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