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/pfn_t.h> 54 #include <linux/writeback.h> 55 #include <linux/memcontrol.h> 56 #include <linux/mmu_notifier.h> 57 #include <linux/kallsyms.h> 58 #include <linux/swapops.h> 59 #include <linux/elf.h> 60 #include <linux/gfp.h> 61 #include <linux/migrate.h> 62 #include <linux/string.h> 63 #include <linux/dma-debug.h> 64 #include <linux/debugfs.h> 65 #include <linux/userfaultfd_k.h> 66 67 #include <asm/io.h> 68 #include <asm/pgalloc.h> 69 #include <asm/uaccess.h> 70 #include <asm/tlb.h> 71 #include <asm/tlbflush.h> 72 #include <asm/pgtable.h> 73 74 #include "internal.h" 75 76 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS 77 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 78 #endif 79 80 #ifndef CONFIG_NEED_MULTIPLE_NODES 81 /* use the per-pgdat data instead for discontigmem - mbligh */ 82 unsigned long max_mapnr; 83 struct page *mem_map; 84 85 EXPORT_SYMBOL(max_mapnr); 86 EXPORT_SYMBOL(mem_map); 87 #endif 88 89 /* 90 * A number of key systems in x86 including ioremap() rely on the assumption 91 * that high_memory defines the upper bound on direct map memory, then end 92 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 93 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 94 * and ZONE_HIGHMEM. 95 */ 96 void * high_memory; 97 98 EXPORT_SYMBOL(high_memory); 99 100 /* 101 * Randomize the address space (stacks, mmaps, brk, etc.). 102 * 103 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 104 * as ancient (libc5 based) binaries can segfault. ) 105 */ 106 int randomize_va_space __read_mostly = 107 #ifdef CONFIG_COMPAT_BRK 108 1; 109 #else 110 2; 111 #endif 112 113 static int __init disable_randmaps(char *s) 114 { 115 randomize_va_space = 0; 116 return 1; 117 } 118 __setup("norandmaps", disable_randmaps); 119 120 unsigned long zero_pfn __read_mostly; 121 unsigned long highest_memmap_pfn __read_mostly; 122 123 EXPORT_SYMBOL(zero_pfn); 124 125 /* 126 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 127 */ 128 static int __init init_zero_pfn(void) 129 { 130 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 131 return 0; 132 } 133 core_initcall(init_zero_pfn); 134 135 136 #if defined(SPLIT_RSS_COUNTING) 137 138 void sync_mm_rss(struct mm_struct *mm) 139 { 140 int i; 141 142 for (i = 0; i < NR_MM_COUNTERS; i++) { 143 if (current->rss_stat.count[i]) { 144 add_mm_counter(mm, i, current->rss_stat.count[i]); 145 current->rss_stat.count[i] = 0; 146 } 147 } 148 current->rss_stat.events = 0; 149 } 150 151 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 152 { 153 struct task_struct *task = current; 154 155 if (likely(task->mm == mm)) 156 task->rss_stat.count[member] += val; 157 else 158 add_mm_counter(mm, member, val); 159 } 160 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 161 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 162 163 /* sync counter once per 64 page faults */ 164 #define TASK_RSS_EVENTS_THRESH (64) 165 static void check_sync_rss_stat(struct task_struct *task) 166 { 167 if (unlikely(task != current)) 168 return; 169 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 170 sync_mm_rss(task->mm); 171 } 172 #else /* SPLIT_RSS_COUNTING */ 173 174 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 175 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 176 177 static void check_sync_rss_stat(struct task_struct *task) 178 { 179 } 180 181 #endif /* SPLIT_RSS_COUNTING */ 182 183 #ifdef HAVE_GENERIC_MMU_GATHER 184 185 static bool tlb_next_batch(struct mmu_gather *tlb) 186 { 187 struct mmu_gather_batch *batch; 188 189 batch = tlb->active; 190 if (batch->next) { 191 tlb->active = batch->next; 192 return true; 193 } 194 195 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT) 196 return false; 197 198 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0); 199 if (!batch) 200 return false; 201 202 tlb->batch_count++; 203 batch->next = NULL; 204 batch->nr = 0; 205 batch->max = MAX_GATHER_BATCH; 206 207 tlb->active->next = batch; 208 tlb->active = batch; 209 210 return true; 211 } 212 213 /* tlb_gather_mmu 214 * Called to initialize an (on-stack) mmu_gather structure for page-table 215 * tear-down from @mm. The @fullmm argument is used when @mm is without 216 * users and we're going to destroy the full address space (exit/execve). 217 */ 218 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end) 219 { 220 tlb->mm = mm; 221 222 /* Is it from 0 to ~0? */ 223 tlb->fullmm = !(start | (end+1)); 224 tlb->need_flush_all = 0; 225 tlb->local.next = NULL; 226 tlb->local.nr = 0; 227 tlb->local.max = ARRAY_SIZE(tlb->__pages); 228 tlb->active = &tlb->local; 229 tlb->batch_count = 0; 230 231 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 232 tlb->batch = NULL; 233 #endif 234 235 __tlb_reset_range(tlb); 236 } 237 238 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb) 239 { 240 if (!tlb->end) 241 return; 242 243 tlb_flush(tlb); 244 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end); 245 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 246 tlb_table_flush(tlb); 247 #endif 248 __tlb_reset_range(tlb); 249 } 250 251 static void tlb_flush_mmu_free(struct mmu_gather *tlb) 252 { 253 struct mmu_gather_batch *batch; 254 255 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) { 256 free_pages_and_swap_cache(batch->pages, batch->nr); 257 batch->nr = 0; 258 } 259 tlb->active = &tlb->local; 260 } 261 262 void tlb_flush_mmu(struct mmu_gather *tlb) 263 { 264 tlb_flush_mmu_tlbonly(tlb); 265 tlb_flush_mmu_free(tlb); 266 } 267 268 /* tlb_finish_mmu 269 * Called at the end of the shootdown operation to free up any resources 270 * that were required. 271 */ 272 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) 273 { 274 struct mmu_gather_batch *batch, *next; 275 276 tlb_flush_mmu(tlb); 277 278 /* keep the page table cache within bounds */ 279 check_pgt_cache(); 280 281 for (batch = tlb->local.next; batch; batch = next) { 282 next = batch->next; 283 free_pages((unsigned long)batch, 0); 284 } 285 tlb->local.next = NULL; 286 } 287 288 /* __tlb_remove_page 289 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while 290 * handling the additional races in SMP caused by other CPUs caching valid 291 * mappings in their TLBs. Returns the number of free page slots left. 292 * When out of page slots we must call tlb_flush_mmu(). 293 */ 294 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) 295 { 296 struct mmu_gather_batch *batch; 297 298 VM_BUG_ON(!tlb->end); 299 300 batch = tlb->active; 301 batch->pages[batch->nr++] = page; 302 if (batch->nr == batch->max) { 303 if (!tlb_next_batch(tlb)) 304 return 0; 305 batch = tlb->active; 306 } 307 VM_BUG_ON_PAGE(batch->nr > batch->max, page); 308 309 return batch->max - batch->nr; 310 } 311 312 #endif /* HAVE_GENERIC_MMU_GATHER */ 313 314 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 315 316 /* 317 * See the comment near struct mmu_table_batch. 318 */ 319 320 static void tlb_remove_table_smp_sync(void *arg) 321 { 322 /* Simply deliver the interrupt */ 323 } 324 325 static void tlb_remove_table_one(void *table) 326 { 327 /* 328 * This isn't an RCU grace period and hence the page-tables cannot be 329 * assumed to be actually RCU-freed. 330 * 331 * It is however sufficient for software page-table walkers that rely on 332 * IRQ disabling. See the comment near struct mmu_table_batch. 333 */ 334 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 335 __tlb_remove_table(table); 336 } 337 338 static void tlb_remove_table_rcu(struct rcu_head *head) 339 { 340 struct mmu_table_batch *batch; 341 int i; 342 343 batch = container_of(head, struct mmu_table_batch, rcu); 344 345 for (i = 0; i < batch->nr; i++) 346 __tlb_remove_table(batch->tables[i]); 347 348 free_page((unsigned long)batch); 349 } 350 351 void tlb_table_flush(struct mmu_gather *tlb) 352 { 353 struct mmu_table_batch **batch = &tlb->batch; 354 355 if (*batch) { 356 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 357 *batch = NULL; 358 } 359 } 360 361 void tlb_remove_table(struct mmu_gather *tlb, void *table) 362 { 363 struct mmu_table_batch **batch = &tlb->batch; 364 365 /* 366 * When there's less then two users of this mm there cannot be a 367 * concurrent page-table walk. 368 */ 369 if (atomic_read(&tlb->mm->mm_users) < 2) { 370 __tlb_remove_table(table); 371 return; 372 } 373 374 if (*batch == NULL) { 375 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); 376 if (*batch == NULL) { 377 tlb_remove_table_one(table); 378 return; 379 } 380 (*batch)->nr = 0; 381 } 382 (*batch)->tables[(*batch)->nr++] = table; 383 if ((*batch)->nr == MAX_TABLE_BATCH) 384 tlb_table_flush(tlb); 385 } 386 387 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ 388 389 /* 390 * Note: this doesn't free the actual pages themselves. That 391 * has been handled earlier when unmapping all the memory regions. 392 */ 393 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 394 unsigned long addr) 395 { 396 pgtable_t token = pmd_pgtable(*pmd); 397 pmd_clear(pmd); 398 pte_free_tlb(tlb, token, addr); 399 atomic_long_dec(&tlb->mm->nr_ptes); 400 } 401 402 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 403 unsigned long addr, unsigned long end, 404 unsigned long floor, unsigned long ceiling) 405 { 406 pmd_t *pmd; 407 unsigned long next; 408 unsigned long start; 409 410 start = addr; 411 pmd = pmd_offset(pud, addr); 412 do { 413 next = pmd_addr_end(addr, end); 414 if (pmd_none_or_clear_bad(pmd)) 415 continue; 416 free_pte_range(tlb, pmd, addr); 417 } while (pmd++, addr = next, addr != end); 418 419 start &= PUD_MASK; 420 if (start < floor) 421 return; 422 if (ceiling) { 423 ceiling &= PUD_MASK; 424 if (!ceiling) 425 return; 426 } 427 if (end - 1 > ceiling - 1) 428 return; 429 430 pmd = pmd_offset(pud, start); 431 pud_clear(pud); 432 pmd_free_tlb(tlb, pmd, start); 433 mm_dec_nr_pmds(tlb->mm); 434 } 435 436 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 437 unsigned long addr, unsigned long end, 438 unsigned long floor, unsigned long ceiling) 439 { 440 pud_t *pud; 441 unsigned long next; 442 unsigned long start; 443 444 start = addr; 445 pud = pud_offset(pgd, addr); 446 do { 447 next = pud_addr_end(addr, end); 448 if (pud_none_or_clear_bad(pud)) 449 continue; 450 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 451 } while (pud++, addr = next, addr != end); 452 453 start &= PGDIR_MASK; 454 if (start < floor) 455 return; 456 if (ceiling) { 457 ceiling &= PGDIR_MASK; 458 if (!ceiling) 459 return; 460 } 461 if (end - 1 > ceiling - 1) 462 return; 463 464 pud = pud_offset(pgd, start); 465 pgd_clear(pgd); 466 pud_free_tlb(tlb, pud, start); 467 } 468 469 /* 470 * This function frees user-level page tables of a process. 471 */ 472 void free_pgd_range(struct mmu_gather *tlb, 473 unsigned long addr, unsigned long end, 474 unsigned long floor, unsigned long ceiling) 475 { 476 pgd_t *pgd; 477 unsigned long next; 478 479 /* 480 * The next few lines have given us lots of grief... 481 * 482 * Why are we testing PMD* at this top level? Because often 483 * there will be no work to do at all, and we'd prefer not to 484 * go all the way down to the bottom just to discover that. 485 * 486 * Why all these "- 1"s? Because 0 represents both the bottom 487 * of the address space and the top of it (using -1 for the 488 * top wouldn't help much: the masks would do the wrong thing). 489 * The rule is that addr 0 and floor 0 refer to the bottom of 490 * the address space, but end 0 and ceiling 0 refer to the top 491 * Comparisons need to use "end - 1" and "ceiling - 1" (though 492 * that end 0 case should be mythical). 493 * 494 * Wherever addr is brought up or ceiling brought down, we must 495 * be careful to reject "the opposite 0" before it confuses the 496 * subsequent tests. But what about where end is brought down 497 * by PMD_SIZE below? no, end can't go down to 0 there. 498 * 499 * Whereas we round start (addr) and ceiling down, by different 500 * masks at different levels, in order to test whether a table 501 * now has no other vmas using it, so can be freed, we don't 502 * bother to round floor or end up - the tests don't need that. 503 */ 504 505 addr &= PMD_MASK; 506 if (addr < floor) { 507 addr += PMD_SIZE; 508 if (!addr) 509 return; 510 } 511 if (ceiling) { 512 ceiling &= PMD_MASK; 513 if (!ceiling) 514 return; 515 } 516 if (end - 1 > ceiling - 1) 517 end -= PMD_SIZE; 518 if (addr > end - 1) 519 return; 520 521 pgd = pgd_offset(tlb->mm, addr); 522 do { 523 next = pgd_addr_end(addr, end); 524 if (pgd_none_or_clear_bad(pgd)) 525 continue; 526 free_pud_range(tlb, pgd, addr, next, floor, ceiling); 527 } while (pgd++, addr = next, addr != end); 528 } 529 530 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 531 unsigned long floor, unsigned long ceiling) 532 { 533 while (vma) { 534 struct vm_area_struct *next = vma->vm_next; 535 unsigned long addr = vma->vm_start; 536 537 /* 538 * Hide vma from rmap and truncate_pagecache before freeing 539 * pgtables 540 */ 541 unlink_anon_vmas(vma); 542 unlink_file_vma(vma); 543 544 if (is_vm_hugetlb_page(vma)) { 545 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 546 floor, next? next->vm_start: ceiling); 547 } else { 548 /* 549 * Optimization: gather nearby vmas into one call down 550 */ 551 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 552 && !is_vm_hugetlb_page(next)) { 553 vma = next; 554 next = vma->vm_next; 555 unlink_anon_vmas(vma); 556 unlink_file_vma(vma); 557 } 558 free_pgd_range(tlb, addr, vma->vm_end, 559 floor, next? next->vm_start: ceiling); 560 } 561 vma = next; 562 } 563 } 564 565 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 566 pmd_t *pmd, unsigned long address) 567 { 568 spinlock_t *ptl; 569 pgtable_t new = pte_alloc_one(mm, address); 570 if (!new) 571 return -ENOMEM; 572 573 /* 574 * Ensure all pte setup (eg. pte page lock and page clearing) are 575 * visible before the pte is made visible to other CPUs by being 576 * put into page tables. 577 * 578 * The other side of the story is the pointer chasing in the page 579 * table walking code (when walking the page table without locking; 580 * ie. most of the time). Fortunately, these data accesses consist 581 * of a chain of data-dependent loads, meaning most CPUs (alpha 582 * being the notable exception) will already guarantee loads are 583 * seen in-order. See the alpha page table accessors for the 584 * smp_read_barrier_depends() barriers in page table walking code. 585 */ 586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 587 588 ptl = pmd_lock(mm, pmd); 589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 590 atomic_long_inc(&mm->nr_ptes); 591 pmd_populate(mm, pmd, new); 592 new = NULL; 593 } 594 spin_unlock(ptl); 595 if (new) 596 pte_free(mm, new); 597 return 0; 598 } 599 600 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address) 601 { 602 pte_t *new = pte_alloc_one_kernel(&init_mm, address); 603 if (!new) 604 return -ENOMEM; 605 606 smp_wmb(); /* See comment in __pte_alloc */ 607 608 spin_lock(&init_mm.page_table_lock); 609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 610 pmd_populate_kernel(&init_mm, pmd, new); 611 new = NULL; 612 } 613 spin_unlock(&init_mm.page_table_lock); 614 if (new) 615 pte_free_kernel(&init_mm, new); 616 return 0; 617 } 618 619 static inline void init_rss_vec(int *rss) 620 { 621 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 622 } 623 624 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 625 { 626 int i; 627 628 if (current->mm == mm) 629 sync_mm_rss(mm); 630 for (i = 0; i < NR_MM_COUNTERS; i++) 631 if (rss[i]) 632 add_mm_counter(mm, i, rss[i]); 633 } 634 635 /* 636 * This function is called to print an error when a bad pte 637 * is found. For example, we might have a PFN-mapped pte in 638 * a region that doesn't allow it. 639 * 640 * The calling function must still handle the error. 641 */ 642 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 643 pte_t pte, struct page *page) 644 { 645 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 646 pud_t *pud = pud_offset(pgd, addr); 647 pmd_t *pmd = pmd_offset(pud, addr); 648 struct address_space *mapping; 649 pgoff_t index; 650 static unsigned long resume; 651 static unsigned long nr_shown; 652 static unsigned long nr_unshown; 653 654 /* 655 * Allow a burst of 60 reports, then keep quiet for that minute; 656 * or allow a steady drip of one report per second. 657 */ 658 if (nr_shown == 60) { 659 if (time_before(jiffies, resume)) { 660 nr_unshown++; 661 return; 662 } 663 if (nr_unshown) { 664 printk(KERN_ALERT 665 "BUG: Bad page map: %lu messages suppressed\n", 666 nr_unshown); 667 nr_unshown = 0; 668 } 669 nr_shown = 0; 670 } 671 if (nr_shown++ == 0) 672 resume = jiffies + 60 * HZ; 673 674 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 675 index = linear_page_index(vma, addr); 676 677 printk(KERN_ALERT 678 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 679 current->comm, 680 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 681 if (page) 682 dump_page(page, "bad pte"); 683 printk(KERN_ALERT 684 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", 685 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 686 /* 687 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y 688 */ 689 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n", 690 vma->vm_file, 691 vma->vm_ops ? vma->vm_ops->fault : NULL, 692 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 693 mapping ? mapping->a_ops->readpage : NULL); 694 dump_stack(); 695 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 696 } 697 698 /* 699 * vm_normal_page -- This function gets the "struct page" associated with a pte. 700 * 701 * "Special" mappings do not wish to be associated with a "struct page" (either 702 * it doesn't exist, or it exists but they don't want to touch it). In this 703 * case, NULL is returned here. "Normal" mappings do have a struct page. 704 * 705 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 706 * pte bit, in which case this function is trivial. Secondly, an architecture 707 * may not have a spare pte bit, which requires a more complicated scheme, 708 * described below. 709 * 710 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 711 * special mapping (even if there are underlying and valid "struct pages"). 712 * COWed pages of a VM_PFNMAP are always normal. 713 * 714 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 715 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 716 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 717 * mapping will always honor the rule 718 * 719 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 720 * 721 * And for normal mappings this is false. 722 * 723 * This restricts such mappings to be a linear translation from virtual address 724 * to pfn. To get around this restriction, we allow arbitrary mappings so long 725 * as the vma is not a COW mapping; in that case, we know that all ptes are 726 * special (because none can have been COWed). 727 * 728 * 729 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 730 * 731 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 732 * page" backing, however the difference is that _all_ pages with a struct 733 * page (that is, those where pfn_valid is true) are refcounted and considered 734 * normal pages by the VM. The disadvantage is that pages are refcounted 735 * (which can be slower and simply not an option for some PFNMAP users). The 736 * advantage is that we don't have to follow the strict linearity rule of 737 * PFNMAP mappings in order to support COWable mappings. 738 * 739 */ 740 #ifdef __HAVE_ARCH_PTE_SPECIAL 741 # define HAVE_PTE_SPECIAL 1 742 #else 743 # define HAVE_PTE_SPECIAL 0 744 #endif 745 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 746 pte_t pte) 747 { 748 unsigned long pfn = pte_pfn(pte); 749 750 if (HAVE_PTE_SPECIAL) { 751 if (likely(!pte_special(pte))) 752 goto check_pfn; 753 if (vma->vm_ops && vma->vm_ops->find_special_page) 754 return vma->vm_ops->find_special_page(vma, addr); 755 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 756 return NULL; 757 if (!is_zero_pfn(pfn)) 758 print_bad_pte(vma, addr, pte, NULL); 759 return NULL; 760 } 761 762 /* !HAVE_PTE_SPECIAL case follows: */ 763 764 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 765 if (vma->vm_flags & VM_MIXEDMAP) { 766 if (!pfn_valid(pfn)) 767 return NULL; 768 goto out; 769 } else { 770 unsigned long off; 771 off = (addr - vma->vm_start) >> PAGE_SHIFT; 772 if (pfn == vma->vm_pgoff + off) 773 return NULL; 774 if (!is_cow_mapping(vma->vm_flags)) 775 return NULL; 776 } 777 } 778 779 if (is_zero_pfn(pfn)) 780 return NULL; 781 check_pfn: 782 if (unlikely(pfn > highest_memmap_pfn)) { 783 print_bad_pte(vma, addr, pte, NULL); 784 return NULL; 785 } 786 787 /* 788 * NOTE! We still have PageReserved() pages in the page tables. 789 * eg. VDSO mappings can cause them to exist. 790 */ 791 out: 792 return pfn_to_page(pfn); 793 } 794 795 /* 796 * copy one vm_area from one task to the other. Assumes the page tables 797 * already present in the new task to be cleared in the whole range 798 * covered by this vma. 799 */ 800 801 static inline unsigned long 802 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 803 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 804 unsigned long addr, int *rss) 805 { 806 unsigned long vm_flags = vma->vm_flags; 807 pte_t pte = *src_pte; 808 struct page *page; 809 810 /* pte contains position in swap or file, so copy. */ 811 if (unlikely(!pte_present(pte))) { 812 swp_entry_t entry = pte_to_swp_entry(pte); 813 814 if (likely(!non_swap_entry(entry))) { 815 if (swap_duplicate(entry) < 0) 816 return entry.val; 817 818 /* make sure dst_mm is on swapoff's mmlist. */ 819 if (unlikely(list_empty(&dst_mm->mmlist))) { 820 spin_lock(&mmlist_lock); 821 if (list_empty(&dst_mm->mmlist)) 822 list_add(&dst_mm->mmlist, 823 &src_mm->mmlist); 824 spin_unlock(&mmlist_lock); 825 } 826 rss[MM_SWAPENTS]++; 827 } else if (is_migration_entry(entry)) { 828 page = migration_entry_to_page(entry); 829 830 rss[mm_counter(page)]++; 831 832 if (is_write_migration_entry(entry) && 833 is_cow_mapping(vm_flags)) { 834 /* 835 * COW mappings require pages in both 836 * parent and child to be set to read. 837 */ 838 make_migration_entry_read(&entry); 839 pte = swp_entry_to_pte(entry); 840 if (pte_swp_soft_dirty(*src_pte)) 841 pte = pte_swp_mksoft_dirty(pte); 842 set_pte_at(src_mm, addr, src_pte, pte); 843 } 844 } 845 goto out_set_pte; 846 } 847 848 /* 849 * If it's a COW mapping, write protect it both 850 * in the parent and the child 851 */ 852 if (is_cow_mapping(vm_flags)) { 853 ptep_set_wrprotect(src_mm, addr, src_pte); 854 pte = pte_wrprotect(pte); 855 } 856 857 /* 858 * If it's a shared mapping, mark it clean in 859 * the child 860 */ 861 if (vm_flags & VM_SHARED) 862 pte = pte_mkclean(pte); 863 pte = pte_mkold(pte); 864 865 page = vm_normal_page(vma, addr, pte); 866 if (page) { 867 get_page(page); 868 page_dup_rmap(page, false); 869 rss[mm_counter(page)]++; 870 } 871 872 out_set_pte: 873 set_pte_at(dst_mm, addr, dst_pte, pte); 874 return 0; 875 } 876 877 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 878 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 879 unsigned long addr, unsigned long end) 880 { 881 pte_t *orig_src_pte, *orig_dst_pte; 882 pte_t *src_pte, *dst_pte; 883 spinlock_t *src_ptl, *dst_ptl; 884 int progress = 0; 885 int rss[NR_MM_COUNTERS]; 886 swp_entry_t entry = (swp_entry_t){0}; 887 888 again: 889 init_rss_vec(rss); 890 891 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 892 if (!dst_pte) 893 return -ENOMEM; 894 src_pte = pte_offset_map(src_pmd, addr); 895 src_ptl = pte_lockptr(src_mm, src_pmd); 896 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 897 orig_src_pte = src_pte; 898 orig_dst_pte = dst_pte; 899 arch_enter_lazy_mmu_mode(); 900 901 do { 902 /* 903 * We are holding two locks at this point - either of them 904 * could generate latencies in another task on another CPU. 905 */ 906 if (progress >= 32) { 907 progress = 0; 908 if (need_resched() || 909 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 910 break; 911 } 912 if (pte_none(*src_pte)) { 913 progress++; 914 continue; 915 } 916 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, 917 vma, addr, rss); 918 if (entry.val) 919 break; 920 progress += 8; 921 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 922 923 arch_leave_lazy_mmu_mode(); 924 spin_unlock(src_ptl); 925 pte_unmap(orig_src_pte); 926 add_mm_rss_vec(dst_mm, rss); 927 pte_unmap_unlock(orig_dst_pte, dst_ptl); 928 cond_resched(); 929 930 if (entry.val) { 931 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) 932 return -ENOMEM; 933 progress = 0; 934 } 935 if (addr != end) 936 goto again; 937 return 0; 938 } 939 940 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 941 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 942 unsigned long addr, unsigned long end) 943 { 944 pmd_t *src_pmd, *dst_pmd; 945 unsigned long next; 946 947 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 948 if (!dst_pmd) 949 return -ENOMEM; 950 src_pmd = pmd_offset(src_pud, addr); 951 do { 952 next = pmd_addr_end(addr, end); 953 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) { 954 int err; 955 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE); 956 err = copy_huge_pmd(dst_mm, src_mm, 957 dst_pmd, src_pmd, addr, vma); 958 if (err == -ENOMEM) 959 return -ENOMEM; 960 if (!err) 961 continue; 962 /* fall through */ 963 } 964 if (pmd_none_or_clear_bad(src_pmd)) 965 continue; 966 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 967 vma, addr, next)) 968 return -ENOMEM; 969 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 970 return 0; 971 } 972 973 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 974 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 975 unsigned long addr, unsigned long end) 976 { 977 pud_t *src_pud, *dst_pud; 978 unsigned long next; 979 980 dst_pud = pud_alloc(dst_mm, dst_pgd, addr); 981 if (!dst_pud) 982 return -ENOMEM; 983 src_pud = pud_offset(src_pgd, addr); 984 do { 985 next = pud_addr_end(addr, end); 986 if (pud_none_or_clear_bad(src_pud)) 987 continue; 988 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 989 vma, addr, next)) 990 return -ENOMEM; 991 } while (dst_pud++, src_pud++, addr = next, addr != end); 992 return 0; 993 } 994 995 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 996 struct vm_area_struct *vma) 997 { 998 pgd_t *src_pgd, *dst_pgd; 999 unsigned long next; 1000 unsigned long addr = vma->vm_start; 1001 unsigned long end = vma->vm_end; 1002 unsigned long mmun_start; /* For mmu_notifiers */ 1003 unsigned long mmun_end; /* For mmu_notifiers */ 1004 bool is_cow; 1005 int ret; 1006 1007 /* 1008 * Don't copy ptes where a page fault will fill them correctly. 1009 * Fork becomes much lighter when there are big shared or private 1010 * readonly mappings. The tradeoff is that copy_page_range is more 1011 * efficient than faulting. 1012 */ 1013 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 1014 !vma->anon_vma) 1015 return 0; 1016 1017 if (is_vm_hugetlb_page(vma)) 1018 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 1019 1020 if (unlikely(vma->vm_flags & VM_PFNMAP)) { 1021 /* 1022 * We do not free on error cases below as remove_vma 1023 * gets called on error from higher level routine 1024 */ 1025 ret = track_pfn_copy(vma); 1026 if (ret) 1027 return ret; 1028 } 1029 1030 /* 1031 * We need to invalidate the secondary MMU mappings only when 1032 * there could be a permission downgrade on the ptes of the 1033 * parent mm. And a permission downgrade will only happen if 1034 * is_cow_mapping() returns true. 1035 */ 1036 is_cow = is_cow_mapping(vma->vm_flags); 1037 mmun_start = addr; 1038 mmun_end = end; 1039 if (is_cow) 1040 mmu_notifier_invalidate_range_start(src_mm, mmun_start, 1041 mmun_end); 1042 1043 ret = 0; 1044 dst_pgd = pgd_offset(dst_mm, addr); 1045 src_pgd = pgd_offset(src_mm, addr); 1046 do { 1047 next = pgd_addr_end(addr, end); 1048 if (pgd_none_or_clear_bad(src_pgd)) 1049 continue; 1050 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, 1051 vma, addr, next))) { 1052 ret = -ENOMEM; 1053 break; 1054 } 1055 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1056 1057 if (is_cow) 1058 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end); 1059 return ret; 1060 } 1061 1062 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1063 struct vm_area_struct *vma, pmd_t *pmd, 1064 unsigned long addr, unsigned long end, 1065 struct zap_details *details) 1066 { 1067 struct mm_struct *mm = tlb->mm; 1068 int force_flush = 0; 1069 int rss[NR_MM_COUNTERS]; 1070 spinlock_t *ptl; 1071 pte_t *start_pte; 1072 pte_t *pte; 1073 swp_entry_t entry; 1074 1075 again: 1076 init_rss_vec(rss); 1077 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1078 pte = start_pte; 1079 arch_enter_lazy_mmu_mode(); 1080 do { 1081 pte_t ptent = *pte; 1082 if (pte_none(ptent)) { 1083 continue; 1084 } 1085 1086 if (pte_present(ptent)) { 1087 struct page *page; 1088 1089 page = vm_normal_page(vma, addr, ptent); 1090 if (unlikely(details) && page) { 1091 /* 1092 * unmap_shared_mapping_pages() wants to 1093 * invalidate cache without truncating: 1094 * unmap shared but keep private pages. 1095 */ 1096 if (details->check_mapping && 1097 details->check_mapping != page->mapping) 1098 continue; 1099 } 1100 ptent = ptep_get_and_clear_full(mm, addr, pte, 1101 tlb->fullmm); 1102 tlb_remove_tlb_entry(tlb, pte, addr); 1103 if (unlikely(!page)) 1104 continue; 1105 1106 if (!PageAnon(page)) { 1107 if (pte_dirty(ptent)) { 1108 force_flush = 1; 1109 set_page_dirty(page); 1110 } 1111 if (pte_young(ptent) && 1112 likely(!(vma->vm_flags & VM_SEQ_READ))) 1113 mark_page_accessed(page); 1114 } 1115 rss[mm_counter(page)]--; 1116 page_remove_rmap(page, false); 1117 if (unlikely(page_mapcount(page) < 0)) 1118 print_bad_pte(vma, addr, ptent, page); 1119 if (unlikely(!__tlb_remove_page(tlb, page))) { 1120 force_flush = 1; 1121 addr += PAGE_SIZE; 1122 break; 1123 } 1124 continue; 1125 } 1126 /* If details->check_mapping, we leave swap entries. */ 1127 if (unlikely(details)) 1128 continue; 1129 1130 entry = pte_to_swp_entry(ptent); 1131 if (!non_swap_entry(entry)) 1132 rss[MM_SWAPENTS]--; 1133 else if (is_migration_entry(entry)) { 1134 struct page *page; 1135 1136 page = migration_entry_to_page(entry); 1137 rss[mm_counter(page)]--; 1138 } 1139 if (unlikely(!free_swap_and_cache(entry))) 1140 print_bad_pte(vma, addr, ptent, NULL); 1141 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1142 } while (pte++, addr += PAGE_SIZE, addr != end); 1143 1144 add_mm_rss_vec(mm, rss); 1145 arch_leave_lazy_mmu_mode(); 1146 1147 /* Do the actual TLB flush before dropping ptl */ 1148 if (force_flush) 1149 tlb_flush_mmu_tlbonly(tlb); 1150 pte_unmap_unlock(start_pte, ptl); 1151 1152 /* 1153 * If we forced a TLB flush (either due to running out of 1154 * batch buffers or because we needed to flush dirty TLB 1155 * entries before releasing the ptl), free the batched 1156 * memory too. Restart if we didn't do everything. 1157 */ 1158 if (force_flush) { 1159 force_flush = 0; 1160 tlb_flush_mmu_free(tlb); 1161 1162 if (addr != end) 1163 goto again; 1164 } 1165 1166 return addr; 1167 } 1168 1169 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1170 struct vm_area_struct *vma, pud_t *pud, 1171 unsigned long addr, unsigned long end, 1172 struct zap_details *details) 1173 { 1174 pmd_t *pmd; 1175 unsigned long next; 1176 1177 pmd = pmd_offset(pud, addr); 1178 do { 1179 next = pmd_addr_end(addr, end); 1180 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1181 if (next - addr != HPAGE_PMD_SIZE) { 1182 #ifdef CONFIG_DEBUG_VM 1183 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) { 1184 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n", 1185 __func__, addr, end, 1186 vma->vm_start, 1187 vma->vm_end); 1188 BUG(); 1189 } 1190 #endif 1191 split_huge_pmd(vma, pmd, addr); 1192 } else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1193 goto next; 1194 /* fall through */ 1195 } 1196 /* 1197 * Here there can be other concurrent MADV_DONTNEED or 1198 * trans huge page faults running, and if the pmd is 1199 * none or trans huge it can change under us. This is 1200 * because MADV_DONTNEED holds the mmap_sem in read 1201 * mode. 1202 */ 1203 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1204 goto next; 1205 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1206 next: 1207 cond_resched(); 1208 } while (pmd++, addr = next, addr != end); 1209 1210 return addr; 1211 } 1212 1213 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1214 struct vm_area_struct *vma, pgd_t *pgd, 1215 unsigned long addr, unsigned long end, 1216 struct zap_details *details) 1217 { 1218 pud_t *pud; 1219 unsigned long next; 1220 1221 pud = pud_offset(pgd, addr); 1222 do { 1223 next = pud_addr_end(addr, end); 1224 if (pud_none_or_clear_bad(pud)) 1225 continue; 1226 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1227 } while (pud++, addr = next, addr != end); 1228 1229 return addr; 1230 } 1231 1232 static void unmap_page_range(struct mmu_gather *tlb, 1233 struct vm_area_struct *vma, 1234 unsigned long addr, unsigned long end, 1235 struct zap_details *details) 1236 { 1237 pgd_t *pgd; 1238 unsigned long next; 1239 1240 if (details && !details->check_mapping) 1241 details = NULL; 1242 1243 BUG_ON(addr >= end); 1244 tlb_start_vma(tlb, vma); 1245 pgd = pgd_offset(vma->vm_mm, addr); 1246 do { 1247 next = pgd_addr_end(addr, end); 1248 if (pgd_none_or_clear_bad(pgd)) 1249 continue; 1250 next = zap_pud_range(tlb, vma, pgd, addr, next, details); 1251 } while (pgd++, addr = next, addr != end); 1252 tlb_end_vma(tlb, vma); 1253 } 1254 1255 1256 static void unmap_single_vma(struct mmu_gather *tlb, 1257 struct vm_area_struct *vma, unsigned long start_addr, 1258 unsigned long end_addr, 1259 struct zap_details *details) 1260 { 1261 unsigned long start = max(vma->vm_start, start_addr); 1262 unsigned long end; 1263 1264 if (start >= vma->vm_end) 1265 return; 1266 end = min(vma->vm_end, end_addr); 1267 if (end <= vma->vm_start) 1268 return; 1269 1270 if (vma->vm_file) 1271 uprobe_munmap(vma, start, end); 1272 1273 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1274 untrack_pfn(vma, 0, 0); 1275 1276 if (start != end) { 1277 if (unlikely(is_vm_hugetlb_page(vma))) { 1278 /* 1279 * It is undesirable to test vma->vm_file as it 1280 * should be non-null for valid hugetlb area. 1281 * However, vm_file will be NULL in the error 1282 * cleanup path of mmap_region. When 1283 * hugetlbfs ->mmap method fails, 1284 * mmap_region() nullifies vma->vm_file 1285 * before calling this function to clean up. 1286 * Since no pte has actually been setup, it is 1287 * safe to do nothing in this case. 1288 */ 1289 if (vma->vm_file) { 1290 i_mmap_lock_write(vma->vm_file->f_mapping); 1291 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1292 i_mmap_unlock_write(vma->vm_file->f_mapping); 1293 } 1294 } else 1295 unmap_page_range(tlb, vma, start, end, details); 1296 } 1297 } 1298 1299 /** 1300 * unmap_vmas - unmap a range of memory covered by a list of vma's 1301 * @tlb: address of the caller's struct mmu_gather 1302 * @vma: the starting vma 1303 * @start_addr: virtual address at which to start unmapping 1304 * @end_addr: virtual address at which to end unmapping 1305 * 1306 * Unmap all pages in the vma list. 1307 * 1308 * Only addresses between `start' and `end' will be unmapped. 1309 * 1310 * The VMA list must be sorted in ascending virtual address order. 1311 * 1312 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1313 * range after unmap_vmas() returns. So the only responsibility here is to 1314 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1315 * drops the lock and schedules. 1316 */ 1317 void unmap_vmas(struct mmu_gather *tlb, 1318 struct vm_area_struct *vma, unsigned long start_addr, 1319 unsigned long end_addr) 1320 { 1321 struct mm_struct *mm = vma->vm_mm; 1322 1323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr); 1324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1325 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1326 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr); 1327 } 1328 1329 /** 1330 * zap_page_range - remove user pages in a given range 1331 * @vma: vm_area_struct holding the applicable pages 1332 * @start: starting address of pages to zap 1333 * @size: number of bytes to zap 1334 * @details: details of shared cache invalidation 1335 * 1336 * Caller must protect the VMA list 1337 */ 1338 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1339 unsigned long size, struct zap_details *details) 1340 { 1341 struct mm_struct *mm = vma->vm_mm; 1342 struct mmu_gather tlb; 1343 unsigned long end = start + size; 1344 1345 lru_add_drain(); 1346 tlb_gather_mmu(&tlb, mm, start, end); 1347 update_hiwater_rss(mm); 1348 mmu_notifier_invalidate_range_start(mm, start, end); 1349 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) 1350 unmap_single_vma(&tlb, vma, start, end, details); 1351 mmu_notifier_invalidate_range_end(mm, start, end); 1352 tlb_finish_mmu(&tlb, start, end); 1353 } 1354 1355 /** 1356 * zap_page_range_single - remove user pages in a given range 1357 * @vma: vm_area_struct holding the applicable pages 1358 * @address: starting address of pages to zap 1359 * @size: number of bytes to zap 1360 * @details: details of shared cache invalidation 1361 * 1362 * The range must fit into one VMA. 1363 */ 1364 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1365 unsigned long size, struct zap_details *details) 1366 { 1367 struct mm_struct *mm = vma->vm_mm; 1368 struct mmu_gather tlb; 1369 unsigned long end = address + size; 1370 1371 lru_add_drain(); 1372 tlb_gather_mmu(&tlb, mm, address, end); 1373 update_hiwater_rss(mm); 1374 mmu_notifier_invalidate_range_start(mm, address, end); 1375 unmap_single_vma(&tlb, vma, address, end, details); 1376 mmu_notifier_invalidate_range_end(mm, address, end); 1377 tlb_finish_mmu(&tlb, address, end); 1378 } 1379 1380 /** 1381 * zap_vma_ptes - remove ptes mapping the vma 1382 * @vma: vm_area_struct holding ptes to be zapped 1383 * @address: starting address of pages to zap 1384 * @size: number of bytes to zap 1385 * 1386 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1387 * 1388 * The entire address range must be fully contained within the vma. 1389 * 1390 * Returns 0 if successful. 1391 */ 1392 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1393 unsigned long size) 1394 { 1395 if (address < vma->vm_start || address + size > vma->vm_end || 1396 !(vma->vm_flags & VM_PFNMAP)) 1397 return -1; 1398 zap_page_range_single(vma, address, size, NULL); 1399 return 0; 1400 } 1401 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1402 1403 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1404 spinlock_t **ptl) 1405 { 1406 pgd_t * pgd = pgd_offset(mm, addr); 1407 pud_t * pud = pud_alloc(mm, pgd, addr); 1408 if (pud) { 1409 pmd_t * pmd = pmd_alloc(mm, pud, addr); 1410 if (pmd) { 1411 VM_BUG_ON(pmd_trans_huge(*pmd)); 1412 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1413 } 1414 } 1415 return NULL; 1416 } 1417 1418 /* 1419 * This is the old fallback for page remapping. 1420 * 1421 * For historical reasons, it only allows reserved pages. Only 1422 * old drivers should use this, and they needed to mark their 1423 * pages reserved for the old functions anyway. 1424 */ 1425 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1426 struct page *page, pgprot_t prot) 1427 { 1428 struct mm_struct *mm = vma->vm_mm; 1429 int retval; 1430 pte_t *pte; 1431 spinlock_t *ptl; 1432 1433 retval = -EINVAL; 1434 if (PageAnon(page)) 1435 goto out; 1436 retval = -ENOMEM; 1437 flush_dcache_page(page); 1438 pte = get_locked_pte(mm, addr, &ptl); 1439 if (!pte) 1440 goto out; 1441 retval = -EBUSY; 1442 if (!pte_none(*pte)) 1443 goto out_unlock; 1444 1445 /* Ok, finally just insert the thing.. */ 1446 get_page(page); 1447 inc_mm_counter_fast(mm, mm_counter_file(page)); 1448 page_add_file_rmap(page); 1449 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1450 1451 retval = 0; 1452 pte_unmap_unlock(pte, ptl); 1453 return retval; 1454 out_unlock: 1455 pte_unmap_unlock(pte, ptl); 1456 out: 1457 return retval; 1458 } 1459 1460 /** 1461 * vm_insert_page - insert single page into user vma 1462 * @vma: user vma to map to 1463 * @addr: target user address of this page 1464 * @page: source kernel page 1465 * 1466 * This allows drivers to insert individual pages they've allocated 1467 * into a user vma. 1468 * 1469 * The page has to be a nice clean _individual_ kernel allocation. 1470 * If you allocate a compound page, you need to have marked it as 1471 * such (__GFP_COMP), or manually just split the page up yourself 1472 * (see split_page()). 1473 * 1474 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1475 * took an arbitrary page protection parameter. This doesn't allow 1476 * that. Your vma protection will have to be set up correctly, which 1477 * means that if you want a shared writable mapping, you'd better 1478 * ask for a shared writable mapping! 1479 * 1480 * The page does not need to be reserved. 1481 * 1482 * Usually this function is called from f_op->mmap() handler 1483 * under mm->mmap_sem write-lock, so it can change vma->vm_flags. 1484 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1485 * function from other places, for example from page-fault handler. 1486 */ 1487 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1488 struct page *page) 1489 { 1490 if (addr < vma->vm_start || addr >= vma->vm_end) 1491 return -EFAULT; 1492 if (!page_count(page)) 1493 return -EINVAL; 1494 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1495 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem)); 1496 BUG_ON(vma->vm_flags & VM_PFNMAP); 1497 vma->vm_flags |= VM_MIXEDMAP; 1498 } 1499 return insert_page(vma, addr, page, vma->vm_page_prot); 1500 } 1501 EXPORT_SYMBOL(vm_insert_page); 1502 1503 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1504 pfn_t pfn, pgprot_t prot) 1505 { 1506 struct mm_struct *mm = vma->vm_mm; 1507 int retval; 1508 pte_t *pte, entry; 1509 spinlock_t *ptl; 1510 1511 retval = -ENOMEM; 1512 pte = get_locked_pte(mm, addr, &ptl); 1513 if (!pte) 1514 goto out; 1515 retval = -EBUSY; 1516 if (!pte_none(*pte)) 1517 goto out_unlock; 1518 1519 /* Ok, finally just insert the thing.. */ 1520 if (pfn_t_devmap(pfn)) 1521 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 1522 else 1523 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 1524 set_pte_at(mm, addr, pte, entry); 1525 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 1526 1527 retval = 0; 1528 out_unlock: 1529 pte_unmap_unlock(pte, ptl); 1530 out: 1531 return retval; 1532 } 1533 1534 /** 1535 * vm_insert_pfn - insert single pfn into user vma 1536 * @vma: user vma to map to 1537 * @addr: target user address of this page 1538 * @pfn: source kernel pfn 1539 * 1540 * Similar to vm_insert_page, this allows drivers to insert individual pages 1541 * they've allocated into a user vma. Same comments apply. 1542 * 1543 * This function should only be called from a vm_ops->fault handler, and 1544 * in that case the handler should return NULL. 1545 * 1546 * vma cannot be a COW mapping. 1547 * 1548 * As this is called only for pages that do not currently exist, we 1549 * do not need to flush old virtual caches or the TLB. 1550 */ 1551 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1552 unsigned long pfn) 1553 { 1554 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 1555 } 1556 EXPORT_SYMBOL(vm_insert_pfn); 1557 1558 /** 1559 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot 1560 * @vma: user vma to map to 1561 * @addr: target user address of this page 1562 * @pfn: source kernel pfn 1563 * @pgprot: pgprot flags for the inserted page 1564 * 1565 * This is exactly like vm_insert_pfn, except that it allows drivers to 1566 * to override pgprot on a per-page basis. 1567 * 1568 * This only makes sense for IO mappings, and it makes no sense for 1569 * cow mappings. In general, using multiple vmas is preferable; 1570 * vm_insert_pfn_prot should only be used if using multiple VMAs is 1571 * impractical. 1572 */ 1573 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 1574 unsigned long pfn, pgprot_t pgprot) 1575 { 1576 int ret; 1577 /* 1578 * Technically, architectures with pte_special can avoid all these 1579 * restrictions (same for remap_pfn_range). However we would like 1580 * consistency in testing and feature parity among all, so we should 1581 * try to keep these invariants in place for everybody. 1582 */ 1583 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 1584 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1585 (VM_PFNMAP|VM_MIXEDMAP)); 1586 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1587 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 1588 1589 if (addr < vma->vm_start || addr >= vma->vm_end) 1590 return -EFAULT; 1591 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV))) 1592 return -EINVAL; 1593 1594 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot); 1595 1596 return ret; 1597 } 1598 EXPORT_SYMBOL(vm_insert_pfn_prot); 1599 1600 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 1601 pfn_t pfn) 1602 { 1603 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP)); 1604 1605 if (addr < vma->vm_start || addr >= vma->vm_end) 1606 return -EFAULT; 1607 1608 /* 1609 * If we don't have pte special, then we have to use the pfn_valid() 1610 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 1611 * refcount the page if pfn_valid is true (hence insert_page rather 1612 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 1613 * without pte special, it would there be refcounted as a normal page. 1614 */ 1615 if (!HAVE_PTE_SPECIAL && pfn_t_valid(pfn)) { 1616 struct page *page; 1617 1618 page = pfn_t_to_page(pfn); 1619 return insert_page(vma, addr, page, vma->vm_page_prot); 1620 } 1621 return insert_pfn(vma, addr, pfn, vma->vm_page_prot); 1622 } 1623 EXPORT_SYMBOL(vm_insert_mixed); 1624 1625 /* 1626 * maps a range of physical memory into the requested pages. the old 1627 * mappings are removed. any references to nonexistent pages results 1628 * in null mappings (currently treated as "copy-on-access") 1629 */ 1630 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1631 unsigned long addr, unsigned long end, 1632 unsigned long pfn, pgprot_t prot) 1633 { 1634 pte_t *pte; 1635 spinlock_t *ptl; 1636 1637 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1638 if (!pte) 1639 return -ENOMEM; 1640 arch_enter_lazy_mmu_mode(); 1641 do { 1642 BUG_ON(!pte_none(*pte)); 1643 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 1644 pfn++; 1645 } while (pte++, addr += PAGE_SIZE, addr != end); 1646 arch_leave_lazy_mmu_mode(); 1647 pte_unmap_unlock(pte - 1, ptl); 1648 return 0; 1649 } 1650 1651 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 1652 unsigned long addr, unsigned long end, 1653 unsigned long pfn, pgprot_t prot) 1654 { 1655 pmd_t *pmd; 1656 unsigned long next; 1657 1658 pfn -= addr >> PAGE_SHIFT; 1659 pmd = pmd_alloc(mm, pud, addr); 1660 if (!pmd) 1661 return -ENOMEM; 1662 VM_BUG_ON(pmd_trans_huge(*pmd)); 1663 do { 1664 next = pmd_addr_end(addr, end); 1665 if (remap_pte_range(mm, pmd, addr, next, 1666 pfn + (addr >> PAGE_SHIFT), prot)) 1667 return -ENOMEM; 1668 } while (pmd++, addr = next, addr != end); 1669 return 0; 1670 } 1671 1672 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, 1673 unsigned long addr, unsigned long end, 1674 unsigned long pfn, pgprot_t prot) 1675 { 1676 pud_t *pud; 1677 unsigned long next; 1678 1679 pfn -= addr >> PAGE_SHIFT; 1680 pud = pud_alloc(mm, pgd, addr); 1681 if (!pud) 1682 return -ENOMEM; 1683 do { 1684 next = pud_addr_end(addr, end); 1685 if (remap_pmd_range(mm, pud, addr, next, 1686 pfn + (addr >> PAGE_SHIFT), prot)) 1687 return -ENOMEM; 1688 } while (pud++, addr = next, addr != end); 1689 return 0; 1690 } 1691 1692 /** 1693 * remap_pfn_range - remap kernel memory to userspace 1694 * @vma: user vma to map to 1695 * @addr: target user address to start at 1696 * @pfn: physical address of kernel memory 1697 * @size: size of map area 1698 * @prot: page protection flags for this mapping 1699 * 1700 * Note: this is only safe if the mm semaphore is held when called. 1701 */ 1702 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 1703 unsigned long pfn, unsigned long size, pgprot_t prot) 1704 { 1705 pgd_t *pgd; 1706 unsigned long next; 1707 unsigned long end = addr + PAGE_ALIGN(size); 1708 struct mm_struct *mm = vma->vm_mm; 1709 int err; 1710 1711 /* 1712 * Physically remapped pages are special. Tell the 1713 * rest of the world about it: 1714 * VM_IO tells people not to look at these pages 1715 * (accesses can have side effects). 1716 * VM_PFNMAP tells the core MM that the base pages are just 1717 * raw PFN mappings, and do not have a "struct page" associated 1718 * with them. 1719 * VM_DONTEXPAND 1720 * Disable vma merging and expanding with mremap(). 1721 * VM_DONTDUMP 1722 * Omit vma from core dump, even when VM_IO turned off. 1723 * 1724 * There's a horrible special case to handle copy-on-write 1725 * behaviour that some programs depend on. We mark the "original" 1726 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 1727 * See vm_normal_page() for details. 1728 */ 1729 if (is_cow_mapping(vma->vm_flags)) { 1730 if (addr != vma->vm_start || end != vma->vm_end) 1731 return -EINVAL; 1732 vma->vm_pgoff = pfn; 1733 } 1734 1735 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 1736 if (err) 1737 return -EINVAL; 1738 1739 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 1740 1741 BUG_ON(addr >= end); 1742 pfn -= addr >> PAGE_SHIFT; 1743 pgd = pgd_offset(mm, addr); 1744 flush_cache_range(vma, addr, end); 1745 do { 1746 next = pgd_addr_end(addr, end); 1747 err = remap_pud_range(mm, pgd, addr, next, 1748 pfn + (addr >> PAGE_SHIFT), prot); 1749 if (err) 1750 break; 1751 } while (pgd++, addr = next, addr != end); 1752 1753 if (err) 1754 untrack_pfn(vma, pfn, PAGE_ALIGN(size)); 1755 1756 return err; 1757 } 1758 EXPORT_SYMBOL(remap_pfn_range); 1759 1760 /** 1761 * vm_iomap_memory - remap memory to userspace 1762 * @vma: user vma to map to 1763 * @start: start of area 1764 * @len: size of area 1765 * 1766 * This is a simplified io_remap_pfn_range() for common driver use. The 1767 * driver just needs to give us the physical memory range to be mapped, 1768 * we'll figure out the rest from the vma information. 1769 * 1770 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 1771 * whatever write-combining details or similar. 1772 */ 1773 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 1774 { 1775 unsigned long vm_len, pfn, pages; 1776 1777 /* Check that the physical memory area passed in looks valid */ 1778 if (start + len < start) 1779 return -EINVAL; 1780 /* 1781 * You *really* shouldn't map things that aren't page-aligned, 1782 * but we've historically allowed it because IO memory might 1783 * just have smaller alignment. 1784 */ 1785 len += start & ~PAGE_MASK; 1786 pfn = start >> PAGE_SHIFT; 1787 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 1788 if (pfn + pages < pfn) 1789 return -EINVAL; 1790 1791 /* We start the mapping 'vm_pgoff' pages into the area */ 1792 if (vma->vm_pgoff > pages) 1793 return -EINVAL; 1794 pfn += vma->vm_pgoff; 1795 pages -= vma->vm_pgoff; 1796 1797 /* Can we fit all of the mapping? */ 1798 vm_len = vma->vm_end - vma->vm_start; 1799 if (vm_len >> PAGE_SHIFT > pages) 1800 return -EINVAL; 1801 1802 /* Ok, let it rip */ 1803 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 1804 } 1805 EXPORT_SYMBOL(vm_iomap_memory); 1806 1807 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 1808 unsigned long addr, unsigned long end, 1809 pte_fn_t fn, void *data) 1810 { 1811 pte_t *pte; 1812 int err; 1813 pgtable_t token; 1814 spinlock_t *uninitialized_var(ptl); 1815 1816 pte = (mm == &init_mm) ? 1817 pte_alloc_kernel(pmd, addr) : 1818 pte_alloc_map_lock(mm, pmd, addr, &ptl); 1819 if (!pte) 1820 return -ENOMEM; 1821 1822 BUG_ON(pmd_huge(*pmd)); 1823 1824 arch_enter_lazy_mmu_mode(); 1825 1826 token = pmd_pgtable(*pmd); 1827 1828 do { 1829 err = fn(pte++, token, addr, data); 1830 if (err) 1831 break; 1832 } while (addr += PAGE_SIZE, addr != end); 1833 1834 arch_leave_lazy_mmu_mode(); 1835 1836 if (mm != &init_mm) 1837 pte_unmap_unlock(pte-1, ptl); 1838 return err; 1839 } 1840 1841 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 1842 unsigned long addr, unsigned long end, 1843 pte_fn_t fn, void *data) 1844 { 1845 pmd_t *pmd; 1846 unsigned long next; 1847 int err; 1848 1849 BUG_ON(pud_huge(*pud)); 1850 1851 pmd = pmd_alloc(mm, pud, addr); 1852 if (!pmd) 1853 return -ENOMEM; 1854 do { 1855 next = pmd_addr_end(addr, end); 1856 err = apply_to_pte_range(mm, pmd, addr, next, fn, data); 1857 if (err) 1858 break; 1859 } while (pmd++, addr = next, addr != end); 1860 return err; 1861 } 1862 1863 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, 1864 unsigned long addr, unsigned long end, 1865 pte_fn_t fn, void *data) 1866 { 1867 pud_t *pud; 1868 unsigned long next; 1869 int err; 1870 1871 pud = pud_alloc(mm, pgd, addr); 1872 if (!pud) 1873 return -ENOMEM; 1874 do { 1875 next = pud_addr_end(addr, end); 1876 err = apply_to_pmd_range(mm, pud, addr, next, fn, data); 1877 if (err) 1878 break; 1879 } while (pud++, addr = next, addr != end); 1880 return err; 1881 } 1882 1883 /* 1884 * Scan a region of virtual memory, filling in page tables as necessary 1885 * and calling a provided function on each leaf page table. 1886 */ 1887 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 1888 unsigned long size, pte_fn_t fn, void *data) 1889 { 1890 pgd_t *pgd; 1891 unsigned long next; 1892 unsigned long end = addr + size; 1893 int err; 1894 1895 BUG_ON(addr >= end); 1896 pgd = pgd_offset(mm, addr); 1897 do { 1898 next = pgd_addr_end(addr, end); 1899 err = apply_to_pud_range(mm, pgd, addr, next, fn, data); 1900 if (err) 1901 break; 1902 } while (pgd++, addr = next, addr != end); 1903 1904 return err; 1905 } 1906 EXPORT_SYMBOL_GPL(apply_to_page_range); 1907 1908 /* 1909 * handle_pte_fault chooses page fault handler according to an entry which was 1910 * read non-atomically. Before making any commitment, on those architectures 1911 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 1912 * parts, do_swap_page must check under lock before unmapping the pte and 1913 * proceeding (but do_wp_page is only called after already making such a check; 1914 * and do_anonymous_page can safely check later on). 1915 */ 1916 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 1917 pte_t *page_table, pte_t orig_pte) 1918 { 1919 int same = 1; 1920 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 1921 if (sizeof(pte_t) > sizeof(unsigned long)) { 1922 spinlock_t *ptl = pte_lockptr(mm, pmd); 1923 spin_lock(ptl); 1924 same = pte_same(*page_table, orig_pte); 1925 spin_unlock(ptl); 1926 } 1927 #endif 1928 pte_unmap(page_table); 1929 return same; 1930 } 1931 1932 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) 1933 { 1934 debug_dma_assert_idle(src); 1935 1936 /* 1937 * If the source page was a PFN mapping, we don't have 1938 * a "struct page" for it. We do a best-effort copy by 1939 * just copying from the original user address. If that 1940 * fails, we just zero-fill it. Live with it. 1941 */ 1942 if (unlikely(!src)) { 1943 void *kaddr = kmap_atomic(dst); 1944 void __user *uaddr = (void __user *)(va & PAGE_MASK); 1945 1946 /* 1947 * This really shouldn't fail, because the page is there 1948 * in the page tables. But it might just be unreadable, 1949 * in which case we just give up and fill the result with 1950 * zeroes. 1951 */ 1952 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) 1953 clear_page(kaddr); 1954 kunmap_atomic(kaddr); 1955 flush_dcache_page(dst); 1956 } else 1957 copy_user_highpage(dst, src, va, vma); 1958 } 1959 1960 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 1961 { 1962 struct file *vm_file = vma->vm_file; 1963 1964 if (vm_file) 1965 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 1966 1967 /* 1968 * Special mappings (e.g. VDSO) do not have any file so fake 1969 * a default GFP_KERNEL for them. 1970 */ 1971 return GFP_KERNEL; 1972 } 1973 1974 /* 1975 * Notify the address space that the page is about to become writable so that 1976 * it can prohibit this or wait for the page to get into an appropriate state. 1977 * 1978 * We do this without the lock held, so that it can sleep if it needs to. 1979 */ 1980 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page, 1981 unsigned long address) 1982 { 1983 struct vm_fault vmf; 1984 int ret; 1985 1986 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 1987 vmf.pgoff = page->index; 1988 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 1989 vmf.gfp_mask = __get_fault_gfp_mask(vma); 1990 vmf.page = page; 1991 vmf.cow_page = NULL; 1992 1993 ret = vma->vm_ops->page_mkwrite(vma, &vmf); 1994 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 1995 return ret; 1996 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 1997 lock_page(page); 1998 if (!page->mapping) { 1999 unlock_page(page); 2000 return 0; /* retry */ 2001 } 2002 ret |= VM_FAULT_LOCKED; 2003 } else 2004 VM_BUG_ON_PAGE(!PageLocked(page), page); 2005 return ret; 2006 } 2007 2008 /* 2009 * Handle write page faults for pages that can be reused in the current vma 2010 * 2011 * This can happen either due to the mapping being with the VM_SHARED flag, 2012 * or due to us being the last reference standing to the page. In either 2013 * case, all we need to do here is to mark the page as writable and update 2014 * any related book-keeping. 2015 */ 2016 static inline int wp_page_reuse(struct mm_struct *mm, 2017 struct vm_area_struct *vma, unsigned long address, 2018 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte, 2019 struct page *page, int page_mkwrite, 2020 int dirty_shared) 2021 __releases(ptl) 2022 { 2023 pte_t entry; 2024 /* 2025 * Clear the pages cpupid information as the existing 2026 * information potentially belongs to a now completely 2027 * unrelated process. 2028 */ 2029 if (page) 2030 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2031 2032 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2033 entry = pte_mkyoung(orig_pte); 2034 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2035 if (ptep_set_access_flags(vma, address, page_table, entry, 1)) 2036 update_mmu_cache(vma, address, page_table); 2037 pte_unmap_unlock(page_table, ptl); 2038 2039 if (dirty_shared) { 2040 struct address_space *mapping; 2041 int dirtied; 2042 2043 if (!page_mkwrite) 2044 lock_page(page); 2045 2046 dirtied = set_page_dirty(page); 2047 VM_BUG_ON_PAGE(PageAnon(page), page); 2048 mapping = page->mapping; 2049 unlock_page(page); 2050 page_cache_release(page); 2051 2052 if ((dirtied || page_mkwrite) && mapping) { 2053 /* 2054 * Some device drivers do not set page.mapping 2055 * but still dirty their pages 2056 */ 2057 balance_dirty_pages_ratelimited(mapping); 2058 } 2059 2060 if (!page_mkwrite) 2061 file_update_time(vma->vm_file); 2062 } 2063 2064 return VM_FAULT_WRITE; 2065 } 2066 2067 /* 2068 * Handle the case of a page which we actually need to copy to a new page. 2069 * 2070 * Called with mmap_sem locked and the old page referenced, but 2071 * without the ptl held. 2072 * 2073 * High level logic flow: 2074 * 2075 * - Allocate a page, copy the content of the old page to the new one. 2076 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2077 * - Take the PTL. If the pte changed, bail out and release the allocated page 2078 * - If the pte is still the way we remember it, update the page table and all 2079 * relevant references. This includes dropping the reference the page-table 2080 * held to the old page, as well as updating the rmap. 2081 * - In any case, unlock the PTL and drop the reference we took to the old page. 2082 */ 2083 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma, 2084 unsigned long address, pte_t *page_table, pmd_t *pmd, 2085 pte_t orig_pte, struct page *old_page) 2086 { 2087 struct page *new_page = NULL; 2088 spinlock_t *ptl = NULL; 2089 pte_t entry; 2090 int page_copied = 0; 2091 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */ 2092 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */ 2093 struct mem_cgroup *memcg; 2094 2095 if (unlikely(anon_vma_prepare(vma))) 2096 goto oom; 2097 2098 if (is_zero_pfn(pte_pfn(orig_pte))) { 2099 new_page = alloc_zeroed_user_highpage_movable(vma, address); 2100 if (!new_page) 2101 goto oom; 2102 } else { 2103 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 2104 if (!new_page) 2105 goto oom; 2106 cow_user_page(new_page, old_page, address, vma); 2107 } 2108 2109 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) 2110 goto oom_free_new; 2111 2112 __SetPageUptodate(new_page); 2113 2114 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2115 2116 /* 2117 * Re-check the pte - we dropped the lock 2118 */ 2119 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2120 if (likely(pte_same(*page_table, orig_pte))) { 2121 if (old_page) { 2122 if (!PageAnon(old_page)) { 2123 dec_mm_counter_fast(mm, 2124 mm_counter_file(old_page)); 2125 inc_mm_counter_fast(mm, MM_ANONPAGES); 2126 } 2127 } else { 2128 inc_mm_counter_fast(mm, MM_ANONPAGES); 2129 } 2130 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2131 entry = mk_pte(new_page, vma->vm_page_prot); 2132 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2133 /* 2134 * Clear the pte entry and flush it first, before updating the 2135 * pte with the new entry. This will avoid a race condition 2136 * seen in the presence of one thread doing SMC and another 2137 * thread doing COW. 2138 */ 2139 ptep_clear_flush_notify(vma, address, page_table); 2140 page_add_new_anon_rmap(new_page, vma, address, false); 2141 mem_cgroup_commit_charge(new_page, memcg, false, false); 2142 lru_cache_add_active_or_unevictable(new_page, vma); 2143 /* 2144 * We call the notify macro here because, when using secondary 2145 * mmu page tables (such as kvm shadow page tables), we want the 2146 * new page to be mapped directly into the secondary page table. 2147 */ 2148 set_pte_at_notify(mm, address, page_table, entry); 2149 update_mmu_cache(vma, address, page_table); 2150 if (old_page) { 2151 /* 2152 * Only after switching the pte to the new page may 2153 * we remove the mapcount here. Otherwise another 2154 * process may come and find the rmap count decremented 2155 * before the pte is switched to the new page, and 2156 * "reuse" the old page writing into it while our pte 2157 * here still points into it and can be read by other 2158 * threads. 2159 * 2160 * The critical issue is to order this 2161 * page_remove_rmap with the ptp_clear_flush above. 2162 * Those stores are ordered by (if nothing else,) 2163 * the barrier present in the atomic_add_negative 2164 * in page_remove_rmap. 2165 * 2166 * Then the TLB flush in ptep_clear_flush ensures that 2167 * no process can access the old page before the 2168 * decremented mapcount is visible. And the old page 2169 * cannot be reused until after the decremented 2170 * mapcount is visible. So transitively, TLBs to 2171 * old page will be flushed before it can be reused. 2172 */ 2173 page_remove_rmap(old_page, false); 2174 } 2175 2176 /* Free the old page.. */ 2177 new_page = old_page; 2178 page_copied = 1; 2179 } else { 2180 mem_cgroup_cancel_charge(new_page, memcg, false); 2181 } 2182 2183 if (new_page) 2184 page_cache_release(new_page); 2185 2186 pte_unmap_unlock(page_table, ptl); 2187 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2188 if (old_page) { 2189 /* 2190 * Don't let another task, with possibly unlocked vma, 2191 * keep the mlocked page. 2192 */ 2193 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 2194 lock_page(old_page); /* LRU manipulation */ 2195 if (PageMlocked(old_page)) 2196 munlock_vma_page(old_page); 2197 unlock_page(old_page); 2198 } 2199 page_cache_release(old_page); 2200 } 2201 return page_copied ? VM_FAULT_WRITE : 0; 2202 oom_free_new: 2203 page_cache_release(new_page); 2204 oom: 2205 if (old_page) 2206 page_cache_release(old_page); 2207 return VM_FAULT_OOM; 2208 } 2209 2210 /* 2211 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 2212 * mapping 2213 */ 2214 static int wp_pfn_shared(struct mm_struct *mm, 2215 struct vm_area_struct *vma, unsigned long address, 2216 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte, 2217 pmd_t *pmd) 2218 { 2219 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 2220 struct vm_fault vmf = { 2221 .page = NULL, 2222 .pgoff = linear_page_index(vma, address), 2223 .virtual_address = (void __user *)(address & PAGE_MASK), 2224 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE, 2225 }; 2226 int ret; 2227 2228 pte_unmap_unlock(page_table, ptl); 2229 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf); 2230 if (ret & VM_FAULT_ERROR) 2231 return ret; 2232 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2233 /* 2234 * We might have raced with another page fault while we 2235 * released the pte_offset_map_lock. 2236 */ 2237 if (!pte_same(*page_table, orig_pte)) { 2238 pte_unmap_unlock(page_table, ptl); 2239 return 0; 2240 } 2241 } 2242 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte, 2243 NULL, 0, 0); 2244 } 2245 2246 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma, 2247 unsigned long address, pte_t *page_table, 2248 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte, 2249 struct page *old_page) 2250 __releases(ptl) 2251 { 2252 int page_mkwrite = 0; 2253 2254 page_cache_get(old_page); 2255 2256 /* 2257 * Only catch write-faults on shared writable pages, 2258 * read-only shared pages can get COWed by 2259 * get_user_pages(.write=1, .force=1). 2260 */ 2261 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 2262 int tmp; 2263 2264 pte_unmap_unlock(page_table, ptl); 2265 tmp = do_page_mkwrite(vma, old_page, address); 2266 if (unlikely(!tmp || (tmp & 2267 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 2268 page_cache_release(old_page); 2269 return tmp; 2270 } 2271 /* 2272 * Since we dropped the lock we need to revalidate 2273 * the PTE as someone else may have changed it. If 2274 * they did, we just return, as we can count on the 2275 * MMU to tell us if they didn't also make it writable. 2276 */ 2277 page_table = pte_offset_map_lock(mm, pmd, address, 2278 &ptl); 2279 if (!pte_same(*page_table, orig_pte)) { 2280 unlock_page(old_page); 2281 pte_unmap_unlock(page_table, ptl); 2282 page_cache_release(old_page); 2283 return 0; 2284 } 2285 page_mkwrite = 1; 2286 } 2287 2288 return wp_page_reuse(mm, vma, address, page_table, ptl, 2289 orig_pte, old_page, page_mkwrite, 1); 2290 } 2291 2292 /* 2293 * This routine handles present pages, when users try to write 2294 * to a shared page. It is done by copying the page to a new address 2295 * and decrementing the shared-page counter for the old page. 2296 * 2297 * Note that this routine assumes that the protection checks have been 2298 * done by the caller (the low-level page fault routine in most cases). 2299 * Thus we can safely just mark it writable once we've done any necessary 2300 * COW. 2301 * 2302 * We also mark the page dirty at this point even though the page will 2303 * change only once the write actually happens. This avoids a few races, 2304 * and potentially makes it more efficient. 2305 * 2306 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2307 * but allow concurrent faults), with pte both mapped and locked. 2308 * We return with mmap_sem still held, but pte unmapped and unlocked. 2309 */ 2310 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 2311 unsigned long address, pte_t *page_table, pmd_t *pmd, 2312 spinlock_t *ptl, pte_t orig_pte) 2313 __releases(ptl) 2314 { 2315 struct page *old_page; 2316 2317 old_page = vm_normal_page(vma, address, orig_pte); 2318 if (!old_page) { 2319 /* 2320 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 2321 * VM_PFNMAP VMA. 2322 * 2323 * We should not cow pages in a shared writeable mapping. 2324 * Just mark the pages writable and/or call ops->pfn_mkwrite. 2325 */ 2326 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2327 (VM_WRITE|VM_SHARED)) 2328 return wp_pfn_shared(mm, vma, address, page_table, ptl, 2329 orig_pte, pmd); 2330 2331 pte_unmap_unlock(page_table, ptl); 2332 return wp_page_copy(mm, vma, address, page_table, pmd, 2333 orig_pte, old_page); 2334 } 2335 2336 /* 2337 * Take out anonymous pages first, anonymous shared vmas are 2338 * not dirty accountable. 2339 */ 2340 if (PageAnon(old_page) && !PageKsm(old_page)) { 2341 if (!trylock_page(old_page)) { 2342 page_cache_get(old_page); 2343 pte_unmap_unlock(page_table, ptl); 2344 lock_page(old_page); 2345 page_table = pte_offset_map_lock(mm, pmd, address, 2346 &ptl); 2347 if (!pte_same(*page_table, orig_pte)) { 2348 unlock_page(old_page); 2349 pte_unmap_unlock(page_table, ptl); 2350 page_cache_release(old_page); 2351 return 0; 2352 } 2353 page_cache_release(old_page); 2354 } 2355 if (reuse_swap_page(old_page)) { 2356 /* 2357 * The page is all ours. Move it to our anon_vma so 2358 * the rmap code will not search our parent or siblings. 2359 * Protected against the rmap code by the page lock. 2360 */ 2361 page_move_anon_rmap(old_page, vma, address); 2362 unlock_page(old_page); 2363 return wp_page_reuse(mm, vma, address, page_table, ptl, 2364 orig_pte, old_page, 0, 0); 2365 } 2366 unlock_page(old_page); 2367 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2368 (VM_WRITE|VM_SHARED))) { 2369 return wp_page_shared(mm, vma, address, page_table, pmd, 2370 ptl, orig_pte, old_page); 2371 } 2372 2373 /* 2374 * Ok, we need to copy. Oh, well.. 2375 */ 2376 page_cache_get(old_page); 2377 2378 pte_unmap_unlock(page_table, ptl); 2379 return wp_page_copy(mm, vma, address, page_table, pmd, 2380 orig_pte, old_page); 2381 } 2382 2383 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 2384 unsigned long start_addr, unsigned long end_addr, 2385 struct zap_details *details) 2386 { 2387 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 2388 } 2389 2390 static inline void unmap_mapping_range_tree(struct rb_root *root, 2391 struct zap_details *details) 2392 { 2393 struct vm_area_struct *vma; 2394 pgoff_t vba, vea, zba, zea; 2395 2396 vma_interval_tree_foreach(vma, root, 2397 details->first_index, details->last_index) { 2398 2399 vba = vma->vm_pgoff; 2400 vea = vba + vma_pages(vma) - 1; 2401 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ 2402 zba = details->first_index; 2403 if (zba < vba) 2404 zba = vba; 2405 zea = details->last_index; 2406 if (zea > vea) 2407 zea = vea; 2408 2409 unmap_mapping_range_vma(vma, 2410 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 2411 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 2412 details); 2413 } 2414 } 2415 2416 /** 2417 * unmap_mapping_range - unmap the portion of all mmaps in the specified 2418 * address_space corresponding to the specified page range in the underlying 2419 * file. 2420 * 2421 * @mapping: the address space containing mmaps to be unmapped. 2422 * @holebegin: byte in first page to unmap, relative to the start of 2423 * the underlying file. This will be rounded down to a PAGE_SIZE 2424 * boundary. Note that this is different from truncate_pagecache(), which 2425 * must keep the partial page. In contrast, we must get rid of 2426 * partial pages. 2427 * @holelen: size of prospective hole in bytes. This will be rounded 2428 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 2429 * end of the file. 2430 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 2431 * but 0 when invalidating pagecache, don't throw away private data. 2432 */ 2433 void unmap_mapping_range(struct address_space *mapping, 2434 loff_t const holebegin, loff_t const holelen, int even_cows) 2435 { 2436 struct zap_details details; 2437 pgoff_t hba = holebegin >> PAGE_SHIFT; 2438 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2439 2440 /* Check for overflow. */ 2441 if (sizeof(holelen) > sizeof(hlen)) { 2442 long long holeend = 2443 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2444 if (holeend & ~(long long)ULONG_MAX) 2445 hlen = ULONG_MAX - hba + 1; 2446 } 2447 2448 details.check_mapping = even_cows? NULL: mapping; 2449 details.first_index = hba; 2450 details.last_index = hba + hlen - 1; 2451 if (details.last_index < details.first_index) 2452 details.last_index = ULONG_MAX; 2453 2454 2455 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */ 2456 i_mmap_lock_write(mapping); 2457 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap))) 2458 unmap_mapping_range_tree(&mapping->i_mmap, &details); 2459 i_mmap_unlock_write(mapping); 2460 } 2461 EXPORT_SYMBOL(unmap_mapping_range); 2462 2463 /* 2464 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2465 * but allow concurrent faults), and pte mapped but not yet locked. 2466 * We return with pte unmapped and unlocked. 2467 * 2468 * We return with the mmap_sem locked or unlocked in the same cases 2469 * as does filemap_fault(). 2470 */ 2471 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, 2472 unsigned long address, pte_t *page_table, pmd_t *pmd, 2473 unsigned int flags, pte_t orig_pte) 2474 { 2475 spinlock_t *ptl; 2476 struct page *page, *swapcache; 2477 struct mem_cgroup *memcg; 2478 swp_entry_t entry; 2479 pte_t pte; 2480 int locked; 2481 int exclusive = 0; 2482 int ret = 0; 2483 2484 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 2485 goto out; 2486 2487 entry = pte_to_swp_entry(orig_pte); 2488 if (unlikely(non_swap_entry(entry))) { 2489 if (is_migration_entry(entry)) { 2490 migration_entry_wait(mm, pmd, address); 2491 } else if (is_hwpoison_entry(entry)) { 2492 ret = VM_FAULT_HWPOISON; 2493 } else { 2494 print_bad_pte(vma, address, orig_pte, NULL); 2495 ret = VM_FAULT_SIGBUS; 2496 } 2497 goto out; 2498 } 2499 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 2500 page = lookup_swap_cache(entry); 2501 if (!page) { 2502 page = swapin_readahead(entry, 2503 GFP_HIGHUSER_MOVABLE, vma, address); 2504 if (!page) { 2505 /* 2506 * Back out if somebody else faulted in this pte 2507 * while we released the pte lock. 2508 */ 2509 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2510 if (likely(pte_same(*page_table, orig_pte))) 2511 ret = VM_FAULT_OOM; 2512 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2513 goto unlock; 2514 } 2515 2516 /* Had to read the page from swap area: Major fault */ 2517 ret = VM_FAULT_MAJOR; 2518 count_vm_event(PGMAJFAULT); 2519 mem_cgroup_count_vm_event(mm, PGMAJFAULT); 2520 } else if (PageHWPoison(page)) { 2521 /* 2522 * hwpoisoned dirty swapcache pages are kept for killing 2523 * owner processes (which may be unknown at hwpoison time) 2524 */ 2525 ret = VM_FAULT_HWPOISON; 2526 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2527 swapcache = page; 2528 goto out_release; 2529 } 2530 2531 swapcache = page; 2532 locked = lock_page_or_retry(page, mm, flags); 2533 2534 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2535 if (!locked) { 2536 ret |= VM_FAULT_RETRY; 2537 goto out_release; 2538 } 2539 2540 /* 2541 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 2542 * release the swapcache from under us. The page pin, and pte_same 2543 * test below, are not enough to exclude that. Even if it is still 2544 * swapcache, we need to check that the page's swap has not changed. 2545 */ 2546 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) 2547 goto out_page; 2548 2549 page = ksm_might_need_to_copy(page, vma, address); 2550 if (unlikely(!page)) { 2551 ret = VM_FAULT_OOM; 2552 page = swapcache; 2553 goto out_page; 2554 } 2555 2556 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) { 2557 ret = VM_FAULT_OOM; 2558 goto out_page; 2559 } 2560 2561 /* 2562 * Back out if somebody else already faulted in this pte. 2563 */ 2564 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2565 if (unlikely(!pte_same(*page_table, orig_pte))) 2566 goto out_nomap; 2567 2568 if (unlikely(!PageUptodate(page))) { 2569 ret = VM_FAULT_SIGBUS; 2570 goto out_nomap; 2571 } 2572 2573 /* 2574 * The page isn't present yet, go ahead with the fault. 2575 * 2576 * Be careful about the sequence of operations here. 2577 * To get its accounting right, reuse_swap_page() must be called 2578 * while the page is counted on swap but not yet in mapcount i.e. 2579 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 2580 * must be called after the swap_free(), or it will never succeed. 2581 */ 2582 2583 inc_mm_counter_fast(mm, MM_ANONPAGES); 2584 dec_mm_counter_fast(mm, MM_SWAPENTS); 2585 pte = mk_pte(page, vma->vm_page_prot); 2586 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { 2587 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 2588 flags &= ~FAULT_FLAG_WRITE; 2589 ret |= VM_FAULT_WRITE; 2590 exclusive = RMAP_EXCLUSIVE; 2591 } 2592 flush_icache_page(vma, page); 2593 if (pte_swp_soft_dirty(orig_pte)) 2594 pte = pte_mksoft_dirty(pte); 2595 set_pte_at(mm, address, page_table, pte); 2596 if (page == swapcache) { 2597 do_page_add_anon_rmap(page, vma, address, exclusive); 2598 mem_cgroup_commit_charge(page, memcg, true, false); 2599 } else { /* ksm created a completely new copy */ 2600 page_add_new_anon_rmap(page, vma, address, false); 2601 mem_cgroup_commit_charge(page, memcg, false, false); 2602 lru_cache_add_active_or_unevictable(page, vma); 2603 } 2604 2605 swap_free(entry); 2606 if (mem_cgroup_swap_full(page) || 2607 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 2608 try_to_free_swap(page); 2609 unlock_page(page); 2610 if (page != swapcache) { 2611 /* 2612 * Hold the lock to avoid the swap entry to be reused 2613 * until we take the PT lock for the pte_same() check 2614 * (to avoid false positives from pte_same). For 2615 * further safety release the lock after the swap_free 2616 * so that the swap count won't change under a 2617 * parallel locked swapcache. 2618 */ 2619 unlock_page(swapcache); 2620 page_cache_release(swapcache); 2621 } 2622 2623 if (flags & FAULT_FLAG_WRITE) { 2624 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); 2625 if (ret & VM_FAULT_ERROR) 2626 ret &= VM_FAULT_ERROR; 2627 goto out; 2628 } 2629 2630 /* No need to invalidate - it was non-present before */ 2631 update_mmu_cache(vma, address, page_table); 2632 unlock: 2633 pte_unmap_unlock(page_table, ptl); 2634 out: 2635 return ret; 2636 out_nomap: 2637 mem_cgroup_cancel_charge(page, memcg, false); 2638 pte_unmap_unlock(page_table, ptl); 2639 out_page: 2640 unlock_page(page); 2641 out_release: 2642 page_cache_release(page); 2643 if (page != swapcache) { 2644 unlock_page(swapcache); 2645 page_cache_release(swapcache); 2646 } 2647 return ret; 2648 } 2649 2650 /* 2651 * This is like a special single-page "expand_{down|up}wards()", 2652 * except we must first make sure that 'address{-|+}PAGE_SIZE' 2653 * doesn't hit another vma. 2654 */ 2655 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) 2656 { 2657 address &= PAGE_MASK; 2658 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { 2659 struct vm_area_struct *prev = vma->vm_prev; 2660 2661 /* 2662 * Is there a mapping abutting this one below? 2663 * 2664 * That's only ok if it's the same stack mapping 2665 * that has gotten split.. 2666 */ 2667 if (prev && prev->vm_end == address) 2668 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; 2669 2670 return expand_downwards(vma, address - PAGE_SIZE); 2671 } 2672 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { 2673 struct vm_area_struct *next = vma->vm_next; 2674 2675 /* As VM_GROWSDOWN but s/below/above/ */ 2676 if (next && next->vm_start == address + PAGE_SIZE) 2677 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; 2678 2679 return expand_upwards(vma, address + PAGE_SIZE); 2680 } 2681 return 0; 2682 } 2683 2684 /* 2685 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2686 * but allow concurrent faults), and pte mapped but not yet locked. 2687 * We return with mmap_sem still held, but pte unmapped and unlocked. 2688 */ 2689 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 2690 unsigned long address, pte_t *page_table, pmd_t *pmd, 2691 unsigned int flags) 2692 { 2693 struct mem_cgroup *memcg; 2694 struct page *page; 2695 spinlock_t *ptl; 2696 pte_t entry; 2697 2698 pte_unmap(page_table); 2699 2700 /* File mapping without ->vm_ops ? */ 2701 if (vma->vm_flags & VM_SHARED) 2702 return VM_FAULT_SIGBUS; 2703 2704 /* Check if we need to add a guard page to the stack */ 2705 if (check_stack_guard_page(vma, address) < 0) 2706 return VM_FAULT_SIGSEGV; 2707 2708 /* Use the zero-page for reads */ 2709 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) { 2710 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), 2711 vma->vm_page_prot)); 2712 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2713 if (!pte_none(*page_table)) 2714 goto unlock; 2715 /* Deliver the page fault to userland, check inside PT lock */ 2716 if (userfaultfd_missing(vma)) { 2717 pte_unmap_unlock(page_table, ptl); 2718 return handle_userfault(vma, address, flags, 2719 VM_UFFD_MISSING); 2720 } 2721 goto setpte; 2722 } 2723 2724 /* Allocate our own private page. */ 2725 if (unlikely(anon_vma_prepare(vma))) 2726 goto oom; 2727 page = alloc_zeroed_user_highpage_movable(vma, address); 2728 if (!page) 2729 goto oom; 2730 2731 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) 2732 goto oom_free_page; 2733 2734 /* 2735 * The memory barrier inside __SetPageUptodate makes sure that 2736 * preceeding stores to the page contents become visible before 2737 * the set_pte_at() write. 2738 */ 2739 __SetPageUptodate(page); 2740 2741 entry = mk_pte(page, vma->vm_page_prot); 2742 if (vma->vm_flags & VM_WRITE) 2743 entry = pte_mkwrite(pte_mkdirty(entry)); 2744 2745 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2746 if (!pte_none(*page_table)) 2747 goto release; 2748 2749 /* Deliver the page fault to userland, check inside PT lock */ 2750 if (userfaultfd_missing(vma)) { 2751 pte_unmap_unlock(page_table, ptl); 2752 mem_cgroup_cancel_charge(page, memcg, false); 2753 page_cache_release(page); 2754 return handle_userfault(vma, address, flags, 2755 VM_UFFD_MISSING); 2756 } 2757 2758 inc_mm_counter_fast(mm, MM_ANONPAGES); 2759 page_add_new_anon_rmap(page, vma, address, false); 2760 mem_cgroup_commit_charge(page, memcg, false, false); 2761 lru_cache_add_active_or_unevictable(page, vma); 2762 setpte: 2763 set_pte_at(mm, address, page_table, entry); 2764 2765 /* No need to invalidate - it was non-present before */ 2766 update_mmu_cache(vma, address, page_table); 2767 unlock: 2768 pte_unmap_unlock(page_table, ptl); 2769 return 0; 2770 release: 2771 mem_cgroup_cancel_charge(page, memcg, false); 2772 page_cache_release(page); 2773 goto unlock; 2774 oom_free_page: 2775 page_cache_release(page); 2776 oom: 2777 return VM_FAULT_OOM; 2778 } 2779 2780 /* 2781 * The mmap_sem must have been held on entry, and may have been 2782 * released depending on flags and vma->vm_ops->fault() return value. 2783 * See filemap_fault() and __lock_page_retry(). 2784 */ 2785 static int __do_fault(struct vm_area_struct *vma, unsigned long address, 2786 pgoff_t pgoff, unsigned int flags, 2787 struct page *cow_page, struct page **page) 2788 { 2789 struct vm_fault vmf; 2790 int ret; 2791 2792 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 2793 vmf.pgoff = pgoff; 2794 vmf.flags = flags; 2795 vmf.page = NULL; 2796 vmf.gfp_mask = __get_fault_gfp_mask(vma); 2797 vmf.cow_page = cow_page; 2798 2799 ret = vma->vm_ops->fault(vma, &vmf); 2800 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2801 return ret; 2802 if (!vmf.page) 2803 goto out; 2804 2805 if (unlikely(PageHWPoison(vmf.page))) { 2806 if (ret & VM_FAULT_LOCKED) 2807 unlock_page(vmf.page); 2808 page_cache_release(vmf.page); 2809 return VM_FAULT_HWPOISON; 2810 } 2811 2812 if (unlikely(!(ret & VM_FAULT_LOCKED))) 2813 lock_page(vmf.page); 2814 else 2815 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page); 2816 2817 out: 2818 *page = vmf.page; 2819 return ret; 2820 } 2821 2822 /** 2823 * do_set_pte - setup new PTE entry for given page and add reverse page mapping. 2824 * 2825 * @vma: virtual memory area 2826 * @address: user virtual address 2827 * @page: page to map 2828 * @pte: pointer to target page table entry 2829 * @write: true, if new entry is writable 2830 * @anon: true, if it's anonymous page 2831 * 2832 * Caller must hold page table lock relevant for @pte. 2833 * 2834 * Target users are page handler itself and implementations of 2835 * vm_ops->map_pages. 2836 */ 2837 void do_set_pte(struct vm_area_struct *vma, unsigned long address, 2838 struct page *page, pte_t *pte, bool write, bool anon) 2839 { 2840 pte_t entry; 2841 2842 flush_icache_page(vma, page); 2843 entry = mk_pte(page, vma->vm_page_prot); 2844 if (write) 2845 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2846 if (anon) { 2847 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 2848 page_add_new_anon_rmap(page, vma, address, false); 2849 } else { 2850 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 2851 page_add_file_rmap(page); 2852 } 2853 set_pte_at(vma->vm_mm, address, pte, entry); 2854 2855 /* no need to invalidate: a not-present page won't be cached */ 2856 update_mmu_cache(vma, address, pte); 2857 } 2858 2859 static unsigned long fault_around_bytes __read_mostly = 2860 rounddown_pow_of_two(65536); 2861 2862 #ifdef CONFIG_DEBUG_FS 2863 static int fault_around_bytes_get(void *data, u64 *val) 2864 { 2865 *val = fault_around_bytes; 2866 return 0; 2867 } 2868 2869 /* 2870 * fault_around_pages() and fault_around_mask() expects fault_around_bytes 2871 * rounded down to nearest page order. It's what do_fault_around() expects to 2872 * see. 2873 */ 2874 static int fault_around_bytes_set(void *data, u64 val) 2875 { 2876 if (val / PAGE_SIZE > PTRS_PER_PTE) 2877 return -EINVAL; 2878 if (val > PAGE_SIZE) 2879 fault_around_bytes = rounddown_pow_of_two(val); 2880 else 2881 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 2882 return 0; 2883 } 2884 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops, 2885 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 2886 2887 static int __init fault_around_debugfs(void) 2888 { 2889 void *ret; 2890 2891 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL, 2892 &fault_around_bytes_fops); 2893 if (!ret) 2894 pr_warn("Failed to create fault_around_bytes in debugfs"); 2895 return 0; 2896 } 2897 late_initcall(fault_around_debugfs); 2898 #endif 2899 2900 /* 2901 * do_fault_around() tries to map few pages around the fault address. The hope 2902 * is that the pages will be needed soon and this will lower the number of 2903 * faults to handle. 2904 * 2905 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 2906 * not ready to be mapped: not up-to-date, locked, etc. 2907 * 2908 * This function is called with the page table lock taken. In the split ptlock 2909 * case the page table lock only protects only those entries which belong to 2910 * the page table corresponding to the fault address. 2911 * 2912 * This function doesn't cross the VMA boundaries, in order to call map_pages() 2913 * only once. 2914 * 2915 * fault_around_pages() defines how many pages we'll try to map. 2916 * do_fault_around() expects it to return a power of two less than or equal to 2917 * PTRS_PER_PTE. 2918 * 2919 * The virtual address of the area that we map is naturally aligned to the 2920 * fault_around_pages() value (and therefore to page order). This way it's 2921 * easier to guarantee that we don't cross page table boundaries. 2922 */ 2923 static void do_fault_around(struct vm_area_struct *vma, unsigned long address, 2924 pte_t *pte, pgoff_t pgoff, unsigned int flags) 2925 { 2926 unsigned long start_addr, nr_pages, mask; 2927 pgoff_t max_pgoff; 2928 struct vm_fault vmf; 2929 int off; 2930 2931 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 2932 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 2933 2934 start_addr = max(address & mask, vma->vm_start); 2935 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 2936 pte -= off; 2937 pgoff -= off; 2938 2939 /* 2940 * max_pgoff is either end of page table or end of vma 2941 * or fault_around_pages() from pgoff, depending what is nearest. 2942 */ 2943 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 2944 PTRS_PER_PTE - 1; 2945 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1, 2946 pgoff + nr_pages - 1); 2947 2948 /* Check if it makes any sense to call ->map_pages */ 2949 while (!pte_none(*pte)) { 2950 if (++pgoff > max_pgoff) 2951 return; 2952 start_addr += PAGE_SIZE; 2953 if (start_addr >= vma->vm_end) 2954 return; 2955 pte++; 2956 } 2957 2958 vmf.virtual_address = (void __user *) start_addr; 2959 vmf.pte = pte; 2960 vmf.pgoff = pgoff; 2961 vmf.max_pgoff = max_pgoff; 2962 vmf.flags = flags; 2963 vmf.gfp_mask = __get_fault_gfp_mask(vma); 2964 vma->vm_ops->map_pages(vma, &vmf); 2965 } 2966 2967 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2968 unsigned long address, pmd_t *pmd, 2969 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 2970 { 2971 struct page *fault_page; 2972 spinlock_t *ptl; 2973 pte_t *pte; 2974 int ret = 0; 2975 2976 /* 2977 * Let's call ->map_pages() first and use ->fault() as fallback 2978 * if page by the offset is not ready to be mapped (cold cache or 2979 * something). 2980 */ 2981 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 2982 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2983 do_fault_around(vma, address, pte, pgoff, flags); 2984 if (!pte_same(*pte, orig_pte)) 2985 goto unlock_out; 2986 pte_unmap_unlock(pte, ptl); 2987 } 2988 2989 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page); 2990 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2991 return ret; 2992 2993 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2994 if (unlikely(!pte_same(*pte, orig_pte))) { 2995 pte_unmap_unlock(pte, ptl); 2996 unlock_page(fault_page); 2997 page_cache_release(fault_page); 2998 return ret; 2999 } 3000 do_set_pte(vma, address, fault_page, pte, false, false); 3001 unlock_page(fault_page); 3002 unlock_out: 3003 pte_unmap_unlock(pte, ptl); 3004 return ret; 3005 } 3006 3007 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3008 unsigned long address, pmd_t *pmd, 3009 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 3010 { 3011 struct page *fault_page, *new_page; 3012 struct mem_cgroup *memcg; 3013 spinlock_t *ptl; 3014 pte_t *pte; 3015 int ret; 3016 3017 if (unlikely(anon_vma_prepare(vma))) 3018 return VM_FAULT_OOM; 3019 3020 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 3021 if (!new_page) 3022 return VM_FAULT_OOM; 3023 3024 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) { 3025 page_cache_release(new_page); 3026 return VM_FAULT_OOM; 3027 } 3028 3029 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page); 3030 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3031 goto uncharge_out; 3032 3033 if (fault_page) 3034 copy_user_highpage(new_page, fault_page, address, vma); 3035 __SetPageUptodate(new_page); 3036 3037 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 3038 if (unlikely(!pte_same(*pte, orig_pte))) { 3039 pte_unmap_unlock(pte, ptl); 3040 if (fault_page) { 3041 unlock_page(fault_page); 3042 page_cache_release(fault_page); 3043 } else { 3044 /* 3045 * The fault handler has no page to lock, so it holds 3046 * i_mmap_lock for read to protect against truncate. 3047 */ 3048 i_mmap_unlock_read(vma->vm_file->f_mapping); 3049 } 3050 goto uncharge_out; 3051 } 3052 do_set_pte(vma, address, new_page, pte, true, true); 3053 mem_cgroup_commit_charge(new_page, memcg, false, false); 3054 lru_cache_add_active_or_unevictable(new_page, vma); 3055 pte_unmap_unlock(pte, ptl); 3056 if (fault_page) { 3057 unlock_page(fault_page); 3058 page_cache_release(fault_page); 3059 } else { 3060 /* 3061 * The fault handler has no page to lock, so it holds 3062 * i_mmap_lock for read to protect against truncate. 3063 */ 3064 i_mmap_unlock_read(vma->vm_file->f_mapping); 3065 } 3066 return ret; 3067 uncharge_out: 3068 mem_cgroup_cancel_charge(new_page, memcg, false); 3069 page_cache_release(new_page); 3070 return ret; 3071 } 3072 3073 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3074 unsigned long address, pmd_t *pmd, 3075 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 3076 { 3077 struct page *fault_page; 3078 struct address_space *mapping; 3079 spinlock_t *ptl; 3080 pte_t *pte; 3081 int dirtied = 0; 3082 int ret, tmp; 3083 3084 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page); 3085 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3086 return ret; 3087 3088 /* 3089 * Check if the backing address space wants to know that the page is 3090 * about to become writable 3091 */ 3092 if (vma->vm_ops->page_mkwrite) { 3093 unlock_page(fault_page); 3094 tmp = do_page_mkwrite(vma, fault_page, address); 3095 if (unlikely(!tmp || 3096 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3097 page_cache_release(fault_page); 3098 return tmp; 3099 } 3100 } 3101 3102 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 3103 if (unlikely(!pte_same(*pte, orig_pte))) { 3104 pte_unmap_unlock(pte, ptl); 3105 unlock_page(fault_page); 3106 page_cache_release(fault_page); 3107 return ret; 3108 } 3109 do_set_pte(vma, address, fault_page, pte, true, false); 3110 pte_unmap_unlock(pte, ptl); 3111 3112 if (set_page_dirty(fault_page)) 3113 dirtied = 1; 3114 /* 3115 * Take a local copy of the address_space - page.mapping may be zeroed 3116 * by truncate after unlock_page(). The address_space itself remains 3117 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 3118 * release semantics to prevent the compiler from undoing this copying. 3119 */ 3120 mapping = page_rmapping(fault_page); 3121 unlock_page(fault_page); 3122 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) { 3123 /* 3124 * Some device drivers do not set page.mapping but still 3125 * dirty their pages 3126 */ 3127 balance_dirty_pages_ratelimited(mapping); 3128 } 3129 3130 if (!vma->vm_ops->page_mkwrite) 3131 file_update_time(vma->vm_file); 3132 3133 return ret; 3134 } 3135 3136 /* 3137 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3138 * but allow concurrent faults). 3139 * The mmap_sem may have been released depending on flags and our 3140 * return value. See filemap_fault() and __lock_page_or_retry(). 3141 */ 3142 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3143 unsigned long address, pte_t *page_table, pmd_t *pmd, 3144 unsigned int flags, pte_t orig_pte) 3145 { 3146 pgoff_t pgoff = (((address & PAGE_MASK) 3147 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; 3148 3149 pte_unmap(page_table); 3150 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */ 3151 if (!vma->vm_ops->fault) 3152 return VM_FAULT_SIGBUS; 3153 if (!(flags & FAULT_FLAG_WRITE)) 3154 return do_read_fault(mm, vma, address, pmd, pgoff, flags, 3155 orig_pte); 3156 if (!(vma->vm_flags & VM_SHARED)) 3157 return do_cow_fault(mm, vma, address, pmd, pgoff, flags, 3158 orig_pte); 3159 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); 3160 } 3161 3162 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 3163 unsigned long addr, int page_nid, 3164 int *flags) 3165 { 3166 get_page(page); 3167 3168 count_vm_numa_event(NUMA_HINT_FAULTS); 3169 if (page_nid == numa_node_id()) { 3170 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 3171 *flags |= TNF_FAULT_LOCAL; 3172 } 3173 3174 return mpol_misplaced(page, vma, addr); 3175 } 3176 3177 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, 3178 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd) 3179 { 3180 struct page *page = NULL; 3181 spinlock_t *ptl; 3182 int page_nid = -1; 3183 int last_cpupid; 3184 int target_nid; 3185 bool migrated = false; 3186 bool was_writable = pte_write(pte); 3187 int flags = 0; 3188 3189 /* A PROT_NONE fault should not end up here */ 3190 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); 3191 3192 /* 3193 * The "pte" at this point cannot be used safely without 3194 * validation through pte_unmap_same(). It's of NUMA type but 3195 * the pfn may be screwed if the read is non atomic. 3196 * 3197 * We can safely just do a "set_pte_at()", because the old 3198 * page table entry is not accessible, so there would be no 3199 * concurrent hardware modifications to the PTE. 3200 */ 3201 ptl = pte_lockptr(mm, pmd); 3202 spin_lock(ptl); 3203 if (unlikely(!pte_same(*ptep, pte))) { 3204 pte_unmap_unlock(ptep, ptl); 3205 goto out; 3206 } 3207 3208 /* Make it present again */ 3209 pte = pte_modify(pte, vma->vm_page_prot); 3210 pte = pte_mkyoung(pte); 3211 if (was_writable) 3212 pte = pte_mkwrite(pte); 3213 set_pte_at(mm, addr, ptep, pte); 3214 update_mmu_cache(vma, addr, ptep); 3215 3216 page = vm_normal_page(vma, addr, pte); 3217 if (!page) { 3218 pte_unmap_unlock(ptep, ptl); 3219 return 0; 3220 } 3221 3222 /* TODO: handle PTE-mapped THP */ 3223 if (PageCompound(page)) { 3224 pte_unmap_unlock(ptep, ptl); 3225 return 0; 3226 } 3227 3228 /* 3229 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 3230 * much anyway since they can be in shared cache state. This misses 3231 * the case where a mapping is writable but the process never writes 3232 * to it but pte_write gets cleared during protection updates and 3233 * pte_dirty has unpredictable behaviour between PTE scan updates, 3234 * background writeback, dirty balancing and application behaviour. 3235 */ 3236 if (!(vma->vm_flags & VM_WRITE)) 3237 flags |= TNF_NO_GROUP; 3238 3239 /* 3240 * Flag if the page is shared between multiple address spaces. This 3241 * is later used when determining whether to group tasks together 3242 */ 3243 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 3244 flags |= TNF_SHARED; 3245 3246 last_cpupid = page_cpupid_last(page); 3247 page_nid = page_to_nid(page); 3248 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags); 3249 pte_unmap_unlock(ptep, ptl); 3250 if (target_nid == -1) { 3251 put_page(page); 3252 goto out; 3253 } 3254 3255 /* Migrate to the requested node */ 3256 migrated = migrate_misplaced_page(page, vma, target_nid); 3257 if (migrated) { 3258 page_nid = target_nid; 3259 flags |= TNF_MIGRATED; 3260 } else 3261 flags |= TNF_MIGRATE_FAIL; 3262 3263 out: 3264 if (page_nid != -1) 3265 task_numa_fault(last_cpupid, page_nid, 1, flags); 3266 return 0; 3267 } 3268 3269 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 3270 unsigned long address, pmd_t *pmd, unsigned int flags) 3271 { 3272 if (vma_is_anonymous(vma)) 3273 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags); 3274 if (vma->vm_ops->pmd_fault) 3275 return vma->vm_ops->pmd_fault(vma, address, pmd, flags); 3276 return VM_FAULT_FALLBACK; 3277 } 3278 3279 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma, 3280 unsigned long address, pmd_t *pmd, pmd_t orig_pmd, 3281 unsigned int flags) 3282 { 3283 if (vma_is_anonymous(vma)) 3284 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd); 3285 if (vma->vm_ops->pmd_fault) 3286 return vma->vm_ops->pmd_fault(vma, address, pmd, flags); 3287 return VM_FAULT_FALLBACK; 3288 } 3289 3290 /* 3291 * These routines also need to handle stuff like marking pages dirty 3292 * and/or accessed for architectures that don't do it in hardware (most 3293 * RISC architectures). The early dirtying is also good on the i386. 3294 * 3295 * There is also a hook called "update_mmu_cache()" that architectures 3296 * with external mmu caches can use to update those (ie the Sparc or 3297 * PowerPC hashed page tables that act as extended TLBs). 3298 * 3299 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3300 * but allow concurrent faults), and pte mapped but not yet locked. 3301 * We return with pte unmapped and unlocked. 3302 * 3303 * The mmap_sem may have been released depending on flags and our 3304 * return value. See filemap_fault() and __lock_page_or_retry(). 3305 */ 3306 static int handle_pte_fault(struct mm_struct *mm, 3307 struct vm_area_struct *vma, unsigned long address, 3308 pte_t *pte, pmd_t *pmd, unsigned int flags) 3309 { 3310 pte_t entry; 3311 spinlock_t *ptl; 3312 3313 /* 3314 * some architectures can have larger ptes than wordsize, 3315 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y, 3316 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses. 3317 * The code below just needs a consistent view for the ifs and 3318 * we later double check anyway with the ptl lock held. So here 3319 * a barrier will do. 3320 */ 3321 entry = *pte; 3322 barrier(); 3323 if (!pte_present(entry)) { 3324 if (pte_none(entry)) { 3325 if (vma_is_anonymous(vma)) 3326 return do_anonymous_page(mm, vma, address, 3327 pte, pmd, flags); 3328 else 3329 return do_fault(mm, vma, address, pte, pmd, 3330 flags, entry); 3331 } 3332 return do_swap_page(mm, vma, address, 3333 pte, pmd, flags, entry); 3334 } 3335 3336 if (pte_protnone(entry)) 3337 return do_numa_page(mm, vma, address, entry, pte, pmd); 3338 3339 ptl = pte_lockptr(mm, pmd); 3340 spin_lock(ptl); 3341 if (unlikely(!pte_same(*pte, entry))) 3342 goto unlock; 3343 if (flags & FAULT_FLAG_WRITE) { 3344 if (!pte_write(entry)) 3345 return do_wp_page(mm, vma, address, 3346 pte, pmd, ptl, entry); 3347 entry = pte_mkdirty(entry); 3348 } 3349 entry = pte_mkyoung(entry); 3350 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) { 3351 update_mmu_cache(vma, address, pte); 3352 } else { 3353 /* 3354 * This is needed only for protection faults but the arch code 3355 * is not yet telling us if this is a protection fault or not. 3356 * This still avoids useless tlb flushes for .text page faults 3357 * with threads. 3358 */ 3359 if (flags & FAULT_FLAG_WRITE) 3360 flush_tlb_fix_spurious_fault(vma, address); 3361 } 3362 unlock: 3363 pte_unmap_unlock(pte, ptl); 3364 return 0; 3365 } 3366 3367 /* 3368 * By the time we get here, we already hold the mm semaphore 3369 * 3370 * The mmap_sem may have been released depending on flags and our 3371 * return value. See filemap_fault() and __lock_page_or_retry(). 3372 */ 3373 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3374 unsigned long address, unsigned int flags) 3375 { 3376 pgd_t *pgd; 3377 pud_t *pud; 3378 pmd_t *pmd; 3379 pte_t *pte; 3380 3381 if (unlikely(is_vm_hugetlb_page(vma))) 3382 return hugetlb_fault(mm, vma, address, flags); 3383 3384 pgd = pgd_offset(mm, address); 3385 pud = pud_alloc(mm, pgd, address); 3386 if (!pud) 3387 return VM_FAULT_OOM; 3388 pmd = pmd_alloc(mm, pud, address); 3389 if (!pmd) 3390 return VM_FAULT_OOM; 3391 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { 3392 int ret = create_huge_pmd(mm, vma, address, pmd, flags); 3393 if (!(ret & VM_FAULT_FALLBACK)) 3394 return ret; 3395 } else { 3396 pmd_t orig_pmd = *pmd; 3397 int ret; 3398 3399 barrier(); 3400 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) { 3401 unsigned int dirty = flags & FAULT_FLAG_WRITE; 3402 3403 if (pmd_protnone(orig_pmd)) 3404 return do_huge_pmd_numa_page(mm, vma, address, 3405 orig_pmd, pmd); 3406 3407 if (dirty && !pmd_write(orig_pmd)) { 3408 ret = wp_huge_pmd(mm, vma, address, pmd, 3409 orig_pmd, flags); 3410 if (!(ret & VM_FAULT_FALLBACK)) 3411 return ret; 3412 } else { 3413 huge_pmd_set_accessed(mm, vma, address, pmd, 3414 orig_pmd, dirty); 3415 return 0; 3416 } 3417 } 3418 } 3419 3420 /* 3421 * Use __pte_alloc instead of pte_alloc_map, because we can't 3422 * run pte_offset_map on the pmd, if an huge pmd could 3423 * materialize from under us from a different thread. 3424 */ 3425 if (unlikely(pmd_none(*pmd)) && 3426 unlikely(__pte_alloc(mm, vma, pmd, address))) 3427 return VM_FAULT_OOM; 3428 /* if an huge pmd materialized from under us just retry later */ 3429 if (unlikely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) 3430 return 0; 3431 /* 3432 * A regular pmd is established and it can't morph into a huge pmd 3433 * from under us anymore at this point because we hold the mmap_sem 3434 * read mode and khugepaged takes it in write mode. So now it's 3435 * safe to run pte_offset_map(). 3436 */ 3437 pte = pte_offset_map(pmd, address); 3438 3439 return handle_pte_fault(mm, vma, address, pte, pmd, flags); 3440 } 3441 3442 /* 3443 * By the time we get here, we already hold the mm semaphore 3444 * 3445 * The mmap_sem may have been released depending on flags and our 3446 * return value. See filemap_fault() and __lock_page_or_retry(). 3447 */ 3448 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3449 unsigned long address, unsigned int flags) 3450 { 3451 int ret; 3452 3453 __set_current_state(TASK_RUNNING); 3454 3455 count_vm_event(PGFAULT); 3456 mem_cgroup_count_vm_event(mm, PGFAULT); 3457 3458 /* do counter updates before entering really critical section. */ 3459 check_sync_rss_stat(current); 3460 3461 /* 3462 * Enable the memcg OOM handling for faults triggered in user 3463 * space. Kernel faults are handled more gracefully. 3464 */ 3465 if (flags & FAULT_FLAG_USER) 3466 mem_cgroup_oom_enable(); 3467 3468 ret = __handle_mm_fault(mm, vma, address, flags); 3469 3470 if (flags & FAULT_FLAG_USER) { 3471 mem_cgroup_oom_disable(); 3472 /* 3473 * The task may have entered a memcg OOM situation but 3474 * if the allocation error was handled gracefully (no 3475 * VM_FAULT_OOM), there is no need to kill anything. 3476 * Just clean up the OOM state peacefully. 3477 */ 3478 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 3479 mem_cgroup_oom_synchronize(false); 3480 } 3481 3482 return ret; 3483 } 3484 EXPORT_SYMBOL_GPL(handle_mm_fault); 3485 3486 #ifndef __PAGETABLE_PUD_FOLDED 3487 /* 3488 * Allocate page upper directory. 3489 * We've already handled the fast-path in-line. 3490 */ 3491 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 3492 { 3493 pud_t *new = pud_alloc_one(mm, address); 3494 if (!new) 3495 return -ENOMEM; 3496 3497 smp_wmb(); /* See comment in __pte_alloc */ 3498 3499 spin_lock(&mm->page_table_lock); 3500 if (pgd_present(*pgd)) /* Another has populated it */ 3501 pud_free(mm, new); 3502 else 3503 pgd_populate(mm, pgd, new); 3504 spin_unlock(&mm->page_table_lock); 3505 return 0; 3506 } 3507 #endif /* __PAGETABLE_PUD_FOLDED */ 3508 3509 #ifndef __PAGETABLE_PMD_FOLDED 3510 /* 3511 * Allocate page middle directory. 3512 * We've already handled the fast-path in-line. 3513 */ 3514 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 3515 { 3516 pmd_t *new = pmd_alloc_one(mm, address); 3517 if (!new) 3518 return -ENOMEM; 3519 3520 smp_wmb(); /* See comment in __pte_alloc */ 3521 3522 spin_lock(&mm->page_table_lock); 3523 #ifndef __ARCH_HAS_4LEVEL_HACK 3524 if (!pud_present(*pud)) { 3525 mm_inc_nr_pmds(mm); 3526 pud_populate(mm, pud, new); 3527 } else /* Another has populated it */ 3528 pmd_free(mm, new); 3529 #else 3530 if (!pgd_present(*pud)) { 3531 mm_inc_nr_pmds(mm); 3532 pgd_populate(mm, pud, new); 3533 } else /* Another has populated it */ 3534 pmd_free(mm, new); 3535 #endif /* __ARCH_HAS_4LEVEL_HACK */ 3536 spin_unlock(&mm->page_table_lock); 3537 return 0; 3538 } 3539 #endif /* __PAGETABLE_PMD_FOLDED */ 3540 3541 static int __follow_pte(struct mm_struct *mm, unsigned long address, 3542 pte_t **ptepp, spinlock_t **ptlp) 3543 { 3544 pgd_t *pgd; 3545 pud_t *pud; 3546 pmd_t *pmd; 3547 pte_t *ptep; 3548 3549 pgd = pgd_offset(mm, address); 3550 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 3551 goto out; 3552 3553 pud = pud_offset(pgd, address); 3554 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 3555 goto out; 3556 3557 pmd = pmd_offset(pud, address); 3558 VM_BUG_ON(pmd_trans_huge(*pmd)); 3559 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 3560 goto out; 3561 3562 /* We cannot handle huge page PFN maps. Luckily they don't exist. */ 3563 if (pmd_huge(*pmd)) 3564 goto out; 3565 3566 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 3567 if (!ptep) 3568 goto out; 3569 if (!pte_present(*ptep)) 3570 goto unlock; 3571 *ptepp = ptep; 3572 return 0; 3573 unlock: 3574 pte_unmap_unlock(ptep, *ptlp); 3575 out: 3576 return -EINVAL; 3577 } 3578 3579 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 3580 pte_t **ptepp, spinlock_t **ptlp) 3581 { 3582 int res; 3583 3584 /* (void) is needed to make gcc happy */ 3585 (void) __cond_lock(*ptlp, 3586 !(res = __follow_pte(mm, address, ptepp, ptlp))); 3587 return res; 3588 } 3589 3590 /** 3591 * follow_pfn - look up PFN at a user virtual address 3592 * @vma: memory mapping 3593 * @address: user virtual address 3594 * @pfn: location to store found PFN 3595 * 3596 * Only IO mappings and raw PFN mappings are allowed. 3597 * 3598 * Returns zero and the pfn at @pfn on success, -ve otherwise. 3599 */ 3600 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 3601 unsigned long *pfn) 3602 { 3603 int ret = -EINVAL; 3604 spinlock_t *ptl; 3605 pte_t *ptep; 3606 3607 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3608 return ret; 3609 3610 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 3611 if (ret) 3612 return ret; 3613 *pfn = pte_pfn(*ptep); 3614 pte_unmap_unlock(ptep, ptl); 3615 return 0; 3616 } 3617 EXPORT_SYMBOL(follow_pfn); 3618 3619 #ifdef CONFIG_HAVE_IOREMAP_PROT 3620 int follow_phys(struct vm_area_struct *vma, 3621 unsigned long address, unsigned int flags, 3622 unsigned long *prot, resource_size_t *phys) 3623 { 3624 int ret = -EINVAL; 3625 pte_t *ptep, pte; 3626 spinlock_t *ptl; 3627 3628 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3629 goto out; 3630 3631 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 3632 goto out; 3633 pte = *ptep; 3634 3635 if ((flags & FOLL_WRITE) && !pte_write(pte)) 3636 goto unlock; 3637 3638 *prot = pgprot_val(pte_pgprot(pte)); 3639 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 3640 3641 ret = 0; 3642 unlock: 3643 pte_unmap_unlock(ptep, ptl); 3644 out: 3645 return ret; 3646 } 3647 3648 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 3649 void *buf, int len, int write) 3650 { 3651 resource_size_t phys_addr; 3652 unsigned long prot = 0; 3653 void __iomem *maddr; 3654 int offset = addr & (PAGE_SIZE-1); 3655 3656 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 3657 return -EINVAL; 3658 3659 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 3660 if (write) 3661 memcpy_toio(maddr + offset, buf, len); 3662 else 3663 memcpy_fromio(buf, maddr + offset, len); 3664 iounmap(maddr); 3665 3666 return len; 3667 } 3668 EXPORT_SYMBOL_GPL(generic_access_phys); 3669 #endif 3670 3671 /* 3672 * Access another process' address space as given in mm. If non-NULL, use the 3673 * given task for page fault accounting. 3674 */ 3675 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, 3676 unsigned long addr, void *buf, int len, int write) 3677 { 3678 struct vm_area_struct *vma; 3679 void *old_buf = buf; 3680 3681 down_read(&mm->mmap_sem); 3682 /* ignore errors, just check how much was successfully transferred */ 3683 while (len) { 3684 int bytes, ret, offset; 3685 void *maddr; 3686 struct page *page = NULL; 3687 3688 ret = get_user_pages(tsk, mm, addr, 1, 3689 write, 1, &page, &vma); 3690 if (ret <= 0) { 3691 #ifndef CONFIG_HAVE_IOREMAP_PROT 3692 break; 3693 #else 3694 /* 3695 * Check if this is a VM_IO | VM_PFNMAP VMA, which 3696 * we can access using slightly different code. 3697 */ 3698 vma = find_vma(mm, addr); 3699 if (!vma || vma->vm_start > addr) 3700 break; 3701 if (vma->vm_ops && vma->vm_ops->access) 3702 ret = vma->vm_ops->access(vma, addr, buf, 3703 len, write); 3704 if (ret <= 0) 3705 break; 3706 bytes = ret; 3707 #endif 3708 } else { 3709 bytes = len; 3710 offset = addr & (PAGE_SIZE-1); 3711 if (bytes > PAGE_SIZE-offset) 3712 bytes = PAGE_SIZE-offset; 3713 3714 maddr = kmap(page); 3715 if (write) { 3716 copy_to_user_page(vma, page, addr, 3717 maddr + offset, buf, bytes); 3718 set_page_dirty_lock(page); 3719 } else { 3720 copy_from_user_page(vma, page, addr, 3721 buf, maddr + offset, bytes); 3722 } 3723 kunmap(page); 3724 page_cache_release(page); 3725 } 3726 len -= bytes; 3727 buf += bytes; 3728 addr += bytes; 3729 } 3730 up_read(&mm->mmap_sem); 3731 3732 return buf - old_buf; 3733 } 3734 3735 /** 3736 * access_remote_vm - access another process' address space 3737 * @mm: the mm_struct of the target address space 3738 * @addr: start address to access 3739 * @buf: source or destination buffer 3740 * @len: number of bytes to transfer 3741 * @write: whether the access is a write 3742 * 3743 * The caller must hold a reference on @mm. 3744 */ 3745 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 3746 void *buf, int len, int write) 3747 { 3748 return __access_remote_vm(NULL, mm, addr, buf, len, write); 3749 } 3750 3751 /* 3752 * Access another process' address space. 3753 * Source/target buffer must be kernel space, 3754 * Do not walk the page table directly, use get_user_pages 3755 */ 3756 int access_process_vm(struct task_struct *tsk, unsigned long addr, 3757 void *buf, int len, int write) 3758 { 3759 struct mm_struct *mm; 3760 int ret; 3761 3762 mm = get_task_mm(tsk); 3763 if (!mm) 3764 return 0; 3765 3766 ret = __access_remote_vm(tsk, mm, addr, buf, len, write); 3767 mmput(mm); 3768 3769 return ret; 3770 } 3771 3772 /* 3773 * Print the name of a VMA. 3774 */ 3775 void print_vma_addr(char *prefix, unsigned long ip) 3776 { 3777 struct mm_struct *mm = current->mm; 3778 struct vm_area_struct *vma; 3779 3780 /* 3781 * Do not print if we are in atomic 3782 * contexts (in exception stacks, etc.): 3783 */ 3784 if (preempt_count()) 3785 return; 3786 3787 down_read(&mm->mmap_sem); 3788 vma = find_vma(mm, ip); 3789 if (vma && vma->vm_file) { 3790 struct file *f = vma->vm_file; 3791 char *buf = (char *)__get_free_page(GFP_KERNEL); 3792 if (buf) { 3793 char *p; 3794 3795 p = file_path(f, buf, PAGE_SIZE); 3796 if (IS_ERR(p)) 3797 p = "?"; 3798 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 3799 vma->vm_start, 3800 vma->vm_end - vma->vm_start); 3801 free_page((unsigned long)buf); 3802 } 3803 } 3804 up_read(&mm->mmap_sem); 3805 } 3806 3807 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 3808 void __might_fault(const char *file, int line) 3809 { 3810 /* 3811 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 3812 * holding the mmap_sem, this is safe because kernel memory doesn't 3813 * get paged out, therefore we'll never actually fault, and the 3814 * below annotations will generate false positives. 3815 */ 3816 if (segment_eq(get_fs(), KERNEL_DS)) 3817 return; 3818 if (pagefault_disabled()) 3819 return; 3820 __might_sleep(file, line, 0); 3821 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 3822 if (current->mm) 3823 might_lock_read(¤t->mm->mmap_sem); 3824 #endif 3825 } 3826 EXPORT_SYMBOL(__might_fault); 3827 #endif 3828 3829 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 3830 static void clear_gigantic_page(struct page *page, 3831 unsigned long addr, 3832 unsigned int pages_per_huge_page) 3833 { 3834 int i; 3835 struct page *p = page; 3836 3837 might_sleep(); 3838 for (i = 0; i < pages_per_huge_page; 3839 i++, p = mem_map_next(p, page, i)) { 3840 cond_resched(); 3841 clear_user_highpage(p, addr + i * PAGE_SIZE); 3842 } 3843 } 3844 void clear_huge_page(struct page *page, 3845 unsigned long addr, unsigned int pages_per_huge_page) 3846 { 3847 int i; 3848 3849 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3850 clear_gigantic_page(page, addr, pages_per_huge_page); 3851 return; 3852 } 3853 3854 might_sleep(); 3855 for (i = 0; i < pages_per_huge_page; i++) { 3856 cond_resched(); 3857 clear_user_highpage(page + i, addr + i * PAGE_SIZE); 3858 } 3859 } 3860 3861 static void copy_user_gigantic_page(struct page *dst, struct page *src, 3862 unsigned long addr, 3863 struct vm_area_struct *vma, 3864 unsigned int pages_per_huge_page) 3865 { 3866 int i; 3867 struct page *dst_base = dst; 3868 struct page *src_base = src; 3869 3870 for (i = 0; i < pages_per_huge_page; ) { 3871 cond_resched(); 3872 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 3873 3874 i++; 3875 dst = mem_map_next(dst, dst_base, i); 3876 src = mem_map_next(src, src_base, i); 3877 } 3878 } 3879 3880 void copy_user_huge_page(struct page *dst, struct page *src, 3881 unsigned long addr, struct vm_area_struct *vma, 3882 unsigned int pages_per_huge_page) 3883 { 3884 int i; 3885 3886 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3887 copy_user_gigantic_page(dst, src, addr, vma, 3888 pages_per_huge_page); 3889 return; 3890 } 3891 3892 might_sleep(); 3893 for (i = 0; i < pages_per_huge_page; i++) { 3894 cond_resched(); 3895 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 3896 } 3897 } 3898 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 3899 3900 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 3901 3902 static struct kmem_cache *page_ptl_cachep; 3903 3904 void __init ptlock_cache_init(void) 3905 { 3906 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 3907 SLAB_PANIC, NULL); 3908 } 3909 3910 bool ptlock_alloc(struct page *page) 3911 { 3912 spinlock_t *ptl; 3913 3914 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 3915 if (!ptl) 3916 return false; 3917 page->ptl = ptl; 3918 return true; 3919 } 3920 3921 void ptlock_free(struct page *page) 3922 { 3923 kmem_cache_free(page_ptl_cachep, page->ptl); 3924 } 3925 #endif 3926