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 #include <asm/pgalloc.h>
  21
  22 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
  23 bool arch_hugetlb_migration_supported(struct hstate *h)
  24 {
  25         size_t pagesize = huge_page_size(h);
  26
  27         switch (pagesize) {
  28 #ifdef CONFIG_ARM64_4K_PAGES
  29         case PUD_SIZE:
  30 #endif
  31         case PMD_SIZE:
  32         case CONT_PMD_SIZE:
  33         case CONT_PTE_SIZE:
  34                 return true;
  35         }
  36         pr_warn("%s: unrecognized huge page size 0x%lx\n",
  37                         __func__, pagesize);
  38         return false;
  39 }
  40 #endif
  41
  42 int pmd_huge(pmd_t pmd)
  43 {
  44         return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
  45 }
  46
  47 int pud_huge(pud_t pud)
  48 {
  49 #ifndef __PAGETABLE_PMD_FOLDED
  50         return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
  51 #else
  52         return 0;
  53 #endif
  54 }
  55
  56 /*
  57  * Select all bits except the pfn
  58  */
  59 static inline pgprot_t pte_pgprot(pte_t pte)
  60 {
  61         unsigned long pfn = pte_pfn(pte);
  62
  63         return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
  64 }
  65
  66 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
  67                            pte_t *ptep, size_t *pgsize)
  68 {
  69         pgd_t *pgdp = pgd_offset(mm, addr);
  70         pud_t *pudp;
  71         pmd_t *pmdp;
  72
  73         *pgsize = PAGE_SIZE;
  74         pudp = pud_offset(pgdp, addr);
  75         pmdp = pmd_offset(pudp, addr);
  76         if ((pte_t *)pmdp == ptep) {
  77                 *pgsize = PMD_SIZE;
  78                 return CONT_PMDS;
  79         }
  80         return CONT_PTES;
  81 }
  82
  83 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
  84 {
  85         int contig_ptes = 0;
  86
  87         *pgsize = size;
  88
  89         switch (size) {
  90 #ifdef CONFIG_ARM64_4K_PAGES
  91         case PUD_SIZE:
  92 #endif
  93         case PMD_SIZE:
  94                 contig_ptes = 1;
  95                 break;
  96         case CONT_PMD_SIZE:
  97                 *pgsize = PMD_SIZE;
  98                 contig_ptes = CONT_PMDS;
  99                 break;
 100         case CONT_PTE_SIZE:
 101                 *pgsize = PAGE_SIZE;
 102                 contig_ptes = CONT_PTES;
 103                 break;
 104         }
 105
 106         return contig_ptes;
 107 }
 108
 109 /*
 110  * Changing some bits of contiguous entries requires us to follow a
 111  * Break-Before-Make approach, breaking the whole contiguous set
 112  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 113  * "Misprogramming of the Contiguous bit", page D4-1762.
 114  *
 115  * This helper performs the break step.
 116  */
 117 static pte_t get_clear_flush(struct mm_struct *mm,
 118                              unsigned long addr,
 119                              pte_t *ptep,
 120                              unsigned long pgsize,
 121                              unsigned long ncontig)
 122 {
 123         pte_t orig_pte = huge_ptep_get(ptep);
 124         bool valid = pte_valid(orig_pte);
 125         unsigned long i, saddr = addr;
 126
 127         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
 128                 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
 129
 130                 /*
 131                  * If HW_AFDBM is enabled, then the HW could turn on
 132                  * the dirty or accessed bit for any page in the set,
 133                  * so check them all.
 134                  */
 135                 if (pte_dirty(pte))
 136                         orig_pte = pte_mkdirty(orig_pte);
 137
 138                 if (pte_young(pte))
 139                         orig_pte = pte_mkyoung(orig_pte);
 140         }
 141
 142         if (valid) {
 143                 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
 144                 flush_tlb_range(&vma, saddr, addr);
 145         }
 146         return orig_pte;
 147 }
 148
 149 /*
 150  * Changing some bits of contiguous entries requires us to follow a
 151  * Break-Before-Make approach, breaking the whole contiguous set
 152  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 153  * "Misprogramming of the Contiguous bit", page D4-1762.
 154  *
 155  * This helper performs the break step for use cases where the
 156  * original pte is not needed.
 157  */
 158 static void clear_flush(struct mm_struct *mm,
 159                              unsigned long addr,
 160                              pte_t *ptep,
 161                              unsigned long pgsize,
 162                              unsigned long ncontig)
 163 {
 164         struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
 165         unsigned long i, saddr = addr;
 166
 167         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 168                 pte_clear(mm, addr, ptep);
 169
 170         flush_tlb_range(&vma, saddr, addr);
 171 }
 172
 173 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 174                             pte_t *ptep, pte_t pte)
 175 {
 176         size_t pgsize;
 177         int i;
 178         int ncontig;
 179         unsigned long pfn, dpfn;
 180         pgprot_t hugeprot;
 181
 182         /*
 183          * Code needs to be expanded to handle huge swap and migration
 184          * entries. Needed for HUGETLB and MEMORY_FAILURE.
 185          */
 186         WARN_ON(!pte_present(pte));
 187
 188         if (!pte_cont(pte)) {
 189                 set_pte_at(mm, addr, ptep, pte);
 190                 return;
 191         }
 192
 193         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 194         pfn = pte_pfn(pte);
 195         dpfn = pgsize >> PAGE_SHIFT;
 196         hugeprot = pte_pgprot(pte);
 197
 198         clear_flush(mm, addr, ptep, pgsize, ncontig);
 199
 200         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 201                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 202 }
 203
 204 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
 205                           pte_t *ptep, pte_t pte, unsigned long sz)
 206 {
 207         int i, ncontig;
 208         size_t pgsize;
 209
 210         ncontig = num_contig_ptes(sz, &pgsize);
 211
 212         for (i = 0; i < ncontig; i++, ptep++)
 213                 set_pte(ptep, pte);
 214 }
 215
 216 pte_t *huge_pte_alloc(struct mm_struct *mm,
 217                       unsigned long addr, unsigned long sz)
 218 {
 219         pgd_t *pgdp;
 220         pud_t *pudp;
 221         pmd_t *pmdp;
 222         pte_t *ptep = NULL;
 223
 224         pgdp = pgd_offset(mm, addr);
 225         pudp = pud_alloc(mm, pgdp, addr);
 226         if (!pudp)
 227                 return NULL;
 228
 229         if (sz == PUD_SIZE) {
 230                 ptep = (pte_t *)pudp;
 231         } else if (sz == (CONT_PTE_SIZE)) {
 232                 pmdp = pmd_alloc(mm, pudp, addr);
 233
 234                 WARN_ON(addr & (sz - 1));
 235                 /*
 236                  * Note that if this code were ever ported to the
 237                  * 32-bit arm platform then it will cause trouble in
 238                  * the case where CONFIG_HIGHPTE is set, since there
 239                  * will be no pte_unmap() to correspond with this
 240                  * pte_alloc_map().
 241                  */
 242                 ptep = pte_alloc_map(mm, pmdp, addr);
 243         } else if (sz == PMD_SIZE) {
 244                 if (IS_ENABLED(CONFIG_ARCH_WANT_HUGE_PMD_SHARE) &&
 245                     pud_none(READ_ONCE(*pudp)))
 246                         ptep = huge_pmd_share(mm, addr, pudp);
 247                 else
 248                         ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
 249         } else if (sz == (CONT_PMD_SIZE)) {
 250                 pmdp = pmd_alloc(mm, pudp, addr);
 251                 WARN_ON(addr & (sz - 1));
 252                 return (pte_t *)pmdp;
 253         }
 254
 255         return ptep;
 256 }
 257
 258 pte_t *huge_pte_offset(struct mm_struct *mm,
 259                        unsigned long addr, unsigned long sz)
 260 {
 261         pgd_t *pgdp;
 262         pud_t *pudp, pud;
 263         pmd_t *pmdp, pmd;
 264
 265         pgdp = pgd_offset(mm, addr);
 266         if (!pgd_present(READ_ONCE(*pgdp)))
 267                 return NULL;
 268
 269         pudp = pud_offset(pgdp, addr);
 270         pud = READ_ONCE(*pudp);
 271         if (sz != PUD_SIZE && pud_none(pud))
 272                 return NULL;
 273         /* hugepage or swap? */
 274         if (pud_huge(pud) || !pud_present(pud))
 275                 return (pte_t *)pudp;
 276         /* table; check the next level */
 277
 278         if (sz == CONT_PMD_SIZE)
 279                 addr &= CONT_PMD_MASK;
 280
 281         pmdp = pmd_offset(pudp, addr);
 282         pmd = READ_ONCE(*pmdp);
 283         if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
 284             pmd_none(pmd))
 285                 return NULL;
 286         if (pmd_huge(pmd) || !pmd_present(pmd))
 287                 return (pte_t *)pmdp;
 288
 289         if (sz == CONT_PTE_SIZE)
 290                 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
 291
 292         return NULL;
 293 }
 294
 295 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
 296                          struct page *page, int writable)
 297 {
 298         size_t pagesize = huge_page_size(hstate_vma(vma));
 299
 300         if (pagesize == CONT_PTE_SIZE) {
 301                 entry = pte_mkcont(entry);
 302         } else if (pagesize == CONT_PMD_SIZE) {
 303                 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
 304         } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
 305                 pr_warn("%s: unrecognized huge page size 0x%lx\n",
 306                         __func__, pagesize);
 307         }
 308         return entry;
 309 }
 310
 311 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
 312                     pte_t *ptep, unsigned long sz)
 313 {
 314         int i, ncontig;
 315         size_t pgsize;
 316
 317         ncontig = num_contig_ptes(sz, &pgsize);
 318
 319         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 320                 pte_clear(mm, addr, ptep);
 321 }
 322
 323 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
 324                               unsigned long addr, pte_t *ptep)
 325 {
 326         int ncontig;
 327         size_t pgsize;
 328         pte_t orig_pte = huge_ptep_get(ptep);
 329
 330         if (!pte_cont(orig_pte))
 331                 return ptep_get_and_clear(mm, addr, ptep);
 332
 333         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 334
 335         return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 336 }
 337
 338 /*
 339  * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
 340  * and write permission.
 341  *
 342  * For a contiguous huge pte range we need to check whether or not write
 343  * permission has to change only on the first pte in the set. Then for
 344  * all the contiguous ptes we need to check whether or not there is a
 345  * discrepancy between dirty or young.
 346  */
 347 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
 348 {
 349         int i;
 350
 351         if (pte_write(pte) != pte_write(huge_ptep_get(ptep)))
 352                 return 1;
 353
 354         for (i = 0; i < ncontig; i++) {
 355                 pte_t orig_pte = huge_ptep_get(ptep + i);
 356
 357                 if (pte_dirty(pte) != pte_dirty(orig_pte))
 358                         return 1;
 359
 360                 if (pte_young(pte) != pte_young(orig_pte))
 361                         return 1;
 362         }
 363
 364         return 0;
 365 }
 366
 367 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
 368                                unsigned long addr, pte_t *ptep,
 369                                pte_t pte, int dirty)
 370 {
 371         int ncontig, i;
 372         size_t pgsize = 0;
 373         unsigned long pfn = pte_pfn(pte), dpfn;
 374         pgprot_t hugeprot;
 375         pte_t orig_pte;
 376
 377         if (!pte_cont(pte))
 378                 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
 379
 380         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 381         dpfn = pgsize >> PAGE_SHIFT;
 382
 383         if (!__cont_access_flags_changed(ptep, pte, ncontig))
 384                 return 0;
 385
 386         orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 387
 388         /* Make sure we don't lose the dirty or young state */
 389         if (pte_dirty(orig_pte))
 390                 pte = pte_mkdirty(pte);
 391
 392         if (pte_young(orig_pte))
 393                 pte = pte_mkyoung(pte);
 394
 395         hugeprot = pte_pgprot(pte);
 396         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 397                 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
 398
 399         return 1;
 400 }
 401
 402 void huge_ptep_set_wrprotect(struct mm_struct *mm,
 403                              unsigned long addr, pte_t *ptep)
 404 {
 405         unsigned long pfn, dpfn;
 406         pgprot_t hugeprot;
 407         int ncontig, i;
 408         size_t pgsize;
 409         pte_t pte;
 410
 411         if (!pte_cont(READ_ONCE(*ptep))) {
 412                 ptep_set_wrprotect(mm, addr, ptep);
 413                 return;
 414         }
 415
 416         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 417         dpfn = pgsize >> PAGE_SHIFT;
 418
 419         pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 420         pte = pte_wrprotect(pte);
 421
 422         hugeprot = pte_pgprot(pte);
 423         pfn = pte_pfn(pte);
 424
 425         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 426                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 427 }
 428
 429 void huge_ptep_clear_flush(struct vm_area_struct *vma,
 430                            unsigned long addr, pte_t *ptep)
 431 {
 432         size_t pgsize;
 433         int ncontig;
 434
 435         if (!pte_cont(READ_ONCE(*ptep))) {
 436                 ptep_clear_flush(vma, addr, ptep);
 437                 return;
 438         }
 439
 440         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 441         clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 442 }
 443
 444 static void __init add_huge_page_size(unsigned long size)
 445 {
 446         if (size_to_hstate(size))
 447                 return;
 448
 449         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
 450 }
 451
 452 static int __init hugetlbpage_init(void)
 453 {
 454 #ifdef CONFIG_ARM64_4K_PAGES
 455         add_huge_page_size(PUD_SIZE);
 456 #endif
 457         add_huge_page_size(CONT_PMD_SIZE);
 458         add_huge_page_size(PMD_SIZE);
 459         add_huge_page_size(CONT_PTE_SIZE);
 460
 461         return 0;
 462 }
 463 arch_initcall(hugetlbpage_init);
 464
 465 static __init int setup_hugepagesz(char *opt)
 466 {
 467         unsigned long ps = memparse(opt, &opt);
 468
 469         switch (ps) {
 470 #ifdef CONFIG_ARM64_4K_PAGES
 471         case PUD_SIZE:
 472 #endif
 473         case CONT_PMD_SIZE:
 474         case PMD_SIZE:
 475         case CONT_PTE_SIZE:
 476                 add_huge_page_size(ps);
 477                 return 1;
 478         }
 479
 480         hugetlb_bad_size();
 481         pr_err("hugepagesz: Unsupported page size %lu K\n", ps >> 10);
 482         return 0;
 483 }
 484 __setup("hugepagesz=", setup_hugepagesz);