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