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/mmu_context.h> 69 #include <asm/pgalloc.h> 70 #include <asm/uaccess.h> 71 #include <asm/tlb.h> 72 #include <asm/tlbflush.h> 73 #include <asm/pgtable.h> 74 75 #include "internal.h" 76 77 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS 78 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 79 #endif 80 81 #ifndef CONFIG_NEED_MULTIPLE_NODES 82 /* use the per-pgdat data instead for discontigmem - mbligh */ 83 unsigned long max_mapnr; 84 struct page *mem_map; 85 86 EXPORT_SYMBOL(max_mapnr); 87 EXPORT_SYMBOL(mem_map); 88 #endif 89 90 /* 91 * A number of key systems in x86 including ioremap() rely on the assumption 92 * that high_memory defines the upper bound on direct map memory, then end 93 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 95 * and ZONE_HIGHMEM. 96 */ 97 void * high_memory; 98 99 EXPORT_SYMBOL(high_memory); 100 101 /* 102 * Randomize the address space (stacks, mmaps, brk, etc.). 103 * 104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 105 * as ancient (libc5 based) binaries can segfault. ) 106 */ 107 int randomize_va_space __read_mostly = 108 #ifdef CONFIG_COMPAT_BRK 109 1; 110 #else 111 2; 112 #endif 113 114 static int __init disable_randmaps(char *s) 115 { 116 randomize_va_space = 0; 117 return 1; 118 } 119 __setup("norandmaps", disable_randmaps); 120 121 unsigned long zero_pfn __read_mostly; 122 unsigned long highest_memmap_pfn __read_mostly; 123 124 EXPORT_SYMBOL(zero_pfn); 125 126 /* 127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 128 */ 129 static int __init init_zero_pfn(void) 130 { 131 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 132 return 0; 133 } 134 core_initcall(init_zero_pfn); 135 136 137 #if defined(SPLIT_RSS_COUNTING) 138 139 void sync_mm_rss(struct mm_struct *mm) 140 { 141 int i; 142 143 for (i = 0; i < NR_MM_COUNTERS; i++) { 144 if (current->rss_stat.count[i]) { 145 add_mm_counter(mm, i, current->rss_stat.count[i]); 146 current->rss_stat.count[i] = 0; 147 } 148 } 149 current->rss_stat.events = 0; 150 } 151 152 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 153 { 154 struct task_struct *task = current; 155 156 if (likely(task->mm == mm)) 157 task->rss_stat.count[member] += val; 158 else 159 add_mm_counter(mm, member, val); 160 } 161 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 162 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 163 164 /* sync counter once per 64 page faults */ 165 #define TASK_RSS_EVENTS_THRESH (64) 166 static void check_sync_rss_stat(struct task_struct *task) 167 { 168 if (unlikely(task != current)) 169 return; 170 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 171 sync_mm_rss(task->mm); 172 } 173 #else /* SPLIT_RSS_COUNTING */ 174 175 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 176 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 177 178 static void check_sync_rss_stat(struct task_struct *task) 179 { 180 } 181 182 #endif /* SPLIT_RSS_COUNTING */ 183 184 #ifdef HAVE_GENERIC_MMU_GATHER 185 186 static bool tlb_next_batch(struct mmu_gather *tlb) 187 { 188 struct mmu_gather_batch *batch; 189 190 batch = tlb->active; 191 if (batch->next) { 192 tlb->active = batch->next; 193 return true; 194 } 195 196 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT) 197 return false; 198 199 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0); 200 if (!batch) 201 return false; 202 203 tlb->batch_count++; 204 batch->next = NULL; 205 batch->nr = 0; 206 batch->max = MAX_GATHER_BATCH; 207 208 tlb->active->next = batch; 209 tlb->active = batch; 210 211 return true; 212 } 213 214 /* tlb_gather_mmu 215 * Called to initialize an (on-stack) mmu_gather structure for page-table 216 * tear-down from @mm. The @fullmm argument is used when @mm is without 217 * users and we're going to destroy the full address space (exit/execve). 218 */ 219 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end) 220 { 221 tlb->mm = mm; 222 223 /* Is it from 0 to ~0? */ 224 tlb->fullmm = !(start | (end+1)); 225 tlb->need_flush_all = 0; 226 tlb->local.next = NULL; 227 tlb->local.nr = 0; 228 tlb->local.max = ARRAY_SIZE(tlb->__pages); 229 tlb->active = &tlb->local; 230 tlb->batch_count = 0; 231 232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 233 tlb->batch = NULL; 234 #endif 235 236 __tlb_reset_range(tlb); 237 } 238 239 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb) 240 { 241 if (!tlb->end) 242 return; 243 244 tlb_flush(tlb); 245 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end); 246 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 247 tlb_table_flush(tlb); 248 #endif 249 __tlb_reset_range(tlb); 250 } 251 252 static void tlb_flush_mmu_free(struct mmu_gather *tlb) 253 { 254 struct mmu_gather_batch *batch; 255 256 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) { 257 free_pages_and_swap_cache(batch->pages, batch->nr); 258 batch->nr = 0; 259 } 260 tlb->active = &tlb->local; 261 } 262 263 void tlb_flush_mmu(struct mmu_gather *tlb) 264 { 265 tlb_flush_mmu_tlbonly(tlb); 266 tlb_flush_mmu_free(tlb); 267 } 268 269 /* tlb_finish_mmu 270 * Called at the end of the shootdown operation to free up any resources 271 * that were required. 272 */ 273 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) 274 { 275 struct mmu_gather_batch *batch, *next; 276 277 tlb_flush_mmu(tlb); 278 279 /* keep the page table cache within bounds */ 280 check_pgt_cache(); 281 282 for (batch = tlb->local.next; batch; batch = next) { 283 next = batch->next; 284 free_pages((unsigned long)batch, 0); 285 } 286 tlb->local.next = NULL; 287 } 288 289 /* __tlb_remove_page 290 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while 291 * handling the additional races in SMP caused by other CPUs caching valid 292 * mappings in their TLBs. Returns the number of free page slots left. 293 * When out of page slots we must call tlb_flush_mmu(). 294 */ 295 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) 296 { 297 struct mmu_gather_batch *batch; 298 299 VM_BUG_ON(!tlb->end); 300 301 batch = tlb->active; 302 batch->pages[batch->nr++] = page; 303 if (batch->nr == batch->max) { 304 if (!tlb_next_batch(tlb)) 305 return 0; 306 batch = tlb->active; 307 } 308 VM_BUG_ON_PAGE(batch->nr > batch->max, page); 309 310 return batch->max - batch->nr; 311 } 312 313 #endif /* HAVE_GENERIC_MMU_GATHER */ 314 315 #ifdef CONFIG_HAVE_RCU_TABLE_FREE 316 317 /* 318 * See the comment near struct mmu_table_batch. 319 */ 320 321 static void tlb_remove_table_smp_sync(void *arg) 322 { 323 /* Simply deliver the interrupt */ 324 } 325 326 static void tlb_remove_table_one(void *table) 327 { 328 /* 329 * This isn't an RCU grace period and hence the page-tables cannot be 330 * assumed to be actually RCU-freed. 331 * 332 * It is however sufficient for software page-table walkers that rely on 333 * IRQ disabling. See the comment near struct mmu_table_batch. 334 */ 335 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 336 __tlb_remove_table(table); 337 } 338 339 static void tlb_remove_table_rcu(struct rcu_head *head) 340 { 341 struct mmu_table_batch *batch; 342 int i; 343 344 batch = container_of(head, struct mmu_table_batch, rcu); 345 346 for (i = 0; i < batch->nr; i++) 347 __tlb_remove_table(batch->tables[i]); 348 349 free_page((unsigned long)batch); 350 } 351 352 void tlb_table_flush(struct mmu_gather *tlb) 353 { 354 struct mmu_table_batch **batch = &tlb->batch; 355 356 if (*batch) { 357 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 358 *batch = NULL; 359 } 360 } 361 362 void tlb_remove_table(struct mmu_gather *tlb, void *table) 363 { 364 struct mmu_table_batch **batch = &tlb->batch; 365 366 /* 367 * When there's less then two users of this mm there cannot be a 368 * concurrent page-table walk. 369 */ 370 if (atomic_read(&tlb->mm->mm_users) < 2) { 371 __tlb_remove_table(table); 372 return; 373 } 374 375 if (*batch == NULL) { 376 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); 377 if (*batch == NULL) { 378 tlb_remove_table_one(table); 379 return; 380 } 381 (*batch)->nr = 0; 382 } 383 (*batch)->tables[(*batch)->nr++] = table; 384 if ((*batch)->nr == MAX_TABLE_BATCH) 385 tlb_table_flush(tlb); 386 } 387 388 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ 389 390 /* 391 * Note: this doesn't free the actual pages themselves. That 392 * has been handled earlier when unmapping all the memory regions. 393 */ 394 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 395 unsigned long addr) 396 { 397 pgtable_t token = pmd_pgtable(*pmd); 398 pmd_clear(pmd); 399 pte_free_tlb(tlb, token, addr); 400 atomic_long_dec(&tlb->mm->nr_ptes); 401 } 402 403 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 404 unsigned long addr, unsigned long end, 405 unsigned long floor, unsigned long ceiling) 406 { 407 pmd_t *pmd; 408 unsigned long next; 409 unsigned long start; 410 411 start = addr; 412 pmd = pmd_offset(pud, addr); 413 do { 414 next = pmd_addr_end(addr, end); 415 if (pmd_none_or_clear_bad(pmd)) 416 continue; 417 free_pte_range(tlb, pmd, addr); 418 } while (pmd++, addr = next, addr != end); 419 420 start &= PUD_MASK; 421 if (start < floor) 422 return; 423 if (ceiling) { 424 ceiling &= PUD_MASK; 425 if (!ceiling) 426 return; 427 } 428 if (end - 1 > ceiling - 1) 429 return; 430 431 pmd = pmd_offset(pud, start); 432 pud_clear(pud); 433 pmd_free_tlb(tlb, pmd, start); 434 mm_dec_nr_pmds(tlb->mm); 435 } 436 437 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 438 unsigned long addr, unsigned long end, 439 unsigned long floor, unsigned long ceiling) 440 { 441 pud_t *pud; 442 unsigned long next; 443 unsigned long start; 444 445 start = addr; 446 pud = pud_offset(pgd, addr); 447 do { 448 next = pud_addr_end(addr, end); 449 if (pud_none_or_clear_bad(pud)) 450 continue; 451 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 452 } while (pud++, addr = next, addr != end); 453 454 start &= PGDIR_MASK; 455 if (start < floor) 456 return; 457 if (ceiling) { 458 ceiling &= PGDIR_MASK; 459 if (!ceiling) 460 return; 461 } 462 if (end - 1 > ceiling - 1) 463 return; 464 465 pud = pud_offset(pgd, start); 466 pgd_clear(pgd); 467 pud_free_tlb(tlb, pud, start); 468 } 469 470 /* 471 * This function frees user-level page tables of a process. 472 */ 473 void free_pgd_range(struct mmu_gather *tlb, 474 unsigned long addr, unsigned long end, 475 unsigned long floor, unsigned long ceiling) 476 { 477 pgd_t *pgd; 478 unsigned long next; 479 480 /* 481 * The next few lines have given us lots of grief... 482 * 483 * Why are we testing PMD* at this top level? Because often 484 * there will be no work to do at all, and we'd prefer not to 485 * go all the way down to the bottom just to discover that. 486 * 487 * Why all these "- 1"s? Because 0 represents both the bottom 488 * of the address space and the top of it (using -1 for the 489 * top wouldn't help much: the masks would do the wrong thing). 490 * The rule is that addr 0 and floor 0 refer to the bottom of 491 * the address space, but end 0 and ceiling 0 refer to the top 492 * Comparisons need to use "end - 1" and "ceiling - 1" (though 493 * that end 0 case should be mythical). 494 * 495 * Wherever addr is brought up or ceiling brought down, we must 496 * be careful to reject "the opposite 0" before it confuses the 497 * subsequent tests. But what about where end is brought down 498 * by PMD_SIZE below? no, end can't go down to 0 there. 499 * 500 * Whereas we round start (addr) and ceiling down, by different 501 * masks at different levels, in order to test whether a table 502 * now has no other vmas using it, so can be freed, we don't 503 * bother to round floor or end up - the tests don't need that. 504 */ 505 506 addr &= PMD_MASK; 507 if (addr < floor) { 508 addr += PMD_SIZE; 509 if (!addr) 510 return; 511 } 512 if (ceiling) { 513 ceiling &= PMD_MASK; 514 if (!ceiling) 515 return; 516 } 517 if (end - 1 > ceiling - 1) 518 end -= PMD_SIZE; 519 if (addr > end - 1) 520 return; 521 522 pgd = pgd_offset(tlb->mm, addr); 523 do { 524 next = pgd_addr_end(addr, end); 525 if (pgd_none_or_clear_bad(pgd)) 526 continue; 527 free_pud_range(tlb, pgd, addr, next, floor, ceiling); 528 } while (pgd++, addr = next, addr != end); 529 } 530 531 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 532 unsigned long floor, unsigned long ceiling) 533 { 534 while (vma) { 535 struct vm_area_struct *next = vma->vm_next; 536 unsigned long addr = vma->vm_start; 537 538 /* 539 * Hide vma from rmap and truncate_pagecache before freeing 540 * pgtables 541 */ 542 unlink_anon_vmas(vma); 543 unlink_file_vma(vma); 544 545 if (is_vm_hugetlb_page(vma)) { 546 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 547 floor, next? next->vm_start: ceiling); 548 } else { 549 /* 550 * Optimization: gather nearby vmas into one call down 551 */ 552 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 553 && !is_vm_hugetlb_page(next)) { 554 vma = next; 555 next = vma->vm_next; 556 unlink_anon_vmas(vma); 557 unlink_file_vma(vma); 558 } 559 free_pgd_range(tlb, addr, vma->vm_end, 560 floor, next? next->vm_start: ceiling); 561 } 562 vma = next; 563 } 564 } 565 566 int __pte_alloc(struct mm_struct *mm, 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 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 665 nr_unshown); 666 nr_unshown = 0; 667 } 668 nr_shown = 0; 669 } 670 if (nr_shown++ == 0) 671 resume = jiffies + 60 * HZ; 672 673 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 674 index = linear_page_index(vma, addr); 675 676 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 677 current->comm, 678 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 679 if (page) 680 dump_page(page, "bad pte"); 681 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", 682 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 683 /* 684 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y 685 */ 686 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n", 687 vma->vm_file, 688 vma->vm_ops ? vma->vm_ops->fault : NULL, 689 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 690 mapping ? mapping->a_ops->readpage : NULL); 691 dump_stack(); 692 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 693 } 694 695 /* 696 * vm_normal_page -- This function gets the "struct page" associated with a pte. 697 * 698 * "Special" mappings do not wish to be associated with a "struct page" (either 699 * it doesn't exist, or it exists but they don't want to touch it). In this 700 * case, NULL is returned here. "Normal" mappings do have a struct page. 701 * 702 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 703 * pte bit, in which case this function is trivial. Secondly, an architecture 704 * may not have a spare pte bit, which requires a more complicated scheme, 705 * described below. 706 * 707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 708 * special mapping (even if there are underlying and valid "struct pages"). 709 * COWed pages of a VM_PFNMAP are always normal. 710 * 711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 714 * mapping will always honor the rule 715 * 716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 717 * 718 * And for normal mappings this is false. 719 * 720 * This restricts such mappings to be a linear translation from virtual address 721 * to pfn. To get around this restriction, we allow arbitrary mappings so long 722 * as the vma is not a COW mapping; in that case, we know that all ptes are 723 * special (because none can have been COWed). 724 * 725 * 726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 727 * 728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 729 * page" backing, however the difference is that _all_ pages with a struct 730 * page (that is, those where pfn_valid is true) are refcounted and considered 731 * normal pages by the VM. The disadvantage is that pages are refcounted 732 * (which can be slower and simply not an option for some PFNMAP users). The 733 * advantage is that we don't have to follow the strict linearity rule of 734 * PFNMAP mappings in order to support COWable mappings. 735 * 736 */ 737 #ifdef __HAVE_ARCH_PTE_SPECIAL 738 # define HAVE_PTE_SPECIAL 1 739 #else 740 # define HAVE_PTE_SPECIAL 0 741 #endif 742 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 743 pte_t pte) 744 { 745 unsigned long pfn = pte_pfn(pte); 746 747 if (HAVE_PTE_SPECIAL) { 748 if (likely(!pte_special(pte))) 749 goto check_pfn; 750 if (vma->vm_ops && vma->vm_ops->find_special_page) 751 return vma->vm_ops->find_special_page(vma, addr); 752 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 753 return NULL; 754 if (!is_zero_pfn(pfn)) 755 print_bad_pte(vma, addr, pte, NULL); 756 return NULL; 757 } 758 759 /* !HAVE_PTE_SPECIAL case follows: */ 760 761 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 762 if (vma->vm_flags & VM_MIXEDMAP) { 763 if (!pfn_valid(pfn)) 764 return NULL; 765 goto out; 766 } else { 767 unsigned long off; 768 off = (addr - vma->vm_start) >> PAGE_SHIFT; 769 if (pfn == vma->vm_pgoff + off) 770 return NULL; 771 if (!is_cow_mapping(vma->vm_flags)) 772 return NULL; 773 } 774 } 775 776 if (is_zero_pfn(pfn)) 777 return NULL; 778 check_pfn: 779 if (unlikely(pfn > highest_memmap_pfn)) { 780 print_bad_pte(vma, addr, pte, NULL); 781 return NULL; 782 } 783 784 /* 785 * NOTE! We still have PageReserved() pages in the page tables. 786 * eg. VDSO mappings can cause them to exist. 787 */ 788 out: 789 return pfn_to_page(pfn); 790 } 791 792 /* 793 * copy one vm_area from one task to the other. Assumes the page tables 794 * already present in the new task to be cleared in the whole range 795 * covered by this vma. 796 */ 797 798 static inline unsigned long 799 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 800 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 801 unsigned long addr, int *rss) 802 { 803 unsigned long vm_flags = vma->vm_flags; 804 pte_t pte = *src_pte; 805 struct page *page; 806 807 /* pte contains position in swap or file, so copy. */ 808 if (unlikely(!pte_present(pte))) { 809 swp_entry_t entry = pte_to_swp_entry(pte); 810 811 if (likely(!non_swap_entry(entry))) { 812 if (swap_duplicate(entry) < 0) 813 return entry.val; 814 815 /* make sure dst_mm is on swapoff's mmlist. */ 816 if (unlikely(list_empty(&dst_mm->mmlist))) { 817 spin_lock(&mmlist_lock); 818 if (list_empty(&dst_mm->mmlist)) 819 list_add(&dst_mm->mmlist, 820 &src_mm->mmlist); 821 spin_unlock(&mmlist_lock); 822 } 823 rss[MM_SWAPENTS]++; 824 } else if (is_migration_entry(entry)) { 825 page = migration_entry_to_page(entry); 826 827 rss[mm_counter(page)]++; 828 829 if (is_write_migration_entry(entry) && 830 is_cow_mapping(vm_flags)) { 831 /* 832 * COW mappings require pages in both 833 * parent and child to be set to read. 834 */ 835 make_migration_entry_read(&entry); 836 pte = swp_entry_to_pte(entry); 837 if (pte_swp_soft_dirty(*src_pte)) 838 pte = pte_swp_mksoft_dirty(pte); 839 set_pte_at(src_mm, addr, src_pte, pte); 840 } 841 } 842 goto out_set_pte; 843 } 844 845 /* 846 * If it's a COW mapping, write protect it both 847 * in the parent and the child 848 */ 849 if (is_cow_mapping(vm_flags)) { 850 ptep_set_wrprotect(src_mm, addr, src_pte); 851 pte = pte_wrprotect(pte); 852 } 853 854 /* 855 * If it's a shared mapping, mark it clean in 856 * the child 857 */ 858 if (vm_flags & VM_SHARED) 859 pte = pte_mkclean(pte); 860 pte = pte_mkold(pte); 861 862 page = vm_normal_page(vma, addr, pte); 863 if (page) { 864 get_page(page); 865 page_dup_rmap(page, false); 866 rss[mm_counter(page)]++; 867 } 868 869 out_set_pte: 870 set_pte_at(dst_mm, addr, dst_pte, pte); 871 return 0; 872 } 873 874 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 875 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 876 unsigned long addr, unsigned long end) 877 { 878 pte_t *orig_src_pte, *orig_dst_pte; 879 pte_t *src_pte, *dst_pte; 880 spinlock_t *src_ptl, *dst_ptl; 881 int progress = 0; 882 int rss[NR_MM_COUNTERS]; 883 swp_entry_t entry = (swp_entry_t){0}; 884 885 again: 886 init_rss_vec(rss); 887 888 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 889 if (!dst_pte) 890 return -ENOMEM; 891 src_pte = pte_offset_map(src_pmd, addr); 892 src_ptl = pte_lockptr(src_mm, src_pmd); 893 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 894 orig_src_pte = src_pte; 895 orig_dst_pte = dst_pte; 896 arch_enter_lazy_mmu_mode(); 897 898 do { 899 /* 900 * We are holding two locks at this point - either of them 901 * could generate latencies in another task on another CPU. 902 */ 903 if (progress >= 32) { 904 progress = 0; 905 if (need_resched() || 906 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 907 break; 908 } 909 if (pte_none(*src_pte)) { 910 progress++; 911 continue; 912 } 913 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, 914 vma, addr, rss); 915 if (entry.val) 916 break; 917 progress += 8; 918 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 919 920 arch_leave_lazy_mmu_mode(); 921 spin_unlock(src_ptl); 922 pte_unmap(orig_src_pte); 923 add_mm_rss_vec(dst_mm, rss); 924 pte_unmap_unlock(orig_dst_pte, dst_ptl); 925 cond_resched(); 926 927 if (entry.val) { 928 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) 929 return -ENOMEM; 930 progress = 0; 931 } 932 if (addr != end) 933 goto again; 934 return 0; 935 } 936 937 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 938 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 939 unsigned long addr, unsigned long end) 940 { 941 pmd_t *src_pmd, *dst_pmd; 942 unsigned long next; 943 944 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 945 if (!dst_pmd) 946 return -ENOMEM; 947 src_pmd = pmd_offset(src_pud, addr); 948 do { 949 next = pmd_addr_end(addr, end); 950 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) { 951 int err; 952 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE); 953 err = copy_huge_pmd(dst_mm, src_mm, 954 dst_pmd, src_pmd, addr, vma); 955 if (err == -ENOMEM) 956 return -ENOMEM; 957 if (!err) 958 continue; 959 /* fall through */ 960 } 961 if (pmd_none_or_clear_bad(src_pmd)) 962 continue; 963 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 964 vma, addr, next)) 965 return -ENOMEM; 966 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 967 return 0; 968 } 969 970 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 971 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 972 unsigned long addr, unsigned long end) 973 { 974 pud_t *src_pud, *dst_pud; 975 unsigned long next; 976 977 dst_pud = pud_alloc(dst_mm, dst_pgd, addr); 978 if (!dst_pud) 979 return -ENOMEM; 980 src_pud = pud_offset(src_pgd, addr); 981 do { 982 next = pud_addr_end(addr, end); 983 if (pud_none_or_clear_bad(src_pud)) 984 continue; 985 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 986 vma, addr, next)) 987 return -ENOMEM; 988 } while (dst_pud++, src_pud++, addr = next, addr != end); 989 return 0; 990 } 991 992 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 993 struct vm_area_struct *vma) 994 { 995 pgd_t *src_pgd, *dst_pgd; 996 unsigned long next; 997 unsigned long addr = vma->vm_start; 998 unsigned long end = vma->vm_end; 999 unsigned long mmun_start; /* For mmu_notifiers */ 1000 unsigned long mmun_end; /* For mmu_notifiers */ 1001 bool is_cow; 1002 int ret; 1003 1004 /* 1005 * Don't copy ptes where a page fault will fill them correctly. 1006 * Fork becomes much lighter when there are big shared or private 1007 * readonly mappings. The tradeoff is that copy_page_range is more 1008 * efficient than faulting. 1009 */ 1010 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 1011 !vma->anon_vma) 1012 return 0; 1013 1014 if (is_vm_hugetlb_page(vma)) 1015 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 1016 1017 if (unlikely(vma->vm_flags & VM_PFNMAP)) { 1018 /* 1019 * We do not free on error cases below as remove_vma 1020 * gets called on error from higher level routine 1021 */ 1022 ret = track_pfn_copy(vma); 1023 if (ret) 1024 return ret; 1025 } 1026 1027 /* 1028 * We need to invalidate the secondary MMU mappings only when 1029 * there could be a permission downgrade on the ptes of the 1030 * parent mm. And a permission downgrade will only happen if 1031 * is_cow_mapping() returns true. 1032 */ 1033 is_cow = is_cow_mapping(vma->vm_flags); 1034 mmun_start = addr; 1035 mmun_end = end; 1036 if (is_cow) 1037 mmu_notifier_invalidate_range_start(src_mm, mmun_start, 1038 mmun_end); 1039 1040 ret = 0; 1041 dst_pgd = pgd_offset(dst_mm, addr); 1042 src_pgd = pgd_offset(src_mm, addr); 1043 do { 1044 next = pgd_addr_end(addr, end); 1045 if (pgd_none_or_clear_bad(src_pgd)) 1046 continue; 1047 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, 1048 vma, addr, next))) { 1049 ret = -ENOMEM; 1050 break; 1051 } 1052 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1053 1054 if (is_cow) 1055 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end); 1056 return ret; 1057 } 1058 1059 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1060 struct vm_area_struct *vma, pmd_t *pmd, 1061 unsigned long addr, unsigned long end, 1062 struct zap_details *details) 1063 { 1064 struct mm_struct *mm = tlb->mm; 1065 int force_flush = 0; 1066 int rss[NR_MM_COUNTERS]; 1067 spinlock_t *ptl; 1068 pte_t *start_pte; 1069 pte_t *pte; 1070 swp_entry_t entry; 1071 1072 again: 1073 init_rss_vec(rss); 1074 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1075 pte = start_pte; 1076 arch_enter_lazy_mmu_mode(); 1077 do { 1078 pte_t ptent = *pte; 1079 if (pte_none(ptent)) { 1080 continue; 1081 } 1082 1083 if (pte_present(ptent)) { 1084 struct page *page; 1085 1086 page = vm_normal_page(vma, addr, ptent); 1087 if (unlikely(details) && page) { 1088 /* 1089 * unmap_shared_mapping_pages() wants to 1090 * invalidate cache without truncating: 1091 * unmap shared but keep private pages. 1092 */ 1093 if (details->check_mapping && 1094 details->check_mapping != page->mapping) 1095 continue; 1096 } 1097 ptent = ptep_get_and_clear_full(mm, addr, pte, 1098 tlb->fullmm); 1099 tlb_remove_tlb_entry(tlb, pte, addr); 1100 if (unlikely(!page)) 1101 continue; 1102 1103 if (!PageAnon(page)) { 1104 if (pte_dirty(ptent)) { 1105 /* 1106 * oom_reaper cannot tear down dirty 1107 * pages 1108 */ 1109 if (unlikely(details && details->ignore_dirty)) 1110 continue; 1111 force_flush = 1; 1112 set_page_dirty(page); 1113 } 1114 if (pte_young(ptent) && 1115 likely(!(vma->vm_flags & VM_SEQ_READ))) 1116 mark_page_accessed(page); 1117 } 1118 rss[mm_counter(page)]--; 1119 page_remove_rmap(page, false); 1120 if (unlikely(page_mapcount(page) < 0)) 1121 print_bad_pte(vma, addr, ptent, page); 1122 if (unlikely(!__tlb_remove_page(tlb, page))) { 1123 force_flush = 1; 1124 addr += PAGE_SIZE; 1125 break; 1126 } 1127 continue; 1128 } 1129 /* only check swap_entries if explicitly asked for in details */ 1130 if (unlikely(details && !details->check_swap_entries)) 1131 continue; 1132 1133 entry = pte_to_swp_entry(ptent); 1134 if (!non_swap_entry(entry)) 1135 rss[MM_SWAPENTS]--; 1136 else if (is_migration_entry(entry)) { 1137 struct page *page; 1138 1139 page = migration_entry_to_page(entry); 1140 rss[mm_counter(page)]--; 1141 } 1142 if (unlikely(!free_swap_and_cache(entry))) 1143 print_bad_pte(vma, addr, ptent, NULL); 1144 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1145 } while (pte++, addr += PAGE_SIZE, addr != end); 1146 1147 add_mm_rss_vec(mm, rss); 1148 arch_leave_lazy_mmu_mode(); 1149 1150 /* Do the actual TLB flush before dropping ptl */ 1151 if (force_flush) 1152 tlb_flush_mmu_tlbonly(tlb); 1153 pte_unmap_unlock(start_pte, ptl); 1154 1155 /* 1156 * If we forced a TLB flush (either due to running out of 1157 * batch buffers or because we needed to flush dirty TLB 1158 * entries before releasing the ptl), free the batched 1159 * memory too. Restart if we didn't do everything. 1160 */ 1161 if (force_flush) { 1162 force_flush = 0; 1163 tlb_flush_mmu_free(tlb); 1164 1165 if (addr != end) 1166 goto again; 1167 } 1168 1169 return addr; 1170 } 1171 1172 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1173 struct vm_area_struct *vma, pud_t *pud, 1174 unsigned long addr, unsigned long end, 1175 struct zap_details *details) 1176 { 1177 pmd_t *pmd; 1178 unsigned long next; 1179 1180 pmd = pmd_offset(pud, addr); 1181 do { 1182 next = pmd_addr_end(addr, end); 1183 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1184 if (next - addr != HPAGE_PMD_SIZE) { 1185 #ifdef CONFIG_DEBUG_VM 1186 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) { 1187 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n", 1188 __func__, addr, end, 1189 vma->vm_start, 1190 vma->vm_end); 1191 BUG(); 1192 } 1193 #endif 1194 split_huge_pmd(vma, pmd, addr); 1195 } else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1196 goto next; 1197 /* fall through */ 1198 } 1199 /* 1200 * Here there can be other concurrent MADV_DONTNEED or 1201 * trans huge page faults running, and if the pmd is 1202 * none or trans huge it can change under us. This is 1203 * because MADV_DONTNEED holds the mmap_sem in read 1204 * mode. 1205 */ 1206 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1207 goto next; 1208 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1209 next: 1210 cond_resched(); 1211 } while (pmd++, addr = next, addr != end); 1212 1213 return addr; 1214 } 1215 1216 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1217 struct vm_area_struct *vma, pgd_t *pgd, 1218 unsigned long addr, unsigned long end, 1219 struct zap_details *details) 1220 { 1221 pud_t *pud; 1222 unsigned long next; 1223 1224 pud = pud_offset(pgd, addr); 1225 do { 1226 next = pud_addr_end(addr, end); 1227 if (pud_none_or_clear_bad(pud)) 1228 continue; 1229 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1230 } while (pud++, addr = next, addr != end); 1231 1232 return addr; 1233 } 1234 1235 void unmap_page_range(struct mmu_gather *tlb, 1236 struct vm_area_struct *vma, 1237 unsigned long addr, unsigned long end, 1238 struct zap_details *details) 1239 { 1240 pgd_t *pgd; 1241 unsigned long next; 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_devmap(pfn) && pfn_t_valid(pfn)) { 1616 struct page *page; 1617 1618 /* 1619 * At this point we are committed to insert_page() 1620 * regardless of whether the caller specified flags that 1621 * result in pfn_t_has_page() == false. 1622 */ 1623 page = pfn_to_page(pfn_t_to_pfn(pfn)); 1624 return insert_page(vma, addr, page, vma->vm_page_prot); 1625 } 1626 return insert_pfn(vma, addr, pfn, vma->vm_page_prot); 1627 } 1628 EXPORT_SYMBOL(vm_insert_mixed); 1629 1630 /* 1631 * maps a range of physical memory into the requested pages. the old 1632 * mappings are removed. any references to nonexistent pages results 1633 * in null mappings (currently treated as "copy-on-access") 1634 */ 1635 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1636 unsigned long addr, unsigned long end, 1637 unsigned long pfn, pgprot_t prot) 1638 { 1639 pte_t *pte; 1640 spinlock_t *ptl; 1641 1642 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1643 if (!pte) 1644 return -ENOMEM; 1645 arch_enter_lazy_mmu_mode(); 1646 do { 1647 BUG_ON(!pte_none(*pte)); 1648 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 1649 pfn++; 1650 } while (pte++, addr += PAGE_SIZE, addr != end); 1651 arch_leave_lazy_mmu_mode(); 1652 pte_unmap_unlock(pte - 1, ptl); 1653 return 0; 1654 } 1655 1656 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 1657 unsigned long addr, unsigned long end, 1658 unsigned long pfn, pgprot_t prot) 1659 { 1660 pmd_t *pmd; 1661 unsigned long next; 1662 1663 pfn -= addr >> PAGE_SHIFT; 1664 pmd = pmd_alloc(mm, pud, addr); 1665 if (!pmd) 1666 return -ENOMEM; 1667 VM_BUG_ON(pmd_trans_huge(*pmd)); 1668 do { 1669 next = pmd_addr_end(addr, end); 1670 if (remap_pte_range(mm, pmd, addr, next, 1671 pfn + (addr >> PAGE_SHIFT), prot)) 1672 return -ENOMEM; 1673 } while (pmd++, addr = next, addr != end); 1674 return 0; 1675 } 1676 1677 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, 1678 unsigned long addr, unsigned long end, 1679 unsigned long pfn, pgprot_t prot) 1680 { 1681 pud_t *pud; 1682 unsigned long next; 1683 1684 pfn -= addr >> PAGE_SHIFT; 1685 pud = pud_alloc(mm, pgd, addr); 1686 if (!pud) 1687 return -ENOMEM; 1688 do { 1689 next = pud_addr_end(addr, end); 1690 if (remap_pmd_range(mm, pud, addr, next, 1691 pfn + (addr >> PAGE_SHIFT), prot)) 1692 return -ENOMEM; 1693 } while (pud++, addr = next, addr != end); 1694 return 0; 1695 } 1696 1697 /** 1698 * remap_pfn_range - remap kernel memory to userspace 1699 * @vma: user vma to map to 1700 * @addr: target user address to start at 1701 * @pfn: physical address of kernel memory 1702 * @size: size of map area 1703 * @prot: page protection flags for this mapping 1704 * 1705 * Note: this is only safe if the mm semaphore is held when called. 1706 */ 1707 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 1708 unsigned long pfn, unsigned long size, pgprot_t prot) 1709 { 1710 pgd_t *pgd; 1711 unsigned long next; 1712 unsigned long end = addr + PAGE_ALIGN(size); 1713 struct mm_struct *mm = vma->vm_mm; 1714 int err; 1715 1716 /* 1717 * Physically remapped pages are special. Tell the 1718 * rest of the world about it: 1719 * VM_IO tells people not to look at these pages 1720 * (accesses can have side effects). 1721 * VM_PFNMAP tells the core MM that the base pages are just 1722 * raw PFN mappings, and do not have a "struct page" associated 1723 * with them. 1724 * VM_DONTEXPAND 1725 * Disable vma merging and expanding with mremap(). 1726 * VM_DONTDUMP 1727 * Omit vma from core dump, even when VM_IO turned off. 1728 * 1729 * There's a horrible special case to handle copy-on-write 1730 * behaviour that some programs depend on. We mark the "original" 1731 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 1732 * See vm_normal_page() for details. 1733 */ 1734 if (is_cow_mapping(vma->vm_flags)) { 1735 if (addr != vma->vm_start || end != vma->vm_end) 1736 return -EINVAL; 1737 vma->vm_pgoff = pfn; 1738 } 1739 1740 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 1741 if (err) 1742 return -EINVAL; 1743 1744 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 1745 1746 BUG_ON(addr >= end); 1747 pfn -= addr >> PAGE_SHIFT; 1748 pgd = pgd_offset(mm, addr); 1749 flush_cache_range(vma, addr, end); 1750 do { 1751 next = pgd_addr_end(addr, end); 1752 err = remap_pud_range(mm, pgd, addr, next, 1753 pfn + (addr >> PAGE_SHIFT), prot); 1754 if (err) 1755 break; 1756 } while (pgd++, addr = next, addr != end); 1757 1758 if (err) 1759 untrack_pfn(vma, pfn, PAGE_ALIGN(size)); 1760 1761 return err; 1762 } 1763 EXPORT_SYMBOL(remap_pfn_range); 1764 1765 /** 1766 * vm_iomap_memory - remap memory to userspace 1767 * @vma: user vma to map to 1768 * @start: start of area 1769 * @len: size of area 1770 * 1771 * This is a simplified io_remap_pfn_range() for common driver use. The 1772 * driver just needs to give us the physical memory range to be mapped, 1773 * we'll figure out the rest from the vma information. 1774 * 1775 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 1776 * whatever write-combining details or similar. 1777 */ 1778 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 1779 { 1780 unsigned long vm_len, pfn, pages; 1781 1782 /* Check that the physical memory area passed in looks valid */ 1783 if (start + len < start) 1784 return -EINVAL; 1785 /* 1786 * You *really* shouldn't map things that aren't page-aligned, 1787 * but we've historically allowed it because IO memory might 1788 * just have smaller alignment. 1789 */ 1790 len += start & ~PAGE_MASK; 1791 pfn = start >> PAGE_SHIFT; 1792 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 1793 if (pfn + pages < pfn) 1794 return -EINVAL; 1795 1796 /* We start the mapping 'vm_pgoff' pages into the area */ 1797 if (vma->vm_pgoff > pages) 1798 return -EINVAL; 1799 pfn += vma->vm_pgoff; 1800 pages -= vma->vm_pgoff; 1801 1802 /* Can we fit all of the mapping? */ 1803 vm_len = vma->vm_end - vma->vm_start; 1804 if (vm_len >> PAGE_SHIFT > pages) 1805 return -EINVAL; 1806 1807 /* Ok, let it rip */ 1808 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 1809 } 1810 EXPORT_SYMBOL(vm_iomap_memory); 1811 1812 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 1813 unsigned long addr, unsigned long end, 1814 pte_fn_t fn, void *data) 1815 { 1816 pte_t *pte; 1817 int err; 1818 pgtable_t token; 1819 spinlock_t *uninitialized_var(ptl); 1820 1821 pte = (mm == &init_mm) ? 1822 pte_alloc_kernel(pmd, addr) : 1823 pte_alloc_map_lock(mm, pmd, addr, &ptl); 1824 if (!pte) 1825 return -ENOMEM; 1826 1827 BUG_ON(pmd_huge(*pmd)); 1828 1829 arch_enter_lazy_mmu_mode(); 1830 1831 token = pmd_pgtable(*pmd); 1832 1833 do { 1834 err = fn(pte++, token, addr, data); 1835 if (err) 1836 break; 1837 } while (addr += PAGE_SIZE, addr != end); 1838 1839 arch_leave_lazy_mmu_mode(); 1840 1841 if (mm != &init_mm) 1842 pte_unmap_unlock(pte-1, ptl); 1843 return err; 1844 } 1845 1846 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 1847 unsigned long addr, unsigned long end, 1848 pte_fn_t fn, void *data) 1849 { 1850 pmd_t *pmd; 1851 unsigned long next; 1852 int err; 1853 1854 BUG_ON(pud_huge(*pud)); 1855 1856 pmd = pmd_alloc(mm, pud, addr); 1857 if (!pmd) 1858 return -ENOMEM; 1859 do { 1860 next = pmd_addr_end(addr, end); 1861 err = apply_to_pte_range(mm, pmd, addr, next, fn, data); 1862 if (err) 1863 break; 1864 } while (pmd++, addr = next, addr != end); 1865 return err; 1866 } 1867 1868 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, 1869 unsigned long addr, unsigned long end, 1870 pte_fn_t fn, void *data) 1871 { 1872 pud_t *pud; 1873 unsigned long next; 1874 int err; 1875 1876 pud = pud_alloc(mm, pgd, addr); 1877 if (!pud) 1878 return -ENOMEM; 1879 do { 1880 next = pud_addr_end(addr, end); 1881 err = apply_to_pmd_range(mm, pud, addr, next, fn, data); 1882 if (err) 1883 break; 1884 } while (pud++, addr = next, addr != end); 1885 return err; 1886 } 1887 1888 /* 1889 * Scan a region of virtual memory, filling in page tables as necessary 1890 * and calling a provided function on each leaf page table. 1891 */ 1892 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 1893 unsigned long size, pte_fn_t fn, void *data) 1894 { 1895 pgd_t *pgd; 1896 unsigned long next; 1897 unsigned long end = addr + size; 1898 int err; 1899 1900 if (WARN_ON(addr >= end)) 1901 return -EINVAL; 1902 1903 pgd = pgd_offset(mm, addr); 1904 do { 1905 next = pgd_addr_end(addr, end); 1906 err = apply_to_pud_range(mm, pgd, addr, next, fn, data); 1907 if (err) 1908 break; 1909 } while (pgd++, addr = next, addr != end); 1910 1911 return err; 1912 } 1913 EXPORT_SYMBOL_GPL(apply_to_page_range); 1914 1915 /* 1916 * handle_pte_fault chooses page fault handler according to an entry which was 1917 * read non-atomically. Before making any commitment, on those architectures 1918 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 1919 * parts, do_swap_page must check under lock before unmapping the pte and 1920 * proceeding (but do_wp_page is only called after already making such a check; 1921 * and do_anonymous_page can safely check later on). 1922 */ 1923 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 1924 pte_t *page_table, pte_t orig_pte) 1925 { 1926 int same = 1; 1927 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 1928 if (sizeof(pte_t) > sizeof(unsigned long)) { 1929 spinlock_t *ptl = pte_lockptr(mm, pmd); 1930 spin_lock(ptl); 1931 same = pte_same(*page_table, orig_pte); 1932 spin_unlock(ptl); 1933 } 1934 #endif 1935 pte_unmap(page_table); 1936 return same; 1937 } 1938 1939 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) 1940 { 1941 debug_dma_assert_idle(src); 1942 1943 /* 1944 * If the source page was a PFN mapping, we don't have 1945 * a "struct page" for it. We do a best-effort copy by 1946 * just copying from the original user address. If that 1947 * fails, we just zero-fill it. Live with it. 1948 */ 1949 if (unlikely(!src)) { 1950 void *kaddr = kmap_atomic(dst); 1951 void __user *uaddr = (void __user *)(va & PAGE_MASK); 1952 1953 /* 1954 * This really shouldn't fail, because the page is there 1955 * in the page tables. But it might just be unreadable, 1956 * in which case we just give up and fill the result with 1957 * zeroes. 1958 */ 1959 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) 1960 clear_page(kaddr); 1961 kunmap_atomic(kaddr); 1962 flush_dcache_page(dst); 1963 } else 1964 copy_user_highpage(dst, src, va, vma); 1965 } 1966 1967 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 1968 { 1969 struct file *vm_file = vma->vm_file; 1970 1971 if (vm_file) 1972 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 1973 1974 /* 1975 * Special mappings (e.g. VDSO) do not have any file so fake 1976 * a default GFP_KERNEL for them. 1977 */ 1978 return GFP_KERNEL; 1979 } 1980 1981 /* 1982 * Notify the address space that the page is about to become writable so that 1983 * it can prohibit this or wait for the page to get into an appropriate state. 1984 * 1985 * We do this without the lock held, so that it can sleep if it needs to. 1986 */ 1987 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page, 1988 unsigned long address) 1989 { 1990 struct vm_fault vmf; 1991 int ret; 1992 1993 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 1994 vmf.pgoff = page->index; 1995 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 1996 vmf.gfp_mask = __get_fault_gfp_mask(vma); 1997 vmf.page = page; 1998 vmf.cow_page = NULL; 1999 2000 ret = vma->vm_ops->page_mkwrite(vma, &vmf); 2001 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2002 return ret; 2003 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2004 lock_page(page); 2005 if (!page->mapping) { 2006 unlock_page(page); 2007 return 0; /* retry */ 2008 } 2009 ret |= VM_FAULT_LOCKED; 2010 } else 2011 VM_BUG_ON_PAGE(!PageLocked(page), page); 2012 return ret; 2013 } 2014 2015 /* 2016 * Handle write page faults for pages that can be reused in the current vma 2017 * 2018 * This can happen either due to the mapping being with the VM_SHARED flag, 2019 * or due to us being the last reference standing to the page. In either 2020 * case, all we need to do here is to mark the page as writable and update 2021 * any related book-keeping. 2022 */ 2023 static inline int wp_page_reuse(struct mm_struct *mm, 2024 struct vm_area_struct *vma, unsigned long address, 2025 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte, 2026 struct page *page, int page_mkwrite, 2027 int dirty_shared) 2028 __releases(ptl) 2029 { 2030 pte_t entry; 2031 /* 2032 * Clear the pages cpupid information as the existing 2033 * information potentially belongs to a now completely 2034 * unrelated process. 2035 */ 2036 if (page) 2037 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2038 2039 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2040 entry = pte_mkyoung(orig_pte); 2041 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2042 if (ptep_set_access_flags(vma, address, page_table, entry, 1)) 2043 update_mmu_cache(vma, address, page_table); 2044 pte_unmap_unlock(page_table, ptl); 2045 2046 if (dirty_shared) { 2047 struct address_space *mapping; 2048 int dirtied; 2049 2050 if (!page_mkwrite) 2051 lock_page(page); 2052 2053 dirtied = set_page_dirty(page); 2054 VM_BUG_ON_PAGE(PageAnon(page), page); 2055 mapping = page->mapping; 2056 unlock_page(page); 2057 put_page(page); 2058 2059 if ((dirtied || page_mkwrite) && mapping) { 2060 /* 2061 * Some device drivers do not set page.mapping 2062 * but still dirty their pages 2063 */ 2064 balance_dirty_pages_ratelimited(mapping); 2065 } 2066 2067 if (!page_mkwrite) 2068 file_update_time(vma->vm_file); 2069 } 2070 2071 return VM_FAULT_WRITE; 2072 } 2073 2074 /* 2075 * Handle the case of a page which we actually need to copy to a new page. 2076 * 2077 * Called with mmap_sem locked and the old page referenced, but 2078 * without the ptl held. 2079 * 2080 * High level logic flow: 2081 * 2082 * - Allocate a page, copy the content of the old page to the new one. 2083 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2084 * - Take the PTL. If the pte changed, bail out and release the allocated page 2085 * - If the pte is still the way we remember it, update the page table and all 2086 * relevant references. This includes dropping the reference the page-table 2087 * held to the old page, as well as updating the rmap. 2088 * - In any case, unlock the PTL and drop the reference we took to the old page. 2089 */ 2090 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma, 2091 unsigned long address, pte_t *page_table, pmd_t *pmd, 2092 pte_t orig_pte, struct page *old_page) 2093 { 2094 struct page *new_page = NULL; 2095 spinlock_t *ptl = NULL; 2096 pte_t entry; 2097 int page_copied = 0; 2098 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */ 2099 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */ 2100 struct mem_cgroup *memcg; 2101 2102 if (unlikely(anon_vma_prepare(vma))) 2103 goto oom; 2104 2105 if (is_zero_pfn(pte_pfn(orig_pte))) { 2106 new_page = alloc_zeroed_user_highpage_movable(vma, address); 2107 if (!new_page) 2108 goto oom; 2109 } else { 2110 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 2111 if (!new_page) 2112 goto oom; 2113 cow_user_page(new_page, old_page, address, vma); 2114 } 2115 2116 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) 2117 goto oom_free_new; 2118 2119 __SetPageUptodate(new_page); 2120 2121 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2122 2123 /* 2124 * Re-check the pte - we dropped the lock 2125 */ 2126 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2127 if (likely(pte_same(*page_table, orig_pte))) { 2128 if (old_page) { 2129 if (!PageAnon(old_page)) { 2130 dec_mm_counter_fast(mm, 2131 mm_counter_file(old_page)); 2132 inc_mm_counter_fast(mm, MM_ANONPAGES); 2133 } 2134 } else { 2135 inc_mm_counter_fast(mm, MM_ANONPAGES); 2136 } 2137 flush_cache_page(vma, address, pte_pfn(orig_pte)); 2138 entry = mk_pte(new_page, vma->vm_page_prot); 2139 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2140 /* 2141 * Clear the pte entry and flush it first, before updating the 2142 * pte with the new entry. This will avoid a race condition 2143 * seen in the presence of one thread doing SMC and another 2144 * thread doing COW. 2145 */ 2146 ptep_clear_flush_notify(vma, address, page_table); 2147 page_add_new_anon_rmap(new_page, vma, address, false); 2148 mem_cgroup_commit_charge(new_page, memcg, false, false); 2149 lru_cache_add_active_or_unevictable(new_page, vma); 2150 /* 2151 * We call the notify macro here because, when using secondary 2152 * mmu page tables (such as kvm shadow page tables), we want the 2153 * new page to be mapped directly into the secondary page table. 2154 */ 2155 set_pte_at_notify(mm, address, page_table, entry); 2156 update_mmu_cache(vma, address, page_table); 2157 if (old_page) { 2158 /* 2159 * Only after switching the pte to the new page may 2160 * we remove the mapcount here. Otherwise another 2161 * process may come and find the rmap count decremented 2162 * before the pte is switched to the new page, and 2163 * "reuse" the old page writing into it while our pte 2164 * here still points into it and can be read by other 2165 * threads. 2166 * 2167 * The critical issue is to order this 2168 * page_remove_rmap with the ptp_clear_flush above. 2169 * Those stores are ordered by (if nothing else,) 2170 * the barrier present in the atomic_add_negative 2171 * in page_remove_rmap. 2172 * 2173 * Then the TLB flush in ptep_clear_flush ensures that 2174 * no process can access the old page before the 2175 * decremented mapcount is visible. And the old page 2176 * cannot be reused until after the decremented 2177 * mapcount is visible. So transitively, TLBs to 2178 * old page will be flushed before it can be reused. 2179 */ 2180 page_remove_rmap(old_page, false); 2181 } 2182 2183 /* Free the old page.. */ 2184 new_page = old_page; 2185 page_copied = 1; 2186 } else { 2187 mem_cgroup_cancel_charge(new_page, memcg, false); 2188 } 2189 2190 if (new_page) 2191 put_page(new_page); 2192 2193 pte_unmap_unlock(page_table, ptl); 2194 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2195 if (old_page) { 2196 /* 2197 * Don't let another task, with possibly unlocked vma, 2198 * keep the mlocked page. 2199 */ 2200 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 2201 lock_page(old_page); /* LRU manipulation */ 2202 if (PageMlocked(old_page)) 2203 munlock_vma_page(old_page); 2204 unlock_page(old_page); 2205 } 2206 put_page(old_page); 2207 } 2208 return page_copied ? VM_FAULT_WRITE : 0; 2209 oom_free_new: 2210 put_page(new_page); 2211 oom: 2212 if (old_page) 2213 put_page(old_page); 2214 return VM_FAULT_OOM; 2215 } 2216 2217 /* 2218 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 2219 * mapping 2220 */ 2221 static int wp_pfn_shared(struct mm_struct *mm, 2222 struct vm_area_struct *vma, unsigned long address, 2223 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte, 2224 pmd_t *pmd) 2225 { 2226 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 2227 struct vm_fault vmf = { 2228 .page = NULL, 2229 .pgoff = linear_page_index(vma, address), 2230 .virtual_address = (void __user *)(address & PAGE_MASK), 2231 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE, 2232 }; 2233 int ret; 2234 2235 pte_unmap_unlock(page_table, ptl); 2236 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf); 2237 if (ret & VM_FAULT_ERROR) 2238 return ret; 2239 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2240 /* 2241 * We might have raced with another page fault while we 2242 * released the pte_offset_map_lock. 2243 */ 2244 if (!pte_same(*page_table, orig_pte)) { 2245 pte_unmap_unlock(page_table, ptl); 2246 return 0; 2247 } 2248 } 2249 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte, 2250 NULL, 0, 0); 2251 } 2252 2253 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma, 2254 unsigned long address, pte_t *page_table, 2255 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte, 2256 struct page *old_page) 2257 __releases(ptl) 2258 { 2259 int page_mkwrite = 0; 2260 2261 get_page(old_page); 2262 2263 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 2264 int tmp; 2265 2266 pte_unmap_unlock(page_table, ptl); 2267 tmp = do_page_mkwrite(vma, old_page, address); 2268 if (unlikely(!tmp || (tmp & 2269 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 2270 put_page(old_page); 2271 return tmp; 2272 } 2273 /* 2274 * Since we dropped the lock we need to revalidate 2275 * the PTE as someone else may have changed it. If 2276 * they did, we just return, as we can count on the 2277 * MMU to tell us if they didn't also make it writable. 2278 */ 2279 page_table = pte_offset_map_lock(mm, pmd, address, 2280 &ptl); 2281 if (!pte_same(*page_table, orig_pte)) { 2282 unlock_page(old_page); 2283 pte_unmap_unlock(page_table, ptl); 2284 put_page(old_page); 2285 return 0; 2286 } 2287 page_mkwrite = 1; 2288 } 2289 2290 return wp_page_reuse(mm, vma, address, page_table, ptl, 2291 orig_pte, old_page, page_mkwrite, 1); 2292 } 2293 2294 /* 2295 * This routine handles present pages, when users try to write 2296 * to a shared page. It is done by copying the page to a new address 2297 * and decrementing the shared-page counter for the old page. 2298 * 2299 * Note that this routine assumes that the protection checks have been 2300 * done by the caller (the low-level page fault routine in most cases). 2301 * Thus we can safely just mark it writable once we've done any necessary 2302 * COW. 2303 * 2304 * We also mark the page dirty at this point even though the page will 2305 * change only once the write actually happens. This avoids a few races, 2306 * and potentially makes it more efficient. 2307 * 2308 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2309 * but allow concurrent faults), with pte both mapped and locked. 2310 * We return with mmap_sem still held, but pte unmapped and unlocked. 2311 */ 2312 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 2313 unsigned long address, pte_t *page_table, pmd_t *pmd, 2314 spinlock_t *ptl, pte_t orig_pte) 2315 __releases(ptl) 2316 { 2317 struct page *old_page; 2318 2319 old_page = vm_normal_page(vma, address, orig_pte); 2320 if (!old_page) { 2321 /* 2322 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 2323 * VM_PFNMAP VMA. 2324 * 2325 * We should not cow pages in a shared writeable mapping. 2326 * Just mark the pages writable and/or call ops->pfn_mkwrite. 2327 */ 2328 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2329 (VM_WRITE|VM_SHARED)) 2330 return wp_pfn_shared(mm, vma, address, page_table, ptl, 2331 orig_pte, pmd); 2332 2333 pte_unmap_unlock(page_table, ptl); 2334 return wp_page_copy(mm, vma, address, page_table, pmd, 2335 orig_pte, old_page); 2336 } 2337 2338 /* 2339 * Take out anonymous pages first, anonymous shared vmas are 2340 * not dirty accountable. 2341 */ 2342 if (PageAnon(old_page) && !PageKsm(old_page)) { 2343 if (!trylock_page(old_page)) { 2344 get_page(old_page); 2345 pte_unmap_unlock(page_table, ptl); 2346 lock_page(old_page); 2347 page_table = pte_offset_map_lock(mm, pmd, address, 2348 &ptl); 2349 if (!pte_same(*page_table, orig_pte)) { 2350 unlock_page(old_page); 2351 pte_unmap_unlock(page_table, ptl); 2352 put_page(old_page); 2353 return 0; 2354 } 2355 put_page(old_page); 2356 } 2357 if (reuse_swap_page(old_page)) { 2358 /* 2359 * The page is all ours. Move it to our anon_vma so 2360 * the rmap code will not search our parent or siblings. 2361 * Protected against the rmap code by the page lock. 2362 */ 2363 page_move_anon_rmap(old_page, vma, address); 2364 unlock_page(old_page); 2365 return wp_page_reuse(mm, vma, address, page_table, ptl, 2366 orig_pte, old_page, 0, 0); 2367 } 2368 unlock_page(old_page); 2369 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2370 (VM_WRITE|VM_SHARED))) { 2371 return wp_page_shared(mm, vma, address, page_table, pmd, 2372 ptl, orig_pte, old_page); 2373 } 2374 2375 /* 2376 * Ok, we need to copy. Oh, well.. 2377 */ 2378 get_page(old_page); 2379 2380 pte_unmap_unlock(page_table, ptl); 2381 return wp_page_copy(mm, vma, address, page_table, pmd, 2382 orig_pte, old_page); 2383 } 2384 2385 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 2386 unsigned long start_addr, unsigned long end_addr, 2387 struct zap_details *details) 2388 { 2389 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 2390 } 2391 2392 static inline void unmap_mapping_range_tree(struct rb_root *root, 2393 struct zap_details *details) 2394 { 2395 struct vm_area_struct *vma; 2396 pgoff_t vba, vea, zba, zea; 2397 2398 vma_interval_tree_foreach(vma, root, 2399 details->first_index, details->last_index) { 2400 2401 vba = vma->vm_pgoff; 2402 vea = vba + vma_pages(vma) - 1; 2403 zba = details->first_index; 2404 if (zba < vba) 2405 zba = vba; 2406 zea = details->last_index; 2407 if (zea > vea) 2408 zea = vea; 2409 2410 unmap_mapping_range_vma(vma, 2411 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 2412 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 2413 details); 2414 } 2415 } 2416 2417 /** 2418 * unmap_mapping_range - unmap the portion of all mmaps in the specified 2419 * address_space corresponding to the specified page range in the underlying 2420 * file. 2421 * 2422 * @mapping: the address space containing mmaps to be unmapped. 2423 * @holebegin: byte in first page to unmap, relative to the start of 2424 * the underlying file. This will be rounded down to a PAGE_SIZE 2425 * boundary. Note that this is different from truncate_pagecache(), which 2426 * must keep the partial page. In contrast, we must get rid of 2427 * partial pages. 2428 * @holelen: size of prospective hole in bytes. This will be rounded 2429 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 2430 * end of the file. 2431 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 2432 * but 0 when invalidating pagecache, don't throw away private data. 2433 */ 2434 void unmap_mapping_range(struct address_space *mapping, 2435 loff_t const holebegin, loff_t const holelen, int even_cows) 2436 { 2437 struct zap_details details = { }; 2438 pgoff_t hba = holebegin >> PAGE_SHIFT; 2439 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2440 2441 /* Check for overflow. */ 2442 if (sizeof(holelen) > sizeof(hlen)) { 2443 long long holeend = 2444 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2445 if (holeend & ~(long long)ULONG_MAX) 2446 hlen = ULONG_MAX - hba + 1; 2447 } 2448 2449 details.check_mapping = even_cows? NULL: mapping; 2450 details.first_index = hba; 2451 details.last_index = hba + hlen - 1; 2452 if (details.last_index < details.first_index) 2453 details.last_index = ULONG_MAX; 2454 2455 2456 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */ 2457 i_mmap_lock_write(mapping); 2458 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap))) 2459 unmap_mapping_range_tree(&mapping->i_mmap, &details); 2460 i_mmap_unlock_write(mapping); 2461 } 2462 EXPORT_SYMBOL(unmap_mapping_range); 2463 2464 /* 2465 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2466 * but allow concurrent faults), and pte mapped but not yet locked. 2467 * We return with pte unmapped and unlocked. 2468 * 2469 * We return with the mmap_sem locked or unlocked in the same cases 2470 * as does filemap_fault(). 2471 */ 2472 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, 2473 unsigned long address, pte_t *page_table, pmd_t *pmd, 2474 unsigned int flags, pte_t orig_pte) 2475 { 2476 spinlock_t *ptl; 2477 struct page *page, *swapcache; 2478 struct mem_cgroup *memcg; 2479 swp_entry_t entry; 2480 pte_t pte; 2481 int locked; 2482 int exclusive = 0; 2483 int ret = 0; 2484 2485 if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) 2486 goto out; 2487 2488 entry = pte_to_swp_entry(orig_pte); 2489 if (unlikely(non_swap_entry(entry))) { 2490 if (is_migration_entry(entry)) { 2491 migration_entry_wait(mm, pmd, address); 2492 } else if (is_hwpoison_entry(entry)) { 2493 ret = VM_FAULT_HWPOISON; 2494 } else { 2495 print_bad_pte(vma, address, orig_pte, NULL); 2496 ret = VM_FAULT_SIGBUS; 2497 } 2498 goto out; 2499 } 2500 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 2501 page = lookup_swap_cache(entry); 2502 if (!page) { 2503 page = swapin_readahead(entry, 2504 GFP_HIGHUSER_MOVABLE, vma, address); 2505 if (!page) { 2506 /* 2507 * Back out if somebody else faulted in this pte 2508 * while we released the pte lock. 2509 */ 2510 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2511 if (likely(pte_same(*page_table, orig_pte))) 2512 ret = VM_FAULT_OOM; 2513 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2514 goto unlock; 2515 } 2516 2517 /* Had to read the page from swap area: Major fault */ 2518 ret = VM_FAULT_MAJOR; 2519 count_vm_event(PGMAJFAULT); 2520 mem_cgroup_count_vm_event(mm, PGMAJFAULT); 2521 } else if (PageHWPoison(page)) { 2522 /* 2523 * hwpoisoned dirty swapcache pages are kept for killing 2524 * owner processes (which may be unknown at hwpoison time) 2525 */ 2526 ret = VM_FAULT_HWPOISON; 2527 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2528 swapcache = page; 2529 goto out_release; 2530 } 2531 2532 swapcache = page; 2533 locked = lock_page_or_retry(page, mm, flags); 2534 2535 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2536 if (!locked) { 2537 ret |= VM_FAULT_RETRY; 2538 goto out_release; 2539 } 2540 2541 /* 2542 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 2543 * release the swapcache from under us. The page pin, and pte_same 2544 * test below, are not enough to exclude that. Even if it is still 2545 * swapcache, we need to check that the page's swap has not changed. 2546 */ 2547 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) 2548 goto out_page; 2549 2550 page = ksm_might_need_to_copy(page, vma, address); 2551 if (unlikely(!page)) { 2552 ret = VM_FAULT_OOM; 2553 page = swapcache; 2554 goto out_page; 2555 } 2556 2557 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) { 2558 ret = VM_FAULT_OOM; 2559 goto out_page; 2560 } 2561 2562 /* 2563 * Back out if somebody else already faulted in this pte. 2564 */ 2565 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2566 if (unlikely(!pte_same(*page_table, orig_pte))) 2567 goto out_nomap; 2568 2569 if (unlikely(!PageUptodate(page))) { 2570 ret = VM_FAULT_SIGBUS; 2571 goto out_nomap; 2572 } 2573 2574 /* 2575 * The page isn't present yet, go ahead with the fault. 2576 * 2577 * Be careful about the sequence of operations here. 2578 * To get its accounting right, reuse_swap_page() must be called 2579 * while the page is counted on swap but not yet in mapcount i.e. 2580 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 2581 * must be called after the swap_free(), or it will never succeed. 2582 */ 2583 2584 inc_mm_counter_fast(mm, MM_ANONPAGES); 2585 dec_mm_counter_fast(mm, MM_SWAPENTS); 2586 pte = mk_pte(page, vma->vm_page_prot); 2587 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { 2588 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 2589 flags &= ~FAULT_FLAG_WRITE; 2590 ret |= VM_FAULT_WRITE; 2591 exclusive = RMAP_EXCLUSIVE; 2592 } 2593 flush_icache_page(vma, page); 2594 if (pte_swp_soft_dirty(orig_pte)) 2595 pte = pte_mksoft_dirty(pte); 2596 set_pte_at(mm, address, page_table, pte); 2597 if (page == swapcache) { 2598 do_page_add_anon_rmap(page, vma, address, exclusive); 2599 mem_cgroup_commit_charge(page, memcg, true, false); 2600 } else { /* ksm created a completely new copy */ 2601 page_add_new_anon_rmap(page, vma, address, false); 2602 mem_cgroup_commit_charge(page, memcg, false, false); 2603 lru_cache_add_active_or_unevictable(page, vma); 2604 } 2605 2606 swap_free(entry); 2607 if (mem_cgroup_swap_full(page) || 2608 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 2609 try_to_free_swap(page); 2610 unlock_page(page); 2611 if (page != swapcache) { 2612 /* 2613 * Hold the lock to avoid the swap entry to be reused 2614 * until we take the PT lock for the pte_same() check 2615 * (to avoid false positives from pte_same). For 2616 * further safety release the lock after the swap_free 2617 * so that the swap count won't change under a 2618 * parallel locked swapcache. 2619 */ 2620 unlock_page(swapcache); 2621 put_page(swapcache); 2622 } 2623 2624 if (flags & FAULT_FLAG_WRITE) { 2625 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); 2626 if (ret & VM_FAULT_ERROR) 2627 ret &= VM_FAULT_ERROR; 2628 goto out; 2629 } 2630 2631 /* No need to invalidate - it was non-present before */ 2632 update_mmu_cache(vma, address, page_table); 2633 unlock: 2634 pte_unmap_unlock(page_table, ptl); 2635 out: 2636 return ret; 2637 out_nomap: 2638 mem_cgroup_cancel_charge(page, memcg, false); 2639 pte_unmap_unlock(page_table, ptl); 2640 out_page: 2641 unlock_page(page); 2642 out_release: 2643 put_page(page); 2644 if (page != swapcache) { 2645 unlock_page(swapcache); 2646 put_page(swapcache); 2647 } 2648 return ret; 2649 } 2650 2651 /* 2652 * This is like a special single-page "expand_{down|up}wards()", 2653 * except we must first make sure that 'address{-|+}PAGE_SIZE' 2654 * doesn't hit another vma. 2655 */ 2656 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) 2657 { 2658 address &= PAGE_MASK; 2659 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { 2660 struct vm_area_struct *prev = vma->vm_prev; 2661 2662 /* 2663 * Is there a mapping abutting this one below? 2664 * 2665 * That's only ok if it's the same stack mapping 2666 * that has gotten split.. 2667 */ 2668 if (prev && prev->vm_end == address) 2669 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; 2670 2671 return expand_downwards(vma, address - PAGE_SIZE); 2672 } 2673 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { 2674 struct vm_area_struct *next = vma->vm_next; 2675 2676 /* As VM_GROWSDOWN but s/below/above/ */ 2677 if (next && next->vm_start == address + PAGE_SIZE) 2678 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; 2679 2680 return expand_upwards(vma, address + PAGE_SIZE); 2681 } 2682 return 0; 2683 } 2684 2685 /* 2686 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2687 * but allow concurrent faults), and pte mapped but not yet locked. 2688 * We return with mmap_sem still held, but pte unmapped and unlocked. 2689 */ 2690 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 2691 unsigned long address, pte_t *page_table, pmd_t *pmd, 2692 unsigned int flags) 2693 { 2694 struct mem_cgroup *memcg; 2695 struct page *page; 2696 spinlock_t *ptl; 2697 pte_t entry; 2698 2699 pte_unmap(page_table); 2700 2701 /* File mapping without ->vm_ops ? */ 2702 if (vma->vm_flags & VM_SHARED) 2703 return VM_FAULT_SIGBUS; 2704 2705 /* Check if we need to add a guard page to the stack */ 2706 if (check_stack_guard_page(vma, address) < 0) 2707 return VM_FAULT_SIGSEGV; 2708 2709 /* Use the zero-page for reads */ 2710 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) { 2711 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), 2712 vma->vm_page_prot)); 2713 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2714 if (!pte_none(*page_table)) 2715 goto unlock; 2716 /* Deliver the page fault to userland, check inside PT lock */ 2717 if (userfaultfd_missing(vma)) { 2718 pte_unmap_unlock(page_table, ptl); 2719 return handle_userfault(vma, address, flags, 2720 VM_UFFD_MISSING); 2721 } 2722 goto setpte; 2723 } 2724 2725 /* Allocate our own private page. */ 2726 if (unlikely(anon_vma_prepare(vma))) 2727 goto oom; 2728 page = alloc_zeroed_user_highpage_movable(vma, address); 2729 if (!page) 2730 goto oom; 2731 2732 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) 2733 goto oom_free_page; 2734 2735 /* 2736 * The memory barrier inside __SetPageUptodate makes sure that 2737 * preceeding stores to the page contents become visible before 2738 * the set_pte_at() write. 2739 */ 2740 __SetPageUptodate(page); 2741 2742 entry = mk_pte(page, vma->vm_page_prot); 2743 if (vma->vm_flags & VM_WRITE) 2744 entry = pte_mkwrite(pte_mkdirty(entry)); 2745 2746 page_table = pte_offset_map_lock(mm, pmd, address, &ptl); 2747 if (!pte_none(*page_table)) 2748 goto release; 2749 2750 /* Deliver the page fault to userland, check inside PT lock */ 2751 if (userfaultfd_missing(vma)) { 2752 pte_unmap_unlock(page_table, ptl); 2753 mem_cgroup_cancel_charge(page, memcg, false); 2754 put_page(page); 2755 return handle_userfault(vma, address, flags, 2756 VM_UFFD_MISSING); 2757 } 2758 2759 inc_mm_counter_fast(mm, MM_ANONPAGES); 2760 page_add_new_anon_rmap(page, vma, address, false); 2761 mem_cgroup_commit_charge(page, memcg, false, false); 2762 lru_cache_add_active_or_unevictable(page, vma); 2763 setpte: 2764 set_pte_at(mm, address, page_table, entry); 2765 2766 /* No need to invalidate - it was non-present before */ 2767 update_mmu_cache(vma, address, page_table); 2768 unlock: 2769 pte_unmap_unlock(page_table, ptl); 2770 return 0; 2771 release: 2772 mem_cgroup_cancel_charge(page, memcg, false); 2773 put_page(page); 2774 goto unlock; 2775 oom_free_page: 2776 put_page(page); 2777 oom: 2778 return VM_FAULT_OOM; 2779 } 2780 2781 /* 2782 * The mmap_sem must have been held on entry, and may have been 2783 * released depending on flags and vma->vm_ops->fault() return value. 2784 * See filemap_fault() and __lock_page_retry(). 2785 */ 2786 static int __do_fault(struct vm_area_struct *vma, unsigned long address, 2787 pgoff_t pgoff, unsigned int flags, 2788 struct page *cow_page, struct page **page) 2789 { 2790 struct vm_fault vmf; 2791 int ret; 2792 2793 vmf.virtual_address = (void __user *)(address & PAGE_MASK); 2794 vmf.pgoff = pgoff; 2795 vmf.flags = flags; 2796 vmf.page = NULL; 2797 vmf.gfp_mask = __get_fault_gfp_mask(vma); 2798 vmf.cow_page = cow_page; 2799 2800 ret = vma->vm_ops->fault(vma, &vmf); 2801 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2802 return ret; 2803 if (!vmf.page) 2804 goto out; 2805 2806 if (unlikely(PageHWPoison(vmf.page))) { 2807 if (ret & VM_FAULT_LOCKED) 2808 unlock_page(vmf.page); 2809 put_page(vmf.page); 2810 return VM_FAULT_HWPOISON; 2811 } 2812 2813 if (unlikely(!(ret & VM_FAULT_LOCKED))) 2814 lock_page(vmf.page); 2815 else 2816 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page); 2817 2818 out: 2819 *page = vmf.page; 2820 return ret; 2821 } 2822 2823 /** 2824 * do_set_pte - setup new PTE entry for given page and add reverse page mapping. 2825 * 2826 * @vma: virtual memory area 2827 * @address: user virtual address 2828 * @page: page to map 2829 * @pte: pointer to target page table entry 2830 * @write: true, if new entry is writable 2831 * @anon: true, if it's anonymous page 2832 * 2833 * Caller must hold page table lock relevant for @pte. 2834 * 2835 * Target users are page handler itself and implementations of 2836 * vm_ops->map_pages. 2837 */ 2838 void do_set_pte(struct vm_area_struct *vma, unsigned long address, 2839 struct page *page, pte_t *pte, bool write, bool anon) 2840 { 2841 pte_t entry; 2842 2843 flush_icache_page(vma, page); 2844 entry = mk_pte(page, vma->vm_page_prot); 2845 if (write) 2846 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2847 if (anon) { 2848 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 2849 page_add_new_anon_rmap(page, vma, address, false); 2850 } else { 2851 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 2852 page_add_file_rmap(page); 2853 } 2854 set_pte_at(vma->vm_mm, address, pte, entry); 2855 2856 /* no need to invalidate: a not-present page won't be cached */ 2857 update_mmu_cache(vma, address, pte); 2858 } 2859 2860 static unsigned long fault_around_bytes __read_mostly = 2861 rounddown_pow_of_two(65536); 2862 2863 #ifdef CONFIG_DEBUG_FS 2864 static int fault_around_bytes_get(void *data, u64 *val) 2865 { 2866 *val = fault_around_bytes; 2867 return 0; 2868 } 2869 2870 /* 2871 * fault_around_pages() and fault_around_mask() expects fault_around_bytes 2872 * rounded down to nearest page order. It's what do_fault_around() expects to 2873 * see. 2874 */ 2875 static int fault_around_bytes_set(void *data, u64 val) 2876 { 2877 if (val / PAGE_SIZE > PTRS_PER_PTE) 2878 return -EINVAL; 2879 if (val > PAGE_SIZE) 2880 fault_around_bytes = rounddown_pow_of_two(val); 2881 else 2882 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 2883 return 0; 2884 } 2885 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops, 2886 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 2887 2888 static int __init fault_around_debugfs(void) 2889 { 2890 void *ret; 2891 2892 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL, 2893 &fault_around_bytes_fops); 2894 if (!ret) 2895 pr_warn("Failed to create fault_around_bytes in debugfs"); 2896 return 0; 2897 } 2898 late_initcall(fault_around_debugfs); 2899 #endif 2900 2901 /* 2902 * do_fault_around() tries to map few pages around the fault address. The hope 2903 * is that the pages will be needed soon and this will lower the number of 2904 * faults to handle. 2905 * 2906 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 2907 * not ready to be mapped: not up-to-date, locked, etc. 2908 * 2909 * This function is called with the page table lock taken. In the split ptlock 2910 * case the page table lock only protects only those entries which belong to 2911 * the page table corresponding to the fault address. 2912 * 2913 * This function doesn't cross the VMA boundaries, in order to call map_pages() 2914 * only once. 2915 * 2916 * fault_around_pages() defines how many pages we'll try to map. 2917 * do_fault_around() expects it to return a power of two less than or equal to 2918 * PTRS_PER_PTE. 2919 * 2920 * The virtual address of the area that we map is naturally aligned to the 2921 * fault_around_pages() value (and therefore to page order). This way it's 2922 * easier to guarantee that we don't cross page table boundaries. 2923 */ 2924 static void do_fault_around(struct vm_area_struct *vma, unsigned long address, 2925 pte_t *pte, pgoff_t pgoff, unsigned int flags) 2926 { 2927 unsigned long start_addr, nr_pages, mask; 2928 pgoff_t max_pgoff; 2929 struct vm_fault vmf; 2930 int off; 2931 2932 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 2933 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 2934 2935 start_addr = max(address & mask, vma->vm_start); 2936 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 2937 pte -= off; 2938 pgoff -= off; 2939 2940 /* 2941 * max_pgoff is either end of page table or end of vma 2942 * or fault_around_pages() from pgoff, depending what is nearest. 2943 */ 2944 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 2945 PTRS_PER_PTE - 1; 2946 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1, 2947 pgoff + nr_pages - 1); 2948 2949 /* Check if it makes any sense to call ->map_pages */ 2950 while (!pte_none(*pte)) { 2951 if (++pgoff > max_pgoff) 2952 return; 2953 start_addr += PAGE_SIZE; 2954 if (start_addr >= vma->vm_end) 2955 return; 2956 pte++; 2957 } 2958 2959 vmf.virtual_address = (void __user *) start_addr; 2960 vmf.pte = pte; 2961 vmf.pgoff = pgoff; 2962 vmf.max_pgoff = max_pgoff; 2963 vmf.flags = flags; 2964 vmf.gfp_mask = __get_fault_gfp_mask(vma); 2965 vma->vm_ops->map_pages(vma, &vmf); 2966 } 2967 2968 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2969 unsigned long address, pmd_t *pmd, 2970 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 2971 { 2972 struct page *fault_page; 2973 spinlock_t *ptl; 2974 pte_t *pte; 2975 int ret = 0; 2976 2977 /* 2978 * Let's call ->map_pages() first and use ->fault() as fallback 2979 * if page by the offset is not ready to be mapped (cold cache or 2980 * something). 2981 */ 2982 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 2983 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2984 do_fault_around(vma, address, pte, pgoff, flags); 2985 if (!pte_same(*pte, orig_pte)) 2986 goto unlock_out; 2987 pte_unmap_unlock(pte, ptl); 2988 } 2989 2990 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page); 2991 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 2992 return ret; 2993 2994 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2995 if (unlikely(!pte_same(*pte, orig_pte))) { 2996 pte_unmap_unlock(pte, ptl); 2997 unlock_page(fault_page); 2998 put_page(fault_page); 2999 return ret; 3000 } 3001 do_set_pte(vma, address, fault_page, pte, false, false); 3002 unlock_page(fault_page); 3003 unlock_out: 3004 pte_unmap_unlock(pte, ptl); 3005 return ret; 3006 } 3007 3008 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3009 unsigned long address, pmd_t *pmd, 3010 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 3011 { 3012 struct page *fault_page, *new_page; 3013 struct mem_cgroup *memcg; 3014 spinlock_t *ptl; 3015 pte_t *pte; 3016 int ret; 3017 3018 if (unlikely(anon_vma_prepare(vma))) 3019 return VM_FAULT_OOM; 3020 3021 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 3022 if (!new_page) 3023 return VM_FAULT_OOM; 3024 3025 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) { 3026 put_page(new_page); 3027 return VM_FAULT_OOM; 3028 } 3029 3030 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page); 3031 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3032 goto uncharge_out; 3033 3034 if (fault_page) 3035 copy_user_highpage(new_page, fault_page, address, vma); 3036 __SetPageUptodate(new_page); 3037 3038 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 3039 if (unlikely(!pte_same(*pte, orig_pte))) { 3040 pte_unmap_unlock(pte, ptl); 3041 if (fault_page) { 3042 unlock_page(fault_page); 3043 put_page(fault_page); 3044 } else { 3045 /* 3046 * The fault handler has no page to lock, so it holds 3047 * i_mmap_lock for read to protect against truncate. 3048 */ 3049 i_mmap_unlock_read(vma->vm_file->f_mapping); 3050 } 3051 goto uncharge_out; 3052 } 3053 do_set_pte(vma, address, new_page, pte, true, true); 3054 mem_cgroup_commit_charge(new_page, memcg, false, false); 3055 lru_cache_add_active_or_unevictable(new_page, vma); 3056 pte_unmap_unlock(pte, ptl); 3057 if (fault_page) { 3058 unlock_page(fault_page); 3059 put_page(fault_page); 3060 } else { 3061 /* 3062 * The fault handler has no page to lock, so it holds 3063 * i_mmap_lock for read to protect against truncate. 3064 */ 3065 i_mmap_unlock_read(vma->vm_file->f_mapping); 3066 } 3067 return ret; 3068 uncharge_out: 3069 mem_cgroup_cancel_charge(new_page, memcg, false); 3070 put_page(new_page); 3071 return ret; 3072 } 3073 3074 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3075 unsigned long address, pmd_t *pmd, 3076 pgoff_t pgoff, unsigned int flags, pte_t orig_pte) 3077 { 3078 struct page *fault_page; 3079 struct address_space *mapping; 3080 spinlock_t *ptl; 3081 pte_t *pte; 3082 int dirtied = 0; 3083 int ret, tmp; 3084 3085 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page); 3086 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3087 return ret; 3088 3089 /* 3090 * Check if the backing address space wants to know that the page is 3091 * about to become writable 3092 */ 3093 if (vma->vm_ops->page_mkwrite) { 3094 unlock_page(fault_page); 3095 tmp = do_page_mkwrite(vma, fault_page, address); 3096 if (unlikely(!tmp || 3097 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3098 put_page(fault_page); 3099 return tmp; 3100 } 3101 } 3102 3103 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 3104 if (unlikely(!pte_same(*pte, orig_pte))) { 3105 pte_unmap_unlock(pte, ptl); 3106 unlock_page(fault_page); 3107 put_page(fault_page); 3108 return ret; 3109 } 3110 do_set_pte(vma, address, fault_page, pte, true, false); 3111 pte_unmap_unlock(pte, ptl); 3112 3113 if (set_page_dirty(fault_page)) 3114 dirtied = 1; 3115 /* 3116 * Take a local copy of the address_space - page.mapping may be zeroed 3117 * by truncate after unlock_page(). The address_space itself remains 3118 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 3119 * release semantics to prevent the compiler from undoing this copying. 3120 */ 3121 mapping = page_rmapping(fault_page); 3122 unlock_page(fault_page); 3123 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) { 3124 /* 3125 * Some device drivers do not set page.mapping but still 3126 * dirty their pages 3127 */ 3128 balance_dirty_pages_ratelimited(mapping); 3129 } 3130 3131 if (!vma->vm_ops->page_mkwrite) 3132 file_update_time(vma->vm_file); 3133 3134 return ret; 3135 } 3136 3137 /* 3138 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3139 * but allow concurrent faults). 3140 * The mmap_sem may have been released depending on flags and our 3141 * return value. See filemap_fault() and __lock_page_or_retry(). 3142 */ 3143 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3144 unsigned long address, pte_t *page_table, pmd_t *pmd, 3145 unsigned int flags, pte_t orig_pte) 3146 { 3147 pgoff_t pgoff = linear_page_index(vma, address); 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 (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 3382 flags & FAULT_FLAG_INSTRUCTION, 3383 flags & FAULT_FLAG_REMOTE)) 3384 return VM_FAULT_SIGSEGV; 3385 3386 if (unlikely(is_vm_hugetlb_page(vma))) 3387 return hugetlb_fault(mm, vma, address, flags); 3388 3389 pgd = pgd_offset(mm, address); 3390 pud = pud_alloc(mm, pgd, address); 3391 if (!pud) 3392 return VM_FAULT_OOM; 3393 pmd = pmd_alloc(mm, pud, address); 3394 if (!pmd) 3395 return VM_FAULT_OOM; 3396 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { 3397 int ret = create_huge_pmd(mm, vma, address, pmd, flags); 3398 if (!(ret & VM_FAULT_FALLBACK)) 3399 return ret; 3400 } else { 3401 pmd_t orig_pmd = *pmd; 3402 int ret; 3403 3404 barrier(); 3405 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) { 3406 unsigned int dirty = flags & FAULT_FLAG_WRITE; 3407 3408 if (pmd_protnone(orig_pmd)) 3409 return do_huge_pmd_numa_page(mm, vma, address, 3410 orig_pmd, pmd); 3411 3412 if (dirty && !pmd_write(orig_pmd)) { 3413 ret = wp_huge_pmd(mm, vma, address, pmd, 3414 orig_pmd, flags); 3415 if (!(ret & VM_FAULT_FALLBACK)) 3416 return ret; 3417 } else { 3418 huge_pmd_set_accessed(mm, vma, address, pmd, 3419 orig_pmd, dirty); 3420 return 0; 3421 } 3422 } 3423 } 3424 3425 /* 3426 * Use pte_alloc() instead of pte_alloc_map, because we can't 3427 * run pte_offset_map on the pmd, if an huge pmd could 3428 * materialize from under us from a different thread. 3429 */ 3430 if (unlikely(pte_alloc(mm, pmd, address))) 3431 return VM_FAULT_OOM; 3432 /* 3433 * If a huge pmd materialized under us just retry later. Use 3434 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd 3435 * didn't become pmd_trans_huge under us and then back to pmd_none, as 3436 * a result of MADV_DONTNEED running immediately after a huge pmd fault 3437 * in a different thread of this mm, in turn leading to a misleading 3438 * pmd_trans_huge() retval. All we have to ensure is that it is a 3439 * regular pmd that we can walk with pte_offset_map() and we can do that 3440 * through an atomic read in C, which is what pmd_trans_unstable() 3441 * provides. 3442 */ 3443 if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd))) 3444 return 0; 3445 /* 3446 * A regular pmd is established and it can't morph into a huge pmd 3447 * from under us anymore at this point because we hold the mmap_sem 3448 * read mode and khugepaged takes it in write mode. So now it's 3449 * safe to run pte_offset_map(). 3450 */ 3451 pte = pte_offset_map(pmd, address); 3452 3453 return handle_pte_fault(mm, vma, address, pte, pmd, flags); 3454 } 3455 3456 /* 3457 * By the time we get here, we already hold the mm semaphore 3458 * 3459 * The mmap_sem may have been released depending on flags and our 3460 * return value. See filemap_fault() and __lock_page_or_retry(). 3461 */ 3462 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3463 unsigned long address, unsigned int flags) 3464 { 3465 int ret; 3466 3467 __set_current_state(TASK_RUNNING); 3468 3469 count_vm_event(PGFAULT); 3470 mem_cgroup_count_vm_event(mm, PGFAULT); 3471 3472 /* do counter updates before entering really critical section. */ 3473 check_sync_rss_stat(current); 3474 3475 /* 3476 * Enable the memcg OOM handling for faults triggered in user 3477 * space. Kernel faults are handled more gracefully. 3478 */ 3479 if (flags & FAULT_FLAG_USER) 3480 mem_cgroup_oom_enable(); 3481 3482 ret = __handle_mm_fault(mm, vma, address, flags); 3483 3484 if (flags & FAULT_FLAG_USER) { 3485 mem_cgroup_oom_disable(); 3486 /* 3487 * The task may have entered a memcg OOM situation but 3488 * if the allocation error was handled gracefully (no 3489 * VM_FAULT_OOM), there is no need to kill anything. 3490 * Just clean up the OOM state peacefully. 3491 */ 3492 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 3493 mem_cgroup_oom_synchronize(false); 3494 } 3495 3496 return ret; 3497 } 3498 EXPORT_SYMBOL_GPL(handle_mm_fault); 3499 3500 #ifndef __PAGETABLE_PUD_FOLDED 3501 /* 3502 * Allocate page upper directory. 3503 * We've already handled the fast-path in-line. 3504 */ 3505 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 3506 { 3507 pud_t *new = pud_alloc_one(mm, address); 3508 if (!new) 3509 return -ENOMEM; 3510 3511 smp_wmb(); /* See comment in __pte_alloc */ 3512 3513 spin_lock(&mm->page_table_lock); 3514 if (pgd_present(*pgd)) /* Another has populated it */ 3515 pud_free(mm, new); 3516 else 3517 pgd_populate(mm, pgd, new); 3518 spin_unlock(&mm->page_table_lock); 3519 return 0; 3520 } 3521 #endif /* __PAGETABLE_PUD_FOLDED */ 3522 3523 #ifndef __PAGETABLE_PMD_FOLDED 3524 /* 3525 * Allocate page middle directory. 3526 * We've already handled the fast-path in-line. 3527 */ 3528 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 3529 { 3530 pmd_t *new = pmd_alloc_one(mm, address); 3531 if (!new) 3532 return -ENOMEM; 3533 3534 smp_wmb(); /* See comment in __pte_alloc */ 3535 3536 spin_lock(&mm->page_table_lock); 3537 #ifndef __ARCH_HAS_4LEVEL_HACK 3538 if (!pud_present(*pud)) { 3539 mm_inc_nr_pmds(mm); 3540 pud_populate(mm, pud, new); 3541 } else /* Another has populated it */ 3542 pmd_free(mm, new); 3543 #else 3544 if (!pgd_present(*pud)) { 3545 mm_inc_nr_pmds(mm); 3546 pgd_populate(mm, pud, new); 3547 } else /* Another has populated it */ 3548 pmd_free(mm, new); 3549 #endif /* __ARCH_HAS_4LEVEL_HACK */ 3550 spin_unlock(&mm->page_table_lock); 3551 return 0; 3552 } 3553 #endif /* __PAGETABLE_PMD_FOLDED */ 3554 3555 static int __follow_pte(struct mm_struct *mm, unsigned long address, 3556 pte_t **ptepp, spinlock_t **ptlp) 3557 { 3558 pgd_t *pgd; 3559 pud_t *pud; 3560 pmd_t *pmd; 3561 pte_t *ptep; 3562 3563 pgd = pgd_offset(mm, address); 3564 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 3565 goto out; 3566 3567 pud = pud_offset(pgd, address); 3568 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 3569 goto out; 3570 3571 pmd = pmd_offset(pud, address); 3572 VM_BUG_ON(pmd_trans_huge(*pmd)); 3573 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 3574 goto out; 3575 3576 /* We cannot handle huge page PFN maps. Luckily they don't exist. */ 3577 if (pmd_huge(*pmd)) 3578 goto out; 3579 3580 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 3581 if (!ptep) 3582 goto out; 3583 if (!pte_present(*ptep)) 3584 goto unlock; 3585 *ptepp = ptep; 3586 return 0; 3587 unlock: 3588 pte_unmap_unlock(ptep, *ptlp); 3589 out: 3590 return -EINVAL; 3591 } 3592 3593 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 3594 pte_t **ptepp, spinlock_t **ptlp) 3595 { 3596 int res; 3597 3598 /* (void) is needed to make gcc happy */ 3599 (void) __cond_lock(*ptlp, 3600 !(res = __follow_pte(mm, address, ptepp, ptlp))); 3601 return res; 3602 } 3603 3604 /** 3605 * follow_pfn - look up PFN at a user virtual address 3606 * @vma: memory mapping 3607 * @address: user virtual address 3608 * @pfn: location to store found PFN 3609 * 3610 * Only IO mappings and raw PFN mappings are allowed. 3611 * 3612 * Returns zero and the pfn at @pfn on success, -ve otherwise. 3613 */ 3614 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 3615 unsigned long *pfn) 3616 { 3617 int ret = -EINVAL; 3618 spinlock_t *ptl; 3619 pte_t *ptep; 3620 3621 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3622 return ret; 3623 3624 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 3625 if (ret) 3626 return ret; 3627 *pfn = pte_pfn(*ptep); 3628 pte_unmap_unlock(ptep, ptl); 3629 return 0; 3630 } 3631 EXPORT_SYMBOL(follow_pfn); 3632 3633 #ifdef CONFIG_HAVE_IOREMAP_PROT 3634 int follow_phys(struct vm_area_struct *vma, 3635 unsigned long address, unsigned int flags, 3636 unsigned long *prot, resource_size_t *phys) 3637 { 3638 int ret = -EINVAL; 3639 pte_t *ptep, pte; 3640 spinlock_t *ptl; 3641 3642 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 3643 goto out; 3644 3645 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 3646 goto out; 3647 pte = *ptep; 3648 3649 if ((flags & FOLL_WRITE) && !pte_write(pte)) 3650 goto unlock; 3651 3652 *prot = pgprot_val(pte_pgprot(pte)); 3653 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 3654 3655 ret = 0; 3656 unlock: 3657 pte_unmap_unlock(ptep, ptl); 3658 out: 3659 return ret; 3660 } 3661 3662 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 3663 void *buf, int len, int write) 3664 { 3665 resource_size_t phys_addr; 3666 unsigned long prot = 0; 3667 void __iomem *maddr; 3668 int offset = addr & (PAGE_SIZE-1); 3669 3670 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 3671 return -EINVAL; 3672 3673 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 3674 if (write) 3675 memcpy_toio(maddr + offset, buf, len); 3676 else 3677 memcpy_fromio(buf, maddr + offset, len); 3678 iounmap(maddr); 3679 3680 return len; 3681 } 3682 EXPORT_SYMBOL_GPL(generic_access_phys); 3683 #endif 3684 3685 /* 3686 * Access another process' address space as given in mm. If non-NULL, use the 3687 * given task for page fault accounting. 3688 */ 3689 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, 3690 unsigned long addr, void *buf, int len, int write) 3691 { 3692 struct vm_area_struct *vma; 3693 void *old_buf = buf; 3694 3695 down_read(&mm->mmap_sem); 3696 /* ignore errors, just check how much was successfully transferred */ 3697 while (len) { 3698 int bytes, ret, offset; 3699 void *maddr; 3700 struct page *page = NULL; 3701 3702 ret = get_user_pages_remote(tsk, mm, addr, 1, 3703 write, 1, &page, &vma); 3704 if (ret <= 0) { 3705 #ifndef CONFIG_HAVE_IOREMAP_PROT 3706 break; 3707 #else 3708 /* 3709 * Check if this is a VM_IO | VM_PFNMAP VMA, which 3710 * we can access using slightly different code. 3711 */ 3712 vma = find_vma(mm, addr); 3713 if (!vma || vma->vm_start > addr) 3714 break; 3715 if (vma->vm_ops && vma->vm_ops->access) 3716 ret = vma->vm_ops->access(vma, addr, buf, 3717 len, write); 3718 if (ret <= 0) 3719 break; 3720 bytes = ret; 3721 #endif 3722 } else { 3723 bytes = len; 3724 offset = addr & (PAGE_SIZE-1); 3725 if (bytes > PAGE_SIZE-offset) 3726 bytes = PAGE_SIZE-offset; 3727 3728 maddr = kmap(page); 3729 if (write) { 3730 copy_to_user_page(vma, page, addr, 3731 maddr + offset, buf, bytes); 3732 set_page_dirty_lock(page); 3733 } else { 3734 copy_from_user_page(vma, page, addr, 3735 buf, maddr + offset, bytes); 3736 } 3737 kunmap(page); 3738 put_page(page); 3739 } 3740 len -= bytes; 3741 buf += bytes; 3742 addr += bytes; 3743 } 3744 up_read(&mm->mmap_sem); 3745 3746 return buf - old_buf; 3747 } 3748 3749 /** 3750 * access_remote_vm - access another process' address space 3751 * @mm: the mm_struct of the target address space 3752 * @addr: start address to access 3753 * @buf: source or destination buffer 3754 * @len: number of bytes to transfer 3755 * @write: whether the access is a write 3756 * 3757 * The caller must hold a reference on @mm. 3758 */ 3759 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 3760 void *buf, int len, int write) 3761 { 3762 return __access_remote_vm(NULL, mm, addr, buf, len, write); 3763 } 3764 3765 /* 3766 * Access another process' address space. 3767 * Source/target buffer must be kernel space, 3768 * Do not walk the page table directly, use get_user_pages 3769 */ 3770 int access_process_vm(struct task_struct *tsk, unsigned long addr, 3771 void *buf, int len, int write) 3772 { 3773 struct mm_struct *mm; 3774 int ret; 3775 3776 mm = get_task_mm(tsk); 3777 if (!mm) 3778 return 0; 3779 3780 ret = __access_remote_vm(tsk, mm, addr, buf, len, write); 3781 mmput(mm); 3782 3783 return ret; 3784 } 3785 3786 /* 3787 * Print the name of a VMA. 3788 */ 3789 void print_vma_addr(char *prefix, unsigned long ip) 3790 { 3791 struct mm_struct *mm = current->mm; 3792 struct vm_area_struct *vma; 3793 3794 /* 3795 * Do not print if we are in atomic 3796 * contexts (in exception stacks, etc.): 3797 */ 3798 if (preempt_count()) 3799 return; 3800 3801 down_read(&mm->mmap_sem); 3802 vma = find_vma(mm, ip); 3803 if (vma && vma->vm_file) { 3804 struct file *f = vma->vm_file; 3805 char *buf = (char *)__get_free_page(GFP_KERNEL); 3806 if (buf) { 3807 char *p; 3808 3809 p = file_path(f, buf, PAGE_SIZE); 3810 if (IS_ERR(p)) 3811 p = "?"; 3812 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 3813 vma->vm_start, 3814 vma->vm_end - vma->vm_start); 3815 free_page((unsigned long)buf); 3816 } 3817 } 3818 up_read(&mm->mmap_sem); 3819 } 3820 3821 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 3822 void __might_fault(const char *file, int line) 3823 { 3824 /* 3825 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 3826 * holding the mmap_sem, this is safe because kernel memory doesn't 3827 * get paged out, therefore we'll never actually fault, and the 3828 * below annotations will generate false positives. 3829 */ 3830 if (segment_eq(get_fs(), KERNEL_DS)) 3831 return; 3832 if (pagefault_disabled()) 3833 return; 3834 __might_sleep(file, line, 0); 3835 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 3836 if (current->mm) 3837 might_lock_read(¤t->mm->mmap_sem); 3838 #endif 3839 } 3840 EXPORT_SYMBOL(__might_fault); 3841 #endif 3842 3843 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 3844 static void clear_gigantic_page(struct page *page, 3845 unsigned long addr, 3846 unsigned int pages_per_huge_page) 3847 { 3848 int i; 3849 struct page *p = page; 3850 3851 might_sleep(); 3852 for (i = 0; i < pages_per_huge_page; 3853 i++, p = mem_map_next(p, page, i)) { 3854 cond_resched(); 3855 clear_user_highpage(p, addr + i * PAGE_SIZE); 3856 } 3857 } 3858 void clear_huge_page(struct page *page, 3859 unsigned long addr, unsigned int pages_per_huge_page) 3860 { 3861 int i; 3862 3863 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3864 clear_gigantic_page(page, addr, pages_per_huge_page); 3865 return; 3866 } 3867 3868 might_sleep(); 3869 for (i = 0; i < pages_per_huge_page; i++) { 3870 cond_resched(); 3871 clear_user_highpage(page + i, addr + i * PAGE_SIZE); 3872 } 3873 } 3874 3875 static void copy_user_gigantic_page(struct page *dst, struct page *src, 3876 unsigned long addr, 3877 struct vm_area_struct *vma, 3878 unsigned int pages_per_huge_page) 3879 { 3880 int i; 3881 struct page *dst_base = dst; 3882 struct page *src_base = src; 3883 3884 for (i = 0; i < pages_per_huge_page; ) { 3885 cond_resched(); 3886 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 3887 3888 i++; 3889 dst = mem_map_next(dst, dst_base, i); 3890 src = mem_map_next(src, src_base, i); 3891 } 3892 } 3893 3894 void copy_user_huge_page(struct page *dst, struct page *src, 3895 unsigned long addr, struct vm_area_struct *vma, 3896 unsigned int pages_per_huge_page) 3897 { 3898 int i; 3899 3900 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 3901 copy_user_gigantic_page(dst, src, addr, vma, 3902 pages_per_huge_page); 3903 return; 3904 } 3905 3906 might_sleep(); 3907 for (i = 0; i < pages_per_huge_page; i++) { 3908 cond_resched(); 3909 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 3910 } 3911 } 3912 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 3913 3914 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 3915 3916 static struct kmem_cache *page_ptl_cachep; 3917 3918 void __init ptlock_cache_init(void) 3919 { 3920 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 3921 SLAB_PANIC, NULL); 3922 } 3923 3924 bool ptlock_alloc(struct page *page) 3925 { 3926 spinlock_t *ptl; 3927 3928 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 3929 if (!ptl) 3930 return false; 3931 page->ptl = ptl; 3932 return true; 3933 } 3934 3935 void ptlock_free(struct page *page) 3936 { 3937 kmem_cache_free(page_ptl_cachep, page->ptl); 3938 } 3939 #endif 3940