arch/arm64/mm/mmu.c

   1 /*
   2  * Based on arch/arm/mm/mmu.c
   3  *
   4  * Copyright (C) 1995-2005 Russell King
   5  * Copyright (C) 2012 ARM Ltd.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include <linux/export.h>
  21 #include <linux/kernel.h>
  22 #include <linux/errno.h>
  23 #include <linux/init.h>
  24 #include <linux/mman.h>
  25 #include <linux/nodemask.h>
  26 #include <linux/memblock.h>
  27 #include <linux/fs.h>
  28 #include <linux/io.h>
  29 #include <linux/slab.h>
  30 #include <linux/stop_machine.h>
  31
  32 #include <asm/cputype.h>
  33 #include <asm/fixmap.h>
  34 #include <asm/sections.h>
  35 #include <asm/setup.h>
  36 #include <asm/sizes.h>
  37 #include <asm/tlb.h>
  38 #include <asm/memblock.h>
  39 #include <asm/mmu_context.h>
  40
  41 #include "mm.h"
  42
  43 /*
  44  * Empty_zero_page is a special page that is used for zero-initialized data
  45  * and COW.
  46  */
  47 struct page *empty_zero_page;
  48 EXPORT_SYMBOL(empty_zero_page);
  49
  50 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  51                               unsigned long size, pgprot_t vma_prot)
  52 {
  53         if (!pfn_valid(pfn))
  54                 return pgprot_noncached(vma_prot);
  55         else if (file->f_flags & O_SYNC)
  56                 return pgprot_writecombine(vma_prot);
  57         return vma_prot;
  58 }
  59 EXPORT_SYMBOL(phys_mem_access_prot);
  60
  61 static void __init *early_alloc(unsigned long sz)
  62 {
  63         void *ptr = __va(memblock_alloc(sz, sz));
  64         BUG_ON(!ptr);
  65         memset(ptr, 0, sz);
  66         return ptr;
  67 }
  68
  69 /*
  70  * remap a PMD into pages
  71  */
  72 static void split_pmd(pmd_t *pmd, pte_t *pte)
  73 {
  74         unsigned long pfn = pmd_pfn(*pmd);
  75         int i = 0;
  76
  77         do {
  78                 /*
  79                  * Need to have the least restrictive permissions available
  80                  * permissions will be fixed up later
  81                  */
  82                 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
  83                 pfn++;
  84         } while (pte++, i++, i < PTRS_PER_PTE);
  85 }
  86
  87 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
  88                                   unsigned long end, unsigned long pfn,
  89                                   pgprot_t prot,
  90                                   void *(*alloc)(unsigned long size))
  91 {
  92         pte_t *pte;
  93
  94         if (pmd_none(*pmd) || pmd_sect(*pmd)) {
  95                 pte = alloc(PTRS_PER_PTE * sizeof(pte_t));
  96                 if (pmd_sect(*pmd))
  97                         split_pmd(pmd, pte);
  98                 __pmd_populate(pmd, __pa(pte), PMD_TYPE_TABLE);
  99                 flush_tlb_all();
 100         }
 101         BUG_ON(pmd_bad(*pmd));
 102
 103         pte = pte_offset_kernel(pmd, addr);
 104         do {
 105                 set_pte(pte, pfn_pte(pfn, prot));
 106                 pfn++;
 107         } while (pte++, addr += PAGE_SIZE, addr != end);
 108 }
 109
 110 void split_pud(pud_t *old_pud, pmd_t *pmd)
 111 {
 112         unsigned long addr = pud_pfn(*old_pud) << PAGE_SHIFT;
 113         pgprot_t prot = __pgprot(pud_val(*old_pud) ^ addr);
 114         int i = 0;
 115
 116         do {
 117                 set_pmd(pmd, __pmd(addr | prot));
 118                 addr += PMD_SIZE;
 119         } while (pmd++, i++, i < PTRS_PER_PMD);
 120 }
 121
 122 static void alloc_init_pmd(struct mm_struct *mm, pud_t *pud,
 123                                   unsigned long addr, unsigned long end,
 124                                   phys_addr_t phys, pgprot_t prot,
 125                                   void *(*alloc)(unsigned long size))
 126 {
 127         pmd_t *pmd;
 128         unsigned long next;
 129
 130         /*
 131          * Check for initial section mappings in the pgd/pud and remove them.
 132          */
 133         if (pud_none(*pud) || pud_sect(*pud)) {
 134                 pmd = alloc(PTRS_PER_PMD * sizeof(pmd_t));
 135                 if (pud_sect(*pud)) {
 136                         /*
 137                          * need to have the 1G of mappings continue to be
 138                          * present
 139                          */
 140                         split_pud(pud, pmd);
 141                 }
 142                 pud_populate(mm, pud, pmd);
 143                 flush_tlb_all();
 144         }
 145         BUG_ON(pud_bad(*pud));
 146
 147         pmd = pmd_offset(pud, addr);
 148         do {
 149                 next = pmd_addr_end(addr, end);
 150                 /* try section mapping first */
 151                 if (((addr | next | phys) & ~SECTION_MASK) == 0) {
 152                         pmd_t old_pmd =*pmd;
 153                         set_pmd(pmd, __pmd(phys |
 154                                            pgprot_val(mk_sect_prot(prot))));
 155                         /*
 156                          * Check for previous table entries created during
 157                          * boot (__create_page_tables) and flush them.
 158                          */
 159                         if (!pmd_none(old_pmd)) {
 160                                 flush_tlb_all();
 161                                 if (pmd_table(old_pmd)) {
 162                                         phys_addr_t table = __pa(pte_offset_map(&old_pmd, 0));
 163                                         if (!WARN_ON_ONCE(slab_is_available()))
 164                                                 memblock_free(table, PAGE_SIZE);
 165                                 }
 166                         }
 167                 } else {
 168                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 169                                        prot, alloc);
 170                 }
 171                 phys += next - addr;
 172         } while (pmd++, addr = next, addr != end);
 173 }
 174
 175 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 176                         unsigned long phys)
 177 {
 178         if (PAGE_SHIFT != 12)
 179                 return false;
 180
 181         if (((addr | next | phys) & ~PUD_MASK) != 0)
 182                 return false;
 183
 184         return true;
 185 }
 186
 187 static void alloc_init_pud(struct mm_struct *mm, pgd_t *pgd,
 188                                   unsigned long addr, unsigned long end,
 189                                   phys_addr_t phys, pgprot_t prot,
 190                                   void *(*alloc)(unsigned long size))
 191 {
 192         pud_t *pud;
 193         unsigned long next;
 194
 195         if (pgd_none(*pgd)) {
 196                 pud = alloc(PTRS_PER_PUD * sizeof(pud_t));
 197                 pgd_populate(mm, pgd, pud);
 198         }
 199         BUG_ON(pgd_bad(*pgd));
 200
 201         pud = pud_offset(pgd, addr);
 202         do {
 203                 next = pud_addr_end(addr, end);
 204
 205                 /*
 206                  * For 4K granule only, attempt to put down a 1GB block
 207                  */
 208                 if (use_1G_block(addr, next, phys)) {
 209                         pud_t old_pud = *pud;
 210                         set_pud(pud, __pud(phys |
 211                                            pgprot_val(mk_sect_prot(prot))));
 212
 213                         /*
 214                          * If we have an old value for a pud, it will
 215                          * be pointing to a pmd table that we no longer
 216                          * need (from swapper_pg_dir).
 217                          *
 218                          * Look up the old pmd table and free it.
 219                          */
 220                         if (!pud_none(old_pud)) {
 221                                 flush_tlb_all();
 222                                 if (pud_table(old_pud)) {
 223                                         phys_addr_t table = __pa(pmd_offset(&old_pud, 0));
 224                                         if (!WARN_ON_ONCE(slab_is_available()))
 225                                                 memblock_free(table, PAGE_SIZE);
 226                                 }
 227                         }
 228                 } else {
 229                         alloc_init_pmd(mm, pud, addr, next, phys, prot, alloc);
 230                 }
 231                 phys += next - addr;
 232         } while (pud++, addr = next, addr != end);
 233 }
 234
 235 /*
 236  * Create the page directory entries and any necessary page tables for the
 237  * mapping specified by 'md'.
 238  */
 239 static void  __create_mapping(struct mm_struct *mm, pgd_t *pgd,
 240                                     phys_addr_t phys, unsigned long virt,
 241                                     phys_addr_t size, pgprot_t prot,
 242                                     void *(*alloc)(unsigned long size))
 243 {
 244         unsigned long addr, length, end, next;
 245
 246         addr = virt & PAGE_MASK;
 247         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 248
 249         end = addr + length;
 250         do {
 251                 next = pgd_addr_end(addr, end);
 252                 alloc_init_pud(mm, pgd, addr, next, phys, prot, alloc);
 253                 phys += next - addr;
 254         } while (pgd++, addr = next, addr != end);
 255 }
 256
 257 static void *late_alloc(unsigned long size)
 258 {
 259         void *ptr;
 260
 261         BUG_ON(size > PAGE_SIZE);
 262         ptr = (void *)__get_free_page(PGALLOC_GFP);
 263         BUG_ON(!ptr);
 264         return ptr;
 265 }
 266
 267 static void __ref create_mapping(phys_addr_t phys, unsigned long virt,
 268                                   phys_addr_t size, pgprot_t prot)
 269 {
 270         if (virt < VMALLOC_START) {
 271                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 272                         &phys, virt);
 273                 return;
 274         }
 275         __create_mapping(&init_mm, pgd_offset_k(virt & PAGE_MASK), phys, virt,
 276                          size, prot, early_alloc);
 277 }
 278
 279 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 280                                unsigned long virt, phys_addr_t size,
 281                                pgprot_t prot)
 282 {
 283         __create_mapping(mm, pgd_offset(mm, virt), phys, virt, size, prot,
 284                                 late_alloc);
 285 }
 286
 287 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 288                                   phys_addr_t size, pgprot_t prot)
 289 {
 290         if (virt < VMALLOC_START) {
 291                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 292                         &phys, virt);
 293                 return;
 294         }
 295
 296         return __create_mapping(&init_mm, pgd_offset_k(virt & PAGE_MASK),
 297                                 phys, virt, size, prot, late_alloc);
 298 }
 299
 300 #ifdef CONFIG_DEBUG_RODATA
 301 static void __init __map_memblock(phys_addr_t start, phys_addr_t end)
 302 {
 303         /*
 304          * Set up the executable regions using the existing section mappings
 305          * for now. This will get more fine grained later once all memory
 306          * is mapped
 307          */
 308         unsigned long kernel_x_start = round_down(__pa(_stext), SECTION_SIZE);
 309         unsigned long kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
 310
 311         if (end < kernel_x_start) {
 312                 create_mapping(start, __phys_to_virt(start),
 313                         end - start, PAGE_KERNEL);
 314         } else if (start >= kernel_x_end) {
 315                 create_mapping(start, __phys_to_virt(start),
 316                         end - start, PAGE_KERNEL);
 317         } else {
 318                 if (start < kernel_x_start)
 319                         create_mapping(start, __phys_to_virt(start),
 320                                 kernel_x_start - start,
 321                                 PAGE_KERNEL);
 322                 create_mapping(kernel_x_start,
 323                                 __phys_to_virt(kernel_x_start),
 324                                 kernel_x_end - kernel_x_start,
 325                                 PAGE_KERNEL_EXEC);
 326                 if (kernel_x_end < end)
 327                         create_mapping(kernel_x_end,
 328                                 __phys_to_virt(kernel_x_end),
 329                                 end - kernel_x_end,
 330                                 PAGE_KERNEL);
 331         }
 332
 333 }
 334 #else
 335 static void __init __map_memblock(phys_addr_t start, phys_addr_t end)
 336 {
 337         create_mapping(start, __phys_to_virt(start), end - start,
 338                         PAGE_KERNEL_EXEC);
 339 }
 340 #endif
 341
 342 static void __init map_mem(void)
 343 {
 344         struct memblock_region *reg;
 345         phys_addr_t limit;
 346
 347         /*
 348          * Temporarily limit the memblock range. We need to do this as
 349          * create_mapping requires puds, pmds and ptes to be allocated from
 350          * memory addressable from the initial direct kernel mapping.
 351          *
 352          * The initial direct kernel mapping, located at swapper_pg_dir, gives
 353          * us PUD_SIZE (4K pages) or PMD_SIZE (64K pages) memory starting from
 354          * PHYS_OFFSET (which must be aligned to 2MB as per
 355          * Documentation/arm64/booting.txt).
 356          */
 357         if (IS_ENABLED(CONFIG_ARM64_64K_PAGES))
 358                 limit = PHYS_OFFSET + PMD_SIZE;
 359         else
 360                 limit = PHYS_OFFSET + PUD_SIZE;
 361         memblock_set_current_limit(limit);
 362
 363         /* map all the memory banks */
 364         for_each_memblock(memory, reg) {
 365                 phys_addr_t start = reg->base;
 366                 phys_addr_t end = start + reg->size;
 367
 368                 if (start >= end)
 369                         break;
 370
 371 #ifndef CONFIG_ARM64_64K_PAGES
 372                 /*
 373                  * For the first memory bank align the start address and
 374                  * current memblock limit to prevent create_mapping() from
 375                  * allocating pte page tables from unmapped memory.
 376                  * When 64K pages are enabled, the pte page table for the
 377                  * first PGDIR_SIZE is already present in swapper_pg_dir.
 378                  */
 379                 if (start < limit)
 380                         start = ALIGN(start, PMD_SIZE);
 381                 if (end < limit) {
 382                         limit = end & PMD_MASK;
 383                         memblock_set_current_limit(limit);
 384                 }
 385 #endif
 386                 __map_memblock(start, end);
 387         }
 388
 389         /* Limit no longer required. */
 390         memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
 391 }
 392
 393 void __init fixup_executable(void)
 394 {
 395 #ifdef CONFIG_DEBUG_RODATA
 396         /* now that we are actually fully mapped, make the start/end more fine grained */
 397         if (!IS_ALIGNED((unsigned long)_stext, SECTION_SIZE)) {
 398                 unsigned long aligned_start = round_down(__pa(_stext),
 399                                                         SECTION_SIZE);
 400
 401                 create_mapping(aligned_start, __phys_to_virt(aligned_start),
 402                                 __pa(_stext) - aligned_start,
 403                                 PAGE_KERNEL);
 404         }
 405
 406         if (!IS_ALIGNED((unsigned long)__init_end, SECTION_SIZE)) {
 407                 unsigned long aligned_end = round_up(__pa(__init_end),
 408                                                         SECTION_SIZE);
 409                 create_mapping(__pa(__init_end), (unsigned long)__init_end,
 410                                 aligned_end - __pa(__init_end),
 411                                 PAGE_KERNEL);
 412         }
 413 #endif
 414 }
 415
 416 #ifdef CONFIG_DEBUG_RODATA
 417 void mark_rodata_ro(void)
 418 {
 419         create_mapping_late(__pa(_stext), (unsigned long)_stext,
 420                                 (unsigned long)_etext - (unsigned long)_stext,
 421                                 PAGE_KERNEL_EXEC | PTE_RDONLY);
 422
 423 }
 424 #endif
 425
 426 void fixup_init(void)
 427 {
 428         create_mapping_late(__pa(__init_begin), (unsigned long)__init_begin,
 429                         (unsigned long)__init_end - (unsigned long)__init_begin,
 430                         PAGE_KERNEL);
 431 }
 432
 433 /*
 434  * paging_init() sets up the page tables, initialises the zone memory
 435  * maps and sets up the zero page.
 436  */
 437 void __init paging_init(void)
 438 {
 439         void *zero_page;
 440
 441         map_mem();
 442         fixup_executable();
 443
 444         /* allocate the zero page. */
 445         zero_page = early_alloc(PAGE_SIZE);
 446
 447         bootmem_init();
 448
 449         empty_zero_page = virt_to_page(zero_page);
 450
 451         /*
 452          * TTBR0 is only used for the identity mapping at this stage. Make it
 453          * point to zero page to avoid speculatively fetching new entries.
 454          */
 455         cpu_set_reserved_ttbr0();
 456         flush_tlb_all();
 457 }
 458
 459 /*
 460  * Enable the identity mapping to allow the MMU disabling.
 461  */
 462 void setup_mm_for_reboot(void)
 463 {
 464         cpu_switch_mm(idmap_pg_dir, &init_mm);
 465         flush_tlb_all();
 466 }
 467
 468 /*
 469  * Check whether a kernel address is valid (derived from arch/x86/).
 470  */
 471 int kern_addr_valid(unsigned long addr)
 472 {
 473         pgd_t *pgd;
 474         pud_t *pud;
 475         pmd_t *pmd;
 476         pte_t *pte;
 477
 478         if ((((long)addr) >> VA_BITS) != -1UL)
 479                 return 0;
 480
 481         pgd = pgd_offset_k(addr);
 482         if (pgd_none(*pgd))
 483                 return 0;
 484
 485         pud = pud_offset(pgd, addr);
 486         if (pud_none(*pud))
 487                 return 0;
 488
 489         if (pud_sect(*pud))
 490                 return pfn_valid(pud_pfn(*pud));
 491
 492         pmd = pmd_offset(pud, addr);
 493         if (pmd_none(*pmd))
 494                 return 0;
 495
 496         if (pmd_sect(*pmd))
 497                 return pfn_valid(pmd_pfn(*pmd));
 498
 499         pte = pte_offset_kernel(pmd, addr);
 500         if (pte_none(*pte))
 501                 return 0;
 502
 503         return pfn_valid(pte_pfn(*pte));
 504 }
 505 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 506 #ifdef CONFIG_ARM64_64K_PAGES
 507 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 508 {
 509         return vmemmap_populate_basepages(start, end, node);
 510 }
 511 #else   /* !CONFIG_ARM64_64K_PAGES */
 512 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 513 {
 514         unsigned long addr = start;
 515         unsigned long next;
 516         pgd_t *pgd;
 517         pud_t *pud;
 518         pmd_t *pmd;
 519
 520         do {
 521                 next = pmd_addr_end(addr, end);
 522
 523                 pgd = vmemmap_pgd_populate(addr, node);
 524                 if (!pgd)
 525                         return -ENOMEM;
 526
 527                 pud = vmemmap_pud_populate(pgd, addr, node);
 528                 if (!pud)
 529                         return -ENOMEM;
 530
 531                 pmd = pmd_offset(pud, addr);
 532                 if (pmd_none(*pmd)) {
 533                         void *p = NULL;
 534
 535                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 536                         if (!p)
 537                                 return -ENOMEM;
 538
 539                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 540                 } else
 541                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 542         } while (addr = next, addr != end);
 543
 544         return 0;
 545 }
 546 #endif  /* CONFIG_ARM64_64K_PAGES */
 547 void vmemmap_free(unsigned long start, unsigned long end)
 548 {
 549 }
 550 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 551
 552 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
 553 #if CONFIG_PGTABLE_LEVELS > 2
 554 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss;
 555 #endif
 556 #if CONFIG_PGTABLE_LEVELS > 3
 557 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss;
 558 #endif
 559
 560 static inline pud_t * fixmap_pud(unsigned long addr)
 561 {
 562         pgd_t *pgd = pgd_offset_k(addr);
 563
 564         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 565
 566         return pud_offset(pgd, addr);
 567 }
 568
 569 static inline pmd_t * fixmap_pmd(unsigned long addr)
 570 {
 571         pud_t *pud = fixmap_pud(addr);
 572
 573         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 574
 575         return pmd_offset(pud, addr);
 576 }
 577
 578 static inline pte_t * fixmap_pte(unsigned long addr)
 579 {
 580         pmd_t *pmd = fixmap_pmd(addr);
 581
 582         BUG_ON(pmd_none(*pmd) || pmd_bad(*pmd));
 583
 584         return pte_offset_kernel(pmd, addr);
 585 }
 586
 587 void __init early_fixmap_init(void)
 588 {
 589         pgd_t *pgd;
 590         pud_t *pud;
 591         pmd_t *pmd;
 592         unsigned long addr = FIXADDR_START;
 593
 594         pgd = pgd_offset_k(addr);
 595         pgd_populate(&init_mm, pgd, bm_pud);
 596         pud = pud_offset(pgd, addr);
 597         pud_populate(&init_mm, pud, bm_pmd);
 598         pmd = pmd_offset(pud, addr);
 599         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 600
 601         /*
 602          * The boot-ioremap range spans multiple pmds, for which
 603          * we are not preparted:
 604          */
 605         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 606                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 607
 608         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 609              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 610                 WARN_ON(1);
 611                 pr_warn("pmd %p != %p, %p\n",
 612                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 613                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 614                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 615                         fix_to_virt(FIX_BTMAP_BEGIN));
 616                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 617                         fix_to_virt(FIX_BTMAP_END));
 618
 619                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 620                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 621         }
 622 }
 623
 624 void __set_fixmap(enum fixed_addresses idx,
 625                                phys_addr_t phys, pgprot_t flags)
 626 {
 627         unsigned long addr = __fix_to_virt(idx);
 628         pte_t *pte;
 629
 630         if (idx >= __end_of_fixed_addresses) {
 631                 BUG();
 632                 return;
 633         }
 634
 635         pte = fixmap_pte(addr);
 636
 637         if (pgprot_val(flags)) {
 638                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 639         } else {
 640                 pte_clear(&init_mm, addr, pte);
 641                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 642         }
 643 }