arch/arm64/mm/hugetlbpage.c

   1 /*
   2  * arch/arm64/mm/hugetlbpage.c
   3  *
   4  * Copyright (C) 2013 Linaro Ltd.
   5  *
   6  * Based on arch/x86/mm/hugetlbpage.c.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 as
  10  * published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/fs.h>
  20 #include <linux/mm.h>
  21 #include <linux/hugetlb.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/err.h>
  24 #include <linux/sysctl.h>
  25 #include <asm/mman.h>
  26 #include <asm/tlb.h>
  27 #include <asm/tlbflush.h>
  28 #include <asm/pgalloc.h>
  29
  30 int pmd_huge(pmd_t pmd)
  31 {
  32         return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
  33 }
  34
  35 int pud_huge(pud_t pud)
  36 {
  37 #ifndef __PAGETABLE_PMD_FOLDED
  38         return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
  39 #else
  40         return 0;
  41 #endif
  42 }
  43
  44 /*
  45  * Select all bits except the pfn
  46  */
  47 static inline pgprot_t pte_pgprot(pte_t pte)
  48 {
  49         unsigned long pfn = pte_pfn(pte);
  50
  51         return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
  52 }
  53
  54 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
  55                            pte_t *ptep, size_t *pgsize)
  56 {
  57         pgd_t *pgd = pgd_offset(mm, addr);
  58         pud_t *pud;
  59         pmd_t *pmd;
  60
  61         *pgsize = PAGE_SIZE;
  62         pud = pud_offset(pgd, addr);
  63         pmd = pmd_offset(pud, addr);
  64         if ((pte_t *)pmd == ptep) {
  65                 *pgsize = PMD_SIZE;
  66                 return CONT_PMDS;
  67         }
  68         return CONT_PTES;
  69 }
  70
  71 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
  72 {
  73         int contig_ptes = 0;
  74
  75         *pgsize = size;
  76
  77         switch (size) {
  78 #ifdef CONFIG_ARM64_4K_PAGES
  79         case PUD_SIZE:
  80 #endif
  81         case PMD_SIZE:
  82                 contig_ptes = 1;
  83                 break;
  84         case CONT_PMD_SIZE:
  85                 *pgsize = PMD_SIZE;
  86                 contig_ptes = CONT_PMDS;
  87                 break;
  88         case CONT_PTE_SIZE:
  89                 *pgsize = PAGE_SIZE;
  90                 contig_ptes = CONT_PTES;
  91                 break;
  92         }
  93
  94         return contig_ptes;
  95 }
  96
  97 /*
  98  * Changing some bits of contiguous entries requires us to follow a
  99  * Break-Before-Make approach, breaking the whole contiguous set
 100  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 101  * "Misprogramming of the Contiguous bit", page D4-1762.
 102  *
 103  * This helper performs the break step.
 104  */
 105 static pte_t get_clear_flush(struct mm_struct *mm,
 106                              unsigned long addr,
 107                              pte_t *ptep,
 108                              unsigned long pgsize,
 109                              unsigned long ncontig)
 110 {
 111         struct vm_area_struct vma = { .vm_mm = mm };
 112         pte_t orig_pte = huge_ptep_get(ptep);
 113         bool valid = pte_valid(orig_pte);
 114         unsigned long i, saddr = addr;
 115
 116         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
 117                 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
 118
 119                 /*
 120                  * If HW_AFDBM is enabled, then the HW could turn on
 121                  * the dirty or accessed bit for any page in the set,
 122                  * so check them all.
 123                  */
 124                 if (pte_dirty(pte))
 125                         orig_pte = pte_mkdirty(orig_pte);
 126
 127                 if (pte_young(pte))
 128                         orig_pte = pte_mkyoung(orig_pte);
 129         }
 130
 131         if (valid)
 132                 flush_tlb_range(&vma, saddr, addr);
 133         return orig_pte;
 134 }
 135
 136 /*
 137  * Changing some bits of contiguous entries requires us to follow a
 138  * Break-Before-Make approach, breaking the whole contiguous set
 139  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
 140  * "Misprogramming of the Contiguous bit", page D4-1762.
 141  *
 142  * This helper performs the break step for use cases where the
 143  * original pte is not needed.
 144  */
 145 static void clear_flush(struct mm_struct *mm,
 146                              unsigned long addr,
 147                              pte_t *ptep,
 148                              unsigned long pgsize,
 149                              unsigned long ncontig)
 150 {
 151         struct vm_area_struct vma = { .vm_mm = mm };
 152         unsigned long i, saddr = addr;
 153
 154         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 155                 pte_clear(mm, addr, ptep);
 156
 157         flush_tlb_range(&vma, saddr, addr);
 158 }
 159
 160 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 161                             pte_t *ptep, pte_t pte)
 162 {
 163         size_t pgsize;
 164         int i;
 165         int ncontig;
 166         unsigned long pfn, dpfn;
 167         pgprot_t hugeprot;
 168
 169         /*
 170          * Code needs to be expanded to handle huge swap and migration
 171          * entries. Needed for HUGETLB and MEMORY_FAILURE.
 172          */
 173         WARN_ON(!pte_present(pte));
 174
 175         if (!pte_cont(pte)) {
 176                 set_pte_at(mm, addr, ptep, pte);
 177                 return;
 178         }
 179
 180         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 181         pfn = pte_pfn(pte);
 182         dpfn = pgsize >> PAGE_SHIFT;
 183         hugeprot = pte_pgprot(pte);
 184
 185         clear_flush(mm, addr, ptep, pgsize, ncontig);
 186
 187         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn) {
 188                 pr_debug("%s: set pte %p to 0x%llx\n", __func__, ptep,
 189                          pte_val(pfn_pte(pfn, hugeprot)));
 190                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 191         }
 192 }
 193
 194 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
 195                           pte_t *ptep, pte_t pte, unsigned long sz)
 196 {
 197         int i, ncontig;
 198         size_t pgsize;
 199
 200         ncontig = num_contig_ptes(sz, &pgsize);
 201
 202         for (i = 0; i < ncontig; i++, ptep++)
 203                 set_pte(ptep, pte);
 204 }
 205
 206 pte_t *huge_pte_alloc(struct mm_struct *mm,
 207                       unsigned long addr, unsigned long sz)
 208 {
 209         pgd_t *pgd;
 210         pud_t *pud;
 211         pte_t *pte = NULL;
 212
 213         pr_debug("%s: addr:0x%lx sz:0x%lx\n", __func__, addr, sz);
 214         pgd = pgd_offset(mm, addr);
 215         pud = pud_alloc(mm, pgd, addr);
 216         if (!pud)
 217                 return NULL;
 218
 219         if (sz == PUD_SIZE) {
 220                 pte = (pte_t *)pud;
 221         } else if (sz == (PAGE_SIZE * CONT_PTES)) {
 222                 pmd_t *pmd = pmd_alloc(mm, pud, addr);
 223
 224                 WARN_ON(addr & (sz - 1));
 225                 /*
 226                  * Note that if this code were ever ported to the
 227                  * 32-bit arm platform then it will cause trouble in
 228                  * the case where CONFIG_HIGHPTE is set, since there
 229                  * will be no pte_unmap() to correspond with this
 230                  * pte_alloc_map().
 231                  */
 232                 pte = pte_alloc_map(mm, pmd, addr);
 233         } else if (sz == PMD_SIZE) {
 234                 if (IS_ENABLED(CONFIG_ARCH_WANT_HUGE_PMD_SHARE) &&
 235                     pud_none(*pud))
 236                         pte = huge_pmd_share(mm, addr, pud);
 237                 else
 238                         pte = (pte_t *)pmd_alloc(mm, pud, addr);
 239         } else if (sz == (PMD_SIZE * CONT_PMDS)) {
 240                 pmd_t *pmd;
 241
 242                 pmd = pmd_alloc(mm, pud, addr);
 243                 WARN_ON(addr & (sz - 1));
 244                 return (pte_t *)pmd;
 245         }
 246
 247         pr_debug("%s: addr:0x%lx sz:0x%lx ret pte=%p/0x%llx\n", __func__, addr,
 248                sz, pte, pte_val(*pte));
 249         return pte;
 250 }
 251
 252 pte_t *huge_pte_offset(struct mm_struct *mm,
 253                        unsigned long addr, unsigned long sz)
 254 {
 255         pgd_t *pgd;
 256         pud_t *pud;
 257         pmd_t *pmd;
 258
 259         pgd = pgd_offset(mm, addr);
 260         pr_debug("%s: addr:0x%lx pgd:%p\n", __func__, addr, pgd);
 261         if (!pgd_present(*pgd))
 262                 return NULL;
 263
 264         pud = pud_offset(pgd, addr);
 265         if (sz != PUD_SIZE && pud_none(*pud))
 266                 return NULL;
 267         /* hugepage or swap? */
 268         if (pud_huge(*pud) || !pud_present(*pud))
 269                 return (pte_t *)pud;
 270         /* table; check the next level */
 271
 272         if (sz == CONT_PMD_SIZE)
 273                 addr &= CONT_PMD_MASK;
 274
 275         pmd = pmd_offset(pud, addr);
 276         if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
 277             pmd_none(*pmd))
 278                 return NULL;
 279         if (pmd_huge(*pmd) || !pmd_present(*pmd))
 280                 return (pte_t *)pmd;
 281
 282         if (sz == CONT_PTE_SIZE) {
 283                 pte_t *pte = pte_offset_kernel(pmd, (addr & CONT_PTE_MASK));
 284                 return pte;
 285         }
 286
 287         return NULL;
 288 }
 289
 290 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
 291                          struct page *page, int writable)
 292 {
 293         size_t pagesize = huge_page_size(hstate_vma(vma));
 294
 295         if (pagesize == CONT_PTE_SIZE) {
 296                 entry = pte_mkcont(entry);
 297         } else if (pagesize == CONT_PMD_SIZE) {
 298                 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
 299         } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
 300                 pr_warn("%s: unrecognized huge page size 0x%lx\n",
 301                         __func__, pagesize);
 302         }
 303         return entry;
 304 }
 305
 306 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
 307                     pte_t *ptep, unsigned long sz)
 308 {
 309         int i, ncontig;
 310         size_t pgsize;
 311
 312         ncontig = num_contig_ptes(sz, &pgsize);
 313
 314         for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
 315                 pte_clear(mm, addr, ptep);
 316 }
 317
 318 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
 319                               unsigned long addr, pte_t *ptep)
 320 {
 321         int ncontig;
 322         size_t pgsize;
 323         pte_t orig_pte = huge_ptep_get(ptep);
 324
 325         if (!pte_cont(orig_pte))
 326                 return ptep_get_and_clear(mm, addr, ptep);
 327
 328         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 329
 330         return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 331 }
 332
 333 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
 334                                unsigned long addr, pte_t *ptep,
 335                                pte_t pte, int dirty)
 336 {
 337         int ncontig, i, changed = 0;
 338         size_t pgsize = 0;
 339         unsigned long pfn = pte_pfn(pte), dpfn;
 340         pgprot_t hugeprot;
 341         pte_t orig_pte;
 342
 343         if (!pte_cont(pte))
 344                 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
 345
 346         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 347         dpfn = pgsize >> PAGE_SHIFT;
 348
 349         orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 350         if (!pte_same(orig_pte, pte))
 351                 changed = 1;
 352
 353         /* Make sure we don't lose the dirty or young state */
 354         if (pte_dirty(orig_pte))
 355                 pte = pte_mkdirty(pte);
 356
 357         if (pte_young(orig_pte))
 358                 pte = pte_mkyoung(pte);
 359
 360         hugeprot = pte_pgprot(pte);
 361         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 362                 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
 363
 364         return changed;
 365 }
 366
 367 void huge_ptep_set_wrprotect(struct mm_struct *mm,
 368                              unsigned long addr, pte_t *ptep)
 369 {
 370         unsigned long pfn, dpfn;
 371         pgprot_t hugeprot;
 372         int ncontig, i;
 373         size_t pgsize;
 374         pte_t pte;
 375
 376         if (!pte_cont(*ptep)) {
 377                 ptep_set_wrprotect(mm, addr, ptep);
 378                 return;
 379         }
 380
 381         ncontig = find_num_contig(mm, addr, ptep, &pgsize);
 382         dpfn = pgsize >> PAGE_SHIFT;
 383
 384         pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
 385         pte = pte_wrprotect(pte);
 386
 387         hugeprot = pte_pgprot(pte);
 388         pfn = pte_pfn(pte);
 389
 390         for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
 391                 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
 392 }
 393
 394 void huge_ptep_clear_flush(struct vm_area_struct *vma,
 395                            unsigned long addr, pte_t *ptep)
 396 {
 397         size_t pgsize;
 398         int ncontig;
 399
 400         if (!pte_cont(*ptep)) {
 401                 ptep_clear_flush(vma, addr, ptep);
 402                 return;
 403         }
 404
 405         ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
 406         clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
 407 }
 408
 409 static __init int setup_hugepagesz(char *opt)
 410 {
 411         unsigned long ps = memparse(opt, &opt);
 412
 413         switch (ps) {
 414 #ifdef CONFIG_ARM64_4K_PAGES
 415         case PUD_SIZE:
 416 #endif
 417         case PMD_SIZE * CONT_PMDS:
 418         case PMD_SIZE:
 419         case PAGE_SIZE * CONT_PTES:
 420                 hugetlb_add_hstate(ilog2(ps) - PAGE_SHIFT);
 421                 return 1;
 422         }
 423
 424         hugetlb_bad_size();
 425         pr_err("hugepagesz: Unsupported page size %lu K\n", ps >> 10);
 426         return 0;
 427 }
 428 __setup("hugepagesz=", setup_hugepagesz);
 429
 430 #ifdef CONFIG_ARM64_64K_PAGES
 431 static __init int add_default_hugepagesz(void)
 432 {
 433         if (size_to_hstate(CONT_PTES * PAGE_SIZE) == NULL)
 434                 hugetlb_add_hstate(CONT_PTE_SHIFT);
 435         return 0;
 436 }
 437 arch_initcall(add_default_hugepagesz);
 438 #endif