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_basic_t)) - \
  34                          __builtin_ffs(sizeof(void *)))
  35
  36 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  37 {
  38         /*
  39          * Only called for hugetlbfs pages, hence can ignore THP and the
  40          * irq disabled walk.
  41          */
  42         return __find_linux_pte(mm->pgd, addr, NULL, NULL);
  43 }
  44
  45 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  46                            unsigned long address, unsigned int pdshift,
  47                            unsigned int pshift, spinlock_t *ptl)
  48 {
  49         struct kmem_cache *cachep;
  50         pte_t *new;
  51         int i;
  52         int num_hugepd;
  53
  54         if (pshift >= pdshift) {
  55                 cachep = PGT_CACHE(PTE_T_ORDER);
  56                 num_hugepd = 1 << (pshift - pdshift);
  57                 new = NULL;
  58         } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
  59                 cachep = NULL;
  60                 num_hugepd = 1;
  61                 new = pte_alloc_one(mm);
  62         } else {
  63                 cachep = PGT_CACHE(pdshift - pshift);
  64                 num_hugepd = 1;
  65                 new = NULL;
  66         }
  67
  68         if (!cachep && !new) {
  69                 WARN_ONCE(1, "No page table cache created for hugetlb tables");
  70                 return -ENOMEM;
  71         }
  72
  73         if (cachep)
  74                 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  75
  76         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  77         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  78
  79         if (!new)
  80                 return -ENOMEM;
  81
  82         /*
  83          * Make sure other cpus find the hugepd set only after a
  84          * properly initialized page table is visible to them.
  85          * For more details look for comment in __pte_alloc().
  86          */
  87         smp_wmb();
  88
  89         spin_lock(ptl);
  90         /*
  91          * We have multiple higher-level entries that point to the same
  92          * actual pte location.  Fill in each as we go and backtrack on error.
  93          * We need all of these so the DTLB pgtable walk code can find the
  94          * right higher-level entry without knowing if it's a hugepage or not.
  95          */
  96         for (i = 0; i < num_hugepd; i++, hpdp++) {
  97                 if (unlikely(!hugepd_none(*hpdp)))
  98                         break;
  99                 hugepd_populate(hpdp, new, pshift);
 100         }
 101         /* If we bailed from the for loop early, an error occurred, clean up */
 102         if (i < num_hugepd) {
 103                 for (i = i - 1 ; i >= 0; i--, hpdp--)
 104                         *hpdp = __hugepd(0);
 105                 if (cachep)
 106                         kmem_cache_free(cachep, new);
 107                 else
 108                         pte_free(mm, new);
 109         } else {
 110                 kmemleak_ignore(new);
 111         }
 112         spin_unlock(ptl);
 113         return 0;
 114 }
 115
 116 /*
 117  * At this point we do the placement change only for BOOK3S 64. This would
 118  * possibly work on other subarchs.
 119  */
 120 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 121 {
 122         pgd_t *pg;
 123         pud_t *pu;
 124         pmd_t *pm;
 125         hugepd_t *hpdp = NULL;
 126         unsigned pshift = __ffs(sz);
 127         unsigned pdshift = PGDIR_SHIFT;
 128         spinlock_t *ptl;
 129
 130         addr &= ~(sz-1);
 131         pg = pgd_offset(mm, addr);
 132
 133 #ifdef CONFIG_PPC_BOOK3S_64
 134         if (pshift == PGDIR_SHIFT)
 135                 /* 16GB huge page */
 136                 return (pte_t *) pg;
 137         else if (pshift > PUD_SHIFT) {
 138                 /*
 139                  * We need to use hugepd table
 140                  */
 141                 ptl = &mm->page_table_lock;
 142                 hpdp = (hugepd_t *)pg;
 143         } else {
 144                 pdshift = PUD_SHIFT;
 145                 pu = pud_alloc(mm, pg, addr);
 146                 if (!pu)
 147                         return NULL;
 148                 if (pshift == PUD_SHIFT)
 149                         return (pte_t *)pu;
 150                 else if (pshift > PMD_SHIFT) {
 151                         ptl = pud_lockptr(mm, pu);
 152                         hpdp = (hugepd_t *)pu;
 153                 } else {
 154                         pdshift = PMD_SHIFT;
 155                         pm = pmd_alloc(mm, pu, addr);
 156                         if (!pm)
 157                                 return NULL;
 158                         if (pshift == PMD_SHIFT)
 159                                 /* 16MB hugepage */
 160                                 return (pte_t *)pm;
 161                         else {
 162                                 ptl = pmd_lockptr(mm, pm);
 163                                 hpdp = (hugepd_t *)pm;
 164                         }
 165                 }
 166         }
 167 #else
 168         if (pshift >= PGDIR_SHIFT) {
 169                 ptl = &mm->page_table_lock;
 170                 hpdp = (hugepd_t *)pg;
 171         } else {
 172                 pdshift = PUD_SHIFT;
 173                 pu = pud_alloc(mm, pg, addr);
 174                 if (!pu)
 175                         return NULL;
 176                 if (pshift >= PUD_SHIFT) {
 177                         ptl = pud_lockptr(mm, pu);
 178                         hpdp = (hugepd_t *)pu;
 179                 } else {
 180                         pdshift = PMD_SHIFT;
 181                         pm = pmd_alloc(mm, pu, addr);
 182                         if (!pm)
 183                                 return NULL;
 184                         ptl = pmd_lockptr(mm, pm);
 185                         hpdp = (hugepd_t *)pm;
 186                 }
 187         }
 188 #endif
 189         if (!hpdp)
 190                 return NULL;
 191
 192         if (IS_ENABLED(CONFIG_PPC_8xx) && sz == SZ_512K)
 193                 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
 194
 195         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 196
 197         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
 198                                                   pdshift, pshift, ptl))
 199                 return NULL;
 200
 201         return hugepte_offset(*hpdp, addr, pdshift);
 202 }
 203
 204 #ifdef CONFIG_PPC_BOOK3S_64
 205 /*
 206  * Tracks gpages after the device tree is scanned and before the
 207  * huge_boot_pages list is ready on pseries.
 208  */
 209 #define MAX_NUMBER_GPAGES       1024
 210 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 211 __initdata static unsigned nr_gpages;
 212
 213 /*
 214  * Build list of addresses of gigantic pages.  This function is used in early
 215  * boot before the buddy allocator is setup.
 216  */
 217 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 218 {
 219         if (!addr)
 220                 return;
 221         while (number_of_pages > 0) {
 222                 gpage_freearray[nr_gpages] = addr;
 223                 nr_gpages++;
 224                 number_of_pages--;
 225                 addr += page_size;
 226         }
 227 }
 228
 229 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 230 {
 231         struct huge_bootmem_page *m;
 232         if (nr_gpages == 0)
 233                 return 0;
 234         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 235         gpage_freearray[nr_gpages] = 0;
 236         list_add(&m->list, &huge_boot_pages);
 237         m->hstate = hstate;
 238         return 1;
 239 }
 240 #endif
 241
 242
 243 int __init alloc_bootmem_huge_page(struct hstate *h)
 244 {
 245
 246 #ifdef CONFIG_PPC_BOOK3S_64
 247         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
 248                 return pseries_alloc_bootmem_huge_page(h);
 249 #endif
 250         return __alloc_bootmem_huge_page(h);
 251 }
 252
 253 #ifndef CONFIG_PPC_BOOK3S_64
 254 #define HUGEPD_FREELIST_SIZE \
 255         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 256
 257 struct hugepd_freelist {
 258         struct rcu_head rcu;
 259         unsigned int index;
 260         void *ptes[];
 261 };
 262
 263 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 264
 265 static void hugepd_free_rcu_callback(struct rcu_head *head)
 266 {
 267         struct hugepd_freelist *batch =
 268                 container_of(head, struct hugepd_freelist, rcu);
 269         unsigned int i;
 270
 271         for (i = 0; i < batch->index; i++)
 272                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
 273
 274         free_page((unsigned long)batch);
 275 }
 276
 277 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 278 {
 279         struct hugepd_freelist **batchp;
 280
 281         batchp = &get_cpu_var(hugepd_freelist_cur);
 282
 283         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 284             mm_is_thread_local(tlb->mm)) {
 285                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
 286                 put_cpu_var(hugepd_freelist_cur);
 287                 return;
 288         }
 289
 290         if (*batchp == NULL) {
 291                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 292                 (*batchp)->index = 0;
 293         }
 294
 295         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 296         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 297                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
 298                 *batchp = NULL;
 299         }
 300         put_cpu_var(hugepd_freelist_cur);
 301 }
 302 #else
 303 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 304 #endif
 305
 306 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 307                               unsigned long start, unsigned long end,
 308                               unsigned long floor, unsigned long ceiling)
 309 {
 310         pte_t *hugepte = hugepd_page(*hpdp);
 311         int i;
 312
 313         unsigned long pdmask = ~((1UL << pdshift) - 1);
 314         unsigned int num_hugepd = 1;
 315         unsigned int shift = hugepd_shift(*hpdp);
 316
 317         /* Note: On fsl the hpdp may be the first of several */
 318         if (shift > pdshift)
 319                 num_hugepd = 1 << (shift - pdshift);
 320
 321         start &= pdmask;
 322         if (start < floor)
 323                 return;
 324         if (ceiling) {
 325                 ceiling &= pdmask;
 326                 if (! ceiling)
 327                         return;
 328         }
 329         if (end - 1 > ceiling - 1)
 330                 return;
 331
 332         for (i = 0; i < num_hugepd; i++, hpdp++)
 333                 *hpdp = __hugepd(0);
 334
 335         if (shift >= pdshift)
 336                 hugepd_free(tlb, hugepte);
 337         else
 338                 pgtable_free_tlb(tlb, hugepte,
 339                                  get_hugepd_cache_index(pdshift - shift));
 340 }
 341
 342 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, unsigned long addr)
 343 {
 344         pgtable_t token = pmd_pgtable(*pmd);
 345
 346         pmd_clear(pmd);
 347         pte_free_tlb(tlb, token, addr);
 348         mm_dec_nr_ptes(tlb->mm);
 349 }
 350
 351 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 352                                    unsigned long addr, unsigned long end,
 353                                    unsigned long floor, unsigned long ceiling)
 354 {
 355         pmd_t *pmd;
 356         unsigned long next;
 357         unsigned long start;
 358
 359         start = addr;
 360         do {
 361                 unsigned long more;
 362
 363                 pmd = pmd_offset(pud, addr);
 364                 next = pmd_addr_end(addr, end);
 365                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 366                         if (pmd_none_or_clear_bad(pmd))
 367                                 continue;
 368
 369                         /*
 370                          * if it is not hugepd pointer, we should already find
 371                          * it cleared.
 372                          */
 373                         WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
 374
 375                         hugetlb_free_pte_range(tlb, pmd, addr);
 376
 377                         continue;
 378                 }
 379                 /*
 380                  * Increment next by the size of the huge mapping since
 381                  * there may be more than one entry at this level for a
 382                  * single hugepage, but all of them point to
 383                  * the same kmem cache that holds the hugepte.
 384                  */
 385                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 386                 if (more > next)
 387                         next = more;
 388
 389                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 390                                   addr, next, floor, ceiling);
 391         } while (addr = next, addr != end);
 392
 393         start &= PUD_MASK;
 394         if (start < floor)
 395                 return;
 396         if (ceiling) {
 397                 ceiling &= PUD_MASK;
 398                 if (!ceiling)
 399                         return;
 400         }
 401         if (end - 1 > ceiling - 1)
 402                 return;
 403
 404         pmd = pmd_offset(pud, start);
 405         pud_clear(pud);
 406         pmd_free_tlb(tlb, pmd, start);
 407         mm_dec_nr_pmds(tlb->mm);
 408 }
 409
 410 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 411                                    unsigned long addr, unsigned long end,
 412                                    unsigned long floor, unsigned long ceiling)
 413 {
 414         pud_t *pud;
 415         unsigned long next;
 416         unsigned long start;
 417
 418         start = addr;
 419         do {
 420                 pud = pud_offset(pgd, addr);
 421                 next = pud_addr_end(addr, end);
 422                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 423                         if (pud_none_or_clear_bad(pud))
 424                                 continue;
 425                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 426                                                ceiling);
 427                 } else {
 428                         unsigned long more;
 429                         /*
 430                          * Increment next by the size of the huge mapping since
 431                          * there may be more than one entry at this level for a
 432                          * single hugepage, but all of them point to
 433                          * the same kmem cache that holds the hugepte.
 434                          */
 435                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 436                         if (more > next)
 437                                 next = more;
 438
 439                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 440                                           addr, next, floor, ceiling);
 441                 }
 442         } while (addr = next, addr != end);
 443
 444         start &= PGDIR_MASK;
 445         if (start < floor)
 446                 return;
 447         if (ceiling) {
 448                 ceiling &= PGDIR_MASK;
 449                 if (!ceiling)
 450                         return;
 451         }
 452         if (end - 1 > ceiling - 1)
 453                 return;
 454
 455         pud = pud_offset(pgd, start);
 456         pgd_clear(pgd);
 457         pud_free_tlb(tlb, pud, start);
 458         mm_dec_nr_puds(tlb->mm);
 459 }
 460
 461 /*
 462  * This function frees user-level page tables of a process.
 463  */
 464 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 465                             unsigned long addr, unsigned long end,
 466                             unsigned long floor, unsigned long ceiling)
 467 {
 468         pgd_t *pgd;
 469         unsigned long next;
 470
 471         /*
 472          * Because there are a number of different possible pagetable
 473          * layouts for hugepage ranges, we limit knowledge of how
 474          * things should be laid out to the allocation path
 475          * (huge_pte_alloc(), above).  Everything else works out the
 476          * structure as it goes from information in the hugepd
 477          * pointers.  That means that we can't here use the
 478          * optimization used in the normal page free_pgd_range(), of
 479          * checking whether we're actually covering a large enough
 480          * range to have to do anything at the top level of the walk
 481          * instead of at the bottom.
 482          *
 483          * To make sense of this, you should probably go read the big
 484          * block comment at the top of the normal free_pgd_range(),
 485          * too.
 486          */
 487
 488         do {
 489                 next = pgd_addr_end(addr, end);
 490                 pgd = pgd_offset(tlb->mm, addr);
 491                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 492                         if (pgd_none_or_clear_bad(pgd))
 493                                 continue;
 494                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 495                 } else {
 496                         unsigned long more;
 497                         /*
 498                          * Increment next by the size of the huge mapping since
 499                          * there may be more than one entry at the pgd level
 500                          * for a single hugepage, but all of them point to the
 501                          * same kmem cache that holds the hugepte.
 502                          */
 503                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 504                         if (more > next)
 505                                 next = more;
 506
 507                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 508                                           addr, next, floor, ceiling);
 509                 }
 510         } while (addr = next, addr != end);
 511 }
 512
 513 struct page *follow_huge_pd(struct vm_area_struct *vma,
 514                             unsigned long address, hugepd_t hpd,
 515                             int flags, int pdshift)
 516 {
 517         pte_t *ptep;
 518         spinlock_t *ptl;
 519         struct page *page = NULL;
 520         unsigned long mask;
 521         int shift = hugepd_shift(hpd);
 522         struct mm_struct *mm = vma->vm_mm;
 523
 524 retry:
 525         /*
 526          * hugepage directory entries are protected by mm->page_table_lock
 527          * Use this instead of huge_pte_lockptr
 528          */
 529         ptl = &mm->page_table_lock;
 530         spin_lock(ptl);
 531
 532         ptep = hugepte_offset(hpd, address, pdshift);
 533         if (pte_present(*ptep)) {
 534                 mask = (1UL << shift) - 1;
 535                 page = pte_page(*ptep);
 536                 page += ((address & mask) >> PAGE_SHIFT);
 537                 if (flags & FOLL_GET)
 538                         get_page(page);
 539         } else {
 540                 if (is_hugetlb_entry_migration(*ptep)) {
 541                         spin_unlock(ptl);
 542                         __migration_entry_wait(mm, ptep, ptl);
 543                         goto retry;
 544                 }
 545         }
 546         spin_unlock(ptl);
 547         return page;
 548 }
 549
 550 #ifdef CONFIG_PPC_MM_SLICES
 551 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 552                                         unsigned long len, unsigned long pgoff,
 553                                         unsigned long flags)
 554 {
 555         struct hstate *hstate = hstate_file(file);
 556         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 557
 558 #ifdef CONFIG_PPC_RADIX_MMU
 559         if (radix_enabled())
 560                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 561                                                        pgoff, flags);
 562 #endif
 563         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 564 }
 565 #endif
 566
 567 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 568 {
 569         /* With radix we don't use slice, so derive it from vma*/
 570         if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
 571                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 572
 573                 return 1UL << mmu_psize_to_shift(psize);
 574         }
 575         return vma_kernel_pagesize(vma);
 576 }
 577
 578 static int __init add_huge_page_size(unsigned long long size)
 579 {
 580         int shift = __ffs(size);
 581         int mmu_psize;
 582
 583         /* Check that it is a page size supported by the hardware and
 584          * that it fits within pagetable and slice limits. */
 585         if (size <= PAGE_SIZE || !is_power_of_2(size))
 586                 return -EINVAL;
 587
 588         mmu_psize = check_and_get_huge_psize(shift);
 589         if (mmu_psize < 0)
 590                 return -EINVAL;
 591
 592         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 593
 594         /* Return if huge page size has already been setup */
 595         if (size_to_hstate(size))
 596                 return 0;
 597
 598         hugetlb_add_hstate(shift - PAGE_SHIFT);
 599
 600         return 0;
 601 }
 602
 603 static int __init hugepage_setup_sz(char *str)
 604 {
 605         unsigned long long size;
 606
 607         size = memparse(str, &str);
 608
 609         if (add_huge_page_size(size) != 0) {
 610                 hugetlb_bad_size();
 611                 pr_err("Invalid huge page size specified(%llu)\n", size);
 612         }
 613
 614         return 1;
 615 }
 616 __setup("hugepagesz=", hugepage_setup_sz);
 617
 618 static int __init hugetlbpage_init(void)
 619 {
 620         bool configured = false;
 621         int psize;
 622
 623         if (hugetlb_disabled) {
 624                 pr_info("HugeTLB support is disabled!\n");
 625                 return 0;
 626         }
 627
 628         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
 629             !mmu_has_feature(MMU_FTR_16M_PAGE))
 630                 return -ENODEV;
 631
 632         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 633                 unsigned shift;
 634                 unsigned pdshift;
 635
 636                 if (!mmu_psize_defs[psize].shift)
 637                         continue;
 638
 639                 shift = mmu_psize_to_shift(psize);
 640
 641 #ifdef CONFIG_PPC_BOOK3S_64
 642                 if (shift > PGDIR_SHIFT)
 643                         continue;
 644                 else if (shift > PUD_SHIFT)
 645                         pdshift = PGDIR_SHIFT;
 646                 else if (shift > PMD_SHIFT)
 647                         pdshift = PUD_SHIFT;
 648                 else
 649                         pdshift = PMD_SHIFT;
 650 #else
 651                 if (shift < PUD_SHIFT)
 652                         pdshift = PMD_SHIFT;
 653                 else if (shift < PGDIR_SHIFT)
 654                         pdshift = PUD_SHIFT;
 655                 else
 656                         pdshift = PGDIR_SHIFT;
 657 #endif
 658
 659                 if (add_huge_page_size(1ULL << shift) < 0)
 660                         continue;
 661                 /*
 662                  * if we have pdshift and shift value same, we don't
 663                  * use pgt cache for hugepd.
 664                  */
 665                 if (pdshift > shift) {
 666                         if (!IS_ENABLED(CONFIG_PPC_8xx))
 667                                 pgtable_cache_add(pdshift - shift);
 668                 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
 669                            IS_ENABLED(CONFIG_PPC_8xx)) {
 670                         pgtable_cache_add(PTE_T_ORDER);
 671                 }
 672
 673                 configured = true;
 674         }
 675
 676         if (configured) {
 677                 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
 678                         hugetlbpage_init_default();
 679         } else
 680                 pr_info("Failed to initialize. Disabling HugeTLB");
 681
 682         return 0;
 683 }
 684
 685 arch_initcall(hugetlbpage_init);
 686
 687 void flush_dcache_icache_hugepage(struct page *page)
 688 {
 689         int i;
 690         void *start;
 691
 692         BUG_ON(!PageCompound(page));
 693
 694         for (i = 0; i < compound_nr(page); i++) {
 695                 if (!PageHighMem(page)) {
 696                         __flush_dcache_icache(page_address(page+i));
 697                 } else {
 698                         start = kmap_atomic(page+i);
 699                         __flush_dcache_icache(start);
 700                         kunmap_atomic(start);
 701                 }
 702         }
 703 }