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
  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 #define PTE_T_ORDER     (__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
  46
  47 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  48 {
  49         /*
  50          * Only called for hugetlbfs pages, hence can ignore THP and the
  51          * irq disabled walk.
  52          */
  53         return __find_linux_pte(mm->pgd, addr, NULL, NULL);
  54 }
  55
  56 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  57                            unsigned long address, unsigned int pdshift,
  58                            unsigned int pshift, spinlock_t *ptl)
  59 {
  60         struct kmem_cache *cachep;
  61         pte_t *new;
  62         int i;
  63         int num_hugepd;
  64
  65         if (pshift >= pdshift) {
  66                 cachep = PGT_CACHE(PTE_T_ORDER);
  67                 num_hugepd = 1 << (pshift - pdshift);
  68         } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
  69                 cachep = PGT_CACHE(PTE_INDEX_SIZE);
  70                 num_hugepd = 1;
  71         } else {
  72                 cachep = PGT_CACHE(pdshift - pshift);
  73                 num_hugepd = 1;
  74         }
  75
  76         new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  77
  78         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  79         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  80
  81         if (! new)
  82                 return -ENOMEM;
  83
  84         /*
  85          * Make sure other cpus find the hugepd set only after a
  86          * properly initialized page table is visible to them.
  87          * For more details look for comment in __pte_alloc().
  88          */
  89         smp_wmb();
  90
  91         spin_lock(ptl);
  92         /*
  93          * We have multiple higher-level entries that point to the same
  94          * actual pte location.  Fill in each as we go and backtrack on error.
  95          * We need all of these so the DTLB pgtable walk code can find the
  96          * right higher-level entry without knowing if it's a hugepage or not.
  97          */
  98         for (i = 0; i < num_hugepd; i++, hpdp++) {
  99                 if (unlikely(!hugepd_none(*hpdp)))
 100                         break;
 101                 else {
 102 #ifdef CONFIG_PPC_BOOK3S_64
 103                         *hpdp = __hugepd(__pa(new) | HUGEPD_VAL_BITS |
 104                                          (shift_to_mmu_psize(pshift) << 2));
 105 #elif defined(CONFIG_PPC_8xx)
 106                         *hpdp = __hugepd(__pa(new) | _PMD_USER |
 107                                          (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
 108                                           _PMD_PAGE_512K) | _PMD_PRESENT);
 109 #else
 110                         /* We use the old format for PPC_FSL_BOOK3E */
 111                         *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
 112 #endif
 113                 }
 114         }
 115         /* If we bailed from the for loop early, an error occurred, clean up */
 116         if (i < num_hugepd) {
 117                 for (i = i - 1 ; i >= 0; i--, hpdp--)
 118                         *hpdp = __hugepd(0);
 119                 kmem_cache_free(cachep, new);
 120         } else {
 121                 kmemleak_ignore(new);
 122         }
 123         spin_unlock(ptl);
 124         return 0;
 125 }
 126
 127 /*
 128  * At this point we do the placement change only for BOOK3S 64. This would
 129  * possibly work on other subarchs.
 130  */
 131 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 132 {
 133         pgd_t *pg;
 134         pud_t *pu;
 135         pmd_t *pm;
 136         hugepd_t *hpdp = NULL;
 137         unsigned pshift = __ffs(sz);
 138         unsigned pdshift = PGDIR_SHIFT;
 139         spinlock_t *ptl;
 140
 141         addr &= ~(sz-1);
 142         pg = pgd_offset(mm, addr);
 143
 144 #ifdef CONFIG_PPC_BOOK3S_64
 145         if (pshift == PGDIR_SHIFT)
 146                 /* 16GB huge page */
 147                 return (pte_t *) pg;
 148         else if (pshift > PUD_SHIFT) {
 149                 /*
 150                  * We need to use hugepd table
 151                  */
 152                 ptl = &mm->page_table_lock;
 153                 hpdp = (hugepd_t *)pg;
 154         } else {
 155                 pdshift = PUD_SHIFT;
 156                 pu = pud_alloc(mm, pg, addr);
 157                 if (pshift == PUD_SHIFT)
 158                         return (pte_t *)pu;
 159                 else if (pshift > PMD_SHIFT) {
 160                         ptl = pud_lockptr(mm, pu);
 161                         hpdp = (hugepd_t *)pu;
 162                 } else {
 163                         pdshift = PMD_SHIFT;
 164                         pm = pmd_alloc(mm, pu, addr);
 165                         if (pshift == PMD_SHIFT)
 166                                 /* 16MB hugepage */
 167                                 return (pte_t *)pm;
 168                         else {
 169                                 ptl = pmd_lockptr(mm, pm);
 170                                 hpdp = (hugepd_t *)pm;
 171                         }
 172                 }
 173         }
 174 #else
 175         if (pshift >= PGDIR_SHIFT) {
 176                 ptl = &mm->page_table_lock;
 177                 hpdp = (hugepd_t *)pg;
 178         } else {
 179                 pdshift = PUD_SHIFT;
 180                 pu = pud_alloc(mm, pg, addr);
 181                 if (pshift >= PUD_SHIFT) {
 182                         ptl = pud_lockptr(mm, pu);
 183                         hpdp = (hugepd_t *)pu;
 184                 } else {
 185                         pdshift = PMD_SHIFT;
 186                         pm = pmd_alloc(mm, pu, addr);
 187                         ptl = pmd_lockptr(mm, pm);
 188                         hpdp = (hugepd_t *)pm;
 189                 }
 190         }
 191 #endif
 192         if (!hpdp)
 193                 return NULL;
 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 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 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[0];
 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 if (IS_ENABLED(CONFIG_PPC_8xx))
 338                 pgtable_free_tlb(tlb, hugepte,
 339                                  get_hugepd_cache_index(PTE_INDEX_SIZE));
 340         else
 341                 pgtable_free_tlb(tlb, hugepte,
 342                                  get_hugepd_cache_index(pdshift - shift));
 343 }
 344
 345 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 346                                    unsigned long addr, unsigned long end,
 347                                    unsigned long floor, unsigned long ceiling)
 348 {
 349         pmd_t *pmd;
 350         unsigned long next;
 351         unsigned long start;
 352
 353         start = addr;
 354         do {
 355                 unsigned long more;
 356
 357                 pmd = pmd_offset(pud, addr);
 358                 next = pmd_addr_end(addr, end);
 359                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 360                         /*
 361                          * if it is not hugepd pointer, we should already find
 362                          * it cleared.
 363                          */
 364                         WARN_ON(!pmd_none_or_clear_bad(pmd));
 365                         continue;
 366                 }
 367                 /*
 368                  * Increment next by the size of the huge mapping since
 369                  * there may be more than one entry at this level for a
 370                  * single hugepage, but all of them point to
 371                  * the same kmem cache that holds the hugepte.
 372                  */
 373                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 374                 if (more > next)
 375                         next = more;
 376
 377                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 378                                   addr, next, floor, ceiling);
 379         } while (addr = next, addr != end);
 380
 381         start &= PUD_MASK;
 382         if (start < floor)
 383                 return;
 384         if (ceiling) {
 385                 ceiling &= PUD_MASK;
 386                 if (!ceiling)
 387                         return;
 388         }
 389         if (end - 1 > ceiling - 1)
 390                 return;
 391
 392         pmd = pmd_offset(pud, start);
 393         pud_clear(pud);
 394         pmd_free_tlb(tlb, pmd, start);
 395         mm_dec_nr_pmds(tlb->mm);
 396 }
 397
 398 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 399                                    unsigned long addr, unsigned long end,
 400                                    unsigned long floor, unsigned long ceiling)
 401 {
 402         pud_t *pud;
 403         unsigned long next;
 404         unsigned long start;
 405
 406         start = addr;
 407         do {
 408                 pud = pud_offset(pgd, addr);
 409                 next = pud_addr_end(addr, end);
 410                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 411                         if (pud_none_or_clear_bad(pud))
 412                                 continue;
 413                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 414                                                ceiling);
 415                 } else {
 416                         unsigned long more;
 417                         /*
 418                          * Increment next by the size of the huge mapping since
 419                          * there may be more than one entry at this level for a
 420                          * single hugepage, but all of them point to
 421                          * the same kmem cache that holds the hugepte.
 422                          */
 423                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 424                         if (more > next)
 425                                 next = more;
 426
 427                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 428                                           addr, next, floor, ceiling);
 429                 }
 430         } while (addr = next, addr != end);
 431
 432         start &= PGDIR_MASK;
 433         if (start < floor)
 434                 return;
 435         if (ceiling) {
 436                 ceiling &= PGDIR_MASK;
 437                 if (!ceiling)
 438                         return;
 439         }
 440         if (end - 1 > ceiling - 1)
 441                 return;
 442
 443         pud = pud_offset(pgd, start);
 444         pgd_clear(pgd);
 445         pud_free_tlb(tlb, pud, start);
 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         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                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 480                         if (pgd_none_or_clear_bad(pgd))
 481                                 continue;
 482                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 483                 } else {
 484                         unsigned long more;
 485                         /*
 486                          * Increment next by the size of the huge mapping since
 487                          * there may be more than one entry at the pgd level
 488                          * for a single hugepage, but all of them point to the
 489                          * same kmem cache that holds the hugepte.
 490                          */
 491                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 492                         if (more > next)
 493                                 next = more;
 494
 495                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 496                                           addr, next, floor, ceiling);
 497                 }
 498         } while (addr = next, addr != end);
 499 }
 500
 501 struct page *follow_huge_pd(struct vm_area_struct *vma,
 502                             unsigned long address, hugepd_t hpd,
 503                             int flags, int pdshift)
 504 {
 505         pte_t *ptep;
 506         spinlock_t *ptl;
 507         struct page *page = NULL;
 508         unsigned long mask;
 509         int shift = hugepd_shift(hpd);
 510         struct mm_struct *mm = vma->vm_mm;
 511
 512 retry:
 513         /*
 514          * hugepage directory entries are protected by mm->page_table_lock
 515          * Use this instead of huge_pte_lockptr
 516          */
 517         ptl = &mm->page_table_lock;
 518         spin_lock(ptl);
 519
 520         ptep = hugepte_offset(hpd, address, pdshift);
 521         if (pte_present(*ptep)) {
 522                 mask = (1UL << shift) - 1;
 523                 page = pte_page(*ptep);
 524                 page += ((address & mask) >> PAGE_SHIFT);
 525                 if (flags & FOLL_GET)
 526                         get_page(page);
 527         } else {
 528                 if (is_hugetlb_entry_migration(*ptep)) {
 529                         spin_unlock(ptl);
 530                         __migration_entry_wait(mm, ptep, ptl);
 531                         goto retry;
 532                 }
 533         }
 534         spin_unlock(ptl);
 535         return page;
 536 }
 537
 538 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 539                                       unsigned long sz)
 540 {
 541         unsigned long __boundary = (addr + sz) & ~(sz-1);
 542         return (__boundary - 1 < end - 1) ? __boundary : end;
 543 }
 544
 545 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
 546                 unsigned long end, int write, struct page **pages, int *nr)
 547 {
 548         pte_t *ptep;
 549         unsigned long sz = 1UL << hugepd_shift(hugepd);
 550         unsigned long next;
 551
 552         ptep = hugepte_offset(hugepd, addr, pdshift);
 553         do {
 554                 next = hugepte_addr_end(addr, end, sz);
 555                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 556                         return 0;
 557         } while (ptep++, addr = next, addr != end);
 558
 559         return 1;
 560 }
 561
 562 #ifdef CONFIG_PPC_MM_SLICES
 563 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 564                                         unsigned long len, unsigned long pgoff,
 565                                         unsigned long flags)
 566 {
 567         struct hstate *hstate = hstate_file(file);
 568         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 569
 570 #ifdef CONFIG_PPC_RADIX_MMU
 571         if (radix_enabled())
 572                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 573                                                        pgoff, flags);
 574 #endif
 575         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 576 }
 577 #endif
 578
 579 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 580 {
 581 #ifdef CONFIG_PPC_MM_SLICES
 582         /* With radix we don't use slice, so derive it from vma*/
 583         if (!radix_enabled()) {
 584                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 585
 586                 return 1UL << mmu_psize_to_shift(psize);
 587         }
 588 #endif
 589         return vma_kernel_pagesize(vma);
 590 }
 591
 592 static inline bool is_power_of_4(unsigned long x)
 593 {
 594         if (is_power_of_2(x))
 595                 return (__ilog2(x) % 2) ? false : true;
 596         return false;
 597 }
 598
 599 static int __init add_huge_page_size(unsigned long long size)
 600 {
 601         int shift = __ffs(size);
 602         int mmu_psize;
 603
 604         /* Check that it is a page size supported by the hardware and
 605          * that it fits within pagetable and slice limits. */
 606         if (size <= PAGE_SIZE)
 607                 return -EINVAL;
 608 #if defined(CONFIG_PPC_FSL_BOOK3E)
 609         if (!is_power_of_4(size))
 610                 return -EINVAL;
 611 #elif !defined(CONFIG_PPC_8xx)
 612         if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
 613                 return -EINVAL;
 614 #endif
 615
 616         if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
 617                 return -EINVAL;
 618
 619 #ifdef CONFIG_PPC_BOOK3S_64
 620         /*
 621          * We need to make sure that for different page sizes reported by
 622          * firmware we only add hugetlb support for page sizes that can be
 623          * supported by linux page table layout.
 624          * For now we have
 625          * Radix: 2M and 1G
 626          * Hash: 16M and 16G
 627          */
 628         if (radix_enabled()) {
 629                 if (mmu_psize != MMU_PAGE_2M && mmu_psize != MMU_PAGE_1G)
 630                         return -EINVAL;
 631         } else {
 632                 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
 633                         return -EINVAL;
 634         }
 635 #endif
 636
 637         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 638
 639         /* Return if huge page size has already been setup */
 640         if (size_to_hstate(size))
 641                 return 0;
 642
 643         hugetlb_add_hstate(shift - PAGE_SHIFT);
 644
 645         return 0;
 646 }
 647
 648 static int __init hugepage_setup_sz(char *str)
 649 {
 650         unsigned long long size;
 651
 652         size = memparse(str, &str);
 653
 654         if (add_huge_page_size(size) != 0) {
 655                 hugetlb_bad_size();
 656                 pr_err("Invalid huge page size specified(%llu)\n", size);
 657         }
 658
 659         return 1;
 660 }
 661 __setup("hugepagesz=", hugepage_setup_sz);
 662
 663 static int __init hugetlbpage_init(void)
 664 {
 665         int psize;
 666
 667         if (hugetlb_disabled) {
 668                 pr_info("HugeTLB support is disabled!\n");
 669                 return 0;
 670         }
 671
 672 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
 673         if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
 674                 return -ENODEV;
 675 #endif
 676         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 677                 unsigned shift;
 678                 unsigned pdshift;
 679
 680                 if (!mmu_psize_defs[psize].shift)
 681                         continue;
 682
 683                 shift = mmu_psize_to_shift(psize);
 684
 685 #ifdef CONFIG_PPC_BOOK3S_64
 686                 if (shift > PGDIR_SHIFT)
 687                         continue;
 688                 else if (shift > PUD_SHIFT)
 689                         pdshift = PGDIR_SHIFT;
 690                 else if (shift > PMD_SHIFT)
 691                         pdshift = PUD_SHIFT;
 692                 else
 693                         pdshift = PMD_SHIFT;
 694 #else
 695                 if (shift < PUD_SHIFT)
 696                         pdshift = PMD_SHIFT;
 697                 else if (shift < PGDIR_SHIFT)
 698                         pdshift = PUD_SHIFT;
 699                 else
 700                         pdshift = PGDIR_SHIFT;
 701 #endif
 702
 703                 if (add_huge_page_size(1ULL << shift) < 0)
 704                         continue;
 705                 /*
 706                  * if we have pdshift and shift value same, we don't
 707                  * use pgt cache for hugepd.
 708                  */
 709                 if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
 710                         pgtable_cache_add(PTE_INDEX_SIZE);
 711                 else if (pdshift > shift)
 712                         pgtable_cache_add(pdshift - shift);
 713 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 714                 else
 715                         pgtable_cache_add(PTE_T_ORDER);
 716 #endif
 717         }
 718
 719 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 720         /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
 721         if (mmu_psize_defs[MMU_PAGE_4M].shift)
 722                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
 723         else if (mmu_psize_defs[MMU_PAGE_512K].shift)
 724                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
 725 #else
 726         /* Set default large page size. Currently, we pick 16M or 1M
 727          * depending on what is available
 728          */
 729         if (mmu_psize_defs[MMU_PAGE_16M].shift)
 730                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
 731         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 732                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
 733         else if (mmu_psize_defs[MMU_PAGE_2M].shift)
 734                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
 735 #endif
 736         return 0;
 737 }
 738
 739 arch_initcall(hugetlbpage_init);
 740
 741 void flush_dcache_icache_hugepage(struct page *page)
 742 {
 743         int i;
 744         void *start;
 745
 746         BUG_ON(!PageCompound(page));
 747
 748         for (i = 0; i < (1UL << compound_order(page)); i++) {
 749                 if (!PageHighMem(page)) {
 750                         __flush_dcache_icache(page_address(page+i));
 751                 } else {
 752                         start = kmap_atomic(page+i);
 753                         __flush_dcache_icache(start);
 754                         kunmap_atomic(start);
 755                 }
 756         }
 757 }
 758
 759 #endif /* CONFIG_HUGETLB_PAGE */
 760
 761 /*
 762  * We have 4 cases for pgds and pmds:
 763  * (1) invalid (all zeroes)
 764  * (2) pointer to next table, as normal; bottom 6 bits == 0
 765  * (3) leaf pte for huge page _PAGE_PTE set
 766  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
 767  *
 768  * So long as we atomically load page table pointers we are safe against teardown,
 769  * we can follow the address down to the the page and take a ref on it.
 770  * This function need to be called with interrupts disabled. We use this variant
 771  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
 772  */
 773 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
 774                         bool *is_thp, unsigned *hpage_shift)
 775 {
 776         pgd_t pgd, *pgdp;
 777         pud_t pud, *pudp;
 778         pmd_t pmd, *pmdp;
 779         pte_t *ret_pte;
 780         hugepd_t *hpdp = NULL;
 781         unsigned pdshift = PGDIR_SHIFT;
 782
 783         if (hpage_shift)
 784                 *hpage_shift = 0;
 785
 786         if (is_thp)
 787                 *is_thp = false;
 788
 789         pgdp = pgdir + pgd_index(ea);
 790         pgd  = READ_ONCE(*pgdp);
 791         /*
 792          * Always operate on the local stack value. This make sure the
 793          * value don't get updated by a parallel THP split/collapse,
 794          * page fault or a page unmap. The return pte_t * is still not
 795          * stable. So should be checked there for above conditions.
 796          */
 797         if (pgd_none(pgd))
 798                 return NULL;
 799         else if (pgd_huge(pgd)) {
 800                 ret_pte = (pte_t *) pgdp;
 801                 goto out;
 802         } else if (is_hugepd(__hugepd(pgd_val(pgd))))
 803                 hpdp = (hugepd_t *)&pgd;
 804         else {
 805                 /*
 806                  * Even if we end up with an unmap, the pgtable will not
 807                  * be freed, because we do an rcu free and here we are
 808                  * irq disabled
 809                  */
 810                 pdshift = PUD_SHIFT;
 811                 pudp = pud_offset(&pgd, ea);
 812                 pud  = READ_ONCE(*pudp);
 813
 814                 if (pud_none(pud))
 815                         return NULL;
 816                 else if (pud_huge(pud)) {
 817                         ret_pte = (pte_t *) pudp;
 818                         goto out;
 819                 } else if (is_hugepd(__hugepd(pud_val(pud))))
 820                         hpdp = (hugepd_t *)&pud;
 821                 else {
 822                         pdshift = PMD_SHIFT;
 823                         pmdp = pmd_offset(&pud, ea);
 824                         pmd  = READ_ONCE(*pmdp);
 825                         /*
 826                          * A hugepage collapse is captured by pmd_none, because
 827                          * it mark the pmd none and do a hpte invalidate.
 828                          */
 829                         if (pmd_none(pmd))
 830                                 return NULL;
 831
 832                         if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
 833                                 if (is_thp)
 834                                         *is_thp = true;
 835                                 ret_pte = (pte_t *) pmdp;
 836                                 goto out;
 837                         }
 838                         /*
 839                          * pmd_large check below will handle the swap pmd pte
 840                          * we need to do both the check because they are config
 841                          * dependent.
 842                          */
 843                         if (pmd_huge(pmd) || pmd_large(pmd)) {
 844                                 ret_pte = (pte_t *) pmdp;
 845                                 goto out;
 846                         } else if (is_hugepd(__hugepd(pmd_val(pmd))))
 847                                 hpdp = (hugepd_t *)&pmd;
 848                         else
 849                                 return pte_offset_kernel(&pmd, ea);
 850                 }
 851         }
 852         if (!hpdp)
 853                 return NULL;
 854
 855         ret_pte = hugepte_offset(*hpdp, ea, pdshift);
 856         pdshift = hugepd_shift(*hpdp);
 857 out:
 858         if (hpage_shift)
 859                 *hpage_shift = pdshift;
 860         return ret_pte;
 861 }
 862 EXPORT_SYMBOL_GPL(__find_linux_pte);
 863
 864 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 865                 unsigned long end, int write, struct page **pages, int *nr)
 866 {
 867         unsigned long pte_end;
 868         struct page *head, *page;
 869         pte_t pte;
 870         int refs;
 871
 872         pte_end = (addr + sz) & ~(sz-1);
 873         if (pte_end < end)
 874                 end = pte_end;
 875
 876         pte = READ_ONCE(*ptep);
 877
 878         if (!pte_access_permitted(pte, write))
 879                 return 0;
 880
 881         /* hugepages are never "special" */
 882         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 883
 884         refs = 0;
 885         head = pte_page(pte);
 886
 887         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
 888         do {
 889                 VM_BUG_ON(compound_head(page) != head);
 890                 pages[*nr] = page;
 891                 (*nr)++;
 892                 page++;
 893                 refs++;
 894         } while (addr += PAGE_SIZE, addr != end);
 895
 896         if (!page_cache_add_speculative(head, refs)) {
 897                 *nr -= refs;
 898                 return 0;
 899         }
 900
 901         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 902                 /* Could be optimized better */
 903                 *nr -= refs;
 904                 while (refs--)
 905                         put_page(head);
 906                 return 0;
 907         }
 908
 909         return 1;
 910 }