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                 if (!pmdp)
 234                         return NULL;
 235
 236                 WARN_ON(addr & (sz - 1));
 237                 /*
 238                  * Note that if this code were ever ported to the
 239                  * 32-bit arm platform then it will cause trouble in
 240                  * the case where CONFIG_HIGHPTE is set, since there
 241                  * will be no pte_unmap() to correspond with this
 242                  * pte_alloc_map().
 243                  */
 244                 ptep = pte_alloc_map(mm, pmdp, addr);
 245         } else if (sz == PMD_SIZE) {
 246                 if (IS_ENABLED(CONFIG_ARCH_WANT_HUGE_PMD_SHARE) &&
 247                     pud_none(READ_ONCE(*pudp)))
 248                         ptep = huge_pmd_share(mm, addr, pudp);
 249                 else
 250                         ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
 251         } else if (sz == (CONT_PMD_SIZE)) {
 252                 pmdp = pmd_alloc(mm, pudp, addr);
 253                 WARN_ON(addr & (sz - 1));
 254                 return (pte_t *)pmdp;
 255         }
 256
 257         return ptep;
 258 }
 259
 260 pte_t *huge_pte_offset(struct mm_struct *mm,
 261                        unsigned long addr, unsigned long sz)
 262 {
 263         pgd_t *pgdp;
 264         pud_t *pudp, pud;
 265         pmd_t *pmdp, pmd;
 266
 267         pgdp = pgd_offset(mm, addr);
 268         if (!pgd_present(READ_ONCE(*pgdp)))
 269                 return NULL;
 270
 271         pudp = pud_offset(pgdp, addr);
 272         pud = READ_ONCE(*pudp);
 273         if (sz != PUD_SIZE && pud_none(pud))
 274                 return NULL;
 275         /* hugepage or swap? */
 276         if (pud_huge(pud) || !pud_present(pud))
 277                 return (pte_t *)pudp;
 278         /* table; check the next level */
 279
 280         if (sz == CONT_PMD_SIZE)
 281                 addr &= CONT_PMD_MASK;
 282
 283         pmdp = pmd_offset(pudp, addr);
 284         pmd = READ_ONCE(*pmdp);
 285         if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
 286             pmd_none(pmd))
 287                 return NULL;
 288         if (pmd_huge(pmd) || !pmd_present(pmd))
 289                 return (pte_t *)pmdp;
 290
 291         if (sz == CONT_PTE_SIZE)
 292                 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
 293
 294         return NULL;
 295 }
 296
 297 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
 298                          struct page *page, int writable)
 299 {
 300         size_t pagesize = huge_page_size(hstate_vma(vma));
 301
 302         if (pagesize == CONT_PTE_SIZE) {
 303                 entry = pte_mkcont(entry);
 304         } else if (pagesize == CONT_PMD_SIZE) {
 305                 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
 306         } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
 307                 pr_warn("%s: unrecognized huge page size 0x%lx\n",
 308                         __func__, pagesize);
 309         }
 310         return entry;
 311 }
 312
 313 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
 314                     pte_t *ptep, unsigned long sz)
 315 {
 316         int i, ncontig;
 317         size_t pgsize;
 318
 319         ncontig = num_contig_ptes(sz, &pgsize);
 320
 321         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 322                 pte_clear(mm, addr, ptep);
 323 }
 324
 325 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
 326                               unsigned long addr, pte_t *ptep)
 327 {
 328         int ncontig;
 329         size_t pgsize;
 330         pte_t orig_pte = huge_ptep_get(ptep);
 331
 332         if (!pte_cont(orig_pte))
 333                 return ptep_get_and_clear(mm, addr, ptep);
 334
 335         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 336
 337         return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 338 }
 339
 340 /*
 341  * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
 342  * and write permission.
 343  *
 344  * For a contiguous huge pte range we need to check whether or not write
 345  * permission has to change only on the first pte in the set. Then for
 346  * all the contiguous ptes we need to check whether or not there is a
 347  * discrepancy between dirty or young.
 348  */
 349 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
 350 {
 351         int i;
 352
 353         if (pte_write(pte) != pte_write(huge_ptep_get(ptep)))
 354                 return 1;
 355
 356         for (i = 0; i < ncontig; i++) {
 357                 pte_t orig_pte = huge_ptep_get(ptep + i);
 358
 359                 if (pte_dirty(pte) != pte_dirty(orig_pte))
 360                         return 1;
 361
 362                 if (pte_young(pte) != pte_young(orig_pte))
 363                         return 1;
 364         }
 365
 366         return 0;
 367 }
 368
 369 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
 370                                unsigned long addr, pte_t *ptep,
 371                                pte_t pte, int dirty)
 372 {
 373         int ncontig, i;
 374         size_t pgsize = 0;
 375         unsigned long pfn = pte_pfn(pte), dpfn;
 376         pgprot_t hugeprot;
 377         pte_t orig_pte;
 378
 379         if (!pte_cont(pte))
 380                 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
 381
 382         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 383         dpfn = pgsize >> PAGE_SHIFT;
 384
 385         if (!__cont_access_flags_changed(ptep, pte, ncontig))
 386                 return 0;
 387
 388         orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 389
 390         /* Make sure we don't lose the dirty or young state */
 391         if (pte_dirty(orig_pte))
 392                 pte = pte_mkdirty(pte);
 393
 394         if (pte_young(orig_pte))
 395                 pte = pte_mkyoung(pte);
 396
 397         hugeprot = pte_pgprot(pte);
 398         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 399                 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
 400
 401         return 1;
 402 }
 403
 404 void huge_ptep_set_wrprotect(struct mm_struct *mm,
 405                              unsigned long addr, pte_t *ptep)
 406 {
 407         unsigned long pfn, dpfn;
 408         pgprot_t hugeprot;
 409         int ncontig, i;
 410         size_t pgsize;
 411         pte_t pte;
 412
 413         if (!pte_cont(READ_ONCE(*ptep))) {
 414                 ptep_set_wrprotect(mm, addr, ptep);
 415                 return;
 416         }
 417
 418         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 419         dpfn = pgsize >> PAGE_SHIFT;
 420
 421         pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 422         pte = pte_wrprotect(pte);
 423
 424         hugeprot = pte_pgprot(pte);
 425         pfn = pte_pfn(pte);
 426
 427         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 428                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 429 }
 430
 431 void huge_ptep_clear_flush(struct vm_area_struct *vma,
 432                            unsigned long addr, pte_t *ptep)
 433 {
 434         size_t pgsize;
 435         int ncontig;
 436
 437         if (!pte_cont(READ_ONCE(*ptep))) {
 438                 ptep_clear_flush(vma, addr, ptep);
 439                 return;
 440         }
 441
 442         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 443         clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 444 }
 445
 446 static void __init add_huge_page_size(unsigned long size)
 447 {
 448         if (size_to_hstate(size))
 449                 return;
 450
 451         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
 452 }
 453
 454 static int __init hugetlbpage_init(void)
 455 {
 456 #ifdef CONFIG_ARM64_4K_PAGES
 457         add_huge_page_size(PUD_SIZE);
 458 #endif
 459         add_huge_page_size(CONT_PMD_SIZE);
 460         add_huge_page_size(PMD_SIZE);
 461         add_huge_page_size(CONT_PTE_SIZE);
 462
 463         return 0;
 464 }
 465 arch_initcall(hugetlbpage_init);
 466
 467 static __init int setup_hugepagesz(char *opt)
 468 {
 469         unsigned long ps = memparse(opt, &opt);
 470
 471         switch (ps) {
 472 #ifdef CONFIG_ARM64_4K_PAGES
 473         case PUD_SIZE:
 474 #endif
 475         case CONT_PMD_SIZE:
 476         case PMD_SIZE:
 477         case CONT_PTE_SIZE:
 478                 add_huge_page_size(ps);
 479                 return 1;
 480         }
 481
 482         hugetlb_bad_size();
 483         pr_err("hugepagesz: Unsupported page size %lu K\n", ps >> 10);
 484         return 0;
 485 }
 486 __setup("hugepagesz=", setup_hugepagesz);