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/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,
 341                                  get_hugepd_cache_index(pdshift - shift));
 342 }
 343
 344 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 345                                    unsigned long addr, unsigned long end,
 346                                    unsigned long floor, unsigned long ceiling)
 347 {
 348         pmd_t *pmd;
 349         unsigned long next;
 350         unsigned long start;
 351
 352         start = addr;
 353         do {
 354                 unsigned long more;
 355
 356                 pmd = pmd_offset(pud, addr);
 357                 next = pmd_addr_end(addr, end);
 358                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 359                         /*
 360                          * if it is not hugepd pointer, we should already find
 361                          * it cleared.
 362                          */
 363                         WARN_ON(!pmd_none_or_clear_bad(pmd));
 364                         continue;
 365                 }
 366                 /*
 367                  * Increment next by the size of the huge mapping since
 368                  * there may be more than one entry at this level for a
 369                  * single hugepage, but all of them point to
 370                  * the same kmem cache that holds the hugepte.
 371                  */
 372                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 373                 if (more > next)
 374                         next = more;
 375
 376                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 377                                   addr, next, floor, ceiling);
 378         } while (addr = next, addr != end);
 379
 380         start &= PUD_MASK;
 381         if (start < floor)
 382                 return;
 383         if (ceiling) {
 384                 ceiling &= PUD_MASK;
 385                 if (!ceiling)
 386                         return;
 387         }
 388         if (end - 1 > ceiling - 1)
 389                 return;
 390
 391         pmd = pmd_offset(pud, start);
 392         pud_clear(pud);
 393         pmd_free_tlb(tlb, pmd, start);
 394         mm_dec_nr_pmds(tlb->mm);
 395 }
 396
 397 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 398                                    unsigned long addr, unsigned long end,
 399                                    unsigned long floor, unsigned long ceiling)
 400 {
 401         pud_t *pud;
 402         unsigned long next;
 403         unsigned long start;
 404
 405         start = addr;
 406         do {
 407                 pud = pud_offset(pgd, addr);
 408                 next = pud_addr_end(addr, end);
 409                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 410                         if (pud_none_or_clear_bad(pud))
 411                                 continue;
 412                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 413                                                ceiling);
 414                 } else {
 415                         unsigned long more;
 416                         /*
 417                          * Increment next by the size of the huge mapping since
 418                          * there may be more than one entry at this level for a
 419                          * single hugepage, but all of them point to
 420                          * the same kmem cache that holds the hugepte.
 421                          */
 422                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 423                         if (more > next)
 424                                 next = more;
 425
 426                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 427                                           addr, next, floor, ceiling);
 428                 }
 429         } while (addr = next, addr != end);
 430
 431         start &= PGDIR_MASK;
 432         if (start < floor)
 433                 return;
 434         if (ceiling) {
 435                 ceiling &= PGDIR_MASK;
 436                 if (!ceiling)
 437                         return;
 438         }
 439         if (end - 1 > ceiling - 1)
 440                 return;
 441
 442         pud = pud_offset(pgd, start);
 443         pgd_clear(pgd);
 444         pud_free_tlb(tlb, pud, start);
 445         mm_dec_nr_puds(tlb->mm);
 446 }
 447
 448 /*
 449  * This function frees user-level page tables of a process.
 450  */
 451 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 452                             unsigned long addr, unsigned long end,
 453                             unsigned long floor, unsigned long ceiling)
 454 {
 455         pgd_t *pgd;
 456         unsigned long next;
 457
 458         /*
 459          * Because there are a number of different possible pagetable
 460          * layouts for hugepage ranges, we limit knowledge of how
 461          * things should be laid out to the allocation path
 462          * (huge_pte_alloc(), above).  Everything else works out the
 463          * structure as it goes from information in the hugepd
 464          * pointers.  That means that we can't here use the
 465          * optimization used in the normal page free_pgd_range(), of
 466          * checking whether we're actually covering a large enough
 467          * range to have to do anything at the top level of the walk
 468          * instead of at the bottom.
 469          *
 470          * To make sense of this, you should probably go read the big
 471          * block comment at the top of the normal free_pgd_range(),
 472          * too.
 473          */
 474
 475         do {
 476                 next = pgd_addr_end(addr, end);
 477                 pgd = pgd_offset(tlb->mm, addr);
 478                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 479                         if (pgd_none_or_clear_bad(pgd))
 480                                 continue;
 481                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 482                 } else {
 483                         unsigned long more;
 484                         /*
 485                          * Increment next by the size of the huge mapping since
 486                          * there may be more than one entry at the pgd level
 487                          * for a single hugepage, but all of them point to the
 488                          * same kmem cache that holds the hugepte.
 489                          */
 490                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 491                         if (more > next)
 492                                 next = more;
 493
 494                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 495                                           addr, next, floor, ceiling);
 496                 }
 497         } while (addr = next, addr != end);
 498 }
 499
 500 struct page *follow_huge_pd(struct vm_area_struct *vma,
 501                             unsigned long address, hugepd_t hpd,
 502                             int flags, int pdshift)
 503 {
 504         pte_t *ptep;
 505         spinlock_t *ptl;
 506         struct page *page = NULL;
 507         unsigned long mask;
 508         int shift = hugepd_shift(hpd);
 509         struct mm_struct *mm = vma->vm_mm;
 510
 511 retry:
 512         /*
 513          * hugepage directory entries are protected by mm->page_table_lock
 514          * Use this instead of huge_pte_lockptr
 515          */
 516         ptl = &mm->page_table_lock;
 517         spin_lock(ptl);
 518
 519         ptep = hugepte_offset(hpd, address, pdshift);
 520         if (pte_present(*ptep)) {
 521                 mask = (1UL << shift) - 1;
 522                 page = pte_page(*ptep);
 523                 page += ((address & mask) >> PAGE_SHIFT);
 524                 if (flags & FOLL_GET)
 525                         get_page(page);
 526         } else {
 527                 if (is_hugetlb_entry_migration(*ptep)) {
 528                         spin_unlock(ptl);
 529                         __migration_entry_wait(mm, ptep, ptl);
 530                         goto retry;
 531                 }
 532         }
 533         spin_unlock(ptl);
 534         return page;
 535 }
 536
 537 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 538                                       unsigned long sz)
 539 {
 540         unsigned long __boundary = (addr + sz) & ~(sz-1);
 541         return (__boundary - 1 < end - 1) ? __boundary : end;
 542 }
 543
 544 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
 545                 unsigned long end, int write, struct page **pages, int *nr)
 546 {
 547         pte_t *ptep;
 548         unsigned long sz = 1UL << hugepd_shift(hugepd);
 549         unsigned long next;
 550
 551         ptep = hugepte_offset(hugepd, addr, pdshift);
 552         do {
 553                 next = hugepte_addr_end(addr, end, sz);
 554                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 555                         return 0;
 556         } while (ptep++, addr = next, addr != end);
 557
 558         return 1;
 559 }
 560
 561 #ifdef CONFIG_PPC_MM_SLICES
 562 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 563                                         unsigned long len, unsigned long pgoff,
 564                                         unsigned long flags)
 565 {
 566         struct hstate *hstate = hstate_file(file);
 567         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 568
 569 #ifdef CONFIG_PPC_RADIX_MMU
 570         if (radix_enabled())
 571                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 572                                                        pgoff, flags);
 573 #endif
 574         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 575 }
 576 #endif
 577
 578 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 579 {
 580 #ifdef CONFIG_PPC_MM_SLICES
 581         /* With radix we don't use slice, so derive it from vma*/
 582         if (!radix_enabled()) {
 583                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 584
 585                 return 1UL << mmu_psize_to_shift(psize);
 586         }
 587 #endif
 588         return vma_kernel_pagesize(vma);
 589 }
 590
 591 static inline bool is_power_of_4(unsigned long x)
 592 {
 593         if (is_power_of_2(x))
 594                 return (__ilog2(x) % 2) ? false : true;
 595         return false;
 596 }
 597
 598 static int __init add_huge_page_size(unsigned long long size)
 599 {
 600         int shift = __ffs(size);
 601         int mmu_psize;
 602
 603         /* Check that it is a page size supported by the hardware and
 604          * that it fits within pagetable and slice limits. */
 605         if (size <= PAGE_SIZE)
 606                 return -EINVAL;
 607 #if defined(CONFIG_PPC_FSL_BOOK3E)
 608         if (!is_power_of_4(size))
 609                 return -EINVAL;
 610 #elif !defined(CONFIG_PPC_8xx)
 611         if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
 612                 return -EINVAL;
 613 #endif
 614
 615         if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
 616                 return -EINVAL;
 617
 618 #ifdef CONFIG_PPC_BOOK3S_64
 619         /*
 620          * We need to make sure that for different page sizes reported by
 621          * firmware we only add hugetlb support for page sizes that can be
 622          * supported by linux page table layout.
 623          * For now we have
 624          * Radix: 2M
 625          * Hash: 16M and 16G
 626          */
 627         if (radix_enabled()) {
 628                 if (mmu_psize != MMU_PAGE_2M) {
 629                         if (cpu_has_feature(CPU_FTR_POWER9_DD1) ||
 630                             (mmu_psize != MMU_PAGE_1G))
 631                                 return -EINVAL;
 632                 }
 633         } else {
 634                 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
 635                         return -EINVAL;
 636         }
 637 #endif
 638
 639         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 640
 641         /* Return if huge page size has already been setup */
 642         if (size_to_hstate(size))
 643                 return 0;
 644
 645         hugetlb_add_hstate(shift - PAGE_SHIFT);
 646
 647         return 0;
 648 }
 649
 650 static int __init hugepage_setup_sz(char *str)
 651 {
 652         unsigned long long size;
 653
 654         size = memparse(str, &str);
 655
 656         if (add_huge_page_size(size) != 0) {
 657                 hugetlb_bad_size();
 658                 pr_err("Invalid huge page size specified(%llu)\n", size);
 659         }
 660
 661         return 1;
 662 }
 663 __setup("hugepagesz=", hugepage_setup_sz);
 664
 665 struct kmem_cache *hugepte_cache;
 666 static int __init hugetlbpage_init(void)
 667 {
 668         int psize;
 669
 670         if (hugetlb_disabled) {
 671                 pr_info("HugeTLB support is disabled!\n");
 672                 return 0;
 673         }
 674
 675 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
 676         if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
 677                 return -ENODEV;
 678 #endif
 679         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 680                 unsigned shift;
 681                 unsigned pdshift;
 682
 683                 if (!mmu_psize_defs[psize].shift)
 684                         continue;
 685
 686                 shift = mmu_psize_to_shift(psize);
 687
 688 #ifdef CONFIG_PPC_BOOK3S_64
 689                 if (shift > PGDIR_SHIFT)
 690                         continue;
 691                 else if (shift > PUD_SHIFT)
 692                         pdshift = PGDIR_SHIFT;
 693                 else if (shift > PMD_SHIFT)
 694                         pdshift = PUD_SHIFT;
 695                 else
 696                         pdshift = PMD_SHIFT;
 697 #else
 698                 if (shift < HUGEPD_PUD_SHIFT)
 699                         pdshift = PMD_SHIFT;
 700                 else if (shift < HUGEPD_PGD_SHIFT)
 701                         pdshift = PUD_SHIFT;
 702                 else
 703                         pdshift = PGDIR_SHIFT;
 704 #endif
 705
 706                 if (add_huge_page_size(1ULL << shift) < 0)
 707                         continue;
 708                 /*
 709                  * if we have pdshift and shift value same, we don't
 710                  * use pgt cache for hugepd.
 711                  */
 712                 if (pdshift > shift)
 713                         pgtable_cache_add(pdshift - shift, NULL);
 714 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 715                 else if (!hugepte_cache) {
 716                         /*
 717                          * Create a kmem cache for hugeptes.  The bottom bits in
 718                          * the pte have size information encoded in them, so
 719                          * align them to allow this
 720                          */
 721                         hugepte_cache = kmem_cache_create("hugepte-cache",
 722                                                           sizeof(pte_t),
 723                                                           HUGEPD_SHIFT_MASK + 1,
 724                                                           0, NULL);
 725                         if (hugepte_cache == NULL)
 726                                 panic("%s: Unable to create kmem cache "
 727                                       "for hugeptes\n", __func__);
 728
 729                 }
 730 #endif
 731         }
 732
 733 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 734         /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
 735         if (mmu_psize_defs[MMU_PAGE_4M].shift)
 736                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
 737         else if (mmu_psize_defs[MMU_PAGE_512K].shift)
 738                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
 739 #else
 740         /* Set default large page size. Currently, we pick 16M or 1M
 741          * depending on what is available
 742          */
 743         if (mmu_psize_defs[MMU_PAGE_16M].shift)
 744                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
 745         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 746                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
 747         else if (mmu_psize_defs[MMU_PAGE_2M].shift)
 748                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
 749 #endif
 750         return 0;
 751 }
 752
 753 arch_initcall(hugetlbpage_init);
 754
 755 void flush_dcache_icache_hugepage(struct page *page)
 756 {
 757         int i;
 758         void *start;
 759
 760         BUG_ON(!PageCompound(page));
 761
 762         for (i = 0; i < (1UL << compound_order(page)); i++) {
 763                 if (!PageHighMem(page)) {
 764                         __flush_dcache_icache(page_address(page+i));
 765                 } else {
 766                         start = kmap_atomic(page+i);
 767                         __flush_dcache_icache(start);
 768                         kunmap_atomic(start);
 769                 }
 770         }
 771 }
 772
 773 #endif /* CONFIG_HUGETLB_PAGE */
 774
 775 /*
 776  * We have 4 cases for pgds and pmds:
 777  * (1) invalid (all zeroes)
 778  * (2) pointer to next table, as normal; bottom 6 bits == 0
 779  * (3) leaf pte for huge page _PAGE_PTE set
 780  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
 781  *
 782  * So long as we atomically load page table pointers we are safe against teardown,
 783  * we can follow the address down to the the page and take a ref on it.
 784  * This function need to be called with interrupts disabled. We use this variant
 785  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
 786  */
 787 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
 788                         bool *is_thp, unsigned *hpage_shift)
 789 {
 790         pgd_t pgd, *pgdp;
 791         pud_t pud, *pudp;
 792         pmd_t pmd, *pmdp;
 793         pte_t *ret_pte;
 794         hugepd_t *hpdp = NULL;
 795         unsigned pdshift = PGDIR_SHIFT;
 796
 797         if (hpage_shift)
 798                 *hpage_shift = 0;
 799
 800         if (is_thp)
 801                 *is_thp = false;
 802
 803         pgdp = pgdir + pgd_index(ea);
 804         pgd  = READ_ONCE(*pgdp);
 805         /*
 806          * Always operate on the local stack value. This make sure the
 807          * value don't get updated by a parallel THP split/collapse,
 808          * page fault or a page unmap. The return pte_t * is still not
 809          * stable. So should be checked there for above conditions.
 810          */
 811         if (pgd_none(pgd))
 812                 return NULL;
 813         else if (pgd_huge(pgd)) {
 814                 ret_pte = (pte_t *) pgdp;
 815                 goto out;
 816         } else if (is_hugepd(__hugepd(pgd_val(pgd))))
 817                 hpdp = (hugepd_t *)&pgd;
 818         else {
 819                 /*
 820                  * Even if we end up with an unmap, the pgtable will not
 821                  * be freed, because we do an rcu free and here we are
 822                  * irq disabled
 823                  */
 824                 pdshift = PUD_SHIFT;
 825                 pudp = pud_offset(&pgd, ea);
 826                 pud  = READ_ONCE(*pudp);
 827
 828                 if (pud_none(pud))
 829                         return NULL;
 830                 else if (pud_huge(pud)) {
 831                         ret_pte = (pte_t *) pudp;
 832                         goto out;
 833                 } else if (is_hugepd(__hugepd(pud_val(pud))))
 834                         hpdp = (hugepd_t *)&pud;
 835                 else {
 836                         pdshift = PMD_SHIFT;
 837                         pmdp = pmd_offset(&pud, ea);
 838                         pmd  = READ_ONCE(*pmdp);
 839                         /*
 840                          * A hugepage collapse is captured by pmd_none, because
 841                          * it mark the pmd none and do a hpte invalidate.
 842                          */
 843                         if (pmd_none(pmd))
 844                                 return NULL;
 845
 846                         if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
 847                                 if (is_thp)
 848                                         *is_thp = true;
 849                                 ret_pte = (pte_t *) pmdp;
 850                                 goto out;
 851                         }
 852
 853                         if (pmd_huge(pmd)) {
 854                                 ret_pte = (pte_t *) pmdp;
 855                                 goto out;
 856                         } else if (is_hugepd(__hugepd(pmd_val(pmd))))
 857                                 hpdp = (hugepd_t *)&pmd;
 858                         else
 859                                 return pte_offset_kernel(&pmd, ea);
 860                 }
 861         }
 862         if (!hpdp)
 863                 return NULL;
 864
 865         ret_pte = hugepte_offset(*hpdp, ea, pdshift);
 866         pdshift = hugepd_shift(*hpdp);
 867 out:
 868         if (hpage_shift)
 869                 *hpage_shift = pdshift;
 870         return ret_pte;
 871 }
 872 EXPORT_SYMBOL_GPL(__find_linux_pte);
 873
 874 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 875                 unsigned long end, int write, struct page **pages, int *nr)
 876 {
 877         unsigned long pte_end;
 878         struct page *head, *page;
 879         pte_t pte;
 880         int refs;
 881
 882         pte_end = (addr + sz) & ~(sz-1);
 883         if (pte_end < end)
 884                 end = pte_end;
 885
 886         pte = READ_ONCE(*ptep);
 887
 888         if (!pte_access_permitted(pte, write))
 889                 return 0;
 890
 891         /* hugepages are never "special" */
 892         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 893
 894         refs = 0;
 895         head = pte_page(pte);
 896
 897         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
 898         do {
 899                 VM_BUG_ON(compound_head(page) != head);
 900                 pages[*nr] = page;
 901                 (*nr)++;
 902                 page++;
 903                 refs++;
 904         } while (addr += PAGE_SIZE, addr != end);
 905
 906         if (!page_cache_add_speculative(head, refs)) {
 907                 *nr -= refs;
 908                 return 0;
 909         }
 910
 911         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 912                 /* Could be optimized better */
 913                 *nr -= refs;
 914                 while (refs--)
 915                         put_page(head);
 916                 return 0;
 917         }
 918
 919         return 1;
 920 }