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