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 <[email protected]>
   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/bootmem.h>
  19 #include <linux/moduleparam.h>
  20 #include <linux/swap.h>
  21 #include <linux/swapops.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
  30 #ifdef CONFIG_HUGETLB_PAGE
  31
  32 #define PAGE_SHIFT_64K  16
  33 #define PAGE_SHIFT_512K 19
  34 #define PAGE_SHIFT_8M   23
  35 #define PAGE_SHIFT_16M  24
  36 #define PAGE_SHIFT_16G  34
  37
  38 bool hugetlb_disabled = false;
  39
  40 unsigned int HPAGE_SHIFT;
  41 EXPORT_SYMBOL(HPAGE_SHIFT);
  42
  43 #define hugepd_none(hpd)        (hpd_val(hpd) == 0)
  44
  45 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  46 {
  47         /*
  48          * Only called for hugetlbfs pages, hence can ignore THP and the
  49          * irq disabled walk.
  50          */
  51         return __find_linux_pte(mm->pgd, addr, NULL, NULL);
  52 }
  53
  54 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  55                            unsigned long address, unsigned int pdshift,
  56                            unsigned int pshift, spinlock_t *ptl)
  57 {
  58         struct kmem_cache *cachep;
  59         pte_t *new;
  60         int i;
  61         int num_hugepd;
  62
  63         if (pshift >= pdshift) {
  64                 cachep = hugepte_cache;
  65                 num_hugepd = 1 << (pshift - pdshift);
  66         } else {
  67                 cachep = PGT_CACHE(pdshift - pshift);
  68                 num_hugepd = 1;
  69         }
  70
  71         new = kmem_cache_zalloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  72
  73         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  74         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  75
  76         if (! new)
  77                 return -ENOMEM;
  78
  79         /*
  80          * Make sure other cpus find the hugepd set only after a
  81          * properly initialized page table is visible to them.
  82          * For more details look for comment in __pte_alloc().
  83          */
  84         smp_wmb();
  85
  86         spin_lock(ptl);
  87         /*
  88          * We have multiple higher-level entries that point to the same
  89          * actual pte location.  Fill in each as we go and backtrack on error.
  90          * We need all of these so the DTLB pgtable walk code can find the
  91          * right higher-level entry without knowing if it's a hugepage or not.
  92          */
  93         for (i = 0; i < num_hugepd; i++, hpdp++) {
  94                 if (unlikely(!hugepd_none(*hpdp)))
  95                         break;
  96                 else {
  97 #ifdef CONFIG_PPC_BOOK3S_64
  98                         *hpdp = __hugepd(__pa(new) |
  99                                          (shift_to_mmu_psize(pshift) << 2));
 100 #elif defined(CONFIG_PPC_8xx)
 101                         *hpdp = __hugepd(__pa(new) | _PMD_USER |
 102                                          (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
 103                                           _PMD_PAGE_512K) | _PMD_PRESENT);
 104 #else
 105                         /* We use the old format for PPC_FSL_BOOK3E */
 106                         *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
 107 #endif
 108                 }
 109         }
 110         /* If we bailed from the for loop early, an error occurred, clean up */
 111         if (i < num_hugepd) {
 112                 for (i = i - 1 ; i >= 0; i--, hpdp--)
 113                         *hpdp = __hugepd(0);
 114                 kmem_cache_free(cachep, new);
 115         }
 116         spin_unlock(ptl);
 117         return 0;
 118 }
 119
 120 /*
 121  * These macros define how to determine which level of the page table holds
 122  * the hpdp.
 123  */
 124 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 125 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
 126 #define HUGEPD_PUD_SHIFT PUD_SHIFT
 127 #endif
 128
 129 /*
 130  * At this point we do the placement change only for BOOK3S 64. This would
 131  * possibly work on other subarchs.
 132  */
 133 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 134 {
 135         pgd_t *pg;
 136         pud_t *pu;
 137         pmd_t *pm;
 138         hugepd_t *hpdp = NULL;
 139         unsigned pshift = __ffs(sz);
 140         unsigned pdshift = PGDIR_SHIFT;
 141         spinlock_t *ptl;
 142
 143         addr &= ~(sz-1);
 144         pg = pgd_offset(mm, addr);
 145
 146 #ifdef CONFIG_PPC_BOOK3S_64
 147         if (pshift == PGDIR_SHIFT)
 148                 /* 16GB huge page */
 149                 return (pte_t *) pg;
 150         else if (pshift > PUD_SHIFT) {
 151                 /*
 152                  * We need to use hugepd table
 153                  */
 154                 ptl = &mm->page_table_lock;
 155                 hpdp = (hugepd_t *)pg;
 156         } else {
 157                 pdshift = PUD_SHIFT;
 158                 pu = pud_alloc(mm, pg, addr);
 159                 if (pshift == PUD_SHIFT)
 160                         return (pte_t *)pu;
 161                 else if (pshift > PMD_SHIFT) {
 162                         ptl = pud_lockptr(mm, pu);
 163                         hpdp = (hugepd_t *)pu;
 164                 } else {
 165                         pdshift = PMD_SHIFT;
 166                         pm = pmd_alloc(mm, pu, addr);
 167                         if (pshift == PMD_SHIFT)
 168                                 /* 16MB hugepage */
 169                                 return (pte_t *)pm;
 170                         else {
 171                                 ptl = pmd_lockptr(mm, pm);
 172                                 hpdp = (hugepd_t *)pm;
 173                         }
 174                 }
 175         }
 176 #else
 177         if (pshift >= HUGEPD_PGD_SHIFT) {
 178                 ptl = &mm->page_table_lock;
 179                 hpdp = (hugepd_t *)pg;
 180         } else {
 181                 pdshift = PUD_SHIFT;
 182                 pu = pud_alloc(mm, pg, addr);
 183                 if (pshift >= HUGEPD_PUD_SHIFT) {
 184                         ptl = pud_lockptr(mm, pu);
 185                         hpdp = (hugepd_t *)pu;
 186                 } else {
 187                         pdshift = PMD_SHIFT;
 188                         pm = pmd_alloc(mm, pu, addr);
 189                         ptl = pmd_lockptr(mm, pm);
 190                         hpdp = (hugepd_t *)pm;
 191                 }
 192         }
 193 #endif
 194         if (!hpdp)
 195                 return NULL;
 196
 197         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 198
 199         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
 200                                                   pdshift, pshift, ptl))
 201                 return NULL;
 202
 203         return hugepte_offset(*hpdp, addr, pdshift);
 204 }
 205
 206 #ifdef CONFIG_PPC_BOOK3S_64
 207 /*
 208  * Tracks gpages after the device tree is scanned and before the
 209  * huge_boot_pages list is ready on pseries.
 210  */
 211 #define MAX_NUMBER_GPAGES       1024
 212 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 213 __initdata static unsigned nr_gpages;
 214
 215 /*
 216  * Build list of addresses of gigantic pages.  This function is used in early
 217  * boot before the buddy allocator is setup.
 218  */
 219 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 220 {
 221         if (!addr)
 222                 return;
 223         while (number_of_pages > 0) {
 224                 gpage_freearray[nr_gpages] = addr;
 225                 nr_gpages++;
 226                 number_of_pages--;
 227                 addr += page_size;
 228         }
 229 }
 230
 231 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 232 {
 233         struct huge_bootmem_page *m;
 234         if (nr_gpages == 0)
 235                 return 0;
 236         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 237         gpage_freearray[nr_gpages] = 0;
 238         list_add(&m->list, &huge_boot_pages);
 239         m->hstate = hstate;
 240         return 1;
 241 }
 242 #endif
 243
 244
 245 int __init alloc_bootmem_huge_page(struct hstate *h)
 246 {
 247
 248 #ifdef CONFIG_PPC_BOOK3S_64
 249         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
 250                 return pseries_alloc_bootmem_huge_page(h);
 251 #endif
 252         return __alloc_bootmem_huge_page(h);
 253 }
 254
 255 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 256 #define HUGEPD_FREELIST_SIZE \
 257         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 258
 259 struct hugepd_freelist {
 260         struct rcu_head rcu;
 261         unsigned int index;
 262         void *ptes[0];
 263 };
 264
 265 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 266
 267 static void hugepd_free_rcu_callback(struct rcu_head *head)
 268 {
 269         struct hugepd_freelist *batch =
 270                 container_of(head, struct hugepd_freelist, rcu);
 271         unsigned int i;
 272
 273         for (i = 0; i < batch->index; i++)
 274                 kmem_cache_free(hugepte_cache, batch->ptes[i]);
 275
 276         free_page((unsigned long)batch);
 277 }
 278
 279 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 280 {
 281         struct hugepd_freelist **batchp;
 282
 283         batchp = &get_cpu_var(hugepd_freelist_cur);
 284
 285         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 286             mm_is_thread_local(tlb->mm)) {
 287                 kmem_cache_free(hugepte_cache, hugepte);
 288                 put_cpu_var(hugepd_freelist_cur);
 289                 return;
 290         }
 291
 292         if (*batchp == NULL) {
 293                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 294                 (*batchp)->index = 0;
 295         }
 296
 297         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 298         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 299                 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
 300                 *batchp = NULL;
 301         }
 302         put_cpu_var(hugepd_freelist_cur);
 303 }
 304 #else
 305 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 306 #endif
 307
 308 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 309                               unsigned long start, unsigned long end,
 310                               unsigned long floor, unsigned long ceiling)
 311 {
 312         pte_t *hugepte = hugepd_page(*hpdp);
 313         int i;
 314
 315         unsigned long pdmask = ~((1UL << pdshift) - 1);
 316         unsigned int num_hugepd = 1;
 317         unsigned int shift = hugepd_shift(*hpdp);
 318
 319         /* Note: On fsl the hpdp may be the first of several */
 320         if (shift > pdshift)
 321                 num_hugepd = 1 << (shift - pdshift);
 322
 323         start &= pdmask;
 324         if (start < floor)
 325                 return;
 326         if (ceiling) {
 327                 ceiling &= pdmask;
 328                 if (! ceiling)
 329                         return;
 330         }
 331         if (end - 1 > ceiling - 1)
 332                 return;
 333
 334         for (i = 0; i < num_hugepd; i++, hpdp++)
 335                 *hpdp = __hugepd(0);
 336
 337         if (shift >= pdshift)
 338                 hugepd_free(tlb, hugepte);
 339         else
 340                 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
 341 }
 342
 343 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 344                                    unsigned long addr, unsigned long end,
 345                                    unsigned long floor, unsigned long ceiling)
 346 {
 347         pmd_t *pmd;
 348         unsigned long next;
 349         unsigned long start;
 350
 351         start = addr;
 352         do {
 353                 unsigned long more;
 354
 355                 pmd = pmd_offset(pud, addr);
 356                 next = pmd_addr_end(addr, end);
 357                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 358                         /*
 359                          * if it is not hugepd pointer, we should already find
 360                          * it cleared.
 361                          */
 362                         WARN_ON(!pmd_none_or_clear_bad(pmd));
 363                         continue;
 364                 }
 365                 /*
 366                  * Increment next by the size of the huge mapping since
 367                  * there may be more than one entry at this level for a
 368                  * single hugepage, but all of them point to
 369                  * the same kmem cache that holds the hugepte.
 370                  */
 371                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 372                 if (more > next)
 373                         next = more;
 374
 375                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 376                                   addr, next, floor, ceiling);
 377         } while (addr = next, addr != end);
 378
 379         start &= PUD_MASK;
 380         if (start < floor)
 381                 return;
 382         if (ceiling) {
 383                 ceiling &= PUD_MASK;
 384                 if (!ceiling)
 385                         return;
 386         }
 387         if (end - 1 > ceiling - 1)
 388                 return;
 389
 390         pmd = pmd_offset(pud, start);
 391         pud_clear(pud);
 392         pmd_free_tlb(tlb, pmd, start);
 393         mm_dec_nr_pmds(tlb->mm);
 394 }
 395
 396 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 397                                    unsigned long addr, unsigned long end,
 398                                    unsigned long floor, unsigned long ceiling)
 399 {
 400         pud_t *pud;
 401         unsigned long next;
 402         unsigned long start;
 403
 404         start = addr;
 405         do {
 406                 pud = pud_offset(pgd, addr);
 407                 next = pud_addr_end(addr, end);
 408                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 409                         if (pud_none_or_clear_bad(pud))
 410                                 continue;
 411                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 412                                                ceiling);
 413                 } else {
 414                         unsigned long more;
 415                         /*
 416                          * Increment next by the size of the huge mapping since
 417                          * there may be more than one entry at this level for a
 418                          * single hugepage, but all of them point to
 419                          * the same kmem cache that holds the hugepte.
 420                          */
 421                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 422                         if (more > next)
 423                                 next = more;
 424
 425                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 426                                           addr, next, floor, ceiling);
 427                 }
 428         } while (addr = next, addr != end);
 429
 430         start &= PGDIR_MASK;
 431         if (start < floor)
 432                 return;
 433         if (ceiling) {
 434                 ceiling &= PGDIR_MASK;
 435                 if (!ceiling)
 436                         return;
 437         }
 438         if (end - 1 > ceiling - 1)
 439                 return;
 440
 441         pud = pud_offset(pgd, start);
 442         pgd_clear(pgd);
 443         pud_free_tlb(tlb, pud, start);
 444         mm_dec_nr_puds(tlb->mm);
 445 }
 446
 447 /*
 448  * This function frees user-level page tables of a process.
 449  */
 450 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 451                             unsigned long addr, unsigned long end,
 452                             unsigned long floor, unsigned long ceiling)
 453 {
 454         pgd_t *pgd;
 455         unsigned long next;
 456
 457         /*
 458          * Because there are a number of different possible pagetable
 459          * layouts for hugepage ranges, we limit knowledge of how
 460          * things should be laid out to the allocation path
 461          * (huge_pte_alloc(), above).  Everything else works out the
 462          * structure as it goes from information in the hugepd
 463          * pointers.  That means that we can't here use the
 464          * optimization used in the normal page free_pgd_range(), of
 465          * checking whether we're actually covering a large enough
 466          * range to have to do anything at the top level of the walk
 467          * instead of at the bottom.
 468          *
 469          * To make sense of this, you should probably go read the big
 470          * block comment at the top of the normal free_pgd_range(),
 471          * too.
 472          */
 473
 474         do {
 475                 next = pgd_addr_end(addr, end);
 476                 pgd = pgd_offset(tlb->mm, addr);
 477                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 478                         if (pgd_none_or_clear_bad(pgd))
 479                                 continue;
 480                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 481                 } else {
 482                         unsigned long more;
 483                         /*
 484                          * Increment next by the size of the huge mapping since
 485                          * there may be more than one entry at the pgd level
 486                          * for a single hugepage, but all of them point to the
 487                          * same kmem cache that holds the hugepte.
 488                          */
 489                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 490                         if (more > next)
 491                                 next = more;
 492
 493                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 494                                           addr, next, floor, ceiling);
 495                 }
 496         } while (addr = next, addr != end);
 497 }
 498
 499 struct page *follow_huge_pd(struct vm_area_struct *vma,
 500                             unsigned long address, hugepd_t hpd,
 501                             int flags, int pdshift)
 502 {
 503         pte_t *ptep;
 504         spinlock_t *ptl;
 505         struct page *page = NULL;
 506         unsigned long mask;
 507         int shift = hugepd_shift(hpd);
 508         struct mm_struct *mm = vma->vm_mm;
 509
 510 retry:
 511         /*
 512          * hugepage directory entries are protected by mm->page_table_lock
 513          * Use this instead of huge_pte_lockptr
 514          */
 515         ptl = &mm->page_table_lock;
 516         spin_lock(ptl);
 517
 518         ptep = hugepte_offset(hpd, address, pdshift);
 519         if (pte_present(*ptep)) {
 520                 mask = (1UL << shift) - 1;
 521                 page = pte_page(*ptep);
 522                 page += ((address & mask) >> PAGE_SHIFT);
 523                 if (flags & FOLL_GET)
 524                         get_page(page);
 525         } else {
 526                 if (is_hugetlb_entry_migration(*ptep)) {
 527                         spin_unlock(ptl);
 528                         __migration_entry_wait(mm, ptep, ptl);
 529                         goto retry;
 530                 }
 531         }
 532         spin_unlock(ptl);
 533         return page;
 534 }
 535
 536 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 537                                       unsigned long sz)
 538 {
 539         unsigned long __boundary = (addr + sz) & ~(sz-1);
 540         return (__boundary - 1 < end - 1) ? __boundary : end;
 541 }
 542
 543 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
 544                 unsigned long end, int write, struct page **pages, int *nr)
 545 {
 546         pte_t *ptep;
 547         unsigned long sz = 1UL << hugepd_shift(hugepd);
 548         unsigned long next;
 549
 550         ptep = hugepte_offset(hugepd, addr, pdshift);
 551         do {
 552                 next = hugepte_addr_end(addr, end, sz);
 553                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 554                         return 0;
 555         } while (ptep++, addr = next, addr != end);
 556
 557         return 1;
 558 }
 559
 560 #ifdef CONFIG_PPC_MM_SLICES
 561 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 562                                         unsigned long len, unsigned long pgoff,
 563                                         unsigned long flags)
 564 {
 565         struct hstate *hstate = hstate_file(file);
 566         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 567
 568 #ifdef CONFIG_PPC_RADIX_MMU
 569         if (radix_enabled())
 570                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 571                                                        pgoff, flags);
 572 #endif
 573         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 574 }
 575 #endif
 576
 577 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 578 {
 579 #ifdef CONFIG_PPC_MM_SLICES
 580         /* With radix we don't use slice, so derive it from vma*/
 581         if (!radix_enabled()) {
 582                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 583
 584                 return 1UL << mmu_psize_to_shift(psize);
 585         }
 586 #endif
 587         return vma_kernel_pagesize(vma);
 588 }
 589
 590 static inline bool is_power_of_4(unsigned long x)
 591 {
 592         if (is_power_of_2(x))
 593                 return (__ilog2(x) % 2) ? false : true;
 594         return false;
 595 }
 596
 597 static int __init add_huge_page_size(unsigned long long size)
 598 {
 599         int shift = __ffs(size);
 600         int mmu_psize;
 601
 602         /* Check that it is a page size supported by the hardware and
 603          * that it fits within pagetable and slice limits. */
 604         if (size <= PAGE_SIZE)
 605                 return -EINVAL;
 606 #if defined(CONFIG_PPC_FSL_BOOK3E)
 607         if (!is_power_of_4(size))
 608                 return -EINVAL;
 609 #elif !defined(CONFIG_PPC_8xx)
 610         if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
 611                 return -EINVAL;
 612 #endif
 613
 614         if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
 615                 return -EINVAL;
 616
 617 #ifdef CONFIG_PPC_BOOK3S_64
 618         /*
 619          * We need to make sure that for different page sizes reported by
 620          * firmware we only add hugetlb support for page sizes that can be
 621          * supported by linux page table layout.
 622          * For now we have
 623          * Radix: 2M
 624          * Hash: 16M and 16G
 625          */
 626         if (radix_enabled()) {
 627                 if (mmu_psize != MMU_PAGE_2M) {
 628                         if (cpu_has_feature(CPU_FTR_POWER9_DD1) ||
 629                             (mmu_psize != MMU_PAGE_1G))
 630                                 return -EINVAL;
 631                 }
 632         } else {
 633                 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
 634                         return -EINVAL;
 635         }
 636 #endif
 637
 638         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 639
 640         /* Return if huge page size has already been setup */
 641         if (size_to_hstate(size))
 642                 return 0;
 643
 644         hugetlb_add_hstate(shift - PAGE_SHIFT);
 645
 646         return 0;
 647 }
 648
 649 static int __init hugepage_setup_sz(char *str)
 650 {
 651         unsigned long long size;
 652
 653         size = memparse(str, &str);
 654
 655         if (add_huge_page_size(size) != 0) {
 656                 hugetlb_bad_size();
 657                 pr_err("Invalid huge page size specified(%llu)\n", size);
 658         }
 659
 660         return 1;
 661 }
 662 __setup("hugepagesz=", hugepage_setup_sz);
 663
 664 struct kmem_cache *hugepte_cache;
 665 static int __init hugetlbpage_init(void)
 666 {
 667         int psize;
 668
 669         if (hugetlb_disabled) {
 670                 pr_info("HugeTLB support is disabled!\n");
 671                 return 0;
 672         }
 673
 674 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
 675         if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
 676                 return -ENODEV;
 677 #endif
 678         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 679                 unsigned shift;
 680                 unsigned pdshift;
 681
 682                 if (!mmu_psize_defs[psize].shift)
 683                         continue;
 684
 685                 shift = mmu_psize_to_shift(psize);
 686
 687 #ifdef CONFIG_PPC_BOOK3S_64
 688                 if (shift > PGDIR_SHIFT)
 689                         continue;
 690                 else if (shift > PUD_SHIFT)
 691                         pdshift = PGDIR_SHIFT;
 692                 else if (shift > PMD_SHIFT)
 693                         pdshift = PUD_SHIFT;
 694                 else
 695                         pdshift = PMD_SHIFT;
 696 #else
 697                 if (shift < HUGEPD_PUD_SHIFT)
 698                         pdshift = PMD_SHIFT;
 699                 else if (shift < HUGEPD_PGD_SHIFT)
 700                         pdshift = PUD_SHIFT;
 701                 else
 702                         pdshift = PGDIR_SHIFT;
 703 #endif
 704
 705                 if (add_huge_page_size(1ULL << shift) < 0)
 706                         continue;
 707                 /*
 708                  * if we have pdshift and shift value same, we don't
 709                  * use pgt cache for hugepd.
 710                  */
 711                 if (pdshift > shift)
 712                         pgtable_cache_add(pdshift - shift, NULL);
 713 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 714                 else if (!hugepte_cache) {
 715                         /*
 716                          * Create a kmem cache for hugeptes.  The bottom bits in
 717                          * the pte have size information encoded in them, so
 718                          * align them to allow this
 719                          */
 720                         hugepte_cache = kmem_cache_create("hugepte-cache",
 721                                                           sizeof(pte_t),
 722                                                           HUGEPD_SHIFT_MASK + 1,
 723                                                           0, NULL);
 724                         if (hugepte_cache == NULL)
 725                                 panic("%s: Unable to create kmem cache "
 726                                       "for hugeptes\n", __func__);
 727
 728                 }
 729 #endif
 730         }
 731
 732 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 733         /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
 734         if (mmu_psize_defs[MMU_PAGE_4M].shift)
 735                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
 736         else if (mmu_psize_defs[MMU_PAGE_512K].shift)
 737                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
 738 #else
 739         /* Set default large page size. Currently, we pick 16M or 1M
 740          * depending on what is available
 741          */
 742         if (mmu_psize_defs[MMU_PAGE_16M].shift)
 743                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
 744         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 745                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
 746         else if (mmu_psize_defs[MMU_PAGE_2M].shift)
 747                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
 748 #endif
 749         return 0;
 750 }
 751
 752 arch_initcall(hugetlbpage_init);
 753
 754 void flush_dcache_icache_hugepage(struct page *page)
 755 {
 756         int i;
 757         void *start;
 758
 759         BUG_ON(!PageCompound(page));
 760
 761         for (i = 0; i < (1UL << compound_order(page)); i++) {
 762                 if (!PageHighMem(page)) {
 763                         __flush_dcache_icache(page_address(page+i));
 764                 } else {
 765                         start = kmap_atomic(page+i);
 766                         __flush_dcache_icache(start);
 767                         kunmap_atomic(start);
 768                 }
 769         }
 770 }
 771
 772 #endif /* CONFIG_HUGETLB_PAGE */
 773
 774 /*
 775  * We have 4 cases for pgds and pmds:
 776  * (1) invalid (all zeroes)
 777  * (2) pointer to next table, as normal; bottom 6 bits == 0
 778  * (3) leaf pte for huge page _PAGE_PTE set
 779  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
 780  *
 781  * So long as we atomically load page table pointers we are safe against teardown,
 782  * we can follow the address down to the the page and take a ref on it.
 783  * This function need to be called with interrupts disabled. We use this variant
 784  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
 785  */
 786 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
 787                         bool *is_thp, unsigned *hpage_shift)
 788 {
 789         pgd_t pgd, *pgdp;
 790         pud_t pud, *pudp;
 791         pmd_t pmd, *pmdp;
 792         pte_t *ret_pte;
 793         hugepd_t *hpdp = NULL;
 794         unsigned pdshift = PGDIR_SHIFT;
 795
 796         if (hpage_shift)
 797                 *hpage_shift = 0;
 798
 799         if (is_thp)
 800                 *is_thp = false;
 801
 802         pgdp = pgdir + pgd_index(ea);
 803         pgd  = READ_ONCE(*pgdp);
 804         /*
 805          * Always operate on the local stack value. This make sure the
 806          * value don't get updated by a parallel THP split/collapse,
 807          * page fault or a page unmap. The return pte_t * is still not
 808          * stable. So should be checked there for above conditions.
 809          */
 810         if (pgd_none(pgd))
 811                 return NULL;
 812         else if (pgd_huge(pgd)) {
 813                 ret_pte = (pte_t *) pgdp;
 814                 goto out;
 815         } else if (is_hugepd(__hugepd(pgd_val(pgd))))
 816                 hpdp = (hugepd_t *)&pgd;
 817         else {
 818                 /*
 819                  * Even if we end up with an unmap, the pgtable will not
 820                  * be freed, because we do an rcu free and here we are
 821                  * irq disabled
 822                  */
 823                 pdshift = PUD_SHIFT;
 824                 pudp = pud_offset(&pgd, ea);
 825                 pud  = READ_ONCE(*pudp);
 826
 827                 if (pud_none(pud))
 828                         return NULL;
 829                 else if (pud_huge(pud)) {
 830                         ret_pte = (pte_t *) pudp;
 831                         goto out;
 832                 } else if (is_hugepd(__hugepd(pud_val(pud))))
 833                         hpdp = (hugepd_t *)&pud;
 834                 else {
 835                         pdshift = PMD_SHIFT;
 836                         pmdp = pmd_offset(&pud, ea);
 837                         pmd  = READ_ONCE(*pmdp);
 838                         /*
 839                          * A hugepage collapse is captured by pmd_none, because
 840                          * it mark the pmd none and do a hpte invalidate.
 841                          */
 842                         if (pmd_none(pmd))
 843                                 return NULL;
 844
 845                         if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
 846                                 if (is_thp)
 847                                         *is_thp = true;
 848                                 ret_pte = (pte_t *) pmdp;
 849                                 goto out;
 850                         }
 851
 852                         if (pmd_huge(pmd)) {
 853                                 ret_pte = (pte_t *) pmdp;
 854                                 goto out;
 855                         } else if (is_hugepd(__hugepd(pmd_val(pmd))))
 856                                 hpdp = (hugepd_t *)&pmd;
 857                         else
 858                                 return pte_offset_kernel(&pmd, ea);
 859                 }
 860         }
 861         if (!hpdp)
 862                 return NULL;
 863
 864         ret_pte = hugepte_offset(*hpdp, ea, pdshift);
 865         pdshift = hugepd_shift(*hpdp);
 866 out:
 867         if (hpage_shift)
 868                 *hpage_shift = pdshift;
 869         return ret_pte;
 870 }
 871 EXPORT_SYMBOL_GPL(__find_linux_pte);
 872
 873 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 874                 unsigned long end, int write, struct page **pages, int *nr)
 875 {
 876         unsigned long pte_end;
 877         struct page *head, *page;
 878         pte_t pte;
 879         int refs;
 880
 881         pte_end = (addr + sz) & ~(sz-1);
 882         if (pte_end < end)
 883                 end = pte_end;
 884
 885         pte = READ_ONCE(*ptep);
 886
 887         if (!pte_access_permitted(pte, write))
 888                 return 0;
 889
 890         /* hugepages are never "special" */
 891         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 892
 893         refs = 0;
 894         head = pte_page(pte);
 895
 896         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
 897         do {
 898                 VM_BUG_ON(compound_head(page) != head);
 899                 pages[*nr] = page;
 900                 (*nr)++;
 901                 page++;
 902                 refs++;
 903         } while (addr += PAGE_SIZE, addr != end);
 904
 905         if (!page_cache_add_speculative(head, refs)) {
 906                 *nr -= refs;
 907                 return 0;
 908         }
 909
 910         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 911                 /* Could be optimized better */
 912                 *nr -= refs;
 913                 while (refs--)
 914                         put_page(head);
 915                 return 0;
 916         }
 917
 918         return 1;
 919 }