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/ksm.h> 49 #include <linux/rmap.h> 50 #include <linux/export.h> 51 #include <linux/delayacct.h> 52 #include <linux/init.h> 53 #include <linux/writeback.h> 54 #include <linux/memcontrol.h> 55 #include <linux/mmu_notifier.h> 56 #include <linux/kallsyms.h> 57 #include <linux/swapops.h> 58 #include <linux/elf.h> 59 #include <linux/gfp.h> 60 61 #include <asm/io.h> 62 #include <asm/pgalloc.h> 63 #include <asm/uaccess.h> 64 #include <asm/tlb.h> 65 #include <asm/tlbflush.h> 66 #include <asm/pgtable.h> 67 68 #include "internal.h" 69 70 #ifndef CONFIG_NEED_MULTIPLE_NODES 71 /* use the per-pgdat data instead for discontigmem - mbligh */ 72 unsigned long max_mapnr; 73 struct page *mem_map; 74 75 EXPORT_SYMBOL(max_mapnr); 76 EXPORT_SYMBOL(mem_map); 77 #endif 78 79 unsigned long num_physpages; 80 /* 81 * A number of key systems in x86 including ioremap() rely on the assumption 82 * that high_memory defines the upper bound on direct map memory, then end 83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 85 * and ZONE_HIGHMEM. 86 */ 87 void * high_memory; 88 89 EXPORT_SYMBOL(num_physpages); 90 EXPORT_SYMBOL(high_memory); 91 92 /* 93 * Randomize the address space (stacks, mmaps, brk, etc.). 94 * 95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 96 * as ancient (libc5 based) binaries can segfault. ) 97 */ 98 int randomize_va_space __read_mostly = 99 #ifdef CONFIG_COMPAT_BRK 100 1; 101 #else 102 2; 103 #endif 104 105 static int __init disable_randmaps(char *s) 106 { 107 randomize_va_space = 0; 108 return 1; 109 } 110 __setup("norandmaps", disable_randmaps); 111 112 unsigned long zero_pfn __read_mostly; 113 unsigned long highest_memmap_pfn __read_mostly; 114 115 /* 116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 117 */ 118 static int __init init_zero_pfn(void) 119 { 120 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 121 return 0; 122 } 123 core_initcall(init_zero_pfn); 124 125 126 #if defined(SPLIT_RSS_COUNTING) 127 128 void sync_mm_rss(struct mm_struct *mm) 129 { 130 int i; 131 132 for (i = 0; i < NR_MM_COUNTERS; i++) { 133 if (current->rss_stat.count[i]) { 134 add_mm_counter(mm, i, current->rss_stat.count[i]); 135 current->rss_stat.count[i] = 0; 136 } 137 } 138 current->rss_stat.events = 0; 139 } 140 141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 142 { 143 struct task_struct *task = current; 144 145 if (likely(task->mm == mm)) 146 task->rss_stat.count[member] += val; 147 else 148 add_mm_counter(mm, member, val); 149 } 150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 152 153 /* sync counter once per 64 page faults */ 154 #define TASK_RSS_EVENTS_THRESH (64) 155 static void check_sync_rss_stat(struct task_struct *task) 156 { 157 if (unlikely(task != current)) 158 return; 159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 160 sync_mm_rss(task->mm); 161 } 162 #else /* SPLIT_RSS_COUNTING */ 163 164 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 165 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 166 167 static void check_sync_rss_stat(struct task_struct *task) 168 { 169 } 170 171 #endif /* SPLIT_RSS_COUNTING */ 172 173 #ifdef HAVE_GENERIC_MMU_GATHER 174 175 static int tlb_next_batch(struct mmu_gather *tlb) 176 { 177 struct mmu_gather_batch *batch; 178 179 batch = tlb->active; 180 if (batch->next) { 181 tlb->active = batch->next; 182 return 1; 183 } 184 185 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0); 186 if (!batch) 187 return 0; 188 189 batch->next = NULL; 190 batch->nr = 0; 191 batch->max = MAX_GATHER_BATCH; 192 193 tlb->active->next = batch; 194 tlb->active = batch; 195 196 return 1; 197 } 198 199 /* tlb_gather_mmu 200 * Called to initialize an (on-stack) mmu_gather structure for page-table 201 * tear-down from @mm. The @fullmm argument is used when @mm is without 202 * users and we're going to destroy the full address space (exit/execve). 203 */ 204 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm) 205 { 206 tlb->mm = mm; 207 208 tlb->fullmm = fullmm; 209 tlb->need_flush = 0; 210 tlb->fast_mode = (num_possible_cpus() == 1); 211 tlb->local.next = NULL; 212 tlb->local.nr = 0; 213 tlb->local.max = ARRAY_SIZE(tlb->__pages); 214 tlb->active = &tlb->local; 215 216 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 217 tlb->batch = NULL; 218 #endif 219 } 220 221 void tlb_flush_mmu(struct mmu_gather *tlb) 222 { 223 struct mmu_gather_batch *batch; 224 225 if (!tlb->need_flush) 226 return; 227 tlb->need_flush = 0; 228 tlb_flush(tlb); 229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 230 tlb_table_flush(tlb); 231 #endif 232 233 if (tlb_fast_mode(tlb)) 234 return; 235 236 for (batch = &tlb->local; batch; batch = batch->next) { 237 free_pages_and_swap_cache(batch->pages, batch->nr); 238 batch->nr = 0; 239 } 240 tlb->active = &tlb->local; 241 } 242 243 /* tlb_finish_mmu 244 * Called at the end of the shootdown operation to free up any resources 245 * that were required. 246 */ 247 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) 248 { 249 struct mmu_gather_batch *batch, *next; 250 251 tlb_flush_mmu(tlb); 252 253 /* keep the page table cache within bounds */ 254 check_pgt_cache(); 255 256 for (batch = tlb->local.next; batch; batch = next) { 257 next = batch->next; 258 free_pages((unsigned long)batch, 0); 259 } 260 tlb->local.next = NULL; 261 } 262 263 /* __tlb_remove_page 264 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while 265 * handling the additional races in SMP caused by other CPUs caching valid 266 * mappings in their TLBs. Returns the number of free page slots left. 267 * When out of page slots we must call tlb_flush_mmu(). 268 */ 269 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) 270 { 271 struct mmu_gather_batch *batch; 272 273 VM_BUG_ON(!tlb->need_flush); 274 275 if (tlb_fast_mode(tlb)) { 276 free_page_and_swap_cache(page); 277 return 1; /* avoid calling tlb_flush_mmu() */ 278 } 279 280 batch = tlb->active; 281 batch->pages[batch->nr++] = page; 282 if (batch->nr == batch->max) { 283 if (!tlb_next_batch(tlb)) 284 return 0; 285 batch = tlb->active; 286 } 287 VM_BUG_ON(batch->nr > batch->max); 288 289 return batch->max - batch->nr; 290 } 291 292 #endif /* HAVE_GENERIC_MMU_GATHER */ 293 294 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 295 296 /* 297 * See the comment near struct mmu_table_batch. 298 */ 299 300 static void tlb_remove_table_smp_sync(void *arg) 301 { 302 /* Simply deliver the interrupt */ 303 } 304 305 static void tlb_remove_table_one(void *table) 306 { 307 /* 308 * This isn't an RCU grace period and hence the page-tables cannot be 309 * assumed to be actually RCU-freed. 310 * 311 * It is however sufficient for software page-table walkers that rely on 312 * IRQ disabling. See the comment near struct mmu_table_batch. 313 */ 314 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 315 __tlb_remove_table(table); 316 } 317 318 static void tlb_remove_table_rcu(struct rcu_head *head) 319 { 320 struct mmu_table_batch *batch; 321 int i; 322 323 batch = container_of(head, struct mmu_table_batch, rcu); 324 325 for (i = 0; i < batch->nr; i++) 326 __tlb_remove_table(batch->tables[i]); 327 328 free_page((unsigned long)batch); 329 } 330 331 void tlb_table_flush(struct mmu_gather *tlb) 332 { 333 struct mmu_table_batch **batch = &tlb->batch; 334 335 if (*batch) { 336 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 337 *batch = NULL; 338 } 339 } 340 341 void tlb_remove_table(struct mmu_gather *tlb, void *table) 342 { 343 struct mmu_table_batch **batch = &tlb->batch; 344 345 tlb->need_flush = 1; 346 347 /* 348 * When there's less then two users of this mm there cannot be a 349 * concurrent page-table walk. 350 */ 351 if (atomic_read(&tlb->mm->mm_users) < 2) { 352 __tlb_remove_table(table); 353 return; 354 } 355 356 if (*batch == NULL) { 357 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); 358 if (*batch == NULL) { 359 tlb_remove_table_one(table); 360 return; 361 } 362 (*batch)->nr = 0; 363 } 364 (*batch)->tables[(*batch)->nr++] = table; 365 if ((*batch)->nr == MAX_TABLE_BATCH) 366 tlb_table_flush(tlb); 367 } 368 369 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ 370 371 /* 372 * If a p?d_bad entry is found while walking page tables, report 373 * the error, before resetting entry to p?d_none. Usually (but 374 * very seldom) called out from the p?d_none_or_clear_bad macros. 375 */ 376 377 void pgd_clear_bad(pgd_t *pgd) 378 { 379 pgd_ERROR(*pgd); 380 pgd_clear(pgd); 381 } 382 383 void pud_clear_bad(pud_t *pud) 384 { 385 pud_ERROR(*pud); 386 pud_clear(pud); 387 } 388 389 void pmd_clear_bad(pmd_t *pmd) 390 { 391 pmd_ERROR(*pmd); 392 pmd_clear(pmd); 393 } 394 395 /* 396 * Note: this doesn't free the actual pages themselves. That 397 * has been handled earlier when unmapping all the memory regions. 398 */ 399 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 400 unsigned long addr) 401 { 402 pgtable_t token = pmd_pgtable(*pmd); 403 pmd_clear(pmd); 404 pte_free_tlb(tlb, token, addr); 405 tlb->mm->nr_ptes--; 406 } 407 408 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 409 unsigned long addr, unsigned long end, 410 unsigned long floor, unsigned long ceiling) 411 { 412 pmd_t *pmd; 413 unsigned long next; 414 unsigned long start; 415 416 start = addr; 417 pmd = pmd_offset(pud, addr); 418 do { 419 next = pmd_addr_end(addr, end); 420 if (pmd_none_or_clear_bad(pmd)) 421 continue; 422 free_pte_range(tlb, pmd, addr); 423 } while (pmd++, addr = next, addr != end); 424 425 start &= PUD_MASK; 426 if (start < floor) 427 return; 428 if (ceiling) { 429 ceiling &= PUD_MASK; 430 if (!ceiling) 431 return; 432 } 433 if (end - 1 > ceiling - 1) 434 return; 435 436 pmd = pmd_offset(pud, start); 437 pud_clear(pud); 438 pmd_free_tlb(tlb, pmd, start); 439 } 440 441 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 442 unsigned long addr, unsigned long end, 443 unsigned long floor, unsigned long ceiling) 444 { 445 pud_t *pud; 446 unsigned long next; 447 unsigned long start; 448 449 start = addr; 450 pud = pud_offset(pgd, addr); 451 do { 452 next = pud_addr_end(addr, end); 453 if (pud_none_or_clear_bad(pud)) 454 continue; 455 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 456 } while (pud++, addr = next, addr != end); 457 458 start &= PGDIR_MASK; 459 if (start < floor) 460 return; 461 if (ceiling) { 462 ceiling &= PGDIR_MASK; 463 if (!ceiling) 464 return; 465 } 466 if (end - 1 > ceiling - 1) 467 return; 468 469 pud = pud_offset(pgd, start); 470 pgd_clear(pgd); 471 pud_free_tlb(tlb, pud, start); 472 } 473 474 /* 475 * This function frees user-level page tables of a process. 476 * 477 * Must be called with pagetable lock held. 478 */ 479 void free_pgd_range(struct mmu_gather *tlb, 480 unsigned long addr, unsigned long end, 481 unsigned long floor, unsigned long ceiling) 482 { 483 pgd_t *pgd; 484 unsigned long next; 485 486 /* 487 * The next few lines have given us lots of grief... 488 * 489 * Why are we testing PMD* at this top level? Because often 490 * there will be no work to do at all, and we'd prefer not to 491 * go all the way down to the bottom just to discover that. 492 * 493 * Why all these "- 1"s? Because 0 represents both the bottom 494 * of the address space and the top of it (using -1 for the 495 * top wouldn't help much: the masks would do the wrong thing). 496 * The rule is that addr 0 and floor 0 refer to the bottom of 497 * the address space, but end 0 and ceiling 0 refer to the top 498 * Comparisons need to use "end - 1" and "ceiling - 1" (though 499 * that end 0 case should be mythical). 500 * 501 * Wherever addr is brought up or ceiling brought down, we must 502 * be careful to reject "the opposite 0" before it confuses the 503 * subsequent tests. But what about where end is brought down 504 * by PMD_SIZE below? no, end can't go down to 0 there. 505 * 506 * Whereas we round start (addr) and ceiling down, by different 507 * masks at different levels, in order to test whether a table 508 * now has no other vmas using it, so can be freed, we don't 509 * bother to round floor or end up - the tests don't need that. 510 */ 511 512 addr &= PMD_MASK; 513 if (addr < floor) { 514 addr += PMD_SIZE; 515 if (!addr) 516 return; 517 } 518 if (ceiling) { 519 ceiling &= PMD_MASK; 520 if (!ceiling) 521 return; 522 } 523 if (end - 1 > ceiling - 1) 524 end -= PMD_SIZE; 525 if (addr > end - 1) 526 return; 527 528 pgd = pgd_offset(tlb->mm, addr); 529 do { 530 next = pgd_addr_end(addr, end); 531 if (pgd_none_or_clear_bad(pgd)) 532 continue; 533 free_pud_range(tlb, pgd, addr, next, floor, ceiling); 534 } while (pgd++, addr = next, addr != end); 535 } 536 537 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 538 unsigned long floor, unsigned long ceiling) 539 { 540 while (vma) { 541 struct vm_area_struct *next = vma->vm_next; 542 unsigned long addr = vma->vm_start; 543 544 /* 545 * Hide vma from rmap and truncate_pagecache before freeing 546 * pgtables 547 */ 548 unlink_anon_vmas(vma); 549 unlink_file_vma(vma); 550 551 if (is_vm_hugetlb_page(vma)) { 552 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 553 floor, next? next->vm_start: ceiling); 554 } else { 555 /* 556 * Optimization: gather nearby vmas into one call down 557 */ 558 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 559 && !is_vm_hugetlb_page(next)) { 560 vma = next; 561 next = vma->vm_next; 562 unlink_anon_vmas(vma); 563 unlink_file_vma(vma); 564 } 565 free_pgd_range(tlb, addr, vma->vm_end, 566 floor, next? next->vm_start: ceiling); 567 } 568 vma = next; 569 } 570 } 571 572 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 573 pmd_t *pmd, unsigned long address) 574 { 575 pgtable_t new = pte_alloc_one(mm, address); 576 int wait_split_huge_page; 577 if (!new) 578 return -ENOMEM; 579 580 /* 581 * Ensure all pte setup (eg. pte page lock and page clearing) are 582 * visible before the pte is made visible to other CPUs by being 583 * put into page tables. 584 * 585 * The other side of the story is the pointer chasing in the page 586 * table walking code (when walking the page table without locking; 587 * ie. most of the time). Fortunately, these data accesses consist 588 * of a chain of data-dependent loads, meaning most CPUs (alpha 589 * being the notable exception) will already guarantee loads are 590 * seen in-order. See the alpha page table accessors for the 591 * smp_read_barrier_depends() barriers in page table walking code. 592 */ 593 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 594 595 spin_lock(&mm->page_table_lock); 596 wait_split_huge_page = 0; 597 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 598 mm->nr_ptes++; 599 pmd_populate(mm, pmd, new); 600 new = NULL; 601 } else if (unlikely(pmd_trans_splitting(*pmd))) 602 wait_split_huge_page = 1; 603 spin_unlock(&mm->page_table_lock); 604 if (new) 605 pte_free(mm, new); 606 if (wait_split_huge_page) 607 wait_split_huge_page(vma->anon_vma, pmd); 608 return 0; 609 } 610 611 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address) 612 { 613 pte_t *new = pte_alloc_one_kernel(&init_mm, address); 614 if (!new) 615 return -ENOMEM; 616 617 smp_wmb(); /* See comment in __pte_alloc */ 618 619 spin_lock(&init_mm.page_table_lock); 620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 621 pmd_populate_kernel(&init_mm, pmd, new); 622 new = NULL; 623 } else 624 VM_BUG_ON(pmd_trans_splitting(*pmd)); 625 spin_unlock(&init_mm.page_table_lock); 626 if (new) 627 pte_free_kernel(&init_mm, new); 628 return 0; 629 } 630 631 static inline void init_rss_vec(int *rss) 632 { 633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 634 } 635 636 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 637 { 638 int i; 639 640 if (current->mm == mm) 641 sync_mm_rss(mm); 642 for (i = 0; i < NR_MM_COUNTERS; i++) 643 if (rss[i]) 644 add_mm_counter(mm, i, rss[i]); 645 } 646 647 /* 648 * This function is called to print an error when a bad pte 649 * is found. For example, we might have a PFN-mapped pte in 650 * a region that doesn't allow it. 651 * 652 * The calling function must still handle the error. 653 */ 654 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 655 pte_t pte, struct page *page) 656 { 657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 658 pud_t *pud = pud_offset(pgd, addr); 659 pmd_t *pmd = pmd_offset(pud, addr); 660 struct address_space *mapping; 661 pgoff_t index; 662 static unsigned long resume; 663 static unsigned long nr_shown; 664 static unsigned long nr_unshown; 665 666 /* 667 * Allow a burst of 60 reports, then keep quiet for that minute; 668 * or allow a steady drip of one report per second. 669 */ 670 if (nr_shown == 60) { 671 if (time_before(jiffies, resume)) { 672 nr_unshown++; 673 return; 674 } 675 if (nr_unshown) { 676 printk(KERN_ALERT 677 "BUG: Bad page map: %lu messages suppressed\n", 678 nr_unshown); 679 nr_unshown = 0; 680 } 681 nr_shown = 0; 682 } 683 if (nr_shown++ == 0) 684 resume = jiffies + 60 * HZ; 685 686 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 687 index = linear_page_index(vma, addr); 688 689 printk(KERN_ALERT 690 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 691 current->comm, 692 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 693 if (page) 694 dump_page(page); 695 printk(KERN_ALERT 696 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", 697 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 698 /* 699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y 700 */ 701 if (vma->vm_ops) 702 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n", 703 (unsigned long)vma->vm_ops->fault); 704 if (vma->vm_file && vma->vm_file->f_op) 705 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n", 706 (unsigned long)vma->vm_file->f_op->mmap); 707 dump_stack(); 708 add_taint(TAINT_BAD_PAGE); 709 } 710 711 static inline int is_cow_mapping(vm_flags_t flags) 712 { 713 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 714 } 715 716 #ifndef is_zero_pfn 717 static inline int is_zero_pfn(unsigned long pfn) 718 { 719 return pfn == zero_pfn; 720 } 721 #endif 722 723 #ifndef my_zero_pfn 724 static inline unsigned long my_zero_pfn(unsigned long addr) 725 { 726 return zero_pfn; 727 } 728 #endif 729 730 /* 731 * vm_normal_page -- This function gets the "struct page" associated with a pte. 732 * 733 * "Special" mappings do not wish to be associated with a "struct page" (either 734 * it doesn't exist, or it exists but they don't want to touch it). In this 735 * case, NULL is returned here. "Normal" mappings do have a struct page. 736 * 737 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 738 * pte bit, in which case this function is trivial. Secondly, an architecture 739 * may not have a spare pte bit, which requires a more complicated scheme, 740 * described below. 741 * 742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 743 * special mapping (even if there are underlying and valid "struct pages"). 744 * COWed pages of a VM_PFNMAP are always normal. 745 * 746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 749 * mapping will always honor the rule 750 * 751 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 752 * 753 * And for normal mappings this is false. 754 * 755 * This restricts such mappings to be a linear translation from virtual address 756 * to pfn. To get around this restriction, we allow arbitrary mappings so long 757 * as the vma is not a COW mapping; in that case, we know that all ptes are 758 * special (because none can have been COWed). 759 * 760 * 761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 762 * 763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 764 * page" backing, however the difference is that _all_ pages with a struct 765 * page (that is, those where pfn_valid is true) are refcounted and considered 766 * normal pages by the VM. The disadvantage is that pages are refcounted 767 * (which can be slower and simply not an option for some PFNMAP users). The 768 * advantage is that we don't have to follow the strict linearity rule of 769 * PFNMAP mappings in order to support COWable mappings. 770 * 771 */ 772 #ifdef __HAVE_ARCH_PTE_SPECIAL 773 # define HAVE_PTE_SPECIAL 1 774 #else 775 # define HAVE_PTE_SPECIAL 0 776 #endif 777 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 778 pte_t pte) 779 { 780 unsigned long pfn = pte_pfn(pte); 781 782 if (HAVE_PTE_SPECIAL) { 783 if (likely(!pte_special(pte))) 784 goto check_pfn; 785 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 786 return NULL; 787 if (!is_zero_pfn(pfn)) 788 print_bad_pte(vma, addr, pte, NULL); 789 return NULL; 790 } 791 792 /* !HAVE_PTE_SPECIAL case follows: */ 793 794 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 795 if (vma->vm_flags & VM_MIXEDMAP) { 796 if (!pfn_valid(pfn)) 797 return NULL; 798 goto out; 799 } else { 800 unsigned long off; 801 off = (addr - vma->vm_start) >> PAGE_SHIFT; 802 if (pfn == vma->vm_pgoff + off) 803 return NULL; 804 if (!is_cow_mapping(vma->vm_flags)) 805 return NULL; 806 } 807 } 808 809 if (is_zero_pfn(pfn)) 810 return NULL; 811 check_pfn: 812 if (unlikely(pfn > highest_memmap_pfn)) { 813 print_bad_pte(vma, addr, pte, NULL); 814 return NULL; 815 } 816 817 /* 818 * NOTE! We still have PageReserved() pages in the page tables. 819 * eg. VDSO mappings can cause them to exist. 820 */ 821 out: 822 return pfn_to_page(pfn); 823 } 824 825 /* 826 * copy one vm_area from one task to the other. Assumes the page tables 827 * already present in the new task to be cleared in the whole range 828 * covered by this vma. 829 */ 830 831 static inline unsigned long 832 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 833 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 834 unsigned long addr, int *rss) 835 { 836 unsigned long vm_flags = vma->vm_flags; 837 pte_t pte = *src_pte; 838 struct page *page; 839 840 /* pte contains position in swap or file, so copy. */ 841 if (unlikely(!pte_present(pte))) { 842 if (!pte_file(pte)) { 843 swp_entry_t entry = pte_to_swp_entry(pte); 844 845 if (swap_duplicate(entry) < 0) 846 return entry.val; 847 848 /* make sure dst_mm is on swapoff's mmlist. */ 849 if (unlikely(list_empty(&dst_mm->mmlist))) { 850 spin_lock(&mmlist_lock); 851 if (list_empty(&dst_mm->mmlist)) 852 list_add(&dst_mm->mmlist, 853 &src_mm->mmlist); 854 spin_unlock(&mmlist_lock); 855 } 856 if (likely(!non_swap_entry(entry))) 857 rss[MM_SWAPENTS]++; 858 else if (is_migration_entry(entry)) { 859 page = migration_entry_to_page(entry); 860 861 if (PageAnon(page)) 862 rss[MM_ANONPAGES]++; 863 else 864 rss[MM_FILEPAGES]++; 865 866 if (is_write_migration_entry(entry) && 867 is_cow_mapping(vm_flags)) { 868 /* 869 * COW mappings require pages in both 870 * parent and child to be set to read. 871 */ 872 make_migration_entry_read(&entry); 873 pte = swp_entry_to_pte(entry); 874 set_pte_at(src_mm, addr, src_pte, pte); 875 } 876 } 877 } 878 goto out_set_pte; 879 } 880 881 /* 882 * If it's a COW mapping, write protect it both 883 * in the parent and the child 884 */ 885 if (is_cow_mapping(vm_flags)) { 886 ptep_set_wrprotect(src_mm, addr, src_pte); 887 pte = pte_wrprotect(pte); 888 } 889 890 /* 891 * If it's a shared mapping, mark it clean in 892 * the child 893 */ 894 if (vm_flags & VM_SHARED) 895 pte = pte_mkclean(pte); 896 pte = pte_mkold(pte); 897 898 page = vm_normal_page(vma, addr, pte); 899 if (page) { 900 get_page(page); 901 page_dup_rmap(page); 902 if (PageAnon(page)) 903 rss[MM_ANONPAGES]++; 904 else 905 rss[MM_FILEPAGES]++; 906 } 907 908 out_set_pte: 909 set_pte_at(dst_mm, addr, dst_pte, pte); 910 return 0; 911 } 912 913 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 914 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 915 unsigned long addr, unsigned long end) 916 { 917 pte_t *orig_src_pte, *orig_dst_pte; 918 pte_t *src_pte, *dst_pte; 919 spinlock_t *src_ptl, *dst_ptl; 920 int progress = 0; 921 int rss[NR_MM_COUNTERS]; 922 swp_entry_t entry = (swp_entry_t){0}; 923 924 again: 925 init_rss_vec(rss); 926 927 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 928 if (!dst_pte) 929 return -ENOMEM; 930 src_pte = pte_offset_map(src_pmd, addr); 931 src_ptl = pte_lockptr(src_mm, src_pmd); 932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 933 orig_src_pte = src_pte; 934 orig_dst_pte = dst_pte; 935 arch_enter_lazy_mmu_mode(); 936 937 do { 938 /* 939 * We are holding two locks at this point - either of them 940 * could generate latencies in another task on another CPU. 941 */ 942 if (progress >= 32) { 943 progress = 0; 944 if (need_resched() || 945 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 946 break; 947 } 948 if (pte_none(*src_pte)) { 949 progress++; 950 continue; 951 } 952 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, 953 vma, addr, rss); 954 if (entry.val) 955 break; 956 progress += 8; 957 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 958 959 arch_leave_lazy_mmu_mode(); 960 spin_unlock(src_ptl); 961 pte_unmap(orig_src_pte); 962 add_mm_rss_vec(dst_mm, rss); 963 pte_unmap_unlock(orig_dst_pte, dst_ptl); 964 cond_resched(); 965 966 if (entry.val) { 967 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) 968 return -ENOMEM; 969 progress = 0; 970 } 971 if (addr != end) 972 goto again; 973 return 0; 974 } 975 976 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 977 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 978 unsigned long addr, unsigned long end) 979 { 980 pmd_t *src_pmd, *dst_pmd; 981 unsigned long next; 982 983 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 984 if (!dst_pmd) 985 return -ENOMEM; 986 src_pmd = pmd_offset(src_pud, addr); 987 do { 988 next = pmd_addr_end(addr, end); 989 if (pmd_trans_huge(*src_pmd)) { 990 int err; 991 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE); 992 err = copy_huge_pmd(dst_mm, src_mm, 993 dst_pmd, src_pmd, addr, vma); 994 if (err == -ENOMEM) 995 return -ENOMEM; 996 if (!err) 997 continue; 998 /* fall through */ 999 } 1000 if (pmd_none_or_clear_bad(src_pmd)) 1001 continue; 1002 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 1003 vma, addr, next)) 1004 return -ENOMEM; 1005 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 1006 return 0; 1007 } 1008 1009 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1010 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 1011 unsigned long addr, unsigned long end) 1012 { 1013 pud_t *src_pud, *dst_pud; 1014 unsigned long next; 1015 1016 dst_pud = pud_alloc(dst_mm, dst_pgd, addr); 1017 if (!dst_pud) 1018 return -ENOMEM; 1019 src_pud = pud_offset(src_pgd, addr); 1020 do { 1021 next = pud_addr_end(addr, end); 1022 if (pud_none_or_clear_bad(src_pud)) 1023 continue; 1024 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 1025 vma, addr, next)) 1026 return -ENOMEM; 1027 } while (dst_pud++, src_pud++, addr = next, addr != end); 1028 return 0; 1029 } 1030 1031 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1032 struct vm_area_struct *vma) 1033 { 1034 pgd_t *src_pgd, *dst_pgd; 1035 unsigned long next; 1036 unsigned long addr = vma->vm_start; 1037 unsigned long end = vma->vm_end; 1038 int ret; 1039 1040 /* 1041 * Don't copy ptes where a page fault will fill them correctly. 1042 * Fork becomes much lighter when there are big shared or private 1043 * readonly mappings. The tradeoff is that copy_page_range is more 1044 * efficient than faulting. 1045 */ 1046 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) { 1047 if (!vma->anon_vma) 1048 return 0; 1049 } 1050 1051 if (is_vm_hugetlb_page(vma)) 1052 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 1053 1054 if (unlikely(is_pfn_mapping(vma))) { 1055 /* 1056 * We do not free on error cases below as remove_vma 1057 * gets called on error from higher level routine 1058 */ 1059 ret = track_pfn_vma_copy(vma); 1060 if (ret) 1061 return ret; 1062 } 1063 1064 /* 1065 * We need to invalidate the secondary MMU mappings only when 1066 * there could be a permission downgrade on the ptes of the 1067 * parent mm. And a permission downgrade will only happen if 1068 * is_cow_mapping() returns true. 1069 */ 1070 if (is_cow_mapping(vma->vm_flags)) 1071 mmu_notifier_invalidate_range_start(src_mm, addr, end); 1072 1073 ret = 0; 1074 dst_pgd = pgd_offset(dst_mm, addr); 1075 src_pgd = pgd_offset(src_mm, addr); 1076 do { 1077 next = pgd_addr_end(addr, end); 1078 if (pgd_none_or_clear_bad(src_pgd)) 1079 continue; 1080 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, 1081 vma, addr, next))) { 1082 ret = -ENOMEM; 1083 break; 1084 } 1085 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1086 1087 if (is_cow_mapping(vma->vm_flags)) 1088 mmu_notifier_invalidate_range_end(src_mm, 1089 vma->vm_start, end); 1090 return ret; 1091 } 1092 1093 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1094 struct vm_area_struct *vma, pmd_t *pmd, 1095 unsigned long addr, unsigned long end, 1096 struct zap_details *details) 1097 { 1098 struct mm_struct *mm = tlb->mm; 1099 int force_flush = 0; 1100 int rss[NR_MM_COUNTERS]; 1101 spinlock_t *ptl; 1102 pte_t *start_pte; 1103 pte_t *pte; 1104 1105 again: 1106 init_rss_vec(rss); 1107 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1108 pte = start_pte; 1109 arch_enter_lazy_mmu_mode(); 1110 do { 1111 pte_t ptent = *pte; 1112 if (pte_none(ptent)) { 1113 continue; 1114 } 1115 1116 if (pte_present(ptent)) { 1117 struct page *page; 1118 1119 page = vm_normal_page(vma, addr, ptent); 1120 if (unlikely(details) && page) { 1121 /* 1122 * unmap_shared_mapping_pages() wants to 1123 * invalidate cache without truncating: 1124 * unmap shared but keep private pages. 1125 */ 1126 if (details->check_mapping && 1127 details->check_mapping != page->mapping) 1128 continue; 1129 /* 1130 * Each page->index must be checked when 1131 * invalidating or truncating nonlinear. 1132 */ 1133 if (details->nonlinear_vma && 1134 (page->index < details->first_index || 1135 page->index > details->last_index)) 1136 continue; 1137 } 1138 ptent = ptep_get_and_clear_full(mm, addr, pte, 1139 tlb->fullmm); 1140 tlb_remove_tlb_entry(tlb, pte, addr); 1141 if (unlikely(!page)) 1142 continue; 1143 if (unlikely(details) && details->nonlinear_vma 1144 && linear_page_index(details->nonlinear_vma, 1145 addr) != page->index) 1146 set_pte_at(mm, addr, pte, 1147 pgoff_to_pte(page->index)); 1148 if (PageAnon(page)) 1149 rss[MM_ANONPAGES]--; 1150 else { 1151 if (pte_dirty(ptent)) 1152 set_page_dirty(page); 1153 if (pte_young(ptent) && 1154 likely(!VM_SequentialReadHint(vma))) 1155 mark_page_accessed(page); 1156 rss[MM_FILEPAGES]--; 1157 } 1158 page_remove_rmap(page); 1159 if (unlikely(page_mapcount(page) < 0)) 1160 print_bad_pte(vma, addr, ptent, page); 1161 force_flush = !__tlb_remove_page(tlb, page); 1162 if (force_flush) 1163 break; 1164 continue; 1165 } 1166 /* 1167 * If details->check_mapping, we leave swap entries; 1168 * if details->nonlinear_vma, we leave file entries. 1169 */ 1170 if (unlikely(details)) 1171 continue; 1172 if (pte_file(ptent)) { 1173 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) 1174 print_bad_pte(vma, addr, ptent, NULL); 1175 } else { 1176 swp_entry_t entry = pte_to_swp_entry(ptent); 1177 1178 if (!non_swap_entry(entry)) 1179 rss[MM_SWAPENTS]--; 1180 else if (is_migration_entry(entry)) { 1181 struct page *page; 1182 1183 page = migration_entry_to_page(entry); 1184 1185 if (PageAnon(page)) 1186 rss[MM_ANONPAGES]--; 1187 else 1188 rss[MM_FILEPAGES]--; 1189 } 1190 if (unlikely(!free_swap_and_cache(entry))) 1191 print_bad_pte(vma, addr, ptent, NULL); 1192 } 1193 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1194 } while (pte++, addr += PAGE_SIZE, addr != end); 1195 1196 add_mm_rss_vec(mm, rss); 1197 arch_leave_lazy_mmu_mode(); 1198 pte_unmap_unlock(start_pte, ptl); 1199 1200 /* 1201 * mmu_gather ran out of room to batch pages, we break out of 1202 * the PTE lock to avoid doing the potential expensive TLB invalidate 1203 * and page-free while holding it. 1204 */ 1205 if (force_flush) { 1206 force_flush = 0; 1207 tlb_flush_mmu(tlb); 1208 if (addr != end) 1209 goto again; 1210 } 1211 1212 return addr; 1213 } 1214 1215 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1216 struct vm_area_struct *vma, pud_t *pud, 1217 unsigned long addr, unsigned long end, 1218 struct zap_details *details) 1219 { 1220 pmd_t *pmd; 1221 unsigned long next; 1222 1223 pmd = pmd_offset(pud, addr); 1224 do { 1225 next = pmd_addr_end(addr, end); 1226 if (pmd_trans_huge(*pmd)) { 1227 if (next - addr != HPAGE_PMD_SIZE) { 1228 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem)); 1229 split_huge_page_pmd(vma->vm_mm, pmd); 1230 } else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1231 goto next; 1232 /* fall through */ 1233 } 1234 /* 1235 * Here there can be other concurrent MADV_DONTNEED or 1236 * trans huge page faults running, and if the pmd is 1237 * none or trans huge it can change under us. This is 1238 * because MADV_DONTNEED holds the mmap_sem in read 1239 * mode. 1240 */ 1241 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1242 goto next; 1243 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1244 next: 1245 cond_resched(); 1246 } while (pmd++, addr = next, addr != end); 1247 1248 return addr; 1249 } 1250 1251 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1252 struct vm_area_struct *vma, pgd_t *pgd, 1253 unsigned long addr, unsigned long end, 1254 struct zap_details *details) 1255 { 1256 pud_t *pud; 1257 unsigned long next; 1258 1259 pud = pud_offset(pgd, addr); 1260 do { 1261 next = pud_addr_end(addr, end); 1262 if (pud_none_or_clear_bad(pud)) 1263 continue; 1264 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1265 } while (pud++, addr = next, addr != end); 1266 1267 return addr; 1268 } 1269 1270 static void unmap_page_range(struct mmu_gather *tlb, 1271 struct vm_area_struct *vma, 1272 unsigned long addr, unsigned long end, 1273 struct zap_details *details) 1274 { 1275 pgd_t *pgd; 1276 unsigned long next; 1277 1278 if (details && !details->check_mapping && !details->nonlinear_vma) 1279 details = NULL; 1280 1281 BUG_ON(addr >= end); 1282 mem_cgroup_uncharge_start(); 1283 tlb_start_vma(tlb, vma); 1284 pgd = pgd_offset(vma->vm_mm, addr); 1285 do { 1286 next = pgd_addr_end(addr, end); 1287 if (pgd_none_or_clear_bad(pgd)) 1288 continue; 1289 next = zap_pud_range(tlb, vma, pgd, addr, next, details); 1290 } while (pgd++, addr = next, addr != end); 1291 tlb_end_vma(tlb, vma); 1292 mem_cgroup_uncharge_end(); 1293 } 1294 1295 1296 static void unmap_single_vma(struct mmu_gather *tlb, 1297 struct vm_area_struct *vma, unsigned long start_addr, 1298 unsigned long end_addr, 1299 struct zap_details *details) 1300 { 1301 unsigned long start = max(vma->vm_start, start_addr); 1302 unsigned long end; 1303 1304 if (start >= vma->vm_end) 1305 return; 1306 end = min(vma->vm_end, end_addr); 1307 if (end <= vma->vm_start) 1308 return; 1309 1310 if (vma->vm_file) 1311 uprobe_munmap(vma, start, end); 1312 1313 if (unlikely(is_pfn_mapping(vma))) 1314 untrack_pfn_vma(vma, 0, 0); 1315 1316 if (start != end) { 1317 if (unlikely(is_vm_hugetlb_page(vma))) { 1318 /* 1319 * It is undesirable to test vma->vm_file as it 1320 * should be non-null for valid hugetlb area. 1321 * However, vm_file will be NULL in the error 1322 * cleanup path of do_mmap_pgoff. When 1323 * hugetlbfs ->mmap method fails, 1324 * do_mmap_pgoff() nullifies vma->vm_file 1325 * before calling this function to clean up. 1326 * Since no pte has actually been setup, it is 1327 * safe to do nothing in this case. 1328 */ 1329 if (vma->vm_file) 1330 unmap_hugepage_range(vma, start, end, NULL); 1331 } else 1332 unmap_page_range(tlb, vma, start, end, details); 1333 } 1334 } 1335 1336 /** 1337 * unmap_vmas - unmap a range of memory covered by a list of vma's 1338 * @tlb: address of the caller's struct mmu_gather 1339 * @vma: the starting vma 1340 * @start_addr: virtual address at which to start unmapping 1341 * @end_addr: virtual address at which to end unmapping 1342 * 1343 * Unmap all pages in the vma list. 1344 * 1345 * Only addresses between `start' and `end' will be unmapped. 1346 * 1347 * The VMA list must be sorted in ascending virtual address order. 1348 * 1349 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1350 * range after unmap_vmas() returns. So the only responsibility here is to 1351 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1352 * drops the lock and schedules. 1353 */ 1354 void unmap_vmas(struct mmu_gather *tlb, 1355 struct vm_area_struct *vma, unsigned long start_addr, 1356 unsigned long end_addr) 1357 { 1358 struct mm_struct *mm = vma->vm_mm; 1359 1360 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr); 1361 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1362 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1363 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr); 1364 } 1365 1366 /** 1367 * zap_page_range - remove user pages in a given range 1368 * @vma: vm_area_struct holding the applicable pages 1369 * @address: starting address of pages to zap 1370 * @size: number of bytes to zap 1371 * @details: details of nonlinear truncation or shared cache invalidation 1372 * 1373 * Caller must protect the VMA list 1374 */ 1375 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1376 unsigned long size, struct zap_details *details) 1377 { 1378 struct mm_struct *mm = vma->vm_mm; 1379 struct mmu_gather tlb; 1380 unsigned long end = start + size; 1381 1382 lru_add_drain(); 1383 tlb_gather_mmu(&tlb, mm, 0); 1384 update_hiwater_rss(mm); 1385 mmu_notifier_invalidate_range_start(mm, start, end); 1386 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) 1387 unmap_single_vma(&tlb, vma, start, end, details); 1388 mmu_notifier_invalidate_range_end(mm, start, end); 1389 tlb_finish_mmu(&tlb, start, end); 1390 } 1391 1392 /** 1393 * zap_page_range_single - remove user pages in a given range 1394 * @vma: vm_area_struct holding the applicable pages 1395 * @address: starting address of pages to zap 1396 * @size: number of bytes to zap 1397 * @details: details of nonlinear truncation or shared cache invalidation 1398 * 1399 * The range must fit into one VMA. 1400 */ 1401 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1402 unsigned long size, struct zap_details *details) 1403 { 1404 struct mm_struct *mm = vma->vm_mm; 1405 struct mmu_gather tlb; 1406 unsigned long end = address + size; 1407 1408 lru_add_drain(); 1409 tlb_gather_mmu(&tlb, mm, 0); 1410 update_hiwater_rss(mm); 1411 mmu_notifier_invalidate_range_start(mm, address, end); 1412 unmap_single_vma(&tlb, vma, address, end, details); 1413 mmu_notifier_invalidate_range_end(mm, address, end); 1414 tlb_finish_mmu(&tlb, address, end); 1415 } 1416 1417 /** 1418 * zap_vma_ptes - remove ptes mapping the vma 1419 * @vma: vm_area_struct holding ptes to be zapped 1420 * @address: starting address of pages to zap 1421 * @size: number of bytes to zap 1422 * 1423 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1424 * 1425 * The entire address range must be fully contained within the vma. 1426 * 1427 * Returns 0 if successful. 1428 */ 1429 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1430 unsigned long size) 1431 { 1432 if (address < vma->vm_start || address + size > vma->vm_end || 1433 !(vma->vm_flags & VM_PFNMAP)) 1434 return -1; 1435 zap_page_range_single(vma, address, size, NULL); 1436 return 0; 1437 } 1438 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1439 1440 /** 1441 * follow_page - look up a page descriptor from a user-virtual address 1442 * @vma: vm_area_struct mapping @address 1443 * @address: virtual address to look up 1444 * @flags: flags modifying lookup behaviour 1445 * 1446 * @flags can have FOLL_ flags set, defined in <linux/mm.h> 1447 * 1448 * Returns the mapped (struct page *), %NULL if no mapping exists, or 1449 * an error pointer if there is a mapping to something not represented 1450 * by a page descriptor (see also vm_normal_page()). 1451 */ 1452 struct page *follow_page(struct vm_area_struct *vma, unsigned long address, 1453 unsigned int flags) 1454 { 1455 pgd_t *pgd; 1456 pud_t *pud; 1457 pmd_t *pmd; 1458 pte_t *ptep, pte; 1459 spinlock_t *ptl; 1460 struct page *page; 1461 struct mm_struct *mm = vma->vm_mm; 1462 1463 page = follow_huge_addr(mm, address, flags & FOLL_WRITE); 1464 if (!IS_ERR(page)) { 1465 BUG_ON(flags & FOLL_GET); 1466 goto out; 1467 } 1468 1469 page = NULL; 1470 pgd = pgd_offset(mm, address); 1471 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 1472 goto no_page_table; 1473 1474 pud = pud_offset(pgd, address); 1475 if (pud_none(*pud)) 1476 goto no_page_table; 1477 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) { 1478 BUG_ON(flags & FOLL_GET); 1479 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE); 1480 goto out; 1481 } 1482 if (unlikely(pud_bad(*pud))) 1483 goto no_page_table; 1484 1485 pmd = pmd_offset(pud, address); 1486 if (pmd_none(*pmd)) 1487 goto no_page_table; 1488 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) { 1489 BUG_ON(flags & FOLL_GET); 1490 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE); 1491 goto out; 1492 } 1493 if (pmd_trans_huge(*pmd)) { 1494 if (flags & FOLL_SPLIT) { 1495 split_huge_page_pmd(mm, pmd); 1496 goto split_fallthrough; 1497 } 1498 spin_lock(&mm->page_table_lock); 1499 if (likely(pmd_trans_huge(*pmd))) { 1500 if (unlikely(pmd_trans_splitting(*pmd))) { 1501 spin_unlock(&mm->page_table_lock); 1502 wait_split_huge_page(vma->anon_vma, pmd); 1503 } else { 1504 page = follow_trans_huge_pmd(mm, address, 1505 pmd, flags); 1506 spin_unlock(&mm->page_table_lock); 1507 goto out; 1508 } 1509 } else 1510 spin_unlock(&mm->page_table_lock); 1511 /* fall through */ 1512 } 1513 split_fallthrough: 1514 if (unlikely(pmd_bad(*pmd))) 1515 goto no_page_table; 1516 1517 ptep = pte_offset_map_lock(mm, pmd, address, &ptl); 1518 1519 pte = *ptep; 1520 if (!pte_present(pte)) 1521 goto no_page; 1522 if ((flags & FOLL_WRITE) && !pte_write(pte)) 1523 goto unlock; 1524 1525 page = vm_normal_page(vma, address, pte); 1526 if (unlikely(!page)) { 1527 if ((flags & FOLL_DUMP) || 1528 !is_zero_pfn(pte_pfn(pte))) 1529 goto bad_page; 1530 page = pte_page(pte); 1531 } 1532 1533 if (flags & FOLL_GET) 1534 get_page_foll(page); 1535 if (flags & FOLL_TOUCH) { 1536 if ((flags & FOLL_WRITE) && 1537 !pte_dirty(pte) && !PageDirty(page)) 1538 set_page_dirty(page); 1539 /* 1540 * pte_mkyoung() would be more correct here, but atomic care 1541 * is needed to avoid losing the dirty bit: it is easier to use 1542 * mark_page_accessed(). 1543 */ 1544 mark_page_accessed(page); 1545 } 1546 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1547 /* 1548 * The preliminary mapping check is mainly to avoid the 1549 * pointless overhead of lock_page on the ZERO_PAGE 1550 * which might bounce very badly if there is contention. 1551 * 1552 * If the page is already locked, we don't need to 1553 * handle it now - vmscan will handle it later if and 1554 * when it attempts to reclaim the page. 1555 */ 1556 if (page->mapping && trylock_page(page)) { 1557 lru_add_drain(); /* push cached pages to LRU */ 1558 /* 1559 * Because we lock page here and migration is 1560 * blocked by the pte's page reference, we need 1561 * only check for file-cache page truncation. 1562 */ 1563 if (page->mapping) 1564 mlock_vma_page(page); 1565 unlock_page(page); 1566 } 1567 } 1568 unlock: 1569 pte_unmap_unlock(ptep, ptl); 1570 out: 1571 return page; 1572 1573 bad_page: 1574 pte_unmap_unlock(ptep, ptl); 1575 return ERR_PTR(-EFAULT); 1576 1577 no_page: 1578 pte_unmap_unlock(ptep, ptl); 1579 if (!pte_none(pte)) 1580 return page; 1581 1582 no_page_table: 1583 /* 1584 * When core dumping an enormous anonymous area that nobody 1585 * has touched so far, we don't want to allocate unnecessary pages or 1586 * page tables. Return error instead of NULL to skip handle_mm_fault, 1587 * then get_dump_page() will return NULL to leave a hole in the dump. 1588 * But we can only make this optimization where a hole would surely 1589 * be zero-filled if handle_mm_fault() actually did handle it. 1590 */ 1591 if ((flags & FOLL_DUMP) && 1592 (!vma->vm_ops || !vma->vm_ops->fault)) 1593 return ERR_PTR(-EFAULT); 1594 return page; 1595 } 1596 1597 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr) 1598 { 1599 return stack_guard_page_start(vma, addr) || 1600 stack_guard_page_end(vma, addr+PAGE_SIZE); 1601 } 1602 1603 /** 1604 * __get_user_pages() - pin user pages in memory 1605 * @tsk: task_struct of target task 1606 * @mm: mm_struct of target mm 1607 * @start: starting user address 1608 * @nr_pages: number of pages from start to pin 1609 * @gup_flags: flags modifying pin behaviour 1610 * @pages: array that receives pointers to the pages pinned. 1611 * Should be at least nr_pages long. Or NULL, if caller 1612 * only intends to ensure the pages are faulted in. 1613 * @vmas: array of pointers to vmas corresponding to each page. 1614 * Or NULL if the caller does not require them. 1615 * @nonblocking: whether waiting for disk IO or mmap_sem contention 1616 * 1617 * Returns number of pages pinned. This may be fewer than the number 1618 * requested. If nr_pages is 0 or negative, returns 0. If no pages 1619 * were pinned, returns -errno. Each page returned must be released 1620 * with a put_page() call when it is finished with. vmas will only 1621 * remain valid while mmap_sem is held. 1622 * 1623 * Must be called with mmap_sem held for read or write. 1624 * 1625 * __get_user_pages walks a process's page tables and takes a reference to 1626 * each struct page that each user address corresponds to at a given 1627 * instant. That is, it takes the page that would be accessed if a user 1628 * thread accesses the given user virtual address at that instant. 1629 * 1630 * This does not guarantee that the page exists in the user mappings when 1631 * __get_user_pages returns, and there may even be a completely different 1632 * page there in some cases (eg. if mmapped pagecache has been invalidated 1633 * and subsequently re faulted). However it does guarantee that the page 1634 * won't be freed completely. And mostly callers simply care that the page 1635 * contains data that was valid *at some point in time*. Typically, an IO 1636 * or similar operation cannot guarantee anything stronger anyway because 1637 * locks can't be held over the syscall boundary. 1638 * 1639 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If 1640 * the page is written to, set_page_dirty (or set_page_dirty_lock, as 1641 * appropriate) must be called after the page is finished with, and 1642 * before put_page is called. 1643 * 1644 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO 1645 * or mmap_sem contention, and if waiting is needed to pin all pages, 1646 * *@nonblocking will be set to 0. 1647 * 1648 * In most cases, get_user_pages or get_user_pages_fast should be used 1649 * instead of __get_user_pages. __get_user_pages should be used only if 1650 * you need some special @gup_flags. 1651 */ 1652 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm, 1653 unsigned long start, int nr_pages, unsigned int gup_flags, 1654 struct page **pages, struct vm_area_struct **vmas, 1655 int *nonblocking) 1656 { 1657 int i; 1658 unsigned long vm_flags; 1659 1660 if (nr_pages <= 0) 1661 return 0; 1662 1663 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET)); 1664 1665 /* 1666 * Require read or write permissions. 1667 * If FOLL_FORCE is set, we only require the "MAY" flags. 1668 */ 1669 vm_flags = (gup_flags & FOLL_WRITE) ? 1670 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); 1671 vm_flags &= (gup_flags & FOLL_FORCE) ? 1672 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); 1673 i = 0; 1674 1675 do { 1676 struct vm_area_struct *vma; 1677 1678 vma = find_extend_vma(mm, start); 1679 if (!vma && in_gate_area(mm, start)) { 1680 unsigned long pg = start & PAGE_MASK; 1681 pgd_t *pgd; 1682 pud_t *pud; 1683 pmd_t *pmd; 1684 pte_t *pte; 1685 1686 /* user gate pages are read-only */ 1687 if (gup_flags & FOLL_WRITE) 1688 return i ? : -EFAULT; 1689 if (pg > TASK_SIZE) 1690 pgd = pgd_offset_k(pg); 1691 else 1692 pgd = pgd_offset_gate(mm, pg); 1693 BUG_ON(pgd_none(*pgd)); 1694 pud = pud_offset(pgd, pg); 1695 BUG_ON(pud_none(*pud)); 1696 pmd = pmd_offset(pud, pg); 1697 if (pmd_none(*pmd)) 1698 return i ? : -EFAULT; 1699 VM_BUG_ON(pmd_trans_huge(*pmd)); 1700 pte = pte_offset_map(pmd, pg); 1701 if (pte_none(*pte)) { 1702 pte_unmap(pte); 1703 return i ? : -EFAULT; 1704 } 1705 vma = get_gate_vma(mm); 1706 if (pages) { 1707 struct page *page; 1708 1709 page = vm_normal_page(vma, start, *pte); 1710 if (!page) { 1711 if (!(gup_flags & FOLL_DUMP) && 1712 is_zero_pfn(pte_pfn(*pte))) 1713 page = pte_page(*pte); 1714 else { 1715 pte_unmap(pte); 1716 return i ? : -EFAULT; 1717 } 1718 } 1719 pages[i] = page; 1720 get_page(page); 1721 } 1722 pte_unmap(pte); 1723 goto next_page; 1724 } 1725 1726 if (!vma || 1727 (vma->vm_flags & (VM_IO | VM_PFNMAP)) || 1728 !(vm_flags & vma->vm_flags)) 1729 return i ? : -EFAULT; 1730 1731 if (is_vm_hugetlb_page(vma)) { 1732 i = follow_hugetlb_page(mm, vma, pages, vmas, 1733 &start, &nr_pages, i, gup_flags); 1734 continue; 1735 } 1736 1737 do { 1738 struct page *page; 1739 unsigned int foll_flags = gup_flags; 1740 1741 /* 1742 * If we have a pending SIGKILL, don't keep faulting 1743 * pages and potentially allocating memory. 1744 */ 1745 if (unlikely(fatal_signal_pending(current))) 1746 return i ? i : -ERESTARTSYS; 1747 1748 cond_resched(); 1749 while (!(page = follow_page(vma, start, foll_flags))) { 1750 int ret; 1751 unsigned int fault_flags = 0; 1752 1753 /* For mlock, just skip the stack guard page. */ 1754 if (foll_flags & FOLL_MLOCK) { 1755 if (stack_guard_page(vma, start)) 1756 goto next_page; 1757 } 1758 if (foll_flags & FOLL_WRITE) 1759 fault_flags |= FAULT_FLAG_WRITE; 1760 if (nonblocking) 1761 fault_flags |= FAULT_FLAG_ALLOW_RETRY; 1762 if (foll_flags & FOLL_NOWAIT) 1763 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT); 1764 1765 ret = handle_mm_fault(mm, vma, start, 1766 fault_flags); 1767 1768 if (ret & VM_FAULT_ERROR) { 1769 if (ret & VM_FAULT_OOM) 1770 return i ? i : -ENOMEM; 1771 if (ret & (VM_FAULT_HWPOISON | 1772 VM_FAULT_HWPOISON_LARGE)) { 1773 if (i) 1774 return i; 1775 else if (gup_flags & FOLL_HWPOISON) 1776 return -EHWPOISON; 1777 else 1778 return -EFAULT; 1779 } 1780 if (ret & VM_FAULT_SIGBUS) 1781 return i ? i : -EFAULT; 1782 BUG(); 1783 } 1784 1785 if (tsk) { 1786 if (ret & VM_FAULT_MAJOR) 1787 tsk->maj_flt++; 1788 else 1789 tsk->min_flt++; 1790 } 1791 1792 if (ret & VM_FAULT_RETRY) { 1793 if (nonblocking) 1794 *nonblocking = 0; 1795 return i; 1796 } 1797 1798 /* 1799 * The VM_FAULT_WRITE bit tells us that 1800 * do_wp_page has broken COW when necessary, 1801 * even if maybe_mkwrite decided not to set 1802 * pte_write. We can thus safely do subsequent 1803 * page lookups as if they were reads. But only 1804 * do so when looping for pte_write is futile: 1805 * in some cases userspace may also be wanting 1806 * to write to the gotten user page, which a 1807 * read fault here might prevent (a readonly 1808 * page might get reCOWed by userspace write). 1809 */ 1810 if ((ret & VM_FAULT_WRITE) && 1811 !(vma->vm_flags & VM_WRITE)) 1812 foll_flags &= ~FOLL_WRITE; 1813 1814 cond_resched(); 1815 } 1816 if (IS_ERR(page)) 1817 return i ? i : PTR_ERR(page); 1818 if (pages) { 1819 pages[i] = page; 1820 1821 flush_anon_page(vma, page, start); 1822 flush_dcache_page(page); 1823 } 1824 next_page: 1825 if (vmas) 1826 vmas[i] = vma; 1827 i++; 1828 start += PAGE_SIZE; 1829 nr_pages--; 1830 } while (nr_pages && start < vma->vm_end); 1831 } while (nr_pages); 1832 return i; 1833 } 1834 EXPORT_SYMBOL(__get_user_pages); 1835 1836 /* 1837 * fixup_user_fault() - manually resolve a user page fault 1838 * @tsk: the task_struct to use for page fault accounting, or 1839 * NULL if faults are not to be recorded. 1840 * @mm: mm_struct of target mm 1841 * @address: user address 1842 * @fault_flags:flags to pass down to handle_mm_fault() 1843 * 1844 * This is meant to be called in the specific scenario where for locking reasons 1845 * we try to access user memory in atomic context (within a pagefault_disable() 1846 * section), this returns -EFAULT, and we want to resolve the user fault before 1847 * trying again. 1848 * 1849 * Typically this is meant to be used by the futex code. 1850 * 1851 * The main difference with get_user_pages() is that this function will 1852 * unconditionally call handle_mm_fault() which will in turn perform all the 1853 * necessary SW fixup of the dirty and young bits in the PTE, while 1854 * handle_mm_fault() only guarantees to update these in the struct page. 1855 * 1856 * This is important for some architectures where those bits also gate the 1857 * access permission to the page because they are maintained in software. On 1858 * such architectures, gup() will not be enough to make a subsequent access 1859 * succeed. 1860 * 1861 * This should be called with the mm_sem held for read. 1862 */ 1863 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm, 1864 unsigned long address, unsigned int fault_flags) 1865 { 1866 struct vm_area_struct *vma; 1867 int ret; 1868 1869 vma = find_extend_vma(mm, address); 1870 if (!vma || address < vma->vm_start) 1871 return -EFAULT; 1872 1873 ret = handle_mm_fault(mm, vma, address, fault_flags); 1874 if (ret & VM_FAULT_ERROR) { 1875 if (ret & VM_FAULT_OOM) 1876 return -ENOMEM; 1877 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE)) 1878 return -EHWPOISON; 1879 if (ret & VM_FAULT_SIGBUS) 1880 return -EFAULT; 1881 BUG(); 1882 } 1883 if (tsk) { 1884 if (ret & VM_FAULT_MAJOR) 1885 tsk->maj_flt++; 1886 else 1887 tsk->min_flt++; 1888 } 1889 return 0; 1890 } 1891 1892 /* 1893 * get_user_pages() - pin user pages in memory 1894 * @tsk: the task_struct to use for page fault accounting, or 1895 * NULL if faults are not to be recorded. 1896 * @mm: mm_struct of target mm 1897 * @start: starting user address 1898 * @nr_pages: number of pages from start to pin 1899 * @write: whether pages will be written to by the caller 1900 * @force: whether to force write access even if user mapping is 1901 * readonly. This will result in the page being COWed even 1902 * in MAP_SHARED mappings. You do not want this. 1903 * @pages: array that receives pointers to the pages pinned. 1904 * Should be at least nr_pages long. Or NULL, if caller 1905 * only intends to ensure the pages are faulted in. 1906 * @vmas: array of pointers to vmas corresponding to each page. 1907 * Or NULL if the caller does not require them. 1908 * 1909 * Returns number of pages pinned. This may be fewer than the number 1910 * requested. If nr_pages is 0 or negative, returns 0. If no pages 1911 * were pinned, returns -errno. Each page returned must be released 1912 * with a put_page() call when it is finished with. vmas will only 1913 * remain valid while mmap_sem is held. 1914 * 1915 * Must be called with mmap_sem held for read or write. 1916 * 1917 * get_user_pages walks a process's page tables and takes a reference to 1918 * each struct page that each user address corresponds to at a given 1919 * instant. That is, it takes the page that would be accessed if a user 1920 * thread accesses the given user virtual address at that instant. 1921 * 1922 * This does not guarantee that the page exists in the user mappings when 1923 * get_user_pages returns, and there may even be a completely different 1924 * page there in some cases (eg. if mmapped pagecache has been invalidated 1925 * and subsequently re faulted). However it does guarantee that the page 1926 * won't be freed completely. And mostly callers simply care that the page 1927 * contains data that was valid *at some point in time*. Typically, an IO 1928 * or similar operation cannot guarantee anything stronger anyway because 1929 * locks can't be held over the syscall boundary. 1930 * 1931 * If write=0, the page must not be written to. If the page is written to, 1932 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called 1933 * after the page is finished with, and before put_page is called. 1934 * 1935 * get_user_pages is typically used for fewer-copy IO operations, to get a 1936 * handle on the memory by some means other than accesses via the user virtual 1937 * addresses. The pages may be submitted for DMA to devices or accessed via 1938 * their kernel linear mapping (via the kmap APIs). Care should be taken to 1939 * use the correct cache flushing APIs. 1940 * 1941 * See also get_user_pages_fast, for performance critical applications. 1942 */ 1943 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm, 1944 unsigned long start, int nr_pages, int write, int force, 1945 struct page **pages, struct vm_area_struct **vmas) 1946 { 1947 int flags = FOLL_TOUCH; 1948 1949 if (pages) 1950 flags |= FOLL_GET; 1951 if (write) 1952 flags |= FOLL_WRITE; 1953 if (force) 1954 flags |= FOLL_FORCE; 1955 1956 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas, 1957 NULL); 1958 } 1959 EXPORT_SYMBOL(get_user_pages); 1960 1961 /** 1962 * get_dump_page() - pin user page in memory while writing it to core dump 1963 * @addr: user address 1964 * 1965 * Returns struct page pointer of user page pinned for dump, 1966 * to be freed afterwards by page_cache_release() or put_page(). 1967 * 1968 * Returns NULL on any kind of failure - a hole must then be inserted into 1969 * the corefile, to preserve alignment with its headers; and also returns 1970 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found - 1971 * allowing a hole to be left in the corefile to save diskspace. 1972 * 1973 * Called without mmap_sem, but after all other threads have been killed. 1974 */ 1975 #ifdef CONFIG_ELF_CORE 1976 struct page *get_dump_page(unsigned long addr) 1977 { 1978 struct vm_area_struct *vma; 1979 struct page *page; 1980 1981 if (__get_user_pages(current, current->mm, addr, 1, 1982 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma, 1983 NULL) < 1) 1984 return NULL; 1985 flush_cache_page(vma, addr, page_to_pfn(page)); 1986 return page; 1987 } 1988 #endif /* CONFIG_ELF_CORE */ 1989 1990 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1991 spinlock_t **ptl) 1992 { 1993 pgd_t * pgd = pgd_offset(mm, addr); 1994 pud_t * pud = pud_alloc(mm, pgd, addr); 1995 if (pud) { 1996 pmd_t * pmd = pmd_alloc(mm, pud, addr); 1997 if (pmd) { 1998 VM_BUG_ON(pmd_trans_huge(*pmd)); 1999 return pte_alloc_map_lock(mm, pmd, addr, ptl); 2000 } 2001 } 2002 return NULL; 2003 } 2004 2005 /* 2006 * This is the old fallback for page remapping. 2007 * 2008 * For historical reasons, it only allows reserved pages. Only 2009 * old drivers should use this, and they needed to mark their 2010 * pages reserved for the old functions anyway. 2011 */ 2012 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 2013 struct page *page, pgprot_t prot) 2014 { 2015 struct mm_struct *mm = vma->vm_mm; 2016 int retval; 2017 pte_t *pte; 2018 spinlock_t *ptl; 2019 2020 retval = -EINVAL; 2021 if (PageAnon(page)) 2022 goto out; 2023 retval = -ENOMEM; 2024 flush_dcache_page(page); 2025 pte = get_locked_pte(mm, addr, &ptl); 2026 if (!pte) 2027 goto out; 2028 retval = -EBUSY; 2029 if (!pte_none(*pte)) 2030 goto out_unlock; 2031 2032 /* Ok, finally just insert the thing.. */ 2033 get_page(page); 2034 inc_mm_counter_fast(mm, MM_FILEPAGES); 2035 page_add_file_rmap(page); 2036 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 2037 2038 retval = 0; 2039 pte_unmap_unlock(pte, ptl); 2040 return retval; 2041 out_unlock: 2042 pte_unmap_unlock(pte, ptl); 2043 out: 2044 return retval; 2045 } 2046 2047 /** 2048 * vm_insert_page - insert single page into user vma 2049 * @vma: user vma to map to 2050 * @addr: target user address of this page 2051 * @page: source kernel page 2052 * 2053 * This allows drivers to insert individual pages they've allocated 2054 * into a user vma. 2055 * 2056 * The page has to be a nice clean _individual_ kernel allocation. 2057 * If you allocate a compound page, you need to have marked it as 2058 * such (__GFP_COMP), or manually just split the page up yourself 2059 * (see split_page()). 2060 * 2061 * NOTE! Traditionally this was done with "remap_pfn_range()" which 2062 * took an arbitrary page protection parameter. This doesn't allow 2063 * that. Your vma protection will have to be set up correctly, which 2064 * means that if you want a shared writable mapping, you'd better 2065 * ask for a shared writable mapping! 2066 * 2067 * The page does not need to be reserved. 2068 */ 2069 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 2070 struct page *page) 2071 { 2072 if (addr < vma->vm_start || addr >= vma->vm_end) 2073 return -EFAULT; 2074 if (!page_count(page)) 2075 return -EINVAL; 2076 vma->vm_flags |= VM_INSERTPAGE; 2077 return insert_page(vma, addr, page, vma->vm_page_prot); 2078 } 2079 EXPORT_SYMBOL(vm_insert_page); 2080 2081 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2082 unsigned long pfn, pgprot_t prot) 2083 { 2084 struct mm_struct *mm = vma->vm_mm; 2085 int retval; 2086 pte_t *pte, entry; 2087 spinlock_t *ptl; 2088 2089 retval = -ENOMEM; 2090 pte = get_locked_pte(mm, addr, &ptl); 2091 if (!pte) 2092 goto out; 2093 retval = -EBUSY; 2094 if (!pte_none(*pte)) 2095 goto out_unlock; 2096 2097 /* Ok, finally just insert the thing.. */ 2098 entry = pte_mkspecial(pfn_pte(pfn, prot)); 2099 set_pte_at(mm, addr, pte, entry); 2100 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 2101 2102 retval = 0; 2103 out_unlock: 2104 pte_unmap_unlock(pte, ptl); 2105 out: 2106 return retval; 2107 } 2108 2109 /** 2110 * vm_insert_pfn - insert single pfn into user vma 2111 * @vma: user vma to map to 2112 * @addr: target user address of this page 2113 * @pfn: source kernel pfn 2114 * 2115 * Similar to vm_inert_page, this allows drivers to insert individual pages 2116 * they've allocated into a user vma. Same comments apply. 2117 * 2118 * This function should only be called from a vm_ops->fault handler, and 2119 * in that case the handler should return NULL. 2120 * 2121 * vma cannot be a COW mapping. 2122 * 2123 * As this is called only for pages that do not currently exist, we 2124 * do not need to flush old virtual caches or the TLB. 2125 */ 2126 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2127 unsigned long pfn) 2128 { 2129 int ret; 2130 pgprot_t pgprot = vma->vm_page_prot; 2131 /* 2132 * Technically, architectures with pte_special can avoid all these 2133 * restrictions (same for remap_pfn_range). However we would like 2134 * consistency in testing and feature parity among all, so we should 2135 * try to keep these invariants in place for everybody. 2136 */ 2137 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 2138 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 2139 (VM_PFNMAP|VM_MIXEDMAP)); 2140 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 2141 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 2142 2143 if (addr < vma->vm_start || addr >= vma->vm_end) 2144 return -EFAULT; 2145 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE)) 2146 return -EINVAL; 2147 2148 ret = insert_pfn(vma, addr, pfn, pgprot); 2149 2150 if (ret) 2151 untrack_pfn_vma(vma, pfn, PAGE_SIZE); 2152 2153 return ret; 2154 } 2155 EXPORT_SYMBOL(vm_insert_pfn); 2156 2157 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 2158 unsigned long pfn) 2159 { 2160 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP)); 2161 2162 if (addr < vma->vm_start || addr >= vma->vm_end) 2163 return -EFAULT; 2164 2165 /* 2166 * If we don't have pte special, then we have to use the pfn_valid() 2167 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2168 * refcount the page if pfn_valid is true (hence insert_page rather 2169 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 2170 * without pte special, it would there be refcounted as a normal page. 2171 */ 2172 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) { 2173 struct page *page; 2174 2175 page = pfn_to_page(pfn); 2176 return insert_page(vma, addr, page, vma->vm_page_prot); 2177 } 2178 return insert_pfn(vma, addr, pfn, vma->vm_page_prot); 2179 } 2180 EXPORT_SYMBOL(vm_insert_mixed); 2181 2182 /* 2183 * maps a range of physical memory into the requested pages. the old 2184 * mappings are removed. any references to nonexistent pages results 2185 * in null mappings (currently treated as "copy-on-access") 2186 */ 2187 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 2188 unsigned long addr, unsigned long end, 2189 unsigned long pfn, pgprot_t prot) 2190 { 2191 pte_t *pte; 2192 spinlock_t *ptl; 2193 2194 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 2195 if (!pte) 2196 return -ENOMEM; 2197 arch_enter_lazy_mmu_mode(); 2198 do { 2199 BUG_ON(!pte_none(*pte)); 2200 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 2201 pfn++; 2202 } while (pte++, addr += PAGE_SIZE, addr != end); 2203 arch_leave_lazy_mmu_mode(); 2204 pte_unmap_unlock(pte - 1, ptl); 2205 return 0; 2206 } 2207 2208 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 2209 unsigned long addr, unsigned long end, 2210 unsigned long pfn, pgprot_t prot) 2211 { 2212 pmd_t *pmd; 2213 unsigned long next; 2214 2215 pfn -= addr >> PAGE_SHIFT; 2216 pmd = pmd_alloc(mm, pud, addr); 2217 if (!pmd) 2218 return -ENOMEM; 2219 VM_BUG_ON(pmd_trans_huge(*pmd)); 2220 do { 2221 next = pmd_addr_end(addr, end); 2222 if (remap_pte_range(mm, pmd, addr, next, 2223 pfn + (addr >> PAGE_SHIFT), prot)) 2224 return -ENOMEM; 2225 } while (pmd++, addr = next, addr != end); 2226 return 0; 2227 } 2228 2229 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, 2230 unsigned long addr, unsigned long end, 2231 unsigned long pfn, pgprot_t prot) 2232 { 2233 pud_t *pud; 2234 unsigned long next; 2235 2236 pfn -= addr >> PAGE_SHIFT; 2237 pud = pud_alloc(mm, pgd, addr); 2238 if (!pud) 2239 return -ENOMEM; 2240 do { 2241 next = pud_addr_end(addr, end); 2242 if (remap_pmd_range(mm, pud, addr, next, 2243 pfn + (addr >> PAGE_SHIFT), prot)) 2244 return -ENOMEM; 2245 } while (pud++, addr = next, addr != end); 2246 return 0; 2247 } 2248 2249 /** 2250 * remap_pfn_range - remap kernel memory to userspace 2251 * @vma: user vma to map to 2252 * @addr: target user address to start at 2253 * @pfn: physical address of kernel memory 2254 * @size: size of map area 2255 * @prot: page protection flags for this mapping 2256 * 2257 * Note: this is only safe if the mm semaphore is held when called. 2258 */ 2259 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 2260 unsigned long pfn, unsigned long size, pgprot_t prot) 2261 { 2262 pgd_t *pgd; 2263 unsigned long next; 2264 unsigned long end = addr + PAGE_ALIGN(size); 2265 struct mm_struct *mm = vma->vm_mm; 2266 int err; 2267 2268 /* 2269 * Physically remapped pages are special. Tell the 2270 * rest of the world about it: 2271 * VM_IO tells people not to look at these pages 2272 * (accesses can have side effects). 2273 * VM_RESERVED is specified all over the place, because 2274 * in 2.4 it kept swapout's vma scan off this vma; but 2275 * in 2.6 the LRU scan won't even find its pages, so this 2276 * flag means no more than count its pages in reserved_vm, 2277 * and omit it from core dump, even when VM_IO turned off. 2278 * VM_PFNMAP tells the core MM that the base pages are just 2279 * raw PFN mappings, and do not have a "struct page" associated 2280 * with them. 2281 * 2282 * There's a horrible special case to handle copy-on-write 2283 * behaviour that some programs depend on. We mark the "original" 2284 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2285 */ 2286 if (addr == vma->vm_start && end == vma->vm_end) { 2287 vma->vm_pgoff = pfn; 2288 vma->vm_flags |= VM_PFN_AT_MMAP; 2289 } else if (is_cow_mapping(vma->vm_flags)) 2290 return -EINVAL; 2291 2292 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP; 2293 2294 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size)); 2295 if (err) { 2296 /* 2297 * To indicate that track_pfn related cleanup is not 2298 * needed from higher level routine calling unmap_vmas 2299 */ 2300 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP); 2301 vma->vm_flags &= ~VM_PFN_AT_MMAP; 2302 return -EINVAL; 2303 } 2304 2305 BUG_ON(addr >= end); 2306 pfn -= addr >> PAGE_SHIFT; 2307 pgd = pgd_offset(mm, addr); 2308 flush_cache_range(vma, addr, end); 2309 do { 2310 next = pgd_addr_end(addr, end); 2311 err = remap_pud_range(mm, pgd, addr, next, 2312 pfn + (addr >> PAGE_SHIFT), prot); 2313 if (err) 2314 break; 2315 } while (pgd++, addr = next, addr != end); 2316 2317 if (err) 2318 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size)); 2319 2320 return err; 2321 } 2322 EXPORT_SYMBOL(remap_pfn_range); 2323 2324 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2325 unsigned long addr, unsigned long end, 2326 pte_fn_t fn, void *data) 2327 { 2328 pte_t *pte; 2329 int err; 2330 pgtable_t token; 2331 spinlock_t *uninitialized_var(ptl); 2332 2333 pte = (mm == &init_mm) ? 2334 pte_alloc_kernel(pmd, addr) : 2335 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2336 if (!pte) 2337 return -ENOMEM; 2338 2339 BUG_ON(pmd_huge(*pmd)); 2340 2341 arch_enter_lazy_mmu_mode(); 2342 2343 token = pmd_pgtable(*pmd); 2344 2345 do { 2346 err = fn(pte++, token, addr, data); 2347 if (err) 2348 break; 2349 } while (addr += PAGE_SIZE, addr != end); 2350 2351 arch_leave_lazy_mmu_mode(); 2352 2353 if (mm != &init_mm) 2354 pte_unmap_unlock(pte-1, ptl); 2355 return err; 2356 } 2357 2358 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2359 unsigned long addr, unsigned long end, 2360 pte_fn_t fn, void *data) 2361 { 2362 pmd_t *pmd; 2363 unsigned long next; 2364 int err; 2365 2366 BUG_ON(pud_huge(*pud)); 2367 2368 pmd = pmd_alloc(mm, pud, addr); 2369 if (!pmd) 2370 return -ENOMEM; 2371 do { 2372 next = pmd_addr_end(addr, end); 2373 err = apply_to_pte_range(mm, pmd, addr, next, fn, data); 2374 if (err) 2375 break; 2376 } while (pmd++, addr = next, addr != end); 2377 return err; 2378 } 2379 2380 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, 2381 unsigned long addr, unsigned long end, 2382 pte_fn_t fn, void *data) 2383 { 2384 pud_t *pud; 2385 unsigned long next; 2386 int err; 2387 2388 pud = pud_alloc(mm, pgd, addr); 2389 if (!pud) 2390 return -ENOMEM; 2391 do { 2392 next = pud_addr_end(addr, end); 2393 err = apply_to_pmd_range(mm, pud, addr, next, fn, data); 2394 if (err) 2395 break; 2396 } while (pud++, addr = next, addr != end); 2397 return err; 2398 } 2399 2400 /* 2401 * Scan a region of virtual memory, filling in page tables as necessary 2402 * and calling a provided function on each leaf page table. 2403 */ 2404 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2405 unsigned long size, pte_fn_t fn, void *data) 2406 { 2407 pgd_t *pgd; 2408 unsigned long next; 2409 unsigned long end = addr + size; 2410 int err; 2411 2412 BUG_ON(addr >= end); 2413 pgd = pgd_offset(mm, addr); 2414 do { 2415 next = pgd_addr_end(addr, end); 2416 err = apply_to_pud_range(mm, pgd, addr, next, fn, data); 2417 if (err) 2418 break; 2419 } while (pgd++, addr = next, addr != end); 2420 2421 return err; 2422 } 2423 EXPORT_SYMBOL_GPL(apply_to_page_range); 2424 2425 /* 2426 * handle_pte_fault chooses page fault handler according to an entry 2427 * which was read non-atomically. Before making any commitment, on 2428 * those architectures or configurations (e.g. i386 with PAE) which 2429 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault 2430 * must check under lock before unmapping the pte and proceeding 2431 * (but do_wp_page is only called after already making such a check; 2432 * and do_anonymous_page can safely check later on). 2433 */ 2434 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 2435 pte_t *page_table, pte_t orig_pte) 2436 { 2437 int same = 1; 2438 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 2439 if (sizeof(pte_t) > sizeof(unsigned long)) { 2440 spinlock_t *ptl = pte_lockptr(mm, pmd); 2441 spin_lock(ptl); 2442 same = pte_same(*page_table, orig_pte); 2443 spin_unlock(ptl); 2444 } 2445 #endif 2446 pte_unmap(page_table); 2447 return same; 2448 } 2449 2450 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) 2451 { 2452 /* 2453 * If the source page was a PFN mapping, we don't have 2454 * a "struct page" for it. We do a best-effort copy by 2455 * just copying from the original user address. If that 2456 * fails, we just zero-fill it. Live with it. 2457 */ 2458 if (unlikely(!src)) { 2459 void *kaddr = kmap_atomic(dst); 2460 void __user *uaddr = (void __user *)(va & PAGE_MASK); 2461 2462 /* 2463 * This really shouldn't fail, because the page is there 2464 * in the page tables. But it might just be unreadable, 2465 * in which case we just give up and fill the result with 2466 * zeroes. 2467 */ 2468 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) 2469 clear_page(kaddr); 2470 kunmap_atomic(kaddr); 2471 flush_dcache_page(dst); 2472 } else 2473 copy_user_highpage(dst, src, va, vma); 2474 } 2475 2476 /* 2477 * This routine handles present pages, when users try to write 2478 * to a shared page. It is done by copying the page to a new address 2479 * and decrementing the shared-page counter for the old page. 2480 * 2481 * Note that this routine assumes that the protection checks have been 2482 * done by the caller (the low-level page fault routine in most cases). 2483 * Thus we can safely just mark it writable once we've done any necessary 2484 * COW. 2485 * 2486 * We also mark the page dirty at this point even though the page will 2487 * change only once the write actually happens. This avoids a few races, 2488 * and potentially makes it more efficient. 2489 * 2490 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2491 * but allow concurrent faults), with pte both mapped and locked. 2492 * We return with mmap_sem still held, but pte unmapped and unlocked. 2493 */ 2494 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 2495 unsigned long address, pte_t *page_table, pmd_t *pmd, 2496 spinlock_t *ptl, pte_t orig_pte) 2497 __releases(ptl) 2498 { 2499 struct page *old_page, *new_page; 2500 pte_t entry; 2501 int ret = 0; 2502 int page_mkwrite = 0; 2503 struct page *dirty_page = NULL; 2504 2505 old_page = vm_normal_page(vma, address, orig_pte); 2506 if (!old_page) { 2507 /* 2508 * VM_MIXEDMAP !pfn_valid() case 2509 * 2510 * We should not cow pages in a shared writeable mapping. 2511 * Just mark the pages writable as we can't do any dirty 2512 * accounting on raw pfn maps. 2513 */ 2514 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2515 (VM_WRITE|VM_SHARED)) 2516 goto reuse; 2517 goto gotten; 2518 } 2519 2520 /* 2521 * Take out anonymous pages first, anonymous shared vmas are 2522 * not dirty accountable. 2523 */ 2524 if (PageAnon(old_page) && !PageKsm(old_page)) { 2525 if (!trylock_page(old_page)) { 2526 page_cache_get(old_page); 2527 pte_unmap_unlock(page_table, ptl); 2528 lock_page(old_page); 2529 page_table = pte_offset_map_lock(mm, pmd, address, 2530 &ptl); 2531 if (!pte_same(*page_table, orig_pte)) { 2532 unlock_page(old_page); 2533 goto unlock; 2534 } 2535 page_cache_release(old_page); 2536 } 2537 if (reuse_swap_page(old_page)) { 2538 /* 2539 * The page is all ours. Move it to our anon_vma so 2540 * the rmap code will not search our parent or siblings. 2541 * Protected against the rmap code by the page lock. 2542 */ 2543 page_move_anon_rmap(old_page, vma, address); 2544 unlock_page(old_page); 2545 goto reuse; 2546 } 2547 unlock_page(old_page); 2548 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2549 (VM_WRITE|VM_SHARED))) { 2550 /* 2551 * Only catch write-faults on shared writable pages, 2552 * read-only shared pages can get COWed by 2553 * get_user_pages(.write=1, .force=1). 2554 */ 2555 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 2556 struct vm_fault vmf; 2557 int tmp; 2558 2559 vmf.virtual_address = (void __user *)(address & 2560 PAGE_MASK); 2561 vmf.pgoff = old_page->index; 2562 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2563 vmf.page = old_page; 2564 2565 /* 2566 * Notify the address space that the page is about to 2567 * become writable so that it can prohibit this or wait 2568 * for the page to get into an appropriate state. 2569 * 2570 * We do this without the lock held, so that it can 2571 * sleep if it needs to. 2572 */ 2573 page_cache_get(old_page); 2574 pte_unmap_unlock(page_table, ptl); 2575 2576 tmp = vma->vm_ops->page_mkwrite(vma, &vmf); 2577 if (unlikely(tmp & 2578 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 2579 ret = tmp; 2580 goto unwritable_page; 2581 } 2582 if (unlikely(!(tmp & VM_FAULT_LOCKED))) { 2583 lock_page(old_page); 2584 if (!old_page->mapping) { 2585 ret = 0; /* retry the fault */ 2586 unlock_page(old_page); 2587 goto unwritable_page; 2588 } 2589 } else 2590 VM_BUG_ON(!PageLocked(old_page)); 2591 2592 /* 2593 * Since we dropped the lock we need to revalidate 2594 * the PTE as someone else may have changed it. If 2595 * they did, we just return, as we can count on the 2596 * MMU to tell us if they didn't also make it writable. 2597 */ 2598 page_table = pte_offset_map_lock(mm, pmd, address, 2599 &ptl); 2600 if (!pte_same(*page_table, orig_pte)) { 2601 unlock_page(old_page); 2602 goto unlock; 2603 } 2604 2605 page_mkwrite = 1; 2606 } 2607 dirty_page = old_page; 2608 get_page(dirty_page); 2609 2610 reuse: 2611 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2612 entry = pte_mkyoung(orig_pte); 2613 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2614 if (ptep_set_access_flags(vma, address, page_table, entry,1)) 2615 update_mmu_cache(vma, address, page_table); 2616 pte_unmap_unlock(page_table, ptl); 2617 ret |= VM_FAULT_WRITE; 2618 2619 if (!dirty_page) 2620 return ret; 2621 2622 /* 2623 * Yes, Virginia, this is actually required to prevent a race 2624 * with clear_page_dirty_for_io() from clearing the page dirty 2625 * bit after it clear all dirty ptes, but before a racing 2626 * do_wp_page installs a dirty pte. 2627 * 2628 * __do_fault is protected similarly. 2629 */ 2630 if (!page_mkwrite) { 2631 wait_on_page_locked(dirty_page); 2632 set_page_dirty_balance(dirty_page, page_mkwrite); 2633 } 2634 put_page(dirty_page); 2635 if (page_mkwrite) { 2636 struct address_space *mapping = dirty_page->mapping; 2637 2638 set_page_dirty(dirty_page); 2639 unlock_page(dirty_page); 2640 page_cache_release(dirty_page); 2641 if (mapping) { 2642 /* 2643 * Some device drivers do not set page.mapping 2644 * but still dirty their pages 2645 */ 2646 balance_dirty_pages_ratelimited(mapping); 2647 } 2648 } 2649 2650 /* file_update_time outside page_lock */ 2651 if (vma->vm_file) 2652 file_update_time(vma->vm_file); 2653 2654 return ret; 2655 } 2656 2657 /* 2658 * Ok, we need to copy. Oh, well.. 2659 */ 2660 page_cache_get(old_page); 2661 gotten: 2662 pte_unmap_unlock(page_table, ptl); 2663 2664 if (unlikely(anon_vma_prepare(vma))) 2665 goto oom; 2666 2667 if (is_zero_pfn(pte_pfn(orig_pte))) { 2668 new_page = alloc_zeroed_user_highpage_movable(vma, address); 2669 if (!new_page) 2670 goto oom; 2671 } else { 2672 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 2673 if (!new_page) 2674 goto oom; 2675 cow_user_page(new_page, old_page, address, vma); 2676 } 2677 __SetPageUptodate(new_page); 2678 2679 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)) 2680 goto oom_free_new; 2681 2682 /* 2683 * Re-check the pte - we dropped the lock 2684 */ 2685 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2686 if (likely(pte_same(*page_table, orig_pte))) { 2687 if (old_page) { 2688 if (!PageAnon(old_page)) { 2689 dec_mm_counter_fast(mm, MM_FILEPAGES); 2690 inc_mm_counter_fast(mm, MM_ANONPAGES); 2691 } 2692 } else 2693 inc_mm_counter_fast(mm, MM_ANONPAGES); 2694 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2695 entry = mk_pte(new_page, vma->vm_page_prot); 2696 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2697 /* 2698 * Clear the pte entry and flush it first, before updating the 2699 * pte with the new entry. This will avoid a race condition 2700 * seen in the presence of one thread doing SMC and another 2701 * thread doing COW. 2702 */ 2703 ptep_clear_flush(vma, address, page_table); 2704 page_add_new_anon_rmap(new_page, vma, address); 2705 /* 2706 * We call the notify macro here because, when using secondary 2707 * mmu page tables (such as kvm shadow page tables), we want the 2708 * new page to be mapped directly into the secondary page table. 2709 */ 2710 set_pte_at_notify(mm, address, page_table, entry); 2711 update_mmu_cache(vma, address, page_table); 2712 if (old_page) { 2713 /* 2714 * Only after switching the pte to the new page may 2715 * we remove the mapcount here. Otherwise another 2716 * process may come and find the rmap count decremented 2717 * before the pte is switched to the new page, and 2718 * "reuse" the old page writing into it while our pte 2719 * here still points into it and can be read by other 2720 * threads. 2721 * 2722 * The critical issue is to order this 2723 * page_remove_rmap with the ptp_clear_flush above. 2724 * Those stores are ordered by (if nothing else,) 2725 * the barrier present in the atomic_add_negative 2726 * in page_remove_rmap. 2727 * 2728 * Then the TLB flush in ptep_clear_flush ensures that 2729 * no process can access the old page before the 2730 * decremented mapcount is visible. And the old page 2731 * cannot be reused until after the decremented 2732 * mapcount is visible. So transitively, TLBs to 2733 * old page will be flushed before it can be reused. 2734 */ 2735 page_remove_rmap(old_page); 2736 } 2737 2738 /* Free the old page.. */ 2739 new_page = old_page; 2740 ret |= VM_FAULT_WRITE; 2741 } else 2742 mem_cgroup_uncharge_page(new_page); 2743 2744 if (new_page) 2745 page_cache_release(new_page); 2746 unlock: 2747 pte_unmap_unlock(page_table, ptl); 2748 if (old_page) { 2749 /* 2750 * Don't let another task, with possibly unlocked vma, 2751 * keep the mlocked page. 2752 */ 2753 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) { 2754 lock_page(old_page); /* LRU manipulation */ 2755 munlock_vma_page(old_page); 2756 unlock_page(old_page); 2757 } 2758 page_cache_release(old_page); 2759 } 2760 return ret; 2761 oom_free_new: 2762 page_cache_release(new_page); 2763 oom: 2764 if (old_page) { 2765 if (page_mkwrite) { 2766 unlock_page(old_page); 2767 page_cache_release(old_page); 2768 } 2769 page_cache_release(old_page); 2770 } 2771 return VM_FAULT_OOM; 2772 2773 unwritable_page: 2774 page_cache_release(old_page); 2775 return ret; 2776 } 2777 2778 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 2779 unsigned long start_addr, unsigned long end_addr, 2780 struct zap_details *details) 2781 { 2782 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 2783 } 2784 2785 static inline void unmap_mapping_range_tree(struct prio_tree_root *root, 2786 struct zap_details *details) 2787 { 2788 struct vm_area_struct *vma; 2789 struct prio_tree_iter iter; 2790 pgoff_t vba, vea, zba, zea; 2791 2792 vma_prio_tree_foreach(vma, &iter, root, 2793 details->first_index, details->last_index) { 2794 2795 vba = vma->vm_pgoff; 2796 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1; 2797 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ 2798 zba = details->first_index; 2799 if (zba < vba) 2800 zba = vba; 2801 zea = details->last_index; 2802 if (zea > vea) 2803 zea = vea; 2804 2805 unmap_mapping_range_vma(vma, 2806 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 2807 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 2808 details); 2809 } 2810 } 2811 2812 static inline void unmap_mapping_range_list(struct list_head *head, 2813 struct zap_details *details) 2814 { 2815 struct vm_area_struct *vma; 2816 2817 /* 2818 * In nonlinear VMAs there is no correspondence between virtual address 2819 * offset and file offset. So we must perform an exhaustive search 2820 * across *all* the pages in each nonlinear VMA, not just the pages 2821 * whose virtual address lies outside the file truncation point. 2822 */ 2823 list_for_each_entry(vma, head, shared.vm_set.list) { 2824 details->nonlinear_vma = vma; 2825 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details); 2826 } 2827 } 2828 2829 /** 2830 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file. 2831 * @mapping: the address space containing mmaps to be unmapped. 2832 * @holebegin: byte in first page to unmap, relative to the start of 2833 * the underlying file. This will be rounded down to a PAGE_SIZE 2834 * boundary. Note that this is different from truncate_pagecache(), which 2835 * must keep the partial page. In contrast, we must get rid of 2836 * partial pages. 2837 * @holelen: size of prospective hole in bytes. This will be rounded 2838 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 2839 * end of the file. 2840 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 2841 * but 0 when invalidating pagecache, don't throw away private data. 2842 */ 2843 void unmap_mapping_range(struct address_space *mapping, 2844 loff_t const holebegin, loff_t const holelen, int even_cows) 2845 { 2846 struct zap_details details; 2847 pgoff_t hba = holebegin >> PAGE_SHIFT; 2848 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2849 2850 /* Check for overflow. */ 2851 if (sizeof(holelen) > sizeof(hlen)) { 2852 long long holeend = 2853 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2854 if (holeend & ~(long long)ULONG_MAX) 2855 hlen = ULONG_MAX - hba + 1; 2856 } 2857 2858 details.check_mapping = even_cows? NULL: mapping; 2859 details.nonlinear_vma = NULL; 2860 details.first_index = hba; 2861 details.last_index = hba + hlen - 1; 2862 if (details.last_index < details.first_index) 2863 details.last_index = ULONG_MAX; 2864 2865 2866 mutex_lock(&mapping->i_mmap_mutex); 2867 if (unlikely(!prio_tree_empty(&mapping->i_mmap))) 2868 unmap_mapping_range_tree(&mapping->i_mmap, &details); 2869 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) 2870 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details); 2871 mutex_unlock(&mapping->i_mmap_mutex); 2872 } 2873 EXPORT_SYMBOL(unmap_mapping_range); 2874 2875 /* 2876 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2877 * but allow concurrent faults), and pte mapped but not yet locked. 2878 * We return with mmap_sem still held, but pte unmapped and unlocked. 2879 */ 2880 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, 2881 unsigned long address, pte_t *page_table, pmd_t *pmd, 2882 unsigned int flags, pte_t orig_pte) 2883 { 2884 spinlock_t *ptl; 2885 struct page *page, *swapcache = NULL; 2886 swp_entry_t entry; 2887 pte_t pte; 2888 int locked; 2889 struct mem_cgroup *ptr; 2890 int exclusive = 0; 2891 int ret = 0; 2892 2893 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 2894 goto out; 2895 2896 entry = pte_to_swp_entry(orig_pte); 2897 if (unlikely(non_swap_entry(entry))) { 2898 if (is_migration_entry(entry)) { 2899 migration_entry_wait(mm, pmd, address); 2900 } else if (is_hwpoison_entry(entry)) { 2901 ret = VM_FAULT_HWPOISON; 2902 } else { 2903 print_bad_pte(vma, address, orig_pte, NULL); 2904 ret = VM_FAULT_SIGBUS; 2905 } 2906 goto out; 2907 } 2908 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 2909 page = lookup_swap_cache(entry); 2910 if (!page) { 2911 page = swapin_readahead(entry, 2912 GFP_HIGHUSER_MOVABLE, vma, address); 2913 if (!page) { 2914 /* 2915 * Back out if somebody else faulted in this pte 2916 * while we released the pte lock. 2917 */ 2918 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2919 if (likely(pte_same(*page_table, orig_pte))) 2920 ret = VM_FAULT_OOM; 2921 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2922 goto unlock; 2923 } 2924 2925 /* Had to read the page from swap area: Major fault */ 2926 ret = VM_FAULT_MAJOR; 2927 count_vm_event(PGMAJFAULT); 2928 mem_cgroup_count_vm_event(mm, PGMAJFAULT); 2929 } else if (PageHWPoison(page)) { 2930 /* 2931 * hwpoisoned dirty swapcache pages are kept for killing 2932 * owner processes (which may be unknown at hwpoison time) 2933 */ 2934 ret = VM_FAULT_HWPOISON; 2935 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2936 goto out_release; 2937 } 2938 2939 locked = lock_page_or_retry(page, mm, flags); 2940 2941 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2942 if (!locked) { 2943 ret |= VM_FAULT_RETRY; 2944 goto out_release; 2945 } 2946 2947 /* 2948 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 2949 * release the swapcache from under us. The page pin, and pte_same 2950 * test below, are not enough to exclude that. Even if it is still 2951 * swapcache, we need to check that the page's swap has not changed. 2952 */ 2953 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) 2954 goto out_page; 2955 2956 if (ksm_might_need_to_copy(page, vma, address)) { 2957 swapcache = page; 2958 page = ksm_does_need_to_copy(page, vma, address); 2959 2960 if (unlikely(!page)) { 2961 ret = VM_FAULT_OOM; 2962 page = swapcache; 2963 swapcache = NULL; 2964 goto out_page; 2965 } 2966 } 2967 2968 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) { 2969 ret = VM_FAULT_OOM; 2970 goto out_page; 2971 } 2972 2973 /* 2974 * Back out if somebody else already faulted in this pte. 2975 */ 2976 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2977 if (unlikely(!pte_same(*page_table, orig_pte))) 2978 goto out_nomap; 2979 2980 if (unlikely(!PageUptodate(page))) { 2981 ret = VM_FAULT_SIGBUS; 2982 goto out_nomap; 2983 } 2984 2985 /* 2986 * The page isn't present yet, go ahead with the fault. 2987 * 2988 * Be careful about the sequence of operations here. 2989 * To get its accounting right, reuse_swap_page() must be called 2990 * while the page is counted on swap but not yet in mapcount i.e. 2991 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 2992 * must be called after the swap_free(), or it will never succeed. 2993 * Because delete_from_swap_page() may be called by reuse_swap_page(), 2994 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry 2995 * in page->private. In this case, a record in swap_cgroup is silently 2996 * discarded at swap_free(). 2997 */ 2998 2999 inc_mm_counter_fast(mm, MM_ANONPAGES); 3000 dec_mm_counter_fast(mm, MM_SWAPENTS); 3001 pte = mk_pte(page, vma->vm_page_prot); 3002 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { 3003 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 3004 flags &= ~FAULT_FLAG_WRITE; 3005 ret |= VM_FAULT_WRITE; 3006 exclusive = 1; 3007 } 3008 flush_icache_page(vma, page); 3009 set_pte_at(mm, address, page_table, pte); 3010 do_page_add_anon_rmap(page, vma, address, exclusive); 3011 /* It's better to call commit-charge after rmap is established */ 3012 mem_cgroup_commit_charge_swapin(page, ptr); 3013 3014 swap_free(entry); 3015 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 3016 try_to_free_swap(page); 3017 unlock_page(page); 3018 if (swapcache) { 3019 /* 3020 * Hold the lock to avoid the swap entry to be reused 3021 * until we take the PT lock for the pte_same() check 3022 * (to avoid false positives from pte_same). For 3023 * further safety release the lock after the swap_free 3024 * so that the swap count won't change under a 3025 * parallel locked swapcache. 3026 */ 3027 unlock_page(swapcache); 3028 page_cache_release(swapcache); 3029 } 3030 3031 if (flags & FAULT_FLAG_WRITE) { 3032 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); 3033 if (ret & VM_FAULT_ERROR) 3034 ret &= VM_FAULT_ERROR; 3035 goto out; 3036 } 3037 3038 /* No need to invalidate - it was non-present before */ 3039 update_mmu_cache(vma, address, page_table); 3040 unlock: 3041 pte_unmap_unlock(page_table, ptl); 3042 out: 3043 return ret; 3044 out_nomap: 3045 mem_cgroup_cancel_charge_swapin(ptr); 3046 pte_unmap_unlock(page_table, ptl); 3047 out_page: 3048 unlock_page(page); 3049 out_release: 3050 page_cache_release(page); 3051 if (swapcache) { 3052 unlock_page(swapcache); 3053 page_cache_release(swapcache); 3054 } 3055 return ret; 3056 } 3057 3058 /* 3059 * This is like a special single-page "expand_{down|up}wards()", 3060 * except we must first make sure that 'address{-|+}PAGE_SIZE' 3061 * doesn't hit another vma. 3062 */ 3063 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) 3064 { 3065 address &= PAGE_MASK; 3066 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { 3067 struct vm_area_struct *prev = vma->vm_prev; 3068 3069 /* 3070 * Is there a mapping abutting this one below? 3071 * 3072 * That's only ok if it's the same stack mapping 3073 * that has gotten split.. 3074 */ 3075 if (prev && prev->vm_end == address) 3076 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; 3077 3078 expand_downwards(vma, address - PAGE_SIZE); 3079 } 3080 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { 3081 struct vm_area_struct *next = vma->vm_next; 3082 3083 /* As VM_GROWSDOWN but s/below/above/ */ 3084 if (next && next->vm_start == address + PAGE_SIZE) 3085 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; 3086 3087 expand_upwards(vma, address + PAGE_SIZE); 3088 } 3089 return 0; 3090 } 3091 3092 /* 3093 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3094 * but allow concurrent faults), and pte mapped but not yet locked. 3095 * We return with mmap_sem still held, but pte unmapped and unlocked. 3096 */ 3097 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 3098 unsigned long address, pte_t *page_table, pmd_t *pmd, 3099 unsigned int flags) 3100 { 3101 struct page *page; 3102 spinlock_t *ptl; 3103 pte_t entry; 3104 3105 pte_unmap(page_table); 3106 3107 /* Check if we need to add a guard page to the stack */ 3108 if (check_stack_guard_page(vma, address) < 0) 3109 return VM_FAULT_SIGBUS; 3110 3111 /* Use the zero-page for reads */ 3112 if (!(flags & FAULT_FLAG_WRITE)) { 3113 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), 3114 vma->vm_page_prot)); 3115 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 3116 if (!pte_none(*page_table)) 3117 goto unlock; 3118 goto setpte; 3119 } 3120 3121 /* Allocate our own private page. */ 3122 if (unlikely(anon_vma_prepare(vma))) 3123 goto oom; 3124 page = alloc_zeroed_user_highpage_movable(vma, address); 3125 if (!page) 3126 goto oom; 3127 __SetPageUptodate(page); 3128 3129 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) 3130 goto oom_free_page; 3131 3132 entry = mk_pte(page, vma->vm_page_prot); 3133 if (vma->vm_flags & VM_WRITE) 3134 entry = pte_mkwrite(pte_mkdirty(entry)); 3135 3136 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 3137 if (!pte_none(*page_table)) 3138 goto release; 3139 3140 inc_mm_counter_fast(mm, MM_ANONPAGES); 3141 page_add_new_anon_rmap(page, vma, address); 3142 setpte: 3143 set_pte_at(mm, address, page_table, entry); 3144 3145 /* No need to invalidate - it was non-present before */ 3146 update_mmu_cache(vma, address, page_table); 3147 unlock: 3148 pte_unmap_unlock(page_table, ptl); 3149 return 0; 3150 release: 3151 mem_cgroup_uncharge_page(page); 3152 page_cache_release(page); 3153 goto unlock; 3154 oom_free_page: 3155 page_cache_release(page); 3156 oom: 3157 return VM_FAULT_OOM; 3158 } 3159 3160 /* 3161 * __do_fault() tries to create a new page mapping. It aggressively 3162 * tries to share with existing pages, but makes a separate copy if 3163 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid 3164 * the next page fault. 3165 * 3166 * As this is called only for pages that do not currently exist, we 3167 * do not need to flush old virtual caches or the TLB. 3168 * 3169 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3170 * but allow concurrent faults), and pte neither mapped nor locked. 3171 * We return with mmap_sem still held, but pte unmapped and unlocked. 3172 */ 3173 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3174 unsigned long address, pmd_t *pmd, 3175 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 3176 { 3177 pte_t *page_table; 3178 spinlock_t *ptl; 3179 struct page *page; 3180 struct page *cow_page; 3181 pte_t entry; 3182 int anon = 0; 3183 struct page *dirty_page = NULL; 3184 struct vm_fault vmf; 3185 int ret; 3186 int page_mkwrite = 0; 3187 3188 /* 3189 * If we do COW later, allocate page befor taking lock_page() 3190 * on the file cache page. This will reduce lock holding time. 3191 */ 3192 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3193 3194 if (unlikely(anon_vma_prepare(vma))) 3195 return VM_FAULT_OOM; 3196 3197 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 3198 if (!cow_page) 3199 return VM_FAULT_OOM; 3200 3201 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) { 3202 page_cache_release(cow_page); 3203 return VM_FAULT_OOM; 3204 } 3205 } else 3206 cow_page = NULL; 3207 3208 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 3209 vmf.pgoff = pgoff; 3210 vmf.flags = flags; 3211 vmf.page = NULL; 3212 3213 ret = vma->vm_ops->fault(vma, &vmf); 3214 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 3215 VM_FAULT_RETRY))) 3216 goto uncharge_out; 3217 3218 if (unlikely(PageHWPoison(vmf.page))) { 3219 if (ret & VM_FAULT_LOCKED) 3220 unlock_page(vmf.page); 3221 ret = VM_FAULT_HWPOISON; 3222 goto uncharge_out; 3223 } 3224 3225 /* 3226 * For consistency in subsequent calls, make the faulted page always 3227 * locked. 3228 */ 3229 if (unlikely(!(ret & VM_FAULT_LOCKED))) 3230 lock_page(vmf.page); 3231 else 3232 VM_BUG_ON(!PageLocked(vmf.page)); 3233 3234 /* 3235 * Should we do an early C-O-W break? 3236 */ 3237 page = vmf.page; 3238 if (flags & FAULT_FLAG_WRITE) { 3239 if (!(vma->vm_flags & VM_SHARED)) { 3240 page = cow_page; 3241 anon = 1; 3242 copy_user_highpage(page, vmf.page, address, vma); 3243 __SetPageUptodate(page); 3244 } else { 3245 /* 3246 * If the page will be shareable, see if the backing 3247 * address space wants to know that the page is about 3248 * to become writable 3249 */ 3250 if (vma->vm_ops->page_mkwrite) { 3251 int tmp; 3252 3253 unlock_page(page); 3254 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 3255 tmp = vma->vm_ops->page_mkwrite(vma, &vmf); 3256 if (unlikely(tmp & 3257 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 3258 ret = tmp; 3259 goto unwritable_page; 3260 } 3261 if (unlikely(!(tmp & VM_FAULT_LOCKED))) { 3262 lock_page(page); 3263 if (!page->mapping) { 3264 ret = 0; /* retry the fault */ 3265 unlock_page(page); 3266 goto unwritable_page; 3267 } 3268 } else 3269 VM_BUG_ON(!PageLocked(page)); 3270 page_mkwrite = 1; 3271 } 3272 } 3273 3274 } 3275 3276 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 3277 3278 /* 3279 * This silly early PAGE_DIRTY setting removes a race 3280 * due to the bad i386 page protection. But it's valid 3281 * for other architectures too. 3282 * 3283 * Note that if FAULT_FLAG_WRITE is set, we either now have 3284 * an exclusive copy of the page, or this is a shared mapping, 3285 * so we can make it writable and dirty to avoid having to 3286 * handle that later. 3287 */ 3288 /* Only go through if we didn't race with anybody else... */ 3289 if (likely(pte_same(*page_table, orig_pte))) { 3290 flush_icache_page(vma, page); 3291 entry = mk_pte(page, vma->vm_page_prot); 3292 if (flags & FAULT_FLAG_WRITE) 3293 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3294 if (anon) { 3295 inc_mm_counter_fast(mm, MM_ANONPAGES); 3296 page_add_new_anon_rmap(page, vma, address); 3297 } else { 3298 inc_mm_counter_fast(mm, MM_FILEPAGES); 3299 page_add_file_rmap(page); 3300 if (flags & FAULT_FLAG_WRITE) { 3301 dirty_page = page; 3302 get_page(dirty_page); 3303 } 3304 } 3305 set_pte_at(mm, address, page_table, entry); 3306 3307 /* no need to invalidate: a not-present page won't be cached */ 3308 update_mmu_cache(vma, address, page_table); 3309 } else { 3310 if (cow_page) 3311 mem_cgroup_uncharge_page(cow_page); 3312 if (anon) 3313 page_cache_release(page); 3314 else 3315 anon = 1; /* no anon but release faulted_page */ 3316 } 3317 3318 pte_unmap_unlock(page_table, ptl); 3319 3320 if (dirty_page) { 3321 struct address_space *mapping = page->mapping; 3322 3323 if (set_page_dirty(dirty_page)) 3324 page_mkwrite = 1; 3325 unlock_page(dirty_page); 3326 put_page(dirty_page); 3327 if (page_mkwrite && mapping) { 3328 /* 3329 * Some device drivers do not set page.mapping but still 3330 * dirty their pages 3331 */ 3332 balance_dirty_pages_ratelimited(mapping); 3333 } 3334 3335 /* file_update_time outside page_lock */ 3336 if (vma->vm_file) 3337 file_update_time(vma->vm_file); 3338 } else { 3339 unlock_page(vmf.page); 3340 if (anon) 3341 page_cache_release(vmf.page); 3342 } 3343 3344 return ret; 3345 3346 unwritable_page: 3347 page_cache_release(page); 3348 return ret; 3349 uncharge_out: 3350 /* fs's fault handler get error */ 3351 if (cow_page) { 3352 mem_cgroup_uncharge_page(cow_page); 3353 page_cache_release(cow_page); 3354 } 3355 return ret; 3356 } 3357 3358 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3359 unsigned long address, pte_t *page_table, pmd_t *pmd, 3360 unsigned int flags, pte_t orig_pte) 3361 { 3362 pgoff_t pgoff = (((address & PAGE_MASK) 3363 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; 3364 3365 pte_unmap(page_table); 3366 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); 3367 } 3368 3369 /* 3370 * Fault of a previously existing named mapping. Repopulate the pte 3371 * from the encoded file_pte if possible. This enables swappable 3372 * nonlinear vmas. 3373 * 3374 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3375 * but allow concurrent faults), and pte mapped but not yet locked. 3376 * We return with mmap_sem still held, but pte unmapped and unlocked. 3377 */ 3378 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3379 unsigned long address, pte_t *page_table, pmd_t *pmd, 3380 unsigned int flags, pte_t orig_pte) 3381 { 3382 pgoff_t pgoff; 3383 3384 flags |= FAULT_FLAG_NONLINEAR; 3385 3386 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 3387 return 0; 3388 3389 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) { 3390 /* 3391 * Page table corrupted: show pte and kill process. 3392 */ 3393 print_bad_pte(vma, address, orig_pte, NULL); 3394 return VM_FAULT_SIGBUS; 3395 } 3396 3397 pgoff = pte_to_pgoff(orig_pte); 3398 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); 3399 } 3400 3401 /* 3402 * These routines also need to handle stuff like marking pages dirty 3403 * and/or accessed for architectures that don't do it in hardware (most 3404 * RISC architectures). The early dirtying is also good on the i386. 3405 * 3406 * There is also a hook called "update_mmu_cache()" that architectures 3407 * with external mmu caches can use to update those (ie the Sparc or 3408 * PowerPC hashed page tables that act as extended TLBs). 3409 * 3410 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3411 * but allow concurrent faults), and pte mapped but not yet locked. 3412 * We return with mmap_sem still held, but pte unmapped and unlocked. 3413 */ 3414 int handle_pte_fault(struct mm_struct *mm, 3415 struct vm_area_struct *vma, unsigned long address, 3416 pte_t *pte, pmd_t *pmd, unsigned int flags) 3417 { 3418 pte_t entry; 3419 spinlock_t *ptl; 3420 3421 entry = *pte; 3422 if (!pte_present(entry)) { 3423 if (pte_none(entry)) { 3424 if (vma->vm_ops) { 3425 if (likely(vma->vm_ops->fault)) 3426 return do_linear_fault(mm, vma, address, 3427 pte, pmd, flags, entry); 3428 } 3429 return do_anonymous_page(mm, vma, address, 3430 pte, pmd, flags); 3431 } 3432 if (pte_file(entry)) 3433 return do_nonlinear_fault(mm, vma, address, 3434 pte, pmd, flags, entry); 3435 return do_swap_page(mm, vma, address, 3436 pte, pmd, flags, entry); 3437 } 3438 3439 ptl = pte_lockptr(mm, pmd); 3440 spin_lock(ptl); 3441 if (unlikely(!pte_same(*pte, entry))) 3442 goto unlock; 3443 if (flags & FAULT_FLAG_WRITE) { 3444 if (!pte_write(entry)) 3445 return do_wp_page(mm, vma, address, 3446 pte, pmd, ptl, entry); 3447 entry = pte_mkdirty(entry); 3448 } 3449 entry = pte_mkyoung(entry); 3450 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) { 3451 update_mmu_cache(vma, address, pte); 3452 } else { 3453 /* 3454 * This is needed only for protection faults but the arch code 3455 * is not yet telling us if this is a protection fault or not. 3456 * This still avoids useless tlb flushes for .text page faults 3457 * with threads. 3458 */ 3459 if (flags & FAULT_FLAG_WRITE) 3460 flush_tlb_fix_spurious_fault(vma, address); 3461 } 3462 unlock: 3463 pte_unmap_unlock(pte, ptl); 3464 return 0; 3465 } 3466 3467 /* 3468 * By the time we get here, we already hold the mm semaphore 3469 */ 3470 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3471 unsigned long address, unsigned int flags) 3472 { 3473 pgd_t *pgd; 3474 pud_t *pud; 3475 pmd_t *pmd; 3476 pte_t *pte; 3477 3478 __set_current_state(TASK_RUNNING); 3479 3480 count_vm_event(PGFAULT); 3481 mem_cgroup_count_vm_event(mm, PGFAULT); 3482 3483 /* do counter updates before entering really critical section. */ 3484 check_sync_rss_stat(current); 3485 3486 if (unlikely(is_vm_hugetlb_page(vma))) 3487 return hugetlb_fault(mm, vma, address, flags); 3488 3489 retry: 3490 pgd = pgd_offset(mm, address); 3491 pud = pud_alloc(mm, pgd, address); 3492 if (!pud) 3493 return VM_FAULT_OOM; 3494 pmd = pmd_alloc(mm, pud, address); 3495 if (!pmd) 3496 return VM_FAULT_OOM; 3497 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { 3498 if (!vma->vm_ops) 3499 return do_huge_pmd_anonymous_page(mm, vma, address, 3500 pmd, flags); 3501 } else { 3502 pmd_t orig_pmd = *pmd; 3503 int ret; 3504 3505 barrier(); 3506 if (pmd_trans_huge(orig_pmd)) { 3507 if (flags & FAULT_FLAG_WRITE && 3508 !pmd_write(orig_pmd) && 3509 !pmd_trans_splitting(orig_pmd)) { 3510 ret = do_huge_pmd_wp_page(mm, vma, address, pmd, 3511 orig_pmd); 3512 /* 3513 * If COW results in an oom, the huge pmd will 3514 * have been split, so retry the fault on the 3515 * pte for a smaller charge. 3516 */ 3517 if (unlikely(ret & VM_FAULT_OOM)) 3518 goto retry; 3519 return ret; 3520 } 3521 return 0; 3522 } 3523 } 3524 3525 /* 3526 * Use __pte_alloc instead of pte_alloc_map, because we can't 3527 * run pte_offset_map on the pmd, if an huge pmd could 3528 * materialize from under us from a different thread. 3529 */ 3530 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address)) 3531 return VM_FAULT_OOM; 3532 /* if an huge pmd materialized from under us just retry later */ 3533 if (unlikely(pmd_trans_huge(*pmd))) 3534 return 0; 3535 /* 3536 * A regular pmd is established and it can't morph into a huge pmd 3537 * from under us anymore at this point because we hold the mmap_sem 3538 * read mode and khugepaged takes it in write mode. So now it's 3539 * safe to run pte_offset_map(). 3540 */ 3541 pte = pte_offset_map(pmd, address); 3542 3543 return handle_pte_fault(mm, vma, address, pte, pmd, flags); 3544 } 3545 3546 #ifndef __PAGETABLE_PUD_FOLDED 3547 /* 3548 * Allocate page upper directory. 3549 * We've already handled the fast-path in-line. 3550 */ 3551 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 3552 { 3553 pud_t *new = pud_alloc_one(mm, address); 3554 if (!new) 3555 return -ENOMEM; 3556 3557 smp_wmb(); /* See comment in __pte_alloc */ 3558 3559 spin_lock(&mm->page_table_lock); 3560 if (pgd_present(*pgd)) /* Another has populated it */ 3561 pud_free(mm, new); 3562 else 3563 pgd_populate(mm, pgd, new); 3564 spin_unlock(&mm->page_table_lock); 3565 return 0; 3566 } 3567 #endif /* __PAGETABLE_PUD_FOLDED */ 3568 3569 #ifndef __PAGETABLE_PMD_FOLDED 3570 /* 3571 * Allocate page middle directory. 3572 * We've already handled the fast-path in-line. 3573 */ 3574 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 3575 { 3576 pmd_t *new = pmd_alloc_one(mm, address); 3577 if (!new) 3578 return -ENOMEM; 3579 3580 smp_wmb(); /* See comment in __pte_alloc */ 3581 3582 spin_lock(&mm->page_table_lock); 3583 #ifndef __ARCH_HAS_4LEVEL_HACK 3584 if (pud_present(*pud)) /* Another has populated it */ 3585 pmd_free(mm, new); 3586 else 3587 pud_populate(mm, pud, new); 3588 #else 3589 if (pgd_present(*pud)) /* Another has populated it */ 3590 pmd_free(mm, new); 3591 else 3592 pgd_populate(mm, pud, new); 3593 #endif /* __ARCH_HAS_4LEVEL_HACK */ 3594 spin_unlock(&mm->page_table_lock); 3595 return 0; 3596 } 3597 #endif /* __PAGETABLE_PMD_FOLDED */ 3598 3599 int make_pages_present(unsigned long addr, unsigned long end) 3600 { 3601 int ret, len, write; 3602 struct vm_area_struct * vma; 3603 3604 vma = find_vma(current->mm, addr); 3605 if (!vma) 3606 return -ENOMEM; 3607 /* 3608 * We want to touch writable mappings with a write fault in order 3609 * to break COW, except for shared mappings because these don't COW 3610 * and we would not want to dirty them for nothing. 3611 */ 3612 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE; 3613 BUG_ON(addr >= end); 3614 BUG_ON(end > vma->vm_end); 3615 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE; 3616 ret = get_user_pages(current, current->mm, addr, 3617 len, write, 0, NULL, NULL); 3618 if (ret < 0) 3619 return ret; 3620 return ret == len ? 0 : -EFAULT; 3621 } 3622 3623 #if !defined(__HAVE_ARCH_GATE_AREA) 3624 3625 #if defined(AT_SYSINFO_EHDR) 3626 static struct vm_area_struct gate_vma; 3627 3628 static int __init gate_vma_init(void) 3629 { 3630 gate_vma.vm_mm = NULL; 3631 gate_vma.vm_start = FIXADDR_USER_START; 3632 gate_vma.vm_end = FIXADDR_USER_END; 3633 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC; 3634 gate_vma.vm_page_prot = __P101; 3635 3636 return 0; 3637 } 3638 __initcall(gate_vma_init); 3639 #endif 3640 3641 struct vm_area_struct *get_gate_vma(struct mm_struct *mm) 3642 { 3643 #ifdef AT_SYSINFO_EHDR 3644 return &gate_vma; 3645 #else 3646 return NULL; 3647 #endif 3648 } 3649 3650 int in_gate_area_no_mm(unsigned long addr) 3651 { 3652 #ifdef AT_SYSINFO_EHDR 3653 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END)) 3654 return 1; 3655 #endif 3656 return 0; 3657 } 3658 3659 #endif /* __HAVE_ARCH_GATE_AREA */ 3660 3661 static int __follow_pte(struct mm_struct *mm, unsigned long address, 3662 pte_t **ptepp, spinlock_t **ptlp) 3663 { 3664 pgd_t *pgd; 3665 pud_t *pud; 3666 pmd_t *pmd; 3667 pte_t *ptep; 3668 3669 pgd = pgd_offset(mm, address); 3670 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 3671 goto out; 3672 3673 pud = pud_offset(pgd, address); 3674 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 3675 goto out; 3676 3677 pmd = pmd_offset(pud, address); 3678 VM_BUG_ON(pmd_trans_huge(*pmd)); 3679 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 3680 goto out; 3681 3682 /* We cannot handle huge page PFN maps. Luckily they don't exist. */ 3683 if (pmd_huge(*pmd)) 3684 goto out; 3685 3686 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 3687 if (!ptep) 3688 goto out; 3689 if (!pte_present(*ptep)) 3690 goto unlock; 3691 *ptepp = ptep; 3692 return 0; 3693 unlock: 3694 pte_unmap_unlock(ptep, *ptlp); 3695 out: 3696 return -EINVAL; 3697 } 3698 3699 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 3700 pte_t **ptepp, spinlock_t **ptlp) 3701 { 3702 int res; 3703 3704 /* (void) is needed to make gcc happy */ 3705 (void) __cond_lock(*ptlp, 3706 !(res = __follow_pte(mm, address, ptepp, ptlp))); 3707 return res; 3708 } 3709 3710 /** 3711 * follow_pfn - look up PFN at a user virtual address 3712 * @vma: memory mapping 3713 * @address: user virtual address 3714 * @pfn: location to store found PFN 3715 * 3716 * Only IO mappings and raw PFN mappings are allowed. 3717 * 3718 * Returns zero and the pfn at @pfn on success, -ve otherwise. 3719 */ 3720 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 3721 unsigned long *pfn) 3722 { 3723 int ret = -EINVAL; 3724 spinlock_t *ptl; 3725 pte_t *ptep; 3726 3727 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3728 return ret; 3729 3730 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 3731 if (ret) 3732 return ret; 3733 *pfn = pte_pfn(*ptep); 3734 pte_unmap_unlock(ptep, ptl); 3735 return 0; 3736 } 3737 EXPORT_SYMBOL(follow_pfn); 3738 3739 #ifdef CONFIG_HAVE_IOREMAP_PROT 3740 int follow_phys(struct vm_area_struct *vma, 3741 unsigned long address, unsigned int flags, 3742 unsigned long *prot, resource_size_t *phys) 3743 { 3744 int ret = -EINVAL; 3745 pte_t *ptep, pte; 3746 spinlock_t *ptl; 3747 3748 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3749 goto out; 3750 3751 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 3752 goto out; 3753 pte = *ptep; 3754 3755 if ((flags & FOLL_WRITE) && !pte_write(pte)) 3756 goto unlock; 3757 3758 *prot = pgprot_val(pte_pgprot(pte)); 3759 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 3760 3761 ret = 0; 3762 unlock: 3763 pte_unmap_unlock(ptep, ptl); 3764 out: 3765 return ret; 3766 } 3767 3768 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 3769 void *buf, int len, int write) 3770 { 3771 resource_size_t phys_addr; 3772 unsigned long prot = 0; 3773 void __iomem *maddr; 3774 int offset = addr & (PAGE_SIZE-1); 3775 3776 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 3777 return -EINVAL; 3778 3779 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot); 3780 if (write) 3781 memcpy_toio(maddr + offset, buf, len); 3782 else 3783 memcpy_fromio(buf, maddr + offset, len); 3784 iounmap(maddr); 3785 3786 return len; 3787 } 3788 #endif 3789 3790 /* 3791 * Access another process' address space as given in mm. If non-NULL, use the 3792 * given task for page fault accounting. 3793 */ 3794 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, 3795 unsigned long addr, void *buf, int len, int write) 3796 { 3797 struct vm_area_struct *vma; 3798 void *old_buf = buf; 3799 3800 down_read(&mm->mmap_sem); 3801 /* ignore errors, just check how much was successfully transferred */ 3802 while (len) { 3803 int bytes, ret, offset; 3804 void *maddr; 3805 struct page *page = NULL; 3806 3807 ret = get_user_pages(tsk, mm, addr, 1, 3808 write, 1, &page, &vma); 3809 if (ret <= 0) { 3810 /* 3811 * Check if this is a VM_IO | VM_PFNMAP VMA, which 3812 * we can access using slightly different code. 3813 */ 3814 #ifdef CONFIG_HAVE_IOREMAP_PROT 3815 vma = find_vma(mm, addr); 3816 if (!vma || vma->vm_start > addr) 3817 break; 3818 if (vma->vm_ops && vma->vm_ops->access) 3819 ret = vma->vm_ops->access(vma, addr, buf, 3820 len, write); 3821 if (ret <= 0) 3822 #endif 3823 break; 3824 bytes = ret; 3825 } else { 3826 bytes = len; 3827 offset = addr & (PAGE_SIZE-1); 3828 if (bytes > PAGE_SIZE-offset) 3829 bytes = PAGE_SIZE-offset; 3830 3831 maddr = kmap(page); 3832 if (write) { 3833 copy_to_user_page(vma, page, addr, 3834 maddr + offset, buf, bytes); 3835 set_page_dirty_lock(page); 3836 } else { 3837 copy_from_user_page(vma, page, addr, 3838 buf, maddr + offset, bytes); 3839 } 3840 kunmap(page); 3841 page_cache_release(page); 3842 } 3843 len -= bytes; 3844 buf += bytes; 3845 addr += bytes; 3846 } 3847 up_read(&mm->mmap_sem); 3848 3849 return buf - old_buf; 3850 } 3851 3852 /** 3853 * access_remote_vm - access another process' address space 3854 * @mm: the mm_struct of the target address space 3855 * @addr: start address to access 3856 * @buf: source or destination buffer 3857 * @len: number of bytes to transfer 3858 * @write: whether the access is a write 3859 * 3860 * The caller must hold a reference on @mm. 3861 */ 3862 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 3863 void *buf, int len, int write) 3864 { 3865 return __access_remote_vm(NULL, mm, addr, buf, len, write); 3866 } 3867 3868 /* 3869 * Access another process' address space. 3870 * Source/target buffer must be kernel space, 3871 * Do not walk the page table directly, use get_user_pages 3872 */ 3873 int access_process_vm(struct task_struct *tsk, unsigned long addr, 3874 void *buf, int len, int write) 3875 { 3876 struct mm_struct *mm; 3877 int ret; 3878 3879 mm = get_task_mm(tsk); 3880 if (!mm) 3881 return 0; 3882 3883 ret = __access_remote_vm(tsk, mm, addr, buf, len, write); 3884 mmput(mm); 3885 3886 return ret; 3887 } 3888 3889 /* 3890 * Print the name of a VMA. 3891 */ 3892 void print_vma_addr(char *prefix, unsigned long ip) 3893 { 3894 struct mm_struct *mm = current->mm; 3895 struct vm_area_struct *vma; 3896 3897 /* 3898 * Do not print if we are in atomic 3899 * contexts (in exception stacks, etc.): 3900 */ 3901 if (preempt_count()) 3902 return; 3903 3904 down_read(&mm->mmap_sem); 3905 vma = find_vma(mm, ip); 3906 if (vma && vma->vm_file) { 3907 struct file *f = vma->vm_file; 3908 char *buf = (char *)__get_free_page(GFP_KERNEL); 3909 if (buf) { 3910 char *p, *s; 3911 3912 p = d_path(&f->f_path, buf, PAGE_SIZE); 3913 if (IS_ERR(p)) 3914 p = "?"; 3915 s = strrchr(p, '/'); 3916 if (s) 3917 p = s+1; 3918 printk("%s%s[%lx+%lx]", prefix, p, 3919 vma->vm_start, 3920 vma->vm_end - vma->vm_start); 3921 free_page((unsigned long)buf); 3922 } 3923 } 3924 up_read(¤t->mm->mmap_sem); 3925 } 3926 3927 #ifdef CONFIG_PROVE_LOCKING 3928 void might_fault(void) 3929 { 3930 /* 3931 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 3932 * holding the mmap_sem, this is safe because kernel memory doesn't 3933 * get paged out, therefore we'll never actually fault, and the 3934 * below annotations will generate false positives. 3935 */ 3936 if (segment_eq(get_fs(), KERNEL_DS)) 3937 return; 3938 3939 might_sleep(); 3940 /* 3941 * it would be nicer only to annotate paths which are not under 3942 * pagefault_disable, however that requires a larger audit and 3943 * providing helpers like get_user_atomic. 3944 */ 3945 if (!in_atomic() && current->mm) 3946 might_lock_read(¤t->mm->mmap_sem); 3947 } 3948 EXPORT_SYMBOL(might_fault); 3949 #endif 3950 3951 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 3952 static void clear_gigantic_page(struct page *page, 3953 unsigned long addr, 3954 unsigned int pages_per_huge_page) 3955 { 3956 int i; 3957 struct page *p = page; 3958 3959 might_sleep(); 3960 for (i = 0; i < pages_per_huge_page; 3961 i++, p = mem_map_next(p, page, i)) { 3962 cond_resched(); 3963 clear_user_highpage(p, addr + i * PAGE_SIZE); 3964 } 3965 } 3966 void clear_huge_page(struct page *page, 3967 unsigned long addr, unsigned int pages_per_huge_page) 3968 { 3969 int i; 3970 3971 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3972 clear_gigantic_page(page, addr, pages_per_huge_page); 3973 return; 3974 } 3975 3976 might_sleep(); 3977 for (i = 0; i < pages_per_huge_page; i++) { 3978 cond_resched(); 3979 clear_user_highpage(page + i, addr + i * PAGE_SIZE); 3980 } 3981 } 3982 3983 static void copy_user_gigantic_page(struct page *dst, struct page *src, 3984 unsigned long addr, 3985 struct vm_area_struct *vma, 3986 unsigned int pages_per_huge_page) 3987 { 3988 int i; 3989 struct page *dst_base = dst; 3990 struct page *src_base = src; 3991 3992 for (i = 0; i < pages_per_huge_page; ) { 3993 cond_resched(); 3994 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 3995 3996 i++; 3997 dst = mem_map_next(dst, dst_base, i); 3998 src = mem_map_next(src, src_base, i); 3999 } 4000 } 4001 4002 void copy_user_huge_page(struct page *dst, struct page *src, 4003 unsigned long addr, struct vm_area_struct *vma, 4004 unsigned int pages_per_huge_page) 4005 { 4006 int i; 4007 4008 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 4009 copy_user_gigantic_page(dst, src, addr, vma, 4010 pages_per_huge_page); 4011 return; 4012 } 4013 4014 might_sleep(); 4015 for (i = 0; i < pages_per_huge_page; i++) { 4016 cond_resched(); 4017 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 4018 } 4019 } 4020 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 4021