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 #include <linux/migrate.h> 61 #include <linux/string.h> 62 #include <linux/dma-debug.h> 63 #include <linux/debugfs.h> 64 65 #include <asm/io.h> 66 #include <asm/pgalloc.h> 67 #include <asm/uaccess.h> 68 #include <asm/tlb.h> 69 #include <asm/tlbflush.h> 70 #include <asm/pgtable.h> 71 72 #include "internal.h" 73 74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS 75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 76 #endif 77 78 #ifndef CONFIG_NEED_MULTIPLE_NODES 79 /* use the per-pgdat data instead for discontigmem - mbligh */ 80 unsigned long max_mapnr; 81 struct page *mem_map; 82 83 EXPORT_SYMBOL(max_mapnr); 84 EXPORT_SYMBOL(mem_map); 85 #endif 86 87 /* 88 * A number of key systems in x86 including ioremap() rely on the assumption 89 * that high_memory defines the upper bound on direct map memory, then end 90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 92 * and ZONE_HIGHMEM. 93 */ 94 void * high_memory; 95 96 EXPORT_SYMBOL(high_memory); 97 98 /* 99 * Randomize the address space (stacks, mmaps, brk, etc.). 100 * 101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 102 * as ancient (libc5 based) binaries can segfault. ) 103 */ 104 int randomize_va_space __read_mostly = 105 #ifdef CONFIG_COMPAT_BRK 106 1; 107 #else 108 2; 109 #endif 110 111 static int __init disable_randmaps(char *s) 112 { 113 randomize_va_space = 0; 114 return 1; 115 } 116 __setup("norandmaps", disable_randmaps); 117 118 unsigned long zero_pfn __read_mostly; 119 unsigned long highest_memmap_pfn __read_mostly; 120 121 EXPORT_SYMBOL(zero_pfn); 122 123 /* 124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 125 */ 126 static int __init init_zero_pfn(void) 127 { 128 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 129 return 0; 130 } 131 core_initcall(init_zero_pfn); 132 133 134 #if defined(SPLIT_RSS_COUNTING) 135 136 void sync_mm_rss(struct mm_struct *mm) 137 { 138 int i; 139 140 for (i = 0; i < NR_MM_COUNTERS; i++) { 141 if (current->rss_stat.count[i]) { 142 add_mm_counter(mm, i, current->rss_stat.count[i]); 143 current->rss_stat.count[i] = 0; 144 } 145 } 146 current->rss_stat.events = 0; 147 } 148 149 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 150 { 151 struct task_struct *task = current; 152 153 if (likely(task->mm == mm)) 154 task->rss_stat.count[member] += val; 155 else 156 add_mm_counter(mm, member, val); 157 } 158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 160 161 /* sync counter once per 64 page faults */ 162 #define TASK_RSS_EVENTS_THRESH (64) 163 static void check_sync_rss_stat(struct task_struct *task) 164 { 165 if (unlikely(task != current)) 166 return; 167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 168 sync_mm_rss(task->mm); 169 } 170 #else /* SPLIT_RSS_COUNTING */ 171 172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 174 175 static void check_sync_rss_stat(struct task_struct *task) 176 { 177 } 178 179 #endif /* SPLIT_RSS_COUNTING */ 180 181 #ifdef HAVE_GENERIC_MMU_GATHER 182 183 static int tlb_next_batch(struct mmu_gather *tlb) 184 { 185 struct mmu_gather_batch *batch; 186 187 batch = tlb->active; 188 if (batch->next) { 189 tlb->active = batch->next; 190 return 1; 191 } 192 193 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT) 194 return 0; 195 196 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0); 197 if (!batch) 198 return 0; 199 200 tlb->batch_count++; 201 batch->next = NULL; 202 batch->nr = 0; 203 batch->max = MAX_GATHER_BATCH; 204 205 tlb->active->next = batch; 206 tlb->active = batch; 207 208 return 1; 209 } 210 211 /* tlb_gather_mmu 212 * Called to initialize an (on-stack) mmu_gather structure for page-table 213 * tear-down from @mm. The @fullmm argument is used when @mm is without 214 * users and we're going to destroy the full address space (exit/execve). 215 */ 216 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end) 217 { 218 tlb->mm = mm; 219 220 /* Is it from 0 to ~0? */ 221 tlb->fullmm = !(start | (end+1)); 222 tlb->need_flush_all = 0; 223 tlb->local.next = NULL; 224 tlb->local.nr = 0; 225 tlb->local.max = ARRAY_SIZE(tlb->__pages); 226 tlb->active = &tlb->local; 227 tlb->batch_count = 0; 228 229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 230 tlb->batch = NULL; 231 #endif 232 233 __tlb_reset_range(tlb); 234 } 235 236 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb) 237 { 238 tlb_flush(tlb); 239 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end); 240 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 241 tlb_table_flush(tlb); 242 #endif 243 __tlb_reset_range(tlb); 244 } 245 246 static void tlb_flush_mmu_free(struct mmu_gather *tlb) 247 { 248 struct mmu_gather_batch *batch; 249 250 for (batch = &tlb->local; batch; batch = batch->next) { 251 free_pages_and_swap_cache(batch->pages, batch->nr); 252 batch->nr = 0; 253 } 254 tlb->active = &tlb->local; 255 } 256 257 void tlb_flush_mmu(struct mmu_gather *tlb) 258 { 259 if (!tlb->end) 260 return; 261 262 tlb_flush_mmu_tlbonly(tlb); 263 tlb_flush_mmu_free(tlb); 264 } 265 266 /* tlb_finish_mmu 267 * Called at the end of the shootdown operation to free up any resources 268 * that were required. 269 */ 270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) 271 { 272 struct mmu_gather_batch *batch, *next; 273 274 tlb_flush_mmu(tlb); 275 276 /* keep the page table cache within bounds */ 277 check_pgt_cache(); 278 279 for (batch = tlb->local.next; batch; batch = next) { 280 next = batch->next; 281 free_pages((unsigned long)batch, 0); 282 } 283 tlb->local.next = NULL; 284 } 285 286 /* __tlb_remove_page 287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while 288 * handling the additional races in SMP caused by other CPUs caching valid 289 * mappings in their TLBs. Returns the number of free page slots left. 290 * When out of page slots we must call tlb_flush_mmu(). 291 */ 292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) 293 { 294 struct mmu_gather_batch *batch; 295 296 VM_BUG_ON(!tlb->end); 297 298 batch = tlb->active; 299 batch->pages[batch->nr++] = page; 300 if (batch->nr == batch->max) { 301 if (!tlb_next_batch(tlb)) 302 return 0; 303 batch = tlb->active; 304 } 305 VM_BUG_ON_PAGE(batch->nr > batch->max, page); 306 307 return batch->max - batch->nr; 308 } 309 310 #endif /* HAVE_GENERIC_MMU_GATHER */ 311 312 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 313 314 /* 315 * See the comment near struct mmu_table_batch. 316 */ 317 318 static void tlb_remove_table_smp_sync(void *arg) 319 { 320 /* Simply deliver the interrupt */ 321 } 322 323 static void tlb_remove_table_one(void *table) 324 { 325 /* 326 * This isn't an RCU grace period and hence the page-tables cannot be 327 * assumed to be actually RCU-freed. 328 * 329 * It is however sufficient for software page-table walkers that rely on 330 * IRQ disabling. See the comment near struct mmu_table_batch. 331 */ 332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 333 __tlb_remove_table(table); 334 } 335 336 static void tlb_remove_table_rcu(struct rcu_head *head) 337 { 338 struct mmu_table_batch *batch; 339 int i; 340 341 batch = container_of(head, struct mmu_table_batch, rcu); 342 343 for (i = 0; i < batch->nr; i++) 344 __tlb_remove_table(batch->tables[i]); 345 346 free_page((unsigned long)batch); 347 } 348 349 void tlb_table_flush(struct mmu_gather *tlb) 350 { 351 struct mmu_table_batch **batch = &tlb->batch; 352 353 if (*batch) { 354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 355 *batch = NULL; 356 } 357 } 358 359 void tlb_remove_table(struct mmu_gather *tlb, void *table) 360 { 361 struct mmu_table_batch **batch = &tlb->batch; 362 363 /* 364 * When there's less then two users of this mm there cannot be a 365 * concurrent page-table walk. 366 */ 367 if (atomic_read(&tlb->mm->mm_users) < 2) { 368 __tlb_remove_table(table); 369 return; 370 } 371 372 if (*batch == NULL) { 373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); 374 if (*batch == NULL) { 375 tlb_remove_table_one(table); 376 return; 377 } 378 (*batch)->nr = 0; 379 } 380 (*batch)->tables[(*batch)->nr++] = table; 381 if ((*batch)->nr == MAX_TABLE_BATCH) 382 tlb_table_flush(tlb); 383 } 384 385 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ 386 387 /* 388 * Note: this doesn't free the actual pages themselves. That 389 * has been handled earlier when unmapping all the memory regions. 390 */ 391 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 392 unsigned long addr) 393 { 394 pgtable_t token = pmd_pgtable(*pmd); 395 pmd_clear(pmd); 396 pte_free_tlb(tlb, token, addr); 397 atomic_long_dec(&tlb->mm->nr_ptes); 398 } 399 400 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 401 unsigned long addr, unsigned long end, 402 unsigned long floor, unsigned long ceiling) 403 { 404 pmd_t *pmd; 405 unsigned long next; 406 unsigned long start; 407 408 start = addr; 409 pmd = pmd_offset(pud, addr); 410 do { 411 next = pmd_addr_end(addr, end); 412 if (pmd_none_or_clear_bad(pmd)) 413 continue; 414 free_pte_range(tlb, pmd, addr); 415 } while (pmd++, addr = next, addr != end); 416 417 start &= PUD_MASK; 418 if (start < floor) 419 return; 420 if (ceiling) { 421 ceiling &= PUD_MASK; 422 if (!ceiling) 423 return; 424 } 425 if (end - 1 > ceiling - 1) 426 return; 427 428 pmd = pmd_offset(pud, start); 429 pud_clear(pud); 430 pmd_free_tlb(tlb, pmd, start); 431 } 432 433 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 434 unsigned long addr, unsigned long end, 435 unsigned long floor, unsigned long ceiling) 436 { 437 pud_t *pud; 438 unsigned long next; 439 unsigned long start; 440 441 start = addr; 442 pud = pud_offset(pgd, addr); 443 do { 444 next = pud_addr_end(addr, end); 445 if (pud_none_or_clear_bad(pud)) 446 continue; 447 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 448 } while (pud++, addr = next, addr != end); 449 450 start &= PGDIR_MASK; 451 if (start < floor) 452 return; 453 if (ceiling) { 454 ceiling &= PGDIR_MASK; 455 if (!ceiling) 456 return; 457 } 458 if (end - 1 > ceiling - 1) 459 return; 460 461 pud = pud_offset(pgd, start); 462 pgd_clear(pgd); 463 pud_free_tlb(tlb, pud, start); 464 } 465 466 /* 467 * This function frees user-level page tables of a process. 468 */ 469 void free_pgd_range(struct mmu_gather *tlb, 470 unsigned long addr, unsigned long end, 471 unsigned long floor, unsigned long ceiling) 472 { 473 pgd_t *pgd; 474 unsigned long next; 475 476 /* 477 * The next few lines have given us lots of grief... 478 * 479 * Why are we testing PMD* at this top level? Because often 480 * there will be no work to do at all, and we'd prefer not to 481 * go all the way down to the bottom just to discover that. 482 * 483 * Why all these "- 1"s? Because 0 represents both the bottom 484 * of the address space and the top of it (using -1 for the 485 * top wouldn't help much: the masks would do the wrong thing). 486 * The rule is that addr 0 and floor 0 refer to the bottom of 487 * the address space, but end 0 and ceiling 0 refer to the top 488 * Comparisons need to use "end - 1" and "ceiling - 1" (though 489 * that end 0 case should be mythical). 490 * 491 * Wherever addr is brought up or ceiling brought down, we must 492 * be careful to reject "the opposite 0" before it confuses the 493 * subsequent tests. But what about where end is brought down 494 * by PMD_SIZE below? no, end can't go down to 0 there. 495 * 496 * Whereas we round start (addr) and ceiling down, by different 497 * masks at different levels, in order to test whether a table 498 * now has no other vmas using it, so can be freed, we don't 499 * bother to round floor or end up - the tests don't need that. 500 */ 501 502 addr &= PMD_MASK; 503 if (addr < floor) { 504 addr += PMD_SIZE; 505 if (!addr) 506 return; 507 } 508 if (ceiling) { 509 ceiling &= PMD_MASK; 510 if (!ceiling) 511 return; 512 } 513 if (end - 1 > ceiling - 1) 514 end -= PMD_SIZE; 515 if (addr > end - 1) 516 return; 517 518 pgd = pgd_offset(tlb->mm, addr); 519 do { 520 next = pgd_addr_end(addr, end); 521 if (pgd_none_or_clear_bad(pgd)) 522 continue; 523 free_pud_range(tlb, pgd, addr, next, floor, ceiling); 524 } while (pgd++, addr = next, addr != end); 525 } 526 527 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 528 unsigned long floor, unsigned long ceiling) 529 { 530 while (vma) { 531 struct vm_area_struct *next = vma->vm_next; 532 unsigned long addr = vma->vm_start; 533 534 /* 535 * Hide vma from rmap and truncate_pagecache before freeing 536 * pgtables 537 */ 538 unlink_anon_vmas(vma); 539 unlink_file_vma(vma); 540 541 if (is_vm_hugetlb_page(vma)) { 542 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 543 floor, next? next->vm_start: ceiling); 544 } else { 545 /* 546 * Optimization: gather nearby vmas into one call down 547 */ 548 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 549 && !is_vm_hugetlb_page(next)) { 550 vma = next; 551 next = vma->vm_next; 552 unlink_anon_vmas(vma); 553 unlink_file_vma(vma); 554 } 555 free_pgd_range(tlb, addr, vma->vm_end, 556 floor, next? next->vm_start: ceiling); 557 } 558 vma = next; 559 } 560 } 561 562 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 563 pmd_t *pmd, unsigned long address) 564 { 565 spinlock_t *ptl; 566 pgtable_t new = pte_alloc_one(mm, address); 567 int wait_split_huge_page; 568 if (!new) 569 return -ENOMEM; 570 571 /* 572 * Ensure all pte setup (eg. pte page lock and page clearing) are 573 * visible before the pte is made visible to other CPUs by being 574 * put into page tables. 575 * 576 * The other side of the story is the pointer chasing in the page 577 * table walking code (when walking the page table without locking; 578 * ie. most of the time). Fortunately, these data accesses consist 579 * of a chain of data-dependent loads, meaning most CPUs (alpha 580 * being the notable exception) will already guarantee loads are 581 * seen in-order. See the alpha page table accessors for the 582 * smp_read_barrier_depends() barriers in page table walking code. 583 */ 584 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 585 586 ptl = pmd_lock(mm, pmd); 587 wait_split_huge_page = 0; 588 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 589 atomic_long_inc(&mm->nr_ptes); 590 pmd_populate(mm, pmd, new); 591 new = NULL; 592 } else if (unlikely(pmd_trans_splitting(*pmd))) 593 wait_split_huge_page = 1; 594 spin_unlock(ptl); 595 if (new) 596 pte_free(mm, new); 597 if (wait_split_huge_page) 598 wait_split_huge_page(vma->anon_vma, pmd); 599 return 0; 600 } 601 602 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address) 603 { 604 pte_t *new = pte_alloc_one_kernel(&init_mm, address); 605 if (!new) 606 return -ENOMEM; 607 608 smp_wmb(); /* See comment in __pte_alloc */ 609 610 spin_lock(&init_mm.page_table_lock); 611 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 612 pmd_populate_kernel(&init_mm, pmd, new); 613 new = NULL; 614 } else 615 VM_BUG_ON(pmd_trans_splitting(*pmd)); 616 spin_unlock(&init_mm.page_table_lock); 617 if (new) 618 pte_free_kernel(&init_mm, new); 619 return 0; 620 } 621 622 static inline void init_rss_vec(int *rss) 623 { 624 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 625 } 626 627 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 628 { 629 int i; 630 631 if (current->mm == mm) 632 sync_mm_rss(mm); 633 for (i = 0; i < NR_MM_COUNTERS; i++) 634 if (rss[i]) 635 add_mm_counter(mm, i, rss[i]); 636 } 637 638 /* 639 * This function is called to print an error when a bad pte 640 * is found. For example, we might have a PFN-mapped pte in 641 * a region that doesn't allow it. 642 * 643 * The calling function must still handle the error. 644 */ 645 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 646 pte_t pte, struct page *page) 647 { 648 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 649 pud_t *pud = pud_offset(pgd, addr); 650 pmd_t *pmd = pmd_offset(pud, addr); 651 struct address_space *mapping; 652 pgoff_t index; 653 static unsigned long resume; 654 static unsigned long nr_shown; 655 static unsigned long nr_unshown; 656 657 /* 658 * Allow a burst of 60 reports, then keep quiet for that minute; 659 * or allow a steady drip of one report per second. 660 */ 661 if (nr_shown == 60) { 662 if (time_before(jiffies, resume)) { 663 nr_unshown++; 664 return; 665 } 666 if (nr_unshown) { 667 printk(KERN_ALERT 668 "BUG: Bad page map: %lu messages suppressed\n", 669 nr_unshown); 670 nr_unshown = 0; 671 } 672 nr_shown = 0; 673 } 674 if (nr_shown++ == 0) 675 resume = jiffies + 60 * HZ; 676 677 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 678 index = linear_page_index(vma, addr); 679 680 printk(KERN_ALERT 681 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 682 current->comm, 683 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 684 if (page) 685 dump_page(page, "bad pte"); 686 printk(KERN_ALERT 687 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", 688 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 689 /* 690 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y 691 */ 692 if (vma->vm_ops) 693 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n", 694 vma->vm_ops->fault); 695 if (vma->vm_file) 696 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n", 697 vma->vm_file->f_op->mmap); 698 dump_stack(); 699 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 700 } 701 702 /* 703 * vm_normal_page -- This function gets the "struct page" associated with a pte. 704 * 705 * "Special" mappings do not wish to be associated with a "struct page" (either 706 * it doesn't exist, or it exists but they don't want to touch it). In this 707 * case, NULL is returned here. "Normal" mappings do have a struct page. 708 * 709 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 710 * pte bit, in which case this function is trivial. Secondly, an architecture 711 * may not have a spare pte bit, which requires a more complicated scheme, 712 * described below. 713 * 714 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 715 * special mapping (even if there are underlying and valid "struct pages"). 716 * COWed pages of a VM_PFNMAP are always normal. 717 * 718 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 719 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 720 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 721 * mapping will always honor the rule 722 * 723 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 724 * 725 * And for normal mappings this is false. 726 * 727 * This restricts such mappings to be a linear translation from virtual address 728 * to pfn. To get around this restriction, we allow arbitrary mappings so long 729 * as the vma is not a COW mapping; in that case, we know that all ptes are 730 * special (because none can have been COWed). 731 * 732 * 733 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 734 * 735 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 736 * page" backing, however the difference is that _all_ pages with a struct 737 * page (that is, those where pfn_valid is true) are refcounted and considered 738 * normal pages by the VM. The disadvantage is that pages are refcounted 739 * (which can be slower and simply not an option for some PFNMAP users). The 740 * advantage is that we don't have to follow the strict linearity rule of 741 * PFNMAP mappings in order to support COWable mappings. 742 * 743 */ 744 #ifdef __HAVE_ARCH_PTE_SPECIAL 745 # define HAVE_PTE_SPECIAL 1 746 #else 747 # define HAVE_PTE_SPECIAL 0 748 #endif 749 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 750 pte_t pte) 751 { 752 unsigned long pfn = pte_pfn(pte); 753 754 if (HAVE_PTE_SPECIAL) { 755 if (likely(!pte_special(pte))) 756 goto check_pfn; 757 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 758 return NULL; 759 if (!is_zero_pfn(pfn)) 760 print_bad_pte(vma, addr, pte, NULL); 761 return NULL; 762 } 763 764 /* !HAVE_PTE_SPECIAL case follows: */ 765 766 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 767 if (vma->vm_flags & VM_MIXEDMAP) { 768 if (!pfn_valid(pfn)) 769 return NULL; 770 goto out; 771 } else { 772 unsigned long off; 773 off = (addr - vma->vm_start) >> PAGE_SHIFT; 774 if (pfn == vma->vm_pgoff + off) 775 return NULL; 776 if (!is_cow_mapping(vma->vm_flags)) 777 return NULL; 778 } 779 } 780 781 if (is_zero_pfn(pfn)) 782 return NULL; 783 check_pfn: 784 if (unlikely(pfn > highest_memmap_pfn)) { 785 print_bad_pte(vma, addr, pte, NULL); 786 return NULL; 787 } 788 789 /* 790 * NOTE! We still have PageReserved() pages in the page tables. 791 * eg. VDSO mappings can cause them to exist. 792 */ 793 out: 794 return pfn_to_page(pfn); 795 } 796 797 /* 798 * copy one vm_area from one task to the other. Assumes the page tables 799 * already present in the new task to be cleared in the whole range 800 * covered by this vma. 801 */ 802 803 static inline unsigned long 804 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 805 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 806 unsigned long addr, int *rss) 807 { 808 unsigned long vm_flags = vma->vm_flags; 809 pte_t pte = *src_pte; 810 struct page *page; 811 812 /* pte contains position in swap or file, so copy. */ 813 if (unlikely(!pte_present(pte))) { 814 if (!pte_file(pte)) { 815 swp_entry_t entry = pte_to_swp_entry(pte); 816 817 if (likely(!non_swap_entry(entry))) { 818 if (swap_duplicate(entry) < 0) 819 return entry.val; 820 821 /* make sure dst_mm is on swapoff's mmlist. */ 822 if (unlikely(list_empty(&dst_mm->mmlist))) { 823 spin_lock(&mmlist_lock); 824 if (list_empty(&dst_mm->mmlist)) 825 list_add(&dst_mm->mmlist, 826 &src_mm->mmlist); 827 spin_unlock(&mmlist_lock); 828 } 829 rss[MM_SWAPENTS]++; 830 } else if (is_migration_entry(entry)) { 831 page = migration_entry_to_page(entry); 832 833 if (PageAnon(page)) 834 rss[MM_ANONPAGES]++; 835 else 836 rss[MM_FILEPAGES]++; 837 838 if (is_write_migration_entry(entry) && 839 is_cow_mapping(vm_flags)) { 840 /* 841 * COW mappings require pages in both 842 * parent and child to be set to read. 843 */ 844 make_migration_entry_read(&entry); 845 pte = swp_entry_to_pte(entry); 846 if (pte_swp_soft_dirty(*src_pte)) 847 pte = pte_swp_mksoft_dirty(pte); 848 set_pte_at(src_mm, addr, src_pte, pte); 849 } 850 } 851 } 852 goto out_set_pte; 853 } 854 855 /* 856 * If it's a COW mapping, write protect it both 857 * in the parent and the child 858 */ 859 if (is_cow_mapping(vm_flags)) { 860 ptep_set_wrprotect(src_mm, addr, src_pte); 861 pte = pte_wrprotect(pte); 862 } 863 864 /* 865 * If it's a shared mapping, mark it clean in 866 * the child 867 */ 868 if (vm_flags & VM_SHARED) 869 pte = pte_mkclean(pte); 870 pte = pte_mkold(pte); 871 872 page = vm_normal_page(vma, addr, pte); 873 if (page) { 874 get_page(page); 875 page_dup_rmap(page); 876 if (PageAnon(page)) 877 rss[MM_ANONPAGES]++; 878 else 879 rss[MM_FILEPAGES]++; 880 } 881 882 out_set_pte: 883 set_pte_at(dst_mm, addr, dst_pte, pte); 884 return 0; 885 } 886 887 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 888 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 889 unsigned long addr, unsigned long end) 890 { 891 pte_t *orig_src_pte, *orig_dst_pte; 892 pte_t *src_pte, *dst_pte; 893 spinlock_t *src_ptl, *dst_ptl; 894 int progress = 0; 895 int rss[NR_MM_COUNTERS]; 896 swp_entry_t entry = (swp_entry_t){0}; 897 898 again: 899 init_rss_vec(rss); 900 901 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 902 if (!dst_pte) 903 return -ENOMEM; 904 src_pte = pte_offset_map(src_pmd, addr); 905 src_ptl = pte_lockptr(src_mm, src_pmd); 906 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 907 orig_src_pte = src_pte; 908 orig_dst_pte = dst_pte; 909 arch_enter_lazy_mmu_mode(); 910 911 do { 912 /* 913 * We are holding two locks at this point - either of them 914 * could generate latencies in another task on another CPU. 915 */ 916 if (progress >= 32) { 917 progress = 0; 918 if (need_resched() || 919 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 920 break; 921 } 922 if (pte_none(*src_pte)) { 923 progress++; 924 continue; 925 } 926 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, 927 vma, addr, rss); 928 if (entry.val) 929 break; 930 progress += 8; 931 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 932 933 arch_leave_lazy_mmu_mode(); 934 spin_unlock(src_ptl); 935 pte_unmap(orig_src_pte); 936 add_mm_rss_vec(dst_mm, rss); 937 pte_unmap_unlock(orig_dst_pte, dst_ptl); 938 cond_resched(); 939 940 if (entry.val) { 941 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) 942 return -ENOMEM; 943 progress = 0; 944 } 945 if (addr != end) 946 goto again; 947 return 0; 948 } 949 950 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 951 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 952 unsigned long addr, unsigned long end) 953 { 954 pmd_t *src_pmd, *dst_pmd; 955 unsigned long next; 956 957 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 958 if (!dst_pmd) 959 return -ENOMEM; 960 src_pmd = pmd_offset(src_pud, addr); 961 do { 962 next = pmd_addr_end(addr, end); 963 if (pmd_trans_huge(*src_pmd)) { 964 int err; 965 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE); 966 err = copy_huge_pmd(dst_mm, src_mm, 967 dst_pmd, src_pmd, addr, vma); 968 if (err == -ENOMEM) 969 return -ENOMEM; 970 if (!err) 971 continue; 972 /* fall through */ 973 } 974 if (pmd_none_or_clear_bad(src_pmd)) 975 continue; 976 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 977 vma, addr, next)) 978 return -ENOMEM; 979 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 980 return 0; 981 } 982 983 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 984 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 985 unsigned long addr, unsigned long end) 986 { 987 pud_t *src_pud, *dst_pud; 988 unsigned long next; 989 990 dst_pud = pud_alloc(dst_mm, dst_pgd, addr); 991 if (!dst_pud) 992 return -ENOMEM; 993 src_pud = pud_offset(src_pgd, addr); 994 do { 995 next = pud_addr_end(addr, end); 996 if (pud_none_or_clear_bad(src_pud)) 997 continue; 998 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 999 vma, addr, next)) 1000 return -ENOMEM; 1001 } while (dst_pud++, src_pud++, addr = next, addr != end); 1002 return 0; 1003 } 1004 1005 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1006 struct vm_area_struct *vma) 1007 { 1008 pgd_t *src_pgd, *dst_pgd; 1009 unsigned long next; 1010 unsigned long addr = vma->vm_start; 1011 unsigned long end = vma->vm_end; 1012 unsigned long mmun_start; /* For mmu_notifiers */ 1013 unsigned long mmun_end; /* For mmu_notifiers */ 1014 bool is_cow; 1015 int ret; 1016 1017 /* 1018 * Don't copy ptes where a page fault will fill them correctly. 1019 * Fork becomes much lighter when there are big shared or private 1020 * readonly mappings. The tradeoff is that copy_page_range is more 1021 * efficient than faulting. 1022 */ 1023 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR | 1024 VM_PFNMAP | VM_MIXEDMAP))) { 1025 if (!vma->anon_vma) 1026 return 0; 1027 } 1028 1029 if (is_vm_hugetlb_page(vma)) 1030 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 1031 1032 if (unlikely(vma->vm_flags & VM_PFNMAP)) { 1033 /* 1034 * We do not free on error cases below as remove_vma 1035 * gets called on error from higher level routine 1036 */ 1037 ret = track_pfn_copy(vma); 1038 if (ret) 1039 return ret; 1040 } 1041 1042 /* 1043 * We need to invalidate the secondary MMU mappings only when 1044 * there could be a permission downgrade on the ptes of the 1045 * parent mm. And a permission downgrade will only happen if 1046 * is_cow_mapping() returns true. 1047 */ 1048 is_cow = is_cow_mapping(vma->vm_flags); 1049 mmun_start = addr; 1050 mmun_end = end; 1051 if (is_cow) 1052 mmu_notifier_invalidate_range_start(src_mm, mmun_start, 1053 mmun_end); 1054 1055 ret = 0; 1056 dst_pgd = pgd_offset(dst_mm, addr); 1057 src_pgd = pgd_offset(src_mm, addr); 1058 do { 1059 next = pgd_addr_end(addr, end); 1060 if (pgd_none_or_clear_bad(src_pgd)) 1061 continue; 1062 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, 1063 vma, addr, next))) { 1064 ret = -ENOMEM; 1065 break; 1066 } 1067 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1068 1069 if (is_cow) 1070 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end); 1071 return ret; 1072 } 1073 1074 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1075 struct vm_area_struct *vma, pmd_t *pmd, 1076 unsigned long addr, unsigned long end, 1077 struct zap_details *details) 1078 { 1079 struct mm_struct *mm = tlb->mm; 1080 int force_flush = 0; 1081 int rss[NR_MM_COUNTERS]; 1082 spinlock_t *ptl; 1083 pte_t *start_pte; 1084 pte_t *pte; 1085 1086 again: 1087 init_rss_vec(rss); 1088 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1089 pte = start_pte; 1090 arch_enter_lazy_mmu_mode(); 1091 do { 1092 pte_t ptent = *pte; 1093 if (pte_none(ptent)) { 1094 continue; 1095 } 1096 1097 if (pte_present(ptent)) { 1098 struct page *page; 1099 1100 page = vm_normal_page(vma, addr, ptent); 1101 if (unlikely(details) && page) { 1102 /* 1103 * unmap_shared_mapping_pages() wants to 1104 * invalidate cache without truncating: 1105 * unmap shared but keep private pages. 1106 */ 1107 if (details->check_mapping && 1108 details->check_mapping != page->mapping) 1109 continue; 1110 /* 1111 * Each page->index must be checked when 1112 * invalidating or truncating nonlinear. 1113 */ 1114 if (details->nonlinear_vma && 1115 (page->index < details->first_index || 1116 page->index > details->last_index)) 1117 continue; 1118 } 1119 ptent = ptep_get_and_clear_full(mm, addr, pte, 1120 tlb->fullmm); 1121 tlb_remove_tlb_entry(tlb, pte, addr); 1122 if (unlikely(!page)) 1123 continue; 1124 if (unlikely(details) && details->nonlinear_vma 1125 && linear_page_index(details->nonlinear_vma, 1126 addr) != page->index) { 1127 pte_t ptfile = pgoff_to_pte(page->index); 1128 if (pte_soft_dirty(ptent)) 1129 ptfile = pte_file_mksoft_dirty(ptfile); 1130 set_pte_at(mm, addr, pte, ptfile); 1131 } 1132 if (PageAnon(page)) 1133 rss[MM_ANONPAGES]--; 1134 else { 1135 if (pte_dirty(ptent)) { 1136 force_flush = 1; 1137 set_page_dirty(page); 1138 } 1139 if (pte_young(ptent) && 1140 likely(!(vma->vm_flags & VM_SEQ_READ))) 1141 mark_page_accessed(page); 1142 rss[MM_FILEPAGES]--; 1143 } 1144 page_remove_rmap(page); 1145 if (unlikely(page_mapcount(page) < 0)) 1146 print_bad_pte(vma, addr, ptent, page); 1147 if (unlikely(!__tlb_remove_page(tlb, page))) { 1148 force_flush = 1; 1149 addr += PAGE_SIZE; 1150 break; 1151 } 1152 continue; 1153 } 1154 /* 1155 * If details->check_mapping, we leave swap entries; 1156 * if details->nonlinear_vma, we leave file entries. 1157 */ 1158 if (unlikely(details)) 1159 continue; 1160 if (pte_file(ptent)) { 1161 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) 1162 print_bad_pte(vma, addr, ptent, NULL); 1163 } else { 1164 swp_entry_t entry = pte_to_swp_entry(ptent); 1165 1166 if (!non_swap_entry(entry)) 1167 rss[MM_SWAPENTS]--; 1168 else if (is_migration_entry(entry)) { 1169 struct page *page; 1170 1171 page = migration_entry_to_page(entry); 1172 1173 if (PageAnon(page)) 1174 rss[MM_ANONPAGES]--; 1175 else 1176 rss[MM_FILEPAGES]--; 1177 } 1178 if (unlikely(!free_swap_and_cache(entry))) 1179 print_bad_pte(vma, addr, ptent, NULL); 1180 } 1181 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1182 } while (pte++, addr += PAGE_SIZE, addr != end); 1183 1184 add_mm_rss_vec(mm, rss); 1185 arch_leave_lazy_mmu_mode(); 1186 1187 /* Do the actual TLB flush before dropping ptl */ 1188 if (force_flush) 1189 tlb_flush_mmu_tlbonly(tlb); 1190 pte_unmap_unlock(start_pte, ptl); 1191 1192 /* 1193 * If we forced a TLB flush (either due to running out of 1194 * batch buffers or because we needed to flush dirty TLB 1195 * entries before releasing the ptl), free the batched 1196 * memory too. Restart if we didn't do everything. 1197 */ 1198 if (force_flush) { 1199 force_flush = 0; 1200 tlb_flush_mmu_free(tlb); 1201 1202 if (addr != end) 1203 goto again; 1204 } 1205 1206 return addr; 1207 } 1208 1209 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1210 struct vm_area_struct *vma, pud_t *pud, 1211 unsigned long addr, unsigned long end, 1212 struct zap_details *details) 1213 { 1214 pmd_t *pmd; 1215 unsigned long next; 1216 1217 pmd = pmd_offset(pud, addr); 1218 do { 1219 next = pmd_addr_end(addr, end); 1220 if (pmd_trans_huge(*pmd)) { 1221 if (next - addr != HPAGE_PMD_SIZE) { 1222 #ifdef CONFIG_DEBUG_VM 1223 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) { 1224 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n", 1225 __func__, addr, end, 1226 vma->vm_start, 1227 vma->vm_end); 1228 BUG(); 1229 } 1230 #endif 1231 split_huge_page_pmd(vma, addr, pmd); 1232 } else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1233 goto next; 1234 /* fall through */ 1235 } 1236 /* 1237 * Here there can be other concurrent MADV_DONTNEED or 1238 * trans huge page faults running, and if the pmd is 1239 * none or trans huge it can change under us. This is 1240 * because MADV_DONTNEED holds the mmap_sem in read 1241 * mode. 1242 */ 1243 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1244 goto next; 1245 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1246 next: 1247 cond_resched(); 1248 } while (pmd++, addr = next, addr != end); 1249 1250 return addr; 1251 } 1252 1253 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1254 struct vm_area_struct *vma, pgd_t *pgd, 1255 unsigned long addr, unsigned long end, 1256 struct zap_details *details) 1257 { 1258 pud_t *pud; 1259 unsigned long next; 1260 1261 pud = pud_offset(pgd, addr); 1262 do { 1263 next = pud_addr_end(addr, end); 1264 if (pud_none_or_clear_bad(pud)) 1265 continue; 1266 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1267 } while (pud++, addr = next, addr != end); 1268 1269 return addr; 1270 } 1271 1272 static void unmap_page_range(struct mmu_gather *tlb, 1273 struct vm_area_struct *vma, 1274 unsigned long addr, unsigned long end, 1275 struct zap_details *details) 1276 { 1277 pgd_t *pgd; 1278 unsigned long next; 1279 1280 if (details && !details->check_mapping && !details->nonlinear_vma) 1281 details = NULL; 1282 1283 BUG_ON(addr >= end); 1284 tlb_start_vma(tlb, vma); 1285 pgd = pgd_offset(vma->vm_mm, addr); 1286 do { 1287 next = pgd_addr_end(addr, end); 1288 if (pgd_none_or_clear_bad(pgd)) 1289 continue; 1290 next = zap_pud_range(tlb, vma, pgd, addr, next, details); 1291 } while (pgd++, addr = next, addr != end); 1292 tlb_end_vma(tlb, vma); 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(vma->vm_flags & VM_PFNMAP)) 1314 untrack_pfn(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 mmap_region. When 1323 * hugetlbfs ->mmap method fails, 1324 * mmap_region() 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 i_mmap_lock_write(vma->vm_file->f_mapping); 1331 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1332 i_mmap_unlock_write(vma->vm_file->f_mapping); 1333 } 1334 } else 1335 unmap_page_range(tlb, vma, start, end, details); 1336 } 1337 } 1338 1339 /** 1340 * unmap_vmas - unmap a range of memory covered by a list of vma's 1341 * @tlb: address of the caller's struct mmu_gather 1342 * @vma: the starting vma 1343 * @start_addr: virtual address at which to start unmapping 1344 * @end_addr: virtual address at which to end unmapping 1345 * 1346 * Unmap all pages in the vma list. 1347 * 1348 * Only addresses between `start' and `end' will be unmapped. 1349 * 1350 * The VMA list must be sorted in ascending virtual address order. 1351 * 1352 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1353 * range after unmap_vmas() returns. So the only responsibility here is to 1354 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1355 * drops the lock and schedules. 1356 */ 1357 void unmap_vmas(struct mmu_gather *tlb, 1358 struct vm_area_struct *vma, unsigned long start_addr, 1359 unsigned long end_addr) 1360 { 1361 struct mm_struct *mm = vma->vm_mm; 1362 1363 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr); 1364 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1365 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1366 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr); 1367 } 1368 1369 /** 1370 * zap_page_range - remove user pages in a given range 1371 * @vma: vm_area_struct holding the applicable pages 1372 * @start: starting address of pages to zap 1373 * @size: number of bytes to zap 1374 * @details: details of nonlinear truncation or shared cache invalidation 1375 * 1376 * Caller must protect the VMA list 1377 */ 1378 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1379 unsigned long size, struct zap_details *details) 1380 { 1381 struct mm_struct *mm = vma->vm_mm; 1382 struct mmu_gather tlb; 1383 unsigned long end = start + size; 1384 1385 lru_add_drain(); 1386 tlb_gather_mmu(&tlb, mm, start, end); 1387 update_hiwater_rss(mm); 1388 mmu_notifier_invalidate_range_start(mm, start, end); 1389 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) 1390 unmap_single_vma(&tlb, vma, start, end, details); 1391 mmu_notifier_invalidate_range_end(mm, start, end); 1392 tlb_finish_mmu(&tlb, start, end); 1393 } 1394 1395 /** 1396 * zap_page_range_single - remove user pages in a given range 1397 * @vma: vm_area_struct holding the applicable pages 1398 * @address: starting address of pages to zap 1399 * @size: number of bytes to zap 1400 * @details: details of nonlinear truncation or shared cache invalidation 1401 * 1402 * The range must fit into one VMA. 1403 */ 1404 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1405 unsigned long size, struct zap_details *details) 1406 { 1407 struct mm_struct *mm = vma->vm_mm; 1408 struct mmu_gather tlb; 1409 unsigned long end = address + size; 1410 1411 lru_add_drain(); 1412 tlb_gather_mmu(&tlb, mm, address, end); 1413 update_hiwater_rss(mm); 1414 mmu_notifier_invalidate_range_start(mm, address, end); 1415 unmap_single_vma(&tlb, vma, address, end, details); 1416 mmu_notifier_invalidate_range_end(mm, address, end); 1417 tlb_finish_mmu(&tlb, address, end); 1418 } 1419 1420 /** 1421 * zap_vma_ptes - remove ptes mapping the vma 1422 * @vma: vm_area_struct holding ptes to be zapped 1423 * @address: starting address of pages to zap 1424 * @size: number of bytes to zap 1425 * 1426 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1427 * 1428 * The entire address range must be fully contained within the vma. 1429 * 1430 * Returns 0 if successful. 1431 */ 1432 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1433 unsigned long size) 1434 { 1435 if (address < vma->vm_start || address + size > vma->vm_end || 1436 !(vma->vm_flags & VM_PFNMAP)) 1437 return -1; 1438 zap_page_range_single(vma, address, size, NULL); 1439 return 0; 1440 } 1441 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1442 1443 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1444 spinlock_t **ptl) 1445 { 1446 pgd_t * pgd = pgd_offset(mm, addr); 1447 pud_t * pud = pud_alloc(mm, pgd, addr); 1448 if (pud) { 1449 pmd_t * pmd = pmd_alloc(mm, pud, addr); 1450 if (pmd) { 1451 VM_BUG_ON(pmd_trans_huge(*pmd)); 1452 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1453 } 1454 } 1455 return NULL; 1456 } 1457 1458 /* 1459 * This is the old fallback for page remapping. 1460 * 1461 * For historical reasons, it only allows reserved pages. Only 1462 * old drivers should use this, and they needed to mark their 1463 * pages reserved for the old functions anyway. 1464 */ 1465 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1466 struct page *page, pgprot_t prot) 1467 { 1468 struct mm_struct *mm = vma->vm_mm; 1469 int retval; 1470 pte_t *pte; 1471 spinlock_t *ptl; 1472 1473 retval = -EINVAL; 1474 if (PageAnon(page)) 1475 goto out; 1476 retval = -ENOMEM; 1477 flush_dcache_page(page); 1478 pte = get_locked_pte(mm, addr, &ptl); 1479 if (!pte) 1480 goto out; 1481 retval = -EBUSY; 1482 if (!pte_none(*pte)) 1483 goto out_unlock; 1484 1485 /* Ok, finally just insert the thing.. */ 1486 get_page(page); 1487 inc_mm_counter_fast(mm, MM_FILEPAGES); 1488 page_add_file_rmap(page); 1489 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1490 1491 retval = 0; 1492 pte_unmap_unlock(pte, ptl); 1493 return retval; 1494 out_unlock: 1495 pte_unmap_unlock(pte, ptl); 1496 out: 1497 return retval; 1498 } 1499 1500 /** 1501 * vm_insert_page - insert single page into user vma 1502 * @vma: user vma to map to 1503 * @addr: target user address of this page 1504 * @page: source kernel page 1505 * 1506 * This allows drivers to insert individual pages they've allocated 1507 * into a user vma. 1508 * 1509 * The page has to be a nice clean _individual_ kernel allocation. 1510 * If you allocate a compound page, you need to have marked it as 1511 * such (__GFP_COMP), or manually just split the page up yourself 1512 * (see split_page()). 1513 * 1514 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1515 * took an arbitrary page protection parameter. This doesn't allow 1516 * that. Your vma protection will have to be set up correctly, which 1517 * means that if you want a shared writable mapping, you'd better 1518 * ask for a shared writable mapping! 1519 * 1520 * The page does not need to be reserved. 1521 * 1522 * Usually this function is called from f_op->mmap() handler 1523 * under mm->mmap_sem write-lock, so it can change vma->vm_flags. 1524 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1525 * function from other places, for example from page-fault handler. 1526 */ 1527 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1528 struct page *page) 1529 { 1530 if (addr < vma->vm_start || addr >= vma->vm_end) 1531 return -EFAULT; 1532 if (!page_count(page)) 1533 return -EINVAL; 1534 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1535 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem)); 1536 BUG_ON(vma->vm_flags & VM_PFNMAP); 1537 vma->vm_flags |= VM_MIXEDMAP; 1538 } 1539 return insert_page(vma, addr, page, vma->vm_page_prot); 1540 } 1541 EXPORT_SYMBOL(vm_insert_page); 1542 1543 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1544 unsigned long pfn, pgprot_t prot) 1545 { 1546 struct mm_struct *mm = vma->vm_mm; 1547 int retval; 1548 pte_t *pte, entry; 1549 spinlock_t *ptl; 1550 1551 retval = -ENOMEM; 1552 pte = get_locked_pte(mm, addr, &ptl); 1553 if (!pte) 1554 goto out; 1555 retval = -EBUSY; 1556 if (!pte_none(*pte)) 1557 goto out_unlock; 1558 1559 /* Ok, finally just insert the thing.. */ 1560 entry = pte_mkspecial(pfn_pte(pfn, prot)); 1561 set_pte_at(mm, addr, pte, entry); 1562 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 1563 1564 retval = 0; 1565 out_unlock: 1566 pte_unmap_unlock(pte, ptl); 1567 out: 1568 return retval; 1569 } 1570 1571 /** 1572 * vm_insert_pfn - insert single pfn into user vma 1573 * @vma: user vma to map to 1574 * @addr: target user address of this page 1575 * @pfn: source kernel pfn 1576 * 1577 * Similar to vm_insert_page, this allows drivers to insert individual pages 1578 * they've allocated into a user vma. Same comments apply. 1579 * 1580 * This function should only be called from a vm_ops->fault handler, and 1581 * in that case the handler should return NULL. 1582 * 1583 * vma cannot be a COW mapping. 1584 * 1585 * As this is called only for pages that do not currently exist, we 1586 * do not need to flush old virtual caches or the TLB. 1587 */ 1588 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1589 unsigned long pfn) 1590 { 1591 int ret; 1592 pgprot_t pgprot = vma->vm_page_prot; 1593 /* 1594 * Technically, architectures with pte_special can avoid all these 1595 * restrictions (same for remap_pfn_range). However we would like 1596 * consistency in testing and feature parity among all, so we should 1597 * try to keep these invariants in place for everybody. 1598 */ 1599 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 1600 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1601 (VM_PFNMAP|VM_MIXEDMAP)); 1602 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1603 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 1604 1605 if (addr < vma->vm_start || addr >= vma->vm_end) 1606 return -EFAULT; 1607 if (track_pfn_insert(vma, &pgprot, pfn)) 1608 return -EINVAL; 1609 1610 ret = insert_pfn(vma, addr, pfn, pgprot); 1611 1612 return ret; 1613 } 1614 EXPORT_SYMBOL(vm_insert_pfn); 1615 1616 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 1617 unsigned long pfn) 1618 { 1619 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP)); 1620 1621 if (addr < vma->vm_start || addr >= vma->vm_end) 1622 return -EFAULT; 1623 1624 /* 1625 * If we don't have pte special, then we have to use the pfn_valid() 1626 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 1627 * refcount the page if pfn_valid is true (hence insert_page rather 1628 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 1629 * without pte special, it would there be refcounted as a normal page. 1630 */ 1631 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) { 1632 struct page *page; 1633 1634 page = pfn_to_page(pfn); 1635 return insert_page(vma, addr, page, vma->vm_page_prot); 1636 } 1637 return insert_pfn(vma, addr, pfn, vma->vm_page_prot); 1638 } 1639 EXPORT_SYMBOL(vm_insert_mixed); 1640 1641 /* 1642 * maps a range of physical memory into the requested pages. the old 1643 * mappings are removed. any references to nonexistent pages results 1644 * in null mappings (currently treated as "copy-on-access") 1645 */ 1646 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1647 unsigned long addr, unsigned long end, 1648 unsigned long pfn, pgprot_t prot) 1649 { 1650 pte_t *pte; 1651 spinlock_t *ptl; 1652 1653 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1654 if (!pte) 1655 return -ENOMEM; 1656 arch_enter_lazy_mmu_mode(); 1657 do { 1658 BUG_ON(!pte_none(*pte)); 1659 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 1660 pfn++; 1661 } while (pte++, addr += PAGE_SIZE, addr != end); 1662 arch_leave_lazy_mmu_mode(); 1663 pte_unmap_unlock(pte - 1, ptl); 1664 return 0; 1665 } 1666 1667 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 1668 unsigned long addr, unsigned long end, 1669 unsigned long pfn, pgprot_t prot) 1670 { 1671 pmd_t *pmd; 1672 unsigned long next; 1673 1674 pfn -= addr >> PAGE_SHIFT; 1675 pmd = pmd_alloc(mm, pud, addr); 1676 if (!pmd) 1677 return -ENOMEM; 1678 VM_BUG_ON(pmd_trans_huge(*pmd)); 1679 do { 1680 next = pmd_addr_end(addr, end); 1681 if (remap_pte_range(mm, pmd, addr, next, 1682 pfn + (addr >> PAGE_SHIFT), prot)) 1683 return -ENOMEM; 1684 } while (pmd++, addr = next, addr != end); 1685 return 0; 1686 } 1687 1688 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, 1689 unsigned long addr, unsigned long end, 1690 unsigned long pfn, pgprot_t prot) 1691 { 1692 pud_t *pud; 1693 unsigned long next; 1694 1695 pfn -= addr >> PAGE_SHIFT; 1696 pud = pud_alloc(mm, pgd, addr); 1697 if (!pud) 1698 return -ENOMEM; 1699 do { 1700 next = pud_addr_end(addr, end); 1701 if (remap_pmd_range(mm, pud, addr, next, 1702 pfn + (addr >> PAGE_SHIFT), prot)) 1703 return -ENOMEM; 1704 } while (pud++, addr = next, addr != end); 1705 return 0; 1706 } 1707 1708 /** 1709 * remap_pfn_range - remap kernel memory to userspace 1710 * @vma: user vma to map to 1711 * @addr: target user address to start at 1712 * @pfn: physical address of kernel memory 1713 * @size: size of map area 1714 * @prot: page protection flags for this mapping 1715 * 1716 * Note: this is only safe if the mm semaphore is held when called. 1717 */ 1718 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 1719 unsigned long pfn, unsigned long size, pgprot_t prot) 1720 { 1721 pgd_t *pgd; 1722 unsigned long next; 1723 unsigned long end = addr + PAGE_ALIGN(size); 1724 struct mm_struct *mm = vma->vm_mm; 1725 int err; 1726 1727 /* 1728 * Physically remapped pages are special. Tell the 1729 * rest of the world about it: 1730 * VM_IO tells people not to look at these pages 1731 * (accesses can have side effects). 1732 * VM_PFNMAP tells the core MM that the base pages are just 1733 * raw PFN mappings, and do not have a "struct page" associated 1734 * with them. 1735 * VM_DONTEXPAND 1736 * Disable vma merging and expanding with mremap(). 1737 * VM_DONTDUMP 1738 * Omit vma from core dump, even when VM_IO turned off. 1739 * 1740 * There's a horrible special case to handle copy-on-write 1741 * behaviour that some programs depend on. We mark the "original" 1742 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 1743 * See vm_normal_page() for details. 1744 */ 1745 if (is_cow_mapping(vma->vm_flags)) { 1746 if (addr != vma->vm_start || end != vma->vm_end) 1747 return -EINVAL; 1748 vma->vm_pgoff = pfn; 1749 } 1750 1751 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 1752 if (err) 1753 return -EINVAL; 1754 1755 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 1756 1757 BUG_ON(addr >= end); 1758 pfn -= addr >> PAGE_SHIFT; 1759 pgd = pgd_offset(mm, addr); 1760 flush_cache_range(vma, addr, end); 1761 do { 1762 next = pgd_addr_end(addr, end); 1763 err = remap_pud_range(mm, pgd, addr, next, 1764 pfn + (addr >> PAGE_SHIFT), prot); 1765 if (err) 1766 break; 1767 } while (pgd++, addr = next, addr != end); 1768 1769 if (err) 1770 untrack_pfn(vma, pfn, PAGE_ALIGN(size)); 1771 1772 return err; 1773 } 1774 EXPORT_SYMBOL(remap_pfn_range); 1775 1776 /** 1777 * vm_iomap_memory - remap memory to userspace 1778 * @vma: user vma to map to 1779 * @start: start of area 1780 * @len: size of area 1781 * 1782 * This is a simplified io_remap_pfn_range() for common driver use. The 1783 * driver just needs to give us the physical memory range to be mapped, 1784 * we'll figure out the rest from the vma information. 1785 * 1786 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 1787 * whatever write-combining details or similar. 1788 */ 1789 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 1790 { 1791 unsigned long vm_len, pfn, pages; 1792 1793 /* Check that the physical memory area passed in looks valid */ 1794 if (start + len < start) 1795 return -EINVAL; 1796 /* 1797 * You *really* shouldn't map things that aren't page-aligned, 1798 * but we've historically allowed it because IO memory might 1799 * just have smaller alignment. 1800 */ 1801 len += start & ~PAGE_MASK; 1802 pfn = start >> PAGE_SHIFT; 1803 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 1804 if (pfn + pages < pfn) 1805 return -EINVAL; 1806 1807 /* We start the mapping 'vm_pgoff' pages into the area */ 1808 if (vma->vm_pgoff > pages) 1809 return -EINVAL; 1810 pfn += vma->vm_pgoff; 1811 pages -= vma->vm_pgoff; 1812 1813 /* Can we fit all of the mapping? */ 1814 vm_len = vma->vm_end - vma->vm_start; 1815 if (vm_len >> PAGE_SHIFT > pages) 1816 return -EINVAL; 1817 1818 /* Ok, let it rip */ 1819 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 1820 } 1821 EXPORT_SYMBOL(vm_iomap_memory); 1822 1823 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 1824 unsigned long addr, unsigned long end, 1825 pte_fn_t fn, void *data) 1826 { 1827 pte_t *pte; 1828 int err; 1829 pgtable_t token; 1830 spinlock_t *uninitialized_var(ptl); 1831 1832 pte = (mm == &init_mm) ? 1833 pte_alloc_kernel(pmd, addr) : 1834 pte_alloc_map_lock(mm, pmd, addr, &ptl); 1835 if (!pte) 1836 return -ENOMEM; 1837 1838 BUG_ON(pmd_huge(*pmd)); 1839 1840 arch_enter_lazy_mmu_mode(); 1841 1842 token = pmd_pgtable(*pmd); 1843 1844 do { 1845 err = fn(pte++, token, addr, data); 1846 if (err) 1847 break; 1848 } while (addr += PAGE_SIZE, addr != end); 1849 1850 arch_leave_lazy_mmu_mode(); 1851 1852 if (mm != &init_mm) 1853 pte_unmap_unlock(pte-1, ptl); 1854 return err; 1855 } 1856 1857 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 1858 unsigned long addr, unsigned long end, 1859 pte_fn_t fn, void *data) 1860 { 1861 pmd_t *pmd; 1862 unsigned long next; 1863 int err; 1864 1865 BUG_ON(pud_huge(*pud)); 1866 1867 pmd = pmd_alloc(mm, pud, addr); 1868 if (!pmd) 1869 return -ENOMEM; 1870 do { 1871 next = pmd_addr_end(addr, end); 1872 err = apply_to_pte_range(mm, pmd, addr, next, fn, data); 1873 if (err) 1874 break; 1875 } while (pmd++, addr = next, addr != end); 1876 return err; 1877 } 1878 1879 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, 1880 unsigned long addr, unsigned long end, 1881 pte_fn_t fn, void *data) 1882 { 1883 pud_t *pud; 1884 unsigned long next; 1885 int err; 1886 1887 pud = pud_alloc(mm, pgd, addr); 1888 if (!pud) 1889 return -ENOMEM; 1890 do { 1891 next = pud_addr_end(addr, end); 1892 err = apply_to_pmd_range(mm, pud, addr, next, fn, data); 1893 if (err) 1894 break; 1895 } while (pud++, addr = next, addr != end); 1896 return err; 1897 } 1898 1899 /* 1900 * Scan a region of virtual memory, filling in page tables as necessary 1901 * and calling a provided function on each leaf page table. 1902 */ 1903 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 1904 unsigned long size, pte_fn_t fn, void *data) 1905 { 1906 pgd_t *pgd; 1907 unsigned long next; 1908 unsigned long end = addr + size; 1909 int err; 1910 1911 BUG_ON(addr >= end); 1912 pgd = pgd_offset(mm, addr); 1913 do { 1914 next = pgd_addr_end(addr, end); 1915 err = apply_to_pud_range(mm, pgd, addr, next, fn, data); 1916 if (err) 1917 break; 1918 } while (pgd++, addr = next, addr != end); 1919 1920 return err; 1921 } 1922 EXPORT_SYMBOL_GPL(apply_to_page_range); 1923 1924 /* 1925 * handle_pte_fault chooses page fault handler according to an entry 1926 * which was read non-atomically. Before making any commitment, on 1927 * those architectures or configurations (e.g. i386 with PAE) which 1928 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault 1929 * must check under lock before unmapping the pte and proceeding 1930 * (but do_wp_page is only called after already making such a check; 1931 * and do_anonymous_page can safely check later on). 1932 */ 1933 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 1934 pte_t *page_table, pte_t orig_pte) 1935 { 1936 int same = 1; 1937 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 1938 if (sizeof(pte_t) > sizeof(unsigned long)) { 1939 spinlock_t *ptl = pte_lockptr(mm, pmd); 1940 spin_lock(ptl); 1941 same = pte_same(*page_table, orig_pte); 1942 spin_unlock(ptl); 1943 } 1944 #endif 1945 pte_unmap(page_table); 1946 return same; 1947 } 1948 1949 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) 1950 { 1951 debug_dma_assert_idle(src); 1952 1953 /* 1954 * If the source page was a PFN mapping, we don't have 1955 * a "struct page" for it. We do a best-effort copy by 1956 * just copying from the original user address. If that 1957 * fails, we just zero-fill it. Live with it. 1958 */ 1959 if (unlikely(!src)) { 1960 void *kaddr = kmap_atomic(dst); 1961 void __user *uaddr = (void __user *)(va & PAGE_MASK); 1962 1963 /* 1964 * This really shouldn't fail, because the page is there 1965 * in the page tables. But it might just be unreadable, 1966 * in which case we just give up and fill the result with 1967 * zeroes. 1968 */ 1969 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) 1970 clear_page(kaddr); 1971 kunmap_atomic(kaddr); 1972 flush_dcache_page(dst); 1973 } else 1974 copy_user_highpage(dst, src, va, vma); 1975 } 1976 1977 /* 1978 * Notify the address space that the page is about to become writable so that 1979 * it can prohibit this or wait for the page to get into an appropriate state. 1980 * 1981 * We do this without the lock held, so that it can sleep if it needs to. 1982 */ 1983 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page, 1984 unsigned long address) 1985 { 1986 struct vm_fault vmf; 1987 int ret; 1988 1989 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 1990 vmf.pgoff = page->index; 1991 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 1992 vmf.page = page; 1993 1994 ret = vma->vm_ops->page_mkwrite(vma, &vmf); 1995 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 1996 return ret; 1997 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 1998 lock_page(page); 1999 if (!page->mapping) { 2000 unlock_page(page); 2001 return 0; /* retry */ 2002 } 2003 ret |= VM_FAULT_LOCKED; 2004 } else 2005 VM_BUG_ON_PAGE(!PageLocked(page), page); 2006 return ret; 2007 } 2008 2009 /* 2010 * This routine handles present pages, when users try to write 2011 * to a shared page. It is done by copying the page to a new address 2012 * and decrementing the shared-page counter for the old page. 2013 * 2014 * Note that this routine assumes that the protection checks have been 2015 * done by the caller (the low-level page fault routine in most cases). 2016 * Thus we can safely just mark it writable once we've done any necessary 2017 * COW. 2018 * 2019 * We also mark the page dirty at this point even though the page will 2020 * change only once the write actually happens. This avoids a few races, 2021 * and potentially makes it more efficient. 2022 * 2023 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2024 * but allow concurrent faults), with pte both mapped and locked. 2025 * We return with mmap_sem still held, but pte unmapped and unlocked. 2026 */ 2027 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 2028 unsigned long address, pte_t *page_table, pmd_t *pmd, 2029 spinlock_t *ptl, pte_t orig_pte) 2030 __releases(ptl) 2031 { 2032 struct page *old_page, *new_page = NULL; 2033 pte_t entry; 2034 int ret = 0; 2035 int page_mkwrite = 0; 2036 struct page *dirty_page = NULL; 2037 unsigned long mmun_start = 0; /* For mmu_notifiers */ 2038 unsigned long mmun_end = 0; /* For mmu_notifiers */ 2039 struct mem_cgroup *memcg; 2040 2041 old_page = vm_normal_page(vma, address, orig_pte); 2042 if (!old_page) { 2043 /* 2044 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 2045 * VM_PFNMAP VMA. 2046 * 2047 * We should not cow pages in a shared writeable mapping. 2048 * Just mark the pages writable as we can't do any dirty 2049 * accounting on raw pfn maps. 2050 */ 2051 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2052 (VM_WRITE|VM_SHARED)) 2053 goto reuse; 2054 goto gotten; 2055 } 2056 2057 /* 2058 * Take out anonymous pages first, anonymous shared vmas are 2059 * not dirty accountable. 2060 */ 2061 if (PageAnon(old_page) && !PageKsm(old_page)) { 2062 if (!trylock_page(old_page)) { 2063 page_cache_get(old_page); 2064 pte_unmap_unlock(page_table, ptl); 2065 lock_page(old_page); 2066 page_table = pte_offset_map_lock(mm, pmd, address, 2067 &ptl); 2068 if (!pte_same(*page_table, orig_pte)) { 2069 unlock_page(old_page); 2070 goto unlock; 2071 } 2072 page_cache_release(old_page); 2073 } 2074 if (reuse_swap_page(old_page)) { 2075 /* 2076 * The page is all ours. Move it to our anon_vma so 2077 * the rmap code will not search our parent or siblings. 2078 * Protected against the rmap code by the page lock. 2079 */ 2080 page_move_anon_rmap(old_page, vma, address); 2081 unlock_page(old_page); 2082 goto reuse; 2083 } 2084 unlock_page(old_page); 2085 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2086 (VM_WRITE|VM_SHARED))) { 2087 /* 2088 * Only catch write-faults on shared writable pages, 2089 * read-only shared pages can get COWed by 2090 * get_user_pages(.write=1, .force=1). 2091 */ 2092 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 2093 int tmp; 2094 page_cache_get(old_page); 2095 pte_unmap_unlock(page_table, ptl); 2096 tmp = do_page_mkwrite(vma, old_page, address); 2097 if (unlikely(!tmp || (tmp & 2098 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 2099 page_cache_release(old_page); 2100 return tmp; 2101 } 2102 /* 2103 * Since we dropped the lock we need to revalidate 2104 * the PTE as someone else may have changed it. If 2105 * they did, we just return, as we can count on the 2106 * MMU to tell us if they didn't also make it writable. 2107 */ 2108 page_table = pte_offset_map_lock(mm, pmd, address, 2109 &ptl); 2110 if (!pte_same(*page_table, orig_pte)) { 2111 unlock_page(old_page); 2112 goto unlock; 2113 } 2114 2115 page_mkwrite = 1; 2116 } 2117 dirty_page = old_page; 2118 get_page(dirty_page); 2119 2120 reuse: 2121 /* 2122 * Clear the pages cpupid information as the existing 2123 * information potentially belongs to a now completely 2124 * unrelated process. 2125 */ 2126 if (old_page) 2127 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1); 2128 2129 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2130 entry = pte_mkyoung(orig_pte); 2131 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2132 if (ptep_set_access_flags(vma, address, page_table, entry,1)) 2133 update_mmu_cache(vma, address, page_table); 2134 pte_unmap_unlock(page_table, ptl); 2135 ret |= VM_FAULT_WRITE; 2136 2137 if (!dirty_page) 2138 return ret; 2139 2140 /* 2141 * Yes, Virginia, this is actually required to prevent a race 2142 * with clear_page_dirty_for_io() from clearing the page dirty 2143 * bit after it clear all dirty ptes, but before a racing 2144 * do_wp_page installs a dirty pte. 2145 * 2146 * do_shared_fault is protected similarly. 2147 */ 2148 if (!page_mkwrite) { 2149 wait_on_page_locked(dirty_page); 2150 set_page_dirty_balance(dirty_page); 2151 /* file_update_time outside page_lock */ 2152 if (vma->vm_file) 2153 file_update_time(vma->vm_file); 2154 } 2155 put_page(dirty_page); 2156 if (page_mkwrite) { 2157 struct address_space *mapping = dirty_page->mapping; 2158 2159 set_page_dirty(dirty_page); 2160 unlock_page(dirty_page); 2161 page_cache_release(dirty_page); 2162 if (mapping) { 2163 /* 2164 * Some device drivers do not set page.mapping 2165 * but still dirty their pages 2166 */ 2167 balance_dirty_pages_ratelimited(mapping); 2168 } 2169 } 2170 2171 return ret; 2172 } 2173 2174 /* 2175 * Ok, we need to copy. Oh, well.. 2176 */ 2177 page_cache_get(old_page); 2178 gotten: 2179 pte_unmap_unlock(page_table, ptl); 2180 2181 if (unlikely(anon_vma_prepare(vma))) 2182 goto oom; 2183 2184 if (is_zero_pfn(pte_pfn(orig_pte))) { 2185 new_page = alloc_zeroed_user_highpage_movable(vma, address); 2186 if (!new_page) 2187 goto oom; 2188 } else { 2189 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 2190 if (!new_page) 2191 goto oom; 2192 cow_user_page(new_page, old_page, address, vma); 2193 } 2194 __SetPageUptodate(new_page); 2195 2196 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) 2197 goto oom_free_new; 2198 2199 mmun_start = address & PAGE_MASK; 2200 mmun_end = mmun_start + PAGE_SIZE; 2201 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2202 2203 /* 2204 * Re-check the pte - we dropped the lock 2205 */ 2206 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2207 if (likely(pte_same(*page_table, orig_pte))) { 2208 if (old_page) { 2209 if (!PageAnon(old_page)) { 2210 dec_mm_counter_fast(mm, MM_FILEPAGES); 2211 inc_mm_counter_fast(mm, MM_ANONPAGES); 2212 } 2213 } else 2214 inc_mm_counter_fast(mm, MM_ANONPAGES); 2215 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2216 entry = mk_pte(new_page, vma->vm_page_prot); 2217 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2218 /* 2219 * Clear the pte entry and flush it first, before updating the 2220 * pte with the new entry. This will avoid a race condition 2221 * seen in the presence of one thread doing SMC and another 2222 * thread doing COW. 2223 */ 2224 ptep_clear_flush_notify(vma, address, page_table); 2225 page_add_new_anon_rmap(new_page, vma, address); 2226 mem_cgroup_commit_charge(new_page, memcg, false); 2227 lru_cache_add_active_or_unevictable(new_page, vma); 2228 /* 2229 * We call the notify macro here because, when using secondary 2230 * mmu page tables (such as kvm shadow page tables), we want the 2231 * new page to be mapped directly into the secondary page table. 2232 */ 2233 set_pte_at_notify(mm, address, page_table, entry); 2234 update_mmu_cache(vma, address, page_table); 2235 if (old_page) { 2236 /* 2237 * Only after switching the pte to the new page may 2238 * we remove the mapcount here. Otherwise another 2239 * process may come and find the rmap count decremented 2240 * before the pte is switched to the new page, and 2241 * "reuse" the old page writing into it while our pte 2242 * here still points into it and can be read by other 2243 * threads. 2244 * 2245 * The critical issue is to order this 2246 * page_remove_rmap with the ptp_clear_flush above. 2247 * Those stores are ordered by (if nothing else,) 2248 * the barrier present in the atomic_add_negative 2249 * in page_remove_rmap. 2250 * 2251 * Then the TLB flush in ptep_clear_flush ensures that 2252 * no process can access the old page before the 2253 * decremented mapcount is visible. And the old page 2254 * cannot be reused until after the decremented 2255 * mapcount is visible. So transitively, TLBs to 2256 * old page will be flushed before it can be reused. 2257 */ 2258 page_remove_rmap(old_page); 2259 } 2260 2261 /* Free the old page.. */ 2262 new_page = old_page; 2263 ret |= VM_FAULT_WRITE; 2264 } else 2265 mem_cgroup_cancel_charge(new_page, memcg); 2266 2267 if (new_page) 2268 page_cache_release(new_page); 2269 unlock: 2270 pte_unmap_unlock(page_table, ptl); 2271 if (mmun_end > mmun_start) 2272 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2273 if (old_page) { 2274 /* 2275 * Don't let another task, with possibly unlocked vma, 2276 * keep the mlocked page. 2277 */ 2278 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) { 2279 lock_page(old_page); /* LRU manipulation */ 2280 munlock_vma_page(old_page); 2281 unlock_page(old_page); 2282 } 2283 page_cache_release(old_page); 2284 } 2285 return ret; 2286 oom_free_new: 2287 page_cache_release(new_page); 2288 oom: 2289 if (old_page) 2290 page_cache_release(old_page); 2291 return VM_FAULT_OOM; 2292 } 2293 2294 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 2295 unsigned long start_addr, unsigned long end_addr, 2296 struct zap_details *details) 2297 { 2298 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 2299 } 2300 2301 static inline void unmap_mapping_range_tree(struct rb_root *root, 2302 struct zap_details *details) 2303 { 2304 struct vm_area_struct *vma; 2305 pgoff_t vba, vea, zba, zea; 2306 2307 vma_interval_tree_foreach(vma, root, 2308 details->first_index, details->last_index) { 2309 2310 vba = vma->vm_pgoff; 2311 vea = vba + vma_pages(vma) - 1; 2312 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ 2313 zba = details->first_index; 2314 if (zba < vba) 2315 zba = vba; 2316 zea = details->last_index; 2317 if (zea > vea) 2318 zea = vea; 2319 2320 unmap_mapping_range_vma(vma, 2321 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 2322 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 2323 details); 2324 } 2325 } 2326 2327 static inline void unmap_mapping_range_list(struct list_head *head, 2328 struct zap_details *details) 2329 { 2330 struct vm_area_struct *vma; 2331 2332 /* 2333 * In nonlinear VMAs there is no correspondence between virtual address 2334 * offset and file offset. So we must perform an exhaustive search 2335 * across *all* the pages in each nonlinear VMA, not just the pages 2336 * whose virtual address lies outside the file truncation point. 2337 */ 2338 list_for_each_entry(vma, head, shared.nonlinear) { 2339 details->nonlinear_vma = vma; 2340 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details); 2341 } 2342 } 2343 2344 /** 2345 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file. 2346 * @mapping: the address space containing mmaps to be unmapped. 2347 * @holebegin: byte in first page to unmap, relative to the start of 2348 * the underlying file. This will be rounded down to a PAGE_SIZE 2349 * boundary. Note that this is different from truncate_pagecache(), which 2350 * must keep the partial page. In contrast, we must get rid of 2351 * partial pages. 2352 * @holelen: size of prospective hole in bytes. This will be rounded 2353 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 2354 * end of the file. 2355 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 2356 * but 0 when invalidating pagecache, don't throw away private data. 2357 */ 2358 void unmap_mapping_range(struct address_space *mapping, 2359 loff_t const holebegin, loff_t const holelen, int even_cows) 2360 { 2361 struct zap_details details; 2362 pgoff_t hba = holebegin >> PAGE_SHIFT; 2363 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2364 2365 /* Check for overflow. */ 2366 if (sizeof(holelen) > sizeof(hlen)) { 2367 long long holeend = 2368 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2369 if (holeend & ~(long long)ULONG_MAX) 2370 hlen = ULONG_MAX - hba + 1; 2371 } 2372 2373 details.check_mapping = even_cows? NULL: mapping; 2374 details.nonlinear_vma = NULL; 2375 details.first_index = hba; 2376 details.last_index = hba + hlen - 1; 2377 if (details.last_index < details.first_index) 2378 details.last_index = ULONG_MAX; 2379 2380 2381 i_mmap_lock_read(mapping); 2382 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap))) 2383 unmap_mapping_range_tree(&mapping->i_mmap, &details); 2384 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) 2385 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details); 2386 i_mmap_unlock_read(mapping); 2387 } 2388 EXPORT_SYMBOL(unmap_mapping_range); 2389 2390 /* 2391 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2392 * but allow concurrent faults), and pte mapped but not yet locked. 2393 * We return with pte unmapped and unlocked. 2394 * 2395 * We return with the mmap_sem locked or unlocked in the same cases 2396 * as does filemap_fault(). 2397 */ 2398 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, 2399 unsigned long address, pte_t *page_table, pmd_t *pmd, 2400 unsigned int flags, pte_t orig_pte) 2401 { 2402 spinlock_t *ptl; 2403 struct page *page, *swapcache; 2404 struct mem_cgroup *memcg; 2405 swp_entry_t entry; 2406 pte_t pte; 2407 int locked; 2408 int exclusive = 0; 2409 int ret = 0; 2410 2411 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 2412 goto out; 2413 2414 entry = pte_to_swp_entry(orig_pte); 2415 if (unlikely(non_swap_entry(entry))) { 2416 if (is_migration_entry(entry)) { 2417 migration_entry_wait(mm, pmd, address); 2418 } else if (is_hwpoison_entry(entry)) { 2419 ret = VM_FAULT_HWPOISON; 2420 } else { 2421 print_bad_pte(vma, address, orig_pte, NULL); 2422 ret = VM_FAULT_SIGBUS; 2423 } 2424 goto out; 2425 } 2426 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 2427 page = lookup_swap_cache(entry); 2428 if (!page) { 2429 page = swapin_readahead(entry, 2430 GFP_HIGHUSER_MOVABLE, vma, address); 2431 if (!page) { 2432 /* 2433 * Back out if somebody else faulted in this pte 2434 * while we released the pte lock. 2435 */ 2436 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2437 if (likely(pte_same(*page_table, orig_pte))) 2438 ret = VM_FAULT_OOM; 2439 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2440 goto unlock; 2441 } 2442 2443 /* Had to read the page from swap area: Major fault */ 2444 ret = VM_FAULT_MAJOR; 2445 count_vm_event(PGMAJFAULT); 2446 mem_cgroup_count_vm_event(mm, PGMAJFAULT); 2447 } else if (PageHWPoison(page)) { 2448 /* 2449 * hwpoisoned dirty swapcache pages are kept for killing 2450 * owner processes (which may be unknown at hwpoison time) 2451 */ 2452 ret = VM_FAULT_HWPOISON; 2453 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2454 swapcache = page; 2455 goto out_release; 2456 } 2457 2458 swapcache = page; 2459 locked = lock_page_or_retry(page, mm, flags); 2460 2461 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2462 if (!locked) { 2463 ret |= VM_FAULT_RETRY; 2464 goto out_release; 2465 } 2466 2467 /* 2468 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 2469 * release the swapcache from under us. The page pin, and pte_same 2470 * test below, are not enough to exclude that. Even if it is still 2471 * swapcache, we need to check that the page's swap has not changed. 2472 */ 2473 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) 2474 goto out_page; 2475 2476 page = ksm_might_need_to_copy(page, vma, address); 2477 if (unlikely(!page)) { 2478 ret = VM_FAULT_OOM; 2479 page = swapcache; 2480 goto out_page; 2481 } 2482 2483 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) { 2484 ret = VM_FAULT_OOM; 2485 goto out_page; 2486 } 2487 2488 /* 2489 * Back out if somebody else already faulted in this pte. 2490 */ 2491 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2492 if (unlikely(!pte_same(*page_table, orig_pte))) 2493 goto out_nomap; 2494 2495 if (unlikely(!PageUptodate(page))) { 2496 ret = VM_FAULT_SIGBUS; 2497 goto out_nomap; 2498 } 2499 2500 /* 2501 * The page isn't present yet, go ahead with the fault. 2502 * 2503 * Be careful about the sequence of operations here. 2504 * To get its accounting right, reuse_swap_page() must be called 2505 * while the page is counted on swap but not yet in mapcount i.e. 2506 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 2507 * must be called after the swap_free(), or it will never succeed. 2508 */ 2509 2510 inc_mm_counter_fast(mm, MM_ANONPAGES); 2511 dec_mm_counter_fast(mm, MM_SWAPENTS); 2512 pte = mk_pte(page, vma->vm_page_prot); 2513 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { 2514 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 2515 flags &= ~FAULT_FLAG_WRITE; 2516 ret |= VM_FAULT_WRITE; 2517 exclusive = 1; 2518 } 2519 flush_icache_page(vma, page); 2520 if (pte_swp_soft_dirty(orig_pte)) 2521 pte = pte_mksoft_dirty(pte); 2522 set_pte_at(mm, address, page_table, pte); 2523 if (page == swapcache) { 2524 do_page_add_anon_rmap(page, vma, address, exclusive); 2525 mem_cgroup_commit_charge(page, memcg, true); 2526 } else { /* ksm created a completely new copy */ 2527 page_add_new_anon_rmap(page, vma, address); 2528 mem_cgroup_commit_charge(page, memcg, false); 2529 lru_cache_add_active_or_unevictable(page, vma); 2530 } 2531 2532 swap_free(entry); 2533 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 2534 try_to_free_swap(page); 2535 unlock_page(page); 2536 if (page != swapcache) { 2537 /* 2538 * Hold the lock to avoid the swap entry to be reused 2539 * until we take the PT lock for the pte_same() check 2540 * (to avoid false positives from pte_same). For 2541 * further safety release the lock after the swap_free 2542 * so that the swap count won't change under a 2543 * parallel locked swapcache. 2544 */ 2545 unlock_page(swapcache); 2546 page_cache_release(swapcache); 2547 } 2548 2549 if (flags & FAULT_FLAG_WRITE) { 2550 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); 2551 if (ret & VM_FAULT_ERROR) 2552 ret &= VM_FAULT_ERROR; 2553 goto out; 2554 } 2555 2556 /* No need to invalidate - it was non-present before */ 2557 update_mmu_cache(vma, address, page_table); 2558 unlock: 2559 pte_unmap_unlock(page_table, ptl); 2560 out: 2561 return ret; 2562 out_nomap: 2563 mem_cgroup_cancel_charge(page, memcg); 2564 pte_unmap_unlock(page_table, ptl); 2565 out_page: 2566 unlock_page(page); 2567 out_release: 2568 page_cache_release(page); 2569 if (page != swapcache) { 2570 unlock_page(swapcache); 2571 page_cache_release(swapcache); 2572 } 2573 return ret; 2574 } 2575 2576 /* 2577 * This is like a special single-page "expand_{down|up}wards()", 2578 * except we must first make sure that 'address{-|+}PAGE_SIZE' 2579 * doesn't hit another vma. 2580 */ 2581 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) 2582 { 2583 address &= PAGE_MASK; 2584 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { 2585 struct vm_area_struct *prev = vma->vm_prev; 2586 2587 /* 2588 * Is there a mapping abutting this one below? 2589 * 2590 * That's only ok if it's the same stack mapping 2591 * that has gotten split.. 2592 */ 2593 if (prev && prev->vm_end == address) 2594 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; 2595 2596 expand_downwards(vma, address - PAGE_SIZE); 2597 } 2598 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { 2599 struct vm_area_struct *next = vma->vm_next; 2600 2601 /* As VM_GROWSDOWN but s/below/above/ */ 2602 if (next && next->vm_start == address + PAGE_SIZE) 2603 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; 2604 2605 expand_upwards(vma, address + PAGE_SIZE); 2606 } 2607 return 0; 2608 } 2609 2610 /* 2611 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2612 * but allow concurrent faults), and pte mapped but not yet locked. 2613 * We return with mmap_sem still held, but pte unmapped and unlocked. 2614 */ 2615 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 2616 unsigned long address, pte_t *page_table, pmd_t *pmd, 2617 unsigned int flags) 2618 { 2619 struct mem_cgroup *memcg; 2620 struct page *page; 2621 spinlock_t *ptl; 2622 pte_t entry; 2623 2624 pte_unmap(page_table); 2625 2626 /* Check if we need to add a guard page to the stack */ 2627 if (check_stack_guard_page(vma, address) < 0) 2628 return VM_FAULT_SIGBUS; 2629 2630 /* Use the zero-page for reads */ 2631 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) { 2632 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), 2633 vma->vm_page_prot)); 2634 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2635 if (!pte_none(*page_table)) 2636 goto unlock; 2637 goto setpte; 2638 } 2639 2640 /* Allocate our own private page. */ 2641 if (unlikely(anon_vma_prepare(vma))) 2642 goto oom; 2643 page = alloc_zeroed_user_highpage_movable(vma, address); 2644 if (!page) 2645 goto oom; 2646 /* 2647 * The memory barrier inside __SetPageUptodate makes sure that 2648 * preceeding stores to the page contents become visible before 2649 * the set_pte_at() write. 2650 */ 2651 __SetPageUptodate(page); 2652 2653 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) 2654 goto oom_free_page; 2655 2656 entry = mk_pte(page, vma->vm_page_prot); 2657 if (vma->vm_flags & VM_WRITE) 2658 entry = pte_mkwrite(pte_mkdirty(entry)); 2659 2660 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2661 if (!pte_none(*page_table)) 2662 goto release; 2663 2664 inc_mm_counter_fast(mm, MM_ANONPAGES); 2665 page_add_new_anon_rmap(page, vma, address); 2666 mem_cgroup_commit_charge(page, memcg, false); 2667 lru_cache_add_active_or_unevictable(page, vma); 2668 setpte: 2669 set_pte_at(mm, address, page_table, entry); 2670 2671 /* No need to invalidate - it was non-present before */ 2672 update_mmu_cache(vma, address, page_table); 2673 unlock: 2674 pte_unmap_unlock(page_table, ptl); 2675 return 0; 2676 release: 2677 mem_cgroup_cancel_charge(page, memcg); 2678 page_cache_release(page); 2679 goto unlock; 2680 oom_free_page: 2681 page_cache_release(page); 2682 oom: 2683 return VM_FAULT_OOM; 2684 } 2685 2686 /* 2687 * The mmap_sem must have been held on entry, and may have been 2688 * released depending on flags and vma->vm_ops->fault() return value. 2689 * See filemap_fault() and __lock_page_retry(). 2690 */ 2691 static int __do_fault(struct vm_area_struct *vma, unsigned long address, 2692 pgoff_t pgoff, unsigned int flags, struct page **page) 2693 { 2694 struct vm_fault vmf; 2695 int ret; 2696 2697 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 2698 vmf.pgoff = pgoff; 2699 vmf.flags = flags; 2700 vmf.page = NULL; 2701 2702 ret = vma->vm_ops->fault(vma, &vmf); 2703 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2704 return ret; 2705 2706 if (unlikely(PageHWPoison(vmf.page))) { 2707 if (ret & VM_FAULT_LOCKED) 2708 unlock_page(vmf.page); 2709 page_cache_release(vmf.page); 2710 return VM_FAULT_HWPOISON; 2711 } 2712 2713 if (unlikely(!(ret & VM_FAULT_LOCKED))) 2714 lock_page(vmf.page); 2715 else 2716 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page); 2717 2718 *page = vmf.page; 2719 return ret; 2720 } 2721 2722 /** 2723 * do_set_pte - setup new PTE entry for given page and add reverse page mapping. 2724 * 2725 * @vma: virtual memory area 2726 * @address: user virtual address 2727 * @page: page to map 2728 * @pte: pointer to target page table entry 2729 * @write: true, if new entry is writable 2730 * @anon: true, if it's anonymous page 2731 * 2732 * Caller must hold page table lock relevant for @pte. 2733 * 2734 * Target users are page handler itself and implementations of 2735 * vm_ops->map_pages. 2736 */ 2737 void do_set_pte(struct vm_area_struct *vma, unsigned long address, 2738 struct page *page, pte_t *pte, bool write, bool anon) 2739 { 2740 pte_t entry; 2741 2742 flush_icache_page(vma, page); 2743 entry = mk_pte(page, vma->vm_page_prot); 2744 if (write) 2745 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2746 else if (pte_file(*pte) && pte_file_soft_dirty(*pte)) 2747 entry = pte_mksoft_dirty(entry); 2748 if (anon) { 2749 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 2750 page_add_new_anon_rmap(page, vma, address); 2751 } else { 2752 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES); 2753 page_add_file_rmap(page); 2754 } 2755 set_pte_at(vma->vm_mm, address, pte, entry); 2756 2757 /* no need to invalidate: a not-present page won't be cached */ 2758 update_mmu_cache(vma, address, pte); 2759 } 2760 2761 static unsigned long fault_around_bytes __read_mostly = 2762 rounddown_pow_of_two(65536); 2763 2764 #ifdef CONFIG_DEBUG_FS 2765 static int fault_around_bytes_get(void *data, u64 *val) 2766 { 2767 *val = fault_around_bytes; 2768 return 0; 2769 } 2770 2771 /* 2772 * fault_around_pages() and fault_around_mask() expects fault_around_bytes 2773 * rounded down to nearest page order. It's what do_fault_around() expects to 2774 * see. 2775 */ 2776 static int fault_around_bytes_set(void *data, u64 val) 2777 { 2778 if (val / PAGE_SIZE > PTRS_PER_PTE) 2779 return -EINVAL; 2780 if (val > PAGE_SIZE) 2781 fault_around_bytes = rounddown_pow_of_two(val); 2782 else 2783 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 2784 return 0; 2785 } 2786 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops, 2787 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 2788 2789 static int __init fault_around_debugfs(void) 2790 { 2791 void *ret; 2792 2793 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL, 2794 &fault_around_bytes_fops); 2795 if (!ret) 2796 pr_warn("Failed to create fault_around_bytes in debugfs"); 2797 return 0; 2798 } 2799 late_initcall(fault_around_debugfs); 2800 #endif 2801 2802 /* 2803 * do_fault_around() tries to map few pages around the fault address. The hope 2804 * is that the pages will be needed soon and this will lower the number of 2805 * faults to handle. 2806 * 2807 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 2808 * not ready to be mapped: not up-to-date, locked, etc. 2809 * 2810 * This function is called with the page table lock taken. In the split ptlock 2811 * case the page table lock only protects only those entries which belong to 2812 * the page table corresponding to the fault address. 2813 * 2814 * This function doesn't cross the VMA boundaries, in order to call map_pages() 2815 * only once. 2816 * 2817 * fault_around_pages() defines how many pages we'll try to map. 2818 * do_fault_around() expects it to return a power of two less than or equal to 2819 * PTRS_PER_PTE. 2820 * 2821 * The virtual address of the area that we map is naturally aligned to the 2822 * fault_around_pages() value (and therefore to page order). This way it's 2823 * easier to guarantee that we don't cross page table boundaries. 2824 */ 2825 static void do_fault_around(struct vm_area_struct *vma, unsigned long address, 2826 pte_t *pte, pgoff_t pgoff, unsigned int flags) 2827 { 2828 unsigned long start_addr, nr_pages, mask; 2829 pgoff_t max_pgoff; 2830 struct vm_fault vmf; 2831 int off; 2832 2833 nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT; 2834 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 2835 2836 start_addr = max(address & mask, vma->vm_start); 2837 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 2838 pte -= off; 2839 pgoff -= off; 2840 2841 /* 2842 * max_pgoff is either end of page table or end of vma 2843 * or fault_around_pages() from pgoff, depending what is nearest. 2844 */ 2845 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 2846 PTRS_PER_PTE - 1; 2847 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1, 2848 pgoff + nr_pages - 1); 2849 2850 /* Check if it makes any sense to call ->map_pages */ 2851 while (!pte_none(*pte)) { 2852 if (++pgoff > max_pgoff) 2853 return; 2854 start_addr += PAGE_SIZE; 2855 if (start_addr >= vma->vm_end) 2856 return; 2857 pte++; 2858 } 2859 2860 vmf.virtual_address = (void __user *) start_addr; 2861 vmf.pte = pte; 2862 vmf.pgoff = pgoff; 2863 vmf.max_pgoff = max_pgoff; 2864 vmf.flags = flags; 2865 vma->vm_ops->map_pages(vma, &vmf); 2866 } 2867 2868 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2869 unsigned long address, pmd_t *pmd, 2870 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 2871 { 2872 struct page *fault_page; 2873 spinlock_t *ptl; 2874 pte_t *pte; 2875 int ret = 0; 2876 2877 /* 2878 * Let's call ->map_pages() first and use ->fault() as fallback 2879 * if page by the offset is not ready to be mapped (cold cache or 2880 * something). 2881 */ 2882 if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) && 2883 fault_around_bytes >> PAGE_SHIFT > 1) { 2884 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2885 do_fault_around(vma, address, pte, pgoff, flags); 2886 if (!pte_same(*pte, orig_pte)) 2887 goto unlock_out; 2888 pte_unmap_unlock(pte, ptl); 2889 } 2890 2891 ret = __do_fault(vma, address, pgoff, flags, &fault_page); 2892 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2893 return ret; 2894 2895 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2896 if (unlikely(!pte_same(*pte, orig_pte))) { 2897 pte_unmap_unlock(pte, ptl); 2898 unlock_page(fault_page); 2899 page_cache_release(fault_page); 2900 return ret; 2901 } 2902 do_set_pte(vma, address, fault_page, pte, false, false); 2903 unlock_page(fault_page); 2904 unlock_out: 2905 pte_unmap_unlock(pte, ptl); 2906 return ret; 2907 } 2908 2909 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2910 unsigned long address, pmd_t *pmd, 2911 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 2912 { 2913 struct page *fault_page, *new_page; 2914 struct mem_cgroup *memcg; 2915 spinlock_t *ptl; 2916 pte_t *pte; 2917 int ret; 2918 2919 if (unlikely(anon_vma_prepare(vma))) 2920 return VM_FAULT_OOM; 2921 2922 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 2923 if (!new_page) 2924 return VM_FAULT_OOM; 2925 2926 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) { 2927 page_cache_release(new_page); 2928 return VM_FAULT_OOM; 2929 } 2930 2931 ret = __do_fault(vma, address, pgoff, flags, &fault_page); 2932 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2933 goto uncharge_out; 2934 2935 copy_user_highpage(new_page, fault_page, address, vma); 2936 __SetPageUptodate(new_page); 2937 2938 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2939 if (unlikely(!pte_same(*pte, orig_pte))) { 2940 pte_unmap_unlock(pte, ptl); 2941 unlock_page(fault_page); 2942 page_cache_release(fault_page); 2943 goto uncharge_out; 2944 } 2945 do_set_pte(vma, address, new_page, pte, true, true); 2946 mem_cgroup_commit_charge(new_page, memcg, false); 2947 lru_cache_add_active_or_unevictable(new_page, vma); 2948 pte_unmap_unlock(pte, ptl); 2949 unlock_page(fault_page); 2950 page_cache_release(fault_page); 2951 return ret; 2952 uncharge_out: 2953 mem_cgroup_cancel_charge(new_page, memcg); 2954 page_cache_release(new_page); 2955 return ret; 2956 } 2957 2958 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2959 unsigned long address, pmd_t *pmd, 2960 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 2961 { 2962 struct page *fault_page; 2963 struct address_space *mapping; 2964 spinlock_t *ptl; 2965 pte_t *pte; 2966 int dirtied = 0; 2967 int ret, tmp; 2968 2969 ret = __do_fault(vma, address, pgoff, flags, &fault_page); 2970 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2971 return ret; 2972 2973 /* 2974 * Check if the backing address space wants to know that the page is 2975 * about to become writable 2976 */ 2977 if (vma->vm_ops->page_mkwrite) { 2978 unlock_page(fault_page); 2979 tmp = do_page_mkwrite(vma, fault_page, address); 2980 if (unlikely(!tmp || 2981 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 2982 page_cache_release(fault_page); 2983 return tmp; 2984 } 2985 } 2986 2987 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2988 if (unlikely(!pte_same(*pte, orig_pte))) { 2989 pte_unmap_unlock(pte, ptl); 2990 unlock_page(fault_page); 2991 page_cache_release(fault_page); 2992 return ret; 2993 } 2994 do_set_pte(vma, address, fault_page, pte, true, false); 2995 pte_unmap_unlock(pte, ptl); 2996 2997 if (set_page_dirty(fault_page)) 2998 dirtied = 1; 2999 /* 3000 * Take a local copy of the address_space - page.mapping may be zeroed 3001 * by truncate after unlock_page(). The address_space itself remains 3002 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 3003 * release semantics to prevent the compiler from undoing this copying. 3004 */ 3005 mapping = fault_page->mapping; 3006 unlock_page(fault_page); 3007 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) { 3008 /* 3009 * Some device drivers do not set page.mapping but still 3010 * dirty their pages 3011 */ 3012 balance_dirty_pages_ratelimited(mapping); 3013 } 3014 3015 /* file_update_time outside page_lock */ 3016 if (vma->vm_file && !vma->vm_ops->page_mkwrite) 3017 file_update_time(vma->vm_file); 3018 3019 return ret; 3020 } 3021 3022 /* 3023 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3024 * but allow concurrent faults). 3025 * The mmap_sem may have been released depending on flags and our 3026 * return value. See filemap_fault() and __lock_page_or_retry(). 3027 */ 3028 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3029 unsigned long address, pte_t *page_table, pmd_t *pmd, 3030 unsigned int flags, pte_t orig_pte) 3031 { 3032 pgoff_t pgoff = (((address & PAGE_MASK) 3033 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; 3034 3035 pte_unmap(page_table); 3036 if (!(flags & FAULT_FLAG_WRITE)) 3037 return do_read_fault(mm, vma, address, pmd, pgoff, flags, 3038 orig_pte); 3039 if (!(vma->vm_flags & VM_SHARED)) 3040 return do_cow_fault(mm, vma, address, pmd, pgoff, flags, 3041 orig_pte); 3042 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); 3043 } 3044 3045 /* 3046 * Fault of a previously existing named mapping. Repopulate the pte 3047 * from the encoded file_pte if possible. This enables swappable 3048 * nonlinear vmas. 3049 * 3050 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3051 * but allow concurrent faults), and pte mapped but not yet locked. 3052 * We return with pte unmapped and unlocked. 3053 * The mmap_sem may have been released depending on flags and our 3054 * return value. See filemap_fault() and __lock_page_or_retry(). 3055 */ 3056 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3057 unsigned long address, pte_t *page_table, pmd_t *pmd, 3058 unsigned int flags, pte_t orig_pte) 3059 { 3060 pgoff_t pgoff; 3061 3062 flags |= FAULT_FLAG_NONLINEAR; 3063 3064 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 3065 return 0; 3066 3067 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) { 3068 /* 3069 * Page table corrupted: show pte and kill process. 3070 */ 3071 print_bad_pte(vma, address, orig_pte, NULL); 3072 return VM_FAULT_SIGBUS; 3073 } 3074 3075 pgoff = pte_to_pgoff(orig_pte); 3076 if (!(flags & FAULT_FLAG_WRITE)) 3077 return do_read_fault(mm, vma, address, pmd, pgoff, flags, 3078 orig_pte); 3079 if (!(vma->vm_flags & VM_SHARED)) 3080 return do_cow_fault(mm, vma, address, pmd, pgoff, flags, 3081 orig_pte); 3082 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); 3083 } 3084 3085 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 3086 unsigned long addr, int page_nid, 3087 int *flags) 3088 { 3089 get_page(page); 3090 3091 count_vm_numa_event(NUMA_HINT_FAULTS); 3092 if (page_nid == numa_node_id()) { 3093 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 3094 *flags |= TNF_FAULT_LOCAL; 3095 } 3096 3097 return mpol_misplaced(page, vma, addr); 3098 } 3099 3100 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, 3101 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd) 3102 { 3103 struct page *page = NULL; 3104 spinlock_t *ptl; 3105 int page_nid = -1; 3106 int last_cpupid; 3107 int target_nid; 3108 bool migrated = false; 3109 int flags = 0; 3110 3111 /* 3112 * The "pte" at this point cannot be used safely without 3113 * validation through pte_unmap_same(). It's of NUMA type but 3114 * the pfn may be screwed if the read is non atomic. 3115 * 3116 * ptep_modify_prot_start is not called as this is clearing 3117 * the _PAGE_NUMA bit and it is not really expected that there 3118 * would be concurrent hardware modifications to the PTE. 3119 */ 3120 ptl = pte_lockptr(mm, pmd); 3121 spin_lock(ptl); 3122 if (unlikely(!pte_same(*ptep, pte))) { 3123 pte_unmap_unlock(ptep, ptl); 3124 goto out; 3125 } 3126 3127 pte = pte_mknonnuma(pte); 3128 set_pte_at(mm, addr, ptep, pte); 3129 update_mmu_cache(vma, addr, ptep); 3130 3131 page = vm_normal_page(vma, addr, pte); 3132 if (!page) { 3133 pte_unmap_unlock(ptep, ptl); 3134 return 0; 3135 } 3136 BUG_ON(is_zero_pfn(page_to_pfn(page))); 3137 3138 /* 3139 * Avoid grouping on DSO/COW pages in specific and RO pages 3140 * in general, RO pages shouldn't hurt as much anyway since 3141 * they can be in shared cache state. 3142 */ 3143 if (!pte_write(pte)) 3144 flags |= TNF_NO_GROUP; 3145 3146 /* 3147 * Flag if the page is shared between multiple address spaces. This 3148 * is later used when determining whether to group tasks together 3149 */ 3150 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 3151 flags |= TNF_SHARED; 3152 3153 last_cpupid = page_cpupid_last(page); 3154 page_nid = page_to_nid(page); 3155 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags); 3156 pte_unmap_unlock(ptep, ptl); 3157 if (target_nid == -1) { 3158 put_page(page); 3159 goto out; 3160 } 3161 3162 /* Migrate to the requested node */ 3163 migrated = migrate_misplaced_page(page, vma, target_nid); 3164 if (migrated) { 3165 page_nid = target_nid; 3166 flags |= TNF_MIGRATED; 3167 } 3168 3169 out: 3170 if (page_nid != -1) 3171 task_numa_fault(last_cpupid, page_nid, 1, flags); 3172 return 0; 3173 } 3174 3175 /* 3176 * These routines also need to handle stuff like marking pages dirty 3177 * and/or accessed for architectures that don't do it in hardware (most 3178 * RISC architectures). The early dirtying is also good on the i386. 3179 * 3180 * There is also a hook called "update_mmu_cache()" that architectures 3181 * with external mmu caches can use to update those (ie the Sparc or 3182 * PowerPC hashed page tables that act as extended TLBs). 3183 * 3184 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3185 * but allow concurrent faults), and pte mapped but not yet locked. 3186 * We return with pte unmapped and unlocked. 3187 * 3188 * The mmap_sem may have been released depending on flags and our 3189 * return value. See filemap_fault() and __lock_page_or_retry(). 3190 */ 3191 static int handle_pte_fault(struct mm_struct *mm, 3192 struct vm_area_struct *vma, unsigned long address, 3193 pte_t *pte, pmd_t *pmd, unsigned int flags) 3194 { 3195 pte_t entry; 3196 spinlock_t *ptl; 3197 3198 /* 3199 * some architectures can have larger ptes than wordsize, 3200 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y, 3201 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses. 3202 * The code below just needs a consistent view for the ifs and 3203 * we later double check anyway with the ptl lock held. So here 3204 * a barrier will do. 3205 */ 3206 entry = *pte; 3207 barrier(); 3208 if (!pte_present(entry)) { 3209 if (pte_none(entry)) { 3210 if (vma->vm_ops) { 3211 if (likely(vma->vm_ops->fault)) 3212 return do_linear_fault(mm, vma, address, 3213 pte, pmd, flags, entry); 3214 } 3215 return do_anonymous_page(mm, vma, address, 3216 pte, pmd, flags); 3217 } 3218 if (pte_file(entry)) 3219 return do_nonlinear_fault(mm, vma, address, 3220 pte, pmd, flags, entry); 3221 return do_swap_page(mm, vma, address, 3222 pte, pmd, flags, entry); 3223 } 3224 3225 if (pte_numa(entry)) 3226 return do_numa_page(mm, vma, address, entry, pte, pmd); 3227 3228 ptl = pte_lockptr(mm, pmd); 3229 spin_lock(ptl); 3230 if (unlikely(!pte_same(*pte, entry))) 3231 goto unlock; 3232 if (flags & FAULT_FLAG_WRITE) { 3233 if (!pte_write(entry)) 3234 return do_wp_page(mm, vma, address, 3235 pte, pmd, ptl, entry); 3236 entry = pte_mkdirty(entry); 3237 } 3238 entry = pte_mkyoung(entry); 3239 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) { 3240 update_mmu_cache(vma, address, pte); 3241 } else { 3242 /* 3243 * This is needed only for protection faults but the arch code 3244 * is not yet telling us if this is a protection fault or not. 3245 * This still avoids useless tlb flushes for .text page faults 3246 * with threads. 3247 */ 3248 if (flags & FAULT_FLAG_WRITE) 3249 flush_tlb_fix_spurious_fault(vma, address); 3250 } 3251 unlock: 3252 pte_unmap_unlock(pte, ptl); 3253 return 0; 3254 } 3255 3256 /* 3257 * By the time we get here, we already hold the mm semaphore 3258 * 3259 * The mmap_sem may have been released depending on flags and our 3260 * return value. See filemap_fault() and __lock_page_or_retry(). 3261 */ 3262 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3263 unsigned long address, unsigned int flags) 3264 { 3265 pgd_t *pgd; 3266 pud_t *pud; 3267 pmd_t *pmd; 3268 pte_t *pte; 3269 3270 if (unlikely(is_vm_hugetlb_page(vma))) 3271 return hugetlb_fault(mm, vma, address, flags); 3272 3273 pgd = pgd_offset(mm, address); 3274 pud = pud_alloc(mm, pgd, address); 3275 if (!pud) 3276 return VM_FAULT_OOM; 3277 pmd = pmd_alloc(mm, pud, address); 3278 if (!pmd) 3279 return VM_FAULT_OOM; 3280 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { 3281 int ret = VM_FAULT_FALLBACK; 3282 if (!vma->vm_ops) 3283 ret = do_huge_pmd_anonymous_page(mm, vma, address, 3284 pmd, flags); 3285 if (!(ret & VM_FAULT_FALLBACK)) 3286 return ret; 3287 } else { 3288 pmd_t orig_pmd = *pmd; 3289 int ret; 3290 3291 barrier(); 3292 if (pmd_trans_huge(orig_pmd)) { 3293 unsigned int dirty = flags & FAULT_FLAG_WRITE; 3294 3295 /* 3296 * If the pmd is splitting, return and retry the 3297 * the fault. Alternative: wait until the split 3298 * is done, and goto retry. 3299 */ 3300 if (pmd_trans_splitting(orig_pmd)) 3301 return 0; 3302 3303 if (pmd_numa(orig_pmd)) 3304 return do_huge_pmd_numa_page(mm, vma, address, 3305 orig_pmd, pmd); 3306 3307 if (dirty && !pmd_write(orig_pmd)) { 3308 ret = do_huge_pmd_wp_page(mm, vma, address, pmd, 3309 orig_pmd); 3310 if (!(ret & VM_FAULT_FALLBACK)) 3311 return ret; 3312 } else { 3313 huge_pmd_set_accessed(mm, vma, address, pmd, 3314 orig_pmd, dirty); 3315 return 0; 3316 } 3317 } 3318 } 3319 3320 /* 3321 * Use __pte_alloc instead of pte_alloc_map, because we can't 3322 * run pte_offset_map on the pmd, if an huge pmd could 3323 * materialize from under us from a different thread. 3324 */ 3325 if (unlikely(pmd_none(*pmd)) && 3326 unlikely(__pte_alloc(mm, vma, pmd, address))) 3327 return VM_FAULT_OOM; 3328 /* if an huge pmd materialized from under us just retry later */ 3329 if (unlikely(pmd_trans_huge(*pmd))) 3330 return 0; 3331 /* 3332 * A regular pmd is established and it can't morph into a huge pmd 3333 * from under us anymore at this point because we hold the mmap_sem 3334 * read mode and khugepaged takes it in write mode. So now it's 3335 * safe to run pte_offset_map(). 3336 */ 3337 pte = pte_offset_map(pmd, address); 3338 3339 return handle_pte_fault(mm, vma, address, pte, pmd, flags); 3340 } 3341 3342 /* 3343 * By the time we get here, we already hold the mm semaphore 3344 * 3345 * The mmap_sem may have been released depending on flags and our 3346 * return value. See filemap_fault() and __lock_page_or_retry(). 3347 */ 3348 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3349 unsigned long address, unsigned int flags) 3350 { 3351 int ret; 3352 3353 __set_current_state(TASK_RUNNING); 3354 3355 count_vm_event(PGFAULT); 3356 mem_cgroup_count_vm_event(mm, PGFAULT); 3357 3358 /* do counter updates before entering really critical section. */ 3359 check_sync_rss_stat(current); 3360 3361 /* 3362 * Enable the memcg OOM handling for faults triggered in user 3363 * space. Kernel faults are handled more gracefully. 3364 */ 3365 if (flags & FAULT_FLAG_USER) 3366 mem_cgroup_oom_enable(); 3367 3368 ret = __handle_mm_fault(mm, vma, address, flags); 3369 3370 if (flags & FAULT_FLAG_USER) { 3371 mem_cgroup_oom_disable(); 3372 /* 3373 * The task may have entered a memcg OOM situation but 3374 * if the allocation error was handled gracefully (no 3375 * VM_FAULT_OOM), there is no need to kill anything. 3376 * Just clean up the OOM state peacefully. 3377 */ 3378 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 3379 mem_cgroup_oom_synchronize(false); 3380 } 3381 3382 return ret; 3383 } 3384 EXPORT_SYMBOL_GPL(handle_mm_fault); 3385 3386 #ifndef __PAGETABLE_PUD_FOLDED 3387 /* 3388 * Allocate page upper directory. 3389 * We've already handled the fast-path in-line. 3390 */ 3391 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 3392 { 3393 pud_t *new = pud_alloc_one(mm, address); 3394 if (!new) 3395 return -ENOMEM; 3396 3397 smp_wmb(); /* See comment in __pte_alloc */ 3398 3399 spin_lock(&mm->page_table_lock); 3400 if (pgd_present(*pgd)) /* Another has populated it */ 3401 pud_free(mm, new); 3402 else 3403 pgd_populate(mm, pgd, new); 3404 spin_unlock(&mm->page_table_lock); 3405 return 0; 3406 } 3407 #endif /* __PAGETABLE_PUD_FOLDED */ 3408 3409 #ifndef __PAGETABLE_PMD_FOLDED 3410 /* 3411 * Allocate page middle directory. 3412 * We've already handled the fast-path in-line. 3413 */ 3414 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 3415 { 3416 pmd_t *new = pmd_alloc_one(mm, address); 3417 if (!new) 3418 return -ENOMEM; 3419 3420 smp_wmb(); /* See comment in __pte_alloc */ 3421 3422 spin_lock(&mm->page_table_lock); 3423 #ifndef __ARCH_HAS_4LEVEL_HACK 3424 if (pud_present(*pud)) /* Another has populated it */ 3425 pmd_free(mm, new); 3426 else 3427 pud_populate(mm, pud, new); 3428 #else 3429 if (pgd_present(*pud)) /* Another has populated it */ 3430 pmd_free(mm, new); 3431 else 3432 pgd_populate(mm, pud, new); 3433 #endif /* __ARCH_HAS_4LEVEL_HACK */ 3434 spin_unlock(&mm->page_table_lock); 3435 return 0; 3436 } 3437 #endif /* __PAGETABLE_PMD_FOLDED */ 3438 3439 static int __follow_pte(struct mm_struct *mm, unsigned long address, 3440 pte_t **ptepp, spinlock_t **ptlp) 3441 { 3442 pgd_t *pgd; 3443 pud_t *pud; 3444 pmd_t *pmd; 3445 pte_t *ptep; 3446 3447 pgd = pgd_offset(mm, address); 3448 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 3449 goto out; 3450 3451 pud = pud_offset(pgd, address); 3452 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 3453 goto out; 3454 3455 pmd = pmd_offset(pud, address); 3456 VM_BUG_ON(pmd_trans_huge(*pmd)); 3457 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 3458 goto out; 3459 3460 /* We cannot handle huge page PFN maps. Luckily they don't exist. */ 3461 if (pmd_huge(*pmd)) 3462 goto out; 3463 3464 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 3465 if (!ptep) 3466 goto out; 3467 if (!pte_present(*ptep)) 3468 goto unlock; 3469 *ptepp = ptep; 3470 return 0; 3471 unlock: 3472 pte_unmap_unlock(ptep, *ptlp); 3473 out: 3474 return -EINVAL; 3475 } 3476 3477 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 3478 pte_t **ptepp, spinlock_t **ptlp) 3479 { 3480 int res; 3481 3482 /* (void) is needed to make gcc happy */ 3483 (void) __cond_lock(*ptlp, 3484 !(res = __follow_pte(mm, address, ptepp, ptlp))); 3485 return res; 3486 } 3487 3488 /** 3489 * follow_pfn - look up PFN at a user virtual address 3490 * @vma: memory mapping 3491 * @address: user virtual address 3492 * @pfn: location to store found PFN 3493 * 3494 * Only IO mappings and raw PFN mappings are allowed. 3495 * 3496 * Returns zero and the pfn at @pfn on success, -ve otherwise. 3497 */ 3498 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 3499 unsigned long *pfn) 3500 { 3501 int ret = -EINVAL; 3502 spinlock_t *ptl; 3503 pte_t *ptep; 3504 3505 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3506 return ret; 3507 3508 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 3509 if (ret) 3510 return ret; 3511 *pfn = pte_pfn(*ptep); 3512 pte_unmap_unlock(ptep, ptl); 3513 return 0; 3514 } 3515 EXPORT_SYMBOL(follow_pfn); 3516 3517 #ifdef CONFIG_HAVE_IOREMAP_PROT 3518 int follow_phys(struct vm_area_struct *vma, 3519 unsigned long address, unsigned int flags, 3520 unsigned long *prot, resource_size_t *phys) 3521 { 3522 int ret = -EINVAL; 3523 pte_t *ptep, pte; 3524 spinlock_t *ptl; 3525 3526 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3527 goto out; 3528 3529 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 3530 goto out; 3531 pte = *ptep; 3532 3533 if ((flags & FOLL_WRITE) && !pte_write(pte)) 3534 goto unlock; 3535 3536 *prot = pgprot_val(pte_pgprot(pte)); 3537 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 3538 3539 ret = 0; 3540 unlock: 3541 pte_unmap_unlock(ptep, ptl); 3542 out: 3543 return ret; 3544 } 3545 3546 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 3547 void *buf, int len, int write) 3548 { 3549 resource_size_t phys_addr; 3550 unsigned long prot = 0; 3551 void __iomem *maddr; 3552 int offset = addr & (PAGE_SIZE-1); 3553 3554 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 3555 return -EINVAL; 3556 3557 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot); 3558 if (write) 3559 memcpy_toio(maddr + offset, buf, len); 3560 else 3561 memcpy_fromio(buf, maddr + offset, len); 3562 iounmap(maddr); 3563 3564 return len; 3565 } 3566 EXPORT_SYMBOL_GPL(generic_access_phys); 3567 #endif 3568 3569 /* 3570 * Access another process' address space as given in mm. If non-NULL, use the 3571 * given task for page fault accounting. 3572 */ 3573 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, 3574 unsigned long addr, void *buf, int len, int write) 3575 { 3576 struct vm_area_struct *vma; 3577 void *old_buf = buf; 3578 3579 down_read(&mm->mmap_sem); 3580 /* ignore errors, just check how much was successfully transferred */ 3581 while (len) { 3582 int bytes, ret, offset; 3583 void *maddr; 3584 struct page *page = NULL; 3585 3586 ret = get_user_pages(tsk, mm, addr, 1, 3587 write, 1, &page, &vma); 3588 if (ret <= 0) { 3589 #ifndef CONFIG_HAVE_IOREMAP_PROT 3590 break; 3591 #else 3592 /* 3593 * Check if this is a VM_IO | VM_PFNMAP VMA, which 3594 * we can access using slightly different code. 3595 */ 3596 vma = find_vma(mm, addr); 3597 if (!vma || vma->vm_start > addr) 3598 break; 3599 if (vma->vm_ops && vma->vm_ops->access) 3600 ret = vma->vm_ops->access(vma, addr, buf, 3601 len, write); 3602 if (ret <= 0) 3603 break; 3604 bytes = ret; 3605 #endif 3606 } else { 3607 bytes = len; 3608 offset = addr & (PAGE_SIZE-1); 3609 if (bytes > PAGE_SIZE-offset) 3610 bytes = PAGE_SIZE-offset; 3611 3612 maddr = kmap(page); 3613 if (write) { 3614 copy_to_user_page(vma, page, addr, 3615 maddr + offset, buf, bytes); 3616 set_page_dirty_lock(page); 3617 } else { 3618 copy_from_user_page(vma, page, addr, 3619 buf, maddr + offset, bytes); 3620 } 3621 kunmap(page); 3622 page_cache_release(page); 3623 } 3624 len -= bytes; 3625 buf += bytes; 3626 addr += bytes; 3627 } 3628 up_read(&mm->mmap_sem); 3629 3630 return buf - old_buf; 3631 } 3632 3633 /** 3634 * access_remote_vm - access another process' address space 3635 * @mm: the mm_struct of the target address space 3636 * @addr: start address to access 3637 * @buf: source or destination buffer 3638 * @len: number of bytes to transfer 3639 * @write: whether the access is a write 3640 * 3641 * The caller must hold a reference on @mm. 3642 */ 3643 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 3644 void *buf, int len, int write) 3645 { 3646 return __access_remote_vm(NULL, mm, addr, buf, len, write); 3647 } 3648 3649 /* 3650 * Access another process' address space. 3651 * Source/target buffer must be kernel space, 3652 * Do not walk the page table directly, use get_user_pages 3653 */ 3654 int access_process_vm(struct task_struct *tsk, unsigned long addr, 3655 void *buf, int len, int write) 3656 { 3657 struct mm_struct *mm; 3658 int ret; 3659 3660 mm = get_task_mm(tsk); 3661 if (!mm) 3662 return 0; 3663 3664 ret = __access_remote_vm(tsk, mm, addr, buf, len, write); 3665 mmput(mm); 3666 3667 return ret; 3668 } 3669 3670 /* 3671 * Print the name of a VMA. 3672 */ 3673 void print_vma_addr(char *prefix, unsigned long ip) 3674 { 3675 struct mm_struct *mm = current->mm; 3676 struct vm_area_struct *vma; 3677 3678 /* 3679 * Do not print if we are in atomic 3680 * contexts (in exception stacks, etc.): 3681 */ 3682 if (preempt_count()) 3683 return; 3684 3685 down_read(&mm->mmap_sem); 3686 vma = find_vma(mm, ip); 3687 if (vma && vma->vm_file) { 3688 struct file *f = vma->vm_file; 3689 char *buf = (char *)__get_free_page(GFP_KERNEL); 3690 if (buf) { 3691 char *p; 3692 3693 p = d_path(&f->f_path, buf, PAGE_SIZE); 3694 if (IS_ERR(p)) 3695 p = "?"; 3696 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 3697 vma->vm_start, 3698 vma->vm_end - vma->vm_start); 3699 free_page((unsigned long)buf); 3700 } 3701 } 3702 up_read(&mm->mmap_sem); 3703 } 3704 3705 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 3706 void might_fault(void) 3707 { 3708 /* 3709 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 3710 * holding the mmap_sem, this is safe because kernel memory doesn't 3711 * get paged out, therefore we'll never actually fault, and the 3712 * below annotations will generate false positives. 3713 */ 3714 if (segment_eq(get_fs(), KERNEL_DS)) 3715 return; 3716 3717 /* 3718 * it would be nicer only to annotate paths which are not under 3719 * pagefault_disable, however that requires a larger audit and 3720 * providing helpers like get_user_atomic. 3721 */ 3722 if (in_atomic()) 3723 return; 3724 3725 __might_sleep(__FILE__, __LINE__, 0); 3726 3727 if (current->mm) 3728 might_lock_read(¤t->mm->mmap_sem); 3729 } 3730 EXPORT_SYMBOL(might_fault); 3731 #endif 3732 3733 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 3734 static void clear_gigantic_page(struct page *page, 3735 unsigned long addr, 3736 unsigned int pages_per_huge_page) 3737 { 3738 int i; 3739 struct page *p = page; 3740 3741 might_sleep(); 3742 for (i = 0; i < pages_per_huge_page; 3743 i++, p = mem_map_next(p, page, i)) { 3744 cond_resched(); 3745 clear_user_highpage(p, addr + i * PAGE_SIZE); 3746 } 3747 } 3748 void clear_huge_page(struct page *page, 3749 unsigned long addr, unsigned int pages_per_huge_page) 3750 { 3751 int i; 3752 3753 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3754 clear_gigantic_page(page, addr, pages_per_huge_page); 3755 return; 3756 } 3757 3758 might_sleep(); 3759 for (i = 0; i < pages_per_huge_page; i++) { 3760 cond_resched(); 3761 clear_user_highpage(page + i, addr + i * PAGE_SIZE); 3762 } 3763 } 3764 3765 static void copy_user_gigantic_page(struct page *dst, struct page *src, 3766 unsigned long addr, 3767 struct vm_area_struct *vma, 3768 unsigned int pages_per_huge_page) 3769 { 3770 int i; 3771 struct page *dst_base = dst; 3772 struct page *src_base = src; 3773 3774 for (i = 0; i < pages_per_huge_page; ) { 3775 cond_resched(); 3776 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 3777 3778 i++; 3779 dst = mem_map_next(dst, dst_base, i); 3780 src = mem_map_next(src, src_base, i); 3781 } 3782 } 3783 3784 void copy_user_huge_page(struct page *dst, struct page *src, 3785 unsigned long addr, struct vm_area_struct *vma, 3786 unsigned int pages_per_huge_page) 3787 { 3788 int i; 3789 3790 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3791 copy_user_gigantic_page(dst, src, addr, vma, 3792 pages_per_huge_page); 3793 return; 3794 } 3795 3796 might_sleep(); 3797 for (i = 0; i < pages_per_huge_page; i++) { 3798 cond_resched(); 3799 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 3800 } 3801 } 3802 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 3803 3804 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 3805 3806 static struct kmem_cache *page_ptl_cachep; 3807 3808 void __init ptlock_cache_init(void) 3809 { 3810 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 3811 SLAB_PANIC, NULL); 3812 } 3813 3814 bool ptlock_alloc(struct page *page) 3815 { 3816 spinlock_t *ptl; 3817 3818 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 3819 if (!ptl) 3820 return false; 3821 page->ptl = ptl; 3822 return true; 3823 } 3824 3825 void ptlock_free(struct page *page) 3826 { 3827 kmem_cache_free(page_ptl_cachep, page->ptl); 3828 } 3829 #endif 3830