arch/arm64/mm/hugetlbpage.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * arch/arm64/mm/hugetlbpage.c
   4  *
   5  * Copyright (C) 2013 Linaro Ltd.
   6  *
   7  * Based on arch/x86/mm/hugetlbpage.c.
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/err.h>
  16 #include <linux/sysctl.h>
  17 #include <asm/mman.h>
  18 #include <asm/tlb.h>
  19 #include <asm/tlbflush.h>
  20
  21 /*
  22  * HugeTLB Support Matrix
  23  *
  24  * ---------------------------------------------------
  25  * | Page Size | CONT PTE |  PMD  | CONT PMD |  PUD  |
  26  * ---------------------------------------------------
  27  * |     4K    |   64K    |   2M  |    32M   |   1G  |
  28  * |    16K    |    2M    |  32M  |     1G   |       |
  29  * |    64K    |    2M    | 512M  |    16G   |       |
  30  * ---------------------------------------------------
  31  */
  32
  33 /*
  34  * Reserve CMA areas for the largest supported gigantic
  35  * huge page when requested. Any other smaller gigantic
  36  * huge pages could still be served from those areas.
  37  */
  38 #ifdef CONFIG_CMA
  39 void __init arm64_hugetlb_cma_reserve(void)
  40 {
  41         int order;
  42
  43 #ifdef CONFIG_ARM64_4K_PAGES
  44         order = PUD_SHIFT - PAGE_SHIFT;
  45 #else
  46         order = CONT_PMD_SHIFT + PMD_SHIFT - PAGE_SHIFT;
  47 #endif
  48         /*
  49          * HugeTLB CMA reservation is required for gigantic
  50          * huge pages which could not be allocated via the
  51          * page allocator. Just warn if there is any change
  52          * breaking this assumption.
  53          */
  54         WARN_ON(order <= MAX_ORDER);
  55         hugetlb_cma_reserve(order);
  56 }
  57 #endif /* CONFIG_CMA */
  58
  59 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
  60 bool arch_hugetlb_migration_supported(struct hstate *h)
  61 {
  62         size_t pagesize = huge_page_size(h);
  63
  64         switch (pagesize) {
  65 #ifdef CONFIG_ARM64_4K_PAGES
  66         case PUD_SIZE:
  67 #endif
  68         case PMD_SIZE:
  69         case CONT_PMD_SIZE:
  70         case CONT_PTE_SIZE:
  71                 return true;
  72         }
  73         pr_warn("%s: unrecognized huge page size 0x%lx\n",
  74                         __func__, pagesize);
  75         return false;
  76 }
  77 #endif
  78
  79 int pmd_huge(pmd_t pmd)
  80 {
  81         return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
  82 }
  83
  84 int pud_huge(pud_t pud)
  85 {
  86 #ifndef __PAGETABLE_PMD_FOLDED
  87         return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
  88 #else
  89         return 0;
  90 #endif
  91 }
  92
  93 /*
  94  * Select all bits except the pfn
  95  */
  96 static inline pgprot_t pte_pgprot(pte_t pte)
  97 {
  98         unsigned long pfn = pte_pfn(pte);
  99
 100         return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
 101 }
 102
 103 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
 104                            pte_t *ptep, size_t *pgsize)
 105 {
 106         pgd_t *pgdp = pgd_offset(mm, addr);
 107         p4d_t *p4dp;
 108         pud_t *pudp;
 109         pmd_t *pmdp;
 110
 111         *pgsize = PAGE_SIZE;
 112         p4dp = p4d_offset(pgdp, addr);
 113         pudp = pud_offset(p4dp, addr);
 114         pmdp = pmd_offset(pudp, addr);
 115         if ((pte_t *)pmdp == ptep) {
 116                 *pgsize = PMD_SIZE;
 117                 return CONT_PMDS;
 118         }
 119         return CONT_PTES;
 120 }
 121
 122 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
 123 {
 124         int contig_ptes = 0;
 125
 126         *pgsize = size;
 127
 128         switch (size) {
 129 #ifdef CONFIG_ARM64_4K_PAGES
 130         case PUD_SIZE:
 131 #endif
 132         case PMD_SIZE:
 133                 contig_ptes = 1;
 134                 break;
 135         case CONT_PMD_SIZE:
 136                 *pgsize = PMD_SIZE;
 137                 contig_ptes = CONT_PMDS;
 138                 break;
 139         case CONT_PTE_SIZE:
 140                 *pgsize = PAGE_SIZE;
 141                 contig_ptes = CONT_PTES;
 142                 break;
 143         }
 144
 145         return contig_ptes;
 146 }
 147
 148 /*
 149  * Changing some bits of contiguous entries requires us to follow a
 150  * Break-Before-Make approach, breaking the whole contiguous set
 151  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 152  * "Misprogramming of the Contiguous bit", page D4-1762.
 153  *
 154  * This helper performs the break step.
 155  */
 156 static pte_t get_clear_flush(struct mm_struct *mm,
 157                              unsigned long addr,
 158                              pte_t *ptep,
 159                              unsigned long pgsize,
 160                              unsigned long ncontig)
 161 {
 162         pte_t orig_pte = huge_ptep_get(ptep);
 163         bool valid = pte_valid(orig_pte);
 164         unsigned long i, saddr = addr;
 165
 166         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
 167                 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
 168
 169                 /*
 170                  * If HW_AFDBM is enabled, then the HW could turn on
 171                  * the dirty or accessed bit for any page in the set,
 172                  * so check them all.
 173                  */
 174                 if (pte_dirty(pte))
 175                         orig_pte = pte_mkdirty(orig_pte);
 176
 177                 if (pte_young(pte))
 178                         orig_pte = pte_mkyoung(orig_pte);
 179         }
 180
 181         if (valid) {
 182                 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
 183                 flush_tlb_range(&vma, saddr, addr);
 184         }
 185         return orig_pte;
 186 }
 187
 188 /*
 189  * Changing some bits of contiguous entries requires us to follow a
 190  * Break-Before-Make approach, breaking the whole contiguous set
 191  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 192  * "Misprogramming of the Contiguous bit", page D4-1762.
 193  *
 194  * This helper performs the break step for use cases where the
 195  * original pte is not needed.
 196  */
 197 static void clear_flush(struct mm_struct *mm,
 198                              unsigned long addr,
 199                              pte_t *ptep,
 200                              unsigned long pgsize,
 201                              unsigned long ncontig)
 202 {
 203         struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
 204         unsigned long i, saddr = addr;
 205
 206         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 207                 pte_clear(mm, addr, ptep);
 208
 209         flush_tlb_range(&vma, saddr, addr);
 210 }
 211
 212 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 213                             pte_t *ptep, pte_t pte)
 214 {
 215         size_t pgsize;
 216         int i;
 217         int ncontig;
 218         unsigned long pfn, dpfn;
 219         pgprot_t hugeprot;
 220
 221         /*
 222          * Code needs to be expanded to handle huge swap and migration
 223          * entries. Needed for HUGETLB and MEMORY_FAILURE.
 224          */
 225         WARN_ON(!pte_present(pte));
 226
 227         if (!pte_cont(pte)) {
 228                 set_pte_at(mm, addr, ptep, pte);
 229                 return;
 230         }
 231
 232         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 233         pfn = pte_pfn(pte);
 234         dpfn = pgsize >> PAGE_SHIFT;
 235         hugeprot = pte_pgprot(pte);
 236
 237         clear_flush(mm, addr, ptep, pgsize, ncontig);
 238
 239         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 240                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 241 }
 242
 243 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
 244                           pte_t *ptep, pte_t pte, unsigned long sz)
 245 {
 246         int i, ncontig;
 247         size_t pgsize;
 248
 249         ncontig = num_contig_ptes(sz, &pgsize);
 250
 251         for (i = 0; i < ncontig; i++, ptep++)
 252                 set_pte(ptep, pte);
 253 }
 254
 255 pte_t *huge_pte_alloc(struct mm_struct *mm,
 256                       unsigned long addr, unsigned long sz)
 257 {
 258         pgd_t *pgdp;
 259         p4d_t *p4dp;
 260         pud_t *pudp;
 261         pmd_t *pmdp;
 262         pte_t *ptep = NULL;
 263
 264         pgdp = pgd_offset(mm, addr);
 265         p4dp = p4d_offset(pgdp, addr);
 266         pudp = pud_alloc(mm, p4dp, addr);
 267         if (!pudp)
 268                 return NULL;
 269
 270         if (sz == PUD_SIZE) {
 271                 ptep = (pte_t *)pudp;
 272         } else if (sz == (CONT_PTE_SIZE)) {
 273                 pmdp = pmd_alloc(mm, pudp, addr);
 274                 if (!pmdp)
 275                         return NULL;
 276
 277                 WARN_ON(addr & (sz - 1));
 278                 /*
 279                  * Note that if this code were ever ported to the
 280                  * 32-bit arm platform then it will cause trouble in
 281                  * the case where CONFIG_HIGHPTE is set, since there
 282                  * will be no pte_unmap() to correspond with this
 283                  * pte_alloc_map().
 284                  */
 285                 ptep = pte_alloc_map(mm, pmdp, addr);
 286         } else if (sz == PMD_SIZE) {
 287                 if (IS_ENABLED(CONFIG_ARCH_WANT_HUGE_PMD_SHARE) &&
 288                     pud_none(READ_ONCE(*pudp)))
 289                         ptep = huge_pmd_share(mm, addr, pudp);
 290                 else
 291                         ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
 292         } else if (sz == (CONT_PMD_SIZE)) {
 293                 pmdp = pmd_alloc(mm, pudp, addr);
 294                 WARN_ON(addr & (sz - 1));
 295                 return (pte_t *)pmdp;
 296         }
 297
 298         return ptep;
 299 }
 300
 301 pte_t *huge_pte_offset(struct mm_struct *mm,
 302                        unsigned long addr, unsigned long sz)
 303 {
 304         pgd_t *pgdp;
 305         p4d_t *p4dp;
 306         pud_t *pudp, pud;
 307         pmd_t *pmdp, pmd;
 308
 309         pgdp = pgd_offset(mm, addr);
 310         if (!pgd_present(READ_ONCE(*pgdp)))
 311                 return NULL;
 312
 313         p4dp = p4d_offset(pgdp, addr);
 314         if (!p4d_present(READ_ONCE(*p4dp)))
 315                 return NULL;
 316
 317         pudp = pud_offset(p4dp, addr);
 318         pud = READ_ONCE(*pudp);
 319         if (sz != PUD_SIZE && pud_none(pud))
 320                 return NULL;
 321         /* hugepage or swap? */
 322         if (pud_huge(pud) || !pud_present(pud))
 323                 return (pte_t *)pudp;
 324         /* table; check the next level */
 325
 326         if (sz == CONT_PMD_SIZE)
 327                 addr &= CONT_PMD_MASK;
 328
 329         pmdp = pmd_offset(pudp, addr);
 330         pmd = READ_ONCE(*pmdp);
 331         if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
 332             pmd_none(pmd))
 333                 return NULL;
 334         if (pmd_huge(pmd) || !pmd_present(pmd))
 335                 return (pte_t *)pmdp;
 336
 337         if (sz == CONT_PTE_SIZE)
 338                 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
 339
 340         return NULL;
 341 }
 342
 343 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
 344                          struct page *page, int writable)
 345 {
 346         size_t pagesize = huge_page_size(hstate_vma(vma));
 347
 348         if (pagesize == CONT_PTE_SIZE) {
 349                 entry = pte_mkcont(entry);
 350         } else if (pagesize == CONT_PMD_SIZE) {
 351                 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
 352         } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
 353                 pr_warn("%s: unrecognized huge page size 0x%lx\n",
 354                         __func__, pagesize);
 355         }
 356         return entry;
 357 }
 358
 359 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
 360                     pte_t *ptep, unsigned long sz)
 361 {
 362         int i, ncontig;
 363         size_t pgsize;
 364
 365         ncontig = num_contig_ptes(sz, &pgsize);
 366
 367         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 368                 pte_clear(mm, addr, ptep);
 369 }
 370
 371 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
 372                               unsigned long addr, pte_t *ptep)
 373 {
 374         int ncontig;
 375         size_t pgsize;
 376         pte_t orig_pte = huge_ptep_get(ptep);
 377
 378         if (!pte_cont(orig_pte))
 379                 return ptep_get_and_clear(mm, addr, ptep);
 380
 381         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 382
 383         return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 384 }
 385
 386 /*
 387  * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
 388  * and write permission.
 389  *
 390  * For a contiguous huge pte range we need to check whether or not write
 391  * permission has to change only on the first pte in the set. Then for
 392  * all the contiguous ptes we need to check whether or not there is a
 393  * discrepancy between dirty or young.
 394  */
 395 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
 396 {
 397         int i;
 398
 399         if (pte_write(pte) != pte_write(huge_ptep_get(ptep)))
 400                 return 1;
 401
 402         for (i = 0; i < ncontig; i++) {
 403                 pte_t orig_pte = huge_ptep_get(ptep + i);
 404
 405                 if (pte_dirty(pte) != pte_dirty(orig_pte))
 406                         return 1;
 407
 408                 if (pte_young(pte) != pte_young(orig_pte))
 409                         return 1;
 410         }
 411
 412         return 0;
 413 }
 414
 415 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
 416                                unsigned long addr, pte_t *ptep,
 417                                pte_t pte, int dirty)
 418 {
 419         int ncontig, i;
 420         size_t pgsize = 0;
 421         unsigned long pfn = pte_pfn(pte), dpfn;
 422         pgprot_t hugeprot;
 423         pte_t orig_pte;
 424
 425         if (!pte_cont(pte))
 426                 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
 427
 428         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 429         dpfn = pgsize >> PAGE_SHIFT;
 430
 431         if (!__cont_access_flags_changed(ptep, pte, ncontig))
 432                 return 0;
 433
 434         orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 435
 436         /* Make sure we don't lose the dirty or young state */
 437         if (pte_dirty(orig_pte))
 438                 pte = pte_mkdirty(pte);
 439
 440         if (pte_young(orig_pte))
 441                 pte = pte_mkyoung(pte);
 442
 443         hugeprot = pte_pgprot(pte);
 444         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 445                 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
 446
 447         return 1;
 448 }
 449
 450 void huge_ptep_set_wrprotect(struct mm_struct *mm,
 451                              unsigned long addr, pte_t *ptep)
 452 {
 453         unsigned long pfn, dpfn;
 454         pgprot_t hugeprot;
 455         int ncontig, i;
 456         size_t pgsize;
 457         pte_t pte;
 458
 459         if (!pte_cont(READ_ONCE(*ptep))) {
 460                 ptep_set_wrprotect(mm, addr, ptep);
 461                 return;
 462         }
 463
 464         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 465         dpfn = pgsize >> PAGE_SHIFT;
 466
 467         pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 468         pte = pte_wrprotect(pte);
 469
 470         hugeprot = pte_pgprot(pte);
 471         pfn = pte_pfn(pte);
 472
 473         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 474                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 475 }
 476
 477 void huge_ptep_clear_flush(struct vm_area_struct *vma,
 478                            unsigned long addr, pte_t *ptep)
 479 {
 480         size_t pgsize;
 481         int ncontig;
 482
 483         if (!pte_cont(READ_ONCE(*ptep))) {
 484                 ptep_clear_flush(vma, addr, ptep);
 485                 return;
 486         }
 487
 488         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 489         clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 490 }
 491
 492 static int __init hugetlbpage_init(void)
 493 {
 494 #ifdef CONFIG_ARM64_4K_PAGES
 495         hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
 496 #endif
 497         hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT);
 498         hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
 499         hugetlb_add_hstate(CONT_PTE_SHIFT - PAGE_SHIFT);
 500
 501         return 0;
 502 }
 503 arch_initcall(hugetlbpage_init);
 504
 505 bool __init arch_hugetlb_valid_size(unsigned long size)
 506 {
 507         switch (size) {
 508 #ifdef CONFIG_ARM64_4K_PAGES
 509         case PUD_SIZE:
 510 #endif
 511         case CONT_PMD_SIZE:
 512         case PMD_SIZE:
 513         case CONT_PTE_SIZE:
 514                 return true;
 515         }
 516
 517         return false;
 518 }