte_alloc().
  76          */
  77         smp_wmb();
  78
  79         spin_lock(ptl);
  80         /*
  81          * We have multiple higher-level entries that point to the same
  82          * actual pte location.  Fill in each as we go and backtrack on error.
  83          * We need all of these so the DTLB pgtable walk code can find the
  84          * right higher-level entry without knowing if it's a hugepage or not.
  85          */
  86         for (i = 0; i < num_hugepd; i++, hpdp++) {
  87                 if (unlikely(!hugepd_none(*hpdp)))
  88                         break;
  89                 hugepd_populate(hpdp, new, pshift);
  90         }
  91         /* If we bailed from the for loop early, an error occurred, clean up */
  92         if (i < num_hugepd) {
  93                 for (i = i - 1 ; i >= 0; i--, hpdp--)
  94                         *hpdp = __hugepd(0);
  95                 kmem_cache_free(cachep, new);
  96         } else {
  97                 kmemleak_ignore(new);
  98         }
  99         spin_unlock(ptl);
 100         return 0;
 101 }
 102
 103 /*
 104  * At this point we do the placement change only for BOOK3S 64. This would
 105  * possibly work on other subarchs.
 106  */
 107 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 108 {
 109         pgd_t *pg;
 110         pud_t *pu;
 111         pmd_t *pm;
 112         hugepd_t *hpdp = NULL;
 113         unsigned pshift = __ffs(sz);
 114         unsigned pdshift = PGDIR_SHIFT;
 115         spinlock_t *ptl;
 116
 117         addr &= ~(sz-1);
 118         pg = pgd_offset(mm, addr);
 119
 120 #ifdef CONFIG_PPC_BOOK3S_64
 121         if (pshift == PGDIR_SHIFT)
 122                 /* 16GB huge page */
 123                 return (pte_t *) pg;
 124         else if (pshift > PUD_SHIFT) {
 125                 /*
 126                  * We need to use hugepd table
 127                  */
 128                 ptl = &mm->page_table_lock;
 129                 hpdp = (hugepd_t *)pg;
 130         } else {
 131                 pdshift = PUD_SHIFT;
 132                 pu = pud_alloc(mm, pg, addr);
 133                 if (!pu)
 134                         return NULL;
 135                 if (pshift == PUD_SHIFT)
 136                         return (pte_t *)pu;
 137                 else if (pshift > PMD_SHIFT) {
 138                         ptl = pud_lockptr(mm, pu);
 139                         hpdp = (hugepd_t *)pu;
 140                 } else {
 141                         pdshift = PMD_SHIFT;
 142                         pm = pmd_alloc(mm, pu, addr);
 143                         if (!pm)
 144                                 return NULL;
 145                         if (pshift == PMD_SHIFT)
 146                                 /* 16MB hugepage */
 147                                 return (pte_t *)pm;
 148                         else {
 149                                 ptl = pmd_lockptr(mm, pm);
 150                                 hpdp = (hugepd_t *)pm;
 151                         }
 152                 }
 153         }
 154 #else
 155         if (pshift >= PGDIR_SHIFT) {
 156                 ptl = &mm->page_table_lock;
 157                 hpdp = (hugepd_t *)pg;
 158         } else {
 159                 pdshift = PUD_SHIFT;
 160                 pu = pud_alloc(mm, pg, addr);
 161                 if (!pu)
 162                         return NULL;
 163                 if (pshift >= PUD_SHIFT) {
 164                         ptl = pud_lockptr(mm, pu);
 165                         hpdp = (hugepd_t *)pu;
 166                 } else {
 167                         pdshift = PMD_SHIFT;
 168                         pm = pmd_alloc(mm, pu, addr);
 169                         if (!pm)
 170                                 return NULL;
 171                         ptl = pmd_lockptr(mm, pm);
 172                         hpdp = (hugepd_t *)pm;
 173                 }
 174         }
 175 #endif
 176         if (!hpdp)
 177                 return NULL;
 178
 179         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 180
 181         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
 182                                                   pdshift, pshift, ptl))
 183                 return NULL;
 184
 185         return hugepte_offset(*hpdp, addr, pdshift);
 186 }
 187
 188 #ifdef CONFIG_PPC_BOOK3S_64
 189 /*
 190  * Tracks gpages after the device tree is scanned and before the
 191  * huge_boot_pages list is ready on pseries.
 192  */
 193 #define MAX_NUMBER_GPAGES       1024
 194 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 195 __initdata static unsigned nr_gpages;
 196
 197 /*
 198  * Build list of addresses of gigantic pages.  This function is used in early
 199  * boot before the buddy allocator is setup.
 200  */
 201 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 202 {
 203         if (!addr)
 204                 return;
 205         while (number_of_pages > 0) {
 206                 gpage_freearray[nr_gpages] = addr;
 207                 nr_gpages++;
 208                 number_of_pages--;
 209                 addr += page_size;
 210         }
 211 }
 212
 213 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 214 {
 215         struct huge_bootmem_page *m;
 216         if (nr_gpages == 0)
 217                 return 0;
 218         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 219         gpage_freearray[nr_gpages] = 0;
 220         list_add(&m->list, &huge_boot_pages);
 221         m->hstate = hstate;
 222         return 1;
 223 }
 224 #endif
 225
 226
 227 int __init alloc_bootmem_huge_page(struct hstate *h)
 228 {
 229
 230 #ifdef CONFIG_PPC_BOOK3S_64
 231         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
 232                 return pseries_alloc_bootmem_huge_page(h);
 233 #endif
 234         return __alloc_bootmem_huge_page(h);
 235 }
 236
 237 #ifndef CONFIG_PPC_BOOK3S_64
 238 #define HUGEPD_FREELIST_SIZE \
 239         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 240
 241 struct hugepd_freelist {
 242         struct rcu_head rcu;
 243         unsigned int index;
 244         void *ptes[0];
 245 };
 246
 247 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 248
 249 static void hugepd_free_rcu_callback(struct rcu_head *head)
 250 {
 251         struct hugepd_freelist *batch =
 252                 container_of(head, struct hugepd_freelist, rcu);
 253         unsigned int i;
 254
 255         for (i = 0; i < batch->index; i++)
 256                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
 257
 258         free_page((unsigned long)batch);
 259 }
 260
 261 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 262 {
 263         struct hugepd_freelist **batchp;
 264
 265         batchp = &get_cpu_var(hugepd_freelist_cur);
 266
 267         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 268             mm_is_thread_local(tlb->mm)) {
 269                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
 270                 put_cpu_var(hugepd_freelist_cur);
 271                 return;
 272         }
 273
 274         if (*batchp == NULL) {
 275                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 276                 (*batchp)->index = 0;
 277         }
 278
 279         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 280         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 281                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
 282                 *batchp = NULL;
 283         }
 284         put_cpu_var(hugepd_freelist_cur);
 285 }
 286 #else
 287 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 288 #endif
 289
 290 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 291                               unsigned long start, unsigned long end,
 292                               unsigned long floor, unsigned long ceiling)
 293 {
 294         pte_t *hugepte = hugepd_page(*hpdp);
 295         int i;
 296
 297         unsigned long pdmask = ~((1UL << pdshift) - 1);
 298         unsigned int num_hugepd = 1;
 299         unsigned int shift = hugepd_shift(*hpdp);
 300
 301         /* Note: On fsl the hpdp may be the first of several */
 302         if (shift > pdshift)
 303                 num_hugepd = 1 << (shift - pdshift);
 304
 305         start &= pdmask;
 306         if (start < floor)
 307                 return;
 308         if (ceiling) {
 309                 ceiling &= pdmask;
 310                 if (! ceiling)
 311                         return;
 312         }
 313         if (end - 1 > ceiling - 1)
 314                 return;
 315
 316         for (i = 0; i < num_hugepd; i++, hpdp++)
 317                 *hpdp = __hugepd(0);
 318
 319         if (shift >= pdshift)
 320                 hugepd_free(tlb, hugepte);
 321         else if (IS_ENABLED(CONFIG_PPC_8xx))
 322                 pgtable_free_tlb(tlb, hugepte,
 323                                  get_hugepd_cache_index(PTE_INDEX_SIZE));
 324         else
 325                 pgtable_free_tlb(tlb, hugepte,
 326                                  get_hugepd_cache_index(pdshift - shift));
 327 }
 328
 329 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 330                                    unsigned long addr, unsigned long end,
 331                                    unsigned long floor, unsigned long ceiling)
 332 {
 333         pmd_t *pmd;
 334         unsigned long next;
 335         unsigned long start;
 336
 337         start = addr;
 338         do {
 339                 unsigned long more;
 340
 341                 pmd = pmd_offset(pud, addr);
 342                 next = pmd_addr_end(addr, end);
 343                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 344                         /*
 345                          * if it is not hugepd pointer, we should already find
 346                          * it cleared.
 347                          */
 348                         WARN_ON(!pmd_none_or_clear_bad(pmd));
 349                         continue;
 350                 }
 351                 /*
 352                  * Increment next by the size of the huge mapping since
 353                  * there may be more than one entry at this level for a
 354                  * single hugepage, but all of them point to
 355                  * the same kmem cache that holds the hugepte.
 356                  */
 357                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 358                 if (more > next)
 359                         next = more;
 360
 361                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 362                                   addr, next, floor, ceiling);
 363         } while (addr = next, addr != end);
 364
 365         start &= PUD_MASK;
 366         if (start < floor)
 367                 return;
 368         if (ceiling) {
 369                 ceiling &= PUD_MASK;
 370                 if (!ceiling)
 371                         return;
 372         }
 373         if (end - 1 > ceiling - 1)
 374                 return;
 375
 376         pmd = pmd_offset(pud, start);
 377         pud_clear(pud);
 378         pmd_free_tlb(tlb, pmd, start);
 379         mm_dec_nr_pmds(tlb->mm);
 380 }
 381
 382 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 383                                    unsigned long addr, unsigned long end,
 384                                    unsigned long floor, unsigned long ceiling)
 385 {
 386         pud_t *pud;
 387         unsigned long next;
 388         unsigned long start;
 389
 390         start = addr;
 391         do {
 392                 pud = pud_offset(pgd, addr);
 393                 next = pud_addr_end(addr, end);
 394                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 395                         if (pud_none_or_clear_bad(pud))
 396                                 continue;
 397                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 398                                                ceiling);
 399                 } else {
 400                         unsigned long more;
 401                         /*
 402                          * Increment next by the size of the huge mapping since
 403                          * there may be more than one entry at this level for a
 404                          * single hugepage, but all of them point to
 405                          * the same kmem cache that holds the hugepte.
 406                          */
 407                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 408                         if (more > next)
 409                                 next = more;
 410
 411                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 412                                           addr, next, floor, ceiling);
 413                 }
 414         } while (addr = next, addr != end);
 415
 416         start &= PGDIR_MASK;
 417         if (start < floor)
 418                 return;
 419         if (ceiling) {
 420                 ceiling &= PGDIR_MASK;
 421                 if (!ceiling)
 422                         return;
 423         }
 424         if (end - 1 > ceiling - 1)
 425                 return;
 426
 427         pud = pud_offset(pgd, start);
 428         pgd_clear(pgd);
 429         pud_free_tlb(tlb, pud, start);
 430         mm_dec_nr_puds(tlb->mm);
 431 }
 432
 433 /*
 434  * This function frees user-level page tables of a process.
 435  */
 436 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 437                             unsigned long addr, unsigned long end,
 438                             unsigned long floor, unsigned long ceiling)
 439 {
 440         pgd_t *pgd;
 441         unsigned long next;
 442
 443         /*
 444          * Because there are a number of different possible pagetable
 445          * layouts for hugepage ranges, we limit knowledge of how
 446          * things should be laid out to the allocation path
 447          * (huge_pte_alloc(), above).  Everything else works out the
 448          * structure as it goes from information in the hugepd
 449          * pointers.  That means that we can't here use the
 450          * optimization used in the normal page free_pgd_range(), of
 451          * checking whether we're actually covering a large enough
 452          * range to have to do anything at the top level of the walk
 453          * instead of at the bottom.
 454          *
 455          * To make sense of this, you should probably go read the big
 456          * block comment at the top of the normal free_pgd_range(),
 457          * too.
 458          */
 459
 460         do {
 461                 next = pgd_addr_end(addr, end);
 462                 pgd = pgd_offset(tlb->mm, addr);
 463                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 464                         if (pgd_none_or_clear_bad(pgd))
 465                                 continue;
 466                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 467                 } else {
 468                         unsigned long more;
 469                         /*
 470                          * Increment next by the size of the huge mapping since
 471                          * there may be more than one entry at the pgd level
 472                          * for a single hugepage, but all of them point to the
 473                          * same kmem cache that holds the hugepte.
 474                          */
 475                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 476                         if (more > next)
 477                                 next = more;
 478
 479                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 480                                           addr, next, floor, ceiling);
 481                 }
 482         } while (addr = next, addr != end);
 483 }
 484
 485 struct page *follow_huge_pd(struct vm_area_struct *vma,
 486                             unsigned long address, hugepd_t hpd,
 487                             int flags, int pdshift)
 488 {
 489         pte_t *ptep;
 490         spinlock_t *ptl;
 491         struct page *page = NULL;
 492         unsigned long mask;
 493         int shift = hugepd_shift(hpd);
 494         struct mm_struct *mm = vma->vm_mm;
 495
 496 retry:
 497         /*
 498          * hugepage directory entries are protected by mm->page_table_lock
 499          * Use this instead of huge_pte_lockptr
 500          */
 501         ptl = &mm->page_table_lock;
 502         spin_lock(ptl);
 503
 504         ptep = hugepte_offset(hpd, address, pdshift);
 505         if (pte_present(*ptep)) {
 506                 mask = (1UL << shift) - 1;
 507                 page = pte_page(*ptep);
 508                 page += ((address & mask) >> PAGE_SHIFT);
 509                 if (flags & FOLL_GET)
 510                         get_page(page);
 511         } else {
 512                 if (is_hugetlb_entry_migration(*ptep)) {
 513                         spin_unlock(ptl);
 514                         __migration_entry_wait(mm, ptep, ptl);
 515                         goto retry;
 516                 }
 517         }
 518         spin_unlock(ptl);
 519         return page;
 520 }
 521
 522 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 523                                       unsigned long sz)
 524 {
 525         unsigned long __boundary = (addr + sz) & ~(sz-1);
 526         return (__boundary - 1 < end - 1) ? __boundary : end;
 527 }
 528
 529 #ifdef CONFIG_PPC_MM_SLICES
 530 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 531                                         unsigned long len, unsigned long pgoff,
 532                                         unsigned long flags)
 533 {
 534         struct hstate *hstate = hstate_file(file);
 535         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 536
 537 #ifdef CONFIG_PPC_RADIX_MMU
 538         if (radix_enabled())
 539                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 540                                                        pgoff, flags);
 541 #endif
 542         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 543 }
 544 #endif
 545
 546 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 547 {
 548         /* With radix we don't use slice, so derive it from vma*/
 549         if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
 550                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 551
 552                 return 1UL << mmu_psize_to_shift(psize);
 553         }
 554         return vma_kernel_pagesize(vma);
 555 }
 556
 557 static int __init add_huge_page_size(unsigned long long size)
 558 {
 559         int shift = __ffs(size);
 560         int mmu_psize;
 561
 562         /* Check that it is a page size supported by the hardware and
 563          * that it fits within pagetable and slice limits. */
 564         if (size <= PAGE_SIZE || !is_power_of_2(size))
 565                 return -EINVAL;
 566
 567         mmu_psize = check_and_get_huge_psize(shift);
 568         if (mmu_psize < 0)
 569                 return -EINVAL;
 570
 571         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 572
 573         /* Return if huge page size has already been setup */
 574         if (size_to_hstate(size))
 575                 return 0;
 576
 577         hugetlb_add_hstate(shift - PAGE_SHIFT);
 578
 579         return 0;
 580 }
 581
 582 static int __init hugepage_setup_sz(char *str)
 583 {
 584         unsigned long long size;
 585
 586         size = memparse(str, &str);
 587
 588         if (add_huge_page_size(size) != 0) {
 589                 hugetlb_bad_size();
 590                 pr_err("Invalid huge page size specified(%llu)\n", size);
 591         }
 592
 593         return 1;
 594 }
 595 __setup("hugepagesz=", hugepage_setup_sz);
 596
 597 static int __init hugetlbpage_init(void)
 598 {
 599         int psize;
 600
 601         if (hugetlb_disabled) {
 602                 pr_info("HugeTLB support is disabled!\n");
 603                 return 0;
 604         }
 605
 606         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
 607             !mmu_has_feature(MMU_FTR_16M_PAGE))
 608                 return -ENODEV;
 609
 610         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 611                 unsigned shift;
 612                 unsigned pdshift;
 613
 614                 if (!mmu_psize_defs[psize].shift)
 615                         continue;
 616
 617                 shift = mmu_psize_to_shift(psize);
 618
 619 #ifdef CONFIG_PPC_BOOK3S_64
 620                 if (shift > PGDIR_SHIFT)
 621                         continue;
 622                 else if (shift > PUD_SHIFT)
 623                         pdshift = PGDIR_SHIFT;
 624                 else if (shift > PMD_SHIFT)
 625                         pdshift = PUD_SHIFT;
 626                 else
 627                         pdshift = PMD_SHIFT;
 628 #else
 629                 if (shift < PUD_SHIFT)
 630                         pdshift = PMD_SHIFT;
 631                 else if (shift < PGDIR_SHIFT)
 632                         pdshift = PUD_SHIFT;
 633                 else
 634                         pdshift = PGDIR_SHIFT;
 635 #endif
 636
 637                 if (add_huge_page_size(1ULL << shift) < 0)
 638                         continue;
 639                 /*
 640                  * if we have pdshift and shift value same, we don't
 641                  * use pgt cache for hugepd.
 642                  */
 643                 if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
 644                         pgtable_cache_add(PTE_INDEX_SIZE);
 645                 else if (pdshift > shift)
 646                         pgtable_cache_add(pdshift - shift);
 647                 else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) || IS_ENABLED(CONFIG_PPC_8xx))
 648                         pgtable_cache_add(PTE_T_ORDER);
 649         }
 650
 651         if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
 652                 hugetlbpage_init_default();
 653
 654         return 0;
 655 }
 656
 657 arch_initcall(hugetlbpage_init);
 658
 659 void flush_dcache_icache_hugepage(struct page *page)
 660 {
 661         int i;
 662         void *start;
 663
 664         BUG_ON(!PageCompound(page));
 665
 666         for (i = 0; i < (1UL << compound_order(page)); i++) {
 667                 if (!PageHighMem(page)) {
 668                         __flush_dcache_icache(page_address(page+i));
 669                 } else {
 670                         start = kmap_atomic(page+i);
 671                         __flush_dcache_icache(start);
 672                         kunmap_atomic(start);
 673                 }
 674         }
 675 }
 676
 677 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 678                        unsigned long end, int write, struct page **pages, int *nr)
 679 {
 680         unsigned long pte_end;
 681         struct page *head, *page;
 682         pte_t pte;
 683         int refs;
 684
 685         pte_end = (addr + sz) & ~(sz-1);
 686         if (pte_end < end)
 687                 end = pte_end;
 688
 689         pte = READ_ONCE(*ptep);
 690
 691         if (!pte_access_permitted(pte, write))
 692                 return 0;
 693
 694         /* hugepages are never "special" */
 695         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 696
 697         refs = 0;
 698         head = pte_page(pte);
 699
 700         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
 701         do {
 702                 VM_BUG_ON(compound_head(page) != head);
 703                 pages[*nr] = page;
 704                 (*nr)++;
 705                 page++;
 706                 refs++;
 707         } while (addr += PAGE_SIZE, addr != end);
 708
 709         if (!page_cache_add_speculative(head, refs)) {
 710                 *nr -= refs;
 711                 return 0;
 712         }
 713
 714         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 715                 /* Could be optimized better */
 716                 *nr -= refs;
 717                 while (refs--)
 718                         put_page(head);
 719                 return 0;
 720         }
 721
 722         return 1;
 723 }
 724
 725 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned int pdshift,
 726                 unsigned long end, int write, struct page **pages, int *nr)
 727 {
 728         pte_t *ptep;
 729         unsigned long sz = 1UL << hugepd_shift(hugepd);
 730         unsigned long next;
 731
 732         ptep = hugepte_offset(hugepd, addr, pdshift);
 733         do {
 734                 next = hugepte_addr_end(addr, end, sz);
 735                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 736                         return 0;
 737         } while (ptep++, addr = next, addr != end);
 738
 739         return 1;
 740 }