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
  28 #ifdef CONFIG_HUGETLB_PAGE
  29
  30 #define PAGE_SHIFT_64K  16
  31 #define PAGE_SHIFT_512K 19
  32 #define PAGE_SHIFT_8M   23
  33 #define PAGE_SHIFT_16M  24
  34 #define PAGE_SHIFT_16G  34
  35
  36 unsigned int HPAGE_SHIFT;
  37
  38 /*
  39  * Tracks gpages after the device tree is scanned and before the
  40  * huge_boot_pages list is ready.  On non-Freescale implementations, this is
  41  * just used to track 16G pages and so is a single array.  FSL-based
  42  * implementations may have more than one gpage size, so we need multiple
  43  * arrays
  44  */
  45 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
  46 #define MAX_NUMBER_GPAGES       128
  47 struct psize_gpages {
  48         u64 gpage_list[MAX_NUMBER_GPAGES];
  49         unsigned int nr_gpages;
  50 };
  51 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
  52 #else
  53 #define MAX_NUMBER_GPAGES       1024
  54 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
  55 static unsigned nr_gpages;
  56 #endif
  57
  58 #define hugepd_none(hpd)        (hpd_val(hpd) == 0)
  59
  60 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  61 {
  62         /* Only called for hugetlbfs pages, hence can ignore THP */
  63         return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
  64 }
  65
  66 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  67                            unsigned long address, unsigned pdshift, unsigned pshift)
  68 {
  69         struct kmem_cache *cachep;
  70         pte_t *new;
  71         int i;
  72         int num_hugepd;
  73
  74         if (pshift >= pdshift) {
  75                 cachep = hugepte_cache;
  76                 num_hugepd = 1 << (pshift - pdshift);
  77         } else {
  78                 cachep = PGT_CACHE(pdshift - pshift);
  79                 num_hugepd = 1;
  80         }
  81
  82         new = kmem_cache_zalloc(cachep, GFP_KERNEL);
  83
  84         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  85         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  86
  87         if (! new)
  88                 return -ENOMEM;
  89
  90         /*
  91          * Make sure other cpus find the hugepd set only after a
  92          * properly initialized page table is visible to them.
  93          * For more details look for comment in __pte_alloc().
  94          */
  95         smp_wmb();
  96
  97         spin_lock(&mm->page_table_lock);
  98
  99         /*
 100          * We have multiple higher-level entries that point to the same
 101          * actual pte location.  Fill in each as we go and backtrack on error.
 102          * We need all of these so the DTLB pgtable walk code can find the
 103          * right higher-level entry without knowing if it's a hugepage or not.
 104          */
 105         for (i = 0; i < num_hugepd; i++, hpdp++) {
 106                 if (unlikely(!hugepd_none(*hpdp)))
 107                         break;
 108                 else {
 109 #ifdef CONFIG_PPC_BOOK3S_64
 110                         *hpdp = __hugepd(__pa(new) |
 111                                          (shift_to_mmu_psize(pshift) << 2));
 112 #elif defined(CONFIG_PPC_8xx)
 113                         *hpdp = __hugepd(__pa(new) |
 114                                          (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
 115                                           _PMD_PAGE_512K) | _PMD_PRESENT);
 116 #else
 117                         /* We use the old format for PPC_FSL_BOOK3E */
 118                         *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
 119 #endif
 120                 }
 121         }
 122         /* If we bailed from the for loop early, an error occurred, clean up */
 123         if (i < num_hugepd) {
 124                 for (i = i - 1 ; i >= 0; i--, hpdp--)
 125                         *hpdp = __hugepd(0);
 126                 kmem_cache_free(cachep, new);
 127         }
 128         spin_unlock(&mm->page_table_lock);
 129         return 0;
 130 }
 131
 132 /*
 133  * These macros define how to determine which level of the page table holds
 134  * the hpdp.
 135  */
 136 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 137 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
 138 #define HUGEPD_PUD_SHIFT PUD_SHIFT
 139 #else
 140 #define HUGEPD_PGD_SHIFT PUD_SHIFT
 141 #define HUGEPD_PUD_SHIFT PMD_SHIFT
 142 #endif
 143
 144 /*
 145  * At this point we do the placement change only for BOOK3S 64. This would
 146  * possibly work on other subarchs.
 147  */
 148 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 149 {
 150         pgd_t *pg;
 151         pud_t *pu;
 152         pmd_t *pm;
 153         hugepd_t *hpdp = NULL;
 154         unsigned pshift = __ffs(sz);
 155         unsigned pdshift = PGDIR_SHIFT;
 156
 157         addr &= ~(sz-1);
 158         pg = pgd_offset(mm, addr);
 159
 160 #ifdef CONFIG_PPC_BOOK3S_64
 161         if (pshift == PGDIR_SHIFT)
 162                 /* 16GB huge page */
 163                 return (pte_t *) pg;
 164         else if (pshift > PUD_SHIFT)
 165                 /*
 166                  * We need to use hugepd table
 167                  */
 168                 hpdp = (hugepd_t *)pg;
 169         else {
 170                 pdshift = PUD_SHIFT;
 171                 pu = pud_alloc(mm, pg, addr);
 172                 if (pshift == PUD_SHIFT)
 173                         return (pte_t *)pu;
 174                 else if (pshift > PMD_SHIFT)
 175                         hpdp = (hugepd_t *)pu;
 176                 else {
 177                         pdshift = PMD_SHIFT;
 178                         pm = pmd_alloc(mm, pu, addr);
 179                         if (pshift == PMD_SHIFT)
 180                                 /* 16MB hugepage */
 181                                 return (pte_t *)pm;
 182                         else
 183                                 hpdp = (hugepd_t *)pm;
 184                 }
 185         }
 186 #else
 187         if (pshift >= HUGEPD_PGD_SHIFT) {
 188                 hpdp = (hugepd_t *)pg;
 189         } else {
 190                 pdshift = PUD_SHIFT;
 191                 pu = pud_alloc(mm, pg, addr);
 192                 if (pshift >= HUGEPD_PUD_SHIFT) {
 193                         hpdp = (hugepd_t *)pu;
 194                 } else {
 195                         pdshift = PMD_SHIFT;
 196                         pm = pmd_alloc(mm, pu, addr);
 197                         hpdp = (hugepd_t *)pm;
 198                 }
 199         }
 200 #endif
 201         if (!hpdp)
 202                 return NULL;
 203
 204         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 205
 206         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
 207                 return NULL;
 208
 209         return hugepte_offset(*hpdp, addr, pdshift);
 210 }
 211
 212 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 213 /* Build list of addresses of gigantic pages.  This function is used in early
 214  * boot before the buddy allocator is setup.
 215  */
 216 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 217 {
 218         unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
 219         int i;
 220
 221         if (addr == 0)
 222                 return;
 223
 224         gpage_freearray[idx].nr_gpages = number_of_pages;
 225
 226         for (i = 0; i < number_of_pages; i++) {
 227                 gpage_freearray[idx].gpage_list[i] = addr;
 228                 addr += page_size;
 229         }
 230 }
 231
 232 /*
 233  * Moves the gigantic page addresses from the temporary list to the
 234  * huge_boot_pages list.
 235  */
 236 int alloc_bootmem_huge_page(struct hstate *hstate)
 237 {
 238         struct huge_bootmem_page *m;
 239         int idx = shift_to_mmu_psize(huge_page_shift(hstate));
 240         int nr_gpages = gpage_freearray[idx].nr_gpages;
 241
 242         if (nr_gpages == 0)
 243                 return 0;
 244
 245 #ifdef CONFIG_HIGHMEM
 246         /*
 247          * If gpages can be in highmem we can't use the trick of storing the
 248          * data structure in the page; allocate space for this
 249          */
 250         m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
 251         m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
 252 #else
 253         m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
 254 #endif
 255
 256         list_add(&m->list, &huge_boot_pages);
 257         gpage_freearray[idx].nr_gpages = nr_gpages;
 258         gpage_freearray[idx].gpage_list[nr_gpages] = 0;
 259         m->hstate = hstate;
 260
 261         return 1;
 262 }
 263 /*
 264  * Scan the command line hugepagesz= options for gigantic pages; store those in
 265  * a list that we use to allocate the memory once all options are parsed.
 266  */
 267
 268 unsigned long gpage_npages[MMU_PAGE_COUNT];
 269
 270 static int __init do_gpage_early_setup(char *param, char *val,
 271                                        const char *unused, void *arg)
 272 {
 273         static phys_addr_t size;
 274         unsigned long npages;
 275
 276         /*
 277          * The hugepagesz and hugepages cmdline options are interleaved.  We
 278          * use the size variable to keep track of whether or not this was done
 279          * properly and skip over instances where it is incorrect.  Other
 280          * command-line parsing code will issue warnings, so we don't need to.
 281          *
 282          */
 283         if ((strcmp(param, "default_hugepagesz") == 0) ||
 284             (strcmp(param, "hugepagesz") == 0)) {
 285                 size = memparse(val, NULL);
 286         } else if (strcmp(param, "hugepages") == 0) {
 287                 if (size != 0) {
 288                         if (sscanf(val, "%lu", &npages) <= 0)
 289                                 npages = 0;
 290                         if (npages > MAX_NUMBER_GPAGES) {
 291                                 pr_warn("MMU: %lu pages requested for page "
 292 #ifdef CONFIG_PHYS_ADDR_T_64BIT
 293                                         "size %llu KB, limiting to "
 294 #else
 295                                         "size %u KB, limiting to "
 296 #endif
 297                                         __stringify(MAX_NUMBER_GPAGES) "\n",
 298                                         npages, size / 1024);
 299                                 npages = MAX_NUMBER_GPAGES;
 300                         }
 301                         gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
 302                         size = 0;
 303                 }
 304         }
 305         return 0;
 306 }
 307
 308
 309 /*
 310  * This function allocates physical space for pages that are larger than the
 311  * buddy allocator can handle.  We want to allocate these in highmem because
 312  * the amount of lowmem is limited.  This means that this function MUST be
 313  * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
 314  * allocate to grab highmem.
 315  */
 316 void __init reserve_hugetlb_gpages(void)
 317 {
 318         static __initdata char cmdline[COMMAND_LINE_SIZE];
 319         phys_addr_t size, base;
 320         int i;
 321
 322         strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
 323         parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
 324                         NULL, &do_gpage_early_setup);
 325
 326         /*
 327          * Walk gpage list in reverse, allocating larger page sizes first.
 328          * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
 329          * When we reach the point in the list where pages are no longer
 330          * considered gpages, we're done.
 331          */
 332         for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
 333                 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
 334                         continue;
 335                 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
 336                         break;
 337
 338                 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
 339                 base = memblock_alloc_base(size * gpage_npages[i], size,
 340                                            MEMBLOCK_ALLOC_ANYWHERE);
 341                 add_gpage(base, size, gpage_npages[i]);
 342         }
 343 }
 344
 345 #else /* !PPC_FSL_BOOK3E */
 346
 347 /* Build list of addresses of gigantic pages.  This function is used in early
 348  * boot before the buddy allocator is setup.
 349  */
 350 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 351 {
 352         if (!addr)
 353                 return;
 354         while (number_of_pages > 0) {
 355                 gpage_freearray[nr_gpages] = addr;
 356                 nr_gpages++;
 357                 number_of_pages--;
 358                 addr += page_size;
 359         }
 360 }
 361
 362 /* Moves the gigantic page addresses from the temporary list to the
 363  * huge_boot_pages list.
 364  */
 365 int alloc_bootmem_huge_page(struct hstate *hstate)
 366 {
 367         struct huge_bootmem_page *m;
 368         if (nr_gpages == 0)
 369                 return 0;
 370         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 371         gpage_freearray[nr_gpages] = 0;
 372         list_add(&m->list, &huge_boot_pages);
 373         m->hstate = hstate;
 374         return 1;
 375 }
 376 #endif
 377
 378 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 379 #define HUGEPD_FREELIST_SIZE \
 380         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 381
 382 struct hugepd_freelist {
 383         struct rcu_head rcu;
 384         unsigned int index;
 385         void *ptes[0];
 386 };
 387
 388 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 389
 390 static void hugepd_free_rcu_callback(struct rcu_head *head)
 391 {
 392         struct hugepd_freelist *batch =
 393                 container_of(head, struct hugepd_freelist, rcu);
 394         unsigned int i;
 395
 396         for (i = 0; i < batch->index; i++)
 397                 kmem_cache_free(hugepte_cache, batch->ptes[i]);
 398
 399         free_page((unsigned long)batch);
 400 }
 401
 402 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 403 {
 404         struct hugepd_freelist **batchp;
 405
 406         batchp = &get_cpu_var(hugepd_freelist_cur);
 407
 408         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 409             cpumask_equal(mm_cpumask(tlb->mm),
 410                           cpumask_of(smp_processor_id()))) {
 411                 kmem_cache_free(hugepte_cache, hugepte);
 412                 put_cpu_var(hugepd_freelist_cur);
 413                 return;
 414         }
 415
 416         if (*batchp == NULL) {
 417                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 418                 (*batchp)->index = 0;
 419         }
 420
 421         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 422         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 423                 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
 424                 *batchp = NULL;
 425         }
 426         put_cpu_var(hugepd_freelist_cur);
 427 }
 428 #else
 429 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 430 #endif
 431
 432 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 433                               unsigned long start, unsigned long end,
 434                               unsigned long floor, unsigned long ceiling)
 435 {
 436         pte_t *hugepte = hugepd_page(*hpdp);
 437         int i;
 438
 439         unsigned long pdmask = ~((1UL << pdshift) - 1);
 440         unsigned int num_hugepd = 1;
 441         unsigned int shift = hugepd_shift(*hpdp);
 442
 443         /* Note: On fsl the hpdp may be the first of several */
 444         if (shift > pdshift)
 445                 num_hugepd = 1 << (shift - pdshift);
 446
 447         start &= pdmask;
 448         if (start < floor)
 449                 return;
 450         if (ceiling) {
 451                 ceiling &= pdmask;
 452                 if (! ceiling)
 453                         return;
 454         }
 455         if (end - 1 > ceiling - 1)
 456                 return;
 457
 458         for (i = 0; i < num_hugepd; i++, hpdp++)
 459                 *hpdp = __hugepd(0);
 460
 461         if (shift >= pdshift)
 462                 hugepd_free(tlb, hugepte);
 463         else
 464                 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
 465 }
 466
 467 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 468                                    unsigned long addr, unsigned long end,
 469                                    unsigned long floor, unsigned long ceiling)
 470 {
 471         pmd_t *pmd;
 472         unsigned long next;
 473         unsigned long start;
 474
 475         start = addr;
 476         do {
 477                 unsigned long more;
 478
 479                 pmd = pmd_offset(pud, addr);
 480                 next = pmd_addr_end(addr, end);
 481                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 482                         /*
 483                          * if it is not hugepd pointer, we should already find
 484                          * it cleared.
 485                          */
 486                         WARN_ON(!pmd_none_or_clear_bad(pmd));
 487                         continue;
 488                 }
 489                 /*
 490                  * Increment next by the size of the huge mapping since
 491                  * there may be more than one entry at this level for a
 492                  * single hugepage, but all of them point to
 493                  * the same kmem cache that holds the hugepte.
 494                  */
 495                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 496                 if (more > next)
 497                         next = more;
 498
 499                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 500                                   addr, next, floor, ceiling);
 501         } while (addr = next, addr != end);
 502
 503         start &= PUD_MASK;
 504         if (start < floor)
 505                 return;
 506         if (ceiling) {
 507                 ceiling &= PUD_MASK;
 508                 if (!ceiling)
 509                         return;
 510         }
 511         if (end - 1 > ceiling - 1)
 512                 return;
 513
 514         pmd = pmd_offset(pud, start);
 515         pud_clear(pud);
 516         pmd_free_tlb(tlb, pmd, start);
 517         mm_dec_nr_pmds(tlb->mm);
 518 }
 519
 520 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 521                                    unsigned long addr, unsigned long end,
 522                                    unsigned long floor, unsigned long ceiling)
 523 {
 524         pud_t *pud;
 525         unsigned long next;
 526         unsigned long start;
 527
 528         start = addr;
 529         do {
 530                 pud = pud_offset(pgd, addr);
 531                 next = pud_addr_end(addr, end);
 532                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 533                         if (pud_none_or_clear_bad(pud))
 534                                 continue;
 535                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 536                                                ceiling);
 537                 } else {
 538                         unsigned long more;
 539                         /*
 540                          * Increment next by the size of the huge mapping since
 541                          * there may be more than one entry at this level for a
 542                          * single hugepage, but all of them point to
 543                          * the same kmem cache that holds the hugepte.
 544                          */
 545                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 546                         if (more > next)
 547                                 next = more;
 548
 549                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 550                                           addr, next, floor, ceiling);
 551                 }
 552         } while (addr = next, addr != end);
 553
 554         start &= PGDIR_MASK;
 555         if (start < floor)
 556                 return;
 557         if (ceiling) {
 558                 ceiling &= PGDIR_MASK;
 559                 if (!ceiling)
 560                         return;
 561         }
 562         if (end - 1 > ceiling - 1)
 563                 return;
 564
 565         pud = pud_offset(pgd, start);
 566         pgd_clear(pgd);
 567         pud_free_tlb(tlb, pud, start);
 568 }
 569
 570 /*
 571  * This function frees user-level page tables of a process.
 572  */
 573 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 574                             unsigned long addr, unsigned long end,
 575                             unsigned long floor, unsigned long ceiling)
 576 {
 577         pgd_t *pgd;
 578         unsigned long next;
 579
 580         /*
 581          * Because there are a number of different possible pagetable
 582          * layouts for hugepage ranges, we limit knowledge of how
 583          * things should be laid out to the allocation path
 584          * (huge_pte_alloc(), above).  Everything else works out the
 585          * structure as it goes from information in the hugepd
 586          * pointers.  That means that we can't here use the
 587          * optimization used in the normal page free_pgd_range(), of
 588          * checking whether we're actually covering a large enough
 589          * range to have to do anything at the top level of the walk
 590          * instead of at the bottom.
 591          *
 592          * To make sense of this, you should probably go read the big
 593          * block comment at the top of the normal free_pgd_range(),
 594          * too.
 595          */
 596
 597         do {
 598                 next = pgd_addr_end(addr, end);
 599                 pgd = pgd_offset(tlb->mm, addr);
 600                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 601                         if (pgd_none_or_clear_bad(pgd))
 602                                 continue;
 603                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 604                 } else {
 605                         unsigned long more;
 606                         /*
 607                          * Increment next by the size of the huge mapping since
 608                          * there may be more than one entry at the pgd level
 609                          * for a single hugepage, but all of them point to the
 610                          * same kmem cache that holds the hugepte.
 611                          */
 612                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 613                         if (more > next)
 614                                 next = more;
 615
 616                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 617                                           addr, next, floor, ceiling);
 618                 }
 619         } while (addr = next, addr != end);
 620 }
 621
 622 struct page *follow_huge_pd(struct vm_area_struct *vma,
 623                             unsigned long address, hugepd_t hpd,
 624                             int flags, int pdshift)
 625 {
 626         pte_t *ptep;
 627         spinlock_t *ptl;
 628         struct page *page = NULL;
 629         unsigned long mask;
 630         int shift = hugepd_shift(hpd);
 631         struct mm_struct *mm = vma->vm_mm;
 632
 633 retry:
 634         ptl = &mm->page_table_lock;
 635         spin_lock(ptl);
 636
 637         ptep = hugepte_offset(hpd, address, pdshift);
 638         if (pte_present(*ptep)) {
 639                 mask = (1UL << shift) - 1;
 640                 page = pte_page(*ptep);
 641                 page += ((address & mask) >> PAGE_SHIFT);
 642                 if (flags & FOLL_GET)
 643                         get_page(page);
 644         } else {
 645                 if (is_hugetlb_entry_migration(*ptep)) {
 646                         spin_unlock(ptl);
 647                         __migration_entry_wait(mm, ptep, ptl);
 648                         goto retry;
 649                 }
 650         }
 651         spin_unlock(ptl);
 652         return page;
 653 }
 654
 655 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 656                                       unsigned long sz)
 657 {
 658         unsigned long __boundary = (addr + sz) & ~(sz-1);
 659         return (__boundary - 1 < end - 1) ? __boundary : end;
 660 }
 661
 662 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
 663                 unsigned long end, int write, struct page **pages, int *nr)
 664 {
 665         pte_t *ptep;
 666         unsigned long sz = 1UL << hugepd_shift(hugepd);
 667         unsigned long next;
 668
 669         ptep = hugepte_offset(hugepd, addr, pdshift);
 670         do {
 671                 next = hugepte_addr_end(addr, end, sz);
 672                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 673                         return 0;
 674         } while (ptep++, addr = next, addr != end);
 675
 676         return 1;
 677 }
 678
 679 #ifdef CONFIG_PPC_MM_SLICES
 680 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 681                                         unsigned long len, unsigned long pgoff,
 682                                         unsigned long flags)
 683 {
 684         struct hstate *hstate = hstate_file(file);
 685         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 686
 687         if (radix_enabled())
 688                 return radix__hugetlb_get_unmapped_area(file, addr, len,
 689                                                        pgoff, flags);
 690         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 691 }
 692 #endif
 693
 694 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 695 {
 696 #ifdef CONFIG_PPC_MM_SLICES
 697         unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 698         /* With radix we don't use slice, so derive it from vma*/
 699         if (!radix_enabled())
 700                 return 1UL << mmu_psize_to_shift(psize);
 701 #endif
 702         if (!is_vm_hugetlb_page(vma))
 703                 return PAGE_SIZE;
 704
 705         return huge_page_size(hstate_vma(vma));
 706 }
 707
 708 static inline bool is_power_of_4(unsigned long x)
 709 {
 710         if (is_power_of_2(x))
 711                 return (__ilog2(x) % 2) ? false : true;
 712         return false;
 713 }
 714
 715 static int __init add_huge_page_size(unsigned long long size)
 716 {
 717         int shift = __ffs(size);
 718         int mmu_psize;
 719
 720         /* Check that it is a page size supported by the hardware and
 721          * that it fits within pagetable and slice limits. */
 722         if (size <= PAGE_SIZE)
 723                 return -EINVAL;
 724 #if defined(CONFIG_PPC_FSL_BOOK3E)
 725         if (!is_power_of_4(size))
 726                 return -EINVAL;
 727 #elif !defined(CONFIG_PPC_8xx)
 728         if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
 729                 return -EINVAL;
 730 #endif
 731
 732         if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
 733                 return -EINVAL;
 734
 735 #ifdef CONFIG_PPC_BOOK3S_64
 736         /*
 737          * We need to make sure that for different page sizes reported by
 738          * firmware we only add hugetlb support for page sizes that can be
 739          * supported by linux page table layout.
 740          * For now we have
 741          * Radix: 2M
 742          * Hash: 16M and 16G
 743          */
 744         if (radix_enabled()) {
 745                 if (mmu_psize != MMU_PAGE_2M)
 746                         return -EINVAL;
 747         } else {
 748                 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
 749                         return -EINVAL;
 750         }
 751 #endif
 752
 753         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 754
 755         /* Return if huge page size has already been setup */
 756         if (size_to_hstate(size))
 757                 return 0;
 758
 759         hugetlb_add_hstate(shift - PAGE_SHIFT);
 760
 761         return 0;
 762 }
 763
 764 static int __init hugepage_setup_sz(char *str)
 765 {
 766         unsigned long long size;
 767
 768         size = memparse(str, &str);
 769
 770         if (add_huge_page_size(size) != 0) {
 771                 hugetlb_bad_size();
 772                 pr_err("Invalid huge page size specified(%llu)\n", size);
 773         }
 774
 775         return 1;
 776 }
 777 __setup("hugepagesz=", hugepage_setup_sz);
 778
 779 struct kmem_cache *hugepte_cache;
 780 static int __init hugetlbpage_init(void)
 781 {
 782         int psize;
 783
 784 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
 785         if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
 786                 return -ENODEV;
 787 #endif
 788         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 789                 unsigned shift;
 790                 unsigned pdshift;
 791
 792                 if (!mmu_psize_defs[psize].shift)
 793                         continue;
 794
 795                 shift = mmu_psize_to_shift(psize);
 796
 797                 if (add_huge_page_size(1ULL << shift) < 0)
 798                         continue;
 799
 800                 if (shift < HUGEPD_PUD_SHIFT)
 801                         pdshift = PMD_SHIFT;
 802                 else if (shift < HUGEPD_PGD_SHIFT)
 803                         pdshift = PUD_SHIFT;
 804                 else
 805                         pdshift = PGDIR_SHIFT;
 806                 /*
 807                  * if we have pdshift and shift value same, we don't
 808                  * use pgt cache for hugepd.
 809                  */
 810                 if (pdshift > shift)
 811                         pgtable_cache_add(pdshift - shift, NULL);
 812 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 813                 else if (!hugepte_cache) {
 814                         /*
 815                          * Create a kmem cache for hugeptes.  The bottom bits in
 816                          * the pte have size information encoded in them, so
 817                          * align them to allow this
 818                          */
 819                         hugepte_cache = kmem_cache_create("hugepte-cache",
 820                                                           sizeof(pte_t),
 821                                                           HUGEPD_SHIFT_MASK + 1,
 822                                                           0, NULL);
 823                         if (hugepte_cache == NULL)
 824                                 panic("%s: Unable to create kmem cache "
 825                                       "for hugeptes\n", __func__);
 826
 827                 }
 828 #endif
 829         }
 830
 831 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 832         /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
 833         if (mmu_psize_defs[MMU_PAGE_4M].shift)
 834                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
 835         else if (mmu_psize_defs[MMU_PAGE_512K].shift)
 836                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
 837 #else
 838         /* Set default large page size. Currently, we pick 16M or 1M
 839          * depending on what is available
 840          */
 841         if (mmu_psize_defs[MMU_PAGE_16M].shift)
 842                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
 843         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 844                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
 845         else if (mmu_psize_defs[MMU_PAGE_2M].shift)
 846                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
 847 #endif
 848         return 0;
 849 }
 850
 851 arch_initcall(hugetlbpage_init);
 852
 853 void flush_dcache_icache_hugepage(struct page *page)
 854 {
 855         int i;
 856         void *start;
 857
 858         BUG_ON(!PageCompound(page));
 859
 860         for (i = 0; i < (1UL << compound_order(page)); i++) {
 861                 if (!PageHighMem(page)) {
 862                         __flush_dcache_icache(page_address(page+i));
 863                 } else {
 864                         start = kmap_atomic(page+i);
 865                         __flush_dcache_icache(start);
 866                         kunmap_atomic(start);
 867                 }
 868         }
 869 }
 870
 871 #endif /* CONFIG_HUGETLB_PAGE */
 872
 873 /*
 874  * We have 4 cases for pgds and pmds:
 875  * (1) invalid (all zeroes)
 876  * (2) pointer to next table, as normal; bottom 6 bits == 0
 877  * (3) leaf pte for huge page _PAGE_PTE set
 878  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
 879  *
 880  * So long as we atomically load page table pointers we are safe against teardown,
 881  * we can follow the address down to the the page and take a ref on it.
 882  * This function need to be called with interrupts disabled. We use this variant
 883  * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
 884  */
 885
 886 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
 887                                    bool *is_thp, unsigned *shift)
 888 {
 889         pgd_t pgd, *pgdp;
 890         pud_t pud, *pudp;
 891         pmd_t pmd, *pmdp;
 892         pte_t *ret_pte;
 893         hugepd_t *hpdp = NULL;
 894         unsigned pdshift = PGDIR_SHIFT;
 895
 896         if (shift)
 897                 *shift = 0;
 898
 899         if (is_thp)
 900                 *is_thp = false;
 901
 902         pgdp = pgdir + pgd_index(ea);
 903         pgd  = READ_ONCE(*pgdp);
 904         /*
 905          * Always operate on the local stack value. This make sure the
 906          * value don't get updated by a parallel THP split/collapse,
 907          * page fault or a page unmap. The return pte_t * is still not
 908          * stable. So should be checked there for above conditions.
 909          */
 910         if (pgd_none(pgd))
 911                 return NULL;
 912         else if (pgd_huge(pgd)) {
 913                 ret_pte = (pte_t *) pgdp;
 914                 goto out;
 915         } else if (is_hugepd(__hugepd(pgd_val(pgd))))
 916                 hpdp = (hugepd_t *)&pgd;
 917         else {
 918                 /*
 919                  * Even if we end up with an unmap, the pgtable will not
 920                  * be freed, because we do an rcu free and here we are
 921                  * irq disabled
 922                  */
 923                 pdshift = PUD_SHIFT;
 924                 pudp = pud_offset(&pgd, ea);
 925                 pud  = READ_ONCE(*pudp);
 926
 927                 if (pud_none(pud))
 928                         return NULL;
 929                 else if (pud_huge(pud)) {
 930                         ret_pte = (pte_t *) pudp;
 931                         goto out;
 932                 } else if (is_hugepd(__hugepd(pud_val(pud))))
 933                         hpdp = (hugepd_t *)&pud;
 934                 else {
 935                         pdshift = PMD_SHIFT;
 936                         pmdp = pmd_offset(&pud, ea);
 937                         pmd  = READ_ONCE(*pmdp);
 938                         /*
 939                          * A hugepage collapse is captured by pmd_none, because
 940                          * it mark the pmd none and do a hpte invalidate.
 941                          */
 942                         if (pmd_none(pmd))
 943                                 return NULL;
 944
 945                         if (pmd_trans_huge(pmd)) {
 946                                 if (is_thp)
 947                                         *is_thp = true;
 948                                 ret_pte = (pte_t *) pmdp;
 949                                 goto out;
 950                         }
 951
 952                         if (pmd_huge(pmd)) {
 953                                 ret_pte = (pte_t *) pmdp;
 954                                 goto out;
 955                         } else if (is_hugepd(__hugepd(pmd_val(pmd))))
 956                                 hpdp = (hugepd_t *)&pmd;
 957                         else
 958                                 return pte_offset_kernel(&pmd, ea);
 959                 }
 960         }
 961         if (!hpdp)
 962                 return NULL;
 963
 964         ret_pte = hugepte_offset(*hpdp, ea, pdshift);
 965         pdshift = hugepd_shift(*hpdp);
 966 out:
 967         if (shift)
 968                 *shift = pdshift;
 969         return ret_pte;
 970 }
 971 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
 972
 973 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 974                 unsigned long end, int write, struct page **pages, int *nr)
 975 {
 976         unsigned long mask;
 977         unsigned long pte_end;
 978         struct page *head, *page;
 979         pte_t pte;
 980         int refs;
 981
 982         pte_end = (addr + sz) & ~(sz-1);
 983         if (pte_end < end)
 984                 end = pte_end;
 985
 986         pte = READ_ONCE(*ptep);
 987         mask = _PAGE_PRESENT | _PAGE_READ;
 988
 989         /*
 990          * On some CPUs like the 8xx, _PAGE_RW hence _PAGE_WRITE is defined
 991          * as 0 and _PAGE_RO has to be set when a page is not writable
 992          */
 993         if (write)
 994                 mask |= _PAGE_WRITE;
 995         else
 996                 mask |= _PAGE_RO;
 997
 998         if ((pte_val(pte) & mask) != mask)
 999                 return 0;
1000
1001         /* hugepages are never "special" */
1002         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1003
1004         refs = 0;
1005         head = pte_page(pte);
1006
1007         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1008         do {
1009                 VM_BUG_ON(compound_head(page) != head);
1010                 pages[*nr] = page;
1011                 (*nr)++;
1012                 page++;
1013                 refs++;
1014         } while (addr += PAGE_SIZE, addr != end);
1015
1016         if (!page_cache_add_speculative(head, refs)) {
1017                 *nr -= refs;
1018                 return 0;
1019         }
1020
1021         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1022                 /* Could be optimized better */
1023                 *nr -= refs;
1024                 while (refs--)
1025                         put_page(head);
1026                 return 0;
1027         }
1028
1029         return 1;
1030 }