1 /* 2 * linux/arch/x86_64/mm/init.c 3 * 4 * Copyright (C) 1995 Linus Torvalds 5 * Copyright (C) 2000 Pavel Machek <pavel@ucw.cz> 6 * Copyright (C) 2002,2003 Andi Kleen <ak@suse.de> 7 */ 8 9 #include <linux/signal.h> 10 #include <linux/sched.h> 11 #include <linux/kernel.h> 12 #include <linux/errno.h> 13 #include <linux/string.h> 14 #include <linux/types.h> 15 #include <linux/ptrace.h> 16 #include <linux/mman.h> 17 #include <linux/mm.h> 18 #include <linux/swap.h> 19 #include <linux/smp.h> 20 #include <linux/init.h> 21 #include <linux/initrd.h> 22 #include <linux/pagemap.h> 23 #include <linux/bootmem.h> 24 #include <linux/memblock.h> 25 #include <linux/proc_fs.h> 26 #include <linux/pci.h> 27 #include <linux/pfn.h> 28 #include <linux/poison.h> 29 #include <linux/dma-mapping.h> 30 #include <linux/memory.h> 31 #include <linux/memory_hotplug.h> 32 #include <linux/memremap.h> 33 #include <linux/nmi.h> 34 #include <linux/gfp.h> 35 #include <linux/kcore.h> 36 37 #include <asm/processor.h> 38 #include <asm/bios_ebda.h> 39 #include <asm/uaccess.h> 40 #include <asm/pgtable.h> 41 #include <asm/pgalloc.h> 42 #include <asm/dma.h> 43 #include <asm/fixmap.h> 44 #include <asm/e820.h> 45 #include <asm/apic.h> 46 #include <asm/tlb.h> 47 #include <asm/mmu_context.h> 48 #include <asm/proto.h> 49 #include <asm/smp.h> 50 #include <asm/sections.h> 51 #include <asm/kdebug.h> 52 #include <asm/numa.h> 53 #include <asm/cacheflush.h> 54 #include <asm/init.h> 55 #include <asm/uv/uv.h> 56 #include <asm/setup.h> 57 58 #include "mm_internal.h" 59 60 #include "ident_map.c" 61 62 /* 63 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the 64 * physical space so we can cache the place of the first one and move 65 * around without checking the pgd every time. 66 */ 67 68 pteval_t __supported_pte_mask __read_mostly = ~0; 69 EXPORT_SYMBOL_GPL(__supported_pte_mask); 70 71 int force_personality32; 72 73 /* 74 * noexec32=on|off 75 * Control non executable heap for 32bit processes. 76 * To control the stack too use noexec=off 77 * 78 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default) 79 * off PROT_READ implies PROT_EXEC 80 */ 81 static int __init nonx32_setup(char *str) 82 { 83 if (!strcmp(str, "on")) 84 force_personality32 &= ~READ_IMPLIES_EXEC; 85 else if (!strcmp(str, "off")) 86 force_personality32 |= READ_IMPLIES_EXEC; 87 return 1; 88 } 89 __setup("noexec32=", nonx32_setup); 90 91 /* 92 * When memory was added/removed make sure all the processes MM have 93 * suitable PGD entries in the local PGD level page. 94 */ 95 void sync_global_pgds(unsigned long start, unsigned long end, int removed) 96 { 97 unsigned long address; 98 99 for (address = start; address <= end; address += PGDIR_SIZE) { 100 const pgd_t *pgd_ref = pgd_offset_k(address); 101 struct page *page; 102 103 /* 104 * When it is called after memory hot remove, pgd_none() 105 * returns true. In this case (removed == 1), we must clear 106 * the PGD entries in the local PGD level page. 107 */ 108 if (pgd_none(*pgd_ref) && !removed) 109 continue; 110 111 spin_lock(&pgd_lock); 112 list_for_each_entry(page, &pgd_list, lru) { 113 pgd_t *pgd; 114 spinlock_t *pgt_lock; 115 116 pgd = (pgd_t *)page_address(page) + pgd_index(address); 117 /* the pgt_lock only for Xen */ 118 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 119 spin_lock(pgt_lock); 120 121 if (!pgd_none(*pgd_ref) && !pgd_none(*pgd)) 122 BUG_ON(pgd_page_vaddr(*pgd) 123 != pgd_page_vaddr(*pgd_ref)); 124 125 if (removed) { 126 if (pgd_none(*pgd_ref) && !pgd_none(*pgd)) 127 pgd_clear(pgd); 128 } else { 129 if (pgd_none(*pgd)) 130 set_pgd(pgd, *pgd_ref); 131 } 132 133 spin_unlock(pgt_lock); 134 } 135 spin_unlock(&pgd_lock); 136 } 137 } 138 139 /* 140 * NOTE: This function is marked __ref because it calls __init function 141 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0. 142 */ 143 static __ref void *spp_getpage(void) 144 { 145 void *ptr; 146 147 if (after_bootmem) 148 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK); 149 else 150 ptr = alloc_bootmem_pages(PAGE_SIZE); 151 152 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) { 153 panic("set_pte_phys: cannot allocate page data %s\n", 154 after_bootmem ? "after bootmem" : ""); 155 } 156 157 pr_debug("spp_getpage %p\n", ptr); 158 159 return ptr; 160 } 161 162 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr) 163 { 164 if (pgd_none(*pgd)) { 165 pud_t *pud = (pud_t *)spp_getpage(); 166 pgd_populate(&init_mm, pgd, pud); 167 if (pud != pud_offset(pgd, 0)) 168 printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n", 169 pud, pud_offset(pgd, 0)); 170 } 171 return pud_offset(pgd, vaddr); 172 } 173 174 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr) 175 { 176 if (pud_none(*pud)) { 177 pmd_t *pmd = (pmd_t *) spp_getpage(); 178 pud_populate(&init_mm, pud, pmd); 179 if (pmd != pmd_offset(pud, 0)) 180 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n", 181 pmd, pmd_offset(pud, 0)); 182 } 183 return pmd_offset(pud, vaddr); 184 } 185 186 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr) 187 { 188 if (pmd_none(*pmd)) { 189 pte_t *pte = (pte_t *) spp_getpage(); 190 pmd_populate_kernel(&init_mm, pmd, pte); 191 if (pte != pte_offset_kernel(pmd, 0)) 192 printk(KERN_ERR "PAGETABLE BUG #02!\n"); 193 } 194 return pte_offset_kernel(pmd, vaddr); 195 } 196 197 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte) 198 { 199 pud_t *pud; 200 pmd_t *pmd; 201 pte_t *pte; 202 203 pud = pud_page + pud_index(vaddr); 204 pmd = fill_pmd(pud, vaddr); 205 pte = fill_pte(pmd, vaddr); 206 207 set_pte(pte, new_pte); 208 209 /* 210 * It's enough to flush this one mapping. 211 * (PGE mappings get flushed as well) 212 */ 213 __flush_tlb_one(vaddr); 214 } 215 216 void set_pte_vaddr(unsigned long vaddr, pte_t pteval) 217 { 218 pgd_t *pgd; 219 pud_t *pud_page; 220 221 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval)); 222 223 pgd = pgd_offset_k(vaddr); 224 if (pgd_none(*pgd)) { 225 printk(KERN_ERR 226 "PGD FIXMAP MISSING, it should be setup in head.S!\n"); 227 return; 228 } 229 pud_page = (pud_t*)pgd_page_vaddr(*pgd); 230 set_pte_vaddr_pud(pud_page, vaddr, pteval); 231 } 232 233 pmd_t * __init populate_extra_pmd(unsigned long vaddr) 234 { 235 pgd_t *pgd; 236 pud_t *pud; 237 238 pgd = pgd_offset_k(vaddr); 239 pud = fill_pud(pgd, vaddr); 240 return fill_pmd(pud, vaddr); 241 } 242 243 pte_t * __init populate_extra_pte(unsigned long vaddr) 244 { 245 pmd_t *pmd; 246 247 pmd = populate_extra_pmd(vaddr); 248 return fill_pte(pmd, vaddr); 249 } 250 251 /* 252 * Create large page table mappings for a range of physical addresses. 253 */ 254 static void __init __init_extra_mapping(unsigned long phys, unsigned long size, 255 enum page_cache_mode cache) 256 { 257 pgd_t *pgd; 258 pud_t *pud; 259 pmd_t *pmd; 260 pgprot_t prot; 261 262 pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) | 263 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache))); 264 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK)); 265 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) { 266 pgd = pgd_offset_k((unsigned long)__va(phys)); 267 if (pgd_none(*pgd)) { 268 pud = (pud_t *) spp_getpage(); 269 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE | 270 _PAGE_USER)); 271 } 272 pud = pud_offset(pgd, (unsigned long)__va(phys)); 273 if (pud_none(*pud)) { 274 pmd = (pmd_t *) spp_getpage(); 275 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | 276 _PAGE_USER)); 277 } 278 pmd = pmd_offset(pud, phys); 279 BUG_ON(!pmd_none(*pmd)); 280 set_pmd(pmd, __pmd(phys | pgprot_val(prot))); 281 } 282 } 283 284 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size) 285 { 286 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB); 287 } 288 289 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size) 290 { 291 __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC); 292 } 293 294 /* 295 * The head.S code sets up the kernel high mapping: 296 * 297 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text) 298 * 299 * phys_base holds the negative offset to the kernel, which is added 300 * to the compile time generated pmds. This results in invalid pmds up 301 * to the point where we hit the physaddr 0 mapping. 302 * 303 * We limit the mappings to the region from _text to _brk_end. _brk_end 304 * is rounded up to the 2MB boundary. This catches the invalid pmds as 305 * well, as they are located before _text: 306 */ 307 void __init cleanup_highmap(void) 308 { 309 unsigned long vaddr = __START_KERNEL_map; 310 unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE; 311 unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1; 312 pmd_t *pmd = level2_kernel_pgt; 313 314 /* 315 * Native path, max_pfn_mapped is not set yet. 316 * Xen has valid max_pfn_mapped set in 317 * arch/x86/xen/mmu.c:xen_setup_kernel_pagetable(). 318 */ 319 if (max_pfn_mapped) 320 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT); 321 322 for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) { 323 if (pmd_none(*pmd)) 324 continue; 325 if (vaddr < (unsigned long) _text || vaddr > end) 326 set_pmd(pmd, __pmd(0)); 327 } 328 } 329 330 static unsigned long __meminit 331 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end, 332 pgprot_t prot) 333 { 334 unsigned long pages = 0, next; 335 unsigned long last_map_addr = end; 336 int i; 337 338 pte_t *pte = pte_page + pte_index(addr); 339 340 for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) { 341 next = (addr & PAGE_MASK) + PAGE_SIZE; 342 if (addr >= end) { 343 if (!after_bootmem && 344 !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) && 345 !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN)) 346 set_pte(pte, __pte(0)); 347 continue; 348 } 349 350 /* 351 * We will re-use the existing mapping. 352 * Xen for example has some special requirements, like mapping 353 * pagetable pages as RO. So assume someone who pre-setup 354 * these mappings are more intelligent. 355 */ 356 if (pte_val(*pte)) { 357 if (!after_bootmem) 358 pages++; 359 continue; 360 } 361 362 if (0) 363 printk(" pte=%p addr=%lx pte=%016lx\n", 364 pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte); 365 pages++; 366 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot)); 367 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE; 368 } 369 370 update_page_count(PG_LEVEL_4K, pages); 371 372 return last_map_addr; 373 } 374 375 static unsigned long __meminit 376 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end, 377 unsigned long page_size_mask, pgprot_t prot) 378 { 379 unsigned long pages = 0, next; 380 unsigned long last_map_addr = end; 381 382 int i = pmd_index(address); 383 384 for (; i < PTRS_PER_PMD; i++, address = next) { 385 pmd_t *pmd = pmd_page + pmd_index(address); 386 pte_t *pte; 387 pgprot_t new_prot = prot; 388 389 next = (address & PMD_MASK) + PMD_SIZE; 390 if (address >= end) { 391 if (!after_bootmem && 392 !e820_any_mapped(address & PMD_MASK, next, E820_RAM) && 393 !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN)) 394 set_pmd(pmd, __pmd(0)); 395 continue; 396 } 397 398 if (pmd_val(*pmd)) { 399 if (!pmd_large(*pmd)) { 400 spin_lock(&init_mm.page_table_lock); 401 pte = (pte_t *)pmd_page_vaddr(*pmd); 402 last_map_addr = phys_pte_init(pte, address, 403 end, prot); 404 spin_unlock(&init_mm.page_table_lock); 405 continue; 406 } 407 /* 408 * If we are ok with PG_LEVEL_2M mapping, then we will 409 * use the existing mapping, 410 * 411 * Otherwise, we will split the large page mapping but 412 * use the same existing protection bits except for 413 * large page, so that we don't violate Intel's TLB 414 * Application note (317080) which says, while changing 415 * the page sizes, new and old translations should 416 * not differ with respect to page frame and 417 * attributes. 418 */ 419 if (page_size_mask & (1 << PG_LEVEL_2M)) { 420 if (!after_bootmem) 421 pages++; 422 last_map_addr = next; 423 continue; 424 } 425 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd)); 426 } 427 428 if (page_size_mask & (1<<PG_LEVEL_2M)) { 429 pages++; 430 spin_lock(&init_mm.page_table_lock); 431 set_pte((pte_t *)pmd, 432 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT, 433 __pgprot(pgprot_val(prot) | _PAGE_PSE))); 434 spin_unlock(&init_mm.page_table_lock); 435 last_map_addr = next; 436 continue; 437 } 438 439 pte = alloc_low_page(); 440 last_map_addr = phys_pte_init(pte, address, end, new_prot); 441 442 spin_lock(&init_mm.page_table_lock); 443 pmd_populate_kernel(&init_mm, pmd, pte); 444 spin_unlock(&init_mm.page_table_lock); 445 } 446 update_page_count(PG_LEVEL_2M, pages); 447 return last_map_addr; 448 } 449 450 static unsigned long __meminit 451 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end, 452 unsigned long page_size_mask) 453 { 454 unsigned long pages = 0, next; 455 unsigned long last_map_addr = end; 456 int i = pud_index(addr); 457 458 for (; i < PTRS_PER_PUD; i++, addr = next) { 459 pud_t *pud = pud_page + pud_index(addr); 460 pmd_t *pmd; 461 pgprot_t prot = PAGE_KERNEL; 462 463 next = (addr & PUD_MASK) + PUD_SIZE; 464 if (addr >= end) { 465 if (!after_bootmem && 466 !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) && 467 !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN)) 468 set_pud(pud, __pud(0)); 469 continue; 470 } 471 472 if (pud_val(*pud)) { 473 if (!pud_large(*pud)) { 474 pmd = pmd_offset(pud, 0); 475 last_map_addr = phys_pmd_init(pmd, addr, end, 476 page_size_mask, prot); 477 __flush_tlb_all(); 478 continue; 479 } 480 /* 481 * If we are ok with PG_LEVEL_1G mapping, then we will 482 * use the existing mapping. 483 * 484 * Otherwise, we will split the gbpage mapping but use 485 * the same existing protection bits except for large 486 * page, so that we don't violate Intel's TLB 487 * Application note (317080) which says, while changing 488 * the page sizes, new and old translations should 489 * not differ with respect to page frame and 490 * attributes. 491 */ 492 if (page_size_mask & (1 << PG_LEVEL_1G)) { 493 if (!after_bootmem) 494 pages++; 495 last_map_addr = next; 496 continue; 497 } 498 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud)); 499 } 500 501 if (page_size_mask & (1<<PG_LEVEL_1G)) { 502 pages++; 503 spin_lock(&init_mm.page_table_lock); 504 set_pte((pte_t *)pud, 505 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT, 506 PAGE_KERNEL_LARGE)); 507 spin_unlock(&init_mm.page_table_lock); 508 last_map_addr = next; 509 continue; 510 } 511 512 pmd = alloc_low_page(); 513 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask, 514 prot); 515 516 spin_lock(&init_mm.page_table_lock); 517 pud_populate(&init_mm, pud, pmd); 518 spin_unlock(&init_mm.page_table_lock); 519 } 520 __flush_tlb_all(); 521 522 update_page_count(PG_LEVEL_1G, pages); 523 524 return last_map_addr; 525 } 526 527 unsigned long __meminit 528 kernel_physical_mapping_init(unsigned long start, 529 unsigned long end, 530 unsigned long page_size_mask) 531 { 532 bool pgd_changed = false; 533 unsigned long next, last_map_addr = end; 534 unsigned long addr; 535 536 start = (unsigned long)__va(start); 537 end = (unsigned long)__va(end); 538 addr = start; 539 540 for (; start < end; start = next) { 541 pgd_t *pgd = pgd_offset_k(start); 542 pud_t *pud; 543 544 next = (start & PGDIR_MASK) + PGDIR_SIZE; 545 546 if (pgd_val(*pgd)) { 547 pud = (pud_t *)pgd_page_vaddr(*pgd); 548 last_map_addr = phys_pud_init(pud, __pa(start), 549 __pa(end), page_size_mask); 550 continue; 551 } 552 553 pud = alloc_low_page(); 554 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end), 555 page_size_mask); 556 557 spin_lock(&init_mm.page_table_lock); 558 pgd_populate(&init_mm, pgd, pud); 559 spin_unlock(&init_mm.page_table_lock); 560 pgd_changed = true; 561 } 562 563 if (pgd_changed) 564 sync_global_pgds(addr, end - 1, 0); 565 566 __flush_tlb_all(); 567 568 return last_map_addr; 569 } 570 571 #ifndef CONFIG_NUMA 572 void __init initmem_init(void) 573 { 574 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0); 575 } 576 #endif 577 578 void __init paging_init(void) 579 { 580 sparse_memory_present_with_active_regions(MAX_NUMNODES); 581 sparse_init(); 582 583 /* 584 * clear the default setting with node 0 585 * note: don't use nodes_clear here, that is really clearing when 586 * numa support is not compiled in, and later node_set_state 587 * will not set it back. 588 */ 589 node_clear_state(0, N_MEMORY); 590 if (N_MEMORY != N_NORMAL_MEMORY) 591 node_clear_state(0, N_NORMAL_MEMORY); 592 593 zone_sizes_init(); 594 } 595 596 /* 597 * Memory hotplug specific functions 598 */ 599 #ifdef CONFIG_MEMORY_HOTPLUG 600 /* 601 * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need 602 * updating. 603 */ 604 static void update_end_of_memory_vars(u64 start, u64 size) 605 { 606 unsigned long end_pfn = PFN_UP(start + size); 607 608 if (end_pfn > max_pfn) { 609 max_pfn = end_pfn; 610 max_low_pfn = end_pfn; 611 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1; 612 } 613 } 614 615 /* 616 * Memory is added always to NORMAL zone. This means you will never get 617 * additional DMA/DMA32 memory. 618 */ 619 int arch_add_memory(int nid, u64 start, u64 size, bool for_device) 620 { 621 struct pglist_data *pgdat = NODE_DATA(nid); 622 struct zone *zone = pgdat->node_zones + 623 zone_for_memory(nid, start, size, ZONE_NORMAL, for_device); 624 unsigned long start_pfn = start >> PAGE_SHIFT; 625 unsigned long nr_pages = size >> PAGE_SHIFT; 626 int ret; 627 628 init_memory_mapping(start, start + size); 629 630 ret = __add_pages(nid, zone, start_pfn, nr_pages); 631 WARN_ON_ONCE(ret); 632 633 /* update max_pfn, max_low_pfn and high_memory */ 634 update_end_of_memory_vars(start, size); 635 636 return ret; 637 } 638 EXPORT_SYMBOL_GPL(arch_add_memory); 639 640 #define PAGE_INUSE 0xFD 641 642 static void __meminit free_pagetable(struct page *page, int order) 643 { 644 unsigned long magic; 645 unsigned int nr_pages = 1 << order; 646 struct vmem_altmap *altmap = to_vmem_altmap((unsigned long) page); 647 648 if (altmap) { 649 vmem_altmap_free(altmap, nr_pages); 650 return; 651 } 652 653 /* bootmem page has reserved flag */ 654 if (PageReserved(page)) { 655 __ClearPageReserved(page); 656 657 magic = (unsigned long)page->lru.next; 658 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) { 659 while (nr_pages--) 660 put_page_bootmem(page++); 661 } else 662 while (nr_pages--) 663 free_reserved_page(page++); 664 } else 665 free_pages((unsigned long)page_address(page), order); 666 } 667 668 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd) 669 { 670 pte_t *pte; 671 int i; 672 673 for (i = 0; i < PTRS_PER_PTE; i++) { 674 pte = pte_start + i; 675 if (pte_val(*pte)) 676 return; 677 } 678 679 /* free a pte talbe */ 680 free_pagetable(pmd_page(*pmd), 0); 681 spin_lock(&init_mm.page_table_lock); 682 pmd_clear(pmd); 683 spin_unlock(&init_mm.page_table_lock); 684 } 685 686 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud) 687 { 688 pmd_t *pmd; 689 int i; 690 691 for (i = 0; i < PTRS_PER_PMD; i++) { 692 pmd = pmd_start + i; 693 if (pmd_val(*pmd)) 694 return; 695 } 696 697 /* free a pmd talbe */ 698 free_pagetable(pud_page(*pud), 0); 699 spin_lock(&init_mm.page_table_lock); 700 pud_clear(pud); 701 spin_unlock(&init_mm.page_table_lock); 702 } 703 704 /* Return true if pgd is changed, otherwise return false. */ 705 static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd) 706 { 707 pud_t *pud; 708 int i; 709 710 for (i = 0; i < PTRS_PER_PUD; i++) { 711 pud = pud_start + i; 712 if (pud_val(*pud)) 713 return false; 714 } 715 716 /* free a pud table */ 717 free_pagetable(pgd_page(*pgd), 0); 718 spin_lock(&init_mm.page_table_lock); 719 pgd_clear(pgd); 720 spin_unlock(&init_mm.page_table_lock); 721 722 return true; 723 } 724 725 static void __meminit 726 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end, 727 bool direct) 728 { 729 unsigned long next, pages = 0; 730 pte_t *pte; 731 void *page_addr; 732 phys_addr_t phys_addr; 733 734 pte = pte_start + pte_index(addr); 735 for (; addr < end; addr = next, pte++) { 736 next = (addr + PAGE_SIZE) & PAGE_MASK; 737 if (next > end) 738 next = end; 739 740 if (!pte_present(*pte)) 741 continue; 742 743 /* 744 * We mapped [0,1G) memory as identity mapping when 745 * initializing, in arch/x86/kernel/head_64.S. These 746 * pagetables cannot be removed. 747 */ 748 phys_addr = pte_val(*pte) + (addr & PAGE_MASK); 749 if (phys_addr < (phys_addr_t)0x40000000) 750 return; 751 752 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) { 753 /* 754 * Do not free direct mapping pages since they were 755 * freed when offlining, or simplely not in use. 756 */ 757 if (!direct) 758 free_pagetable(pte_page(*pte), 0); 759 760 spin_lock(&init_mm.page_table_lock); 761 pte_clear(&init_mm, addr, pte); 762 spin_unlock(&init_mm.page_table_lock); 763 764 /* For non-direct mapping, pages means nothing. */ 765 pages++; 766 } else { 767 /* 768 * If we are here, we are freeing vmemmap pages since 769 * direct mapped memory ranges to be freed are aligned. 770 * 771 * If we are not removing the whole page, it means 772 * other page structs in this page are being used and 773 * we canot remove them. So fill the unused page_structs 774 * with 0xFD, and remove the page when it is wholly 775 * filled with 0xFD. 776 */ 777 memset((void *)addr, PAGE_INUSE, next - addr); 778 779 page_addr = page_address(pte_page(*pte)); 780 if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) { 781 free_pagetable(pte_page(*pte), 0); 782 783 spin_lock(&init_mm.page_table_lock); 784 pte_clear(&init_mm, addr, pte); 785 spin_unlock(&init_mm.page_table_lock); 786 } 787 } 788 } 789 790 /* Call free_pte_table() in remove_pmd_table(). */ 791 flush_tlb_all(); 792 if (direct) 793 update_page_count(PG_LEVEL_4K, -pages); 794 } 795 796 static void __meminit 797 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end, 798 bool direct) 799 { 800 unsigned long next, pages = 0; 801 pte_t *pte_base; 802 pmd_t *pmd; 803 void *page_addr; 804 805 pmd = pmd_start + pmd_index(addr); 806 for (; addr < end; addr = next, pmd++) { 807 next = pmd_addr_end(addr, end); 808 809 if (!pmd_present(*pmd)) 810 continue; 811 812 if (pmd_large(*pmd)) { 813 if (IS_ALIGNED(addr, PMD_SIZE) && 814 IS_ALIGNED(next, PMD_SIZE)) { 815 if (!direct) 816 free_pagetable(pmd_page(*pmd), 817 get_order(PMD_SIZE)); 818 819 spin_lock(&init_mm.page_table_lock); 820 pmd_clear(pmd); 821 spin_unlock(&init_mm.page_table_lock); 822 pages++; 823 } else { 824 /* If here, we are freeing vmemmap pages. */ 825 memset((void *)addr, PAGE_INUSE, next - addr); 826 827 page_addr = page_address(pmd_page(*pmd)); 828 if (!memchr_inv(page_addr, PAGE_INUSE, 829 PMD_SIZE)) { 830 free_pagetable(pmd_page(*pmd), 831 get_order(PMD_SIZE)); 832 833 spin_lock(&init_mm.page_table_lock); 834 pmd_clear(pmd); 835 spin_unlock(&init_mm.page_table_lock); 836 } 837 } 838 839 continue; 840 } 841 842 pte_base = (pte_t *)pmd_page_vaddr(*pmd); 843 remove_pte_table(pte_base, addr, next, direct); 844 free_pte_table(pte_base, pmd); 845 } 846 847 /* Call free_pmd_table() in remove_pud_table(). */ 848 if (direct) 849 update_page_count(PG_LEVEL_2M, -pages); 850 } 851 852 static void __meminit 853 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end, 854 bool direct) 855 { 856 unsigned long next, pages = 0; 857 pmd_t *pmd_base; 858 pud_t *pud; 859 void *page_addr; 860 861 pud = pud_start + pud_index(addr); 862 for (; addr < end; addr = next, pud++) { 863 next = pud_addr_end(addr, end); 864 865 if (!pud_present(*pud)) 866 continue; 867 868 if (pud_large(*pud)) { 869 if (IS_ALIGNED(addr, PUD_SIZE) && 870 IS_ALIGNED(next, PUD_SIZE)) { 871 if (!direct) 872 free_pagetable(pud_page(*pud), 873 get_order(PUD_SIZE)); 874 875 spin_lock(&init_mm.page_table_lock); 876 pud_clear(pud); 877 spin_unlock(&init_mm.page_table_lock); 878 pages++; 879 } else { 880 /* If here, we are freeing vmemmap pages. */ 881 memset((void *)addr, PAGE_INUSE, next - addr); 882 883 page_addr = page_address(pud_page(*pud)); 884 if (!memchr_inv(page_addr, PAGE_INUSE, 885 PUD_SIZE)) { 886 free_pagetable(pud_page(*pud), 887 get_order(PUD_SIZE)); 888 889 spin_lock(&init_mm.page_table_lock); 890 pud_clear(pud); 891 spin_unlock(&init_mm.page_table_lock); 892 } 893 } 894 895 continue; 896 } 897 898 pmd_base = (pmd_t *)pud_page_vaddr(*pud); 899 remove_pmd_table(pmd_base, addr, next, direct); 900 free_pmd_table(pmd_base, pud); 901 } 902 903 if (direct) 904 update_page_count(PG_LEVEL_1G, -pages); 905 } 906 907 /* start and end are both virtual address. */ 908 static void __meminit 909 remove_pagetable(unsigned long start, unsigned long end, bool direct) 910 { 911 unsigned long next; 912 unsigned long addr; 913 pgd_t *pgd; 914 pud_t *pud; 915 bool pgd_changed = false; 916 917 for (addr = start; addr < end; addr = next) { 918 next = pgd_addr_end(addr, end); 919 920 pgd = pgd_offset_k(addr); 921 if (!pgd_present(*pgd)) 922 continue; 923 924 pud = (pud_t *)pgd_page_vaddr(*pgd); 925 remove_pud_table(pud, addr, next, direct); 926 if (free_pud_table(pud, pgd)) 927 pgd_changed = true; 928 } 929 930 if (pgd_changed) 931 sync_global_pgds(start, end - 1, 1); 932 933 flush_tlb_all(); 934 } 935 936 void __ref vmemmap_free(unsigned long start, unsigned long end) 937 { 938 remove_pagetable(start, end, false); 939 } 940 941 #ifdef CONFIG_MEMORY_HOTREMOVE 942 static void __meminit 943 kernel_physical_mapping_remove(unsigned long start, unsigned long end) 944 { 945 start = (unsigned long)__va(start); 946 end = (unsigned long)__va(end); 947 948 remove_pagetable(start, end, true); 949 } 950 951 int __ref arch_remove_memory(u64 start, u64 size) 952 { 953 unsigned long start_pfn = start >> PAGE_SHIFT; 954 unsigned long nr_pages = size >> PAGE_SHIFT; 955 struct page *page = pfn_to_page(start_pfn); 956 struct vmem_altmap *altmap; 957 struct zone *zone; 958 int ret; 959 960 /* With altmap the first mapped page is offset from @start */ 961 altmap = to_vmem_altmap((unsigned long) page); 962 if (altmap) 963 page += vmem_altmap_offset(altmap); 964 zone = page_zone(page); 965 ret = __remove_pages(zone, start_pfn, nr_pages); 966 WARN_ON_ONCE(ret); 967 kernel_physical_mapping_remove(start, start + size); 968 969 return ret; 970 } 971 #endif 972 #endif /* CONFIG_MEMORY_HOTPLUG */ 973 974 static struct kcore_list kcore_vsyscall; 975 976 static void __init register_page_bootmem_info(void) 977 { 978 #ifdef CONFIG_NUMA 979 int i; 980 981 for_each_online_node(i) 982 register_page_bootmem_info_node(NODE_DATA(i)); 983 #endif 984 } 985 986 void __init mem_init(void) 987 { 988 pci_iommu_alloc(); 989 990 /* clear_bss() already clear the empty_zero_page */ 991 992 register_page_bootmem_info(); 993 994 /* this will put all memory onto the freelists */ 995 free_all_bootmem(); 996 after_bootmem = 1; 997 998 /* Register memory areas for /proc/kcore */ 999 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, 1000 PAGE_SIZE, KCORE_OTHER); 1001 1002 mem_init_print_info(NULL); 1003 } 1004 1005 const int rodata_test_data = 0xC3; 1006 EXPORT_SYMBOL_GPL(rodata_test_data); 1007 1008 int kernel_set_to_readonly; 1009 1010 void set_kernel_text_rw(void) 1011 { 1012 unsigned long start = PFN_ALIGN(_text); 1013 unsigned long end = PFN_ALIGN(__stop___ex_table); 1014 1015 if (!kernel_set_to_readonly) 1016 return; 1017 1018 pr_debug("Set kernel text: %lx - %lx for read write\n", 1019 start, end); 1020 1021 /* 1022 * Make the kernel identity mapping for text RW. Kernel text 1023 * mapping will always be RO. Refer to the comment in 1024 * static_protections() in pageattr.c 1025 */ 1026 set_memory_rw(start, (end - start) >> PAGE_SHIFT); 1027 } 1028 1029 void set_kernel_text_ro(void) 1030 { 1031 unsigned long start = PFN_ALIGN(_text); 1032 unsigned long end = PFN_ALIGN(__stop___ex_table); 1033 1034 if (!kernel_set_to_readonly) 1035 return; 1036 1037 pr_debug("Set kernel text: %lx - %lx for read only\n", 1038 start, end); 1039 1040 /* 1041 * Set the kernel identity mapping for text RO. 1042 */ 1043 set_memory_ro(start, (end - start) >> PAGE_SHIFT); 1044 } 1045 1046 void mark_rodata_ro(void) 1047 { 1048 unsigned long start = PFN_ALIGN(_text); 1049 unsigned long rodata_start = PFN_ALIGN(__start_rodata); 1050 unsigned long end = (unsigned long) &__end_rodata_hpage_align; 1051 unsigned long text_end = PFN_ALIGN(&__stop___ex_table); 1052 unsigned long rodata_end = PFN_ALIGN(&__end_rodata); 1053 unsigned long all_end; 1054 1055 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n", 1056 (end - start) >> 10); 1057 set_memory_ro(start, (end - start) >> PAGE_SHIFT); 1058 1059 kernel_set_to_readonly = 1; 1060 1061 /* 1062 * The rodata/data/bss/brk section (but not the kernel text!) 1063 * should also be not-executable. 1064 * 1065 * We align all_end to PMD_SIZE because the existing mapping 1066 * is a full PMD. If we would align _brk_end to PAGE_SIZE we 1067 * split the PMD and the reminder between _brk_end and the end 1068 * of the PMD will remain mapped executable. 1069 * 1070 * Any PMD which was setup after the one which covers _brk_end 1071 * has been zapped already via cleanup_highmem(). 1072 */ 1073 all_end = roundup((unsigned long)_brk_end, PMD_SIZE); 1074 set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT); 1075 1076 rodata_test(); 1077 1078 #ifdef CONFIG_CPA_DEBUG 1079 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end); 1080 set_memory_rw(start, (end-start) >> PAGE_SHIFT); 1081 1082 printk(KERN_INFO "Testing CPA: again\n"); 1083 set_memory_ro(start, (end-start) >> PAGE_SHIFT); 1084 #endif 1085 1086 free_init_pages("unused kernel", 1087 (unsigned long) __va(__pa_symbol(text_end)), 1088 (unsigned long) __va(__pa_symbol(rodata_start))); 1089 free_init_pages("unused kernel", 1090 (unsigned long) __va(__pa_symbol(rodata_end)), 1091 (unsigned long) __va(__pa_symbol(_sdata))); 1092 1093 debug_checkwx(); 1094 } 1095 1096 int kern_addr_valid(unsigned long addr) 1097 { 1098 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT; 1099 pgd_t *pgd; 1100 pud_t *pud; 1101 pmd_t *pmd; 1102 pte_t *pte; 1103 1104 if (above != 0 && above != -1UL) 1105 return 0; 1106 1107 pgd = pgd_offset_k(addr); 1108 if (pgd_none(*pgd)) 1109 return 0; 1110 1111 pud = pud_offset(pgd, addr); 1112 if (pud_none(*pud)) 1113 return 0; 1114 1115 if (pud_large(*pud)) 1116 return pfn_valid(pud_pfn(*pud)); 1117 1118 pmd = pmd_offset(pud, addr); 1119 if (pmd_none(*pmd)) 1120 return 0; 1121 1122 if (pmd_large(*pmd)) 1123 return pfn_valid(pmd_pfn(*pmd)); 1124 1125 pte = pte_offset_kernel(pmd, addr); 1126 if (pte_none(*pte)) 1127 return 0; 1128 1129 return pfn_valid(pte_pfn(*pte)); 1130 } 1131 1132 static unsigned long probe_memory_block_size(void) 1133 { 1134 unsigned long bz = MIN_MEMORY_BLOCK_SIZE; 1135 1136 /* if system is UV or has 64GB of RAM or more, use large blocks */ 1137 if (is_uv_system() || ((max_pfn << PAGE_SHIFT) >= (64UL << 30))) 1138 bz = 2UL << 30; /* 2GB */ 1139 1140 pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20); 1141 1142 return bz; 1143 } 1144 1145 static unsigned long memory_block_size_probed; 1146 unsigned long memory_block_size_bytes(void) 1147 { 1148 if (!memory_block_size_probed) 1149 memory_block_size_probed = probe_memory_block_size(); 1150 1151 return memory_block_size_probed; 1152 } 1153 1154 #ifdef CONFIG_SPARSEMEM_VMEMMAP 1155 /* 1156 * Initialise the sparsemem vmemmap using huge-pages at the PMD level. 1157 */ 1158 static long __meminitdata addr_start, addr_end; 1159 static void __meminitdata *p_start, *p_end; 1160 static int __meminitdata node_start; 1161 1162 static int __meminit vmemmap_populate_hugepages(unsigned long start, 1163 unsigned long end, int node, struct vmem_altmap *altmap) 1164 { 1165 unsigned long addr; 1166 unsigned long next; 1167 pgd_t *pgd; 1168 pud_t *pud; 1169 pmd_t *pmd; 1170 1171 for (addr = start; addr < end; addr = next) { 1172 next = pmd_addr_end(addr, end); 1173 1174 pgd = vmemmap_pgd_populate(addr, node); 1175 if (!pgd) 1176 return -ENOMEM; 1177 1178 pud = vmemmap_pud_populate(pgd, addr, node); 1179 if (!pud) 1180 return -ENOMEM; 1181 1182 pmd = pmd_offset(pud, addr); 1183 if (pmd_none(*pmd)) { 1184 void *p; 1185 1186 p = __vmemmap_alloc_block_buf(PMD_SIZE, node, altmap); 1187 if (p) { 1188 pte_t entry; 1189 1190 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, 1191 PAGE_KERNEL_LARGE); 1192 set_pmd(pmd, __pmd(pte_val(entry))); 1193 1194 /* check to see if we have contiguous blocks */ 1195 if (p_end != p || node_start != node) { 1196 if (p_start) 1197 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n", 1198 addr_start, addr_end-1, p_start, p_end-1, node_start); 1199 addr_start = addr; 1200 node_start = node; 1201 p_start = p; 1202 } 1203 1204 addr_end = addr + PMD_SIZE; 1205 p_end = p + PMD_SIZE; 1206 continue; 1207 } else if (altmap) 1208 return -ENOMEM; /* no fallback */ 1209 } else if (pmd_large(*pmd)) { 1210 vmemmap_verify((pte_t *)pmd, node, addr, next); 1211 continue; 1212 } 1213 pr_warn_once("vmemmap: falling back to regular page backing\n"); 1214 if (vmemmap_populate_basepages(addr, next, node)) 1215 return -ENOMEM; 1216 } 1217 return 0; 1218 } 1219 1220 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node) 1221 { 1222 struct vmem_altmap *altmap = to_vmem_altmap(start); 1223 int err; 1224 1225 if (boot_cpu_has(X86_FEATURE_PSE)) 1226 err = vmemmap_populate_hugepages(start, end, node, altmap); 1227 else if (altmap) { 1228 pr_err_once("%s: no cpu support for altmap allocations\n", 1229 __func__); 1230 err = -ENOMEM; 1231 } else 1232 err = vmemmap_populate_basepages(start, end, node); 1233 if (!err) 1234 sync_global_pgds(start, end - 1, 0); 1235 return err; 1236 } 1237 1238 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE) 1239 void register_page_bootmem_memmap(unsigned long section_nr, 1240 struct page *start_page, unsigned long size) 1241 { 1242 unsigned long addr = (unsigned long)start_page; 1243 unsigned long end = (unsigned long)(start_page + size); 1244 unsigned long next; 1245 pgd_t *pgd; 1246 pud_t *pud; 1247 pmd_t *pmd; 1248 unsigned int nr_pages; 1249 struct page *page; 1250 1251 for (; addr < end; addr = next) { 1252 pte_t *pte = NULL; 1253 1254 pgd = pgd_offset_k(addr); 1255 if (pgd_none(*pgd)) { 1256 next = (addr + PAGE_SIZE) & PAGE_MASK; 1257 continue; 1258 } 1259 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO); 1260 1261 pud = pud_offset(pgd, addr); 1262 if (pud_none(*pud)) { 1263 next = (addr + PAGE_SIZE) & PAGE_MASK; 1264 continue; 1265 } 1266 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO); 1267 1268 if (!boot_cpu_has(X86_FEATURE_PSE)) { 1269 next = (addr + PAGE_SIZE) & PAGE_MASK; 1270 pmd = pmd_offset(pud, addr); 1271 if (pmd_none(*pmd)) 1272 continue; 1273 get_page_bootmem(section_nr, pmd_page(*pmd), 1274 MIX_SECTION_INFO); 1275 1276 pte = pte_offset_kernel(pmd, addr); 1277 if (pte_none(*pte)) 1278 continue; 1279 get_page_bootmem(section_nr, pte_page(*pte), 1280 SECTION_INFO); 1281 } else { 1282 next = pmd_addr_end(addr, end); 1283 1284 pmd = pmd_offset(pud, addr); 1285 if (pmd_none(*pmd)) 1286 continue; 1287 1288 nr_pages = 1 << (get_order(PMD_SIZE)); 1289 page = pmd_page(*pmd); 1290 while (nr_pages--) 1291 get_page_bootmem(section_nr, page++, 1292 SECTION_INFO); 1293 } 1294 } 1295 } 1296 #endif 1297 1298 void __meminit vmemmap_populate_print_last(void) 1299 { 1300 if (p_start) { 1301 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n", 1302 addr_start, addr_end-1, p_start, p_end-1, node_start); 1303 p_start = NULL; 1304 p_end = NULL; 1305 node_start = 0; 1306 } 1307 } 1308 #endif 1309