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/pgalloc.h>
  23 #include <asm/tlb.h>
  24 #include <asm/setup.h>
  25 #include <asm/hugetlb.h>
  26 #include <asm/pte-walk.h>
  27
  28 bool hugetlb_disabled = false;
  29
  30 #define hugepd_none(hpd)        (hpd_val(hpd) == 0)
  31
  32 #define PTE_T_ORDER     (__builtin_ffs(sizeof(pte_basic_t)) - \
  33                          __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 {
  57                 cachep = PGT_CACHE(pdshift - pshift);
  58                 num_hugepd = 1;
  59         }
  60
  61         if (!cachep) {
  62                 WARN_ONCE(1, "No page table cache created for hugetlb tables");
  63                 return -ENOMEM;
  64         }
  65
  66         new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  67
  68         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  69         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  70
  71         if (!new)
  72                 return -ENOMEM;
  73
  74         /*
  75          * Make sure other cpus find the hugepd set only after a
  76          * properly initialized page table is visible to them.
  77          * For more details look for comment in __pte_alloc().
  78          */
  79         smp_wmb();
  80
  81         spin_lock(ptl);
  82         /*
  83          * We have multiple higher-level entries that point to the same
  84          * actual pte location.  Fill in each as we go and backtrack on error.
  85          * We need all of these so the DTLB pgtable walk code can find the
  86          * right higher-level entry without knowing if it's a hugepage or not.
  87          */
  88         for (i = 0; i < num_hugepd; i++, hpdp++) {
  89                 if (unlikely(!hugepd_none(*hpdp)))
  90                         break;
  91                 hugepd_populate(hpdp, new, pshift);
  92         }
  93         /* If we bailed from the for loop early, an error occurred, clean up */
  94         if (i < num_hugepd) {
  95                 for (i = i - 1 ; i >= 0; i--, hpdp--)
  96                         *hpdp = __hugepd(0);
  97                 kmem_cache_free(cachep, new);
  98         } else {
  99                 kmemleak_ignore(new);
 100         }
 101         spin_unlock(ptl);
 102         return 0;
 103 }
 104
 105 /*
 106  * At this point we do the placement change only for BOOK3S 64. This would
 107  * possibly work on other subarchs.
 108  */
 109 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 110 {
 111         pgd_t *pg;
 112         p4d_t *p4;
 113         pud_t *pu;
 114         pmd_t *pm;
 115         hugepd_t *hpdp = NULL;
 116         unsigned pshift = __ffs(sz);
 117         unsigned pdshift = PGDIR_SHIFT;
 118         spinlock_t *ptl;
 119
 120         addr &= ~(sz-1);
 121         pg = pgd_offset(mm, addr);
 122         p4 = p4d_offset(pg, addr);
 123
 124 #ifdef CONFIG_PPC_BOOK3S_64
 125         if (pshift == PGDIR_SHIFT)
 126                 /* 16GB huge page */
 127                 return (pte_t *) p4;
 128         else if (pshift > PUD_SHIFT) {
 129                 /*
 130                  * We need to use hugepd table
 131                  */
 132                 ptl = &mm->page_table_lock;
 133                 hpdp = (hugepd_t *)p4;
 134         } else {
 135                 pdshift = PUD_SHIFT;
 136                 pu = pud_alloc(mm, p4, addr);
 137                 if (!pu)
 138                         return NULL;
 139                 if (pshift == PUD_SHIFT)
 140                         return (pte_t *)pu;
 141                 else if (pshift > PMD_SHIFT) {
 142                         ptl = pud_lockptr(mm, pu);
 143                         hpdp = (hugepd_t *)pu;
 144                 } else {
 145                         pdshift = PMD_SHIFT;
 146                         pm = pmd_alloc(mm, pu, addr);
 147                         if (!pm)
 148                                 return NULL;
 149                         if (pshift == PMD_SHIFT)
 150                                 /* 16MB hugepage */
 151                                 return (pte_t *)pm;
 152                         else {
 153                                 ptl = pmd_lockptr(mm, pm);
 154                                 hpdp = (hugepd_t *)pm;
 155                         }
 156                 }
 157         }
 158 #else
 159         if (pshift >= PGDIR_SHIFT) {
 160                 ptl = &mm->page_table_lock;
 161                 hpdp = (hugepd_t *)p4;
 162         } else {
 163                 pdshift = PUD_SHIFT;
 164                 pu = pud_alloc(mm, p4, addr);
 165                 if (!pu)
 166                         return NULL;
 167                 if (pshift >= PUD_SHIFT) {
 168                         ptl = pud_lockptr(mm, pu);
 169                         hpdp = (hugepd_t *)pu;
 170                 } else {
 171                         pdshift = PMD_SHIFT;
 172                         pm = pmd_alloc(mm, pu, addr);
 173                         if (!pm)
 174                                 return NULL;
 175                         ptl = pmd_lockptr(mm, pm);
 176                         hpdp = (hugepd_t *)pm;
 177                 }
 178         }
 179 #endif
 180         if (!hpdp)
 181                 return NULL;
 182
 183         if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
 184                 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
 185
 186         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 187
 188         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
 189                                                   pdshift, pshift, ptl))
 190                 return NULL;
 191
 192         return hugepte_offset(*hpdp, addr, pdshift);
 193 }
 194
 195 #ifdef CONFIG_PPC_BOOK3S_64
 196 /*
 197  * Tracks gpages after the device tree is scanned and before the
 198  * huge_boot_pages list is ready on pseries.
 199  */
 200 #define MAX_NUMBER_GPAGES       1024
 201 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 202 __initdata static unsigned nr_gpages;
 203
 204 /*
 205  * Build list of addresses of gigantic pages.  This function is used in early
 206  * boot before the buddy allocator is setup.
 207  */
 208 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 209 {
 210         if (!addr)
 211                 return;
 212         while (number_of_pages > 0) {
 213                 gpage_freearray[nr_gpages] = addr;
 214                 nr_gpages++;
 215                 number_of_pages--;
 216                 addr += page_size;
 217         }
 218 }
 219
 220 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 221 {
 222         struct huge_bootmem_page *m;
 223         if (nr_gpages == 0)
 224                 return 0;
 225         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 226         gpage_freearray[nr_gpages] = 0;
 227         list_add(&m->list, &huge_boot_pages);
 228         m->hstate = hstate;
 229         return 1;
 230 }
 231 #endif
 232
 233
 234 int __init alloc_bootmem_huge_page(struct hstate *h)
 235 {
 236
 237 #ifdef CONFIG_PPC_BOOK3S_64
 238         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
 239                 return pseries_alloc_bootmem_huge_page(h);
 240 #endif
 241         return __alloc_bootmem_huge_page(h);
 242 }
 243
 244 #ifndef CONFIG_PPC_BOOK3S_64
 245 #define HUGEPD_FREELIST_SIZE \
 246         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 247
 248 struct hugepd_freelist {
 249         struct rcu_head rcu;
 250         unsigned int index;
 251         void *ptes[];
 252 };
 253
 254 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 255
 256 static void hugepd_free_rcu_callback(struct rcu_head *head)
 257 {
 258         struct hugepd_freelist *batch =
 259                 container_of(head, struct hugepd_freelist, rcu);
 260         unsigned int i;
 261
 262         for (i = 0; i < batch->index; i++)
 263                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
 264
 265         free_page((unsigned long)batch);
 266 }
 267
 268 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 269 {
 270         struct hugepd_freelist **batchp;
 271
 272         batchp = &get_cpu_var(hugepd_freelist_cur);
 273
 274         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 275             mm_is_thread_local(tlb->mm)) {
 276                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
 277                 put_cpu_var(hugepd_freelist_cur);
 278                 return;
 279         }
 280
 281         if (*batchp == NULL) {
 282                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 283                 (*batchp)->index = 0;
 284         }
 285
 286         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 287         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 288                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
 289                 *batchp = NULL;
 290         }
 291         put_cpu_var(hugepd_freelist_cur);
 292 }
 293 #else
 294 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 295 #endif
 296
 297 /* Return true when the entry to be freed maps more than the area being freed */
 298 static bool range_is_outside_limits(unsigned long start, unsigned long end,
 299                                     unsigned long floor, unsigned long ceiling,
 300                                     unsigned long mask)
 301 {
 302         if ((start & mask) < floor)
 303                 return true;
 304         if (ceiling) {
 305                 ceiling &= mask;
 306                 if (!ceiling)
 307                         return true;
 308         }
 309         return end - 1 > ceiling - 1;
 310 }
 311
 312 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 313                               unsigned long start, unsigned long end,
 314                               unsigned long floor, unsigned long ceiling)
 315 {
 316         pte_t *hugepte = hugepd_page(*hpdp);
 317         int i;
 318
 319         unsigned long pdmask = ~((1UL << pdshift) - 1);
 320         unsigned int num_hugepd = 1;
 321         unsigned int shift = hugepd_shift(*hpdp);
 322
 323         /* Note: On fsl the hpdp may be the first of several */
 324         if (shift > pdshift)
 325                 num_hugepd = 1 << (shift - pdshift);
 326
 327         if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
 328                 return;
 329
 330         for (i = 0; i < num_hugepd; i++, hpdp++)
 331                 *hpdp = __hugepd(0);
 332
 333         if (shift >= pdshift)
 334                 hugepd_free(tlb, hugepte);
 335         else
 336                 pgtable_free_tlb(tlb, hugepte,
 337                                  get_hugepd_cache_index(pdshift - shift));
 338 }
 339
 340 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 341                                    unsigned long addr, unsigned long end,
 342                                    unsigned long floor, unsigned long ceiling)
 343 {
 344         pgtable_t token = pmd_pgtable(*pmd);
 345
 346         if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
 347                 return;
 348
 349         pmd_clear(pmd);
 350         pte_free_tlb(tlb, token, addr);
 351         mm_dec_nr_ptes(tlb->mm);
 352 }
 353
 354 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 355                                    unsigned long addr, unsigned long end,
 356                                    unsigned long floor, unsigned long ceiling)
 357 {
 358         pmd_t *pmd;
 359         unsigned long next;
 360         unsigned long start;
 361
 362         start = addr;
 363         do {
 364                 unsigned long more;
 365
 366                 pmd = pmd_offset(pud, addr);
 367                 next = pmd_addr_end(addr, end);
 368                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 369                         if (pmd_none_or_clear_bad(pmd))
 370                                 continue;
 371
 372                         /*
 373                          * if it is not hugepd pointer, we should already find
 374                          * it cleared.
 375                          */
 376                         WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
 377
 378                         hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
 379
 380                         continue;
 381                 }
 382                 /*
 383                  * Increment next by the size of the huge mapping since
 384                  * there may be more than one entry at this level for a
 385                  * single hugepage, but all of them point to
 386                  * the same kmem cache that holds the hugepte.
 387                  */
 388                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 389                 if (more > next)
 390                         next = more;
 391
 392                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 393                                   addr, next, floor, ceiling);
 394         } while (addr = next, addr != end);
 395
 396         if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
 397                 return;
 398
 399         pmd = pmd_offset(pud, start & PUD_MASK);
 400         pud_clear(pud);
 401         pmd_free_tlb(tlb, pmd, start & PUD_MASK);
 402         mm_dec_nr_pmds(tlb->mm);
 403 }
 404
 405 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 406                                    unsigned long addr, unsigned long end,
 407                                    unsigned long floor, unsigned long ceiling)
 408 {
 409         pud_t *pud;
 410         unsigned long next;
 411         unsigned long start;
 412
 413         start = addr;
 414         do {
 415                 pud = pud_offset(p4d, addr);
 416                 next = pud_addr_end(addr, end);
 417                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 418                         if (pud_none_or_clear_bad(pud))
 419                                 continue;
 420                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 421                                                ceiling);
 422                 } else {
 423                         unsigned long more;
 424                         /*
 425                          * Increment next by the size of the huge mapping since
 426                          * there may be more than one entry at this level for a
 427                          * single hugepage, but all of them point to
 428                          * the same kmem cache that holds the hugepte.
 429                          */
 430                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 431                         if (more > next)
 432                                 next = more;
 433
 434                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 435                                           addr, next, floor, ceiling);
 436                 }
 437         } while (addr = next, addr != end);
 438
 439         if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
 440                 return;
 441
 442         pud = pud_offset(p4d, start & PGDIR_MASK);
 443         p4d_clear(p4d);
 444         pud_free_tlb(tlb, pud, start & PGDIR_MASK);
 445         mm_dec_nr_puds(tlb->mm);
 446 }
 447
 448 /*
 449  * This function frees user-level page tables of a process.
 450  */
 451 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 452                             unsigned long addr, unsigned long end,
 453                             unsigned long floor, unsigned long ceiling)
 454 {
 455         pgd_t *pgd;
 456         p4d_t *p4d;
 457         unsigned long next;
 458
 459         /*
 460          * Because there are a number of different possible pagetable
 461          * layouts for hugepage ranges, we limit knowledge of how
 462          * things should be laid out to the allocation path
 463          * (huge_pte_alloc(), above).  Everything else works out the
 464          * structure as it goes from information in the hugepd
 465          * pointers.  That means that we can't here use the
 466          * optimization used in the normal page free_pgd_range(), of
 467          * checking whether we're actually covering a large enough
 468          * range to have to do anything at the top level of the walk
 469          * instead of at the bottom.
 470          *
 471          * To make sense of this, you should probably go read the big
 472          * block comment at the top of the normal free_pgd_range(),
 473          * too.
 474          */
 475
 476         do {
 477                 next = pgd_addr_end(addr, end);
 478                 pgd = pgd_offset(tlb->mm, addr);
 479                 p4d = p4d_offset(pgd, addr);
 480                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 481                         if (p4d_none_or_clear_bad(p4d))
 482                                 continue;
 483                         hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 484                 } else {
 485                         unsigned long more;
 486                         /*
 487                          * Increment next by the size of the huge mapping since
 488                          * there may be more than one entry at the pgd level
 489                          * for a single hugepage, but all of them point to the
 490                          * same kmem cache that holds the hugepte.
 491                          */
 492                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 493                         if (more > next)
 494                                 next = more;
 495
 496                         free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
 497                                           addr, next, floor, ceiling);
 498                 }
 499         } while (addr = next, addr != end);
 500 }
 501
 502 struct page *follow_huge_pd(struct vm_area_struct *vma,
 503                             unsigned long address, hugepd_t hpd,
 504                             int flags, int pdshift)
 505 {
 506         pte_t *ptep;
 507         spinlock_t *ptl;
 508         struct page *page = NULL;
 509         unsigned long mask;
 510         int shift = hugepd_shift(hpd);
 511         struct mm_struct *mm = vma->vm_mm;
 512
 513 retry:
 514         /*
 515          * hugepage directory entries are protected by mm->page_table_lock
 516          * Use this instead of huge_pte_lockptr
 517          */
 518         ptl = &mm->page_table_lock;
 519         spin_lock(ptl);
 520
 521         ptep = hugepte_offset(hpd, address, pdshift);
 522         if (pte_present(*ptep)) {
 523                 mask = (1UL << shift) - 1;
 524                 page = pte_page(*ptep);
 525                 page += ((address & mask) >> PAGE_SHIFT);
 526                 if (flags & FOLL_GET)
 527                         get_page(page);
 528         } else {
 529                 if (is_hugetlb_entry_migration(*ptep)) {
 530                         spin_unlock(ptl);
 531                         __migration_entry_wait(mm, ptep, ptl);
 532                         goto retry;
 533                 }
 534         }
 535         spin_unlock(ptl);
 536         return page;
 537 }
 538
 539 #ifdef CONFIG_PPC_MM_SLICES
 540 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 541                                         unsigned long len, unsigned long pgoff,
 542                                         unsigned long flags)
 543 {
 544         struct hstate *hstate = hstate_file(file);
 545         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 546
 547 #ifdef CONFIG_PPC_RADIX_MMU
 548         if (radix_enabled())
 549                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 550                                                        pgoff, flags);
 551 #endif
 552         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 553 }
 554 #endif
 555
 556 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 557 {
 558         /* With radix we don't use slice, so derive it from vma*/
 559         if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
 560                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 561
 562                 return 1UL << mmu_psize_to_shift(psize);
 563         }
 564         return vma_kernel_pagesize(vma);
 565 }
 566
 567 bool __init arch_hugetlb_valid_size(unsigned long size)
 568 {
 569         int shift = __ffs(size);
 570         int mmu_psize;
 571
 572         /* Check that it is a page size supported by the hardware and
 573          * that it fits within pagetable and slice limits. */
 574         if (size <= PAGE_SIZE || !is_power_of_2(size))
 575                 return false;
 576
 577         mmu_psize = check_and_get_huge_psize(shift);
 578         if (mmu_psize < 0)
 579                 return false;
 580
 581         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 582
 583         return true;
 584 }
 585
 586 static int __init add_huge_page_size(unsigned long long size)
 587 {
 588         int shift = __ffs(size);
 589
 590         if (!arch_hugetlb_valid_size((unsigned long)size))
 591                 return -EINVAL;
 592
 593         hugetlb_add_hstate(shift - PAGE_SHIFT);
 594         return 0;
 595 }
 596
 597 static int __init hugetlbpage_init(void)
 598 {
 599         bool configured = false;
 600         int psize;
 601
 602         if (hugetlb_disabled) {
 603                 pr_info("HugeTLB support is disabled!\n");
 604                 return 0;
 605         }
 606
 607         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
 608             !mmu_has_feature(MMU_FTR_16M_PAGE))
 609                 return -ENODEV;
 610
 611         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 612                 unsigned shift;
 613                 unsigned pdshift;
 614
 615                 if (!mmu_psize_defs[psize].shift)
 616                         continue;
 617
 618                 shift = mmu_psize_to_shift(psize);
 619
 620 #ifdef CONFIG_PPC_BOOK3S_64
 621                 if (shift > PGDIR_SHIFT)
 622                         continue;
 623                 else if (shift > PUD_SHIFT)
 624                         pdshift = PGDIR_SHIFT;
 625                 else if (shift > PMD_SHIFT)
 626                         pdshift = PUD_SHIFT;
 627                 else
 628                         pdshift = PMD_SHIFT;
 629 #else
 630                 if (shift < PUD_SHIFT)
 631                         pdshift = PMD_SHIFT;
 632                 else if (shift < PGDIR_SHIFT)
 633                         pdshift = PUD_SHIFT;
 634                 else
 635                         pdshift = PGDIR_SHIFT;
 636 #endif
 637
 638                 if (add_huge_page_size(1ULL << shift) < 0)
 639                         continue;
 640                 /*
 641                  * if we have pdshift and shift value same, we don't
 642                  * use pgt cache for hugepd.
 643                  */
 644                 if (pdshift > shift) {
 645                         if (!IS_ENABLED(CONFIG_PPC_8xx))
 646                                 pgtable_cache_add(pdshift - shift);
 647                 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
 648                            IS_ENABLED(CONFIG_PPC_8xx)) {
 649                         pgtable_cache_add(PTE_T_ORDER);
 650                 }
 651
 652                 configured = true;
 653         }
 654
 655         if (configured) {
 656                 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
 657                         hugetlbpage_init_default();
 658         } else
 659                 pr_info("Failed to initialize. Disabling HugeTLB");
 660
 661         return 0;
 662 }
 663
 664 arch_initcall(hugetlbpage_init);
 665
 666 void flush_dcache_icache_hugepage(struct page *page)
 667 {
 668         int i;
 669         void *start;
 670
 671         BUG_ON(!PageCompound(page));
 672
 673         for (i = 0; i < compound_nr(page); i++) {
 674                 if (!PageHighMem(page)) {
 675                         __flush_dcache_icache(page_address(page+i));
 676                 } else {
 677                         start = kmap_atomic(page+i);
 678                         __flush_dcache_icache(start);
 679                         kunmap_atomic(start);
 680                 }
 681         }
 682 }
 683
 684 void __init gigantic_hugetlb_cma_reserve(void)
 685 {
 686         unsigned long order = 0;
 687
 688         if (radix_enabled())
 689                 order = PUD_SHIFT - PAGE_SHIFT;
 690         else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
 691                 /*
 692                  * For pseries we do use ibm,expected#pages for reserving 16G pages.
 693                  */
 694                 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
 695
 696         if (order) {
 697                 VM_WARN_ON(order < MAX_ORDER);
 698                 hugetlb_cma_reserve(order);
 699         }
 700 }