memory.c revision 8f4e2101fd7df9031a754eedb82e2060b51f8c45
1/* 2 * linux/mm/memory.c 3 * 4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 5 */ 6 7/* 8 * demand-loading started 01.12.91 - seems it is high on the list of 9 * things wanted, and it should be easy to implement. - Linus 10 */ 11 12/* 13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared 14 * pages started 02.12.91, seems to work. - Linus. 15 * 16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it 17 * would have taken more than the 6M I have free, but it worked well as 18 * far as I could see. 19 * 20 * Also corrected some "invalidate()"s - I wasn't doing enough of them. 21 */ 22 23/* 24 * Real VM (paging to/from disk) started 18.12.91. Much more work and 25 * thought has to go into this. Oh, well.. 26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why. 27 * Found it. Everything seems to work now. 28 * 20.12.91 - Ok, making the swap-device changeable like the root. 29 */ 30 31/* 32 * 05.04.94 - Multi-page memory management added for v1.1. 33 * Idea by Alex Bligh (alex@cconcepts.co.uk) 34 * 35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG 36 * (Gerhard.Wichert@pdb.siemens.de) 37 * 38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen) 39 */ 40 41#include <linux/kernel_stat.h> 42#include <linux/mm.h> 43#include <linux/hugetlb.h> 44#include <linux/mman.h> 45#include <linux/swap.h> 46#include <linux/highmem.h> 47#include <linux/pagemap.h> 48#include <linux/rmap.h> 49#include <linux/module.h> 50#include <linux/init.h> 51 52#include <asm/pgalloc.h> 53#include <asm/uaccess.h> 54#include <asm/tlb.h> 55#include <asm/tlbflush.h> 56#include <asm/pgtable.h> 57 58#include <linux/swapops.h> 59#include <linux/elf.h> 60 61#ifndef CONFIG_NEED_MULTIPLE_NODES 62/* use the per-pgdat data instead for discontigmem - mbligh */ 63unsigned long max_mapnr; 64struct page *mem_map; 65 66EXPORT_SYMBOL(max_mapnr); 67EXPORT_SYMBOL(mem_map); 68#endif 69 70unsigned long num_physpages; 71/* 72 * A number of key systems in x86 including ioremap() rely on the assumption 73 * that high_memory defines the upper bound on direct map memory, then end 74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 76 * and ZONE_HIGHMEM. 77 */ 78void * high_memory; 79unsigned long vmalloc_earlyreserve; 80 81EXPORT_SYMBOL(num_physpages); 82EXPORT_SYMBOL(high_memory); 83EXPORT_SYMBOL(vmalloc_earlyreserve); 84 85/* 86 * If a p?d_bad entry is found while walking page tables, report 87 * the error, before resetting entry to p?d_none. Usually (but 88 * very seldom) called out from the p?d_none_or_clear_bad macros. 89 */ 90 91void pgd_clear_bad(pgd_t *pgd) 92{ 93 pgd_ERROR(*pgd); 94 pgd_clear(pgd); 95} 96 97void pud_clear_bad(pud_t *pud) 98{ 99 pud_ERROR(*pud); 100 pud_clear(pud); 101} 102 103void pmd_clear_bad(pmd_t *pmd) 104{ 105 pmd_ERROR(*pmd); 106 pmd_clear(pmd); 107} 108 109/* 110 * Note: this doesn't free the actual pages themselves. That 111 * has been handled earlier when unmapping all the memory regions. 112 */ 113static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd) 114{ 115 struct page *page = pmd_page(*pmd); 116 pmd_clear(pmd); 117 pte_free_tlb(tlb, page); 118 dec_page_state(nr_page_table_pages); 119 tlb->mm->nr_ptes--; 120} 121 122static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 123 unsigned long addr, unsigned long end, 124 unsigned long floor, unsigned long ceiling) 125{ 126 pmd_t *pmd; 127 unsigned long next; 128 unsigned long start; 129 130 start = addr; 131 pmd = pmd_offset(pud, addr); 132 do { 133 next = pmd_addr_end(addr, end); 134 if (pmd_none_or_clear_bad(pmd)) 135 continue; 136 free_pte_range(tlb, pmd); 137 } while (pmd++, addr = next, addr != end); 138 139 start &= PUD_MASK; 140 if (start < floor) 141 return; 142 if (ceiling) { 143 ceiling &= PUD_MASK; 144 if (!ceiling) 145 return; 146 } 147 if (end - 1 > ceiling - 1) 148 return; 149 150 pmd = pmd_offset(pud, start); 151 pud_clear(pud); 152 pmd_free_tlb(tlb, pmd); 153} 154 155static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 156 unsigned long addr, unsigned long end, 157 unsigned long floor, unsigned long ceiling) 158{ 159 pud_t *pud; 160 unsigned long next; 161 unsigned long start; 162 163 start = addr; 164 pud = pud_offset(pgd, addr); 165 do { 166 next = pud_addr_end(addr, end); 167 if (pud_none_or_clear_bad(pud)) 168 continue; 169 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 170 } while (pud++, addr = next, addr != end); 171 172 start &= PGDIR_MASK; 173 if (start < floor) 174 return; 175 if (ceiling) { 176 ceiling &= PGDIR_MASK; 177 if (!ceiling) 178 return; 179 } 180 if (end - 1 > ceiling - 1) 181 return; 182 183 pud = pud_offset(pgd, start); 184 pgd_clear(pgd); 185 pud_free_tlb(tlb, pud); 186} 187 188/* 189 * This function frees user-level page tables of a process. 190 * 191 * Must be called with pagetable lock held. 192 */ 193void free_pgd_range(struct mmu_gather **tlb, 194 unsigned long addr, unsigned long end, 195 unsigned long floor, unsigned long ceiling) 196{ 197 pgd_t *pgd; 198 unsigned long next; 199 unsigned long start; 200 201 /* 202 * The next few lines have given us lots of grief... 203 * 204 * Why are we testing PMD* at this top level? Because often 205 * there will be no work to do at all, and we'd prefer not to 206 * go all the way down to the bottom just to discover that. 207 * 208 * Why all these "- 1"s? Because 0 represents both the bottom 209 * of the address space and the top of it (using -1 for the 210 * top wouldn't help much: the masks would do the wrong thing). 211 * The rule is that addr 0 and floor 0 refer to the bottom of 212 * the address space, but end 0 and ceiling 0 refer to the top 213 * Comparisons need to use "end - 1" and "ceiling - 1" (though 214 * that end 0 case should be mythical). 215 * 216 * Wherever addr is brought up or ceiling brought down, we must 217 * be careful to reject "the opposite 0" before it confuses the 218 * subsequent tests. But what about where end is brought down 219 * by PMD_SIZE below? no, end can't go down to 0 there. 220 * 221 * Whereas we round start (addr) and ceiling down, by different 222 * masks at different levels, in order to test whether a table 223 * now has no other vmas using it, so can be freed, we don't 224 * bother to round floor or end up - the tests don't need that. 225 */ 226 227 addr &= PMD_MASK; 228 if (addr < floor) { 229 addr += PMD_SIZE; 230 if (!addr) 231 return; 232 } 233 if (ceiling) { 234 ceiling &= PMD_MASK; 235 if (!ceiling) 236 return; 237 } 238 if (end - 1 > ceiling - 1) 239 end -= PMD_SIZE; 240 if (addr > end - 1) 241 return; 242 243 start = addr; 244 pgd = pgd_offset((*tlb)->mm, addr); 245 do { 246 next = pgd_addr_end(addr, end); 247 if (pgd_none_or_clear_bad(pgd)) 248 continue; 249 free_pud_range(*tlb, pgd, addr, next, floor, ceiling); 250 } while (pgd++, addr = next, addr != end); 251 252 if (!(*tlb)->fullmm) 253 flush_tlb_pgtables((*tlb)->mm, start, end); 254} 255 256void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma, 257 unsigned long floor, unsigned long ceiling) 258{ 259 while (vma) { 260 struct vm_area_struct *next = vma->vm_next; 261 unsigned long addr = vma->vm_start; 262 263 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) { 264 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 265 floor, next? next->vm_start: ceiling); 266 } else { 267 /* 268 * Optimization: gather nearby vmas into one call down 269 */ 270 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 271 && !is_hugepage_only_range(vma->vm_mm, next->vm_start, 272 HPAGE_SIZE)) { 273 vma = next; 274 next = vma->vm_next; 275 } 276 free_pgd_range(tlb, addr, vma->vm_end, 277 floor, next? next->vm_start: ceiling); 278 } 279 vma = next; 280 } 281} 282 283int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address) 284{ 285 struct page *new = pte_alloc_one(mm, address); 286 if (!new) 287 return -ENOMEM; 288 289 spin_lock(&mm->page_table_lock); 290 if (pmd_present(*pmd)) /* Another has populated it */ 291 pte_free(new); 292 else { 293 mm->nr_ptes++; 294 inc_page_state(nr_page_table_pages); 295 pmd_populate(mm, pmd, new); 296 } 297 spin_unlock(&mm->page_table_lock); 298 return 0; 299} 300 301int __pte_alloc_kernel(pmd_t *pmd, unsigned long address) 302{ 303 pte_t *new = pte_alloc_one_kernel(&init_mm, address); 304 if (!new) 305 return -ENOMEM; 306 307 spin_lock(&init_mm.page_table_lock); 308 if (pmd_present(*pmd)) /* Another has populated it */ 309 pte_free_kernel(new); 310 else 311 pmd_populate_kernel(&init_mm, pmd, new); 312 spin_unlock(&init_mm.page_table_lock); 313 return 0; 314} 315 316static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss) 317{ 318 if (file_rss) 319 add_mm_counter(mm, file_rss, file_rss); 320 if (anon_rss) 321 add_mm_counter(mm, anon_rss, anon_rss); 322} 323 324/* 325 * This function is called to print an error when a pte in a 326 * !VM_RESERVED region is found pointing to an invalid pfn (which 327 * is an error. 328 * 329 * The calling function must still handle the error. 330 */ 331void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr) 332{ 333 printk(KERN_ERR "Bad pte = %08llx, process = %s, " 334 "vm_flags = %lx, vaddr = %lx\n", 335 (long long)pte_val(pte), 336 (vma->vm_mm == current->mm ? current->comm : "???"), 337 vma->vm_flags, vaddr); 338 dump_stack(); 339} 340 341/* 342 * copy one vm_area from one task to the other. Assumes the page tables 343 * already present in the new task to be cleared in the whole range 344 * covered by this vma. 345 */ 346 347static inline void 348copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 349 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 350 unsigned long addr, int *rss) 351{ 352 unsigned long vm_flags = vma->vm_flags; 353 pte_t pte = *src_pte; 354 struct page *page; 355 unsigned long pfn; 356 357 /* pte contains position in swap or file, so copy. */ 358 if (unlikely(!pte_present(pte))) { 359 if (!pte_file(pte)) { 360 swap_duplicate(pte_to_swp_entry(pte)); 361 /* make sure dst_mm is on swapoff's mmlist. */ 362 if (unlikely(list_empty(&dst_mm->mmlist))) { 363 spin_lock(&mmlist_lock); 364 list_add(&dst_mm->mmlist, &src_mm->mmlist); 365 spin_unlock(&mmlist_lock); 366 } 367 } 368 goto out_set_pte; 369 } 370 371 /* If the region is VM_RESERVED, the mapping is not 372 * mapped via rmap - duplicate the pte as is. 373 */ 374 if (vm_flags & VM_RESERVED) 375 goto out_set_pte; 376 377 pfn = pte_pfn(pte); 378 /* If the pte points outside of valid memory but 379 * the region is not VM_RESERVED, we have a problem. 380 */ 381 if (unlikely(!pfn_valid(pfn))) { 382 print_bad_pte(vma, pte, addr); 383 goto out_set_pte; /* try to do something sane */ 384 } 385 386 page = pfn_to_page(pfn); 387 388 /* 389 * If it's a COW mapping, write protect it both 390 * in the parent and the child 391 */ 392 if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) { 393 ptep_set_wrprotect(src_mm, addr, src_pte); 394 pte = *src_pte; 395 } 396 397 /* 398 * If it's a shared mapping, mark it clean in 399 * the child 400 */ 401 if (vm_flags & VM_SHARED) 402 pte = pte_mkclean(pte); 403 pte = pte_mkold(pte); 404 get_page(page); 405 page_dup_rmap(page); 406 rss[!!PageAnon(page)]++; 407 408out_set_pte: 409 set_pte_at(dst_mm, addr, dst_pte, pte); 410} 411 412static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 413 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 414 unsigned long addr, unsigned long end) 415{ 416 pte_t *src_pte, *dst_pte; 417 spinlock_t *src_ptl, *dst_ptl; 418 int progress = 0; 419 int rss[2]; 420 421again: 422 rss[1] = rss[0] = 0; 423 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 424 if (!dst_pte) 425 return -ENOMEM; 426 src_pte = pte_offset_map_nested(src_pmd, addr); 427 src_ptl = &src_mm->page_table_lock; 428 spin_lock(src_ptl); 429 430 do { 431 /* 432 * We are holding two locks at this point - either of them 433 * could generate latencies in another task on another CPU. 434 */ 435 if (progress >= 32) { 436 progress = 0; 437 if (need_resched() || 438 need_lockbreak(src_ptl) || 439 need_lockbreak(dst_ptl)) 440 break; 441 } 442 if (pte_none(*src_pte)) { 443 progress++; 444 continue; 445 } 446 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss); 447 progress += 8; 448 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 449 450 spin_unlock(src_ptl); 451 pte_unmap_nested(src_pte - 1); 452 add_mm_rss(dst_mm, rss[0], rss[1]); 453 pte_unmap_unlock(dst_pte - 1, dst_ptl); 454 cond_resched(); 455 if (addr != end) 456 goto again; 457 return 0; 458} 459 460static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 461 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 462 unsigned long addr, unsigned long end) 463{ 464 pmd_t *src_pmd, *dst_pmd; 465 unsigned long next; 466 467 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 468 if (!dst_pmd) 469 return -ENOMEM; 470 src_pmd = pmd_offset(src_pud, addr); 471 do { 472 next = pmd_addr_end(addr, end); 473 if (pmd_none_or_clear_bad(src_pmd)) 474 continue; 475 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 476 vma, addr, next)) 477 return -ENOMEM; 478 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 479 return 0; 480} 481 482static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 483 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 484 unsigned long addr, unsigned long end) 485{ 486 pud_t *src_pud, *dst_pud; 487 unsigned long next; 488 489 dst_pud = pud_alloc(dst_mm, dst_pgd, addr); 490 if (!dst_pud) 491 return -ENOMEM; 492 src_pud = pud_offset(src_pgd, addr); 493 do { 494 next = pud_addr_end(addr, end); 495 if (pud_none_or_clear_bad(src_pud)) 496 continue; 497 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 498 vma, addr, next)) 499 return -ENOMEM; 500 } while (dst_pud++, src_pud++, addr = next, addr != end); 501 return 0; 502} 503 504int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 505 struct vm_area_struct *vma) 506{ 507 pgd_t *src_pgd, *dst_pgd; 508 unsigned long next; 509 unsigned long addr = vma->vm_start; 510 unsigned long end = vma->vm_end; 511 512 /* 513 * Don't copy ptes where a page fault will fill them correctly. 514 * Fork becomes much lighter when there are big shared or private 515 * readonly mappings. The tradeoff is that copy_page_range is more 516 * efficient than faulting. 517 */ 518 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) { 519 if (!vma->anon_vma) 520 return 0; 521 } 522 523 if (is_vm_hugetlb_page(vma)) 524 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 525 526 dst_pgd = pgd_offset(dst_mm, addr); 527 src_pgd = pgd_offset(src_mm, addr); 528 do { 529 next = pgd_addr_end(addr, end); 530 if (pgd_none_or_clear_bad(src_pgd)) 531 continue; 532 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, 533 vma, addr, next)) 534 return -ENOMEM; 535 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 536 return 0; 537} 538 539static void zap_pte_range(struct mmu_gather *tlb, 540 struct vm_area_struct *vma, pmd_t *pmd, 541 unsigned long addr, unsigned long end, 542 struct zap_details *details) 543{ 544 struct mm_struct *mm = tlb->mm; 545 pte_t *pte; 546 int file_rss = 0; 547 int anon_rss = 0; 548 549 pte = pte_offset_map(pmd, addr); 550 do { 551 pte_t ptent = *pte; 552 if (pte_none(ptent)) 553 continue; 554 if (pte_present(ptent)) { 555 struct page *page = NULL; 556 if (!(vma->vm_flags & VM_RESERVED)) { 557 unsigned long pfn = pte_pfn(ptent); 558 if (unlikely(!pfn_valid(pfn))) 559 print_bad_pte(vma, ptent, addr); 560 else 561 page = pfn_to_page(pfn); 562 } 563 if (unlikely(details) && page) { 564 /* 565 * unmap_shared_mapping_pages() wants to 566 * invalidate cache without truncating: 567 * unmap shared but keep private pages. 568 */ 569 if (details->check_mapping && 570 details->check_mapping != page->mapping) 571 continue; 572 /* 573 * Each page->index must be checked when 574 * invalidating or truncating nonlinear. 575 */ 576 if (details->nonlinear_vma && 577 (page->index < details->first_index || 578 page->index > details->last_index)) 579 continue; 580 } 581 ptent = ptep_get_and_clear_full(mm, addr, pte, 582 tlb->fullmm); 583 tlb_remove_tlb_entry(tlb, pte, addr); 584 if (unlikely(!page)) 585 continue; 586 if (unlikely(details) && details->nonlinear_vma 587 && linear_page_index(details->nonlinear_vma, 588 addr) != page->index) 589 set_pte_at(mm, addr, pte, 590 pgoff_to_pte(page->index)); 591 if (PageAnon(page)) 592 anon_rss--; 593 else { 594 if (pte_dirty(ptent)) 595 set_page_dirty(page); 596 if (pte_young(ptent)) 597 mark_page_accessed(page); 598 file_rss--; 599 } 600 page_remove_rmap(page); 601 tlb_remove_page(tlb, page); 602 continue; 603 } 604 /* 605 * If details->check_mapping, we leave swap entries; 606 * if details->nonlinear_vma, we leave file entries. 607 */ 608 if (unlikely(details)) 609 continue; 610 if (!pte_file(ptent)) 611 free_swap_and_cache(pte_to_swp_entry(ptent)); 612 pte_clear_full(mm, addr, pte, tlb->fullmm); 613 } while (pte++, addr += PAGE_SIZE, addr != end); 614 615 add_mm_rss(mm, file_rss, anon_rss); 616 pte_unmap(pte - 1); 617} 618 619static inline void zap_pmd_range(struct mmu_gather *tlb, 620 struct vm_area_struct *vma, pud_t *pud, 621 unsigned long addr, unsigned long end, 622 struct zap_details *details) 623{ 624 pmd_t *pmd; 625 unsigned long next; 626 627 pmd = pmd_offset(pud, addr); 628 do { 629 next = pmd_addr_end(addr, end); 630 if (pmd_none_or_clear_bad(pmd)) 631 continue; 632 zap_pte_range(tlb, vma, pmd, addr, next, details); 633 } while (pmd++, addr = next, addr != end); 634} 635 636static inline void zap_pud_range(struct mmu_gather *tlb, 637 struct vm_area_struct *vma, pgd_t *pgd, 638 unsigned long addr, unsigned long end, 639 struct zap_details *details) 640{ 641 pud_t *pud; 642 unsigned long next; 643 644 pud = pud_offset(pgd, addr); 645 do { 646 next = pud_addr_end(addr, end); 647 if (pud_none_or_clear_bad(pud)) 648 continue; 649 zap_pmd_range(tlb, vma, pud, addr, next, details); 650 } while (pud++, addr = next, addr != end); 651} 652 653static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 654 unsigned long addr, unsigned long end, 655 struct zap_details *details) 656{ 657 pgd_t *pgd; 658 unsigned long next; 659 660 if (details && !details->check_mapping && !details->nonlinear_vma) 661 details = NULL; 662 663 BUG_ON(addr >= end); 664 tlb_start_vma(tlb, vma); 665 pgd = pgd_offset(vma->vm_mm, addr); 666 do { 667 next = pgd_addr_end(addr, end); 668 if (pgd_none_or_clear_bad(pgd)) 669 continue; 670 zap_pud_range(tlb, vma, pgd, addr, next, details); 671 } while (pgd++, addr = next, addr != end); 672 tlb_end_vma(tlb, vma); 673} 674 675#ifdef CONFIG_PREEMPT 676# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE) 677#else 678/* No preempt: go for improved straight-line efficiency */ 679# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE) 680#endif 681 682/** 683 * unmap_vmas - unmap a range of memory covered by a list of vma's 684 * @tlbp: address of the caller's struct mmu_gather 685 * @mm: the controlling mm_struct 686 * @vma: the starting vma 687 * @start_addr: virtual address at which to start unmapping 688 * @end_addr: virtual address at which to end unmapping 689 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here 690 * @details: details of nonlinear truncation or shared cache invalidation 691 * 692 * Returns the end address of the unmapping (restart addr if interrupted). 693 * 694 * Unmap all pages in the vma list. Called under page_table_lock. 695 * 696 * We aim to not hold page_table_lock for too long (for scheduling latency 697 * reasons). So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to 698 * return the ending mmu_gather to the caller. 699 * 700 * Only addresses between `start' and `end' will be unmapped. 701 * 702 * The VMA list must be sorted in ascending virtual address order. 703 * 704 * unmap_vmas() assumes that the caller will flush the whole unmapped address 705 * range after unmap_vmas() returns. So the only responsibility here is to 706 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 707 * drops the lock and schedules. 708 */ 709unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm, 710 struct vm_area_struct *vma, unsigned long start_addr, 711 unsigned long end_addr, unsigned long *nr_accounted, 712 struct zap_details *details) 713{ 714 unsigned long zap_bytes = ZAP_BLOCK_SIZE; 715 unsigned long tlb_start = 0; /* For tlb_finish_mmu */ 716 int tlb_start_valid = 0; 717 unsigned long start = start_addr; 718 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL; 719 int fullmm = (*tlbp)->fullmm; 720 721 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) { 722 unsigned long end; 723 724 start = max(vma->vm_start, start_addr); 725 if (start >= vma->vm_end) 726 continue; 727 end = min(vma->vm_end, end_addr); 728 if (end <= vma->vm_start) 729 continue; 730 731 if (vma->vm_flags & VM_ACCOUNT) 732 *nr_accounted += (end - start) >> PAGE_SHIFT; 733 734 while (start != end) { 735 unsigned long block; 736 737 if (!tlb_start_valid) { 738 tlb_start = start; 739 tlb_start_valid = 1; 740 } 741 742 if (is_vm_hugetlb_page(vma)) { 743 block = end - start; 744 unmap_hugepage_range(vma, start, end); 745 } else { 746 block = min(zap_bytes, end - start); 747 unmap_page_range(*tlbp, vma, start, 748 start + block, details); 749 } 750 751 start += block; 752 zap_bytes -= block; 753 if ((long)zap_bytes > 0) 754 continue; 755 756 tlb_finish_mmu(*tlbp, tlb_start, start); 757 758 if (need_resched() || 759 need_lockbreak(&mm->page_table_lock) || 760 (i_mmap_lock && need_lockbreak(i_mmap_lock))) { 761 if (i_mmap_lock) { 762 /* must reset count of rss freed */ 763 *tlbp = tlb_gather_mmu(mm, fullmm); 764 goto out; 765 } 766 spin_unlock(&mm->page_table_lock); 767 cond_resched(); 768 spin_lock(&mm->page_table_lock); 769 } 770 771 *tlbp = tlb_gather_mmu(mm, fullmm); 772 tlb_start_valid = 0; 773 zap_bytes = ZAP_BLOCK_SIZE; 774 } 775 } 776out: 777 return start; /* which is now the end (or restart) address */ 778} 779 780/** 781 * zap_page_range - remove user pages in a given range 782 * @vma: vm_area_struct holding the applicable pages 783 * @address: starting address of pages to zap 784 * @size: number of bytes to zap 785 * @details: details of nonlinear truncation or shared cache invalidation 786 */ 787unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address, 788 unsigned long size, struct zap_details *details) 789{ 790 struct mm_struct *mm = vma->vm_mm; 791 struct mmu_gather *tlb; 792 unsigned long end = address + size; 793 unsigned long nr_accounted = 0; 794 795 if (is_vm_hugetlb_page(vma)) { 796 zap_hugepage_range(vma, address, size); 797 return end; 798 } 799 800 lru_add_drain(); 801 spin_lock(&mm->page_table_lock); 802 tlb = tlb_gather_mmu(mm, 0); 803 update_hiwater_rss(mm); 804 end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details); 805 tlb_finish_mmu(tlb, address, end); 806 spin_unlock(&mm->page_table_lock); 807 return end; 808} 809 810/* 811 * Do a quick page-table lookup for a single page. 812 * mm->page_table_lock must be held. 813 */ 814static struct page *__follow_page(struct mm_struct *mm, unsigned long address, 815 int read, int write, int accessed) 816{ 817 pgd_t *pgd; 818 pud_t *pud; 819 pmd_t *pmd; 820 pte_t *ptep, pte; 821 unsigned long pfn; 822 struct page *page; 823 824 page = follow_huge_addr(mm, address, write); 825 if (! IS_ERR(page)) 826 return page; 827 828 pgd = pgd_offset(mm, address); 829 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 830 goto out; 831 832 pud = pud_offset(pgd, address); 833 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 834 goto out; 835 836 pmd = pmd_offset(pud, address); 837 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 838 goto out; 839 if (pmd_huge(*pmd)) 840 return follow_huge_pmd(mm, address, pmd, write); 841 842 ptep = pte_offset_map(pmd, address); 843 if (!ptep) 844 goto out; 845 846 pte = *ptep; 847 pte_unmap(ptep); 848 if (pte_present(pte)) { 849 if (write && !pte_write(pte)) 850 goto out; 851 if (read && !pte_read(pte)) 852 goto out; 853 pfn = pte_pfn(pte); 854 if (pfn_valid(pfn)) { 855 page = pfn_to_page(pfn); 856 if (accessed) { 857 if (write && !pte_dirty(pte) &&!PageDirty(page)) 858 set_page_dirty(page); 859 mark_page_accessed(page); 860 } 861 return page; 862 } 863 } 864 865out: 866 return NULL; 867} 868 869inline struct page * 870follow_page(struct mm_struct *mm, unsigned long address, int write) 871{ 872 return __follow_page(mm, address, 0, write, 1); 873} 874 875/* 876 * check_user_page_readable() can be called frm niterrupt context by oprofile, 877 * so we need to avoid taking any non-irq-safe locks 878 */ 879int check_user_page_readable(struct mm_struct *mm, unsigned long address) 880{ 881 return __follow_page(mm, address, 1, 0, 0) != NULL; 882} 883EXPORT_SYMBOL(check_user_page_readable); 884 885static inline int 886untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma, 887 unsigned long address) 888{ 889 pgd_t *pgd; 890 pud_t *pud; 891 pmd_t *pmd; 892 893 /* Check if the vma is for an anonymous mapping. */ 894 if (vma->vm_ops && vma->vm_ops->nopage) 895 return 0; 896 897 /* Check if page directory entry exists. */ 898 pgd = pgd_offset(mm, address); 899 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 900 return 1; 901 902 pud = pud_offset(pgd, address); 903 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 904 return 1; 905 906 /* Check if page middle directory entry exists. */ 907 pmd = pmd_offset(pud, address); 908 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 909 return 1; 910 911 /* There is a pte slot for 'address' in 'mm'. */ 912 return 0; 913} 914 915int get_user_pages(struct task_struct *tsk, struct mm_struct *mm, 916 unsigned long start, int len, int write, int force, 917 struct page **pages, struct vm_area_struct **vmas) 918{ 919 int i; 920 unsigned int flags; 921 922 /* 923 * Require read or write permissions. 924 * If 'force' is set, we only require the "MAY" flags. 925 */ 926 flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); 927 flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); 928 i = 0; 929 930 do { 931 struct vm_area_struct * vma; 932 933 vma = find_extend_vma(mm, start); 934 if (!vma && in_gate_area(tsk, start)) { 935 unsigned long pg = start & PAGE_MASK; 936 struct vm_area_struct *gate_vma = get_gate_vma(tsk); 937 pgd_t *pgd; 938 pud_t *pud; 939 pmd_t *pmd; 940 pte_t *pte; 941 if (write) /* user gate pages are read-only */ 942 return i ? : -EFAULT; 943 if (pg > TASK_SIZE) 944 pgd = pgd_offset_k(pg); 945 else 946 pgd = pgd_offset_gate(mm, pg); 947 BUG_ON(pgd_none(*pgd)); 948 pud = pud_offset(pgd, pg); 949 BUG_ON(pud_none(*pud)); 950 pmd = pmd_offset(pud, pg); 951 if (pmd_none(*pmd)) 952 return i ? : -EFAULT; 953 pte = pte_offset_map(pmd, pg); 954 if (pte_none(*pte)) { 955 pte_unmap(pte); 956 return i ? : -EFAULT; 957 } 958 if (pages) { 959 pages[i] = pte_page(*pte); 960 get_page(pages[i]); 961 } 962 pte_unmap(pte); 963 if (vmas) 964 vmas[i] = gate_vma; 965 i++; 966 start += PAGE_SIZE; 967 len--; 968 continue; 969 } 970 971 if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED)) 972 || !(flags & vma->vm_flags)) 973 return i ? : -EFAULT; 974 975 if (is_vm_hugetlb_page(vma)) { 976 i = follow_hugetlb_page(mm, vma, pages, vmas, 977 &start, &len, i); 978 continue; 979 } 980 spin_lock(&mm->page_table_lock); 981 do { 982 int write_access = write; 983 struct page *page; 984 985 cond_resched_lock(&mm->page_table_lock); 986 while (!(page = follow_page(mm, start, write_access))) { 987 int ret; 988 989 /* 990 * Shortcut for anonymous pages. We don't want 991 * to force the creation of pages tables for 992 * insanely big anonymously mapped areas that 993 * nobody touched so far. This is important 994 * for doing a core dump for these mappings. 995 */ 996 if (!write && untouched_anonymous_page(mm,vma,start)) { 997 page = ZERO_PAGE(start); 998 break; 999 } 1000 spin_unlock(&mm->page_table_lock); 1001 ret = __handle_mm_fault(mm, vma, start, write_access); 1002 1003 /* 1004 * The VM_FAULT_WRITE bit tells us that do_wp_page has 1005 * broken COW when necessary, even if maybe_mkwrite 1006 * decided not to set pte_write. We can thus safely do 1007 * subsequent page lookups as if they were reads. 1008 */ 1009 if (ret & VM_FAULT_WRITE) 1010 write_access = 0; 1011 1012 switch (ret & ~VM_FAULT_WRITE) { 1013 case VM_FAULT_MINOR: 1014 tsk->min_flt++; 1015 break; 1016 case VM_FAULT_MAJOR: 1017 tsk->maj_flt++; 1018 break; 1019 case VM_FAULT_SIGBUS: 1020 return i ? i : -EFAULT; 1021 case VM_FAULT_OOM: 1022 return i ? i : -ENOMEM; 1023 default: 1024 BUG(); 1025 } 1026 spin_lock(&mm->page_table_lock); 1027 } 1028 if (pages) { 1029 pages[i] = page; 1030 flush_dcache_page(page); 1031 page_cache_get(page); 1032 } 1033 if (vmas) 1034 vmas[i] = vma; 1035 i++; 1036 start += PAGE_SIZE; 1037 len--; 1038 } while (len && start < vma->vm_end); 1039 spin_unlock(&mm->page_table_lock); 1040 } while (len); 1041 return i; 1042} 1043EXPORT_SYMBOL(get_user_pages); 1044 1045static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1046 unsigned long addr, unsigned long end, pgprot_t prot) 1047{ 1048 pte_t *pte; 1049 spinlock_t *ptl; 1050 1051 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1052 if (!pte) 1053 return -ENOMEM; 1054 do { 1055 struct page *page = ZERO_PAGE(addr); 1056 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot)); 1057 page_cache_get(page); 1058 page_add_file_rmap(page); 1059 inc_mm_counter(mm, file_rss); 1060 BUG_ON(!pte_none(*pte)); 1061 set_pte_at(mm, addr, pte, zero_pte); 1062 } while (pte++, addr += PAGE_SIZE, addr != end); 1063 pte_unmap_unlock(pte - 1, ptl); 1064 return 0; 1065} 1066 1067static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud, 1068 unsigned long addr, unsigned long end, pgprot_t prot) 1069{ 1070 pmd_t *pmd; 1071 unsigned long next; 1072 1073 pmd = pmd_alloc(mm, pud, addr); 1074 if (!pmd) 1075 return -ENOMEM; 1076 do { 1077 next = pmd_addr_end(addr, end); 1078 if (zeromap_pte_range(mm, pmd, addr, next, prot)) 1079 return -ENOMEM; 1080 } while (pmd++, addr = next, addr != end); 1081 return 0; 1082} 1083 1084static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd, 1085 unsigned long addr, unsigned long end, pgprot_t prot) 1086{ 1087 pud_t *pud; 1088 unsigned long next; 1089 1090 pud = pud_alloc(mm, pgd, addr); 1091 if (!pud) 1092 return -ENOMEM; 1093 do { 1094 next = pud_addr_end(addr, end); 1095 if (zeromap_pmd_range(mm, pud, addr, next, prot)) 1096 return -ENOMEM; 1097 } while (pud++, addr = next, addr != end); 1098 return 0; 1099} 1100 1101int zeromap_page_range(struct vm_area_struct *vma, 1102 unsigned long addr, unsigned long size, pgprot_t prot) 1103{ 1104 pgd_t *pgd; 1105 unsigned long next; 1106 unsigned long end = addr + size; 1107 struct mm_struct *mm = vma->vm_mm; 1108 int err; 1109 1110 BUG_ON(addr >= end); 1111 pgd = pgd_offset(mm, addr); 1112 flush_cache_range(vma, addr, end); 1113 do { 1114 next = pgd_addr_end(addr, end); 1115 err = zeromap_pud_range(mm, pgd, addr, next, prot); 1116 if (err) 1117 break; 1118 } while (pgd++, addr = next, addr != end); 1119 return err; 1120} 1121 1122/* 1123 * maps a range of physical memory into the requested pages. the old 1124 * mappings are removed. any references to nonexistent pages results 1125 * in null mappings (currently treated as "copy-on-access") 1126 */ 1127static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1128 unsigned long addr, unsigned long end, 1129 unsigned long pfn, pgprot_t prot) 1130{ 1131 pte_t *pte; 1132 spinlock_t *ptl; 1133 1134 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1135 if (!pte) 1136 return -ENOMEM; 1137 do { 1138 BUG_ON(!pte_none(*pte)); 1139 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot)); 1140 pfn++; 1141 } while (pte++, addr += PAGE_SIZE, addr != end); 1142 pte_unmap_unlock(pte - 1, ptl); 1143 return 0; 1144} 1145 1146static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 1147 unsigned long addr, unsigned long end, 1148 unsigned long pfn, pgprot_t prot) 1149{ 1150 pmd_t *pmd; 1151 unsigned long next; 1152 1153 pfn -= addr >> PAGE_SHIFT; 1154 pmd = pmd_alloc(mm, pud, addr); 1155 if (!pmd) 1156 return -ENOMEM; 1157 do { 1158 next = pmd_addr_end(addr, end); 1159 if (remap_pte_range(mm, pmd, addr, next, 1160 pfn + (addr >> PAGE_SHIFT), prot)) 1161 return -ENOMEM; 1162 } while (pmd++, addr = next, addr != end); 1163 return 0; 1164} 1165 1166static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, 1167 unsigned long addr, unsigned long end, 1168 unsigned long pfn, pgprot_t prot) 1169{ 1170 pud_t *pud; 1171 unsigned long next; 1172 1173 pfn -= addr >> PAGE_SHIFT; 1174 pud = pud_alloc(mm, pgd, addr); 1175 if (!pud) 1176 return -ENOMEM; 1177 do { 1178 next = pud_addr_end(addr, end); 1179 if (remap_pmd_range(mm, pud, addr, next, 1180 pfn + (addr >> PAGE_SHIFT), prot)) 1181 return -ENOMEM; 1182 } while (pud++, addr = next, addr != end); 1183 return 0; 1184} 1185 1186/* Note: this is only safe if the mm semaphore is held when called. */ 1187int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 1188 unsigned long pfn, unsigned long size, pgprot_t prot) 1189{ 1190 pgd_t *pgd; 1191 unsigned long next; 1192 unsigned long end = addr + PAGE_ALIGN(size); 1193 struct mm_struct *mm = vma->vm_mm; 1194 int err; 1195 1196 /* 1197 * Physically remapped pages are special. Tell the 1198 * rest of the world about it: 1199 * VM_IO tells people not to look at these pages 1200 * (accesses can have side effects). 1201 * VM_RESERVED tells the core MM not to "manage" these pages 1202 * (e.g. refcount, mapcount, try to swap them out). 1203 */ 1204 vma->vm_flags |= VM_IO | VM_RESERVED; 1205 1206 BUG_ON(addr >= end); 1207 pfn -= addr >> PAGE_SHIFT; 1208 pgd = pgd_offset(mm, addr); 1209 flush_cache_range(vma, addr, end); 1210 do { 1211 next = pgd_addr_end(addr, end); 1212 err = remap_pud_range(mm, pgd, addr, next, 1213 pfn + (addr >> PAGE_SHIFT), prot); 1214 if (err) 1215 break; 1216 } while (pgd++, addr = next, addr != end); 1217 return err; 1218} 1219EXPORT_SYMBOL(remap_pfn_range); 1220 1221/* 1222 * handle_pte_fault chooses page fault handler according to an entry 1223 * which was read non-atomically. Before making any commitment, on 1224 * those architectures or configurations (e.g. i386 with PAE) which 1225 * might give a mix of unmatched parts, do_swap_page and do_file_page 1226 * must check under lock before unmapping the pte and proceeding 1227 * (but do_wp_page is only called after already making such a check; 1228 * and do_anonymous_page and do_no_page can safely check later on). 1229 */ 1230static inline int pte_unmap_same(struct mm_struct *mm, 1231 pte_t *page_table, pte_t orig_pte) 1232{ 1233 int same = 1; 1234#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 1235 if (sizeof(pte_t) > sizeof(unsigned long)) { 1236 spin_lock(&mm->page_table_lock); 1237 same = pte_same(*page_table, orig_pte); 1238 spin_unlock(&mm->page_table_lock); 1239 } 1240#endif 1241 pte_unmap(page_table); 1242 return same; 1243} 1244 1245/* 1246 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when 1247 * servicing faults for write access. In the normal case, do always want 1248 * pte_mkwrite. But get_user_pages can cause write faults for mappings 1249 * that do not have writing enabled, when used by access_process_vm. 1250 */ 1251static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma) 1252{ 1253 if (likely(vma->vm_flags & VM_WRITE)) 1254 pte = pte_mkwrite(pte); 1255 return pte; 1256} 1257 1258/* 1259 * This routine handles present pages, when users try to write 1260 * to a shared page. It is done by copying the page to a new address 1261 * and decrementing the shared-page counter for the old page. 1262 * 1263 * Note that this routine assumes that the protection checks have been 1264 * done by the caller (the low-level page fault routine in most cases). 1265 * Thus we can safely just mark it writable once we've done any necessary 1266 * COW. 1267 * 1268 * We also mark the page dirty at this point even though the page will 1269 * change only once the write actually happens. This avoids a few races, 1270 * and potentially makes it more efficient. 1271 * 1272 * We enter with non-exclusive mmap_sem (to exclude vma changes, 1273 * but allow concurrent faults), with pte both mapped and locked. 1274 * We return with mmap_sem still held, but pte unmapped and unlocked. 1275 */ 1276static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 1277 unsigned long address, pte_t *page_table, pmd_t *pmd, 1278 spinlock_t *ptl, pte_t orig_pte) 1279{ 1280 struct page *old_page, *new_page; 1281 unsigned long pfn = pte_pfn(orig_pte); 1282 pte_t entry; 1283 int ret = VM_FAULT_MINOR; 1284 1285 BUG_ON(vma->vm_flags & VM_RESERVED); 1286 1287 if (unlikely(!pfn_valid(pfn))) { 1288 /* 1289 * Page table corrupted: show pte and kill process. 1290 */ 1291 print_bad_pte(vma, orig_pte, address); 1292 ret = VM_FAULT_OOM; 1293 goto unlock; 1294 } 1295 old_page = pfn_to_page(pfn); 1296 1297 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) { 1298 int reuse = can_share_swap_page(old_page); 1299 unlock_page(old_page); 1300 if (reuse) { 1301 flush_cache_page(vma, address, pfn); 1302 entry = pte_mkyoung(orig_pte); 1303 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1304 ptep_set_access_flags(vma, address, page_table, entry, 1); 1305 update_mmu_cache(vma, address, entry); 1306 lazy_mmu_prot_update(entry); 1307 ret |= VM_FAULT_WRITE; 1308 goto unlock; 1309 } 1310 } 1311 1312 /* 1313 * Ok, we need to copy. Oh, well.. 1314 */ 1315 page_cache_get(old_page); 1316 pte_unmap_unlock(page_table, ptl); 1317 1318 if (unlikely(anon_vma_prepare(vma))) 1319 goto oom; 1320 if (old_page == ZERO_PAGE(address)) { 1321 new_page = alloc_zeroed_user_highpage(vma, address); 1322 if (!new_page) 1323 goto oom; 1324 } else { 1325 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address); 1326 if (!new_page) 1327 goto oom; 1328 copy_user_highpage(new_page, old_page, address); 1329 } 1330 1331 /* 1332 * Re-check the pte - we dropped the lock 1333 */ 1334 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 1335 if (likely(pte_same(*page_table, orig_pte))) { 1336 page_remove_rmap(old_page); 1337 if (!PageAnon(old_page)) { 1338 inc_mm_counter(mm, anon_rss); 1339 dec_mm_counter(mm, file_rss); 1340 } 1341 flush_cache_page(vma, address, pfn); 1342 entry = mk_pte(new_page, vma->vm_page_prot); 1343 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1344 ptep_establish(vma, address, page_table, entry); 1345 update_mmu_cache(vma, address, entry); 1346 lazy_mmu_prot_update(entry); 1347 lru_cache_add_active(new_page); 1348 page_add_anon_rmap(new_page, vma, address); 1349 1350 /* Free the old page.. */ 1351 new_page = old_page; 1352 ret |= VM_FAULT_WRITE; 1353 } 1354 page_cache_release(new_page); 1355 page_cache_release(old_page); 1356unlock: 1357 pte_unmap_unlock(page_table, ptl); 1358 return ret; 1359oom: 1360 page_cache_release(old_page); 1361 return VM_FAULT_OOM; 1362} 1363 1364/* 1365 * Helper functions for unmap_mapping_range(). 1366 * 1367 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __ 1368 * 1369 * We have to restart searching the prio_tree whenever we drop the lock, 1370 * since the iterator is only valid while the lock is held, and anyway 1371 * a later vma might be split and reinserted earlier while lock dropped. 1372 * 1373 * The list of nonlinear vmas could be handled more efficiently, using 1374 * a placeholder, but handle it in the same way until a need is shown. 1375 * It is important to search the prio_tree before nonlinear list: a vma 1376 * may become nonlinear and be shifted from prio_tree to nonlinear list 1377 * while the lock is dropped; but never shifted from list to prio_tree. 1378 * 1379 * In order to make forward progress despite restarting the search, 1380 * vm_truncate_count is used to mark a vma as now dealt with, so we can 1381 * quickly skip it next time around. Since the prio_tree search only 1382 * shows us those vmas affected by unmapping the range in question, we 1383 * can't efficiently keep all vmas in step with mapping->truncate_count: 1384 * so instead reset them all whenever it wraps back to 0 (then go to 1). 1385 * mapping->truncate_count and vma->vm_truncate_count are protected by 1386 * i_mmap_lock. 1387 * 1388 * In order to make forward progress despite repeatedly restarting some 1389 * large vma, note the restart_addr from unmap_vmas when it breaks out: 1390 * and restart from that address when we reach that vma again. It might 1391 * have been split or merged, shrunk or extended, but never shifted: so 1392 * restart_addr remains valid so long as it remains in the vma's range. 1393 * unmap_mapping_range forces truncate_count to leap over page-aligned 1394 * values so we can save vma's restart_addr in its truncate_count field. 1395 */ 1396#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK)) 1397 1398static void reset_vma_truncate_counts(struct address_space *mapping) 1399{ 1400 struct vm_area_struct *vma; 1401 struct prio_tree_iter iter; 1402 1403 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX) 1404 vma->vm_truncate_count = 0; 1405 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) 1406 vma->vm_truncate_count = 0; 1407} 1408 1409static int unmap_mapping_range_vma(struct vm_area_struct *vma, 1410 unsigned long start_addr, unsigned long end_addr, 1411 struct zap_details *details) 1412{ 1413 unsigned long restart_addr; 1414 int need_break; 1415 1416again: 1417 restart_addr = vma->vm_truncate_count; 1418 if (is_restart_addr(restart_addr) && start_addr < restart_addr) { 1419 start_addr = restart_addr; 1420 if (start_addr >= end_addr) { 1421 /* Top of vma has been split off since last time */ 1422 vma->vm_truncate_count = details->truncate_count; 1423 return 0; 1424 } 1425 } 1426 1427 restart_addr = zap_page_range(vma, start_addr, 1428 end_addr - start_addr, details); 1429 1430 /* 1431 * We cannot rely on the break test in unmap_vmas: 1432 * on the one hand, we don't want to restart our loop 1433 * just because that broke out for the page_table_lock; 1434 * on the other hand, it does no test when vma is small. 1435 */ 1436 need_break = need_resched() || 1437 need_lockbreak(details->i_mmap_lock); 1438 1439 if (restart_addr >= end_addr) { 1440 /* We have now completed this vma: mark it so */ 1441 vma->vm_truncate_count = details->truncate_count; 1442 if (!need_break) 1443 return 0; 1444 } else { 1445 /* Note restart_addr in vma's truncate_count field */ 1446 vma->vm_truncate_count = restart_addr; 1447 if (!need_break) 1448 goto again; 1449 } 1450 1451 spin_unlock(details->i_mmap_lock); 1452 cond_resched(); 1453 spin_lock(details->i_mmap_lock); 1454 return -EINTR; 1455} 1456 1457static inline void unmap_mapping_range_tree(struct prio_tree_root *root, 1458 struct zap_details *details) 1459{ 1460 struct vm_area_struct *vma; 1461 struct prio_tree_iter iter; 1462 pgoff_t vba, vea, zba, zea; 1463 1464restart: 1465 vma_prio_tree_foreach(vma, &iter, root, 1466 details->first_index, details->last_index) { 1467 /* Skip quickly over those we have already dealt with */ 1468 if (vma->vm_truncate_count == details->truncate_count) 1469 continue; 1470 1471 vba = vma->vm_pgoff; 1472 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1; 1473 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ 1474 zba = details->first_index; 1475 if (zba < vba) 1476 zba = vba; 1477 zea = details->last_index; 1478 if (zea > vea) 1479 zea = vea; 1480 1481 if (unmap_mapping_range_vma(vma, 1482 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 1483 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 1484 details) < 0) 1485 goto restart; 1486 } 1487} 1488 1489static inline void unmap_mapping_range_list(struct list_head *head, 1490 struct zap_details *details) 1491{ 1492 struct vm_area_struct *vma; 1493 1494 /* 1495 * In nonlinear VMAs there is no correspondence between virtual address 1496 * offset and file offset. So we must perform an exhaustive search 1497 * across *all* the pages in each nonlinear VMA, not just the pages 1498 * whose virtual address lies outside the file truncation point. 1499 */ 1500restart: 1501 list_for_each_entry(vma, head, shared.vm_set.list) { 1502 /* Skip quickly over those we have already dealt with */ 1503 if (vma->vm_truncate_count == details->truncate_count) 1504 continue; 1505 details->nonlinear_vma = vma; 1506 if (unmap_mapping_range_vma(vma, vma->vm_start, 1507 vma->vm_end, details) < 0) 1508 goto restart; 1509 } 1510} 1511 1512/** 1513 * unmap_mapping_range - unmap the portion of all mmaps 1514 * in the specified address_space corresponding to the specified 1515 * page range in the underlying file. 1516 * @mapping: the address space containing mmaps to be unmapped. 1517 * @holebegin: byte in first page to unmap, relative to the start of 1518 * the underlying file. This will be rounded down to a PAGE_SIZE 1519 * boundary. Note that this is different from vmtruncate(), which 1520 * must keep the partial page. In contrast, we must get rid of 1521 * partial pages. 1522 * @holelen: size of prospective hole in bytes. This will be rounded 1523 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 1524 * end of the file. 1525 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 1526 * but 0 when invalidating pagecache, don't throw away private data. 1527 */ 1528void unmap_mapping_range(struct address_space *mapping, 1529 loff_t const holebegin, loff_t const holelen, int even_cows) 1530{ 1531 struct zap_details details; 1532 pgoff_t hba = holebegin >> PAGE_SHIFT; 1533 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 1534 1535 /* Check for overflow. */ 1536 if (sizeof(holelen) > sizeof(hlen)) { 1537 long long holeend = 1538 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 1539 if (holeend & ~(long long)ULONG_MAX) 1540 hlen = ULONG_MAX - hba + 1; 1541 } 1542 1543 details.check_mapping = even_cows? NULL: mapping; 1544 details.nonlinear_vma = NULL; 1545 details.first_index = hba; 1546 details.last_index = hba + hlen - 1; 1547 if (details.last_index < details.first_index) 1548 details.last_index = ULONG_MAX; 1549 details.i_mmap_lock = &mapping->i_mmap_lock; 1550 1551 spin_lock(&mapping->i_mmap_lock); 1552 1553 /* serialize i_size write against truncate_count write */ 1554 smp_wmb(); 1555 /* Protect against page faults, and endless unmapping loops */ 1556 mapping->truncate_count++; 1557 /* 1558 * For archs where spin_lock has inclusive semantics like ia64 1559 * this smp_mb() will prevent to read pagetable contents 1560 * before the truncate_count increment is visible to 1561 * other cpus. 1562 */ 1563 smp_mb(); 1564 if (unlikely(is_restart_addr(mapping->truncate_count))) { 1565 if (mapping->truncate_count == 0) 1566 reset_vma_truncate_counts(mapping); 1567 mapping->truncate_count++; 1568 } 1569 details.truncate_count = mapping->truncate_count; 1570 1571 if (unlikely(!prio_tree_empty(&mapping->i_mmap))) 1572 unmap_mapping_range_tree(&mapping->i_mmap, &details); 1573 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) 1574 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details); 1575 spin_unlock(&mapping->i_mmap_lock); 1576} 1577EXPORT_SYMBOL(unmap_mapping_range); 1578 1579/* 1580 * Handle all mappings that got truncated by a "truncate()" 1581 * system call. 1582 * 1583 * NOTE! We have to be ready to update the memory sharing 1584 * between the file and the memory map for a potential last 1585 * incomplete page. Ugly, but necessary. 1586 */ 1587int vmtruncate(struct inode * inode, loff_t offset) 1588{ 1589 struct address_space *mapping = inode->i_mapping; 1590 unsigned long limit; 1591 1592 if (inode->i_size < offset) 1593 goto do_expand; 1594 /* 1595 * truncation of in-use swapfiles is disallowed - it would cause 1596 * subsequent swapout to scribble on the now-freed blocks. 1597 */ 1598 if (IS_SWAPFILE(inode)) 1599 goto out_busy; 1600 i_size_write(inode, offset); 1601 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1); 1602 truncate_inode_pages(mapping, offset); 1603 goto out_truncate; 1604 1605do_expand: 1606 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur; 1607 if (limit != RLIM_INFINITY && offset > limit) 1608 goto out_sig; 1609 if (offset > inode->i_sb->s_maxbytes) 1610 goto out_big; 1611 i_size_write(inode, offset); 1612 1613out_truncate: 1614 if (inode->i_op && inode->i_op->truncate) 1615 inode->i_op->truncate(inode); 1616 return 0; 1617out_sig: 1618 send_sig(SIGXFSZ, current, 0); 1619out_big: 1620 return -EFBIG; 1621out_busy: 1622 return -ETXTBSY; 1623} 1624 1625EXPORT_SYMBOL(vmtruncate); 1626 1627/* 1628 * Primitive swap readahead code. We simply read an aligned block of 1629 * (1 << page_cluster) entries in the swap area. This method is chosen 1630 * because it doesn't cost us any seek time. We also make sure to queue 1631 * the 'original' request together with the readahead ones... 1632 * 1633 * This has been extended to use the NUMA policies from the mm triggering 1634 * the readahead. 1635 * 1636 * Caller must hold down_read on the vma->vm_mm if vma is not NULL. 1637 */ 1638void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma) 1639{ 1640#ifdef CONFIG_NUMA 1641 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL; 1642#endif 1643 int i, num; 1644 struct page *new_page; 1645 unsigned long offset; 1646 1647 /* 1648 * Get the number of handles we should do readahead io to. 1649 */ 1650 num = valid_swaphandles(entry, &offset); 1651 for (i = 0; i < num; offset++, i++) { 1652 /* Ok, do the async read-ahead now */ 1653 new_page = read_swap_cache_async(swp_entry(swp_type(entry), 1654 offset), vma, addr); 1655 if (!new_page) 1656 break; 1657 page_cache_release(new_page); 1658#ifdef CONFIG_NUMA 1659 /* 1660 * Find the next applicable VMA for the NUMA policy. 1661 */ 1662 addr += PAGE_SIZE; 1663 if (addr == 0) 1664 vma = NULL; 1665 if (vma) { 1666 if (addr >= vma->vm_end) { 1667 vma = next_vma; 1668 next_vma = vma ? vma->vm_next : NULL; 1669 } 1670 if (vma && addr < vma->vm_start) 1671 vma = NULL; 1672 } else { 1673 if (next_vma && addr >= next_vma->vm_start) { 1674 vma = next_vma; 1675 next_vma = vma->vm_next; 1676 } 1677 } 1678#endif 1679 } 1680 lru_add_drain(); /* Push any new pages onto the LRU now */ 1681} 1682 1683/* 1684 * We enter with non-exclusive mmap_sem (to exclude vma changes, 1685 * but allow concurrent faults), and pte mapped but not yet locked. 1686 * We return with mmap_sem still held, but pte unmapped and unlocked. 1687 */ 1688static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, 1689 unsigned long address, pte_t *page_table, pmd_t *pmd, 1690 int write_access, pte_t orig_pte) 1691{ 1692 spinlock_t *ptl; 1693 struct page *page; 1694 swp_entry_t entry; 1695 pte_t pte; 1696 int ret = VM_FAULT_MINOR; 1697 1698 if (!pte_unmap_same(mm, page_table, orig_pte)) 1699 goto out; 1700 1701 entry = pte_to_swp_entry(orig_pte); 1702 page = lookup_swap_cache(entry); 1703 if (!page) { 1704 swapin_readahead(entry, address, vma); 1705 page = read_swap_cache_async(entry, vma, address); 1706 if (!page) { 1707 /* 1708 * Back out if somebody else faulted in this pte 1709 * while we released the pte lock. 1710 */ 1711 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 1712 if (likely(pte_same(*page_table, orig_pte))) 1713 ret = VM_FAULT_OOM; 1714 goto unlock; 1715 } 1716 1717 /* Had to read the page from swap area: Major fault */ 1718 ret = VM_FAULT_MAJOR; 1719 inc_page_state(pgmajfault); 1720 grab_swap_token(); 1721 } 1722 1723 mark_page_accessed(page); 1724 lock_page(page); 1725 1726 /* 1727 * Back out if somebody else already faulted in this pte. 1728 */ 1729 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 1730 if (unlikely(!pte_same(*page_table, orig_pte))) 1731 goto out_nomap; 1732 1733 if (unlikely(!PageUptodate(page))) { 1734 ret = VM_FAULT_SIGBUS; 1735 goto out_nomap; 1736 } 1737 1738 /* The page isn't present yet, go ahead with the fault. */ 1739 1740 inc_mm_counter(mm, anon_rss); 1741 pte = mk_pte(page, vma->vm_page_prot); 1742 if (write_access && can_share_swap_page(page)) { 1743 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 1744 write_access = 0; 1745 } 1746 1747 flush_icache_page(vma, page); 1748 set_pte_at(mm, address, page_table, pte); 1749 page_add_anon_rmap(page, vma, address); 1750 1751 swap_free(entry); 1752 if (vm_swap_full()) 1753 remove_exclusive_swap_page(page); 1754 unlock_page(page); 1755 1756 if (write_access) { 1757 if (do_wp_page(mm, vma, address, 1758 page_table, pmd, ptl, pte) == VM_FAULT_OOM) 1759 ret = VM_FAULT_OOM; 1760 goto out; 1761 } 1762 1763 /* No need to invalidate - it was non-present before */ 1764 update_mmu_cache(vma, address, pte); 1765 lazy_mmu_prot_update(pte); 1766unlock: 1767 pte_unmap_unlock(page_table, ptl); 1768out: 1769 return ret; 1770out_nomap: 1771 pte_unmap_unlock(page_table, ptl); 1772 unlock_page(page); 1773 page_cache_release(page); 1774 return ret; 1775} 1776 1777/* 1778 * We enter with non-exclusive mmap_sem (to exclude vma changes, 1779 * but allow concurrent faults), and pte mapped but not yet locked. 1780 * We return with mmap_sem still held, but pte unmapped and unlocked. 1781 */ 1782static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 1783 unsigned long address, pte_t *page_table, pmd_t *pmd, 1784 int write_access) 1785{ 1786 struct page *page; 1787 spinlock_t *ptl; 1788 pte_t entry; 1789 1790 if (write_access) { 1791 /* Allocate our own private page. */ 1792 pte_unmap(page_table); 1793 1794 if (unlikely(anon_vma_prepare(vma))) 1795 goto oom; 1796 page = alloc_zeroed_user_highpage(vma, address); 1797 if (!page) 1798 goto oom; 1799 1800 entry = mk_pte(page, vma->vm_page_prot); 1801 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1802 1803 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 1804 if (!pte_none(*page_table)) 1805 goto release; 1806 inc_mm_counter(mm, anon_rss); 1807 lru_cache_add_active(page); 1808 SetPageReferenced(page); 1809 page_add_anon_rmap(page, vma, address); 1810 } else { 1811 /* Map the ZERO_PAGE - vm_page_prot is readonly */ 1812 page = ZERO_PAGE(address); 1813 page_cache_get(page); 1814 entry = mk_pte(page, vma->vm_page_prot); 1815 1816 ptl = &mm->page_table_lock; 1817 spin_lock(ptl); 1818 if (!pte_none(*page_table)) 1819 goto release; 1820 inc_mm_counter(mm, file_rss); 1821 page_add_file_rmap(page); 1822 } 1823 1824 set_pte_at(mm, address, page_table, entry); 1825 1826 /* No need to invalidate - it was non-present before */ 1827 update_mmu_cache(vma, address, entry); 1828 lazy_mmu_prot_update(entry); 1829unlock: 1830 pte_unmap_unlock(page_table, ptl); 1831 return VM_FAULT_MINOR; 1832release: 1833 page_cache_release(page); 1834 goto unlock; 1835oom: 1836 return VM_FAULT_OOM; 1837} 1838 1839/* 1840 * do_no_page() tries to create a new page mapping. It aggressively 1841 * tries to share with existing pages, but makes a separate copy if 1842 * the "write_access" parameter is true in order to avoid the next 1843 * page fault. 1844 * 1845 * As this is called only for pages that do not currently exist, we 1846 * do not need to flush old virtual caches or the TLB. 1847 * 1848 * We enter with non-exclusive mmap_sem (to exclude vma changes, 1849 * but allow concurrent faults), and pte mapped but not yet locked. 1850 * We return with mmap_sem still held, but pte unmapped and unlocked. 1851 */ 1852static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 1853 unsigned long address, pte_t *page_table, pmd_t *pmd, 1854 int write_access) 1855{ 1856 spinlock_t *ptl; 1857 struct page *new_page; 1858 struct address_space *mapping = NULL; 1859 pte_t entry; 1860 unsigned int sequence = 0; 1861 int ret = VM_FAULT_MINOR; 1862 int anon = 0; 1863 1864 pte_unmap(page_table); 1865 1866 if (vma->vm_file) { 1867 mapping = vma->vm_file->f_mapping; 1868 sequence = mapping->truncate_count; 1869 smp_rmb(); /* serializes i_size against truncate_count */ 1870 } 1871retry: 1872 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret); 1873 /* 1874 * No smp_rmb is needed here as long as there's a full 1875 * spin_lock/unlock sequence inside the ->nopage callback 1876 * (for the pagecache lookup) that acts as an implicit 1877 * smp_mb() and prevents the i_size read to happen 1878 * after the next truncate_count read. 1879 */ 1880 1881 /* no page was available -- either SIGBUS or OOM */ 1882 if (new_page == NOPAGE_SIGBUS) 1883 return VM_FAULT_SIGBUS; 1884 if (new_page == NOPAGE_OOM) 1885 return VM_FAULT_OOM; 1886 1887 /* 1888 * Should we do an early C-O-W break? 1889 */ 1890 if (write_access && !(vma->vm_flags & VM_SHARED)) { 1891 struct page *page; 1892 1893 if (unlikely(anon_vma_prepare(vma))) 1894 goto oom; 1895 page = alloc_page_vma(GFP_HIGHUSER, vma, address); 1896 if (!page) 1897 goto oom; 1898 copy_user_highpage(page, new_page, address); 1899 page_cache_release(new_page); 1900 new_page = page; 1901 anon = 1; 1902 } 1903 1904 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 1905 /* 1906 * For a file-backed vma, someone could have truncated or otherwise 1907 * invalidated this page. If unmap_mapping_range got called, 1908 * retry getting the page. 1909 */ 1910 if (mapping && unlikely(sequence != mapping->truncate_count)) { 1911 pte_unmap_unlock(page_table, ptl); 1912 page_cache_release(new_page); 1913 cond_resched(); 1914 sequence = mapping->truncate_count; 1915 smp_rmb(); 1916 goto retry; 1917 } 1918 1919 /* 1920 * This silly early PAGE_DIRTY setting removes a race 1921 * due to the bad i386 page protection. But it's valid 1922 * for other architectures too. 1923 * 1924 * Note that if write_access is true, we either now have 1925 * an exclusive copy of the page, or this is a shared mapping, 1926 * so we can make it writable and dirty to avoid having to 1927 * handle that later. 1928 */ 1929 /* Only go through if we didn't race with anybody else... */ 1930 if (pte_none(*page_table)) { 1931 flush_icache_page(vma, new_page); 1932 entry = mk_pte(new_page, vma->vm_page_prot); 1933 if (write_access) 1934 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1935 set_pte_at(mm, address, page_table, entry); 1936 if (anon) { 1937 inc_mm_counter(mm, anon_rss); 1938 lru_cache_add_active(new_page); 1939 page_add_anon_rmap(new_page, vma, address); 1940 } else if (!(vma->vm_flags & VM_RESERVED)) { 1941 inc_mm_counter(mm, file_rss); 1942 page_add_file_rmap(new_page); 1943 } 1944 } else { 1945 /* One of our sibling threads was faster, back out. */ 1946 page_cache_release(new_page); 1947 goto unlock; 1948 } 1949 1950 /* no need to invalidate: a not-present page shouldn't be cached */ 1951 update_mmu_cache(vma, address, entry); 1952 lazy_mmu_prot_update(entry); 1953unlock: 1954 pte_unmap_unlock(page_table, ptl); 1955 return ret; 1956oom: 1957 page_cache_release(new_page); 1958 return VM_FAULT_OOM; 1959} 1960 1961/* 1962 * Fault of a previously existing named mapping. Repopulate the pte 1963 * from the encoded file_pte if possible. This enables swappable 1964 * nonlinear vmas. 1965 * 1966 * We enter with non-exclusive mmap_sem (to exclude vma changes, 1967 * but allow concurrent faults), and pte mapped but not yet locked. 1968 * We return with mmap_sem still held, but pte unmapped and unlocked. 1969 */ 1970static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma, 1971 unsigned long address, pte_t *page_table, pmd_t *pmd, 1972 int write_access, pte_t orig_pte) 1973{ 1974 pgoff_t pgoff; 1975 int err; 1976 1977 if (!pte_unmap_same(mm, page_table, orig_pte)) 1978 return VM_FAULT_MINOR; 1979 1980 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) { 1981 /* 1982 * Page table corrupted: show pte and kill process. 1983 */ 1984 print_bad_pte(vma, orig_pte, address); 1985 return VM_FAULT_OOM; 1986 } 1987 /* We can then assume vm->vm_ops && vma->vm_ops->populate */ 1988 1989 pgoff = pte_to_pgoff(orig_pte); 1990 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, 1991 vma->vm_page_prot, pgoff, 0); 1992 if (err == -ENOMEM) 1993 return VM_FAULT_OOM; 1994 if (err) 1995 return VM_FAULT_SIGBUS; 1996 return VM_FAULT_MAJOR; 1997} 1998 1999/* 2000 * These routines also need to handle stuff like marking pages dirty 2001 * and/or accessed for architectures that don't do it in hardware (most 2002 * RISC architectures). The early dirtying is also good on the i386. 2003 * 2004 * There is also a hook called "update_mmu_cache()" that architectures 2005 * with external mmu caches can use to update those (ie the Sparc or 2006 * PowerPC hashed page tables that act as extended TLBs). 2007 * 2008 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2009 * but allow concurrent faults), and pte mapped but not yet locked. 2010 * We return with mmap_sem still held, but pte unmapped and unlocked. 2011 */ 2012static inline int handle_pte_fault(struct mm_struct *mm, 2013 struct vm_area_struct *vma, unsigned long address, 2014 pte_t *pte, pmd_t *pmd, int write_access) 2015{ 2016 pte_t entry; 2017 spinlock_t *ptl; 2018 2019 entry = *pte; 2020 if (!pte_present(entry)) { 2021 if (pte_none(entry)) { 2022 if (!vma->vm_ops || !vma->vm_ops->nopage) 2023 return do_anonymous_page(mm, vma, address, 2024 pte, pmd, write_access); 2025 return do_no_page(mm, vma, address, 2026 pte, pmd, write_access); 2027 } 2028 if (pte_file(entry)) 2029 return do_file_page(mm, vma, address, 2030 pte, pmd, write_access, entry); 2031 return do_swap_page(mm, vma, address, 2032 pte, pmd, write_access, entry); 2033 } 2034 2035 ptl = &mm->page_table_lock; 2036 spin_lock(ptl); 2037 if (unlikely(!pte_same(*pte, entry))) 2038 goto unlock; 2039 if (write_access) { 2040 if (!pte_write(entry)) 2041 return do_wp_page(mm, vma, address, 2042 pte, pmd, ptl, entry); 2043 entry = pte_mkdirty(entry); 2044 } 2045 entry = pte_mkyoung(entry); 2046 ptep_set_access_flags(vma, address, pte, entry, write_access); 2047 update_mmu_cache(vma, address, entry); 2048 lazy_mmu_prot_update(entry); 2049unlock: 2050 pte_unmap_unlock(pte, ptl); 2051 return VM_FAULT_MINOR; 2052} 2053 2054/* 2055 * By the time we get here, we already hold the mm semaphore 2056 */ 2057int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2058 unsigned long address, int write_access) 2059{ 2060 pgd_t *pgd; 2061 pud_t *pud; 2062 pmd_t *pmd; 2063 pte_t *pte; 2064 2065 __set_current_state(TASK_RUNNING); 2066 2067 inc_page_state(pgfault); 2068 2069 if (unlikely(is_vm_hugetlb_page(vma))) 2070 return hugetlb_fault(mm, vma, address, write_access); 2071 2072 pgd = pgd_offset(mm, address); 2073 pud = pud_alloc(mm, pgd, address); 2074 if (!pud) 2075 return VM_FAULT_OOM; 2076 pmd = pmd_alloc(mm, pud, address); 2077 if (!pmd) 2078 return VM_FAULT_OOM; 2079 pte = pte_alloc_map(mm, pmd, address); 2080 if (!pte) 2081 return VM_FAULT_OOM; 2082 2083 return handle_pte_fault(mm, vma, address, pte, pmd, write_access); 2084} 2085 2086#ifndef __PAGETABLE_PUD_FOLDED 2087/* 2088 * Allocate page upper directory. 2089 * We've already handled the fast-path in-line. 2090 */ 2091int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 2092{ 2093 pud_t *new = pud_alloc_one(mm, address); 2094 if (!new) 2095 return -ENOMEM; 2096 2097 spin_lock(&mm->page_table_lock); 2098 if (pgd_present(*pgd)) /* Another has populated it */ 2099 pud_free(new); 2100 else 2101 pgd_populate(mm, pgd, new); 2102 spin_unlock(&mm->page_table_lock); 2103 return 0; 2104} 2105#endif /* __PAGETABLE_PUD_FOLDED */ 2106 2107#ifndef __PAGETABLE_PMD_FOLDED 2108/* 2109 * Allocate page middle directory. 2110 * We've already handled the fast-path in-line. 2111 */ 2112int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 2113{ 2114 pmd_t *new = pmd_alloc_one(mm, address); 2115 if (!new) 2116 return -ENOMEM; 2117 2118 spin_lock(&mm->page_table_lock); 2119#ifndef __ARCH_HAS_4LEVEL_HACK 2120 if (pud_present(*pud)) /* Another has populated it */ 2121 pmd_free(new); 2122 else 2123 pud_populate(mm, pud, new); 2124#else 2125 if (pgd_present(*pud)) /* Another has populated it */ 2126 pmd_free(new); 2127 else 2128 pgd_populate(mm, pud, new); 2129#endif /* __ARCH_HAS_4LEVEL_HACK */ 2130 spin_unlock(&mm->page_table_lock); 2131 return 0; 2132} 2133#endif /* __PAGETABLE_PMD_FOLDED */ 2134 2135int make_pages_present(unsigned long addr, unsigned long end) 2136{ 2137 int ret, len, write; 2138 struct vm_area_struct * vma; 2139 2140 vma = find_vma(current->mm, addr); 2141 if (!vma) 2142 return -1; 2143 write = (vma->vm_flags & VM_WRITE) != 0; 2144 if (addr >= end) 2145 BUG(); 2146 if (end > vma->vm_end) 2147 BUG(); 2148 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE; 2149 ret = get_user_pages(current, current->mm, addr, 2150 len, write, 0, NULL, NULL); 2151 if (ret < 0) 2152 return ret; 2153 return ret == len ? 0 : -1; 2154} 2155 2156/* 2157 * Map a vmalloc()-space virtual address to the physical page. 2158 */ 2159struct page * vmalloc_to_page(void * vmalloc_addr) 2160{ 2161 unsigned long addr = (unsigned long) vmalloc_addr; 2162 struct page *page = NULL; 2163 pgd_t *pgd = pgd_offset_k(addr); 2164 pud_t *pud; 2165 pmd_t *pmd; 2166 pte_t *ptep, pte; 2167 2168 if (!pgd_none(*pgd)) { 2169 pud = pud_offset(pgd, addr); 2170 if (!pud_none(*pud)) { 2171 pmd = pmd_offset(pud, addr); 2172 if (!pmd_none(*pmd)) { 2173 ptep = pte_offset_map(pmd, addr); 2174 pte = *ptep; 2175 if (pte_present(pte)) 2176 page = pte_page(pte); 2177 pte_unmap(ptep); 2178 } 2179 } 2180 } 2181 return page; 2182} 2183 2184EXPORT_SYMBOL(vmalloc_to_page); 2185 2186/* 2187 * Map a vmalloc()-space virtual address to the physical page frame number. 2188 */ 2189unsigned long vmalloc_to_pfn(void * vmalloc_addr) 2190{ 2191 return page_to_pfn(vmalloc_to_page(vmalloc_addr)); 2192} 2193 2194EXPORT_SYMBOL(vmalloc_to_pfn); 2195 2196#if !defined(__HAVE_ARCH_GATE_AREA) 2197 2198#if defined(AT_SYSINFO_EHDR) 2199static struct vm_area_struct gate_vma; 2200 2201static int __init gate_vma_init(void) 2202{ 2203 gate_vma.vm_mm = NULL; 2204 gate_vma.vm_start = FIXADDR_USER_START; 2205 gate_vma.vm_end = FIXADDR_USER_END; 2206 gate_vma.vm_page_prot = PAGE_READONLY; 2207 gate_vma.vm_flags = VM_RESERVED; 2208 return 0; 2209} 2210__initcall(gate_vma_init); 2211#endif 2212 2213struct vm_area_struct *get_gate_vma(struct task_struct *tsk) 2214{ 2215#ifdef AT_SYSINFO_EHDR 2216 return &gate_vma; 2217#else 2218 return NULL; 2219#endif 2220} 2221 2222int in_gate_area_no_task(unsigned long addr) 2223{ 2224#ifdef AT_SYSINFO_EHDR 2225 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END)) 2226 return 1; 2227#endif 2228 return 0; 2229} 2230 2231#endif /* __HAVE_ARCH_GATE_AREA */ 2232