1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/memory.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 */ 7 8 /* 9 * demand-loading started 01.12.91 - seems it is high on the list of 10 * things wanted, and it should be easy to implement. - Linus 11 */ 12 13 /* 14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared 15 * pages started 02.12.91, seems to work. - Linus. 16 * 17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it 18 * would have taken more than the 6M I have free, but it worked well as 19 * far as I could see. 20 * 21 * Also corrected some "invalidate()"s - I wasn't doing enough of them. 22 */ 23 24 /* 25 * Real VM (paging to/from disk) started 18.12.91. Much more work and 26 * thought has to go into this. Oh, well.. 27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why. 28 * Found it. Everything seems to work now. 29 * 20.12.91 - Ok, making the swap-device changeable like the root. 30 */ 31 32 /* 33 * 05.04.94 - Multi-page memory management added for v1.1. 34 * Idea by Alex Bligh (alex@cconcepts.co.uk) 35 * 36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG 37 * (Gerhard.Wichert@pdb.siemens.de) 38 * 39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen) 40 */ 41 42 #include <linux/kernel_stat.h> 43 #include <linux/mm.h> 44 #include <linux/sched/mm.h> 45 #include <linux/sched/coredump.h> 46 #include <linux/sched/numa_balancing.h> 47 #include <linux/sched/task.h> 48 #include <linux/hugetlb.h> 49 #include <linux/mman.h> 50 #include <linux/swap.h> 51 #include <linux/highmem.h> 52 #include <linux/pagemap.h> 53 #include <linux/memremap.h> 54 #include <linux/ksm.h> 55 #include <linux/rmap.h> 56 #include <linux/export.h> 57 #include <linux/delayacct.h> 58 #include <linux/init.h> 59 #include <linux/pfn_t.h> 60 #include <linux/writeback.h> 61 #include <linux/memcontrol.h> 62 #include <linux/mmu_notifier.h> 63 #include <linux/swapops.h> 64 #include <linux/elf.h> 65 #include <linux/gfp.h> 66 #include <linux/migrate.h> 67 #include <linux/string.h> 68 #include <linux/debugfs.h> 69 #include <linux/userfaultfd_k.h> 70 #include <linux/dax.h> 71 #include <linux/oom.h> 72 #include <linux/numa.h> 73 #include <linux/perf_event.h> 74 #include <linux/ptrace.h> 75 #include <linux/vmalloc.h> 76 77 #include <trace/events/kmem.h> 78 79 #include <asm/io.h> 80 #include <asm/mmu_context.h> 81 #include <asm/pgalloc.h> 82 #include <linux/uaccess.h> 83 #include <asm/tlb.h> 84 #include <asm/tlbflush.h> 85 86 #include "pgalloc-track.h" 87 #include "internal.h" 88 89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) 90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 91 #endif 92 93 #ifndef CONFIG_NEED_MULTIPLE_NODES 94 /* use the per-pgdat data instead for discontigmem - mbligh */ 95 unsigned long max_mapnr; 96 EXPORT_SYMBOL(max_mapnr); 97 98 struct page *mem_map; 99 EXPORT_SYMBOL(mem_map); 100 #endif 101 102 /* 103 * A number of key systems in x86 including ioremap() rely on the assumption 104 * that high_memory defines the upper bound on direct map memory, then end 105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 107 * and ZONE_HIGHMEM. 108 */ 109 void *high_memory; 110 EXPORT_SYMBOL(high_memory); 111 112 /* 113 * Randomize the address space (stacks, mmaps, brk, etc.). 114 * 115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 116 * as ancient (libc5 based) binaries can segfault. ) 117 */ 118 int randomize_va_space __read_mostly = 119 #ifdef CONFIG_COMPAT_BRK 120 1; 121 #else 122 2; 123 #endif 124 125 #ifndef arch_faults_on_old_pte 126 static inline bool arch_faults_on_old_pte(void) 127 { 128 /* 129 * Those arches which don't have hw access flag feature need to 130 * implement their own helper. By default, "true" means pagefault 131 * will be hit on old pte. 132 */ 133 return true; 134 } 135 #endif 136 137 #ifndef arch_wants_old_prefaulted_pte 138 static inline bool arch_wants_old_prefaulted_pte(void) 139 { 140 /* 141 * Transitioning a PTE from 'old' to 'young' can be expensive on 142 * some architectures, even if it's performed in hardware. By 143 * default, "false" means prefaulted entries will be 'young'. 144 */ 145 return false; 146 } 147 #endif 148 149 static int __init disable_randmaps(char *s) 150 { 151 randomize_va_space = 0; 152 return 1; 153 } 154 __setup("norandmaps", disable_randmaps); 155 156 unsigned long zero_pfn __read_mostly; 157 EXPORT_SYMBOL(zero_pfn); 158 159 unsigned long highest_memmap_pfn __read_mostly; 160 161 /* 162 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 163 */ 164 static int __init init_zero_pfn(void) 165 { 166 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 167 return 0; 168 } 169 core_initcall(init_zero_pfn); 170 171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count) 172 { 173 trace_rss_stat(mm, member, count); 174 } 175 176 #if defined(SPLIT_RSS_COUNTING) 177 178 void sync_mm_rss(struct mm_struct *mm) 179 { 180 int i; 181 182 for (i = 0; i < NR_MM_COUNTERS; i++) { 183 if (current->rss_stat.count[i]) { 184 add_mm_counter(mm, i, current->rss_stat.count[i]); 185 current->rss_stat.count[i] = 0; 186 } 187 } 188 current->rss_stat.events = 0; 189 } 190 191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 192 { 193 struct task_struct *task = current; 194 195 if (likely(task->mm == mm)) 196 task->rss_stat.count[member] += val; 197 else 198 add_mm_counter(mm, member, val); 199 } 200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 202 203 /* sync counter once per 64 page faults */ 204 #define TASK_RSS_EVENTS_THRESH (64) 205 static void check_sync_rss_stat(struct task_struct *task) 206 { 207 if (unlikely(task != current)) 208 return; 209 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 210 sync_mm_rss(task->mm); 211 } 212 #else /* SPLIT_RSS_COUNTING */ 213 214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 216 217 static void check_sync_rss_stat(struct task_struct *task) 218 { 219 } 220 221 #endif /* SPLIT_RSS_COUNTING */ 222 223 /* 224 * Note: this doesn't free the actual pages themselves. That 225 * has been handled earlier when unmapping all the memory regions. 226 */ 227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 228 unsigned long addr) 229 { 230 pgtable_t token = pmd_pgtable(*pmd); 231 pmd_clear(pmd); 232 pte_free_tlb(tlb, token, addr); 233 mm_dec_nr_ptes(tlb->mm); 234 } 235 236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 237 unsigned long addr, unsigned long end, 238 unsigned long floor, unsigned long ceiling) 239 { 240 pmd_t *pmd; 241 unsigned long next; 242 unsigned long start; 243 244 start = addr; 245 pmd = pmd_offset(pud, addr); 246 do { 247 next = pmd_addr_end(addr, end); 248 if (pmd_none_or_clear_bad(pmd)) 249 continue; 250 free_pte_range(tlb, pmd, addr); 251 } while (pmd++, addr = next, addr != end); 252 253 start &= PUD_MASK; 254 if (start < floor) 255 return; 256 if (ceiling) { 257 ceiling &= PUD_MASK; 258 if (!ceiling) 259 return; 260 } 261 if (end - 1 > ceiling - 1) 262 return; 263 264 pmd = pmd_offset(pud, start); 265 pud_clear(pud); 266 pmd_free_tlb(tlb, pmd, start); 267 mm_dec_nr_pmds(tlb->mm); 268 } 269 270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 271 unsigned long addr, unsigned long end, 272 unsigned long floor, unsigned long ceiling) 273 { 274 pud_t *pud; 275 unsigned long next; 276 unsigned long start; 277 278 start = addr; 279 pud = pud_offset(p4d, addr); 280 do { 281 next = pud_addr_end(addr, end); 282 if (pud_none_or_clear_bad(pud)) 283 continue; 284 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 285 } while (pud++, addr = next, addr != end); 286 287 start &= P4D_MASK; 288 if (start < floor) 289 return; 290 if (ceiling) { 291 ceiling &= P4D_MASK; 292 if (!ceiling) 293 return; 294 } 295 if (end - 1 > ceiling - 1) 296 return; 297 298 pud = pud_offset(p4d, start); 299 p4d_clear(p4d); 300 pud_free_tlb(tlb, pud, start); 301 mm_dec_nr_puds(tlb->mm); 302 } 303 304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 305 unsigned long addr, unsigned long end, 306 unsigned long floor, unsigned long ceiling) 307 { 308 p4d_t *p4d; 309 unsigned long next; 310 unsigned long start; 311 312 start = addr; 313 p4d = p4d_offset(pgd, addr); 314 do { 315 next = p4d_addr_end(addr, end); 316 if (p4d_none_or_clear_bad(p4d)) 317 continue; 318 free_pud_range(tlb, p4d, addr, next, floor, ceiling); 319 } while (p4d++, addr = next, addr != end); 320 321 start &= PGDIR_MASK; 322 if (start < floor) 323 return; 324 if (ceiling) { 325 ceiling &= PGDIR_MASK; 326 if (!ceiling) 327 return; 328 } 329 if (end - 1 > ceiling - 1) 330 return; 331 332 p4d = p4d_offset(pgd, start); 333 pgd_clear(pgd); 334 p4d_free_tlb(tlb, p4d, start); 335 } 336 337 /* 338 * This function frees user-level page tables of a process. 339 */ 340 void free_pgd_range(struct mmu_gather *tlb, 341 unsigned long addr, unsigned long end, 342 unsigned long floor, unsigned long ceiling) 343 { 344 pgd_t *pgd; 345 unsigned long next; 346 347 /* 348 * The next few lines have given us lots of grief... 349 * 350 * Why are we testing PMD* at this top level? Because often 351 * there will be no work to do at all, and we'd prefer not to 352 * go all the way down to the bottom just to discover that. 353 * 354 * Why all these "- 1"s? Because 0 represents both the bottom 355 * of the address space and the top of it (using -1 for the 356 * top wouldn't help much: the masks would do the wrong thing). 357 * The rule is that addr 0 and floor 0 refer to the bottom of 358 * the address space, but end 0 and ceiling 0 refer to the top 359 * Comparisons need to use "end - 1" and "ceiling - 1" (though 360 * that end 0 case should be mythical). 361 * 362 * Wherever addr is brought up or ceiling brought down, we must 363 * be careful to reject "the opposite 0" before it confuses the 364 * subsequent tests. But what about where end is brought down 365 * by PMD_SIZE below? no, end can't go down to 0 there. 366 * 367 * Whereas we round start (addr) and ceiling down, by different 368 * masks at different levels, in order to test whether a table 369 * now has no other vmas using it, so can be freed, we don't 370 * bother to round floor or end up - the tests don't need that. 371 */ 372 373 addr &= PMD_MASK; 374 if (addr < floor) { 375 addr += PMD_SIZE; 376 if (!addr) 377 return; 378 } 379 if (ceiling) { 380 ceiling &= PMD_MASK; 381 if (!ceiling) 382 return; 383 } 384 if (end - 1 > ceiling - 1) 385 end -= PMD_SIZE; 386 if (addr > end - 1) 387 return; 388 /* 389 * We add page table cache pages with PAGE_SIZE, 390 * (see pte_free_tlb()), flush the tlb if we need 391 */ 392 tlb_change_page_size(tlb, PAGE_SIZE); 393 pgd = pgd_offset(tlb->mm, addr); 394 do { 395 next = pgd_addr_end(addr, end); 396 if (pgd_none_or_clear_bad(pgd)) 397 continue; 398 free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 399 } while (pgd++, addr = next, addr != end); 400 } 401 402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 403 unsigned long floor, unsigned long ceiling) 404 { 405 while (vma) { 406 struct vm_area_struct *next = vma->vm_next; 407 unsigned long addr = vma->vm_start; 408 409 /* 410 * Hide vma from rmap and truncate_pagecache before freeing 411 * pgtables 412 */ 413 unlink_anon_vmas(vma); 414 unlink_file_vma(vma); 415 416 if (is_vm_hugetlb_page(vma)) { 417 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 418 floor, next ? next->vm_start : ceiling); 419 } else { 420 /* 421 * Optimization: gather nearby vmas into one call down 422 */ 423 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 424 && !is_vm_hugetlb_page(next)) { 425 vma = next; 426 next = vma->vm_next; 427 unlink_anon_vmas(vma); 428 unlink_file_vma(vma); 429 } 430 free_pgd_range(tlb, addr, vma->vm_end, 431 floor, next ? next->vm_start : ceiling); 432 } 433 vma = next; 434 } 435 } 436 437 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 438 { 439 spinlock_t *ptl; 440 pgtable_t new = pte_alloc_one(mm); 441 if (!new) 442 return -ENOMEM; 443 444 /* 445 * Ensure all pte setup (eg. pte page lock and page clearing) are 446 * visible before the pte is made visible to other CPUs by being 447 * put into page tables. 448 * 449 * The other side of the story is the pointer chasing in the page 450 * table walking code (when walking the page table without locking; 451 * ie. most of the time). Fortunately, these data accesses consist 452 * of a chain of data-dependent loads, meaning most CPUs (alpha 453 * being the notable exception) will already guarantee loads are 454 * seen in-order. See the alpha page table accessors for the 455 * smp_rmb() barriers in page table walking code. 456 */ 457 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 458 459 ptl = pmd_lock(mm, pmd); 460 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 461 mm_inc_nr_ptes(mm); 462 pmd_populate(mm, pmd, new); 463 new = NULL; 464 } 465 spin_unlock(ptl); 466 if (new) 467 pte_free(mm, new); 468 return 0; 469 } 470 471 int __pte_alloc_kernel(pmd_t *pmd) 472 { 473 pte_t *new = pte_alloc_one_kernel(&init_mm); 474 if (!new) 475 return -ENOMEM; 476 477 smp_wmb(); /* See comment in __pte_alloc */ 478 479 spin_lock(&init_mm.page_table_lock); 480 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 481 pmd_populate_kernel(&init_mm, pmd, new); 482 new = NULL; 483 } 484 spin_unlock(&init_mm.page_table_lock); 485 if (new) 486 pte_free_kernel(&init_mm, new); 487 return 0; 488 } 489 490 static inline void init_rss_vec(int *rss) 491 { 492 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 493 } 494 495 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 496 { 497 int i; 498 499 if (current->mm == mm) 500 sync_mm_rss(mm); 501 for (i = 0; i < NR_MM_COUNTERS; i++) 502 if (rss[i]) 503 add_mm_counter(mm, i, rss[i]); 504 } 505 506 /* 507 * This function is called to print an error when a bad pte 508 * is found. For example, we might have a PFN-mapped pte in 509 * a region that doesn't allow it. 510 * 511 * The calling function must still handle the error. 512 */ 513 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 514 pte_t pte, struct page *page) 515 { 516 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 517 p4d_t *p4d = p4d_offset(pgd, addr); 518 pud_t *pud = pud_offset(p4d, addr); 519 pmd_t *pmd = pmd_offset(pud, addr); 520 struct address_space *mapping; 521 pgoff_t index; 522 static unsigned long resume; 523 static unsigned long nr_shown; 524 static unsigned long nr_unshown; 525 526 /* 527 * Allow a burst of 60 reports, then keep quiet for that minute; 528 * or allow a steady drip of one report per second. 529 */ 530 if (nr_shown == 60) { 531 if (time_before(jiffies, resume)) { 532 nr_unshown++; 533 return; 534 } 535 if (nr_unshown) { 536 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 537 nr_unshown); 538 nr_unshown = 0; 539 } 540 nr_shown = 0; 541 } 542 if (nr_shown++ == 0) 543 resume = jiffies + 60 * HZ; 544 545 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 546 index = linear_page_index(vma, addr); 547 548 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 549 current->comm, 550 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 551 if (page) 552 dump_page(page, "bad pte"); 553 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", 554 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 555 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n", 556 vma->vm_file, 557 vma->vm_ops ? vma->vm_ops->fault : NULL, 558 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 559 mapping ? mapping->a_ops->readpage : NULL); 560 dump_stack(); 561 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 562 } 563 564 /* 565 * vm_normal_page -- This function gets the "struct page" associated with a pte. 566 * 567 * "Special" mappings do not wish to be associated with a "struct page" (either 568 * it doesn't exist, or it exists but they don't want to touch it). In this 569 * case, NULL is returned here. "Normal" mappings do have a struct page. 570 * 571 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 572 * pte bit, in which case this function is trivial. Secondly, an architecture 573 * may not have a spare pte bit, which requires a more complicated scheme, 574 * described below. 575 * 576 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 577 * special mapping (even if there are underlying and valid "struct pages"). 578 * COWed pages of a VM_PFNMAP are always normal. 579 * 580 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 581 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 582 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 583 * mapping will always honor the rule 584 * 585 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 586 * 587 * And for normal mappings this is false. 588 * 589 * This restricts such mappings to be a linear translation from virtual address 590 * to pfn. To get around this restriction, we allow arbitrary mappings so long 591 * as the vma is not a COW mapping; in that case, we know that all ptes are 592 * special (because none can have been COWed). 593 * 594 * 595 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 596 * 597 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 598 * page" backing, however the difference is that _all_ pages with a struct 599 * page (that is, those where pfn_valid is true) are refcounted and considered 600 * normal pages by the VM. The disadvantage is that pages are refcounted 601 * (which can be slower and simply not an option for some PFNMAP users). The 602 * advantage is that we don't have to follow the strict linearity rule of 603 * PFNMAP mappings in order to support COWable mappings. 604 * 605 */ 606 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 607 pte_t pte) 608 { 609 unsigned long pfn = pte_pfn(pte); 610 611 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 612 if (likely(!pte_special(pte))) 613 goto check_pfn; 614 if (vma->vm_ops && vma->vm_ops->find_special_page) 615 return vma->vm_ops->find_special_page(vma, addr); 616 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 617 return NULL; 618 if (is_zero_pfn(pfn)) 619 return NULL; 620 if (pte_devmap(pte)) 621 return NULL; 622 623 print_bad_pte(vma, addr, pte, NULL); 624 return NULL; 625 } 626 627 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 628 629 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 630 if (vma->vm_flags & VM_MIXEDMAP) { 631 if (!pfn_valid(pfn)) 632 return NULL; 633 goto out; 634 } else { 635 unsigned long off; 636 off = (addr - vma->vm_start) >> PAGE_SHIFT; 637 if (pfn == vma->vm_pgoff + off) 638 return NULL; 639 if (!is_cow_mapping(vma->vm_flags)) 640 return NULL; 641 } 642 } 643 644 if (is_zero_pfn(pfn)) 645 return NULL; 646 647 check_pfn: 648 if (unlikely(pfn > highest_memmap_pfn)) { 649 print_bad_pte(vma, addr, pte, NULL); 650 return NULL; 651 } 652 653 /* 654 * NOTE! We still have PageReserved() pages in the page tables. 655 * eg. VDSO mappings can cause them to exist. 656 */ 657 out: 658 return pfn_to_page(pfn); 659 } 660 661 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 662 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 663 pmd_t pmd) 664 { 665 unsigned long pfn = pmd_pfn(pmd); 666 667 /* 668 * There is no pmd_special() but there may be special pmds, e.g. 669 * in a direct-access (dax) mapping, so let's just replicate the 670 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 671 */ 672 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 673 if (vma->vm_flags & VM_MIXEDMAP) { 674 if (!pfn_valid(pfn)) 675 return NULL; 676 goto out; 677 } else { 678 unsigned long off; 679 off = (addr - vma->vm_start) >> PAGE_SHIFT; 680 if (pfn == vma->vm_pgoff + off) 681 return NULL; 682 if (!is_cow_mapping(vma->vm_flags)) 683 return NULL; 684 } 685 } 686 687 if (pmd_devmap(pmd)) 688 return NULL; 689 if (is_huge_zero_pmd(pmd)) 690 return NULL; 691 if (unlikely(pfn > highest_memmap_pfn)) 692 return NULL; 693 694 /* 695 * NOTE! We still have PageReserved() pages in the page tables. 696 * eg. VDSO mappings can cause them to exist. 697 */ 698 out: 699 return pfn_to_page(pfn); 700 } 701 #endif 702 703 /* 704 * copy one vm_area from one task to the other. Assumes the page tables 705 * already present in the new task to be cleared in the whole range 706 * covered by this vma. 707 */ 708 709 static unsigned long 710 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 711 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 712 unsigned long addr, int *rss) 713 { 714 unsigned long vm_flags = vma->vm_flags; 715 pte_t pte = *src_pte; 716 struct page *page; 717 swp_entry_t entry = pte_to_swp_entry(pte); 718 719 if (likely(!non_swap_entry(entry))) { 720 if (swap_duplicate(entry) < 0) 721 return entry.val; 722 723 /* make sure dst_mm is on swapoff's mmlist. */ 724 if (unlikely(list_empty(&dst_mm->mmlist))) { 725 spin_lock(&mmlist_lock); 726 if (list_empty(&dst_mm->mmlist)) 727 list_add(&dst_mm->mmlist, 728 &src_mm->mmlist); 729 spin_unlock(&mmlist_lock); 730 } 731 rss[MM_SWAPENTS]++; 732 } else if (is_migration_entry(entry)) { 733 page = migration_entry_to_page(entry); 734 735 rss[mm_counter(page)]++; 736 737 if (is_write_migration_entry(entry) && 738 is_cow_mapping(vm_flags)) { 739 /* 740 * COW mappings require pages in both 741 * parent and child to be set to read. 742 */ 743 make_migration_entry_read(&entry); 744 pte = swp_entry_to_pte(entry); 745 if (pte_swp_soft_dirty(*src_pte)) 746 pte = pte_swp_mksoft_dirty(pte); 747 if (pte_swp_uffd_wp(*src_pte)) 748 pte = pte_swp_mkuffd_wp(pte); 749 set_pte_at(src_mm, addr, src_pte, pte); 750 } 751 } else if (is_device_private_entry(entry)) { 752 page = device_private_entry_to_page(entry); 753 754 /* 755 * Update rss count even for unaddressable pages, as 756 * they should treated just like normal pages in this 757 * respect. 758 * 759 * We will likely want to have some new rss counters 760 * for unaddressable pages, at some point. But for now 761 * keep things as they are. 762 */ 763 get_page(page); 764 rss[mm_counter(page)]++; 765 page_dup_rmap(page, false); 766 767 /* 768 * We do not preserve soft-dirty information, because so 769 * far, checkpoint/restore is the only feature that 770 * requires that. And checkpoint/restore does not work 771 * when a device driver is involved (you cannot easily 772 * save and restore device driver state). 773 */ 774 if (is_write_device_private_entry(entry) && 775 is_cow_mapping(vm_flags)) { 776 make_device_private_entry_read(&entry); 777 pte = swp_entry_to_pte(entry); 778 if (pte_swp_uffd_wp(*src_pte)) 779 pte = pte_swp_mkuffd_wp(pte); 780 set_pte_at(src_mm, addr, src_pte, pte); 781 } 782 } 783 set_pte_at(dst_mm, addr, dst_pte, pte); 784 return 0; 785 } 786 787 /* 788 * Copy a present and normal page if necessary. 789 * 790 * NOTE! The usual case is that this doesn't need to do 791 * anything, and can just return a positive value. That 792 * will let the caller know that it can just increase 793 * the page refcount and re-use the pte the traditional 794 * way. 795 * 796 * But _if_ we need to copy it because it needs to be 797 * pinned in the parent (and the child should get its own 798 * copy rather than just a reference to the same page), 799 * we'll do that here and return zero to let the caller 800 * know we're done. 801 * 802 * And if we need a pre-allocated page but don't yet have 803 * one, return a negative error to let the preallocation 804 * code know so that it can do so outside the page table 805 * lock. 806 */ 807 static inline int 808 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 809 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 810 struct page **prealloc, pte_t pte, struct page *page) 811 { 812 struct mm_struct *src_mm = src_vma->vm_mm; 813 struct page *new_page; 814 815 if (!is_cow_mapping(src_vma->vm_flags)) 816 return 1; 817 818 /* 819 * What we want to do is to check whether this page may 820 * have been pinned by the parent process. If so, 821 * instead of wrprotect the pte on both sides, we copy 822 * the page immediately so that we'll always guarantee 823 * the pinned page won't be randomly replaced in the 824 * future. 825 * 826 * The page pinning checks are just "has this mm ever 827 * seen pinning", along with the (inexact) check of 828 * the page count. That might give false positives for 829 * for pinning, but it will work correctly. 830 */ 831 if (likely(!atomic_read(&src_mm->has_pinned))) 832 return 1; 833 if (likely(!page_maybe_dma_pinned(page))) 834 return 1; 835 836 new_page = *prealloc; 837 if (!new_page) 838 return -EAGAIN; 839 840 /* 841 * We have a prealloc page, all good! Take it 842 * over and copy the page & arm it. 843 */ 844 *prealloc = NULL; 845 copy_user_highpage(new_page, page, addr, src_vma); 846 __SetPageUptodate(new_page); 847 page_add_new_anon_rmap(new_page, dst_vma, addr, false); 848 lru_cache_add_inactive_or_unevictable(new_page, dst_vma); 849 rss[mm_counter(new_page)]++; 850 851 /* All done, just insert the new page copy in the child */ 852 pte = mk_pte(new_page, dst_vma->vm_page_prot); 853 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); 854 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 855 return 0; 856 } 857 858 /* 859 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page 860 * is required to copy this pte. 861 */ 862 static inline int 863 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 864 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 865 struct page **prealloc) 866 { 867 struct mm_struct *src_mm = src_vma->vm_mm; 868 unsigned long vm_flags = src_vma->vm_flags; 869 pte_t pte = *src_pte; 870 struct page *page; 871 872 page = vm_normal_page(src_vma, addr, pte); 873 if (page) { 874 int retval; 875 876 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte, 877 addr, rss, prealloc, pte, page); 878 if (retval <= 0) 879 return retval; 880 881 get_page(page); 882 page_dup_rmap(page, false); 883 rss[mm_counter(page)]++; 884 } 885 886 /* 887 * If it's a COW mapping, write protect it both 888 * in the parent and the child 889 */ 890 if (is_cow_mapping(vm_flags) && pte_write(pte)) { 891 ptep_set_wrprotect(src_mm, addr, src_pte); 892 pte = pte_wrprotect(pte); 893 } 894 895 /* 896 * If it's a shared mapping, mark it clean in 897 * the child 898 */ 899 if (vm_flags & VM_SHARED) 900 pte = pte_mkclean(pte); 901 pte = pte_mkold(pte); 902 903 /* 904 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA 905 * does not have the VM_UFFD_WP, which means that the uffd 906 * fork event is not enabled. 907 */ 908 if (!(vm_flags & VM_UFFD_WP)) 909 pte = pte_clear_uffd_wp(pte); 910 911 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 912 return 0; 913 } 914 915 static inline struct page * 916 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma, 917 unsigned long addr) 918 { 919 struct page *new_page; 920 921 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr); 922 if (!new_page) 923 return NULL; 924 925 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) { 926 put_page(new_page); 927 return NULL; 928 } 929 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 930 931 return new_page; 932 } 933 934 static int 935 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 936 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 937 unsigned long end) 938 { 939 struct mm_struct *dst_mm = dst_vma->vm_mm; 940 struct mm_struct *src_mm = src_vma->vm_mm; 941 pte_t *orig_src_pte, *orig_dst_pte; 942 pte_t *src_pte, *dst_pte; 943 spinlock_t *src_ptl, *dst_ptl; 944 int progress, ret = 0; 945 int rss[NR_MM_COUNTERS]; 946 swp_entry_t entry = (swp_entry_t){0}; 947 struct page *prealloc = NULL; 948 949 again: 950 progress = 0; 951 init_rss_vec(rss); 952 953 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 954 if (!dst_pte) { 955 ret = -ENOMEM; 956 goto out; 957 } 958 src_pte = pte_offset_map(src_pmd, addr); 959 src_ptl = pte_lockptr(src_mm, src_pmd); 960 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 961 orig_src_pte = src_pte; 962 orig_dst_pte = dst_pte; 963 arch_enter_lazy_mmu_mode(); 964 965 do { 966 /* 967 * We are holding two locks at this point - either of them 968 * could generate latencies in another task on another CPU. 969 */ 970 if (progress >= 32) { 971 progress = 0; 972 if (need_resched() || 973 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 974 break; 975 } 976 if (pte_none(*src_pte)) { 977 progress++; 978 continue; 979 } 980 if (unlikely(!pte_present(*src_pte))) { 981 entry.val = copy_nonpresent_pte(dst_mm, src_mm, 982 dst_pte, src_pte, 983 src_vma, addr, rss); 984 if (entry.val) 985 break; 986 progress += 8; 987 continue; 988 } 989 /* copy_present_pte() will clear `*prealloc' if consumed */ 990 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte, 991 addr, rss, &prealloc); 992 /* 993 * If we need a pre-allocated page for this pte, drop the 994 * locks, allocate, and try again. 995 */ 996 if (unlikely(ret == -EAGAIN)) 997 break; 998 if (unlikely(prealloc)) { 999 /* 1000 * pre-alloc page cannot be reused by next time so as 1001 * to strictly follow mempolicy (e.g., alloc_page_vma() 1002 * will allocate page according to address). This 1003 * could only happen if one pinned pte changed. 1004 */ 1005 put_page(prealloc); 1006 prealloc = NULL; 1007 } 1008 progress += 8; 1009 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 1010 1011 arch_leave_lazy_mmu_mode(); 1012 spin_unlock(src_ptl); 1013 pte_unmap(orig_src_pte); 1014 add_mm_rss_vec(dst_mm, rss); 1015 pte_unmap_unlock(orig_dst_pte, dst_ptl); 1016 cond_resched(); 1017 1018 if (entry.val) { 1019 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { 1020 ret = -ENOMEM; 1021 goto out; 1022 } 1023 entry.val = 0; 1024 } else if (ret) { 1025 WARN_ON_ONCE(ret != -EAGAIN); 1026 prealloc = page_copy_prealloc(src_mm, src_vma, addr); 1027 if (!prealloc) 1028 return -ENOMEM; 1029 /* We've captured and resolved the error. Reset, try again. */ 1030 ret = 0; 1031 } 1032 if (addr != end) 1033 goto again; 1034 out: 1035 if (unlikely(prealloc)) 1036 put_page(prealloc); 1037 return ret; 1038 } 1039 1040 static inline int 1041 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1042 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1043 unsigned long end) 1044 { 1045 struct mm_struct *dst_mm = dst_vma->vm_mm; 1046 struct mm_struct *src_mm = src_vma->vm_mm; 1047 pmd_t *src_pmd, *dst_pmd; 1048 unsigned long next; 1049 1050 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 1051 if (!dst_pmd) 1052 return -ENOMEM; 1053 src_pmd = pmd_offset(src_pud, addr); 1054 do { 1055 next = pmd_addr_end(addr, end); 1056 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 1057 || pmd_devmap(*src_pmd)) { 1058 int err; 1059 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); 1060 err = copy_huge_pmd(dst_mm, src_mm, 1061 dst_pmd, src_pmd, addr, src_vma); 1062 if (err == -ENOMEM) 1063 return -ENOMEM; 1064 if (!err) 1065 continue; 1066 /* fall through */ 1067 } 1068 if (pmd_none_or_clear_bad(src_pmd)) 1069 continue; 1070 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, 1071 addr, next)) 1072 return -ENOMEM; 1073 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 1074 return 0; 1075 } 1076 1077 static inline int 1078 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1079 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, 1080 unsigned long end) 1081 { 1082 struct mm_struct *dst_mm = dst_vma->vm_mm; 1083 struct mm_struct *src_mm = src_vma->vm_mm; 1084 pud_t *src_pud, *dst_pud; 1085 unsigned long next; 1086 1087 dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 1088 if (!dst_pud) 1089 return -ENOMEM; 1090 src_pud = pud_offset(src_p4d, addr); 1091 do { 1092 next = pud_addr_end(addr, end); 1093 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 1094 int err; 1095 1096 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); 1097 err = copy_huge_pud(dst_mm, src_mm, 1098 dst_pud, src_pud, addr, src_vma); 1099 if (err == -ENOMEM) 1100 return -ENOMEM; 1101 if (!err) 1102 continue; 1103 /* fall through */ 1104 } 1105 if (pud_none_or_clear_bad(src_pud)) 1106 continue; 1107 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, 1108 addr, next)) 1109 return -ENOMEM; 1110 } while (dst_pud++, src_pud++, addr = next, addr != end); 1111 return 0; 1112 } 1113 1114 static inline int 1115 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1116 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, 1117 unsigned long end) 1118 { 1119 struct mm_struct *dst_mm = dst_vma->vm_mm; 1120 p4d_t *src_p4d, *dst_p4d; 1121 unsigned long next; 1122 1123 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 1124 if (!dst_p4d) 1125 return -ENOMEM; 1126 src_p4d = p4d_offset(src_pgd, addr); 1127 do { 1128 next = p4d_addr_end(addr, end); 1129 if (p4d_none_or_clear_bad(src_p4d)) 1130 continue; 1131 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, 1132 addr, next)) 1133 return -ENOMEM; 1134 } while (dst_p4d++, src_p4d++, addr = next, addr != end); 1135 return 0; 1136 } 1137 1138 int 1139 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1140 { 1141 pgd_t *src_pgd, *dst_pgd; 1142 unsigned long next; 1143 unsigned long addr = src_vma->vm_start; 1144 unsigned long end = src_vma->vm_end; 1145 struct mm_struct *dst_mm = dst_vma->vm_mm; 1146 struct mm_struct *src_mm = src_vma->vm_mm; 1147 struct mmu_notifier_range range; 1148 bool is_cow; 1149 int ret; 1150 1151 /* 1152 * Don't copy ptes where a page fault will fill them correctly. 1153 * Fork becomes much lighter when there are big shared or private 1154 * readonly mappings. The tradeoff is that copy_page_range is more 1155 * efficient than faulting. 1156 */ 1157 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 1158 !src_vma->anon_vma) 1159 return 0; 1160 1161 if (is_vm_hugetlb_page(src_vma)) 1162 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma); 1163 1164 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { 1165 /* 1166 * We do not free on error cases below as remove_vma 1167 * gets called on error from higher level routine 1168 */ 1169 ret = track_pfn_copy(src_vma); 1170 if (ret) 1171 return ret; 1172 } 1173 1174 /* 1175 * We need to invalidate the secondary MMU mappings only when 1176 * there could be a permission downgrade on the ptes of the 1177 * parent mm. And a permission downgrade will only happen if 1178 * is_cow_mapping() returns true. 1179 */ 1180 is_cow = is_cow_mapping(src_vma->vm_flags); 1181 1182 if (is_cow) { 1183 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 1184 0, src_vma, src_mm, addr, end); 1185 mmu_notifier_invalidate_range_start(&range); 1186 /* 1187 * Disabling preemption is not needed for the write side, as 1188 * the read side doesn't spin, but goes to the mmap_lock. 1189 * 1190 * Use the raw variant of the seqcount_t write API to avoid 1191 * lockdep complaining about preemptibility. 1192 */ 1193 mmap_assert_write_locked(src_mm); 1194 raw_write_seqcount_begin(&src_mm->write_protect_seq); 1195 } 1196 1197 ret = 0; 1198 dst_pgd = pgd_offset(dst_mm, addr); 1199 src_pgd = pgd_offset(src_mm, addr); 1200 do { 1201 next = pgd_addr_end(addr, end); 1202 if (pgd_none_or_clear_bad(src_pgd)) 1203 continue; 1204 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, 1205 addr, next))) { 1206 ret = -ENOMEM; 1207 break; 1208 } 1209 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1210 1211 if (is_cow) { 1212 raw_write_seqcount_end(&src_mm->write_protect_seq); 1213 mmu_notifier_invalidate_range_end(&range); 1214 } 1215 return ret; 1216 } 1217 1218 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1219 struct vm_area_struct *vma, pmd_t *pmd, 1220 unsigned long addr, unsigned long end, 1221 struct zap_details *details) 1222 { 1223 struct mm_struct *mm = tlb->mm; 1224 int force_flush = 0; 1225 int rss[NR_MM_COUNTERS]; 1226 spinlock_t *ptl; 1227 pte_t *start_pte; 1228 pte_t *pte; 1229 swp_entry_t entry; 1230 1231 tlb_change_page_size(tlb, PAGE_SIZE); 1232 again: 1233 init_rss_vec(rss); 1234 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1235 pte = start_pte; 1236 flush_tlb_batched_pending(mm); 1237 arch_enter_lazy_mmu_mode(); 1238 do { 1239 pte_t ptent = *pte; 1240 if (pte_none(ptent)) 1241 continue; 1242 1243 if (need_resched()) 1244 break; 1245 1246 if (pte_present(ptent)) { 1247 struct page *page; 1248 1249 page = vm_normal_page(vma, addr, ptent); 1250 if (unlikely(details) && page) { 1251 /* 1252 * unmap_shared_mapping_pages() wants to 1253 * invalidate cache without truncating: 1254 * unmap shared but keep private pages. 1255 */ 1256 if (details->check_mapping && 1257 details->check_mapping != page_rmapping(page)) 1258 continue; 1259 } 1260 ptent = ptep_get_and_clear_full(mm, addr, pte, 1261 tlb->fullmm); 1262 tlb_remove_tlb_entry(tlb, pte, addr); 1263 if (unlikely(!page)) 1264 continue; 1265 1266 if (!PageAnon(page)) { 1267 if (pte_dirty(ptent)) { 1268 force_flush = 1; 1269 set_page_dirty(page); 1270 } 1271 if (pte_young(ptent) && 1272 likely(!(vma->vm_flags & VM_SEQ_READ))) 1273 mark_page_accessed(page); 1274 } 1275 rss[mm_counter(page)]--; 1276 page_remove_rmap(page, false); 1277 if (unlikely(page_mapcount(page) < 0)) 1278 print_bad_pte(vma, addr, ptent, page); 1279 if (unlikely(__tlb_remove_page(tlb, page))) { 1280 force_flush = 1; 1281 addr += PAGE_SIZE; 1282 break; 1283 } 1284 continue; 1285 } 1286 1287 entry = pte_to_swp_entry(ptent); 1288 if (is_device_private_entry(entry)) { 1289 struct page *page = device_private_entry_to_page(entry); 1290 1291 if (unlikely(details && details->check_mapping)) { 1292 /* 1293 * unmap_shared_mapping_pages() wants to 1294 * invalidate cache without truncating: 1295 * unmap shared but keep private pages. 1296 */ 1297 if (details->check_mapping != 1298 page_rmapping(page)) 1299 continue; 1300 } 1301 1302 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1303 rss[mm_counter(page)]--; 1304 page_remove_rmap(page, false); 1305 put_page(page); 1306 continue; 1307 } 1308 1309 /* If details->check_mapping, we leave swap entries. */ 1310 if (unlikely(details)) 1311 continue; 1312 1313 if (!non_swap_entry(entry)) 1314 rss[MM_SWAPENTS]--; 1315 else if (is_migration_entry(entry)) { 1316 struct page *page; 1317 1318 page = migration_entry_to_page(entry); 1319 rss[mm_counter(page)]--; 1320 } 1321 if (unlikely(!free_swap_and_cache(entry))) 1322 print_bad_pte(vma, addr, ptent, NULL); 1323 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1324 } while (pte++, addr += PAGE_SIZE, addr != end); 1325 1326 add_mm_rss_vec(mm, rss); 1327 arch_leave_lazy_mmu_mode(); 1328 1329 /* Do the actual TLB flush before dropping ptl */ 1330 if (force_flush) 1331 tlb_flush_mmu_tlbonly(tlb); 1332 pte_unmap_unlock(start_pte, ptl); 1333 1334 /* 1335 * If we forced a TLB flush (either due to running out of 1336 * batch buffers or because we needed to flush dirty TLB 1337 * entries before releasing the ptl), free the batched 1338 * memory too. Restart if we didn't do everything. 1339 */ 1340 if (force_flush) { 1341 force_flush = 0; 1342 tlb_flush_mmu(tlb); 1343 } 1344 1345 if (addr != end) { 1346 cond_resched(); 1347 goto again; 1348 } 1349 1350 return addr; 1351 } 1352 1353 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1354 struct vm_area_struct *vma, pud_t *pud, 1355 unsigned long addr, unsigned long end, 1356 struct zap_details *details) 1357 { 1358 pmd_t *pmd; 1359 unsigned long next; 1360 1361 pmd = pmd_offset(pud, addr); 1362 do { 1363 next = pmd_addr_end(addr, end); 1364 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1365 if (next - addr != HPAGE_PMD_SIZE) 1366 __split_huge_pmd(vma, pmd, addr, false, NULL); 1367 else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1368 goto next; 1369 /* fall through */ 1370 } 1371 /* 1372 * Here there can be other concurrent MADV_DONTNEED or 1373 * trans huge page faults running, and if the pmd is 1374 * none or trans huge it can change under us. This is 1375 * because MADV_DONTNEED holds the mmap_lock in read 1376 * mode. 1377 */ 1378 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1379 goto next; 1380 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1381 next: 1382 cond_resched(); 1383 } while (pmd++, addr = next, addr != end); 1384 1385 return addr; 1386 } 1387 1388 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1389 struct vm_area_struct *vma, p4d_t *p4d, 1390 unsigned long addr, unsigned long end, 1391 struct zap_details *details) 1392 { 1393 pud_t *pud; 1394 unsigned long next; 1395 1396 pud = pud_offset(p4d, addr); 1397 do { 1398 next = pud_addr_end(addr, end); 1399 if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1400 if (next - addr != HPAGE_PUD_SIZE) { 1401 mmap_assert_locked(tlb->mm); 1402 split_huge_pud(vma, pud, addr); 1403 } else if (zap_huge_pud(tlb, vma, pud, addr)) 1404 goto next; 1405 /* fall through */ 1406 } 1407 if (pud_none_or_clear_bad(pud)) 1408 continue; 1409 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1410 next: 1411 cond_resched(); 1412 } while (pud++, addr = next, addr != end); 1413 1414 return addr; 1415 } 1416 1417 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1418 struct vm_area_struct *vma, pgd_t *pgd, 1419 unsigned long addr, unsigned long end, 1420 struct zap_details *details) 1421 { 1422 p4d_t *p4d; 1423 unsigned long next; 1424 1425 p4d = p4d_offset(pgd, addr); 1426 do { 1427 next = p4d_addr_end(addr, end); 1428 if (p4d_none_or_clear_bad(p4d)) 1429 continue; 1430 next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1431 } while (p4d++, addr = next, addr != end); 1432 1433 return addr; 1434 } 1435 1436 void unmap_page_range(struct mmu_gather *tlb, 1437 struct vm_area_struct *vma, 1438 unsigned long addr, unsigned long end, 1439 struct zap_details *details) 1440 { 1441 pgd_t *pgd; 1442 unsigned long next; 1443 1444 BUG_ON(addr >= end); 1445 tlb_start_vma(tlb, vma); 1446 pgd = pgd_offset(vma->vm_mm, addr); 1447 do { 1448 next = pgd_addr_end(addr, end); 1449 if (pgd_none_or_clear_bad(pgd)) 1450 continue; 1451 next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 1452 } while (pgd++, addr = next, addr != end); 1453 tlb_end_vma(tlb, vma); 1454 } 1455 1456 1457 static void unmap_single_vma(struct mmu_gather *tlb, 1458 struct vm_area_struct *vma, unsigned long start_addr, 1459 unsigned long end_addr, 1460 struct zap_details *details) 1461 { 1462 unsigned long start = max(vma->vm_start, start_addr); 1463 unsigned long end; 1464 1465 if (start >= vma->vm_end) 1466 return; 1467 end = min(vma->vm_end, end_addr); 1468 if (end <= vma->vm_start) 1469 return; 1470 1471 if (vma->vm_file) 1472 uprobe_munmap(vma, start, end); 1473 1474 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1475 untrack_pfn(vma, 0, 0); 1476 1477 if (start != end) { 1478 if (unlikely(is_vm_hugetlb_page(vma))) { 1479 /* 1480 * It is undesirable to test vma->vm_file as it 1481 * should be non-null for valid hugetlb area. 1482 * However, vm_file will be NULL in the error 1483 * cleanup path of mmap_region. When 1484 * hugetlbfs ->mmap method fails, 1485 * mmap_region() nullifies vma->vm_file 1486 * before calling this function to clean up. 1487 * Since no pte has actually been setup, it is 1488 * safe to do nothing in this case. 1489 */ 1490 if (vma->vm_file) { 1491 i_mmap_lock_write(vma->vm_file->f_mapping); 1492 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1493 i_mmap_unlock_write(vma->vm_file->f_mapping); 1494 } 1495 } else 1496 unmap_page_range(tlb, vma, start, end, details); 1497 } 1498 } 1499 1500 /** 1501 * unmap_vmas - unmap a range of memory covered by a list of vma's 1502 * @tlb: address of the caller's struct mmu_gather 1503 * @vma: the starting vma 1504 * @start_addr: virtual address at which to start unmapping 1505 * @end_addr: virtual address at which to end unmapping 1506 * 1507 * Unmap all pages in the vma list. 1508 * 1509 * Only addresses between `start' and `end' will be unmapped. 1510 * 1511 * The VMA list must be sorted in ascending virtual address order. 1512 * 1513 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1514 * range after unmap_vmas() returns. So the only responsibility here is to 1515 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1516 * drops the lock and schedules. 1517 */ 1518 void unmap_vmas(struct mmu_gather *tlb, 1519 struct vm_area_struct *vma, unsigned long start_addr, 1520 unsigned long end_addr) 1521 { 1522 struct mmu_notifier_range range; 1523 1524 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, 1525 start_addr, end_addr); 1526 mmu_notifier_invalidate_range_start(&range); 1527 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1528 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1529 mmu_notifier_invalidate_range_end(&range); 1530 } 1531 1532 /** 1533 * zap_page_range - remove user pages in a given range 1534 * @vma: vm_area_struct holding the applicable pages 1535 * @start: starting address of pages to zap 1536 * @size: number of bytes to zap 1537 * 1538 * Caller must protect the VMA list 1539 */ 1540 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1541 unsigned long size) 1542 { 1543 struct mmu_notifier_range range; 1544 struct mmu_gather tlb; 1545 1546 lru_add_drain(); 1547 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1548 start, start + size); 1549 tlb_gather_mmu(&tlb, vma->vm_mm); 1550 update_hiwater_rss(vma->vm_mm); 1551 mmu_notifier_invalidate_range_start(&range); 1552 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next) 1553 unmap_single_vma(&tlb, vma, start, range.end, NULL); 1554 mmu_notifier_invalidate_range_end(&range); 1555 tlb_finish_mmu(&tlb); 1556 } 1557 1558 /** 1559 * zap_page_range_single - remove user pages in a given range 1560 * @vma: vm_area_struct holding the applicable pages 1561 * @address: starting address of pages to zap 1562 * @size: number of bytes to zap 1563 * @details: details of shared cache invalidation 1564 * 1565 * The range must fit into one VMA. 1566 */ 1567 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1568 unsigned long size, struct zap_details *details) 1569 { 1570 struct mmu_notifier_range range; 1571 struct mmu_gather tlb; 1572 1573 lru_add_drain(); 1574 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1575 address, address + size); 1576 tlb_gather_mmu(&tlb, vma->vm_mm); 1577 update_hiwater_rss(vma->vm_mm); 1578 mmu_notifier_invalidate_range_start(&range); 1579 unmap_single_vma(&tlb, vma, address, range.end, details); 1580 mmu_notifier_invalidate_range_end(&range); 1581 tlb_finish_mmu(&tlb); 1582 } 1583 1584 /** 1585 * zap_vma_ptes - remove ptes mapping the vma 1586 * @vma: vm_area_struct holding ptes to be zapped 1587 * @address: starting address of pages to zap 1588 * @size: number of bytes to zap 1589 * 1590 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1591 * 1592 * The entire address range must be fully contained within the vma. 1593 * 1594 */ 1595 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1596 unsigned long size) 1597 { 1598 if (address < vma->vm_start || address + size > vma->vm_end || 1599 !(vma->vm_flags & VM_PFNMAP)) 1600 return; 1601 1602 zap_page_range_single(vma, address, size, NULL); 1603 } 1604 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1605 1606 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) 1607 { 1608 pgd_t *pgd; 1609 p4d_t *p4d; 1610 pud_t *pud; 1611 pmd_t *pmd; 1612 1613 pgd = pgd_offset(mm, addr); 1614 p4d = p4d_alloc(mm, pgd, addr); 1615 if (!p4d) 1616 return NULL; 1617 pud = pud_alloc(mm, p4d, addr); 1618 if (!pud) 1619 return NULL; 1620 pmd = pmd_alloc(mm, pud, addr); 1621 if (!pmd) 1622 return NULL; 1623 1624 VM_BUG_ON(pmd_trans_huge(*pmd)); 1625 return pmd; 1626 } 1627 1628 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1629 spinlock_t **ptl) 1630 { 1631 pmd_t *pmd = walk_to_pmd(mm, addr); 1632 1633 if (!pmd) 1634 return NULL; 1635 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1636 } 1637 1638 static int validate_page_before_insert(struct page *page) 1639 { 1640 if (PageAnon(page) || PageSlab(page) || page_has_type(page)) 1641 return -EINVAL; 1642 flush_dcache_page(page); 1643 return 0; 1644 } 1645 1646 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte, 1647 unsigned long addr, struct page *page, pgprot_t prot) 1648 { 1649 if (!pte_none(*pte)) 1650 return -EBUSY; 1651 /* Ok, finally just insert the thing.. */ 1652 get_page(page); 1653 inc_mm_counter_fast(mm, mm_counter_file(page)); 1654 page_add_file_rmap(page, false); 1655 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1656 return 0; 1657 } 1658 1659 /* 1660 * This is the old fallback for page remapping. 1661 * 1662 * For historical reasons, it only allows reserved pages. Only 1663 * old drivers should use this, and they needed to mark their 1664 * pages reserved for the old functions anyway. 1665 */ 1666 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1667 struct page *page, pgprot_t prot) 1668 { 1669 struct mm_struct *mm = vma->vm_mm; 1670 int retval; 1671 pte_t *pte; 1672 spinlock_t *ptl; 1673 1674 retval = validate_page_before_insert(page); 1675 if (retval) 1676 goto out; 1677 retval = -ENOMEM; 1678 pte = get_locked_pte(mm, addr, &ptl); 1679 if (!pte) 1680 goto out; 1681 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot); 1682 pte_unmap_unlock(pte, ptl); 1683 out: 1684 return retval; 1685 } 1686 1687 #ifdef pte_index 1688 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte, 1689 unsigned long addr, struct page *page, pgprot_t prot) 1690 { 1691 int err; 1692 1693 if (!page_count(page)) 1694 return -EINVAL; 1695 err = validate_page_before_insert(page); 1696 if (err) 1697 return err; 1698 return insert_page_into_pte_locked(mm, pte, addr, page, prot); 1699 } 1700 1701 /* insert_pages() amortizes the cost of spinlock operations 1702 * when inserting pages in a loop. Arch *must* define pte_index. 1703 */ 1704 static int insert_pages(struct vm_area_struct *vma, unsigned long addr, 1705 struct page **pages, unsigned long *num, pgprot_t prot) 1706 { 1707 pmd_t *pmd = NULL; 1708 pte_t *start_pte, *pte; 1709 spinlock_t *pte_lock; 1710 struct mm_struct *const mm = vma->vm_mm; 1711 unsigned long curr_page_idx = 0; 1712 unsigned long remaining_pages_total = *num; 1713 unsigned long pages_to_write_in_pmd; 1714 int ret; 1715 more: 1716 ret = -EFAULT; 1717 pmd = walk_to_pmd(mm, addr); 1718 if (!pmd) 1719 goto out; 1720 1721 pages_to_write_in_pmd = min_t(unsigned long, 1722 remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); 1723 1724 /* Allocate the PTE if necessary; takes PMD lock once only. */ 1725 ret = -ENOMEM; 1726 if (pte_alloc(mm, pmd)) 1727 goto out; 1728 1729 while (pages_to_write_in_pmd) { 1730 int pte_idx = 0; 1731 const int batch_size = min_t(int, pages_to_write_in_pmd, 8); 1732 1733 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); 1734 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { 1735 int err = insert_page_in_batch_locked(mm, pte, 1736 addr, pages[curr_page_idx], prot); 1737 if (unlikely(err)) { 1738 pte_unmap_unlock(start_pte, pte_lock); 1739 ret = err; 1740 remaining_pages_total -= pte_idx; 1741 goto out; 1742 } 1743 addr += PAGE_SIZE; 1744 ++curr_page_idx; 1745 } 1746 pte_unmap_unlock(start_pte, pte_lock); 1747 pages_to_write_in_pmd -= batch_size; 1748 remaining_pages_total -= batch_size; 1749 } 1750 if (remaining_pages_total) 1751 goto more; 1752 ret = 0; 1753 out: 1754 *num = remaining_pages_total; 1755 return ret; 1756 } 1757 #endif /* ifdef pte_index */ 1758 1759 /** 1760 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. 1761 * @vma: user vma to map to 1762 * @addr: target start user address of these pages 1763 * @pages: source kernel pages 1764 * @num: in: number of pages to map. out: number of pages that were *not* 1765 * mapped. (0 means all pages were successfully mapped). 1766 * 1767 * Preferred over vm_insert_page() when inserting multiple pages. 1768 * 1769 * In case of error, we may have mapped a subset of the provided 1770 * pages. It is the caller's responsibility to account for this case. 1771 * 1772 * The same restrictions apply as in vm_insert_page(). 1773 */ 1774 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, 1775 struct page **pages, unsigned long *num) 1776 { 1777 #ifdef pte_index 1778 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; 1779 1780 if (addr < vma->vm_start || end_addr >= vma->vm_end) 1781 return -EFAULT; 1782 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1783 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1784 BUG_ON(vma->vm_flags & VM_PFNMAP); 1785 vma->vm_flags |= VM_MIXEDMAP; 1786 } 1787 /* Defer page refcount checking till we're about to map that page. */ 1788 return insert_pages(vma, addr, pages, num, vma->vm_page_prot); 1789 #else 1790 unsigned long idx = 0, pgcount = *num; 1791 int err = -EINVAL; 1792 1793 for (; idx < pgcount; ++idx) { 1794 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]); 1795 if (err) 1796 break; 1797 } 1798 *num = pgcount - idx; 1799 return err; 1800 #endif /* ifdef pte_index */ 1801 } 1802 EXPORT_SYMBOL(vm_insert_pages); 1803 1804 /** 1805 * vm_insert_page - insert single page into user vma 1806 * @vma: user vma to map to 1807 * @addr: target user address of this page 1808 * @page: source kernel page 1809 * 1810 * This allows drivers to insert individual pages they've allocated 1811 * into a user vma. 1812 * 1813 * The page has to be a nice clean _individual_ kernel allocation. 1814 * If you allocate a compound page, you need to have marked it as 1815 * such (__GFP_COMP), or manually just split the page up yourself 1816 * (see split_page()). 1817 * 1818 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1819 * took an arbitrary page protection parameter. This doesn't allow 1820 * that. Your vma protection will have to be set up correctly, which 1821 * means that if you want a shared writable mapping, you'd better 1822 * ask for a shared writable mapping! 1823 * 1824 * The page does not need to be reserved. 1825 * 1826 * Usually this function is called from f_op->mmap() handler 1827 * under mm->mmap_lock write-lock, so it can change vma->vm_flags. 1828 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1829 * function from other places, for example from page-fault handler. 1830 * 1831 * Return: %0 on success, negative error code otherwise. 1832 */ 1833 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1834 struct page *page) 1835 { 1836 if (addr < vma->vm_start || addr >= vma->vm_end) 1837 return -EFAULT; 1838 if (!page_count(page)) 1839 return -EINVAL; 1840 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1841 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1842 BUG_ON(vma->vm_flags & VM_PFNMAP); 1843 vma->vm_flags |= VM_MIXEDMAP; 1844 } 1845 return insert_page(vma, addr, page, vma->vm_page_prot); 1846 } 1847 EXPORT_SYMBOL(vm_insert_page); 1848 1849 /* 1850 * __vm_map_pages - maps range of kernel pages into user vma 1851 * @vma: user vma to map to 1852 * @pages: pointer to array of source kernel pages 1853 * @num: number of pages in page array 1854 * @offset: user's requested vm_pgoff 1855 * 1856 * This allows drivers to map range of kernel pages into a user vma. 1857 * 1858 * Return: 0 on success and error code otherwise. 1859 */ 1860 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1861 unsigned long num, unsigned long offset) 1862 { 1863 unsigned long count = vma_pages(vma); 1864 unsigned long uaddr = vma->vm_start; 1865 int ret, i; 1866 1867 /* Fail if the user requested offset is beyond the end of the object */ 1868 if (offset >= num) 1869 return -ENXIO; 1870 1871 /* Fail if the user requested size exceeds available object size */ 1872 if (count > num - offset) 1873 return -ENXIO; 1874 1875 for (i = 0; i < count; i++) { 1876 ret = vm_insert_page(vma, uaddr, pages[offset + i]); 1877 if (ret < 0) 1878 return ret; 1879 uaddr += PAGE_SIZE; 1880 } 1881 1882 return 0; 1883 } 1884 1885 /** 1886 * vm_map_pages - maps range of kernel pages starts with non zero offset 1887 * @vma: user vma to map to 1888 * @pages: pointer to array of source kernel pages 1889 * @num: number of pages in page array 1890 * 1891 * Maps an object consisting of @num pages, catering for the user's 1892 * requested vm_pgoff 1893 * 1894 * If we fail to insert any page into the vma, the function will return 1895 * immediately leaving any previously inserted pages present. Callers 1896 * from the mmap handler may immediately return the error as their caller 1897 * will destroy the vma, removing any successfully inserted pages. Other 1898 * callers should make their own arrangements for calling unmap_region(). 1899 * 1900 * Context: Process context. Called by mmap handlers. 1901 * Return: 0 on success and error code otherwise. 1902 */ 1903 int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1904 unsigned long num) 1905 { 1906 return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 1907 } 1908 EXPORT_SYMBOL(vm_map_pages); 1909 1910 /** 1911 * vm_map_pages_zero - map range of kernel pages starts with zero offset 1912 * @vma: user vma to map to 1913 * @pages: pointer to array of source kernel pages 1914 * @num: number of pages in page array 1915 * 1916 * Similar to vm_map_pages(), except that it explicitly sets the offset 1917 * to 0. This function is intended for the drivers that did not consider 1918 * vm_pgoff. 1919 * 1920 * Context: Process context. Called by mmap handlers. 1921 * Return: 0 on success and error code otherwise. 1922 */ 1923 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 1924 unsigned long num) 1925 { 1926 return __vm_map_pages(vma, pages, num, 0); 1927 } 1928 EXPORT_SYMBOL(vm_map_pages_zero); 1929 1930 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1931 pfn_t pfn, pgprot_t prot, bool mkwrite) 1932 { 1933 struct mm_struct *mm = vma->vm_mm; 1934 pte_t *pte, entry; 1935 spinlock_t *ptl; 1936 1937 pte = get_locked_pte(mm, addr, &ptl); 1938 if (!pte) 1939 return VM_FAULT_OOM; 1940 if (!pte_none(*pte)) { 1941 if (mkwrite) { 1942 /* 1943 * For read faults on private mappings the PFN passed 1944 * in may not match the PFN we have mapped if the 1945 * mapped PFN is a writeable COW page. In the mkwrite 1946 * case we are creating a writable PTE for a shared 1947 * mapping and we expect the PFNs to match. If they 1948 * don't match, we are likely racing with block 1949 * allocation and mapping invalidation so just skip the 1950 * update. 1951 */ 1952 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) { 1953 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte))); 1954 goto out_unlock; 1955 } 1956 entry = pte_mkyoung(*pte); 1957 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1958 if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 1959 update_mmu_cache(vma, addr, pte); 1960 } 1961 goto out_unlock; 1962 } 1963 1964 /* Ok, finally just insert the thing.. */ 1965 if (pfn_t_devmap(pfn)) 1966 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 1967 else 1968 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 1969 1970 if (mkwrite) { 1971 entry = pte_mkyoung(entry); 1972 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1973 } 1974 1975 set_pte_at(mm, addr, pte, entry); 1976 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 1977 1978 out_unlock: 1979 pte_unmap_unlock(pte, ptl); 1980 return VM_FAULT_NOPAGE; 1981 } 1982 1983 /** 1984 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 1985 * @vma: user vma to map to 1986 * @addr: target user address of this page 1987 * @pfn: source kernel pfn 1988 * @pgprot: pgprot flags for the inserted page 1989 * 1990 * This is exactly like vmf_insert_pfn(), except that it allows drivers 1991 * to override pgprot on a per-page basis. 1992 * 1993 * This only makes sense for IO mappings, and it makes no sense for 1994 * COW mappings. In general, using multiple vmas is preferable; 1995 * vmf_insert_pfn_prot should only be used if using multiple VMAs is 1996 * impractical. 1997 * 1998 * See vmf_insert_mixed_prot() for a discussion of the implication of using 1999 * a value of @pgprot different from that of @vma->vm_page_prot. 2000 * 2001 * Context: Process context. May allocate using %GFP_KERNEL. 2002 * Return: vm_fault_t value. 2003 */ 2004 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 2005 unsigned long pfn, pgprot_t pgprot) 2006 { 2007 /* 2008 * Technically, architectures with pte_special can avoid all these 2009 * restrictions (same for remap_pfn_range). However we would like 2010 * consistency in testing and feature parity among all, so we should 2011 * try to keep these invariants in place for everybody. 2012 */ 2013 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 2014 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 2015 (VM_PFNMAP|VM_MIXEDMAP)); 2016 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 2017 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 2018 2019 if (addr < vma->vm_start || addr >= vma->vm_end) 2020 return VM_FAULT_SIGBUS; 2021 2022 if (!pfn_modify_allowed(pfn, pgprot)) 2023 return VM_FAULT_SIGBUS; 2024 2025 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 2026 2027 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 2028 false); 2029 } 2030 EXPORT_SYMBOL(vmf_insert_pfn_prot); 2031 2032 /** 2033 * vmf_insert_pfn - insert single pfn into user vma 2034 * @vma: user vma to map to 2035 * @addr: target user address of this page 2036 * @pfn: source kernel pfn 2037 * 2038 * Similar to vm_insert_page, this allows drivers to insert individual pages 2039 * they've allocated into a user vma. Same comments apply. 2040 * 2041 * This function should only be called from a vm_ops->fault handler, and 2042 * in that case the handler should return the result of this function. 2043 * 2044 * vma cannot be a COW mapping. 2045 * 2046 * As this is called only for pages that do not currently exist, we 2047 * do not need to flush old virtual caches or the TLB. 2048 * 2049 * Context: Process context. May allocate using %GFP_KERNEL. 2050 * Return: vm_fault_t value. 2051 */ 2052 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2053 unsigned long pfn) 2054 { 2055 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 2056 } 2057 EXPORT_SYMBOL(vmf_insert_pfn); 2058 2059 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 2060 { 2061 /* these checks mirror the abort conditions in vm_normal_page */ 2062 if (vma->vm_flags & VM_MIXEDMAP) 2063 return true; 2064 if (pfn_t_devmap(pfn)) 2065 return true; 2066 if (pfn_t_special(pfn)) 2067 return true; 2068 if (is_zero_pfn(pfn_t_to_pfn(pfn))) 2069 return true; 2070 return false; 2071 } 2072 2073 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 2074 unsigned long addr, pfn_t pfn, pgprot_t pgprot, 2075 bool mkwrite) 2076 { 2077 int err; 2078 2079 BUG_ON(!vm_mixed_ok(vma, pfn)); 2080 2081 if (addr < vma->vm_start || addr >= vma->vm_end) 2082 return VM_FAULT_SIGBUS; 2083 2084 track_pfn_insert(vma, &pgprot, pfn); 2085 2086 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 2087 return VM_FAULT_SIGBUS; 2088 2089 /* 2090 * If we don't have pte special, then we have to use the pfn_valid() 2091 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2092 * refcount the page if pfn_valid is true (hence insert_page rather 2093 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 2094 * without pte special, it would there be refcounted as a normal page. 2095 */ 2096 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 2097 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 2098 struct page *page; 2099 2100 /* 2101 * At this point we are committed to insert_page() 2102 * regardless of whether the caller specified flags that 2103 * result in pfn_t_has_page() == false. 2104 */ 2105 page = pfn_to_page(pfn_t_to_pfn(pfn)); 2106 err = insert_page(vma, addr, page, pgprot); 2107 } else { 2108 return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 2109 } 2110 2111 if (err == -ENOMEM) 2112 return VM_FAULT_OOM; 2113 if (err < 0 && err != -EBUSY) 2114 return VM_FAULT_SIGBUS; 2115 2116 return VM_FAULT_NOPAGE; 2117 } 2118 2119 /** 2120 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot 2121 * @vma: user vma to map to 2122 * @addr: target user address of this page 2123 * @pfn: source kernel pfn 2124 * @pgprot: pgprot flags for the inserted page 2125 * 2126 * This is exactly like vmf_insert_mixed(), except that it allows drivers 2127 * to override pgprot on a per-page basis. 2128 * 2129 * Typically this function should be used by drivers to set caching- and 2130 * encryption bits different than those of @vma->vm_page_prot, because 2131 * the caching- or encryption mode may not be known at mmap() time. 2132 * This is ok as long as @vma->vm_page_prot is not used by the core vm 2133 * to set caching and encryption bits for those vmas (except for COW pages). 2134 * This is ensured by core vm only modifying these page table entries using 2135 * functions that don't touch caching- or encryption bits, using pte_modify() 2136 * if needed. (See for example mprotect()). 2137 * Also when new page-table entries are created, this is only done using the 2138 * fault() callback, and never using the value of vma->vm_page_prot, 2139 * except for page-table entries that point to anonymous pages as the result 2140 * of COW. 2141 * 2142 * Context: Process context. May allocate using %GFP_KERNEL. 2143 * Return: vm_fault_t value. 2144 */ 2145 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr, 2146 pfn_t pfn, pgprot_t pgprot) 2147 { 2148 return __vm_insert_mixed(vma, addr, pfn, pgprot, false); 2149 } 2150 EXPORT_SYMBOL(vmf_insert_mixed_prot); 2151 2152 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 2153 pfn_t pfn) 2154 { 2155 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false); 2156 } 2157 EXPORT_SYMBOL(vmf_insert_mixed); 2158 2159 /* 2160 * If the insertion of PTE failed because someone else already added a 2161 * different entry in the mean time, we treat that as success as we assume 2162 * the same entry was actually inserted. 2163 */ 2164 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 2165 unsigned long addr, pfn_t pfn) 2166 { 2167 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true); 2168 } 2169 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 2170 2171 /* 2172 * maps a range of physical memory into the requested pages. the old 2173 * mappings are removed. any references to nonexistent pages results 2174 * in null mappings (currently treated as "copy-on-access") 2175 */ 2176 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 2177 unsigned long addr, unsigned long end, 2178 unsigned long pfn, pgprot_t prot) 2179 { 2180 pte_t *pte; 2181 spinlock_t *ptl; 2182 int err = 0; 2183 2184 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 2185 if (!pte) 2186 return -ENOMEM; 2187 arch_enter_lazy_mmu_mode(); 2188 do { 2189 BUG_ON(!pte_none(*pte)); 2190 if (!pfn_modify_allowed(pfn, prot)) { 2191 err = -EACCES; 2192 break; 2193 } 2194 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 2195 pfn++; 2196 } while (pte++, addr += PAGE_SIZE, addr != end); 2197 arch_leave_lazy_mmu_mode(); 2198 pte_unmap_unlock(pte - 1, ptl); 2199 return err; 2200 } 2201 2202 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 2203 unsigned long addr, unsigned long end, 2204 unsigned long pfn, pgprot_t prot) 2205 { 2206 pmd_t *pmd; 2207 unsigned long next; 2208 int err; 2209 2210 pfn -= addr >> PAGE_SHIFT; 2211 pmd = pmd_alloc(mm, pud, addr); 2212 if (!pmd) 2213 return -ENOMEM; 2214 VM_BUG_ON(pmd_trans_huge(*pmd)); 2215 do { 2216 next = pmd_addr_end(addr, end); 2217 err = remap_pte_range(mm, pmd, addr, next, 2218 pfn + (addr >> PAGE_SHIFT), prot); 2219 if (err) 2220 return err; 2221 } while (pmd++, addr = next, addr != end); 2222 return 0; 2223 } 2224 2225 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 2226 unsigned long addr, unsigned long end, 2227 unsigned long pfn, pgprot_t prot) 2228 { 2229 pud_t *pud; 2230 unsigned long next; 2231 int err; 2232 2233 pfn -= addr >> PAGE_SHIFT; 2234 pud = pud_alloc(mm, p4d, addr); 2235 if (!pud) 2236 return -ENOMEM; 2237 do { 2238 next = pud_addr_end(addr, end); 2239 err = remap_pmd_range(mm, pud, addr, next, 2240 pfn + (addr >> PAGE_SHIFT), prot); 2241 if (err) 2242 return err; 2243 } while (pud++, addr = next, addr != end); 2244 return 0; 2245 } 2246 2247 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2248 unsigned long addr, unsigned long end, 2249 unsigned long pfn, pgprot_t prot) 2250 { 2251 p4d_t *p4d; 2252 unsigned long next; 2253 int err; 2254 2255 pfn -= addr >> PAGE_SHIFT; 2256 p4d = p4d_alloc(mm, pgd, addr); 2257 if (!p4d) 2258 return -ENOMEM; 2259 do { 2260 next = p4d_addr_end(addr, end); 2261 err = remap_pud_range(mm, p4d, addr, next, 2262 pfn + (addr >> PAGE_SHIFT), prot); 2263 if (err) 2264 return err; 2265 } while (p4d++, addr = next, addr != end); 2266 return 0; 2267 } 2268 2269 /** 2270 * remap_pfn_range - remap kernel memory to userspace 2271 * @vma: user vma to map to 2272 * @addr: target page aligned user address to start at 2273 * @pfn: page frame number of kernel physical memory address 2274 * @size: size of mapping area 2275 * @prot: page protection flags for this mapping 2276 * 2277 * Note: this is only safe if the mm semaphore is held when called. 2278 * 2279 * Return: %0 on success, negative error code otherwise. 2280 */ 2281 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 2282 unsigned long pfn, unsigned long size, pgprot_t prot) 2283 { 2284 pgd_t *pgd; 2285 unsigned long next; 2286 unsigned long end = addr + PAGE_ALIGN(size); 2287 struct mm_struct *mm = vma->vm_mm; 2288 unsigned long remap_pfn = pfn; 2289 int err; 2290 2291 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) 2292 return -EINVAL; 2293 2294 /* 2295 * Physically remapped pages are special. Tell the 2296 * rest of the world about it: 2297 * VM_IO tells people not to look at these pages 2298 * (accesses can have side effects). 2299 * VM_PFNMAP tells the core MM that the base pages are just 2300 * raw PFN mappings, and do not have a "struct page" associated 2301 * with them. 2302 * VM_DONTEXPAND 2303 * Disable vma merging and expanding with mremap(). 2304 * VM_DONTDUMP 2305 * Omit vma from core dump, even when VM_IO turned off. 2306 * 2307 * There's a horrible special case to handle copy-on-write 2308 * behaviour that some programs depend on. We mark the "original" 2309 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2310 * See vm_normal_page() for details. 2311 */ 2312 if (is_cow_mapping(vma->vm_flags)) { 2313 if (addr != vma->vm_start || end != vma->vm_end) 2314 return -EINVAL; 2315 vma->vm_pgoff = pfn; 2316 } 2317 2318 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size)); 2319 if (err) 2320 return -EINVAL; 2321 2322 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 2323 2324 BUG_ON(addr >= end); 2325 pfn -= addr >> PAGE_SHIFT; 2326 pgd = pgd_offset(mm, addr); 2327 flush_cache_range(vma, addr, end); 2328 do { 2329 next = pgd_addr_end(addr, end); 2330 err = remap_p4d_range(mm, pgd, addr, next, 2331 pfn + (addr >> PAGE_SHIFT), prot); 2332 if (err) 2333 break; 2334 } while (pgd++, addr = next, addr != end); 2335 2336 if (err) 2337 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size)); 2338 2339 return err; 2340 } 2341 EXPORT_SYMBOL(remap_pfn_range); 2342 2343 /** 2344 * vm_iomap_memory - remap memory to userspace 2345 * @vma: user vma to map to 2346 * @start: start of the physical memory to be mapped 2347 * @len: size of area 2348 * 2349 * This is a simplified io_remap_pfn_range() for common driver use. The 2350 * driver just needs to give us the physical memory range to be mapped, 2351 * we'll figure out the rest from the vma information. 2352 * 2353 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 2354 * whatever write-combining details or similar. 2355 * 2356 * Return: %0 on success, negative error code otherwise. 2357 */ 2358 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 2359 { 2360 unsigned long vm_len, pfn, pages; 2361 2362 /* Check that the physical memory area passed in looks valid */ 2363 if (start + len < start) 2364 return -EINVAL; 2365 /* 2366 * You *really* shouldn't map things that aren't page-aligned, 2367 * but we've historically allowed it because IO memory might 2368 * just have smaller alignment. 2369 */ 2370 len += start & ~PAGE_MASK; 2371 pfn = start >> PAGE_SHIFT; 2372 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 2373 if (pfn + pages < pfn) 2374 return -EINVAL; 2375 2376 /* We start the mapping 'vm_pgoff' pages into the area */ 2377 if (vma->vm_pgoff > pages) 2378 return -EINVAL; 2379 pfn += vma->vm_pgoff; 2380 pages -= vma->vm_pgoff; 2381 2382 /* Can we fit all of the mapping? */ 2383 vm_len = vma->vm_end - vma->vm_start; 2384 if (vm_len >> PAGE_SHIFT > pages) 2385 return -EINVAL; 2386 2387 /* Ok, let it rip */ 2388 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 2389 } 2390 EXPORT_SYMBOL(vm_iomap_memory); 2391 2392 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2393 unsigned long addr, unsigned long end, 2394 pte_fn_t fn, void *data, bool create, 2395 pgtbl_mod_mask *mask) 2396 { 2397 pte_t *pte; 2398 int err = 0; 2399 spinlock_t *ptl; 2400 2401 if (create) { 2402 pte = (mm == &init_mm) ? 2403 pte_alloc_kernel_track(pmd, addr, mask) : 2404 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2405 if (!pte) 2406 return -ENOMEM; 2407 } else { 2408 pte = (mm == &init_mm) ? 2409 pte_offset_kernel(pmd, addr) : 2410 pte_offset_map_lock(mm, pmd, addr, &ptl); 2411 } 2412 2413 BUG_ON(pmd_huge(*pmd)); 2414 2415 arch_enter_lazy_mmu_mode(); 2416 2417 if (fn) { 2418 do { 2419 if (create || !pte_none(*pte)) { 2420 err = fn(pte++, addr, data); 2421 if (err) 2422 break; 2423 } 2424 } while (addr += PAGE_SIZE, addr != end); 2425 } 2426 *mask |= PGTBL_PTE_MODIFIED; 2427 2428 arch_leave_lazy_mmu_mode(); 2429 2430 if (mm != &init_mm) 2431 pte_unmap_unlock(pte-1, ptl); 2432 return err; 2433 } 2434 2435 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2436 unsigned long addr, unsigned long end, 2437 pte_fn_t fn, void *data, bool create, 2438 pgtbl_mod_mask *mask) 2439 { 2440 pmd_t *pmd; 2441 unsigned long next; 2442 int err = 0; 2443 2444 BUG_ON(pud_huge(*pud)); 2445 2446 if (create) { 2447 pmd = pmd_alloc_track(mm, pud, addr, mask); 2448 if (!pmd) 2449 return -ENOMEM; 2450 } else { 2451 pmd = pmd_offset(pud, addr); 2452 } 2453 do { 2454 next = pmd_addr_end(addr, end); 2455 if (create || !pmd_none_or_clear_bad(pmd)) { 2456 err = apply_to_pte_range(mm, pmd, addr, next, fn, data, 2457 create, mask); 2458 if (err) 2459 break; 2460 } 2461 } while (pmd++, addr = next, addr != end); 2462 return err; 2463 } 2464 2465 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2466 unsigned long addr, unsigned long end, 2467 pte_fn_t fn, void *data, bool create, 2468 pgtbl_mod_mask *mask) 2469 { 2470 pud_t *pud; 2471 unsigned long next; 2472 int err = 0; 2473 2474 if (create) { 2475 pud = pud_alloc_track(mm, p4d, addr, mask); 2476 if (!pud) 2477 return -ENOMEM; 2478 } else { 2479 pud = pud_offset(p4d, addr); 2480 } 2481 do { 2482 next = pud_addr_end(addr, end); 2483 if (create || !pud_none_or_clear_bad(pud)) { 2484 err = apply_to_pmd_range(mm, pud, addr, next, fn, data, 2485 create, mask); 2486 if (err) 2487 break; 2488 } 2489 } while (pud++, addr = next, addr != end); 2490 return err; 2491 } 2492 2493 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2494 unsigned long addr, unsigned long end, 2495 pte_fn_t fn, void *data, bool create, 2496 pgtbl_mod_mask *mask) 2497 { 2498 p4d_t *p4d; 2499 unsigned long next; 2500 int err = 0; 2501 2502 if (create) { 2503 p4d = p4d_alloc_track(mm, pgd, addr, mask); 2504 if (!p4d) 2505 return -ENOMEM; 2506 } else { 2507 p4d = p4d_offset(pgd, addr); 2508 } 2509 do { 2510 next = p4d_addr_end(addr, end); 2511 if (create || !p4d_none_or_clear_bad(p4d)) { 2512 err = apply_to_pud_range(mm, p4d, addr, next, fn, data, 2513 create, mask); 2514 if (err) 2515 break; 2516 } 2517 } while (p4d++, addr = next, addr != end); 2518 return err; 2519 } 2520 2521 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2522 unsigned long size, pte_fn_t fn, 2523 void *data, bool create) 2524 { 2525 pgd_t *pgd; 2526 unsigned long start = addr, next; 2527 unsigned long end = addr + size; 2528 pgtbl_mod_mask mask = 0; 2529 int err = 0; 2530 2531 if (WARN_ON(addr >= end)) 2532 return -EINVAL; 2533 2534 pgd = pgd_offset(mm, addr); 2535 do { 2536 next = pgd_addr_end(addr, end); 2537 if (!create && pgd_none_or_clear_bad(pgd)) 2538 continue; 2539 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask); 2540 if (err) 2541 break; 2542 } while (pgd++, addr = next, addr != end); 2543 2544 if (mask & ARCH_PAGE_TABLE_SYNC_MASK) 2545 arch_sync_kernel_mappings(start, start + size); 2546 2547 return err; 2548 } 2549 2550 /* 2551 * Scan a region of virtual memory, filling in page tables as necessary 2552 * and calling a provided function on each leaf page table. 2553 */ 2554 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2555 unsigned long size, pte_fn_t fn, void *data) 2556 { 2557 return __apply_to_page_range(mm, addr, size, fn, data, true); 2558 } 2559 EXPORT_SYMBOL_GPL(apply_to_page_range); 2560 2561 /* 2562 * Scan a region of virtual memory, calling a provided function on 2563 * each leaf page table where it exists. 2564 * 2565 * Unlike apply_to_page_range, this does _not_ fill in page tables 2566 * where they are absent. 2567 */ 2568 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, 2569 unsigned long size, pte_fn_t fn, void *data) 2570 { 2571 return __apply_to_page_range(mm, addr, size, fn, data, false); 2572 } 2573 EXPORT_SYMBOL_GPL(apply_to_existing_page_range); 2574 2575 /* 2576 * handle_pte_fault chooses page fault handler according to an entry which was 2577 * read non-atomically. Before making any commitment, on those architectures 2578 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 2579 * parts, do_swap_page must check under lock before unmapping the pte and 2580 * proceeding (but do_wp_page is only called after already making such a check; 2581 * and do_anonymous_page can safely check later on). 2582 */ 2583 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 2584 pte_t *page_table, pte_t orig_pte) 2585 { 2586 int same = 1; 2587 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) 2588 if (sizeof(pte_t) > sizeof(unsigned long)) { 2589 spinlock_t *ptl = pte_lockptr(mm, pmd); 2590 spin_lock(ptl); 2591 same = pte_same(*page_table, orig_pte); 2592 spin_unlock(ptl); 2593 } 2594 #endif 2595 pte_unmap(page_table); 2596 return same; 2597 } 2598 2599 static inline bool cow_user_page(struct page *dst, struct page *src, 2600 struct vm_fault *vmf) 2601 { 2602 bool ret; 2603 void *kaddr; 2604 void __user *uaddr; 2605 bool locked = false; 2606 struct vm_area_struct *vma = vmf->vma; 2607 struct mm_struct *mm = vma->vm_mm; 2608 unsigned long addr = vmf->address; 2609 2610 if (likely(src)) { 2611 copy_user_highpage(dst, src, addr, vma); 2612 return true; 2613 } 2614 2615 /* 2616 * If the source page was a PFN mapping, we don't have 2617 * a "struct page" for it. We do a best-effort copy by 2618 * just copying from the original user address. If that 2619 * fails, we just zero-fill it. Live with it. 2620 */ 2621 kaddr = kmap_atomic(dst); 2622 uaddr = (void __user *)(addr & PAGE_MASK); 2623 2624 /* 2625 * On architectures with software "accessed" bits, we would 2626 * take a double page fault, so mark it accessed here. 2627 */ 2628 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) { 2629 pte_t entry; 2630 2631 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2632 locked = true; 2633 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2634 /* 2635 * Other thread has already handled the fault 2636 * and update local tlb only 2637 */ 2638 update_mmu_tlb(vma, addr, vmf->pte); 2639 ret = false; 2640 goto pte_unlock; 2641 } 2642 2643 entry = pte_mkyoung(vmf->orig_pte); 2644 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) 2645 update_mmu_cache(vma, addr, vmf->pte); 2646 } 2647 2648 /* 2649 * This really shouldn't fail, because the page is there 2650 * in the page tables. But it might just be unreadable, 2651 * in which case we just give up and fill the result with 2652 * zeroes. 2653 */ 2654 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2655 if (locked) 2656 goto warn; 2657 2658 /* Re-validate under PTL if the page is still mapped */ 2659 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2660 locked = true; 2661 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2662 /* The PTE changed under us, update local tlb */ 2663 update_mmu_tlb(vma, addr, vmf->pte); 2664 ret = false; 2665 goto pte_unlock; 2666 } 2667 2668 /* 2669 * The same page can be mapped back since last copy attempt. 2670 * Try to copy again under PTL. 2671 */ 2672 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2673 /* 2674 * Give a warn in case there can be some obscure 2675 * use-case 2676 */ 2677 warn: 2678 WARN_ON_ONCE(1); 2679 clear_page(kaddr); 2680 } 2681 } 2682 2683 ret = true; 2684 2685 pte_unlock: 2686 if (locked) 2687 pte_unmap_unlock(vmf->pte, vmf->ptl); 2688 kunmap_atomic(kaddr); 2689 flush_dcache_page(dst); 2690 2691 return ret; 2692 } 2693 2694 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 2695 { 2696 struct file *vm_file = vma->vm_file; 2697 2698 if (vm_file) 2699 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 2700 2701 /* 2702 * Special mappings (e.g. VDSO) do not have any file so fake 2703 * a default GFP_KERNEL for them. 2704 */ 2705 return GFP_KERNEL; 2706 } 2707 2708 /* 2709 * Notify the address space that the page is about to become writable so that 2710 * it can prohibit this or wait for the page to get into an appropriate state. 2711 * 2712 * We do this without the lock held, so that it can sleep if it needs to. 2713 */ 2714 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf) 2715 { 2716 vm_fault_t ret; 2717 struct page *page = vmf->page; 2718 unsigned int old_flags = vmf->flags; 2719 2720 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2721 2722 if (vmf->vma->vm_file && 2723 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) 2724 return VM_FAULT_SIGBUS; 2725 2726 ret = vmf->vma->vm_ops->page_mkwrite(vmf); 2727 /* Restore original flags so that caller is not surprised */ 2728 vmf->flags = old_flags; 2729 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2730 return ret; 2731 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2732 lock_page(page); 2733 if (!page->mapping) { 2734 unlock_page(page); 2735 return 0; /* retry */ 2736 } 2737 ret |= VM_FAULT_LOCKED; 2738 } else 2739 VM_BUG_ON_PAGE(!PageLocked(page), page); 2740 return ret; 2741 } 2742 2743 /* 2744 * Handle dirtying of a page in shared file mapping on a write fault. 2745 * 2746 * The function expects the page to be locked and unlocks it. 2747 */ 2748 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) 2749 { 2750 struct vm_area_struct *vma = vmf->vma; 2751 struct address_space *mapping; 2752 struct page *page = vmf->page; 2753 bool dirtied; 2754 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 2755 2756 dirtied = set_page_dirty(page); 2757 VM_BUG_ON_PAGE(PageAnon(page), page); 2758 /* 2759 * Take a local copy of the address_space - page.mapping may be zeroed 2760 * by truncate after unlock_page(). The address_space itself remains 2761 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 2762 * release semantics to prevent the compiler from undoing this copying. 2763 */ 2764 mapping = page_rmapping(page); 2765 unlock_page(page); 2766 2767 if (!page_mkwrite) 2768 file_update_time(vma->vm_file); 2769 2770 /* 2771 * Throttle page dirtying rate down to writeback speed. 2772 * 2773 * mapping may be NULL here because some device drivers do not 2774 * set page.mapping but still dirty their pages 2775 * 2776 * Drop the mmap_lock before waiting on IO, if we can. The file 2777 * is pinning the mapping, as per above. 2778 */ 2779 if ((dirtied || page_mkwrite) && mapping) { 2780 struct file *fpin; 2781 2782 fpin = maybe_unlock_mmap_for_io(vmf, NULL); 2783 balance_dirty_pages_ratelimited(mapping); 2784 if (fpin) { 2785 fput(fpin); 2786 return VM_FAULT_RETRY; 2787 } 2788 } 2789 2790 return 0; 2791 } 2792 2793 /* 2794 * Handle write page faults for pages that can be reused in the current vma 2795 * 2796 * This can happen either due to the mapping being with the VM_SHARED flag, 2797 * or due to us being the last reference standing to the page. In either 2798 * case, all we need to do here is to mark the page as writable and update 2799 * any related book-keeping. 2800 */ 2801 static inline void wp_page_reuse(struct vm_fault *vmf) 2802 __releases(vmf->ptl) 2803 { 2804 struct vm_area_struct *vma = vmf->vma; 2805 struct page *page = vmf->page; 2806 pte_t entry; 2807 /* 2808 * Clear the pages cpupid information as the existing 2809 * information potentially belongs to a now completely 2810 * unrelated process. 2811 */ 2812 if (page) 2813 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2814 2815 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2816 entry = pte_mkyoung(vmf->orig_pte); 2817 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2818 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 2819 update_mmu_cache(vma, vmf->address, vmf->pte); 2820 pte_unmap_unlock(vmf->pte, vmf->ptl); 2821 count_vm_event(PGREUSE); 2822 } 2823 2824 /* 2825 * Handle the case of a page which we actually need to copy to a new page. 2826 * 2827 * Called with mmap_lock locked and the old page referenced, but 2828 * without the ptl held. 2829 * 2830 * High level logic flow: 2831 * 2832 * - Allocate a page, copy the content of the old page to the new one. 2833 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2834 * - Take the PTL. If the pte changed, bail out and release the allocated page 2835 * - If the pte is still the way we remember it, update the page table and all 2836 * relevant references. This includes dropping the reference the page-table 2837 * held to the old page, as well as updating the rmap. 2838 * - In any case, unlock the PTL and drop the reference we took to the old page. 2839 */ 2840 static vm_fault_t wp_page_copy(struct vm_fault *vmf) 2841 { 2842 struct vm_area_struct *vma = vmf->vma; 2843 struct mm_struct *mm = vma->vm_mm; 2844 struct page *old_page = vmf->page; 2845 struct page *new_page = NULL; 2846 pte_t entry; 2847 int page_copied = 0; 2848 struct mmu_notifier_range range; 2849 2850 if (unlikely(anon_vma_prepare(vma))) 2851 goto oom; 2852 2853 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) { 2854 new_page = alloc_zeroed_user_highpage_movable(vma, 2855 vmf->address); 2856 if (!new_page) 2857 goto oom; 2858 } else { 2859 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 2860 vmf->address); 2861 if (!new_page) 2862 goto oom; 2863 2864 if (!cow_user_page(new_page, old_page, vmf)) { 2865 /* 2866 * COW failed, if the fault was solved by other, 2867 * it's fine. If not, userspace would re-fault on 2868 * the same address and we will handle the fault 2869 * from the second attempt. 2870 */ 2871 put_page(new_page); 2872 if (old_page) 2873 put_page(old_page); 2874 return 0; 2875 } 2876 } 2877 2878 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL)) 2879 goto oom_free_new; 2880 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 2881 2882 __SetPageUptodate(new_page); 2883 2884 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 2885 vmf->address & PAGE_MASK, 2886 (vmf->address & PAGE_MASK) + PAGE_SIZE); 2887 mmu_notifier_invalidate_range_start(&range); 2888 2889 /* 2890 * Re-check the pte - we dropped the lock 2891 */ 2892 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 2893 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2894 if (old_page) { 2895 if (!PageAnon(old_page)) { 2896 dec_mm_counter_fast(mm, 2897 mm_counter_file(old_page)); 2898 inc_mm_counter_fast(mm, MM_ANONPAGES); 2899 } 2900 } else { 2901 inc_mm_counter_fast(mm, MM_ANONPAGES); 2902 } 2903 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2904 entry = mk_pte(new_page, vma->vm_page_prot); 2905 entry = pte_sw_mkyoung(entry); 2906 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2907 2908 /* 2909 * Clear the pte entry and flush it first, before updating the 2910 * pte with the new entry, to keep TLBs on different CPUs in 2911 * sync. This code used to set the new PTE then flush TLBs, but 2912 * that left a window where the new PTE could be loaded into 2913 * some TLBs while the old PTE remains in others. 2914 */ 2915 ptep_clear_flush_notify(vma, vmf->address, vmf->pte); 2916 page_add_new_anon_rmap(new_page, vma, vmf->address, false); 2917 lru_cache_add_inactive_or_unevictable(new_page, vma); 2918 /* 2919 * We call the notify macro here because, when using secondary 2920 * mmu page tables (such as kvm shadow page tables), we want the 2921 * new page to be mapped directly into the secondary page table. 2922 */ 2923 set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 2924 update_mmu_cache(vma, vmf->address, vmf->pte); 2925 if (old_page) { 2926 /* 2927 * Only after switching the pte to the new page may 2928 * we remove the mapcount here. Otherwise another 2929 * process may come and find the rmap count decremented 2930 * before the pte is switched to the new page, and 2931 * "reuse" the old page writing into it while our pte 2932 * here still points into it and can be read by other 2933 * threads. 2934 * 2935 * The critical issue is to order this 2936 * page_remove_rmap with the ptp_clear_flush above. 2937 * Those stores are ordered by (if nothing else,) 2938 * the barrier present in the atomic_add_negative 2939 * in page_remove_rmap. 2940 * 2941 * Then the TLB flush in ptep_clear_flush ensures that 2942 * no process can access the old page before the 2943 * decremented mapcount is visible. And the old page 2944 * cannot be reused until after the decremented 2945 * mapcount is visible. So transitively, TLBs to 2946 * old page will be flushed before it can be reused. 2947 */ 2948 page_remove_rmap(old_page, false); 2949 } 2950 2951 /* Free the old page.. */ 2952 new_page = old_page; 2953 page_copied = 1; 2954 } else { 2955 update_mmu_tlb(vma, vmf->address, vmf->pte); 2956 } 2957 2958 if (new_page) 2959 put_page(new_page); 2960 2961 pte_unmap_unlock(vmf->pte, vmf->ptl); 2962 /* 2963 * No need to double call mmu_notifier->invalidate_range() callback as 2964 * the above ptep_clear_flush_notify() did already call it. 2965 */ 2966 mmu_notifier_invalidate_range_only_end(&range); 2967 if (old_page) { 2968 /* 2969 * Don't let another task, with possibly unlocked vma, 2970 * keep the mlocked page. 2971 */ 2972 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 2973 lock_page(old_page); /* LRU manipulation */ 2974 if (PageMlocked(old_page)) 2975 munlock_vma_page(old_page); 2976 unlock_page(old_page); 2977 } 2978 put_page(old_page); 2979 } 2980 return page_copied ? VM_FAULT_WRITE : 0; 2981 oom_free_new: 2982 put_page(new_page); 2983 oom: 2984 if (old_page) 2985 put_page(old_page); 2986 return VM_FAULT_OOM; 2987 } 2988 2989 /** 2990 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 2991 * writeable once the page is prepared 2992 * 2993 * @vmf: structure describing the fault 2994 * 2995 * This function handles all that is needed to finish a write page fault in a 2996 * shared mapping due to PTE being read-only once the mapped page is prepared. 2997 * It handles locking of PTE and modifying it. 2998 * 2999 * The function expects the page to be locked or other protection against 3000 * concurrent faults / writeback (such as DAX radix tree locks). 3001 * 3002 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before 3003 * we acquired PTE lock. 3004 */ 3005 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf) 3006 { 3007 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 3008 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 3009 &vmf->ptl); 3010 /* 3011 * We might have raced with another page fault while we released the 3012 * pte_offset_map_lock. 3013 */ 3014 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 3015 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 3016 pte_unmap_unlock(vmf->pte, vmf->ptl); 3017 return VM_FAULT_NOPAGE; 3018 } 3019 wp_page_reuse(vmf); 3020 return 0; 3021 } 3022 3023 /* 3024 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 3025 * mapping 3026 */ 3027 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 3028 { 3029 struct vm_area_struct *vma = vmf->vma; 3030 3031 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 3032 vm_fault_t ret; 3033 3034 pte_unmap_unlock(vmf->pte, vmf->ptl); 3035 vmf->flags |= FAULT_FLAG_MKWRITE; 3036 ret = vma->vm_ops->pfn_mkwrite(vmf); 3037 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 3038 return ret; 3039 return finish_mkwrite_fault(vmf); 3040 } 3041 wp_page_reuse(vmf); 3042 return VM_FAULT_WRITE; 3043 } 3044 3045 static vm_fault_t wp_page_shared(struct vm_fault *vmf) 3046 __releases(vmf->ptl) 3047 { 3048 struct vm_area_struct *vma = vmf->vma; 3049 vm_fault_t ret = VM_FAULT_WRITE; 3050 3051 get_page(vmf->page); 3052 3053 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 3054 vm_fault_t tmp; 3055 3056 pte_unmap_unlock(vmf->pte, vmf->ptl); 3057 tmp = do_page_mkwrite(vmf); 3058 if (unlikely(!tmp || (tmp & 3059 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3060 put_page(vmf->page); 3061 return tmp; 3062 } 3063 tmp = finish_mkwrite_fault(vmf); 3064 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 3065 unlock_page(vmf->page); 3066 put_page(vmf->page); 3067 return tmp; 3068 } 3069 } else { 3070 wp_page_reuse(vmf); 3071 lock_page(vmf->page); 3072 } 3073 ret |= fault_dirty_shared_page(vmf); 3074 put_page(vmf->page); 3075 3076 return ret; 3077 } 3078 3079 /* 3080 * This routine handles present pages, when users try to write 3081 * to a shared page. It is done by copying the page to a new address 3082 * and decrementing the shared-page counter for the old page. 3083 * 3084 * Note that this routine assumes that the protection checks have been 3085 * done by the caller (the low-level page fault routine in most cases). 3086 * Thus we can safely just mark it writable once we've done any necessary 3087 * COW. 3088 * 3089 * We also mark the page dirty at this point even though the page will 3090 * change only once the write actually happens. This avoids a few races, 3091 * and potentially makes it more efficient. 3092 * 3093 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3094 * but allow concurrent faults), with pte both mapped and locked. 3095 * We return with mmap_lock still held, but pte unmapped and unlocked. 3096 */ 3097 static vm_fault_t do_wp_page(struct vm_fault *vmf) 3098 __releases(vmf->ptl) 3099 { 3100 struct vm_area_struct *vma = vmf->vma; 3101 3102 if (userfaultfd_pte_wp(vma, *vmf->pte)) { 3103 pte_unmap_unlock(vmf->pte, vmf->ptl); 3104 return handle_userfault(vmf, VM_UFFD_WP); 3105 } 3106 3107 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 3108 if (!vmf->page) { 3109 /* 3110 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 3111 * VM_PFNMAP VMA. 3112 * 3113 * We should not cow pages in a shared writeable mapping. 3114 * Just mark the pages writable and/or call ops->pfn_mkwrite. 3115 */ 3116 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3117 (VM_WRITE|VM_SHARED)) 3118 return wp_pfn_shared(vmf); 3119 3120 pte_unmap_unlock(vmf->pte, vmf->ptl); 3121 return wp_page_copy(vmf); 3122 } 3123 3124 /* 3125 * Take out anonymous pages first, anonymous shared vmas are 3126 * not dirty accountable. 3127 */ 3128 if (PageAnon(vmf->page)) { 3129 struct page *page = vmf->page; 3130 3131 /* PageKsm() doesn't necessarily raise the page refcount */ 3132 if (PageKsm(page) || page_count(page) != 1) 3133 goto copy; 3134 if (!trylock_page(page)) 3135 goto copy; 3136 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) { 3137 unlock_page(page); 3138 goto copy; 3139 } 3140 /* 3141 * Ok, we've got the only map reference, and the only 3142 * page count reference, and the page is locked, 3143 * it's dark out, and we're wearing sunglasses. Hit it. 3144 */ 3145 unlock_page(page); 3146 wp_page_reuse(vmf); 3147 return VM_FAULT_WRITE; 3148 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3149 (VM_WRITE|VM_SHARED))) { 3150 return wp_page_shared(vmf); 3151 } 3152 copy: 3153 /* 3154 * Ok, we need to copy. Oh, well.. 3155 */ 3156 get_page(vmf->page); 3157 3158 pte_unmap_unlock(vmf->pte, vmf->ptl); 3159 return wp_page_copy(vmf); 3160 } 3161 3162 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 3163 unsigned long start_addr, unsigned long end_addr, 3164 struct zap_details *details) 3165 { 3166 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 3167 } 3168 3169 static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 3170 struct zap_details *details) 3171 { 3172 struct vm_area_struct *vma; 3173 pgoff_t vba, vea, zba, zea; 3174 3175 vma_interval_tree_foreach(vma, root, 3176 details->first_index, details->last_index) { 3177 3178 vba = vma->vm_pgoff; 3179 vea = vba + vma_pages(vma) - 1; 3180 zba = details->first_index; 3181 if (zba < vba) 3182 zba = vba; 3183 zea = details->last_index; 3184 if (zea > vea) 3185 zea = vea; 3186 3187 unmap_mapping_range_vma(vma, 3188 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 3189 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 3190 details); 3191 } 3192 } 3193 3194 /** 3195 * unmap_mapping_pages() - Unmap pages from processes. 3196 * @mapping: The address space containing pages to be unmapped. 3197 * @start: Index of first page to be unmapped. 3198 * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 3199 * @even_cows: Whether to unmap even private COWed pages. 3200 * 3201 * Unmap the pages in this address space from any userspace process which 3202 * has them mmaped. Generally, you want to remove COWed pages as well when 3203 * a file is being truncated, but not when invalidating pages from the page 3204 * cache. 3205 */ 3206 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 3207 pgoff_t nr, bool even_cows) 3208 { 3209 struct zap_details details = { }; 3210 3211 details.check_mapping = even_cows ? NULL : mapping; 3212 details.first_index = start; 3213 details.last_index = start + nr - 1; 3214 if (details.last_index < details.first_index) 3215 details.last_index = ULONG_MAX; 3216 3217 i_mmap_lock_write(mapping); 3218 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3219 unmap_mapping_range_tree(&mapping->i_mmap, &details); 3220 i_mmap_unlock_write(mapping); 3221 } 3222 3223 /** 3224 * unmap_mapping_range - unmap the portion of all mmaps in the specified 3225 * address_space corresponding to the specified byte range in the underlying 3226 * file. 3227 * 3228 * @mapping: the address space containing mmaps to be unmapped. 3229 * @holebegin: byte in first page to unmap, relative to the start of 3230 * the underlying file. This will be rounded down to a PAGE_SIZE 3231 * boundary. Note that this is different from truncate_pagecache(), which 3232 * must keep the partial page. In contrast, we must get rid of 3233 * partial pages. 3234 * @holelen: size of prospective hole in bytes. This will be rounded 3235 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 3236 * end of the file. 3237 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 3238 * but 0 when invalidating pagecache, don't throw away private data. 3239 */ 3240 void unmap_mapping_range(struct address_space *mapping, 3241 loff_t const holebegin, loff_t const holelen, int even_cows) 3242 { 3243 pgoff_t hba = holebegin >> PAGE_SHIFT; 3244 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3245 3246 /* Check for overflow. */ 3247 if (sizeof(holelen) > sizeof(hlen)) { 3248 long long holeend = 3249 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3250 if (holeend & ~(long long)ULONG_MAX) 3251 hlen = ULONG_MAX - hba + 1; 3252 } 3253 3254 unmap_mapping_pages(mapping, hba, hlen, even_cows); 3255 } 3256 EXPORT_SYMBOL(unmap_mapping_range); 3257 3258 /* 3259 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3260 * but allow concurrent faults), and pte mapped but not yet locked. 3261 * We return with pte unmapped and unlocked. 3262 * 3263 * We return with the mmap_lock locked or unlocked in the same cases 3264 * as does filemap_fault(). 3265 */ 3266 vm_fault_t do_swap_page(struct vm_fault *vmf) 3267 { 3268 struct vm_area_struct *vma = vmf->vma; 3269 struct page *page = NULL, *swapcache; 3270 swp_entry_t entry; 3271 pte_t pte; 3272 int locked; 3273 int exclusive = 0; 3274 vm_fault_t ret = 0; 3275 void *shadow = NULL; 3276 3277 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) 3278 goto out; 3279 3280 entry = pte_to_swp_entry(vmf->orig_pte); 3281 if (unlikely(non_swap_entry(entry))) { 3282 if (is_migration_entry(entry)) { 3283 migration_entry_wait(vma->vm_mm, vmf->pmd, 3284 vmf->address); 3285 } else if (is_device_private_entry(entry)) { 3286 vmf->page = device_private_entry_to_page(entry); 3287 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 3288 } else if (is_hwpoison_entry(entry)) { 3289 ret = VM_FAULT_HWPOISON; 3290 } else { 3291 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 3292 ret = VM_FAULT_SIGBUS; 3293 } 3294 goto out; 3295 } 3296 3297 3298 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 3299 page = lookup_swap_cache(entry, vma, vmf->address); 3300 swapcache = page; 3301 3302 if (!page) { 3303 struct swap_info_struct *si = swp_swap_info(entry); 3304 3305 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && 3306 __swap_count(entry) == 1) { 3307 /* skip swapcache */ 3308 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 3309 vmf->address); 3310 if (page) { 3311 int err; 3312 3313 __SetPageLocked(page); 3314 __SetPageSwapBacked(page); 3315 set_page_private(page, entry.val); 3316 3317 /* Tell memcg to use swap ownership records */ 3318 SetPageSwapCache(page); 3319 err = mem_cgroup_charge(page, vma->vm_mm, 3320 GFP_KERNEL); 3321 ClearPageSwapCache(page); 3322 if (err) { 3323 ret = VM_FAULT_OOM; 3324 goto out_page; 3325 } 3326 3327 shadow = get_shadow_from_swap_cache(entry); 3328 if (shadow) 3329 workingset_refault(page, shadow); 3330 3331 lru_cache_add(page); 3332 swap_readpage(page, true); 3333 } 3334 } else { 3335 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 3336 vmf); 3337 swapcache = page; 3338 } 3339 3340 if (!page) { 3341 /* 3342 * Back out if somebody else faulted in this pte 3343 * while we released the pte lock. 3344 */ 3345 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3346 vmf->address, &vmf->ptl); 3347 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3348 ret = VM_FAULT_OOM; 3349 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3350 goto unlock; 3351 } 3352 3353 /* Had to read the page from swap area: Major fault */ 3354 ret = VM_FAULT_MAJOR; 3355 count_vm_event(PGMAJFAULT); 3356 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 3357 } else if (PageHWPoison(page)) { 3358 /* 3359 * hwpoisoned dirty swapcache pages are kept for killing 3360 * owner processes (which may be unknown at hwpoison time) 3361 */ 3362 ret = VM_FAULT_HWPOISON; 3363 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3364 goto out_release; 3365 } 3366 3367 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags); 3368 3369 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3370 if (!locked) { 3371 ret |= VM_FAULT_RETRY; 3372 goto out_release; 3373 } 3374 3375 /* 3376 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 3377 * release the swapcache from under us. The page pin, and pte_same 3378 * test below, are not enough to exclude that. Even if it is still 3379 * swapcache, we need to check that the page's swap has not changed. 3380 */ 3381 if (unlikely((!PageSwapCache(page) || 3382 page_private(page) != entry.val)) && swapcache) 3383 goto out_page; 3384 3385 page = ksm_might_need_to_copy(page, vma, vmf->address); 3386 if (unlikely(!page)) { 3387 ret = VM_FAULT_OOM; 3388 page = swapcache; 3389 goto out_page; 3390 } 3391 3392 cgroup_throttle_swaprate(page, GFP_KERNEL); 3393 3394 /* 3395 * Back out if somebody else already faulted in this pte. 3396 */ 3397 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3398 &vmf->ptl); 3399 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) 3400 goto out_nomap; 3401 3402 if (unlikely(!PageUptodate(page))) { 3403 ret = VM_FAULT_SIGBUS; 3404 goto out_nomap; 3405 } 3406 3407 /* 3408 * The page isn't present yet, go ahead with the fault. 3409 * 3410 * Be careful about the sequence of operations here. 3411 * To get its accounting right, reuse_swap_page() must be called 3412 * while the page is counted on swap but not yet in mapcount i.e. 3413 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 3414 * must be called after the swap_free(), or it will never succeed. 3415 */ 3416 3417 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3418 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS); 3419 pte = mk_pte(page, vma->vm_page_prot); 3420 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) { 3421 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 3422 vmf->flags &= ~FAULT_FLAG_WRITE; 3423 ret |= VM_FAULT_WRITE; 3424 exclusive = RMAP_EXCLUSIVE; 3425 } 3426 flush_icache_page(vma, page); 3427 if (pte_swp_soft_dirty(vmf->orig_pte)) 3428 pte = pte_mksoft_dirty(pte); 3429 if (pte_swp_uffd_wp(vmf->orig_pte)) { 3430 pte = pte_mkuffd_wp(pte); 3431 pte = pte_wrprotect(pte); 3432 } 3433 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 3434 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 3435 vmf->orig_pte = pte; 3436 3437 /* ksm created a completely new copy */ 3438 if (unlikely(page != swapcache && swapcache)) { 3439 page_add_new_anon_rmap(page, vma, vmf->address, false); 3440 lru_cache_add_inactive_or_unevictable(page, vma); 3441 } else { 3442 do_page_add_anon_rmap(page, vma, vmf->address, exclusive); 3443 } 3444 3445 swap_free(entry); 3446 if (mem_cgroup_swap_full(page) || 3447 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 3448 try_to_free_swap(page); 3449 unlock_page(page); 3450 if (page != swapcache && swapcache) { 3451 /* 3452 * Hold the lock to avoid the swap entry to be reused 3453 * until we take the PT lock for the pte_same() check 3454 * (to avoid false positives from pte_same). For 3455 * further safety release the lock after the swap_free 3456 * so that the swap count won't change under a 3457 * parallel locked swapcache. 3458 */ 3459 unlock_page(swapcache); 3460 put_page(swapcache); 3461 } 3462 3463 if (vmf->flags & FAULT_FLAG_WRITE) { 3464 ret |= do_wp_page(vmf); 3465 if (ret & VM_FAULT_ERROR) 3466 ret &= VM_FAULT_ERROR; 3467 goto out; 3468 } 3469 3470 /* No need to invalidate - it was non-present before */ 3471 update_mmu_cache(vma, vmf->address, vmf->pte); 3472 unlock: 3473 pte_unmap_unlock(vmf->pte, vmf->ptl); 3474 out: 3475 return ret; 3476 out_nomap: 3477 pte_unmap_unlock(vmf->pte, vmf->ptl); 3478 out_page: 3479 unlock_page(page); 3480 out_release: 3481 put_page(page); 3482 if (page != swapcache && swapcache) { 3483 unlock_page(swapcache); 3484 put_page(swapcache); 3485 } 3486 return ret; 3487 } 3488 3489 /* 3490 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3491 * but allow concurrent faults), and pte mapped but not yet locked. 3492 * We return with mmap_lock still held, but pte unmapped and unlocked. 3493 */ 3494 static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 3495 { 3496 struct vm_area_struct *vma = vmf->vma; 3497 struct page *page; 3498 vm_fault_t ret = 0; 3499 pte_t entry; 3500 3501 /* File mapping without ->vm_ops ? */ 3502 if (vma->vm_flags & VM_SHARED) 3503 return VM_FAULT_SIGBUS; 3504 3505 /* 3506 * Use pte_alloc() instead of pte_alloc_map(). We can't run 3507 * pte_offset_map() on pmds where a huge pmd might be created 3508 * from a different thread. 3509 * 3510 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 3511 * parallel threads are excluded by other means. 3512 * 3513 * Here we only have mmap_read_lock(mm). 3514 */ 3515 if (pte_alloc(vma->vm_mm, vmf->pmd)) 3516 return VM_FAULT_OOM; 3517 3518 /* See comment in handle_pte_fault() */ 3519 if (unlikely(pmd_trans_unstable(vmf->pmd))) 3520 return 0; 3521 3522 /* Use the zero-page for reads */ 3523 if (!(vmf->flags & FAULT_FLAG_WRITE) && 3524 !mm_forbids_zeropage(vma->vm_mm)) { 3525 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 3526 vma->vm_page_prot)); 3527 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3528 vmf->address, &vmf->ptl); 3529 if (!pte_none(*vmf->pte)) { 3530 update_mmu_tlb(vma, vmf->address, vmf->pte); 3531 goto unlock; 3532 } 3533 ret = check_stable_address_space(vma->vm_mm); 3534 if (ret) 3535 goto unlock; 3536 /* Deliver the page fault to userland, check inside PT lock */ 3537 if (userfaultfd_missing(vma)) { 3538 pte_unmap_unlock(vmf->pte, vmf->ptl); 3539 return handle_userfault(vmf, VM_UFFD_MISSING); 3540 } 3541 goto setpte; 3542 } 3543 3544 /* Allocate our own private page. */ 3545 if (unlikely(anon_vma_prepare(vma))) 3546 goto oom; 3547 page = alloc_zeroed_user_highpage_movable(vma, vmf->address); 3548 if (!page) 3549 goto oom; 3550 3551 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) 3552 goto oom_free_page; 3553 cgroup_throttle_swaprate(page, GFP_KERNEL); 3554 3555 /* 3556 * The memory barrier inside __SetPageUptodate makes sure that 3557 * preceding stores to the page contents become visible before 3558 * the set_pte_at() write. 3559 */ 3560 __SetPageUptodate(page); 3561 3562 entry = mk_pte(page, vma->vm_page_prot); 3563 entry = pte_sw_mkyoung(entry); 3564 if (vma->vm_flags & VM_WRITE) 3565 entry = pte_mkwrite(pte_mkdirty(entry)); 3566 3567 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3568 &vmf->ptl); 3569 if (!pte_none(*vmf->pte)) { 3570 update_mmu_cache(vma, vmf->address, vmf->pte); 3571 goto release; 3572 } 3573 3574 ret = check_stable_address_space(vma->vm_mm); 3575 if (ret) 3576 goto release; 3577 3578 /* Deliver the page fault to userland, check inside PT lock */ 3579 if (userfaultfd_missing(vma)) { 3580 pte_unmap_unlock(vmf->pte, vmf->ptl); 3581 put_page(page); 3582 return handle_userfault(vmf, VM_UFFD_MISSING); 3583 } 3584 3585 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3586 page_add_new_anon_rmap(page, vma, vmf->address, false); 3587 lru_cache_add_inactive_or_unevictable(page, vma); 3588 setpte: 3589 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3590 3591 /* No need to invalidate - it was non-present before */ 3592 update_mmu_cache(vma, vmf->address, vmf->pte); 3593 unlock: 3594 pte_unmap_unlock(vmf->pte, vmf->ptl); 3595 return ret; 3596 release: 3597 put_page(page); 3598 goto unlock; 3599 oom_free_page: 3600 put_page(page); 3601 oom: 3602 return VM_FAULT_OOM; 3603 } 3604 3605 /* 3606 * The mmap_lock must have been held on entry, and may have been 3607 * released depending on flags and vma->vm_ops->fault() return value. 3608 * See filemap_fault() and __lock_page_retry(). 3609 */ 3610 static vm_fault_t __do_fault(struct vm_fault *vmf) 3611 { 3612 struct vm_area_struct *vma = vmf->vma; 3613 vm_fault_t ret; 3614 3615 /* 3616 * Preallocate pte before we take page_lock because this might lead to 3617 * deadlocks for memcg reclaim which waits for pages under writeback: 3618 * lock_page(A) 3619 * SetPageWriteback(A) 3620 * unlock_page(A) 3621 * lock_page(B) 3622 * lock_page(B) 3623 * pte_alloc_one 3624 * shrink_page_list 3625 * wait_on_page_writeback(A) 3626 * SetPageWriteback(B) 3627 * unlock_page(B) 3628 * # flush A, B to clear the writeback 3629 */ 3630 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 3631 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3632 if (!vmf->prealloc_pte) 3633 return VM_FAULT_OOM; 3634 smp_wmb(); /* See comment in __pte_alloc() */ 3635 } 3636 3637 ret = vma->vm_ops->fault(vmf); 3638 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 3639 VM_FAULT_DONE_COW))) 3640 return ret; 3641 3642 if (unlikely(PageHWPoison(vmf->page))) { 3643 if (ret & VM_FAULT_LOCKED) 3644 unlock_page(vmf->page); 3645 put_page(vmf->page); 3646 vmf->page = NULL; 3647 return VM_FAULT_HWPOISON; 3648 } 3649 3650 if (unlikely(!(ret & VM_FAULT_LOCKED))) 3651 lock_page(vmf->page); 3652 else 3653 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page); 3654 3655 return ret; 3656 } 3657 3658 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3659 static void deposit_prealloc_pte(struct vm_fault *vmf) 3660 { 3661 struct vm_area_struct *vma = vmf->vma; 3662 3663 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3664 /* 3665 * We are going to consume the prealloc table, 3666 * count that as nr_ptes. 3667 */ 3668 mm_inc_nr_ptes(vma->vm_mm); 3669 vmf->prealloc_pte = NULL; 3670 } 3671 3672 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3673 { 3674 struct vm_area_struct *vma = vmf->vma; 3675 bool write = vmf->flags & FAULT_FLAG_WRITE; 3676 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 3677 pmd_t entry; 3678 int i; 3679 vm_fault_t ret = VM_FAULT_FALLBACK; 3680 3681 if (!transhuge_vma_suitable(vma, haddr)) 3682 return ret; 3683 3684 page = compound_head(page); 3685 if (compound_order(page) != HPAGE_PMD_ORDER) 3686 return ret; 3687 3688 /* 3689 * Archs like ppc64 need additonal space to store information 3690 * related to pte entry. Use the preallocated table for that. 3691 */ 3692 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 3693 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3694 if (!vmf->prealloc_pte) 3695 return VM_FAULT_OOM; 3696 smp_wmb(); /* See comment in __pte_alloc() */ 3697 } 3698 3699 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3700 if (unlikely(!pmd_none(*vmf->pmd))) 3701 goto out; 3702 3703 for (i = 0; i < HPAGE_PMD_NR; i++) 3704 flush_icache_page(vma, page + i); 3705 3706 entry = mk_huge_pmd(page, vma->vm_page_prot); 3707 if (write) 3708 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 3709 3710 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR); 3711 page_add_file_rmap(page, true); 3712 /* 3713 * deposit and withdraw with pmd lock held 3714 */ 3715 if (arch_needs_pgtable_deposit()) 3716 deposit_prealloc_pte(vmf); 3717 3718 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 3719 3720 update_mmu_cache_pmd(vma, haddr, vmf->pmd); 3721 3722 /* fault is handled */ 3723 ret = 0; 3724 count_vm_event(THP_FILE_MAPPED); 3725 out: 3726 spin_unlock(vmf->ptl); 3727 return ret; 3728 } 3729 #else 3730 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3731 { 3732 return VM_FAULT_FALLBACK; 3733 } 3734 #endif 3735 3736 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr) 3737 { 3738 struct vm_area_struct *vma = vmf->vma; 3739 bool write = vmf->flags & FAULT_FLAG_WRITE; 3740 bool prefault = vmf->address != addr; 3741 pte_t entry; 3742 3743 flush_icache_page(vma, page); 3744 entry = mk_pte(page, vma->vm_page_prot); 3745 3746 if (prefault && arch_wants_old_prefaulted_pte()) 3747 entry = pte_mkold(entry); 3748 else 3749 entry = pte_sw_mkyoung(entry); 3750 3751 if (write) 3752 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3753 /* copy-on-write page */ 3754 if (write && !(vma->vm_flags & VM_SHARED)) { 3755 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3756 page_add_new_anon_rmap(page, vma, addr, false); 3757 lru_cache_add_inactive_or_unevictable(page, vma); 3758 } else { 3759 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 3760 page_add_file_rmap(page, false); 3761 } 3762 set_pte_at(vma->vm_mm, addr, vmf->pte, entry); 3763 } 3764 3765 /** 3766 * finish_fault - finish page fault once we have prepared the page to fault 3767 * 3768 * @vmf: structure describing the fault 3769 * 3770 * This function handles all that is needed to finish a page fault once the 3771 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 3772 * given page, adds reverse page mapping, handles memcg charges and LRU 3773 * addition. 3774 * 3775 * The function expects the page to be locked and on success it consumes a 3776 * reference of a page being mapped (for the PTE which maps it). 3777 * 3778 * Return: %0 on success, %VM_FAULT_ code in case of error. 3779 */ 3780 vm_fault_t finish_fault(struct vm_fault *vmf) 3781 { 3782 struct vm_area_struct *vma = vmf->vma; 3783 struct page *page; 3784 vm_fault_t ret; 3785 3786 /* Did we COW the page? */ 3787 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 3788 page = vmf->cow_page; 3789 else 3790 page = vmf->page; 3791 3792 /* 3793 * check even for read faults because we might have lost our CoWed 3794 * page 3795 */ 3796 if (!(vma->vm_flags & VM_SHARED)) { 3797 ret = check_stable_address_space(vma->vm_mm); 3798 if (ret) 3799 return ret; 3800 } 3801 3802 if (pmd_none(*vmf->pmd)) { 3803 if (PageTransCompound(page)) { 3804 ret = do_set_pmd(vmf, page); 3805 if (ret != VM_FAULT_FALLBACK) 3806 return ret; 3807 } 3808 3809 if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) 3810 return VM_FAULT_OOM; 3811 } 3812 3813 /* See comment in handle_pte_fault() */ 3814 if (pmd_devmap_trans_unstable(vmf->pmd)) 3815 return 0; 3816 3817 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3818 vmf->address, &vmf->ptl); 3819 ret = 0; 3820 /* Re-check under ptl */ 3821 if (likely(pte_none(*vmf->pte))) 3822 do_set_pte(vmf, page, vmf->address); 3823 else 3824 ret = VM_FAULT_NOPAGE; 3825 3826 update_mmu_tlb(vma, vmf->address, vmf->pte); 3827 pte_unmap_unlock(vmf->pte, vmf->ptl); 3828 return ret; 3829 } 3830 3831 static unsigned long fault_around_bytes __read_mostly = 3832 rounddown_pow_of_two(65536); 3833 3834 #ifdef CONFIG_DEBUG_FS 3835 static int fault_around_bytes_get(void *data, u64 *val) 3836 { 3837 *val = fault_around_bytes; 3838 return 0; 3839 } 3840 3841 /* 3842 * fault_around_bytes must be rounded down to the nearest page order as it's 3843 * what do_fault_around() expects to see. 3844 */ 3845 static int fault_around_bytes_set(void *data, u64 val) 3846 { 3847 if (val / PAGE_SIZE > PTRS_PER_PTE) 3848 return -EINVAL; 3849 if (val > PAGE_SIZE) 3850 fault_around_bytes = rounddown_pow_of_two(val); 3851 else 3852 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 3853 return 0; 3854 } 3855 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 3856 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 3857 3858 static int __init fault_around_debugfs(void) 3859 { 3860 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 3861 &fault_around_bytes_fops); 3862 return 0; 3863 } 3864 late_initcall(fault_around_debugfs); 3865 #endif 3866 3867 /* 3868 * do_fault_around() tries to map few pages around the fault address. The hope 3869 * is that the pages will be needed soon and this will lower the number of 3870 * faults to handle. 3871 * 3872 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 3873 * not ready to be mapped: not up-to-date, locked, etc. 3874 * 3875 * This function is called with the page table lock taken. In the split ptlock 3876 * case the page table lock only protects only those entries which belong to 3877 * the page table corresponding to the fault address. 3878 * 3879 * This function doesn't cross the VMA boundaries, in order to call map_pages() 3880 * only once. 3881 * 3882 * fault_around_bytes defines how many bytes we'll try to map. 3883 * do_fault_around() expects it to be set to a power of two less than or equal 3884 * to PTRS_PER_PTE. 3885 * 3886 * The virtual address of the area that we map is naturally aligned to 3887 * fault_around_bytes rounded down to the machine page size 3888 * (and therefore to page order). This way it's easier to guarantee 3889 * that we don't cross page table boundaries. 3890 */ 3891 static vm_fault_t do_fault_around(struct vm_fault *vmf) 3892 { 3893 unsigned long address = vmf->address, nr_pages, mask; 3894 pgoff_t start_pgoff = vmf->pgoff; 3895 pgoff_t end_pgoff; 3896 int off; 3897 3898 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 3899 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 3900 3901 address = max(address & mask, vmf->vma->vm_start); 3902 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 3903 start_pgoff -= off; 3904 3905 /* 3906 * end_pgoff is either the end of the page table, the end of 3907 * the vma or nr_pages from start_pgoff, depending what is nearest. 3908 */ 3909 end_pgoff = start_pgoff - 3910 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 3911 PTRS_PER_PTE - 1; 3912 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1, 3913 start_pgoff + nr_pages - 1); 3914 3915 if (pmd_none(*vmf->pmd)) { 3916 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 3917 if (!vmf->prealloc_pte) 3918 return VM_FAULT_OOM; 3919 smp_wmb(); /* See comment in __pte_alloc() */ 3920 } 3921 3922 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff); 3923 } 3924 3925 static vm_fault_t do_read_fault(struct vm_fault *vmf) 3926 { 3927 struct vm_area_struct *vma = vmf->vma; 3928 vm_fault_t ret = 0; 3929 3930 /* 3931 * Let's call ->map_pages() first and use ->fault() as fallback 3932 * if page by the offset is not ready to be mapped (cold cache or 3933 * something). 3934 */ 3935 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 3936 ret = do_fault_around(vmf); 3937 if (ret) 3938 return ret; 3939 } 3940 3941 ret = __do_fault(vmf); 3942 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3943 return ret; 3944 3945 ret |= finish_fault(vmf); 3946 unlock_page(vmf->page); 3947 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3948 put_page(vmf->page); 3949 return ret; 3950 } 3951 3952 static vm_fault_t do_cow_fault(struct vm_fault *vmf) 3953 { 3954 struct vm_area_struct *vma = vmf->vma; 3955 vm_fault_t ret; 3956 3957 if (unlikely(anon_vma_prepare(vma))) 3958 return VM_FAULT_OOM; 3959 3960 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address); 3961 if (!vmf->cow_page) 3962 return VM_FAULT_OOM; 3963 3964 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) { 3965 put_page(vmf->cow_page); 3966 return VM_FAULT_OOM; 3967 } 3968 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL); 3969 3970 ret = __do_fault(vmf); 3971 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3972 goto uncharge_out; 3973 if (ret & VM_FAULT_DONE_COW) 3974 return ret; 3975 3976 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 3977 __SetPageUptodate(vmf->cow_page); 3978 3979 ret |= finish_fault(vmf); 3980 unlock_page(vmf->page); 3981 put_page(vmf->page); 3982 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3983 goto uncharge_out; 3984 return ret; 3985 uncharge_out: 3986 put_page(vmf->cow_page); 3987 return ret; 3988 } 3989 3990 static vm_fault_t do_shared_fault(struct vm_fault *vmf) 3991 { 3992 struct vm_area_struct *vma = vmf->vma; 3993 vm_fault_t ret, tmp; 3994 3995 ret = __do_fault(vmf); 3996 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3997 return ret; 3998 3999 /* 4000 * Check if the backing address space wants to know that the page is 4001 * about to become writable 4002 */ 4003 if (vma->vm_ops->page_mkwrite) { 4004 unlock_page(vmf->page); 4005 tmp = do_page_mkwrite(vmf); 4006 if (unlikely(!tmp || 4007 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 4008 put_page(vmf->page); 4009 return tmp; 4010 } 4011 } 4012 4013 ret |= finish_fault(vmf); 4014 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 4015 VM_FAULT_RETRY))) { 4016 unlock_page(vmf->page); 4017 put_page(vmf->page); 4018 return ret; 4019 } 4020 4021 ret |= fault_dirty_shared_page(vmf); 4022 return ret; 4023 } 4024 4025 /* 4026 * We enter with non-exclusive mmap_lock (to exclude vma changes, 4027 * but allow concurrent faults). 4028 * The mmap_lock may have been released depending on flags and our 4029 * return value. See filemap_fault() and __lock_page_or_retry(). 4030 * If mmap_lock is released, vma may become invalid (for example 4031 * by other thread calling munmap()). 4032 */ 4033 static vm_fault_t do_fault(struct vm_fault *vmf) 4034 { 4035 struct vm_area_struct *vma = vmf->vma; 4036 struct mm_struct *vm_mm = vma->vm_mm; 4037 vm_fault_t ret; 4038 4039 /* 4040 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 4041 */ 4042 if (!vma->vm_ops->fault) { 4043 /* 4044 * If we find a migration pmd entry or a none pmd entry, which 4045 * should never happen, return SIGBUS 4046 */ 4047 if (unlikely(!pmd_present(*vmf->pmd))) 4048 ret = VM_FAULT_SIGBUS; 4049 else { 4050 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, 4051 vmf->pmd, 4052 vmf->address, 4053 &vmf->ptl); 4054 /* 4055 * Make sure this is not a temporary clearing of pte 4056 * by holding ptl and checking again. A R/M/W update 4057 * of pte involves: take ptl, clearing the pte so that 4058 * we don't have concurrent modification by hardware 4059 * followed by an update. 4060 */ 4061 if (unlikely(pte_none(*vmf->pte))) 4062 ret = VM_FAULT_SIGBUS; 4063 else 4064 ret = VM_FAULT_NOPAGE; 4065 4066 pte_unmap_unlock(vmf->pte, vmf->ptl); 4067 } 4068 } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 4069 ret = do_read_fault(vmf); 4070 else if (!(vma->vm_flags & VM_SHARED)) 4071 ret = do_cow_fault(vmf); 4072 else 4073 ret = do_shared_fault(vmf); 4074 4075 /* preallocated pagetable is unused: free it */ 4076 if (vmf->prealloc_pte) { 4077 pte_free(vm_mm, vmf->prealloc_pte); 4078 vmf->prealloc_pte = NULL; 4079 } 4080 return ret; 4081 } 4082 4083 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 4084 unsigned long addr, int page_nid, 4085 int *flags) 4086 { 4087 get_page(page); 4088 4089 count_vm_numa_event(NUMA_HINT_FAULTS); 4090 if (page_nid == numa_node_id()) { 4091 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 4092 *flags |= TNF_FAULT_LOCAL; 4093 } 4094 4095 return mpol_misplaced(page, vma, addr); 4096 } 4097 4098 static vm_fault_t do_numa_page(struct vm_fault *vmf) 4099 { 4100 struct vm_area_struct *vma = vmf->vma; 4101 struct page *page = NULL; 4102 int page_nid = NUMA_NO_NODE; 4103 int last_cpupid; 4104 int target_nid; 4105 bool migrated = false; 4106 pte_t pte, old_pte; 4107 bool was_writable = pte_savedwrite(vmf->orig_pte); 4108 int flags = 0; 4109 4110 /* 4111 * The "pte" at this point cannot be used safely without 4112 * validation through pte_unmap_same(). It's of NUMA type but 4113 * the pfn may be screwed if the read is non atomic. 4114 */ 4115 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd); 4116 spin_lock(vmf->ptl); 4117 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4118 pte_unmap_unlock(vmf->pte, vmf->ptl); 4119 goto out; 4120 } 4121 4122 /* 4123 * Make it present again, Depending on how arch implementes non 4124 * accessible ptes, some can allow access by kernel mode. 4125 */ 4126 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 4127 pte = pte_modify(old_pte, vma->vm_page_prot); 4128 pte = pte_mkyoung(pte); 4129 if (was_writable) 4130 pte = pte_mkwrite(pte); 4131 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 4132 update_mmu_cache(vma, vmf->address, vmf->pte); 4133 4134 page = vm_normal_page(vma, vmf->address, pte); 4135 if (!page) { 4136 pte_unmap_unlock(vmf->pte, vmf->ptl); 4137 return 0; 4138 } 4139 4140 /* TODO: handle PTE-mapped THP */ 4141 if (PageCompound(page)) { 4142 pte_unmap_unlock(vmf->pte, vmf->ptl); 4143 return 0; 4144 } 4145 4146 /* 4147 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 4148 * much anyway since they can be in shared cache state. This misses 4149 * the case where a mapping is writable but the process never writes 4150 * to it but pte_write gets cleared during protection updates and 4151 * pte_dirty has unpredictable behaviour between PTE scan updates, 4152 * background writeback, dirty balancing and application behaviour. 4153 */ 4154 if (!pte_write(pte)) 4155 flags |= TNF_NO_GROUP; 4156 4157 /* 4158 * Flag if the page is shared between multiple address spaces. This 4159 * is later used when determining whether to group tasks together 4160 */ 4161 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 4162 flags |= TNF_SHARED; 4163 4164 last_cpupid = page_cpupid_last(page); 4165 page_nid = page_to_nid(page); 4166 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid, 4167 &flags); 4168 pte_unmap_unlock(vmf->pte, vmf->ptl); 4169 if (target_nid == NUMA_NO_NODE) { 4170 put_page(page); 4171 goto out; 4172 } 4173 4174 /* Migrate to the requested node */ 4175 migrated = migrate_misplaced_page(page, vma, target_nid); 4176 if (migrated) { 4177 page_nid = target_nid; 4178 flags |= TNF_MIGRATED; 4179 } else 4180 flags |= TNF_MIGRATE_FAIL; 4181 4182 out: 4183 if (page_nid != NUMA_NO_NODE) 4184 task_numa_fault(last_cpupid, page_nid, 1, flags); 4185 return 0; 4186 } 4187 4188 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 4189 { 4190 if (vma_is_anonymous(vmf->vma)) 4191 return do_huge_pmd_anonymous_page(vmf); 4192 if (vmf->vma->vm_ops->huge_fault) 4193 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4194 return VM_FAULT_FALLBACK; 4195 } 4196 4197 /* `inline' is required to avoid gcc 4.1.2 build error */ 4198 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd) 4199 { 4200 if (vma_is_anonymous(vmf->vma)) { 4201 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd)) 4202 return handle_userfault(vmf, VM_UFFD_WP); 4203 return do_huge_pmd_wp_page(vmf, orig_pmd); 4204 } 4205 if (vmf->vma->vm_ops->huge_fault) { 4206 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4207 4208 if (!(ret & VM_FAULT_FALLBACK)) 4209 return ret; 4210 } 4211 4212 /* COW or write-notify handled on pte level: split pmd. */ 4213 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL); 4214 4215 return VM_FAULT_FALLBACK; 4216 } 4217 4218 static vm_fault_t create_huge_pud(struct vm_fault *vmf) 4219 { 4220 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 4221 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 4222 /* No support for anonymous transparent PUD pages yet */ 4223 if (vma_is_anonymous(vmf->vma)) 4224 goto split; 4225 if (vmf->vma->vm_ops->huge_fault) { 4226 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4227 4228 if (!(ret & VM_FAULT_FALLBACK)) 4229 return ret; 4230 } 4231 split: 4232 /* COW or write-notify not handled on PUD level: split pud.*/ 4233 __split_huge_pud(vmf->vma, vmf->pud, vmf->address); 4234 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4235 return VM_FAULT_FALLBACK; 4236 } 4237 4238 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 4239 { 4240 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4241 /* No support for anonymous transparent PUD pages yet */ 4242 if (vma_is_anonymous(vmf->vma)) 4243 return VM_FAULT_FALLBACK; 4244 if (vmf->vma->vm_ops->huge_fault) 4245 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4246 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4247 return VM_FAULT_FALLBACK; 4248 } 4249 4250 /* 4251 * These routines also need to handle stuff like marking pages dirty 4252 * and/or accessed for architectures that don't do it in hardware (most 4253 * RISC architectures). The early dirtying is also good on the i386. 4254 * 4255 * There is also a hook called "update_mmu_cache()" that architectures 4256 * with external mmu caches can use to update those (ie the Sparc or 4257 * PowerPC hashed page tables that act as extended TLBs). 4258 * 4259 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow 4260 * concurrent faults). 4261 * 4262 * The mmap_lock may have been released depending on flags and our return value. 4263 * See filemap_fault() and __lock_page_or_retry(). 4264 */ 4265 static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 4266 { 4267 pte_t entry; 4268 4269 if (unlikely(pmd_none(*vmf->pmd))) { 4270 /* 4271 * Leave __pte_alloc() until later: because vm_ops->fault may 4272 * want to allocate huge page, and if we expose page table 4273 * for an instant, it will be difficult to retract from 4274 * concurrent faults and from rmap lookups. 4275 */ 4276 vmf->pte = NULL; 4277 } else { 4278 /* 4279 * If a huge pmd materialized under us just retry later. Use 4280 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead 4281 * of pmd_trans_huge() to ensure the pmd didn't become 4282 * pmd_trans_huge under us and then back to pmd_none, as a 4283 * result of MADV_DONTNEED running immediately after a huge pmd 4284 * fault in a different thread of this mm, in turn leading to a 4285 * misleading pmd_trans_huge() retval. All we have to ensure is 4286 * that it is a regular pmd that we can walk with 4287 * pte_offset_map() and we can do that through an atomic read 4288 * in C, which is what pmd_trans_unstable() provides. 4289 */ 4290 if (pmd_devmap_trans_unstable(vmf->pmd)) 4291 return 0; 4292 /* 4293 * A regular pmd is established and it can't morph into a huge 4294 * pmd from under us anymore at this point because we hold the 4295 * mmap_lock read mode and khugepaged takes it in write mode. 4296 * So now it's safe to run pte_offset_map(). 4297 */ 4298 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4299 vmf->orig_pte = *vmf->pte; 4300 4301 /* 4302 * some architectures can have larger ptes than wordsize, 4303 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and 4304 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic 4305 * accesses. The code below just needs a consistent view 4306 * for the ifs and we later double check anyway with the 4307 * ptl lock held. So here a barrier will do. 4308 */ 4309 barrier(); 4310 if (pte_none(vmf->orig_pte)) { 4311 pte_unmap(vmf->pte); 4312 vmf->pte = NULL; 4313 } 4314 } 4315 4316 if (!vmf->pte) { 4317 if (vma_is_anonymous(vmf->vma)) 4318 return do_anonymous_page(vmf); 4319 else 4320 return do_fault(vmf); 4321 } 4322 4323 if (!pte_present(vmf->orig_pte)) 4324 return do_swap_page(vmf); 4325 4326 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 4327 return do_numa_page(vmf); 4328 4329 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 4330 spin_lock(vmf->ptl); 4331 entry = vmf->orig_pte; 4332 if (unlikely(!pte_same(*vmf->pte, entry))) { 4333 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 4334 goto unlock; 4335 } 4336 if (vmf->flags & FAULT_FLAG_WRITE) { 4337 if (!pte_write(entry)) 4338 return do_wp_page(vmf); 4339 entry = pte_mkdirty(entry); 4340 } 4341 entry = pte_mkyoung(entry); 4342 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 4343 vmf->flags & FAULT_FLAG_WRITE)) { 4344 update_mmu_cache(vmf->vma, vmf->address, vmf->pte); 4345 } else { 4346 /* Skip spurious TLB flush for retried page fault */ 4347 if (vmf->flags & FAULT_FLAG_TRIED) 4348 goto unlock; 4349 /* 4350 * This is needed only for protection faults but the arch code 4351 * is not yet telling us if this is a protection fault or not. 4352 * This still avoids useless tlb flushes for .text page faults 4353 * with threads. 4354 */ 4355 if (vmf->flags & FAULT_FLAG_WRITE) 4356 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address); 4357 } 4358 unlock: 4359 pte_unmap_unlock(vmf->pte, vmf->ptl); 4360 return 0; 4361 } 4362 4363 /* 4364 * By the time we get here, we already hold the mm semaphore 4365 * 4366 * The mmap_lock may have been released depending on flags and our 4367 * return value. See filemap_fault() and __lock_page_or_retry(). 4368 */ 4369 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 4370 unsigned long address, unsigned int flags) 4371 { 4372 struct vm_fault vmf = { 4373 .vma = vma, 4374 .address = address & PAGE_MASK, 4375 .flags = flags, 4376 .pgoff = linear_page_index(vma, address), 4377 .gfp_mask = __get_fault_gfp_mask(vma), 4378 }; 4379 unsigned int dirty = flags & FAULT_FLAG_WRITE; 4380 struct mm_struct *mm = vma->vm_mm; 4381 pgd_t *pgd; 4382 p4d_t *p4d; 4383 vm_fault_t ret; 4384 4385 pgd = pgd_offset(mm, address); 4386 p4d = p4d_alloc(mm, pgd, address); 4387 if (!p4d) 4388 return VM_FAULT_OOM; 4389 4390 vmf.pud = pud_alloc(mm, p4d, address); 4391 if (!vmf.pud) 4392 return VM_FAULT_OOM; 4393 retry_pud: 4394 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) { 4395 ret = create_huge_pud(&vmf); 4396 if (!(ret & VM_FAULT_FALLBACK)) 4397 return ret; 4398 } else { 4399 pud_t orig_pud = *vmf.pud; 4400 4401 barrier(); 4402 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 4403 4404 /* NUMA case for anonymous PUDs would go here */ 4405 4406 if (dirty && !pud_write(orig_pud)) { 4407 ret = wp_huge_pud(&vmf, orig_pud); 4408 if (!(ret & VM_FAULT_FALLBACK)) 4409 return ret; 4410 } else { 4411 huge_pud_set_accessed(&vmf, orig_pud); 4412 return 0; 4413 } 4414 } 4415 } 4416 4417 vmf.pmd = pmd_alloc(mm, vmf.pud, address); 4418 if (!vmf.pmd) 4419 return VM_FAULT_OOM; 4420 4421 /* Huge pud page fault raced with pmd_alloc? */ 4422 if (pud_trans_unstable(vmf.pud)) 4423 goto retry_pud; 4424 4425 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) { 4426 ret = create_huge_pmd(&vmf); 4427 if (!(ret & VM_FAULT_FALLBACK)) 4428 return ret; 4429 } else { 4430 pmd_t orig_pmd = *vmf.pmd; 4431 4432 barrier(); 4433 if (unlikely(is_swap_pmd(orig_pmd))) { 4434 VM_BUG_ON(thp_migration_supported() && 4435 !is_pmd_migration_entry(orig_pmd)); 4436 if (is_pmd_migration_entry(orig_pmd)) 4437 pmd_migration_entry_wait(mm, vmf.pmd); 4438 return 0; 4439 } 4440 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) { 4441 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma)) 4442 return do_huge_pmd_numa_page(&vmf, orig_pmd); 4443 4444 if (dirty && !pmd_write(orig_pmd)) { 4445 ret = wp_huge_pmd(&vmf, orig_pmd); 4446 if (!(ret & VM_FAULT_FALLBACK)) 4447 return ret; 4448 } else { 4449 huge_pmd_set_accessed(&vmf, orig_pmd); 4450 return 0; 4451 } 4452 } 4453 } 4454 4455 return handle_pte_fault(&vmf); 4456 } 4457 4458 /** 4459 * mm_account_fault - Do page fault accountings 4460 * 4461 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting 4462 * of perf event counters, but we'll still do the per-task accounting to 4463 * the task who triggered this page fault. 4464 * @address: the faulted address. 4465 * @flags: the fault flags. 4466 * @ret: the fault retcode. 4467 * 4468 * This will take care of most of the page fault accountings. Meanwhile, it 4469 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter 4470 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should 4471 * still be in per-arch page fault handlers at the entry of page fault. 4472 */ 4473 static inline void mm_account_fault(struct pt_regs *regs, 4474 unsigned long address, unsigned int flags, 4475 vm_fault_t ret) 4476 { 4477 bool major; 4478 4479 /* 4480 * We don't do accounting for some specific faults: 4481 * 4482 * - Unsuccessful faults (e.g. when the address wasn't valid). That 4483 * includes arch_vma_access_permitted() failing before reaching here. 4484 * So this is not a "this many hardware page faults" counter. We 4485 * should use the hw profiling for that. 4486 * 4487 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted 4488 * once they're completed. 4489 */ 4490 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY)) 4491 return; 4492 4493 /* 4494 * We define the fault as a major fault when the final successful fault 4495 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't 4496 * handle it immediately previously). 4497 */ 4498 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); 4499 4500 if (major) 4501 current->maj_flt++; 4502 else 4503 current->min_flt++; 4504 4505 /* 4506 * If the fault is done for GUP, regs will be NULL. We only do the 4507 * accounting for the per thread fault counters who triggered the 4508 * fault, and we skip the perf event updates. 4509 */ 4510 if (!regs) 4511 return; 4512 4513 if (major) 4514 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 4515 else 4516 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 4517 } 4518 4519 /* 4520 * By the time we get here, we already hold the mm semaphore 4521 * 4522 * The mmap_lock may have been released depending on flags and our 4523 * return value. See filemap_fault() and __lock_page_or_retry(). 4524 */ 4525 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 4526 unsigned int flags, struct pt_regs *regs) 4527 { 4528 vm_fault_t ret; 4529 4530 __set_current_state(TASK_RUNNING); 4531 4532 count_vm_event(PGFAULT); 4533 count_memcg_event_mm(vma->vm_mm, PGFAULT); 4534 4535 /* do counter updates before entering really critical section. */ 4536 check_sync_rss_stat(current); 4537 4538 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 4539 flags & FAULT_FLAG_INSTRUCTION, 4540 flags & FAULT_FLAG_REMOTE)) 4541 return VM_FAULT_SIGSEGV; 4542 4543 /* 4544 * Enable the memcg OOM handling for faults triggered in user 4545 * space. Kernel faults are handled more gracefully. 4546 */ 4547 if (flags & FAULT_FLAG_USER) 4548 mem_cgroup_enter_user_fault(); 4549 4550 if (unlikely(is_vm_hugetlb_page(vma))) 4551 ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 4552 else 4553 ret = __handle_mm_fault(vma, address, flags); 4554 4555 if (flags & FAULT_FLAG_USER) { 4556 mem_cgroup_exit_user_fault(); 4557 /* 4558 * The task may have entered a memcg OOM situation but 4559 * if the allocation error was handled gracefully (no 4560 * VM_FAULT_OOM), there is no need to kill anything. 4561 * Just clean up the OOM state peacefully. 4562 */ 4563 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 4564 mem_cgroup_oom_synchronize(false); 4565 } 4566 4567 mm_account_fault(regs, address, flags, ret); 4568 4569 return ret; 4570 } 4571 EXPORT_SYMBOL_GPL(handle_mm_fault); 4572 4573 #ifndef __PAGETABLE_P4D_FOLDED 4574 /* 4575 * Allocate p4d page table. 4576 * We've already handled the fast-path in-line. 4577 */ 4578 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 4579 { 4580 p4d_t *new = p4d_alloc_one(mm, address); 4581 if (!new) 4582 return -ENOMEM; 4583 4584 smp_wmb(); /* See comment in __pte_alloc */ 4585 4586 spin_lock(&mm->page_table_lock); 4587 if (pgd_present(*pgd)) /* Another has populated it */ 4588 p4d_free(mm, new); 4589 else 4590 pgd_populate(mm, pgd, new); 4591 spin_unlock(&mm->page_table_lock); 4592 return 0; 4593 } 4594 #endif /* __PAGETABLE_P4D_FOLDED */ 4595 4596 #ifndef __PAGETABLE_PUD_FOLDED 4597 /* 4598 * Allocate page upper directory. 4599 * We've already handled the fast-path in-line. 4600 */ 4601 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 4602 { 4603 pud_t *new = pud_alloc_one(mm, address); 4604 if (!new) 4605 return -ENOMEM; 4606 4607 smp_wmb(); /* See comment in __pte_alloc */ 4608 4609 spin_lock(&mm->page_table_lock); 4610 if (!p4d_present(*p4d)) { 4611 mm_inc_nr_puds(mm); 4612 p4d_populate(mm, p4d, new); 4613 } else /* Another has populated it */ 4614 pud_free(mm, new); 4615 spin_unlock(&mm->page_table_lock); 4616 return 0; 4617 } 4618 #endif /* __PAGETABLE_PUD_FOLDED */ 4619 4620 #ifndef __PAGETABLE_PMD_FOLDED 4621 /* 4622 * Allocate page middle directory. 4623 * We've already handled the fast-path in-line. 4624 */ 4625 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 4626 { 4627 spinlock_t *ptl; 4628 pmd_t *new = pmd_alloc_one(mm, address); 4629 if (!new) 4630 return -ENOMEM; 4631 4632 smp_wmb(); /* See comment in __pte_alloc */ 4633 4634 ptl = pud_lock(mm, pud); 4635 if (!pud_present(*pud)) { 4636 mm_inc_nr_pmds(mm); 4637 pud_populate(mm, pud, new); 4638 } else /* Another has populated it */ 4639 pmd_free(mm, new); 4640 spin_unlock(ptl); 4641 return 0; 4642 } 4643 #endif /* __PAGETABLE_PMD_FOLDED */ 4644 4645 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address, 4646 struct mmu_notifier_range *range, pte_t **ptepp, 4647 pmd_t **pmdpp, spinlock_t **ptlp) 4648 { 4649 pgd_t *pgd; 4650 p4d_t *p4d; 4651 pud_t *pud; 4652 pmd_t *pmd; 4653 pte_t *ptep; 4654 4655 pgd = pgd_offset(mm, address); 4656 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 4657 goto out; 4658 4659 p4d = p4d_offset(pgd, address); 4660 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 4661 goto out; 4662 4663 pud = pud_offset(p4d, address); 4664 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 4665 goto out; 4666 4667 pmd = pmd_offset(pud, address); 4668 VM_BUG_ON(pmd_trans_huge(*pmd)); 4669 4670 if (pmd_huge(*pmd)) { 4671 if (!pmdpp) 4672 goto out; 4673 4674 if (range) { 4675 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, 4676 NULL, mm, address & PMD_MASK, 4677 (address & PMD_MASK) + PMD_SIZE); 4678 mmu_notifier_invalidate_range_start(range); 4679 } 4680 *ptlp = pmd_lock(mm, pmd); 4681 if (pmd_huge(*pmd)) { 4682 *pmdpp = pmd; 4683 return 0; 4684 } 4685 spin_unlock(*ptlp); 4686 if (range) 4687 mmu_notifier_invalidate_range_end(range); 4688 } 4689 4690 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 4691 goto out; 4692 4693 if (range) { 4694 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 4695 address & PAGE_MASK, 4696 (address & PAGE_MASK) + PAGE_SIZE); 4697 mmu_notifier_invalidate_range_start(range); 4698 } 4699 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 4700 if (!pte_present(*ptep)) 4701 goto unlock; 4702 *ptepp = ptep; 4703 return 0; 4704 unlock: 4705 pte_unmap_unlock(ptep, *ptlp); 4706 if (range) 4707 mmu_notifier_invalidate_range_end(range); 4708 out: 4709 return -EINVAL; 4710 } 4711 4712 /** 4713 * follow_pte - look up PTE at a user virtual address 4714 * @mm: the mm_struct of the target address space 4715 * @address: user virtual address 4716 * @ptepp: location to store found PTE 4717 * @ptlp: location to store the lock for the PTE 4718 * 4719 * On a successful return, the pointer to the PTE is stored in @ptepp; 4720 * the corresponding lock is taken and its location is stored in @ptlp. 4721 * The contents of the PTE are only stable until @ptlp is released; 4722 * any further use, if any, must be protected against invalidation 4723 * with MMU notifiers. 4724 * 4725 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore 4726 * should be taken for read. 4727 * 4728 * KVM uses this function. While it is arguably less bad than ``follow_pfn``, 4729 * it is not a good general-purpose API. 4730 * 4731 * Return: zero on success, -ve otherwise. 4732 */ 4733 int follow_pte(struct mm_struct *mm, unsigned long address, 4734 pte_t **ptepp, spinlock_t **ptlp) 4735 { 4736 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp); 4737 } 4738 EXPORT_SYMBOL_GPL(follow_pte); 4739 4740 /** 4741 * follow_pfn - look up PFN at a user virtual address 4742 * @vma: memory mapping 4743 * @address: user virtual address 4744 * @pfn: location to store found PFN 4745 * 4746 * Only IO mappings and raw PFN mappings are allowed. 4747 * 4748 * This function does not allow the caller to read the permissions 4749 * of the PTE. Do not use it. 4750 * 4751 * Return: zero and the pfn at @pfn on success, -ve otherwise. 4752 */ 4753 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 4754 unsigned long *pfn) 4755 { 4756 int ret = -EINVAL; 4757 spinlock_t *ptl; 4758 pte_t *ptep; 4759 4760 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4761 return ret; 4762 4763 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 4764 if (ret) 4765 return ret; 4766 *pfn = pte_pfn(*ptep); 4767 pte_unmap_unlock(ptep, ptl); 4768 return 0; 4769 } 4770 EXPORT_SYMBOL(follow_pfn); 4771 4772 #ifdef CONFIG_HAVE_IOREMAP_PROT 4773 int follow_phys(struct vm_area_struct *vma, 4774 unsigned long address, unsigned int flags, 4775 unsigned long *prot, resource_size_t *phys) 4776 { 4777 int ret = -EINVAL; 4778 pte_t *ptep, pte; 4779 spinlock_t *ptl; 4780 4781 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4782 goto out; 4783 4784 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 4785 goto out; 4786 pte = *ptep; 4787 4788 if ((flags & FOLL_WRITE) && !pte_write(pte)) 4789 goto unlock; 4790 4791 *prot = pgprot_val(pte_pgprot(pte)); 4792 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 4793 4794 ret = 0; 4795 unlock: 4796 pte_unmap_unlock(ptep, ptl); 4797 out: 4798 return ret; 4799 } 4800 4801 /** 4802 * generic_access_phys - generic implementation for iomem mmap access 4803 * @vma: the vma to access 4804 * @addr: userspace addres, not relative offset within @vma 4805 * @buf: buffer to read/write 4806 * @len: length of transfer 4807 * @write: set to FOLL_WRITE when writing, otherwise reading 4808 * 4809 * This is a generic implementation for &vm_operations_struct.access for an 4810 * iomem mapping. This callback is used by access_process_vm() when the @vma is 4811 * not page based. 4812 */ 4813 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 4814 void *buf, int len, int write) 4815 { 4816 resource_size_t phys_addr; 4817 unsigned long prot = 0; 4818 void __iomem *maddr; 4819 pte_t *ptep, pte; 4820 spinlock_t *ptl; 4821 int offset = offset_in_page(addr); 4822 int ret = -EINVAL; 4823 4824 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4825 return -EINVAL; 4826 4827 retry: 4828 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 4829 return -EINVAL; 4830 pte = *ptep; 4831 pte_unmap_unlock(ptep, ptl); 4832 4833 prot = pgprot_val(pte_pgprot(pte)); 4834 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 4835 4836 if ((write & FOLL_WRITE) && !pte_write(pte)) 4837 return -EINVAL; 4838 4839 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 4840 if (!maddr) 4841 return -ENOMEM; 4842 4843 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 4844 goto out_unmap; 4845 4846 if (!pte_same(pte, *ptep)) { 4847 pte_unmap_unlock(ptep, ptl); 4848 iounmap(maddr); 4849 4850 goto retry; 4851 } 4852 4853 if (write) 4854 memcpy_toio(maddr + offset, buf, len); 4855 else 4856 memcpy_fromio(buf, maddr + offset, len); 4857 ret = len; 4858 pte_unmap_unlock(ptep, ptl); 4859 out_unmap: 4860 iounmap(maddr); 4861 4862 return ret; 4863 } 4864 EXPORT_SYMBOL_GPL(generic_access_phys); 4865 #endif 4866 4867 /* 4868 * Access another process' address space as given in mm. 4869 */ 4870 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf, 4871 int len, unsigned int gup_flags) 4872 { 4873 struct vm_area_struct *vma; 4874 void *old_buf = buf; 4875 int write = gup_flags & FOLL_WRITE; 4876 4877 if (mmap_read_lock_killable(mm)) 4878 return 0; 4879 4880 /* ignore errors, just check how much was successfully transferred */ 4881 while (len) { 4882 int bytes, ret, offset; 4883 void *maddr; 4884 struct page *page = NULL; 4885 4886 ret = get_user_pages_remote(mm, addr, 1, 4887 gup_flags, &page, &vma, NULL); 4888 if (ret <= 0) { 4889 #ifndef CONFIG_HAVE_IOREMAP_PROT 4890 break; 4891 #else 4892 /* 4893 * Check if this is a VM_IO | VM_PFNMAP VMA, which 4894 * we can access using slightly different code. 4895 */ 4896 vma = find_vma(mm, addr); 4897 if (!vma || vma->vm_start > addr) 4898 break; 4899 if (vma->vm_ops && vma->vm_ops->access) 4900 ret = vma->vm_ops->access(vma, addr, buf, 4901 len, write); 4902 if (ret <= 0) 4903 break; 4904 bytes = ret; 4905 #endif 4906 } else { 4907 bytes = len; 4908 offset = addr & (PAGE_SIZE-1); 4909 if (bytes > PAGE_SIZE-offset) 4910 bytes = PAGE_SIZE-offset; 4911 4912 maddr = kmap(page); 4913 if (write) { 4914 copy_to_user_page(vma, page, addr, 4915 maddr + offset, buf, bytes); 4916 set_page_dirty_lock(page); 4917 } else { 4918 copy_from_user_page(vma, page, addr, 4919 buf, maddr + offset, bytes); 4920 } 4921 kunmap(page); 4922 put_page(page); 4923 } 4924 len -= bytes; 4925 buf += bytes; 4926 addr += bytes; 4927 } 4928 mmap_read_unlock(mm); 4929 4930 return buf - old_buf; 4931 } 4932 4933 /** 4934 * access_remote_vm - access another process' address space 4935 * @mm: the mm_struct of the target address space 4936 * @addr: start address to access 4937 * @buf: source or destination buffer 4938 * @len: number of bytes to transfer 4939 * @gup_flags: flags modifying lookup behaviour 4940 * 4941 * The caller must hold a reference on @mm. 4942 * 4943 * Return: number of bytes copied from source to destination. 4944 */ 4945 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 4946 void *buf, int len, unsigned int gup_flags) 4947 { 4948 return __access_remote_vm(mm, addr, buf, len, gup_flags); 4949 } 4950 4951 /* 4952 * Access another process' address space. 4953 * Source/target buffer must be kernel space, 4954 * Do not walk the page table directly, use get_user_pages 4955 */ 4956 int access_process_vm(struct task_struct *tsk, unsigned long addr, 4957 void *buf, int len, unsigned int gup_flags) 4958 { 4959 struct mm_struct *mm; 4960 int ret; 4961 4962 mm = get_task_mm(tsk); 4963 if (!mm) 4964 return 0; 4965 4966 ret = __access_remote_vm(mm, addr, buf, len, gup_flags); 4967 4968 mmput(mm); 4969 4970 return ret; 4971 } 4972 EXPORT_SYMBOL_GPL(access_process_vm); 4973 4974 /* 4975 * Print the name of a VMA. 4976 */ 4977 void print_vma_addr(char *prefix, unsigned long ip) 4978 { 4979 struct mm_struct *mm = current->mm; 4980 struct vm_area_struct *vma; 4981 4982 /* 4983 * we might be running from an atomic context so we cannot sleep 4984 */ 4985 if (!mmap_read_trylock(mm)) 4986 return; 4987 4988 vma = find_vma(mm, ip); 4989 if (vma && vma->vm_file) { 4990 struct file *f = vma->vm_file; 4991 char *buf = (char *)__get_free_page(GFP_NOWAIT); 4992 if (buf) { 4993 char *p; 4994 4995 p = file_path(f, buf, PAGE_SIZE); 4996 if (IS_ERR(p)) 4997 p = "?"; 4998 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 4999 vma->vm_start, 5000 vma->vm_end - vma->vm_start); 5001 free_page((unsigned long)buf); 5002 } 5003 } 5004 mmap_read_unlock(mm); 5005 } 5006 5007 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5008 void __might_fault(const char *file, int line) 5009 { 5010 /* 5011 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 5012 * holding the mmap_lock, this is safe because kernel memory doesn't 5013 * get paged out, therefore we'll never actually fault, and the 5014 * below annotations will generate false positives. 5015 */ 5016 if (uaccess_kernel()) 5017 return; 5018 if (pagefault_disabled()) 5019 return; 5020 __might_sleep(file, line, 0); 5021 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5022 if (current->mm) 5023 might_lock_read(¤t->mm->mmap_lock); 5024 #endif 5025 } 5026 EXPORT_SYMBOL(__might_fault); 5027 #endif 5028 5029 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 5030 /* 5031 * Process all subpages of the specified huge page with the specified 5032 * operation. The target subpage will be processed last to keep its 5033 * cache lines hot. 5034 */ 5035 static inline void process_huge_page( 5036 unsigned long addr_hint, unsigned int pages_per_huge_page, 5037 void (*process_subpage)(unsigned long addr, int idx, void *arg), 5038 void *arg) 5039 { 5040 int i, n, base, l; 5041 unsigned long addr = addr_hint & 5042 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5043 5044 /* Process target subpage last to keep its cache lines hot */ 5045 might_sleep(); 5046 n = (addr_hint - addr) / PAGE_SIZE; 5047 if (2 * n <= pages_per_huge_page) { 5048 /* If target subpage in first half of huge page */ 5049 base = 0; 5050 l = n; 5051 /* Process subpages at the end of huge page */ 5052 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 5053 cond_resched(); 5054 process_subpage(addr + i * PAGE_SIZE, i, arg); 5055 } 5056 } else { 5057 /* If target subpage in second half of huge page */ 5058 base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 5059 l = pages_per_huge_page - n; 5060 /* Process subpages at the begin of huge page */ 5061 for (i = 0; i < base; i++) { 5062 cond_resched(); 5063 process_subpage(addr + i * PAGE_SIZE, i, arg); 5064 } 5065 } 5066 /* 5067 * Process remaining subpages in left-right-left-right pattern 5068 * towards the target subpage 5069 */ 5070 for (i = 0; i < l; i++) { 5071 int left_idx = base + i; 5072 int right_idx = base + 2 * l - 1 - i; 5073 5074 cond_resched(); 5075 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 5076 cond_resched(); 5077 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 5078 } 5079 } 5080 5081 static void clear_gigantic_page(struct page *page, 5082 unsigned long addr, 5083 unsigned int pages_per_huge_page) 5084 { 5085 int i; 5086 struct page *p = page; 5087 5088 might_sleep(); 5089 for (i = 0; i < pages_per_huge_page; 5090 i++, p = mem_map_next(p, page, i)) { 5091 cond_resched(); 5092 clear_user_highpage(p, addr + i * PAGE_SIZE); 5093 } 5094 } 5095 5096 static void clear_subpage(unsigned long addr, int idx, void *arg) 5097 { 5098 struct page *page = arg; 5099 5100 clear_user_highpage(page + idx, addr); 5101 } 5102 5103 void clear_huge_page(struct page *page, 5104 unsigned long addr_hint, unsigned int pages_per_huge_page) 5105 { 5106 unsigned long addr = addr_hint & 5107 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5108 5109 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5110 clear_gigantic_page(page, addr, pages_per_huge_page); 5111 return; 5112 } 5113 5114 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 5115 } 5116 5117 static void copy_user_gigantic_page(struct page *dst, struct page *src, 5118 unsigned long addr, 5119 struct vm_area_struct *vma, 5120 unsigned int pages_per_huge_page) 5121 { 5122 int i; 5123 struct page *dst_base = dst; 5124 struct page *src_base = src; 5125 5126 for (i = 0; i < pages_per_huge_page; ) { 5127 cond_resched(); 5128 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 5129 5130 i++; 5131 dst = mem_map_next(dst, dst_base, i); 5132 src = mem_map_next(src, src_base, i); 5133 } 5134 } 5135 5136 struct copy_subpage_arg { 5137 struct page *dst; 5138 struct page *src; 5139 struct vm_area_struct *vma; 5140 }; 5141 5142 static void copy_subpage(unsigned long addr, int idx, void *arg) 5143 { 5144 struct copy_subpage_arg *copy_arg = arg; 5145 5146 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx, 5147 addr, copy_arg->vma); 5148 } 5149 5150 void copy_user_huge_page(struct page *dst, struct page *src, 5151 unsigned long addr_hint, struct vm_area_struct *vma, 5152 unsigned int pages_per_huge_page) 5153 { 5154 unsigned long addr = addr_hint & 5155 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5156 struct copy_subpage_arg arg = { 5157 .dst = dst, 5158 .src = src, 5159 .vma = vma, 5160 }; 5161 5162 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5163 copy_user_gigantic_page(dst, src, addr, vma, 5164 pages_per_huge_page); 5165 return; 5166 } 5167 5168 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 5169 } 5170 5171 long copy_huge_page_from_user(struct page *dst_page, 5172 const void __user *usr_src, 5173 unsigned int pages_per_huge_page, 5174 bool allow_pagefault) 5175 { 5176 void *src = (void *)usr_src; 5177 void *page_kaddr; 5178 unsigned long i, rc = 0; 5179 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE; 5180 5181 for (i = 0; i < pages_per_huge_page; i++) { 5182 if (allow_pagefault) 5183 page_kaddr = kmap(dst_page + i); 5184 else 5185 page_kaddr = kmap_atomic(dst_page + i); 5186 rc = copy_from_user(page_kaddr, 5187 (const void __user *)(src + i * PAGE_SIZE), 5188 PAGE_SIZE); 5189 if (allow_pagefault) 5190 kunmap(dst_page + i); 5191 else 5192 kunmap_atomic(page_kaddr); 5193 5194 ret_val -= (PAGE_SIZE - rc); 5195 if (rc) 5196 break; 5197 5198 cond_resched(); 5199 } 5200 return ret_val; 5201 } 5202 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 5203 5204 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 5205 5206 static struct kmem_cache *page_ptl_cachep; 5207 5208 void __init ptlock_cache_init(void) 5209 { 5210 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 5211 SLAB_PANIC, NULL); 5212 } 5213 5214 bool ptlock_alloc(struct page *page) 5215 { 5216 spinlock_t *ptl; 5217 5218 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 5219 if (!ptl) 5220 return false; 5221 page->ptl = ptl; 5222 return true; 5223 } 5224 5225 void ptlock_free(struct page *page) 5226 { 5227 kmem_cache_free(page_ptl_cachep, page->ptl); 5228 } 5229 #endif 5230