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/moduleparam.h>
  19 #include <linux/swap.h>
  20 #include <linux/swapops.h>
  21 #include <linux/kmemleak.h>
  22 #include <asm/pgalloc.h>
  23 #include <asm/tlb.h>
  24 #include <asm/setup.h>
  25 #include <asm/hugetlb.h>
  26 #include <asm/pte-walk.h>
  27 #include <asm/firmware.h>
  28
  29 bool hugetlb_disabled = false;
  30
  31 #define hugepd_none(hpd)        (hpd_val(hpd) == 0)
  32
  33 #define PTE_T_ORDER     (__builtin_ffs(sizeof(pte_basic_t)) - \
  34                          __builtin_ffs(sizeof(void *)))
  35
  36 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  37 {
  38         /*
  39          * Only called for hugetlbfs pages, hence can ignore THP and the
  40          * irq disabled walk.
  41          */
  42         return __find_linux_pte(mm->pgd, addr, NULL, NULL);
  43 }
  44
  45 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  46                            unsigned long address, unsigned int pdshift,
  47                            unsigned int pshift, spinlock_t *ptl)
  48 {
  49         struct kmem_cache *cachep;
  50         pte_t *new;
  51         int i;
  52         int num_hugepd;
  53
  54         if (pshift >= pdshift) {
  55                 cachep = PGT_CACHE(PTE_T_ORDER);
  56                 num_hugepd = 1 << (pshift - pdshift);
  57         } else {
  58                 cachep = PGT_CACHE(pdshift - pshift);
  59                 num_hugepd = 1;
  60         }
  61
  62         if (!cachep) {
  63                 WARN_ONCE(1, "No page table cache created for hugetlb tables");
  64                 return -ENOMEM;
  65         }
  66
  67         new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
  68
  69         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  70         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  71
  72         if (!new)
  73                 return -ENOMEM;
  74
  75         /*
  76          * Make sure other cpus find the hugepd set only after a
  77          * properly initialized page table is visible to them.
  78          * For more details look for comment in __pte_alloc().
  79          */
  80         smp_wmb();
  81
  82         spin_lock(ptl);
  83         /*
  84          * We have multiple higher-level entries that point to the same
  85          * actual pte location.  Fill in each as we go and backtrack on error.
  86          * We need all of these so the DTLB pgtable walk code can find the
  87          * right higher-level entry without knowing if it's a hugepage or not.
  88          */
  89         for (i = 0; i < num_hugepd; i++, hpdp++) {
  90                 if (unlikely(!hugepd_none(*hpdp)))
  91                         break;
  92                 hugepd_populate(hpdp, new, pshift);
  93         }
  94         /* If we bailed from the for loop early, an error occurred, clean up */
  95         if (i < num_hugepd) {
  96                 for (i = i - 1 ; i >= 0; i--, hpdp--)
  97                         *hpdp = __hugepd(0);
  98                 kmem_cache_free(cachep, new);
  99         } else {
 100                 kmemleak_ignore(new);
 101         }
 102         spin_unlock(ptl);
 103         return 0;
 104 }
 105
 106 /*
 107  * At this point we do the placement change only for BOOK3S 64. This would
 108  * possibly work on other subarchs.
 109  */
 110 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 111                       unsigned long addr, unsigned long sz)
 112 {
 113         pgd_t *pg;
 114         p4d_t *p4;
 115         pud_t *pu;
 116         pmd_t *pm;
 117         hugepd_t *hpdp = NULL;
 118         unsigned pshift = __ffs(sz);
 119         unsigned pdshift = PGDIR_SHIFT;
 120         spinlock_t *ptl;
 121
 122         addr &= ~(sz-1);
 123         pg = pgd_offset(mm, addr);
 124         p4 = p4d_offset(pg, addr);
 125
 126 #ifdef CONFIG_PPC_BOOK3S_64
 127         if (pshift == PGDIR_SHIFT)
 128                 /* 16GB huge page */
 129                 return (pte_t *) p4;
 130         else if (pshift > PUD_SHIFT) {
 131                 /*
 132                  * We need to use hugepd table
 133                  */
 134                 ptl = &mm->page_table_lock;
 135                 hpdp = (hugepd_t *)p4;
 136         } else {
 137                 pdshift = PUD_SHIFT;
 138                 pu = pud_alloc(mm, p4, addr);
 139                 if (!pu)
 140                         return NULL;
 141                 if (pshift == PUD_SHIFT)
 142                         return (pte_t *)pu;
 143                 else if (pshift > PMD_SHIFT) {
 144                         ptl = pud_lockptr(mm, pu);
 145                         hpdp = (hugepd_t *)pu;
 146                 } else {
 147                         pdshift = PMD_SHIFT;
 148                         pm = pmd_alloc(mm, pu, addr);
 149                         if (!pm)
 150                                 return NULL;
 151                         if (pshift == PMD_SHIFT)
 152                                 /* 16MB hugepage */
 153                                 return (pte_t *)pm;
 154                         else {
 155                                 ptl = pmd_lockptr(mm, pm);
 156                                 hpdp = (hugepd_t *)pm;
 157                         }
 158                 }
 159         }
 160 #else
 161         if (pshift >= PGDIR_SHIFT) {
 162                 ptl = &mm->page_table_lock;
 163                 hpdp = (hugepd_t *)p4;
 164         } else {
 165                 pdshift = PUD_SHIFT;
 166                 pu = pud_alloc(mm, p4, addr);
 167                 if (!pu)
 168                         return NULL;
 169                 if (pshift >= PUD_SHIFT) {
 170                         ptl = pud_lockptr(mm, pu);
 171                         hpdp = (hugepd_t *)pu;
 172                 } else {
 173                         pdshift = PMD_SHIFT;
 174                         pm = pmd_alloc(mm, pu, addr);
 175                         if (!pm)
 176                                 return NULL;
 177                         ptl = pmd_lockptr(mm, pm);
 178                         hpdp = (hugepd_t *)pm;
 179                 }
 180         }
 181 #endif
 182         if (!hpdp)
 183                 return NULL;
 184
 185         if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
 186                 return pte_alloc_huge(mm, (pmd_t *)hpdp, addr);
 187
 188         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 189
 190         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
 191                                                   pdshift, pshift, ptl))
 192                 return NULL;
 193
 194         return hugepte_offset(*hpdp, addr, pdshift);
 195 }
 196
 197 #ifdef CONFIG_PPC_BOOK3S_64
 198 /*
 199  * Tracks gpages after the device tree is scanned and before the
 200  * huge_boot_pages list is ready on pseries.
 201  */
 202 #define MAX_NUMBER_GPAGES       1024
 203 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 204 __initdata static unsigned nr_gpages;
 205
 206 /*
 207  * Build list of addresses of gigantic pages.  This function is used in early
 208  * boot before the buddy allocator is setup.
 209  */
 210 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 211 {
 212         if (!addr)
 213                 return;
 214         while (number_of_pages > 0) {
 215                 gpage_freearray[nr_gpages] = addr;
 216                 nr_gpages++;
 217                 number_of_pages--;
 218                 addr += page_size;
 219         }
 220 }
 221
 222 static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 223 {
 224         struct huge_bootmem_page *m;
 225         if (nr_gpages == 0)
 226                 return 0;
 227         m = phys_to_virt(gpage_freearray[--nr_gpages]);
 228         gpage_freearray[nr_gpages] = 0;
 229         list_add(&m->list, &huge_boot_pages);
 230         m->hstate = hstate;
 231         return 1;
 232 }
 233
 234 bool __init hugetlb_node_alloc_supported(void)
 235 {
 236         return false;
 237 }
 238 #endif
 239
 240
 241 int __init alloc_bootmem_huge_page(struct hstate *h, int nid)
 242 {
 243
 244 #ifdef CONFIG_PPC_BOOK3S_64
 245         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
 246                 return pseries_alloc_bootmem_huge_page(h);
 247 #endif
 248         return __alloc_bootmem_huge_page(h, nid);
 249 }
 250
 251 #ifndef CONFIG_PPC_BOOK3S_64
 252 #define HUGEPD_FREELIST_SIZE \
 253         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
 254
 255 struct hugepd_freelist {
 256         struct rcu_head rcu;
 257         unsigned int index;
 258         void *ptes[];
 259 };
 260
 261 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
 262
 263 static void hugepd_free_rcu_callback(struct rcu_head *head)
 264 {
 265         struct hugepd_freelist *batch =
 266                 container_of(head, struct hugepd_freelist, rcu);
 267         unsigned int i;
 268
 269         for (i = 0; i < batch->index; i++)
 270                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
 271
 272         free_page((unsigned long)batch);
 273 }
 274
 275 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 276 {
 277         struct hugepd_freelist **batchp;
 278
 279         batchp = &get_cpu_var(hugepd_freelist_cur);
 280
 281         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 282             mm_is_thread_local(tlb->mm)) {
 283                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
 284                 put_cpu_var(hugepd_freelist_cur);
 285                 return;
 286         }
 287
 288         if (*batchp == NULL) {
 289                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 290                 (*batchp)->index = 0;
 291         }
 292
 293         (*batchp)->ptes[(*batchp)->index++] = hugepte;
 294         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 295                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
 296                 *batchp = NULL;
 297         }
 298         put_cpu_var(hugepd_freelist_cur);
 299 }
 300 #else
 301 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 302 #endif
 303
 304 /* Return true when the entry to be freed maps more than the area being freed */
 305 static bool range_is_outside_limits(unsigned long start, unsigned long end,
 306                                     unsigned long floor, unsigned long ceiling,
 307                                     unsigned long mask)
 308 {
 309         if ((start & mask) < floor)
 310                 return true;
 311         if (ceiling) {
 312                 ceiling &= mask;
 313                 if (!ceiling)
 314                         return true;
 315         }
 316         return end - 1 > ceiling - 1;
 317 }
 318
 319 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 320                               unsigned long start, unsigned long end,
 321                               unsigned long floor, unsigned long ceiling)
 322 {
 323         pte_t *hugepte = hugepd_page(*hpdp);
 324         int i;
 325
 326         unsigned long pdmask = ~((1UL << pdshift) - 1);
 327         unsigned int num_hugepd = 1;
 328         unsigned int shift = hugepd_shift(*hpdp);
 329
 330         /* Note: On fsl the hpdp may be the first of several */
 331         if (shift > pdshift)
 332                 num_hugepd = 1 << (shift - pdshift);
 333
 334         if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
 335                 return;
 336
 337         for (i = 0; i < num_hugepd; i++, hpdp++)
 338                 *hpdp = __hugepd(0);
 339
 340         if (shift >= pdshift)
 341                 hugepd_free(tlb, hugepte);
 342         else
 343                 pgtable_free_tlb(tlb, hugepte,
 344                                  get_hugepd_cache_index(pdshift - shift));
 345 }
 346
 347 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 348                                    unsigned long addr, unsigned long end,
 349                                    unsigned long floor, unsigned long ceiling)
 350 {
 351         pgtable_t token = pmd_pgtable(*pmd);
 352
 353         if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
 354                 return;
 355
 356         pmd_clear(pmd);
 357         pte_free_tlb(tlb, token, addr);
 358         mm_dec_nr_ptes(tlb->mm);
 359 }
 360
 361 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 362                                    unsigned long addr, unsigned long end,
 363                                    unsigned long floor, unsigned long ceiling)
 364 {
 365         pmd_t *pmd;
 366         unsigned long next;
 367         unsigned long start;
 368
 369         start = addr;
 370         do {
 371                 unsigned long more;
 372
 373                 pmd = pmd_offset(pud, addr);
 374                 next = pmd_addr_end(addr, end);
 375                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 376                         if (pmd_none_or_clear_bad(pmd))
 377                                 continue;
 378
 379                         /*
 380                          * if it is not hugepd pointer, we should already find
 381                          * it cleared.
 382                          */
 383                         WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
 384
 385                         hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
 386
 387                         continue;
 388                 }
 389                 /*
 390                  * Increment next by the size of the huge mapping since
 391                  * there may be more than one entry at this level for a
 392                  * single hugepage, but all of them point to
 393                  * the same kmem cache that holds the hugepte.
 394                  */
 395                 more = addr + (1UL << hugepd_shift(*(hugepd_t *)pmd));
 396                 if (more > next)
 397                         next = more;
 398
 399                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 400                                   addr, next, floor, ceiling);
 401         } while (addr = next, addr != end);
 402
 403         if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
 404                 return;
 405
 406         pmd = pmd_offset(pud, start & PUD_MASK);
 407         pud_clear(pud);
 408         pmd_free_tlb(tlb, pmd, start & PUD_MASK);
 409         mm_dec_nr_pmds(tlb->mm);
 410 }
 411
 412 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 413                                    unsigned long addr, unsigned long end,
 414                                    unsigned long floor, unsigned long ceiling)
 415 {
 416         pud_t *pud;
 417         unsigned long next;
 418         unsigned long start;
 419
 420         start = addr;
 421         do {
 422                 pud = pud_offset(p4d, addr);
 423                 next = pud_addr_end(addr, end);
 424                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 425                         if (pud_none_or_clear_bad(pud))
 426                                 continue;
 427                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 428                                                ceiling);
 429                 } else {
 430                         unsigned long more;
 431                         /*
 432                          * Increment next by the size of the huge mapping since
 433                          * there may be more than one entry at this level for a
 434                          * single hugepage, but all of them point to
 435                          * the same kmem cache that holds the hugepte.
 436                          */
 437                         more = addr + (1UL << hugepd_shift(*(hugepd_t *)pud));
 438                         if (more > next)
 439                                 next = more;
 440
 441                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 442                                           addr, next, floor, ceiling);
 443                 }
 444         } while (addr = next, addr != end);
 445
 446         if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
 447                 return;
 448
 449         pud = pud_offset(p4d, start & PGDIR_MASK);
 450         p4d_clear(p4d);
 451         pud_free_tlb(tlb, pud, start & PGDIR_MASK);
 452         mm_dec_nr_puds(tlb->mm);
 453 }
 454
 455 /*
 456  * This function frees user-level page tables of a process.
 457  */
 458 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 459                             unsigned long addr, unsigned long end,
 460                             unsigned long floor, unsigned long ceiling)
 461 {
 462         pgd_t *pgd;
 463         p4d_t *p4d;
 464         unsigned long next;
 465
 466         /*
 467          * Because there are a number of different possible pagetable
 468          * layouts for hugepage ranges, we limit knowledge of how
 469          * things should be laid out to the allocation path
 470          * (huge_pte_alloc(), above).  Everything else works out the
 471          * structure as it goes from information in the hugepd
 472          * pointers.  That means that we can't here use the
 473          * optimization used in the normal page free_pgd_range(), of
 474          * checking whether we're actually covering a large enough
 475          * range to have to do anything at the top level of the walk
 476          * instead of at the bottom.
 477          *
 478          * To make sense of this, you should probably go read the big
 479          * block comment at the top of the normal free_pgd_range(),
 480          * too.
 481          */
 482
 483         do {
 484                 next = pgd_addr_end(addr, end);
 485                 pgd = pgd_offset(tlb->mm, addr);
 486                 p4d = p4d_offset(pgd, addr);
 487                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 488                         if (p4d_none_or_clear_bad(p4d))
 489                                 continue;
 490                         hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 491                 } else {
 492                         unsigned long more;
 493                         /*
 494                          * Increment next by the size of the huge mapping since
 495                          * there may be more than one entry at the pgd level
 496                          * for a single hugepage, but all of them point to the
 497                          * same kmem cache that holds the hugepte.
 498                          */
 499                         more = addr + (1UL << hugepd_shift(*(hugepd_t *)pgd));
 500                         if (more > next)
 501                                 next = more;
 502
 503                         free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
 504                                           addr, next, floor, ceiling);
 505                 }
 506         } while (addr = next, addr != end);
 507 }
 508
 509 bool __init arch_hugetlb_valid_size(unsigned long size)
 510 {
 511         int shift = __ffs(size);
 512         int mmu_psize;
 513
 514         /* Check that it is a page size supported by the hardware and
 515          * that it fits within pagetable and slice limits. */
 516         if (size <= PAGE_SIZE || !is_power_of_2(size))
 517                 return false;
 518
 519         mmu_psize = check_and_get_huge_psize(shift);
 520         if (mmu_psize < 0)
 521                 return false;
 522
 523         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 524
 525         return true;
 526 }
 527
 528 static int __init add_huge_page_size(unsigned long long size)
 529 {
 530         int shift = __ffs(size);
 531
 532         if (!arch_hugetlb_valid_size((unsigned long)size))
 533                 return -EINVAL;
 534
 535         hugetlb_add_hstate(shift - PAGE_SHIFT);
 536         return 0;
 537 }
 538
 539 static int __init hugetlbpage_init(void)
 540 {
 541         bool configured = false;
 542         int psize;
 543
 544         if (hugetlb_disabled) {
 545                 pr_info("HugeTLB support is disabled!\n");
 546                 return 0;
 547         }
 548
 549         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
 550             !mmu_has_feature(MMU_FTR_16M_PAGE))
 551                 return -ENODEV;
 552
 553         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 554                 unsigned shift;
 555                 unsigned pdshift;
 556
 557                 if (!mmu_psize_defs[psize].shift)
 558                         continue;
 559
 560                 shift = mmu_psize_to_shift(psize);
 561
 562 #ifdef CONFIG_PPC_BOOK3S_64
 563                 if (shift > PGDIR_SHIFT)
 564                         continue;
 565                 else if (shift > PUD_SHIFT)
 566                         pdshift = PGDIR_SHIFT;
 567                 else if (shift > PMD_SHIFT)
 568                         pdshift = PUD_SHIFT;
 569                 else
 570                         pdshift = PMD_SHIFT;
 571 #else
 572                 if (shift < PUD_SHIFT)
 573                         pdshift = PMD_SHIFT;
 574                 else if (shift < PGDIR_SHIFT)
 575                         pdshift = PUD_SHIFT;
 576                 else
 577                         pdshift = PGDIR_SHIFT;
 578 #endif
 579
 580                 if (add_huge_page_size(1ULL << shift) < 0)
 581                         continue;
 582                 /*
 583                  * if we have pdshift and shift value same, we don't
 584                  * use pgt cache for hugepd.
 585                  */
 586                 if (pdshift > shift) {
 587                         if (!IS_ENABLED(CONFIG_PPC_8xx))
 588                                 pgtable_cache_add(pdshift - shift);
 589                 } else if (IS_ENABLED(CONFIG_PPC_E500) ||
 590                            IS_ENABLED(CONFIG_PPC_8xx)) {
 591                         pgtable_cache_add(PTE_T_ORDER);
 592                 }
 593
 594                 configured = true;
 595         }
 596
 597         if (!configured)
 598                 pr_info("Failed to initialize. Disabling HugeTLB");
 599
 600         return 0;
 601 }
 602
 603 arch_initcall(hugetlbpage_init);
 604
 605 void __init gigantic_hugetlb_cma_reserve(void)
 606 {
 607         unsigned long order = 0;
 608
 609         if (radix_enabled())
 610                 order = PUD_SHIFT - PAGE_SHIFT;
 611         else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
 612                 /*
 613                  * For pseries we do use ibm,expected#pages for reserving 16G pages.
 614                  */
 615                 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
 616
 617         if (order) {
 618                 VM_WARN_ON(order <= MAX_ORDER);
 619                 hugetlb_cma_reserve(order);
 620         }
 621 }