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