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 static int __init disable_randmaps(char *s) 138 { 139 randomize_va_space = 0; 140 return 1; 141 } 142 __setup("norandmaps", disable_randmaps); 143 144 unsigned long zero_pfn __read_mostly; 145 EXPORT_SYMBOL(zero_pfn); 146 147 unsigned long highest_memmap_pfn __read_mostly; 148 149 /* 150 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 151 */ 152 static int __init init_zero_pfn(void) 153 { 154 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 155 return 0; 156 } 157 core_initcall(init_zero_pfn); 158 159 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count) 160 { 161 trace_rss_stat(mm, member, count); 162 } 163 164 #if defined(SPLIT_RSS_COUNTING) 165 166 void sync_mm_rss(struct mm_struct *mm) 167 { 168 int i; 169 170 for (i = 0; i < NR_MM_COUNTERS; i++) { 171 if (current->rss_stat.count[i]) { 172 add_mm_counter(mm, i, current->rss_stat.count[i]); 173 current->rss_stat.count[i] = 0; 174 } 175 } 176 current->rss_stat.events = 0; 177 } 178 179 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 180 { 181 struct task_struct *task = current; 182 183 if (likely(task->mm == mm)) 184 task->rss_stat.count[member] += val; 185 else 186 add_mm_counter(mm, member, val); 187 } 188 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 189 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 190 191 /* sync counter once per 64 page faults */ 192 #define TASK_RSS_EVENTS_THRESH (64) 193 static void check_sync_rss_stat(struct task_struct *task) 194 { 195 if (unlikely(task != current)) 196 return; 197 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 198 sync_mm_rss(task->mm); 199 } 200 #else /* SPLIT_RSS_COUNTING */ 201 202 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 203 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 204 205 static void check_sync_rss_stat(struct task_struct *task) 206 { 207 } 208 209 #endif /* SPLIT_RSS_COUNTING */ 210 211 /* 212 * Note: this doesn't free the actual pages themselves. That 213 * has been handled earlier when unmapping all the memory regions. 214 */ 215 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 216 unsigned long addr) 217 { 218 pgtable_t token = pmd_pgtable(*pmd); 219 pmd_clear(pmd); 220 pte_free_tlb(tlb, token, addr); 221 mm_dec_nr_ptes(tlb->mm); 222 } 223 224 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 225 unsigned long addr, unsigned long end, 226 unsigned long floor, unsigned long ceiling) 227 { 228 pmd_t *pmd; 229 unsigned long next; 230 unsigned long start; 231 232 start = addr; 233 pmd = pmd_offset(pud, addr); 234 do { 235 next = pmd_addr_end(addr, end); 236 if (pmd_none_or_clear_bad(pmd)) 237 continue; 238 free_pte_range(tlb, pmd, addr); 239 } while (pmd++, addr = next, addr != end); 240 241 start &= PUD_MASK; 242 if (start < floor) 243 return; 244 if (ceiling) { 245 ceiling &= PUD_MASK; 246 if (!ceiling) 247 return; 248 } 249 if (end - 1 > ceiling - 1) 250 return; 251 252 pmd = pmd_offset(pud, start); 253 pud_clear(pud); 254 pmd_free_tlb(tlb, pmd, start); 255 mm_dec_nr_pmds(tlb->mm); 256 } 257 258 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 259 unsigned long addr, unsigned long end, 260 unsigned long floor, unsigned long ceiling) 261 { 262 pud_t *pud; 263 unsigned long next; 264 unsigned long start; 265 266 start = addr; 267 pud = pud_offset(p4d, addr); 268 do { 269 next = pud_addr_end(addr, end); 270 if (pud_none_or_clear_bad(pud)) 271 continue; 272 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 273 } while (pud++, addr = next, addr != end); 274 275 start &= P4D_MASK; 276 if (start < floor) 277 return; 278 if (ceiling) { 279 ceiling &= P4D_MASK; 280 if (!ceiling) 281 return; 282 } 283 if (end - 1 > ceiling - 1) 284 return; 285 286 pud = pud_offset(p4d, start); 287 p4d_clear(p4d); 288 pud_free_tlb(tlb, pud, start); 289 mm_dec_nr_puds(tlb->mm); 290 } 291 292 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 293 unsigned long addr, unsigned long end, 294 unsigned long floor, unsigned long ceiling) 295 { 296 p4d_t *p4d; 297 unsigned long next; 298 unsigned long start; 299 300 start = addr; 301 p4d = p4d_offset(pgd, addr); 302 do { 303 next = p4d_addr_end(addr, end); 304 if (p4d_none_or_clear_bad(p4d)) 305 continue; 306 free_pud_range(tlb, p4d, addr, next, floor, ceiling); 307 } while (p4d++, addr = next, addr != end); 308 309 start &= PGDIR_MASK; 310 if (start < floor) 311 return; 312 if (ceiling) { 313 ceiling &= PGDIR_MASK; 314 if (!ceiling) 315 return; 316 } 317 if (end - 1 > ceiling - 1) 318 return; 319 320 p4d = p4d_offset(pgd, start); 321 pgd_clear(pgd); 322 p4d_free_tlb(tlb, p4d, start); 323 } 324 325 /* 326 * This function frees user-level page tables of a process. 327 */ 328 void free_pgd_range(struct mmu_gather *tlb, 329 unsigned long addr, unsigned long end, 330 unsigned long floor, unsigned long ceiling) 331 { 332 pgd_t *pgd; 333 unsigned long next; 334 335 /* 336 * The next few lines have given us lots of grief... 337 * 338 * Why are we testing PMD* at this top level? Because often 339 * there will be no work to do at all, and we'd prefer not to 340 * go all the way down to the bottom just to discover that. 341 * 342 * Why all these "- 1"s? Because 0 represents both the bottom 343 * of the address space and the top of it (using -1 for the 344 * top wouldn't help much: the masks would do the wrong thing). 345 * The rule is that addr 0 and floor 0 refer to the bottom of 346 * the address space, but end 0 and ceiling 0 refer to the top 347 * Comparisons need to use "end - 1" and "ceiling - 1" (though 348 * that end 0 case should be mythical). 349 * 350 * Wherever addr is brought up or ceiling brought down, we must 351 * be careful to reject "the opposite 0" before it confuses the 352 * subsequent tests. But what about where end is brought down 353 * by PMD_SIZE below? no, end can't go down to 0 there. 354 * 355 * Whereas we round start (addr) and ceiling down, by different 356 * masks at different levels, in order to test whether a table 357 * now has no other vmas using it, so can be freed, we don't 358 * bother to round floor or end up - the tests don't need that. 359 */ 360 361 addr &= PMD_MASK; 362 if (addr < floor) { 363 addr += PMD_SIZE; 364 if (!addr) 365 return; 366 } 367 if (ceiling) { 368 ceiling &= PMD_MASK; 369 if (!ceiling) 370 return; 371 } 372 if (end - 1 > ceiling - 1) 373 end -= PMD_SIZE; 374 if (addr > end - 1) 375 return; 376 /* 377 * We add page table cache pages with PAGE_SIZE, 378 * (see pte_free_tlb()), flush the tlb if we need 379 */ 380 tlb_change_page_size(tlb, PAGE_SIZE); 381 pgd = pgd_offset(tlb->mm, addr); 382 do { 383 next = pgd_addr_end(addr, end); 384 if (pgd_none_or_clear_bad(pgd)) 385 continue; 386 free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 387 } while (pgd++, addr = next, addr != end); 388 } 389 390 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 391 unsigned long floor, unsigned long ceiling) 392 { 393 while (vma) { 394 struct vm_area_struct *next = vma->vm_next; 395 unsigned long addr = vma->vm_start; 396 397 /* 398 * Hide vma from rmap and truncate_pagecache before freeing 399 * pgtables 400 */ 401 unlink_anon_vmas(vma); 402 unlink_file_vma(vma); 403 404 if (is_vm_hugetlb_page(vma)) { 405 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 406 floor, next ? next->vm_start : ceiling); 407 } else { 408 /* 409 * Optimization: gather nearby vmas into one call down 410 */ 411 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 412 && !is_vm_hugetlb_page(next)) { 413 vma = next; 414 next = vma->vm_next; 415 unlink_anon_vmas(vma); 416 unlink_file_vma(vma); 417 } 418 free_pgd_range(tlb, addr, vma->vm_end, 419 floor, next ? next->vm_start : ceiling); 420 } 421 vma = next; 422 } 423 } 424 425 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 426 { 427 spinlock_t *ptl; 428 pgtable_t new = pte_alloc_one(mm); 429 if (!new) 430 return -ENOMEM; 431 432 /* 433 * Ensure all pte setup (eg. pte page lock and page clearing) are 434 * visible before the pte is made visible to other CPUs by being 435 * put into page tables. 436 * 437 * The other side of the story is the pointer chasing in the page 438 * table walking code (when walking the page table without locking; 439 * ie. most of the time). Fortunately, these data accesses consist 440 * of a chain of data-dependent loads, meaning most CPUs (alpha 441 * being the notable exception) will already guarantee loads are 442 * seen in-order. See the alpha page table accessors for the 443 * smp_rmb() barriers in page table walking code. 444 */ 445 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 446 447 ptl = pmd_lock(mm, pmd); 448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 449 mm_inc_nr_ptes(mm); 450 pmd_populate(mm, pmd, new); 451 new = NULL; 452 } 453 spin_unlock(ptl); 454 if (new) 455 pte_free(mm, new); 456 return 0; 457 } 458 459 int __pte_alloc_kernel(pmd_t *pmd) 460 { 461 pte_t *new = pte_alloc_one_kernel(&init_mm); 462 if (!new) 463 return -ENOMEM; 464 465 smp_wmb(); /* See comment in __pte_alloc */ 466 467 spin_lock(&init_mm.page_table_lock); 468 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 469 pmd_populate_kernel(&init_mm, pmd, new); 470 new = NULL; 471 } 472 spin_unlock(&init_mm.page_table_lock); 473 if (new) 474 pte_free_kernel(&init_mm, new); 475 return 0; 476 } 477 478 static inline void init_rss_vec(int *rss) 479 { 480 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 481 } 482 483 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 484 { 485 int i; 486 487 if (current->mm == mm) 488 sync_mm_rss(mm); 489 for (i = 0; i < NR_MM_COUNTERS; i++) 490 if (rss[i]) 491 add_mm_counter(mm, i, rss[i]); 492 } 493 494 /* 495 * This function is called to print an error when a bad pte 496 * is found. For example, we might have a PFN-mapped pte in 497 * a region that doesn't allow it. 498 * 499 * The calling function must still handle the error. 500 */ 501 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 502 pte_t pte, struct page *page) 503 { 504 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 505 p4d_t *p4d = p4d_offset(pgd, addr); 506 pud_t *pud = pud_offset(p4d, addr); 507 pmd_t *pmd = pmd_offset(pud, addr); 508 struct address_space *mapping; 509 pgoff_t index; 510 static unsigned long resume; 511 static unsigned long nr_shown; 512 static unsigned long nr_unshown; 513 514 /* 515 * Allow a burst of 60 reports, then keep quiet for that minute; 516 * or allow a steady drip of one report per second. 517 */ 518 if (nr_shown == 60) { 519 if (time_before(jiffies, resume)) { 520 nr_unshown++; 521 return; 522 } 523 if (nr_unshown) { 524 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 525 nr_unshown); 526 nr_unshown = 0; 527 } 528 nr_shown = 0; 529 } 530 if (nr_shown++ == 0) 531 resume = jiffies + 60 * HZ; 532 533 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 534 index = linear_page_index(vma, addr); 535 536 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 537 current->comm, 538 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 539 if (page) 540 dump_page(page, "bad pte"); 541 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", 542 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 543 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n", 544 vma->vm_file, 545 vma->vm_ops ? vma->vm_ops->fault : NULL, 546 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 547 mapping ? mapping->a_ops->readpage : NULL); 548 dump_stack(); 549 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 550 } 551 552 /* 553 * vm_normal_page -- This function gets the "struct page" associated with a pte. 554 * 555 * "Special" mappings do not wish to be associated with a "struct page" (either 556 * it doesn't exist, or it exists but they don't want to touch it). In this 557 * case, NULL is returned here. "Normal" mappings do have a struct page. 558 * 559 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 560 * pte bit, in which case this function is trivial. Secondly, an architecture 561 * may not have a spare pte bit, which requires a more complicated scheme, 562 * described below. 563 * 564 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 565 * special mapping (even if there are underlying and valid "struct pages"). 566 * COWed pages of a VM_PFNMAP are always normal. 567 * 568 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 569 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 570 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 571 * mapping will always honor the rule 572 * 573 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 574 * 575 * And for normal mappings this is false. 576 * 577 * This restricts such mappings to be a linear translation from virtual address 578 * to pfn. To get around this restriction, we allow arbitrary mappings so long 579 * as the vma is not a COW mapping; in that case, we know that all ptes are 580 * special (because none can have been COWed). 581 * 582 * 583 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 584 * 585 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 586 * page" backing, however the difference is that _all_ pages with a struct 587 * page (that is, those where pfn_valid is true) are refcounted and considered 588 * normal pages by the VM. The disadvantage is that pages are refcounted 589 * (which can be slower and simply not an option for some PFNMAP users). The 590 * advantage is that we don't have to follow the strict linearity rule of 591 * PFNMAP mappings in order to support COWable mappings. 592 * 593 */ 594 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 595 pte_t pte) 596 { 597 unsigned long pfn = pte_pfn(pte); 598 599 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 600 if (likely(!pte_special(pte))) 601 goto check_pfn; 602 if (vma->vm_ops && vma->vm_ops->find_special_page) 603 return vma->vm_ops->find_special_page(vma, addr); 604 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 605 return NULL; 606 if (is_zero_pfn(pfn)) 607 return NULL; 608 if (pte_devmap(pte)) 609 return NULL; 610 611 print_bad_pte(vma, addr, pte, NULL); 612 return NULL; 613 } 614 615 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 616 617 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 618 if (vma->vm_flags & VM_MIXEDMAP) { 619 if (!pfn_valid(pfn)) 620 return NULL; 621 goto out; 622 } else { 623 unsigned long off; 624 off = (addr - vma->vm_start) >> PAGE_SHIFT; 625 if (pfn == vma->vm_pgoff + off) 626 return NULL; 627 if (!is_cow_mapping(vma->vm_flags)) 628 return NULL; 629 } 630 } 631 632 if (is_zero_pfn(pfn)) 633 return NULL; 634 635 check_pfn: 636 if (unlikely(pfn > highest_memmap_pfn)) { 637 print_bad_pte(vma, addr, pte, NULL); 638 return NULL; 639 } 640 641 /* 642 * NOTE! We still have PageReserved() pages in the page tables. 643 * eg. VDSO mappings can cause them to exist. 644 */ 645 out: 646 return pfn_to_page(pfn); 647 } 648 649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 650 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 651 pmd_t pmd) 652 { 653 unsigned long pfn = pmd_pfn(pmd); 654 655 /* 656 * There is no pmd_special() but there may be special pmds, e.g. 657 * in a direct-access (dax) mapping, so let's just replicate the 658 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 659 */ 660 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 661 if (vma->vm_flags & VM_MIXEDMAP) { 662 if (!pfn_valid(pfn)) 663 return NULL; 664 goto out; 665 } else { 666 unsigned long off; 667 off = (addr - vma->vm_start) >> PAGE_SHIFT; 668 if (pfn == vma->vm_pgoff + off) 669 return NULL; 670 if (!is_cow_mapping(vma->vm_flags)) 671 return NULL; 672 } 673 } 674 675 if (pmd_devmap(pmd)) 676 return NULL; 677 if (is_huge_zero_pmd(pmd)) 678 return NULL; 679 if (unlikely(pfn > highest_memmap_pfn)) 680 return NULL; 681 682 /* 683 * NOTE! We still have PageReserved() pages in the page tables. 684 * eg. VDSO mappings can cause them to exist. 685 */ 686 out: 687 return pfn_to_page(pfn); 688 } 689 #endif 690 691 /* 692 * copy one vm_area from one task to the other. Assumes the page tables 693 * already present in the new task to be cleared in the whole range 694 * covered by this vma. 695 */ 696 697 static unsigned long 698 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 699 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 700 unsigned long addr, int *rss) 701 { 702 unsigned long vm_flags = vma->vm_flags; 703 pte_t pte = *src_pte; 704 struct page *page; 705 swp_entry_t entry = pte_to_swp_entry(pte); 706 707 if (likely(!non_swap_entry(entry))) { 708 if (swap_duplicate(entry) < 0) 709 return entry.val; 710 711 /* make sure dst_mm is on swapoff's mmlist. */ 712 if (unlikely(list_empty(&dst_mm->mmlist))) { 713 spin_lock(&mmlist_lock); 714 if (list_empty(&dst_mm->mmlist)) 715 list_add(&dst_mm->mmlist, 716 &src_mm->mmlist); 717 spin_unlock(&mmlist_lock); 718 } 719 rss[MM_SWAPENTS]++; 720 } else if (is_migration_entry(entry)) { 721 page = migration_entry_to_page(entry); 722 723 rss[mm_counter(page)]++; 724 725 if (is_write_migration_entry(entry) && 726 is_cow_mapping(vm_flags)) { 727 /* 728 * COW mappings require pages in both 729 * parent and child to be set to read. 730 */ 731 make_migration_entry_read(&entry); 732 pte = swp_entry_to_pte(entry); 733 if (pte_swp_soft_dirty(*src_pte)) 734 pte = pte_swp_mksoft_dirty(pte); 735 if (pte_swp_uffd_wp(*src_pte)) 736 pte = pte_swp_mkuffd_wp(pte); 737 set_pte_at(src_mm, addr, src_pte, pte); 738 } 739 } else if (is_device_private_entry(entry)) { 740 page = device_private_entry_to_page(entry); 741 742 /* 743 * Update rss count even for unaddressable pages, as 744 * they should treated just like normal pages in this 745 * respect. 746 * 747 * We will likely want to have some new rss counters 748 * for unaddressable pages, at some point. But for now 749 * keep things as they are. 750 */ 751 get_page(page); 752 rss[mm_counter(page)]++; 753 page_dup_rmap(page, false); 754 755 /* 756 * We do not preserve soft-dirty information, because so 757 * far, checkpoint/restore is the only feature that 758 * requires that. And checkpoint/restore does not work 759 * when a device driver is involved (you cannot easily 760 * save and restore device driver state). 761 */ 762 if (is_write_device_private_entry(entry) && 763 is_cow_mapping(vm_flags)) { 764 make_device_private_entry_read(&entry); 765 pte = swp_entry_to_pte(entry); 766 if (pte_swp_uffd_wp(*src_pte)) 767 pte = pte_swp_mkuffd_wp(pte); 768 set_pte_at(src_mm, addr, src_pte, pte); 769 } 770 } 771 set_pte_at(dst_mm, addr, dst_pte, pte); 772 return 0; 773 } 774 775 /* 776 * Copy a present and normal page if necessary. 777 * 778 * NOTE! The usual case is that this doesn't need to do 779 * anything, and can just return a positive value. That 780 * will let the caller know that it can just increase 781 * the page refcount and re-use the pte the traditional 782 * way. 783 * 784 * But _if_ we need to copy it because it needs to be 785 * pinned in the parent (and the child should get its own 786 * copy rather than just a reference to the same page), 787 * we'll do that here and return zero to let the caller 788 * know we're done. 789 * 790 * And if we need a pre-allocated page but don't yet have 791 * one, return a negative error to let the preallocation 792 * code know so that it can do so outside the page table 793 * lock. 794 */ 795 static inline int 796 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 797 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 798 struct page **prealloc, pte_t pte, struct page *page) 799 { 800 struct mm_struct *src_mm = src_vma->vm_mm; 801 struct page *new_page; 802 803 if (!is_cow_mapping(src_vma->vm_flags)) 804 return 1; 805 806 /* 807 * What we want to do is to check whether this page may 808 * have been pinned by the parent process. If so, 809 * instead of wrprotect the pte on both sides, we copy 810 * the page immediately so that we'll always guarantee 811 * the pinned page won't be randomly replaced in the 812 * future. 813 * 814 * The page pinning checks are just "has this mm ever 815 * seen pinning", along with the (inexact) check of 816 * the page count. That might give false positives for 817 * for pinning, but it will work correctly. 818 */ 819 if (likely(!atomic_read(&src_mm->has_pinned))) 820 return 1; 821 if (likely(!page_maybe_dma_pinned(page))) 822 return 1; 823 824 new_page = *prealloc; 825 if (!new_page) 826 return -EAGAIN; 827 828 /* 829 * We have a prealloc page, all good! Take it 830 * over and copy the page & arm it. 831 */ 832 *prealloc = NULL; 833 copy_user_highpage(new_page, page, addr, src_vma); 834 __SetPageUptodate(new_page); 835 page_add_new_anon_rmap(new_page, dst_vma, addr, false); 836 lru_cache_add_inactive_or_unevictable(new_page, dst_vma); 837 rss[mm_counter(new_page)]++; 838 839 /* All done, just insert the new page copy in the child */ 840 pte = mk_pte(new_page, dst_vma->vm_page_prot); 841 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); 842 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 843 return 0; 844 } 845 846 /* 847 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page 848 * is required to copy this pte. 849 */ 850 static inline int 851 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 852 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 853 struct page **prealloc) 854 { 855 struct mm_struct *src_mm = src_vma->vm_mm; 856 unsigned long vm_flags = src_vma->vm_flags; 857 pte_t pte = *src_pte; 858 struct page *page; 859 860 page = vm_normal_page(src_vma, addr, pte); 861 if (page) { 862 int retval; 863 864 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte, 865 addr, rss, prealloc, pte, page); 866 if (retval <= 0) 867 return retval; 868 869 get_page(page); 870 page_dup_rmap(page, false); 871 rss[mm_counter(page)]++; 872 } 873 874 /* 875 * If it's a COW mapping, write protect it both 876 * in the parent and the child 877 */ 878 if (is_cow_mapping(vm_flags) && pte_write(pte)) { 879 ptep_set_wrprotect(src_mm, addr, src_pte); 880 pte = pte_wrprotect(pte); 881 } 882 883 /* 884 * If it's a shared mapping, mark it clean in 885 * the child 886 */ 887 if (vm_flags & VM_SHARED) 888 pte = pte_mkclean(pte); 889 pte = pte_mkold(pte); 890 891 /* 892 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA 893 * does not have the VM_UFFD_WP, which means that the uffd 894 * fork event is not enabled. 895 */ 896 if (!(vm_flags & VM_UFFD_WP)) 897 pte = pte_clear_uffd_wp(pte); 898 899 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 900 return 0; 901 } 902 903 static inline struct page * 904 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma, 905 unsigned long addr) 906 { 907 struct page *new_page; 908 909 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr); 910 if (!new_page) 911 return NULL; 912 913 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) { 914 put_page(new_page); 915 return NULL; 916 } 917 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 918 919 return new_page; 920 } 921 922 static int 923 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 924 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 925 unsigned long end) 926 { 927 struct mm_struct *dst_mm = dst_vma->vm_mm; 928 struct mm_struct *src_mm = src_vma->vm_mm; 929 pte_t *orig_src_pte, *orig_dst_pte; 930 pte_t *src_pte, *dst_pte; 931 spinlock_t *src_ptl, *dst_ptl; 932 int progress, ret = 0; 933 int rss[NR_MM_COUNTERS]; 934 swp_entry_t entry = (swp_entry_t){0}; 935 struct page *prealloc = NULL; 936 937 again: 938 progress = 0; 939 init_rss_vec(rss); 940 941 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 942 if (!dst_pte) { 943 ret = -ENOMEM; 944 goto out; 945 } 946 src_pte = pte_offset_map(src_pmd, addr); 947 src_ptl = pte_lockptr(src_mm, src_pmd); 948 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 949 orig_src_pte = src_pte; 950 orig_dst_pte = dst_pte; 951 arch_enter_lazy_mmu_mode(); 952 953 do { 954 /* 955 * We are holding two locks at this point - either of them 956 * could generate latencies in another task on another CPU. 957 */ 958 if (progress >= 32) { 959 progress = 0; 960 if (need_resched() || 961 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 962 break; 963 } 964 if (pte_none(*src_pte)) { 965 progress++; 966 continue; 967 } 968 if (unlikely(!pte_present(*src_pte))) { 969 entry.val = copy_nonpresent_pte(dst_mm, src_mm, 970 dst_pte, src_pte, 971 src_vma, addr, rss); 972 if (entry.val) 973 break; 974 progress += 8; 975 continue; 976 } 977 /* copy_present_pte() will clear `*prealloc' if consumed */ 978 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte, 979 addr, rss, &prealloc); 980 /* 981 * If we need a pre-allocated page for this pte, drop the 982 * locks, allocate, and try again. 983 */ 984 if (unlikely(ret == -EAGAIN)) 985 break; 986 if (unlikely(prealloc)) { 987 /* 988 * pre-alloc page cannot be reused by next time so as 989 * to strictly follow mempolicy (e.g., alloc_page_vma() 990 * will allocate page according to address). This 991 * could only happen if one pinned pte changed. 992 */ 993 put_page(prealloc); 994 prealloc = NULL; 995 } 996 progress += 8; 997 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 998 999 arch_leave_lazy_mmu_mode(); 1000 spin_unlock(src_ptl); 1001 pte_unmap(orig_src_pte); 1002 add_mm_rss_vec(dst_mm, rss); 1003 pte_unmap_unlock(orig_dst_pte, dst_ptl); 1004 cond_resched(); 1005 1006 if (entry.val) { 1007 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { 1008 ret = -ENOMEM; 1009 goto out; 1010 } 1011 entry.val = 0; 1012 } else if (ret) { 1013 WARN_ON_ONCE(ret != -EAGAIN); 1014 prealloc = page_copy_prealloc(src_mm, src_vma, addr); 1015 if (!prealloc) 1016 return -ENOMEM; 1017 /* We've captured and resolved the error. Reset, try again. */ 1018 ret = 0; 1019 } 1020 if (addr != end) 1021 goto again; 1022 out: 1023 if (unlikely(prealloc)) 1024 put_page(prealloc); 1025 return ret; 1026 } 1027 1028 static inline int 1029 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1030 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1031 unsigned long end) 1032 { 1033 struct mm_struct *dst_mm = dst_vma->vm_mm; 1034 struct mm_struct *src_mm = src_vma->vm_mm; 1035 pmd_t *src_pmd, *dst_pmd; 1036 unsigned long next; 1037 1038 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 1039 if (!dst_pmd) 1040 return -ENOMEM; 1041 src_pmd = pmd_offset(src_pud, addr); 1042 do { 1043 next = pmd_addr_end(addr, end); 1044 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 1045 || pmd_devmap(*src_pmd)) { 1046 int err; 1047 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); 1048 err = copy_huge_pmd(dst_mm, src_mm, 1049 dst_pmd, src_pmd, addr, src_vma); 1050 if (err == -ENOMEM) 1051 return -ENOMEM; 1052 if (!err) 1053 continue; 1054 /* fall through */ 1055 } 1056 if (pmd_none_or_clear_bad(src_pmd)) 1057 continue; 1058 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, 1059 addr, next)) 1060 return -ENOMEM; 1061 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 1062 return 0; 1063 } 1064 1065 static inline int 1066 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1067 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, 1068 unsigned long end) 1069 { 1070 struct mm_struct *dst_mm = dst_vma->vm_mm; 1071 struct mm_struct *src_mm = src_vma->vm_mm; 1072 pud_t *src_pud, *dst_pud; 1073 unsigned long next; 1074 1075 dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 1076 if (!dst_pud) 1077 return -ENOMEM; 1078 src_pud = pud_offset(src_p4d, addr); 1079 do { 1080 next = pud_addr_end(addr, end); 1081 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 1082 int err; 1083 1084 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); 1085 err = copy_huge_pud(dst_mm, src_mm, 1086 dst_pud, src_pud, addr, src_vma); 1087 if (err == -ENOMEM) 1088 return -ENOMEM; 1089 if (!err) 1090 continue; 1091 /* fall through */ 1092 } 1093 if (pud_none_or_clear_bad(src_pud)) 1094 continue; 1095 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, 1096 addr, next)) 1097 return -ENOMEM; 1098 } while (dst_pud++, src_pud++, addr = next, addr != end); 1099 return 0; 1100 } 1101 1102 static inline int 1103 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1104 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, 1105 unsigned long end) 1106 { 1107 struct mm_struct *dst_mm = dst_vma->vm_mm; 1108 p4d_t *src_p4d, *dst_p4d; 1109 unsigned long next; 1110 1111 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 1112 if (!dst_p4d) 1113 return -ENOMEM; 1114 src_p4d = p4d_offset(src_pgd, addr); 1115 do { 1116 next = p4d_addr_end(addr, end); 1117 if (p4d_none_or_clear_bad(src_p4d)) 1118 continue; 1119 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, 1120 addr, next)) 1121 return -ENOMEM; 1122 } while (dst_p4d++, src_p4d++, addr = next, addr != end); 1123 return 0; 1124 } 1125 1126 int 1127 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1128 { 1129 pgd_t *src_pgd, *dst_pgd; 1130 unsigned long next; 1131 unsigned long addr = src_vma->vm_start; 1132 unsigned long end = src_vma->vm_end; 1133 struct mm_struct *dst_mm = dst_vma->vm_mm; 1134 struct mm_struct *src_mm = src_vma->vm_mm; 1135 struct mmu_notifier_range range; 1136 bool is_cow; 1137 int ret; 1138 1139 /* 1140 * Don't copy ptes where a page fault will fill them correctly. 1141 * Fork becomes much lighter when there are big shared or private 1142 * readonly mappings. The tradeoff is that copy_page_range is more 1143 * efficient than faulting. 1144 */ 1145 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 1146 !src_vma->anon_vma) 1147 return 0; 1148 1149 if (is_vm_hugetlb_page(src_vma)) 1150 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma); 1151 1152 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { 1153 /* 1154 * We do not free on error cases below as remove_vma 1155 * gets called on error from higher level routine 1156 */ 1157 ret = track_pfn_copy(src_vma); 1158 if (ret) 1159 return ret; 1160 } 1161 1162 /* 1163 * We need to invalidate the secondary MMU mappings only when 1164 * there could be a permission downgrade on the ptes of the 1165 * parent mm. And a permission downgrade will only happen if 1166 * is_cow_mapping() returns true. 1167 */ 1168 is_cow = is_cow_mapping(src_vma->vm_flags); 1169 1170 if (is_cow) { 1171 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 1172 0, src_vma, src_mm, addr, end); 1173 mmu_notifier_invalidate_range_start(&range); 1174 /* 1175 * Disabling preemption is not needed for the write side, as 1176 * the read side doesn't spin, but goes to the mmap_lock. 1177 * 1178 * Use the raw variant of the seqcount_t write API to avoid 1179 * lockdep complaining about preemptibility. 1180 */ 1181 mmap_assert_write_locked(src_mm); 1182 raw_write_seqcount_begin(&src_mm->write_protect_seq); 1183 } 1184 1185 ret = 0; 1186 dst_pgd = pgd_offset(dst_mm, addr); 1187 src_pgd = pgd_offset(src_mm, addr); 1188 do { 1189 next = pgd_addr_end(addr, end); 1190 if (pgd_none_or_clear_bad(src_pgd)) 1191 continue; 1192 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, 1193 addr, next))) { 1194 ret = -ENOMEM; 1195 break; 1196 } 1197 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1198 1199 if (is_cow) { 1200 raw_write_seqcount_end(&src_mm->write_protect_seq); 1201 mmu_notifier_invalidate_range_end(&range); 1202 } 1203 return ret; 1204 } 1205 1206 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1207 struct vm_area_struct *vma, pmd_t *pmd, 1208 unsigned long addr, unsigned long end, 1209 struct zap_details *details) 1210 { 1211 struct mm_struct *mm = tlb->mm; 1212 int force_flush = 0; 1213 int rss[NR_MM_COUNTERS]; 1214 spinlock_t *ptl; 1215 pte_t *start_pte; 1216 pte_t *pte; 1217 swp_entry_t entry; 1218 1219 tlb_change_page_size(tlb, PAGE_SIZE); 1220 again: 1221 init_rss_vec(rss); 1222 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1223 pte = start_pte; 1224 flush_tlb_batched_pending(mm); 1225 arch_enter_lazy_mmu_mode(); 1226 do { 1227 pte_t ptent = *pte; 1228 if (pte_none(ptent)) 1229 continue; 1230 1231 if (need_resched()) 1232 break; 1233 1234 if (pte_present(ptent)) { 1235 struct page *page; 1236 1237 page = vm_normal_page(vma, addr, ptent); 1238 if (unlikely(details) && page) { 1239 /* 1240 * unmap_shared_mapping_pages() wants to 1241 * invalidate cache without truncating: 1242 * unmap shared but keep private pages. 1243 */ 1244 if (details->check_mapping && 1245 details->check_mapping != page_rmapping(page)) 1246 continue; 1247 } 1248 ptent = ptep_get_and_clear_full(mm, addr, pte, 1249 tlb->fullmm); 1250 tlb_remove_tlb_entry(tlb, pte, addr); 1251 if (unlikely(!page)) 1252 continue; 1253 1254 if (!PageAnon(page)) { 1255 if (pte_dirty(ptent)) { 1256 force_flush = 1; 1257 set_page_dirty(page); 1258 } 1259 if (pte_young(ptent) && 1260 likely(!(vma->vm_flags & VM_SEQ_READ))) 1261 mark_page_accessed(page); 1262 } 1263 rss[mm_counter(page)]--; 1264 page_remove_rmap(page, false); 1265 if (unlikely(page_mapcount(page) < 0)) 1266 print_bad_pte(vma, addr, ptent, page); 1267 if (unlikely(__tlb_remove_page(tlb, page))) { 1268 force_flush = 1; 1269 addr += PAGE_SIZE; 1270 break; 1271 } 1272 continue; 1273 } 1274 1275 entry = pte_to_swp_entry(ptent); 1276 if (is_device_private_entry(entry)) { 1277 struct page *page = device_private_entry_to_page(entry); 1278 1279 if (unlikely(details && details->check_mapping)) { 1280 /* 1281 * unmap_shared_mapping_pages() wants to 1282 * invalidate cache without truncating: 1283 * unmap shared but keep private pages. 1284 */ 1285 if (details->check_mapping != 1286 page_rmapping(page)) 1287 continue; 1288 } 1289 1290 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1291 rss[mm_counter(page)]--; 1292 page_remove_rmap(page, false); 1293 put_page(page); 1294 continue; 1295 } 1296 1297 /* If details->check_mapping, we leave swap entries. */ 1298 if (unlikely(details)) 1299 continue; 1300 1301 if (!non_swap_entry(entry)) 1302 rss[MM_SWAPENTS]--; 1303 else if (is_migration_entry(entry)) { 1304 struct page *page; 1305 1306 page = migration_entry_to_page(entry); 1307 rss[mm_counter(page)]--; 1308 } 1309 if (unlikely(!free_swap_and_cache(entry))) 1310 print_bad_pte(vma, addr, ptent, NULL); 1311 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1312 } while (pte++, addr += PAGE_SIZE, addr != end); 1313 1314 add_mm_rss_vec(mm, rss); 1315 arch_leave_lazy_mmu_mode(); 1316 1317 /* Do the actual TLB flush before dropping ptl */ 1318 if (force_flush) 1319 tlb_flush_mmu_tlbonly(tlb); 1320 pte_unmap_unlock(start_pte, ptl); 1321 1322 /* 1323 * If we forced a TLB flush (either due to running out of 1324 * batch buffers or because we needed to flush dirty TLB 1325 * entries before releasing the ptl), free the batched 1326 * memory too. Restart if we didn't do everything. 1327 */ 1328 if (force_flush) { 1329 force_flush = 0; 1330 tlb_flush_mmu(tlb); 1331 } 1332 1333 if (addr != end) { 1334 cond_resched(); 1335 goto again; 1336 } 1337 1338 return addr; 1339 } 1340 1341 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1342 struct vm_area_struct *vma, pud_t *pud, 1343 unsigned long addr, unsigned long end, 1344 struct zap_details *details) 1345 { 1346 pmd_t *pmd; 1347 unsigned long next; 1348 1349 pmd = pmd_offset(pud, addr); 1350 do { 1351 next = pmd_addr_end(addr, end); 1352 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1353 if (next - addr != HPAGE_PMD_SIZE) 1354 __split_huge_pmd(vma, pmd, addr, false, NULL); 1355 else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1356 goto next; 1357 /* fall through */ 1358 } 1359 /* 1360 * Here there can be other concurrent MADV_DONTNEED or 1361 * trans huge page faults running, and if the pmd is 1362 * none or trans huge it can change under us. This is 1363 * because MADV_DONTNEED holds the mmap_lock in read 1364 * mode. 1365 */ 1366 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1367 goto next; 1368 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1369 next: 1370 cond_resched(); 1371 } while (pmd++, addr = next, addr != end); 1372 1373 return addr; 1374 } 1375 1376 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1377 struct vm_area_struct *vma, p4d_t *p4d, 1378 unsigned long addr, unsigned long end, 1379 struct zap_details *details) 1380 { 1381 pud_t *pud; 1382 unsigned long next; 1383 1384 pud = pud_offset(p4d, addr); 1385 do { 1386 next = pud_addr_end(addr, end); 1387 if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1388 if (next - addr != HPAGE_PUD_SIZE) { 1389 mmap_assert_locked(tlb->mm); 1390 split_huge_pud(vma, pud, addr); 1391 } else if (zap_huge_pud(tlb, vma, pud, addr)) 1392 goto next; 1393 /* fall through */ 1394 } 1395 if (pud_none_or_clear_bad(pud)) 1396 continue; 1397 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1398 next: 1399 cond_resched(); 1400 } while (pud++, addr = next, addr != end); 1401 1402 return addr; 1403 } 1404 1405 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1406 struct vm_area_struct *vma, pgd_t *pgd, 1407 unsigned long addr, unsigned long end, 1408 struct zap_details *details) 1409 { 1410 p4d_t *p4d; 1411 unsigned long next; 1412 1413 p4d = p4d_offset(pgd, addr); 1414 do { 1415 next = p4d_addr_end(addr, end); 1416 if (p4d_none_or_clear_bad(p4d)) 1417 continue; 1418 next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1419 } while (p4d++, addr = next, addr != end); 1420 1421 return addr; 1422 } 1423 1424 void unmap_page_range(struct mmu_gather *tlb, 1425 struct vm_area_struct *vma, 1426 unsigned long addr, unsigned long end, 1427 struct zap_details *details) 1428 { 1429 pgd_t *pgd; 1430 unsigned long next; 1431 1432 BUG_ON(addr >= end); 1433 tlb_start_vma(tlb, vma); 1434 pgd = pgd_offset(vma->vm_mm, addr); 1435 do { 1436 next = pgd_addr_end(addr, end); 1437 if (pgd_none_or_clear_bad(pgd)) 1438 continue; 1439 next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 1440 } while (pgd++, addr = next, addr != end); 1441 tlb_end_vma(tlb, vma); 1442 } 1443 1444 1445 static void unmap_single_vma(struct mmu_gather *tlb, 1446 struct vm_area_struct *vma, unsigned long start_addr, 1447 unsigned long end_addr, 1448 struct zap_details *details) 1449 { 1450 unsigned long start = max(vma->vm_start, start_addr); 1451 unsigned long end; 1452 1453 if (start >= vma->vm_end) 1454 return; 1455 end = min(vma->vm_end, end_addr); 1456 if (end <= vma->vm_start) 1457 return; 1458 1459 if (vma->vm_file) 1460 uprobe_munmap(vma, start, end); 1461 1462 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1463 untrack_pfn(vma, 0, 0); 1464 1465 if (start != end) { 1466 if (unlikely(is_vm_hugetlb_page(vma))) { 1467 /* 1468 * It is undesirable to test vma->vm_file as it 1469 * should be non-null for valid hugetlb area. 1470 * However, vm_file will be NULL in the error 1471 * cleanup path of mmap_region. When 1472 * hugetlbfs ->mmap method fails, 1473 * mmap_region() nullifies vma->vm_file 1474 * before calling this function to clean up. 1475 * Since no pte has actually been setup, it is 1476 * safe to do nothing in this case. 1477 */ 1478 if (vma->vm_file) { 1479 i_mmap_lock_write(vma->vm_file->f_mapping); 1480 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1481 i_mmap_unlock_write(vma->vm_file->f_mapping); 1482 } 1483 } else 1484 unmap_page_range(tlb, vma, start, end, details); 1485 } 1486 } 1487 1488 /** 1489 * unmap_vmas - unmap a range of memory covered by a list of vma's 1490 * @tlb: address of the caller's struct mmu_gather 1491 * @vma: the starting vma 1492 * @start_addr: virtual address at which to start unmapping 1493 * @end_addr: virtual address at which to end unmapping 1494 * 1495 * Unmap all pages in the vma list. 1496 * 1497 * Only addresses between `start' and `end' will be unmapped. 1498 * 1499 * The VMA list must be sorted in ascending virtual address order. 1500 * 1501 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1502 * range after unmap_vmas() returns. So the only responsibility here is to 1503 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1504 * drops the lock and schedules. 1505 */ 1506 void unmap_vmas(struct mmu_gather *tlb, 1507 struct vm_area_struct *vma, unsigned long start_addr, 1508 unsigned long end_addr) 1509 { 1510 struct mmu_notifier_range range; 1511 1512 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, 1513 start_addr, end_addr); 1514 mmu_notifier_invalidate_range_start(&range); 1515 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1516 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1517 mmu_notifier_invalidate_range_end(&range); 1518 } 1519 1520 /** 1521 * zap_page_range - remove user pages in a given range 1522 * @vma: vm_area_struct holding the applicable pages 1523 * @start: starting address of pages to zap 1524 * @size: number of bytes to zap 1525 * 1526 * Caller must protect the VMA list 1527 */ 1528 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1529 unsigned long size) 1530 { 1531 struct mmu_notifier_range range; 1532 struct mmu_gather tlb; 1533 1534 lru_add_drain(); 1535 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1536 start, start + size); 1537 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end); 1538 update_hiwater_rss(vma->vm_mm); 1539 mmu_notifier_invalidate_range_start(&range); 1540 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next) 1541 unmap_single_vma(&tlb, vma, start, range.end, NULL); 1542 mmu_notifier_invalidate_range_end(&range); 1543 tlb_finish_mmu(&tlb, start, range.end); 1544 } 1545 1546 /** 1547 * zap_page_range_single - remove user pages in a given range 1548 * @vma: vm_area_struct holding the applicable pages 1549 * @address: starting address of pages to zap 1550 * @size: number of bytes to zap 1551 * @details: details of shared cache invalidation 1552 * 1553 * The range must fit into one VMA. 1554 */ 1555 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1556 unsigned long size, struct zap_details *details) 1557 { 1558 struct mmu_notifier_range range; 1559 struct mmu_gather tlb; 1560 1561 lru_add_drain(); 1562 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1563 address, address + size); 1564 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end); 1565 update_hiwater_rss(vma->vm_mm); 1566 mmu_notifier_invalidate_range_start(&range); 1567 unmap_single_vma(&tlb, vma, address, range.end, details); 1568 mmu_notifier_invalidate_range_end(&range); 1569 tlb_finish_mmu(&tlb, address, range.end); 1570 } 1571 1572 /** 1573 * zap_vma_ptes - remove ptes mapping the vma 1574 * @vma: vm_area_struct holding ptes to be zapped 1575 * @address: starting address of pages to zap 1576 * @size: number of bytes to zap 1577 * 1578 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1579 * 1580 * The entire address range must be fully contained within the vma. 1581 * 1582 */ 1583 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1584 unsigned long size) 1585 { 1586 if (address < vma->vm_start || address + size > vma->vm_end || 1587 !(vma->vm_flags & VM_PFNMAP)) 1588 return; 1589 1590 zap_page_range_single(vma, address, size, NULL); 1591 } 1592 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1593 1594 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) 1595 { 1596 pgd_t *pgd; 1597 p4d_t *p4d; 1598 pud_t *pud; 1599 pmd_t *pmd; 1600 1601 pgd = pgd_offset(mm, addr); 1602 p4d = p4d_alloc(mm, pgd, addr); 1603 if (!p4d) 1604 return NULL; 1605 pud = pud_alloc(mm, p4d, addr); 1606 if (!pud) 1607 return NULL; 1608 pmd = pmd_alloc(mm, pud, addr); 1609 if (!pmd) 1610 return NULL; 1611 1612 VM_BUG_ON(pmd_trans_huge(*pmd)); 1613 return pmd; 1614 } 1615 1616 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1617 spinlock_t **ptl) 1618 { 1619 pmd_t *pmd = walk_to_pmd(mm, addr); 1620 1621 if (!pmd) 1622 return NULL; 1623 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1624 } 1625 1626 static int validate_page_before_insert(struct page *page) 1627 { 1628 if (PageAnon(page) || PageSlab(page) || page_has_type(page)) 1629 return -EINVAL; 1630 flush_dcache_page(page); 1631 return 0; 1632 } 1633 1634 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte, 1635 unsigned long addr, struct page *page, pgprot_t prot) 1636 { 1637 if (!pte_none(*pte)) 1638 return -EBUSY; 1639 /* Ok, finally just insert the thing.. */ 1640 get_page(page); 1641 inc_mm_counter_fast(mm, mm_counter_file(page)); 1642 page_add_file_rmap(page, false); 1643 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1644 return 0; 1645 } 1646 1647 /* 1648 * This is the old fallback for page remapping. 1649 * 1650 * For historical reasons, it only allows reserved pages. Only 1651 * old drivers should use this, and they needed to mark their 1652 * pages reserved for the old functions anyway. 1653 */ 1654 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1655 struct page *page, pgprot_t prot) 1656 { 1657 struct mm_struct *mm = vma->vm_mm; 1658 int retval; 1659 pte_t *pte; 1660 spinlock_t *ptl; 1661 1662 retval = validate_page_before_insert(page); 1663 if (retval) 1664 goto out; 1665 retval = -ENOMEM; 1666 pte = get_locked_pte(mm, addr, &ptl); 1667 if (!pte) 1668 goto out; 1669 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot); 1670 pte_unmap_unlock(pte, ptl); 1671 out: 1672 return retval; 1673 } 1674 1675 #ifdef pte_index 1676 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte, 1677 unsigned long addr, struct page *page, pgprot_t prot) 1678 { 1679 int err; 1680 1681 if (!page_count(page)) 1682 return -EINVAL; 1683 err = validate_page_before_insert(page); 1684 if (err) 1685 return err; 1686 return insert_page_into_pte_locked(mm, pte, addr, page, prot); 1687 } 1688 1689 /* insert_pages() amortizes the cost of spinlock operations 1690 * when inserting pages in a loop. Arch *must* define pte_index. 1691 */ 1692 static int insert_pages(struct vm_area_struct *vma, unsigned long addr, 1693 struct page **pages, unsigned long *num, pgprot_t prot) 1694 { 1695 pmd_t *pmd = NULL; 1696 pte_t *start_pte, *pte; 1697 spinlock_t *pte_lock; 1698 struct mm_struct *const mm = vma->vm_mm; 1699 unsigned long curr_page_idx = 0; 1700 unsigned long remaining_pages_total = *num; 1701 unsigned long pages_to_write_in_pmd; 1702 int ret; 1703 more: 1704 ret = -EFAULT; 1705 pmd = walk_to_pmd(mm, addr); 1706 if (!pmd) 1707 goto out; 1708 1709 pages_to_write_in_pmd = min_t(unsigned long, 1710 remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); 1711 1712 /* Allocate the PTE if necessary; takes PMD lock once only. */ 1713 ret = -ENOMEM; 1714 if (pte_alloc(mm, pmd)) 1715 goto out; 1716 1717 while (pages_to_write_in_pmd) { 1718 int pte_idx = 0; 1719 const int batch_size = min_t(int, pages_to_write_in_pmd, 8); 1720 1721 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); 1722 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { 1723 int err = insert_page_in_batch_locked(mm, pte, 1724 addr, pages[curr_page_idx], prot); 1725 if (unlikely(err)) { 1726 pte_unmap_unlock(start_pte, pte_lock); 1727 ret = err; 1728 remaining_pages_total -= pte_idx; 1729 goto out; 1730 } 1731 addr += PAGE_SIZE; 1732 ++curr_page_idx; 1733 } 1734 pte_unmap_unlock(start_pte, pte_lock); 1735 pages_to_write_in_pmd -= batch_size; 1736 remaining_pages_total -= batch_size; 1737 } 1738 if (remaining_pages_total) 1739 goto more; 1740 ret = 0; 1741 out: 1742 *num = remaining_pages_total; 1743 return ret; 1744 } 1745 #endif /* ifdef pte_index */ 1746 1747 /** 1748 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. 1749 * @vma: user vma to map to 1750 * @addr: target start user address of these pages 1751 * @pages: source kernel pages 1752 * @num: in: number of pages to map. out: number of pages that were *not* 1753 * mapped. (0 means all pages were successfully mapped). 1754 * 1755 * Preferred over vm_insert_page() when inserting multiple pages. 1756 * 1757 * In case of error, we may have mapped a subset of the provided 1758 * pages. It is the caller's responsibility to account for this case. 1759 * 1760 * The same restrictions apply as in vm_insert_page(). 1761 */ 1762 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, 1763 struct page **pages, unsigned long *num) 1764 { 1765 #ifdef pte_index 1766 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; 1767 1768 if (addr < vma->vm_start || end_addr >= vma->vm_end) 1769 return -EFAULT; 1770 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1771 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1772 BUG_ON(vma->vm_flags & VM_PFNMAP); 1773 vma->vm_flags |= VM_MIXEDMAP; 1774 } 1775 /* Defer page refcount checking till we're about to map that page. */ 1776 return insert_pages(vma, addr, pages, num, vma->vm_page_prot); 1777 #else 1778 unsigned long idx = 0, pgcount = *num; 1779 int err = -EINVAL; 1780 1781 for (; idx < pgcount; ++idx) { 1782 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]); 1783 if (err) 1784 break; 1785 } 1786 *num = pgcount - idx; 1787 return err; 1788 #endif /* ifdef pte_index */ 1789 } 1790 EXPORT_SYMBOL(vm_insert_pages); 1791 1792 /** 1793 * vm_insert_page - insert single page into user vma 1794 * @vma: user vma to map to 1795 * @addr: target user address of this page 1796 * @page: source kernel page 1797 * 1798 * This allows drivers to insert individual pages they've allocated 1799 * into a user vma. 1800 * 1801 * The page has to be a nice clean _individual_ kernel allocation. 1802 * If you allocate a compound page, you need to have marked it as 1803 * such (__GFP_COMP), or manually just split the page up yourself 1804 * (see split_page()). 1805 * 1806 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1807 * took an arbitrary page protection parameter. This doesn't allow 1808 * that. Your vma protection will have to be set up correctly, which 1809 * means that if you want a shared writable mapping, you'd better 1810 * ask for a shared writable mapping! 1811 * 1812 * The page does not need to be reserved. 1813 * 1814 * Usually this function is called from f_op->mmap() handler 1815 * under mm->mmap_lock write-lock, so it can change vma->vm_flags. 1816 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1817 * function from other places, for example from page-fault handler. 1818 * 1819 * Return: %0 on success, negative error code otherwise. 1820 */ 1821 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1822 struct page *page) 1823 { 1824 if (addr < vma->vm_start || addr >= vma->vm_end) 1825 return -EFAULT; 1826 if (!page_count(page)) 1827 return -EINVAL; 1828 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1829 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1830 BUG_ON(vma->vm_flags & VM_PFNMAP); 1831 vma->vm_flags |= VM_MIXEDMAP; 1832 } 1833 return insert_page(vma, addr, page, vma->vm_page_prot); 1834 } 1835 EXPORT_SYMBOL(vm_insert_page); 1836 1837 /* 1838 * __vm_map_pages - maps range of kernel pages into user vma 1839 * @vma: user vma to map to 1840 * @pages: pointer to array of source kernel pages 1841 * @num: number of pages in page array 1842 * @offset: user's requested vm_pgoff 1843 * 1844 * This allows drivers to map range of kernel pages into a user vma. 1845 * 1846 * Return: 0 on success and error code otherwise. 1847 */ 1848 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1849 unsigned long num, unsigned long offset) 1850 { 1851 unsigned long count = vma_pages(vma); 1852 unsigned long uaddr = vma->vm_start; 1853 int ret, i; 1854 1855 /* Fail if the user requested offset is beyond the end of the object */ 1856 if (offset >= num) 1857 return -ENXIO; 1858 1859 /* Fail if the user requested size exceeds available object size */ 1860 if (count > num - offset) 1861 return -ENXIO; 1862 1863 for (i = 0; i < count; i++) { 1864 ret = vm_insert_page(vma, uaddr, pages[offset + i]); 1865 if (ret < 0) 1866 return ret; 1867 uaddr += PAGE_SIZE; 1868 } 1869 1870 return 0; 1871 } 1872 1873 /** 1874 * vm_map_pages - maps range of kernel pages starts with non zero offset 1875 * @vma: user vma to map to 1876 * @pages: pointer to array of source kernel pages 1877 * @num: number of pages in page array 1878 * 1879 * Maps an object consisting of @num pages, catering for the user's 1880 * requested vm_pgoff 1881 * 1882 * If we fail to insert any page into the vma, the function will return 1883 * immediately leaving any previously inserted pages present. Callers 1884 * from the mmap handler may immediately return the error as their caller 1885 * will destroy the vma, removing any successfully inserted pages. Other 1886 * callers should make their own arrangements for calling unmap_region(). 1887 * 1888 * Context: Process context. Called by mmap handlers. 1889 * Return: 0 on success and error code otherwise. 1890 */ 1891 int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1892 unsigned long num) 1893 { 1894 return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 1895 } 1896 EXPORT_SYMBOL(vm_map_pages); 1897 1898 /** 1899 * vm_map_pages_zero - map range of kernel pages starts with zero offset 1900 * @vma: user vma to map to 1901 * @pages: pointer to array of source kernel pages 1902 * @num: number of pages in page array 1903 * 1904 * Similar to vm_map_pages(), except that it explicitly sets the offset 1905 * to 0. This function is intended for the drivers that did not consider 1906 * vm_pgoff. 1907 * 1908 * Context: Process context. Called by mmap handlers. 1909 * Return: 0 on success and error code otherwise. 1910 */ 1911 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 1912 unsigned long num) 1913 { 1914 return __vm_map_pages(vma, pages, num, 0); 1915 } 1916 EXPORT_SYMBOL(vm_map_pages_zero); 1917 1918 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1919 pfn_t pfn, pgprot_t prot, bool mkwrite) 1920 { 1921 struct mm_struct *mm = vma->vm_mm; 1922 pte_t *pte, entry; 1923 spinlock_t *ptl; 1924 1925 pte = get_locked_pte(mm, addr, &ptl); 1926 if (!pte) 1927 return VM_FAULT_OOM; 1928 if (!pte_none(*pte)) { 1929 if (mkwrite) { 1930 /* 1931 * For read faults on private mappings the PFN passed 1932 * in may not match the PFN we have mapped if the 1933 * mapped PFN is a writeable COW page. In the mkwrite 1934 * case we are creating a writable PTE for a shared 1935 * mapping and we expect the PFNs to match. If they 1936 * don't match, we are likely racing with block 1937 * allocation and mapping invalidation so just skip the 1938 * update. 1939 */ 1940 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) { 1941 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte))); 1942 goto out_unlock; 1943 } 1944 entry = pte_mkyoung(*pte); 1945 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1946 if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 1947 update_mmu_cache(vma, addr, pte); 1948 } 1949 goto out_unlock; 1950 } 1951 1952 /* Ok, finally just insert the thing.. */ 1953 if (pfn_t_devmap(pfn)) 1954 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 1955 else 1956 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 1957 1958 if (mkwrite) { 1959 entry = pte_mkyoung(entry); 1960 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1961 } 1962 1963 set_pte_at(mm, addr, pte, entry); 1964 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 1965 1966 out_unlock: 1967 pte_unmap_unlock(pte, ptl); 1968 return VM_FAULT_NOPAGE; 1969 } 1970 1971 /** 1972 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 1973 * @vma: user vma to map to 1974 * @addr: target user address of this page 1975 * @pfn: source kernel pfn 1976 * @pgprot: pgprot flags for the inserted page 1977 * 1978 * This is exactly like vmf_insert_pfn(), except that it allows drivers 1979 * to override pgprot on a per-page basis. 1980 * 1981 * This only makes sense for IO mappings, and it makes no sense for 1982 * COW mappings. In general, using multiple vmas is preferable; 1983 * vmf_insert_pfn_prot should only be used if using multiple VMAs is 1984 * impractical. 1985 * 1986 * See vmf_insert_mixed_prot() for a discussion of the implication of using 1987 * a value of @pgprot different from that of @vma->vm_page_prot. 1988 * 1989 * Context: Process context. May allocate using %GFP_KERNEL. 1990 * Return: vm_fault_t value. 1991 */ 1992 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 1993 unsigned long pfn, pgprot_t pgprot) 1994 { 1995 /* 1996 * Technically, architectures with pte_special can avoid all these 1997 * restrictions (same for remap_pfn_range). However we would like 1998 * consistency in testing and feature parity among all, so we should 1999 * try to keep these invariants in place for everybody. 2000 */ 2001 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 2002 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 2003 (VM_PFNMAP|VM_MIXEDMAP)); 2004 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 2005 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 2006 2007 if (addr < vma->vm_start || addr >= vma->vm_end) 2008 return VM_FAULT_SIGBUS; 2009 2010 if (!pfn_modify_allowed(pfn, pgprot)) 2011 return VM_FAULT_SIGBUS; 2012 2013 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 2014 2015 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 2016 false); 2017 } 2018 EXPORT_SYMBOL(vmf_insert_pfn_prot); 2019 2020 /** 2021 * vmf_insert_pfn - insert single pfn into user vma 2022 * @vma: user vma to map to 2023 * @addr: target user address of this page 2024 * @pfn: source kernel pfn 2025 * 2026 * Similar to vm_insert_page, this allows drivers to insert individual pages 2027 * they've allocated into a user vma. Same comments apply. 2028 * 2029 * This function should only be called from a vm_ops->fault handler, and 2030 * in that case the handler should return the result of this function. 2031 * 2032 * vma cannot be a COW mapping. 2033 * 2034 * As this is called only for pages that do not currently exist, we 2035 * do not need to flush old virtual caches or the TLB. 2036 * 2037 * Context: Process context. May allocate using %GFP_KERNEL. 2038 * Return: vm_fault_t value. 2039 */ 2040 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2041 unsigned long pfn) 2042 { 2043 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 2044 } 2045 EXPORT_SYMBOL(vmf_insert_pfn); 2046 2047 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 2048 { 2049 /* these checks mirror the abort conditions in vm_normal_page */ 2050 if (vma->vm_flags & VM_MIXEDMAP) 2051 return true; 2052 if (pfn_t_devmap(pfn)) 2053 return true; 2054 if (pfn_t_special(pfn)) 2055 return true; 2056 if (is_zero_pfn(pfn_t_to_pfn(pfn))) 2057 return true; 2058 return false; 2059 } 2060 2061 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 2062 unsigned long addr, pfn_t pfn, pgprot_t pgprot, 2063 bool mkwrite) 2064 { 2065 int err; 2066 2067 BUG_ON(!vm_mixed_ok(vma, pfn)); 2068 2069 if (addr < vma->vm_start || addr >= vma->vm_end) 2070 return VM_FAULT_SIGBUS; 2071 2072 track_pfn_insert(vma, &pgprot, pfn); 2073 2074 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 2075 return VM_FAULT_SIGBUS; 2076 2077 /* 2078 * If we don't have pte special, then we have to use the pfn_valid() 2079 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2080 * refcount the page if pfn_valid is true (hence insert_page rather 2081 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 2082 * without pte special, it would there be refcounted as a normal page. 2083 */ 2084 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 2085 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 2086 struct page *page; 2087 2088 /* 2089 * At this point we are committed to insert_page() 2090 * regardless of whether the caller specified flags that 2091 * result in pfn_t_has_page() == false. 2092 */ 2093 page = pfn_to_page(pfn_t_to_pfn(pfn)); 2094 err = insert_page(vma, addr, page, pgprot); 2095 } else { 2096 return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 2097 } 2098 2099 if (err == -ENOMEM) 2100 return VM_FAULT_OOM; 2101 if (err < 0 && err != -EBUSY) 2102 return VM_FAULT_SIGBUS; 2103 2104 return VM_FAULT_NOPAGE; 2105 } 2106 2107 /** 2108 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot 2109 * @vma: user vma to map to 2110 * @addr: target user address of this page 2111 * @pfn: source kernel pfn 2112 * @pgprot: pgprot flags for the inserted page 2113 * 2114 * This is exactly like vmf_insert_mixed(), except that it allows drivers 2115 * to override pgprot on a per-page basis. 2116 * 2117 * Typically this function should be used by drivers to set caching- and 2118 * encryption bits different than those of @vma->vm_page_prot, because 2119 * the caching- or encryption mode may not be known at mmap() time. 2120 * This is ok as long as @vma->vm_page_prot is not used by the core vm 2121 * to set caching and encryption bits for those vmas (except for COW pages). 2122 * This is ensured by core vm only modifying these page table entries using 2123 * functions that don't touch caching- or encryption bits, using pte_modify() 2124 * if needed. (See for example mprotect()). 2125 * Also when new page-table entries are created, this is only done using the 2126 * fault() callback, and never using the value of vma->vm_page_prot, 2127 * except for page-table entries that point to anonymous pages as the result 2128 * of COW. 2129 * 2130 * Context: Process context. May allocate using %GFP_KERNEL. 2131 * Return: vm_fault_t value. 2132 */ 2133 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr, 2134 pfn_t pfn, pgprot_t pgprot) 2135 { 2136 return __vm_insert_mixed(vma, addr, pfn, pgprot, false); 2137 } 2138 EXPORT_SYMBOL(vmf_insert_mixed_prot); 2139 2140 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 2141 pfn_t pfn) 2142 { 2143 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false); 2144 } 2145 EXPORT_SYMBOL(vmf_insert_mixed); 2146 2147 /* 2148 * If the insertion of PTE failed because someone else already added a 2149 * different entry in the mean time, we treat that as success as we assume 2150 * the same entry was actually inserted. 2151 */ 2152 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 2153 unsigned long addr, pfn_t pfn) 2154 { 2155 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true); 2156 } 2157 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 2158 2159 /* 2160 * maps a range of physical memory into the requested pages. the old 2161 * mappings are removed. any references to nonexistent pages results 2162 * in null mappings (currently treated as "copy-on-access") 2163 */ 2164 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 2165 unsigned long addr, unsigned long end, 2166 unsigned long pfn, pgprot_t prot) 2167 { 2168 pte_t *pte; 2169 spinlock_t *ptl; 2170 int err = 0; 2171 2172 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 2173 if (!pte) 2174 return -ENOMEM; 2175 arch_enter_lazy_mmu_mode(); 2176 do { 2177 BUG_ON(!pte_none(*pte)); 2178 if (!pfn_modify_allowed(pfn, prot)) { 2179 err = -EACCES; 2180 break; 2181 } 2182 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 2183 pfn++; 2184 } while (pte++, addr += PAGE_SIZE, addr != end); 2185 arch_leave_lazy_mmu_mode(); 2186 pte_unmap_unlock(pte - 1, ptl); 2187 return err; 2188 } 2189 2190 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 2191 unsigned long addr, unsigned long end, 2192 unsigned long pfn, pgprot_t prot) 2193 { 2194 pmd_t *pmd; 2195 unsigned long next; 2196 int err; 2197 2198 pfn -= addr >> PAGE_SHIFT; 2199 pmd = pmd_alloc(mm, pud, addr); 2200 if (!pmd) 2201 return -ENOMEM; 2202 VM_BUG_ON(pmd_trans_huge(*pmd)); 2203 do { 2204 next = pmd_addr_end(addr, end); 2205 err = remap_pte_range(mm, pmd, addr, next, 2206 pfn + (addr >> PAGE_SHIFT), prot); 2207 if (err) 2208 return err; 2209 } while (pmd++, addr = next, addr != end); 2210 return 0; 2211 } 2212 2213 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 2214 unsigned long addr, unsigned long end, 2215 unsigned long pfn, pgprot_t prot) 2216 { 2217 pud_t *pud; 2218 unsigned long next; 2219 int err; 2220 2221 pfn -= addr >> PAGE_SHIFT; 2222 pud = pud_alloc(mm, p4d, addr); 2223 if (!pud) 2224 return -ENOMEM; 2225 do { 2226 next = pud_addr_end(addr, end); 2227 err = remap_pmd_range(mm, pud, addr, next, 2228 pfn + (addr >> PAGE_SHIFT), prot); 2229 if (err) 2230 return err; 2231 } while (pud++, addr = next, addr != end); 2232 return 0; 2233 } 2234 2235 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2236 unsigned long addr, unsigned long end, 2237 unsigned long pfn, pgprot_t prot) 2238 { 2239 p4d_t *p4d; 2240 unsigned long next; 2241 int err; 2242 2243 pfn -= addr >> PAGE_SHIFT; 2244 p4d = p4d_alloc(mm, pgd, addr); 2245 if (!p4d) 2246 return -ENOMEM; 2247 do { 2248 next = p4d_addr_end(addr, end); 2249 err = remap_pud_range(mm, p4d, addr, next, 2250 pfn + (addr >> PAGE_SHIFT), prot); 2251 if (err) 2252 return err; 2253 } while (p4d++, addr = next, addr != end); 2254 return 0; 2255 } 2256 2257 /** 2258 * remap_pfn_range - remap kernel memory to userspace 2259 * @vma: user vma to map to 2260 * @addr: target page aligned user address to start at 2261 * @pfn: page frame number of kernel physical memory address 2262 * @size: size of mapping area 2263 * @prot: page protection flags for this mapping 2264 * 2265 * Note: this is only safe if the mm semaphore is held when called. 2266 * 2267 * Return: %0 on success, negative error code otherwise. 2268 */ 2269 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 2270 unsigned long pfn, unsigned long size, pgprot_t prot) 2271 { 2272 pgd_t *pgd; 2273 unsigned long next; 2274 unsigned long end = addr + PAGE_ALIGN(size); 2275 struct mm_struct *mm = vma->vm_mm; 2276 unsigned long remap_pfn = pfn; 2277 int err; 2278 2279 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) 2280 return -EINVAL; 2281 2282 /* 2283 * Physically remapped pages are special. Tell the 2284 * rest of the world about it: 2285 * VM_IO tells people not to look at these pages 2286 * (accesses can have side effects). 2287 * VM_PFNMAP tells the core MM that the base pages are just 2288 * raw PFN mappings, and do not have a "struct page" associated 2289 * with them. 2290 * VM_DONTEXPAND 2291 * Disable vma merging and expanding with mremap(). 2292 * VM_DONTDUMP 2293 * Omit vma from core dump, even when VM_IO turned off. 2294 * 2295 * There's a horrible special case to handle copy-on-write 2296 * behaviour that some programs depend on. We mark the "original" 2297 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2298 * See vm_normal_page() for details. 2299 */ 2300 if (is_cow_mapping(vma->vm_flags)) { 2301 if (addr != vma->vm_start || end != vma->vm_end) 2302 return -EINVAL; 2303 vma->vm_pgoff = pfn; 2304 } 2305 2306 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size)); 2307 if (err) 2308 return -EINVAL; 2309 2310 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 2311 2312 BUG_ON(addr >= end); 2313 pfn -= addr >> PAGE_SHIFT; 2314 pgd = pgd_offset(mm, addr); 2315 flush_cache_range(vma, addr, end); 2316 do { 2317 next = pgd_addr_end(addr, end); 2318 err = remap_p4d_range(mm, pgd, addr, next, 2319 pfn + (addr >> PAGE_SHIFT), prot); 2320 if (err) 2321 break; 2322 } while (pgd++, addr = next, addr != end); 2323 2324 if (err) 2325 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size)); 2326 2327 return err; 2328 } 2329 EXPORT_SYMBOL(remap_pfn_range); 2330 2331 /** 2332 * vm_iomap_memory - remap memory to userspace 2333 * @vma: user vma to map to 2334 * @start: start of the physical memory to be mapped 2335 * @len: size of area 2336 * 2337 * This is a simplified io_remap_pfn_range() for common driver use. The 2338 * driver just needs to give us the physical memory range to be mapped, 2339 * we'll figure out the rest from the vma information. 2340 * 2341 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 2342 * whatever write-combining details or similar. 2343 * 2344 * Return: %0 on success, negative error code otherwise. 2345 */ 2346 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 2347 { 2348 unsigned long vm_len, pfn, pages; 2349 2350 /* Check that the physical memory area passed in looks valid */ 2351 if (start + len < start) 2352 return -EINVAL; 2353 /* 2354 * You *really* shouldn't map things that aren't page-aligned, 2355 * but we've historically allowed it because IO memory might 2356 * just have smaller alignment. 2357 */ 2358 len += start & ~PAGE_MASK; 2359 pfn = start >> PAGE_SHIFT; 2360 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 2361 if (pfn + pages < pfn) 2362 return -EINVAL; 2363 2364 /* We start the mapping 'vm_pgoff' pages into the area */ 2365 if (vma->vm_pgoff > pages) 2366 return -EINVAL; 2367 pfn += vma->vm_pgoff; 2368 pages -= vma->vm_pgoff; 2369 2370 /* Can we fit all of the mapping? */ 2371 vm_len = vma->vm_end - vma->vm_start; 2372 if (vm_len >> PAGE_SHIFT > pages) 2373 return -EINVAL; 2374 2375 /* Ok, let it rip */ 2376 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 2377 } 2378 EXPORT_SYMBOL(vm_iomap_memory); 2379 2380 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2381 unsigned long addr, unsigned long end, 2382 pte_fn_t fn, void *data, bool create, 2383 pgtbl_mod_mask *mask) 2384 { 2385 pte_t *pte; 2386 int err = 0; 2387 spinlock_t *ptl; 2388 2389 if (create) { 2390 pte = (mm == &init_mm) ? 2391 pte_alloc_kernel_track(pmd, addr, mask) : 2392 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2393 if (!pte) 2394 return -ENOMEM; 2395 } else { 2396 pte = (mm == &init_mm) ? 2397 pte_offset_kernel(pmd, addr) : 2398 pte_offset_map_lock(mm, pmd, addr, &ptl); 2399 } 2400 2401 BUG_ON(pmd_huge(*pmd)); 2402 2403 arch_enter_lazy_mmu_mode(); 2404 2405 if (fn) { 2406 do { 2407 if (create || !pte_none(*pte)) { 2408 err = fn(pte++, addr, data); 2409 if (err) 2410 break; 2411 } 2412 } while (addr += PAGE_SIZE, addr != end); 2413 } 2414 *mask |= PGTBL_PTE_MODIFIED; 2415 2416 arch_leave_lazy_mmu_mode(); 2417 2418 if (mm != &init_mm) 2419 pte_unmap_unlock(pte-1, ptl); 2420 return err; 2421 } 2422 2423 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2424 unsigned long addr, unsigned long end, 2425 pte_fn_t fn, void *data, bool create, 2426 pgtbl_mod_mask *mask) 2427 { 2428 pmd_t *pmd; 2429 unsigned long next; 2430 int err = 0; 2431 2432 BUG_ON(pud_huge(*pud)); 2433 2434 if (create) { 2435 pmd = pmd_alloc_track(mm, pud, addr, mask); 2436 if (!pmd) 2437 return -ENOMEM; 2438 } else { 2439 pmd = pmd_offset(pud, addr); 2440 } 2441 do { 2442 next = pmd_addr_end(addr, end); 2443 if (create || !pmd_none_or_clear_bad(pmd)) { 2444 err = apply_to_pte_range(mm, pmd, addr, next, fn, data, 2445 create, mask); 2446 if (err) 2447 break; 2448 } 2449 } while (pmd++, addr = next, addr != end); 2450 return err; 2451 } 2452 2453 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2454 unsigned long addr, unsigned long end, 2455 pte_fn_t fn, void *data, bool create, 2456 pgtbl_mod_mask *mask) 2457 { 2458 pud_t *pud; 2459 unsigned long next; 2460 int err = 0; 2461 2462 if (create) { 2463 pud = pud_alloc_track(mm, p4d, addr, mask); 2464 if (!pud) 2465 return -ENOMEM; 2466 } else { 2467 pud = pud_offset(p4d, addr); 2468 } 2469 do { 2470 next = pud_addr_end(addr, end); 2471 if (create || !pud_none_or_clear_bad(pud)) { 2472 err = apply_to_pmd_range(mm, pud, addr, next, fn, data, 2473 create, mask); 2474 if (err) 2475 break; 2476 } 2477 } while (pud++, addr = next, addr != end); 2478 return err; 2479 } 2480 2481 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2482 unsigned long addr, unsigned long end, 2483 pte_fn_t fn, void *data, bool create, 2484 pgtbl_mod_mask *mask) 2485 { 2486 p4d_t *p4d; 2487 unsigned long next; 2488 int err = 0; 2489 2490 if (create) { 2491 p4d = p4d_alloc_track(mm, pgd, addr, mask); 2492 if (!p4d) 2493 return -ENOMEM; 2494 } else { 2495 p4d = p4d_offset(pgd, addr); 2496 } 2497 do { 2498 next = p4d_addr_end(addr, end); 2499 if (create || !p4d_none_or_clear_bad(p4d)) { 2500 err = apply_to_pud_range(mm, p4d, addr, next, fn, data, 2501 create, mask); 2502 if (err) 2503 break; 2504 } 2505 } while (p4d++, addr = next, addr != end); 2506 return err; 2507 } 2508 2509 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2510 unsigned long size, pte_fn_t fn, 2511 void *data, bool create) 2512 { 2513 pgd_t *pgd; 2514 unsigned long start = addr, next; 2515 unsigned long end = addr + size; 2516 pgtbl_mod_mask mask = 0; 2517 int err = 0; 2518 2519 if (WARN_ON(addr >= end)) 2520 return -EINVAL; 2521 2522 pgd = pgd_offset(mm, addr); 2523 do { 2524 next = pgd_addr_end(addr, end); 2525 if (!create && pgd_none_or_clear_bad(pgd)) 2526 continue; 2527 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask); 2528 if (err) 2529 break; 2530 } while (pgd++, addr = next, addr != end); 2531 2532 if (mask & ARCH_PAGE_TABLE_SYNC_MASK) 2533 arch_sync_kernel_mappings(start, start + size); 2534 2535 return err; 2536 } 2537 2538 /* 2539 * Scan a region of virtual memory, filling in page tables as necessary 2540 * and calling a provided function on each leaf page table. 2541 */ 2542 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2543 unsigned long size, pte_fn_t fn, void *data) 2544 { 2545 return __apply_to_page_range(mm, addr, size, fn, data, true); 2546 } 2547 EXPORT_SYMBOL_GPL(apply_to_page_range); 2548 2549 /* 2550 * Scan a region of virtual memory, calling a provided function on 2551 * each leaf page table where it exists. 2552 * 2553 * Unlike apply_to_page_range, this does _not_ fill in page tables 2554 * where they are absent. 2555 */ 2556 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, 2557 unsigned long size, pte_fn_t fn, void *data) 2558 { 2559 return __apply_to_page_range(mm, addr, size, fn, data, false); 2560 } 2561 EXPORT_SYMBOL_GPL(apply_to_existing_page_range); 2562 2563 /* 2564 * handle_pte_fault chooses page fault handler according to an entry which was 2565 * read non-atomically. Before making any commitment, on those architectures 2566 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 2567 * parts, do_swap_page must check under lock before unmapping the pte and 2568 * proceeding (but do_wp_page is only called after already making such a check; 2569 * and do_anonymous_page can safely check later on). 2570 */ 2571 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 2572 pte_t *page_table, pte_t orig_pte) 2573 { 2574 int same = 1; 2575 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) 2576 if (sizeof(pte_t) > sizeof(unsigned long)) { 2577 spinlock_t *ptl = pte_lockptr(mm, pmd); 2578 spin_lock(ptl); 2579 same = pte_same(*page_table, orig_pte); 2580 spin_unlock(ptl); 2581 } 2582 #endif 2583 pte_unmap(page_table); 2584 return same; 2585 } 2586 2587 static inline bool cow_user_page(struct page *dst, struct page *src, 2588 struct vm_fault *vmf) 2589 { 2590 bool ret; 2591 void *kaddr; 2592 void __user *uaddr; 2593 bool locked = false; 2594 struct vm_area_struct *vma = vmf->vma; 2595 struct mm_struct *mm = vma->vm_mm; 2596 unsigned long addr = vmf->address; 2597 2598 if (likely(src)) { 2599 copy_user_highpage(dst, src, addr, vma); 2600 return true; 2601 } 2602 2603 /* 2604 * If the source page was a PFN mapping, we don't have 2605 * a "struct page" for it. We do a best-effort copy by 2606 * just copying from the original user address. If that 2607 * fails, we just zero-fill it. Live with it. 2608 */ 2609 kaddr = kmap_atomic(dst); 2610 uaddr = (void __user *)(addr & PAGE_MASK); 2611 2612 /* 2613 * On architectures with software "accessed" bits, we would 2614 * take a double page fault, so mark it accessed here. 2615 */ 2616 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) { 2617 pte_t entry; 2618 2619 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2620 locked = true; 2621 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2622 /* 2623 * Other thread has already handled the fault 2624 * and update local tlb only 2625 */ 2626 update_mmu_tlb(vma, addr, vmf->pte); 2627 ret = false; 2628 goto pte_unlock; 2629 } 2630 2631 entry = pte_mkyoung(vmf->orig_pte); 2632 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) 2633 update_mmu_cache(vma, addr, vmf->pte); 2634 } 2635 2636 /* 2637 * This really shouldn't fail, because the page is there 2638 * in the page tables. But it might just be unreadable, 2639 * in which case we just give up and fill the result with 2640 * zeroes. 2641 */ 2642 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2643 if (locked) 2644 goto warn; 2645 2646 /* Re-validate under PTL if the page is still mapped */ 2647 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2648 locked = true; 2649 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2650 /* The PTE changed under us, update local tlb */ 2651 update_mmu_tlb(vma, addr, vmf->pte); 2652 ret = false; 2653 goto pte_unlock; 2654 } 2655 2656 /* 2657 * The same page can be mapped back since last copy attempt. 2658 * Try to copy again under PTL. 2659 */ 2660 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2661 /* 2662 * Give a warn in case there can be some obscure 2663 * use-case 2664 */ 2665 warn: 2666 WARN_ON_ONCE(1); 2667 clear_page(kaddr); 2668 } 2669 } 2670 2671 ret = true; 2672 2673 pte_unlock: 2674 if (locked) 2675 pte_unmap_unlock(vmf->pte, vmf->ptl); 2676 kunmap_atomic(kaddr); 2677 flush_dcache_page(dst); 2678 2679 return ret; 2680 } 2681 2682 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 2683 { 2684 struct file *vm_file = vma->vm_file; 2685 2686 if (vm_file) 2687 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 2688 2689 /* 2690 * Special mappings (e.g. VDSO) do not have any file so fake 2691 * a default GFP_KERNEL for them. 2692 */ 2693 return GFP_KERNEL; 2694 } 2695 2696 /* 2697 * Notify the address space that the page is about to become writable so that 2698 * it can prohibit this or wait for the page to get into an appropriate state. 2699 * 2700 * We do this without the lock held, so that it can sleep if it needs to. 2701 */ 2702 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf) 2703 { 2704 vm_fault_t ret; 2705 struct page *page = vmf->page; 2706 unsigned int old_flags = vmf->flags; 2707 2708 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2709 2710 if (vmf->vma->vm_file && 2711 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) 2712 return VM_FAULT_SIGBUS; 2713 2714 ret = vmf->vma->vm_ops->page_mkwrite(vmf); 2715 /* Restore original flags so that caller is not surprised */ 2716 vmf->flags = old_flags; 2717 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2718 return ret; 2719 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2720 lock_page(page); 2721 if (!page->mapping) { 2722 unlock_page(page); 2723 return 0; /* retry */ 2724 } 2725 ret |= VM_FAULT_LOCKED; 2726 } else 2727 VM_BUG_ON_PAGE(!PageLocked(page), page); 2728 return ret; 2729 } 2730 2731 /* 2732 * Handle dirtying of a page in shared file mapping on a write fault. 2733 * 2734 * The function expects the page to be locked and unlocks it. 2735 */ 2736 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) 2737 { 2738 struct vm_area_struct *vma = vmf->vma; 2739 struct address_space *mapping; 2740 struct page *page = vmf->page; 2741 bool dirtied; 2742 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 2743 2744 dirtied = set_page_dirty(page); 2745 VM_BUG_ON_PAGE(PageAnon(page), page); 2746 /* 2747 * Take a local copy of the address_space - page.mapping may be zeroed 2748 * by truncate after unlock_page(). The address_space itself remains 2749 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 2750 * release semantics to prevent the compiler from undoing this copying. 2751 */ 2752 mapping = page_rmapping(page); 2753 unlock_page(page); 2754 2755 if (!page_mkwrite) 2756 file_update_time(vma->vm_file); 2757 2758 /* 2759 * Throttle page dirtying rate down to writeback speed. 2760 * 2761 * mapping may be NULL here because some device drivers do not 2762 * set page.mapping but still dirty their pages 2763 * 2764 * Drop the mmap_lock before waiting on IO, if we can. The file 2765 * is pinning the mapping, as per above. 2766 */ 2767 if ((dirtied || page_mkwrite) && mapping) { 2768 struct file *fpin; 2769 2770 fpin = maybe_unlock_mmap_for_io(vmf, NULL); 2771 balance_dirty_pages_ratelimited(mapping); 2772 if (fpin) { 2773 fput(fpin); 2774 return VM_FAULT_RETRY; 2775 } 2776 } 2777 2778 return 0; 2779 } 2780 2781 /* 2782 * Handle write page faults for pages that can be reused in the current vma 2783 * 2784 * This can happen either due to the mapping being with the VM_SHARED flag, 2785 * or due to us being the last reference standing to the page. In either 2786 * case, all we need to do here is to mark the page as writable and update 2787 * any related book-keeping. 2788 */ 2789 static inline void wp_page_reuse(struct vm_fault *vmf) 2790 __releases(vmf->ptl) 2791 { 2792 struct vm_area_struct *vma = vmf->vma; 2793 struct page *page = vmf->page; 2794 pte_t entry; 2795 /* 2796 * Clear the pages cpupid information as the existing 2797 * information potentially belongs to a now completely 2798 * unrelated process. 2799 */ 2800 if (page) 2801 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2802 2803 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2804 entry = pte_mkyoung(vmf->orig_pte); 2805 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2806 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 2807 update_mmu_cache(vma, vmf->address, vmf->pte); 2808 pte_unmap_unlock(vmf->pte, vmf->ptl); 2809 count_vm_event(PGREUSE); 2810 } 2811 2812 /* 2813 * Handle the case of a page which we actually need to copy to a new page. 2814 * 2815 * Called with mmap_lock locked and the old page referenced, but 2816 * without the ptl held. 2817 * 2818 * High level logic flow: 2819 * 2820 * - Allocate a page, copy the content of the old page to the new one. 2821 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2822 * - Take the PTL. If the pte changed, bail out and release the allocated page 2823 * - If the pte is still the way we remember it, update the page table and all 2824 * relevant references. This includes dropping the reference the page-table 2825 * held to the old page, as well as updating the rmap. 2826 * - In any case, unlock the PTL and drop the reference we took to the old page. 2827 */ 2828 static vm_fault_t wp_page_copy(struct vm_fault *vmf) 2829 { 2830 struct vm_area_struct *vma = vmf->vma; 2831 struct mm_struct *mm = vma->vm_mm; 2832 struct page *old_page = vmf->page; 2833 struct page *new_page = NULL; 2834 pte_t entry; 2835 int page_copied = 0; 2836 struct mmu_notifier_range range; 2837 2838 if (unlikely(anon_vma_prepare(vma))) 2839 goto oom; 2840 2841 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) { 2842 new_page = alloc_zeroed_user_highpage_movable(vma, 2843 vmf->address); 2844 if (!new_page) 2845 goto oom; 2846 } else { 2847 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 2848 vmf->address); 2849 if (!new_page) 2850 goto oom; 2851 2852 if (!cow_user_page(new_page, old_page, vmf)) { 2853 /* 2854 * COW failed, if the fault was solved by other, 2855 * it's fine. If not, userspace would re-fault on 2856 * the same address and we will handle the fault 2857 * from the second attempt. 2858 */ 2859 put_page(new_page); 2860 if (old_page) 2861 put_page(old_page); 2862 return 0; 2863 } 2864 } 2865 2866 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL)) 2867 goto oom_free_new; 2868 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 2869 2870 __SetPageUptodate(new_page); 2871 2872 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 2873 vmf->address & PAGE_MASK, 2874 (vmf->address & PAGE_MASK) + PAGE_SIZE); 2875 mmu_notifier_invalidate_range_start(&range); 2876 2877 /* 2878 * Re-check the pte - we dropped the lock 2879 */ 2880 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 2881 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2882 if (old_page) { 2883 if (!PageAnon(old_page)) { 2884 dec_mm_counter_fast(mm, 2885 mm_counter_file(old_page)); 2886 inc_mm_counter_fast(mm, MM_ANONPAGES); 2887 } 2888 } else { 2889 inc_mm_counter_fast(mm, MM_ANONPAGES); 2890 } 2891 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2892 entry = mk_pte(new_page, vma->vm_page_prot); 2893 entry = pte_sw_mkyoung(entry); 2894 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2895 /* 2896 * Clear the pte entry and flush it first, before updating the 2897 * pte with the new entry. This will avoid a race condition 2898 * seen in the presence of one thread doing SMC and another 2899 * thread doing COW. 2900 */ 2901 ptep_clear_flush_notify(vma, vmf->address, vmf->pte); 2902 page_add_new_anon_rmap(new_page, vma, vmf->address, false); 2903 lru_cache_add_inactive_or_unevictable(new_page, vma); 2904 /* 2905 * We call the notify macro here because, when using secondary 2906 * mmu page tables (such as kvm shadow page tables), we want the 2907 * new page to be mapped directly into the secondary page table. 2908 */ 2909 set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 2910 update_mmu_cache(vma, vmf->address, vmf->pte); 2911 if (old_page) { 2912 /* 2913 * Only after switching the pte to the new page may 2914 * we remove the mapcount here. Otherwise another 2915 * process may come and find the rmap count decremented 2916 * before the pte is switched to the new page, and 2917 * "reuse" the old page writing into it while our pte 2918 * here still points into it and can be read by other 2919 * threads. 2920 * 2921 * The critical issue is to order this 2922 * page_remove_rmap with the ptp_clear_flush above. 2923 * Those stores are ordered by (if nothing else,) 2924 * the barrier present in the atomic_add_negative 2925 * in page_remove_rmap. 2926 * 2927 * Then the TLB flush in ptep_clear_flush ensures that 2928 * no process can access the old page before the 2929 * decremented mapcount is visible. And the old page 2930 * cannot be reused until after the decremented 2931 * mapcount is visible. So transitively, TLBs to 2932 * old page will be flushed before it can be reused. 2933 */ 2934 page_remove_rmap(old_page, false); 2935 } 2936 2937 /* Free the old page.. */ 2938 new_page = old_page; 2939 page_copied = 1; 2940 } else { 2941 update_mmu_tlb(vma, vmf->address, vmf->pte); 2942 } 2943 2944 if (new_page) 2945 put_page(new_page); 2946 2947 pte_unmap_unlock(vmf->pte, vmf->ptl); 2948 /* 2949 * No need to double call mmu_notifier->invalidate_range() callback as 2950 * the above ptep_clear_flush_notify() did already call it. 2951 */ 2952 mmu_notifier_invalidate_range_only_end(&range); 2953 if (old_page) { 2954 /* 2955 * Don't let another task, with possibly unlocked vma, 2956 * keep the mlocked page. 2957 */ 2958 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 2959 lock_page(old_page); /* LRU manipulation */ 2960 if (PageMlocked(old_page)) 2961 munlock_vma_page(old_page); 2962 unlock_page(old_page); 2963 } 2964 put_page(old_page); 2965 } 2966 return page_copied ? VM_FAULT_WRITE : 0; 2967 oom_free_new: 2968 put_page(new_page); 2969 oom: 2970 if (old_page) 2971 put_page(old_page); 2972 return VM_FAULT_OOM; 2973 } 2974 2975 /** 2976 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 2977 * writeable once the page is prepared 2978 * 2979 * @vmf: structure describing the fault 2980 * 2981 * This function handles all that is needed to finish a write page fault in a 2982 * shared mapping due to PTE being read-only once the mapped page is prepared. 2983 * It handles locking of PTE and modifying it. 2984 * 2985 * The function expects the page to be locked or other protection against 2986 * concurrent faults / writeback (such as DAX radix tree locks). 2987 * 2988 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before 2989 * we acquired PTE lock. 2990 */ 2991 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf) 2992 { 2993 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 2994 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 2995 &vmf->ptl); 2996 /* 2997 * We might have raced with another page fault while we released the 2998 * pte_offset_map_lock. 2999 */ 3000 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 3001 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 3002 pte_unmap_unlock(vmf->pte, vmf->ptl); 3003 return VM_FAULT_NOPAGE; 3004 } 3005 wp_page_reuse(vmf); 3006 return 0; 3007 } 3008 3009 /* 3010 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 3011 * mapping 3012 */ 3013 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 3014 { 3015 struct vm_area_struct *vma = vmf->vma; 3016 3017 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 3018 vm_fault_t ret; 3019 3020 pte_unmap_unlock(vmf->pte, vmf->ptl); 3021 vmf->flags |= FAULT_FLAG_MKWRITE; 3022 ret = vma->vm_ops->pfn_mkwrite(vmf); 3023 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 3024 return ret; 3025 return finish_mkwrite_fault(vmf); 3026 } 3027 wp_page_reuse(vmf); 3028 return VM_FAULT_WRITE; 3029 } 3030 3031 static vm_fault_t wp_page_shared(struct vm_fault *vmf) 3032 __releases(vmf->ptl) 3033 { 3034 struct vm_area_struct *vma = vmf->vma; 3035 vm_fault_t ret = VM_FAULT_WRITE; 3036 3037 get_page(vmf->page); 3038 3039 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 3040 vm_fault_t tmp; 3041 3042 pte_unmap_unlock(vmf->pte, vmf->ptl); 3043 tmp = do_page_mkwrite(vmf); 3044 if (unlikely(!tmp || (tmp & 3045 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3046 put_page(vmf->page); 3047 return tmp; 3048 } 3049 tmp = finish_mkwrite_fault(vmf); 3050 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 3051 unlock_page(vmf->page); 3052 put_page(vmf->page); 3053 return tmp; 3054 } 3055 } else { 3056 wp_page_reuse(vmf); 3057 lock_page(vmf->page); 3058 } 3059 ret |= fault_dirty_shared_page(vmf); 3060 put_page(vmf->page); 3061 3062 return ret; 3063 } 3064 3065 /* 3066 * This routine handles present pages, when users try to write 3067 * to a shared page. It is done by copying the page to a new address 3068 * and decrementing the shared-page counter for the old page. 3069 * 3070 * Note that this routine assumes that the protection checks have been 3071 * done by the caller (the low-level page fault routine in most cases). 3072 * Thus we can safely just mark it writable once we've done any necessary 3073 * COW. 3074 * 3075 * We also mark the page dirty at this point even though the page will 3076 * change only once the write actually happens. This avoids a few races, 3077 * and potentially makes it more efficient. 3078 * 3079 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3080 * but allow concurrent faults), with pte both mapped and locked. 3081 * We return with mmap_lock still held, but pte unmapped and unlocked. 3082 */ 3083 static vm_fault_t do_wp_page(struct vm_fault *vmf) 3084 __releases(vmf->ptl) 3085 { 3086 struct vm_area_struct *vma = vmf->vma; 3087 3088 if (userfaultfd_pte_wp(vma, *vmf->pte)) { 3089 pte_unmap_unlock(vmf->pte, vmf->ptl); 3090 return handle_userfault(vmf, VM_UFFD_WP); 3091 } 3092 3093 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 3094 if (!vmf->page) { 3095 /* 3096 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 3097 * VM_PFNMAP VMA. 3098 * 3099 * We should not cow pages in a shared writeable mapping. 3100 * Just mark the pages writable and/or call ops->pfn_mkwrite. 3101 */ 3102 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3103 (VM_WRITE|VM_SHARED)) 3104 return wp_pfn_shared(vmf); 3105 3106 pte_unmap_unlock(vmf->pte, vmf->ptl); 3107 return wp_page_copy(vmf); 3108 } 3109 3110 /* 3111 * Take out anonymous pages first, anonymous shared vmas are 3112 * not dirty accountable. 3113 */ 3114 if (PageAnon(vmf->page)) { 3115 struct page *page = vmf->page; 3116 3117 /* PageKsm() doesn't necessarily raise the page refcount */ 3118 if (PageKsm(page) || page_count(page) != 1) 3119 goto copy; 3120 if (!trylock_page(page)) 3121 goto copy; 3122 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) { 3123 unlock_page(page); 3124 goto copy; 3125 } 3126 /* 3127 * Ok, we've got the only map reference, and the only 3128 * page count reference, and the page is locked, 3129 * it's dark out, and we're wearing sunglasses. Hit it. 3130 */ 3131 unlock_page(page); 3132 wp_page_reuse(vmf); 3133 return VM_FAULT_WRITE; 3134 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3135 (VM_WRITE|VM_SHARED))) { 3136 return wp_page_shared(vmf); 3137 } 3138 copy: 3139 /* 3140 * Ok, we need to copy. Oh, well.. 3141 */ 3142 get_page(vmf->page); 3143 3144 pte_unmap_unlock(vmf->pte, vmf->ptl); 3145 return wp_page_copy(vmf); 3146 } 3147 3148 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 3149 unsigned long start_addr, unsigned long end_addr, 3150 struct zap_details *details) 3151 { 3152 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 3153 } 3154 3155 static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 3156 struct zap_details *details) 3157 { 3158 struct vm_area_struct *vma; 3159 pgoff_t vba, vea, zba, zea; 3160 3161 vma_interval_tree_foreach(vma, root, 3162 details->first_index, details->last_index) { 3163 3164 vba = vma->vm_pgoff; 3165 vea = vba + vma_pages(vma) - 1; 3166 zba = details->first_index; 3167 if (zba < vba) 3168 zba = vba; 3169 zea = details->last_index; 3170 if (zea > vea) 3171 zea = vea; 3172 3173 unmap_mapping_range_vma(vma, 3174 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 3175 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 3176 details); 3177 } 3178 } 3179 3180 /** 3181 * unmap_mapping_pages() - Unmap pages from processes. 3182 * @mapping: The address space containing pages to be unmapped. 3183 * @start: Index of first page to be unmapped. 3184 * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 3185 * @even_cows: Whether to unmap even private COWed pages. 3186 * 3187 * Unmap the pages in this address space from any userspace process which 3188 * has them mmaped. Generally, you want to remove COWed pages as well when 3189 * a file is being truncated, but not when invalidating pages from the page 3190 * cache. 3191 */ 3192 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 3193 pgoff_t nr, bool even_cows) 3194 { 3195 struct zap_details details = { }; 3196 3197 details.check_mapping = even_cows ? NULL : mapping; 3198 details.first_index = start; 3199 details.last_index = start + nr - 1; 3200 if (details.last_index < details.first_index) 3201 details.last_index = ULONG_MAX; 3202 3203 i_mmap_lock_write(mapping); 3204 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3205 unmap_mapping_range_tree(&mapping->i_mmap, &details); 3206 i_mmap_unlock_write(mapping); 3207 } 3208 3209 /** 3210 * unmap_mapping_range - unmap the portion of all mmaps in the specified 3211 * address_space corresponding to the specified byte range in the underlying 3212 * file. 3213 * 3214 * @mapping: the address space containing mmaps to be unmapped. 3215 * @holebegin: byte in first page to unmap, relative to the start of 3216 * the underlying file. This will be rounded down to a PAGE_SIZE 3217 * boundary. Note that this is different from truncate_pagecache(), which 3218 * must keep the partial page. In contrast, we must get rid of 3219 * partial pages. 3220 * @holelen: size of prospective hole in bytes. This will be rounded 3221 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 3222 * end of the file. 3223 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 3224 * but 0 when invalidating pagecache, don't throw away private data. 3225 */ 3226 void unmap_mapping_range(struct address_space *mapping, 3227 loff_t const holebegin, loff_t const holelen, int even_cows) 3228 { 3229 pgoff_t hba = holebegin >> PAGE_SHIFT; 3230 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3231 3232 /* Check for overflow. */ 3233 if (sizeof(holelen) > sizeof(hlen)) { 3234 long long holeend = 3235 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3236 if (holeend & ~(long long)ULONG_MAX) 3237 hlen = ULONG_MAX - hba + 1; 3238 } 3239 3240 unmap_mapping_pages(mapping, hba, hlen, even_cows); 3241 } 3242 EXPORT_SYMBOL(unmap_mapping_range); 3243 3244 /* 3245 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3246 * but allow concurrent faults), and pte mapped but not yet locked. 3247 * We return with pte unmapped and unlocked. 3248 * 3249 * We return with the mmap_lock locked or unlocked in the same cases 3250 * as does filemap_fault(). 3251 */ 3252 vm_fault_t do_swap_page(struct vm_fault *vmf) 3253 { 3254 struct vm_area_struct *vma = vmf->vma; 3255 struct page *page = NULL, *swapcache; 3256 swp_entry_t entry; 3257 pte_t pte; 3258 int locked; 3259 int exclusive = 0; 3260 vm_fault_t ret = 0; 3261 void *shadow = NULL; 3262 3263 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) 3264 goto out; 3265 3266 entry = pte_to_swp_entry(vmf->orig_pte); 3267 if (unlikely(non_swap_entry(entry))) { 3268 if (is_migration_entry(entry)) { 3269 migration_entry_wait(vma->vm_mm, vmf->pmd, 3270 vmf->address); 3271 } else if (is_device_private_entry(entry)) { 3272 vmf->page = device_private_entry_to_page(entry); 3273 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 3274 } else if (is_hwpoison_entry(entry)) { 3275 ret = VM_FAULT_HWPOISON; 3276 } else { 3277 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 3278 ret = VM_FAULT_SIGBUS; 3279 } 3280 goto out; 3281 } 3282 3283 3284 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 3285 page = lookup_swap_cache(entry, vma, vmf->address); 3286 swapcache = page; 3287 3288 if (!page) { 3289 struct swap_info_struct *si = swp_swap_info(entry); 3290 3291 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && 3292 __swap_count(entry) == 1) { 3293 /* skip swapcache */ 3294 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 3295 vmf->address); 3296 if (page) { 3297 int err; 3298 3299 __SetPageLocked(page); 3300 __SetPageSwapBacked(page); 3301 set_page_private(page, entry.val); 3302 3303 /* Tell memcg to use swap ownership records */ 3304 SetPageSwapCache(page); 3305 err = mem_cgroup_charge(page, vma->vm_mm, 3306 GFP_KERNEL); 3307 ClearPageSwapCache(page); 3308 if (err) { 3309 ret = VM_FAULT_OOM; 3310 goto out_page; 3311 } 3312 3313 shadow = get_shadow_from_swap_cache(entry); 3314 if (shadow) 3315 workingset_refault(page, shadow); 3316 3317 lru_cache_add(page); 3318 swap_readpage(page, true); 3319 } 3320 } else { 3321 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 3322 vmf); 3323 swapcache = page; 3324 } 3325 3326 if (!page) { 3327 /* 3328 * Back out if somebody else faulted in this pte 3329 * while we released the pte lock. 3330 */ 3331 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3332 vmf->address, &vmf->ptl); 3333 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3334 ret = VM_FAULT_OOM; 3335 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3336 goto unlock; 3337 } 3338 3339 /* Had to read the page from swap area: Major fault */ 3340 ret = VM_FAULT_MAJOR; 3341 count_vm_event(PGMAJFAULT); 3342 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 3343 } else if (PageHWPoison(page)) { 3344 /* 3345 * hwpoisoned dirty swapcache pages are kept for killing 3346 * owner processes (which may be unknown at hwpoison time) 3347 */ 3348 ret = VM_FAULT_HWPOISON; 3349 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3350 goto out_release; 3351 } 3352 3353 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags); 3354 3355 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 3356 if (!locked) { 3357 ret |= VM_FAULT_RETRY; 3358 goto out_release; 3359 } 3360 3361 /* 3362 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 3363 * release the swapcache from under us. The page pin, and pte_same 3364 * test below, are not enough to exclude that. Even if it is still 3365 * swapcache, we need to check that the page's swap has not changed. 3366 */ 3367 if (unlikely((!PageSwapCache(page) || 3368 page_private(page) != entry.val)) && swapcache) 3369 goto out_page; 3370 3371 page = ksm_might_need_to_copy(page, vma, vmf->address); 3372 if (unlikely(!page)) { 3373 ret = VM_FAULT_OOM; 3374 page = swapcache; 3375 goto out_page; 3376 } 3377 3378 cgroup_throttle_swaprate(page, GFP_KERNEL); 3379 3380 /* 3381 * Back out if somebody else already faulted in this pte. 3382 */ 3383 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3384 &vmf->ptl); 3385 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) 3386 goto out_nomap; 3387 3388 if (unlikely(!PageUptodate(page))) { 3389 ret = VM_FAULT_SIGBUS; 3390 goto out_nomap; 3391 } 3392 3393 /* 3394 * The page isn't present yet, go ahead with the fault. 3395 * 3396 * Be careful about the sequence of operations here. 3397 * To get its accounting right, reuse_swap_page() must be called 3398 * while the page is counted on swap but not yet in mapcount i.e. 3399 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 3400 * must be called after the swap_free(), or it will never succeed. 3401 */ 3402 3403 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3404 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS); 3405 pte = mk_pte(page, vma->vm_page_prot); 3406 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) { 3407 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 3408 vmf->flags &= ~FAULT_FLAG_WRITE; 3409 ret |= VM_FAULT_WRITE; 3410 exclusive = RMAP_EXCLUSIVE; 3411 } 3412 flush_icache_page(vma, page); 3413 if (pte_swp_soft_dirty(vmf->orig_pte)) 3414 pte = pte_mksoft_dirty(pte); 3415 if (pte_swp_uffd_wp(vmf->orig_pte)) { 3416 pte = pte_mkuffd_wp(pte); 3417 pte = pte_wrprotect(pte); 3418 } 3419 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 3420 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 3421 vmf->orig_pte = pte; 3422 3423 /* ksm created a completely new copy */ 3424 if (unlikely(page != swapcache && swapcache)) { 3425 page_add_new_anon_rmap(page, vma, vmf->address, false); 3426 lru_cache_add_inactive_or_unevictable(page, vma); 3427 } else { 3428 do_page_add_anon_rmap(page, vma, vmf->address, exclusive); 3429 } 3430 3431 swap_free(entry); 3432 if (mem_cgroup_swap_full(page) || 3433 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 3434 try_to_free_swap(page); 3435 unlock_page(page); 3436 if (page != swapcache && swapcache) { 3437 /* 3438 * Hold the lock to avoid the swap entry to be reused 3439 * until we take the PT lock for the pte_same() check 3440 * (to avoid false positives from pte_same). For 3441 * further safety release the lock after the swap_free 3442 * so that the swap count won't change under a 3443 * parallel locked swapcache. 3444 */ 3445 unlock_page(swapcache); 3446 put_page(swapcache); 3447 } 3448 3449 if (vmf->flags & FAULT_FLAG_WRITE) { 3450 ret |= do_wp_page(vmf); 3451 if (ret & VM_FAULT_ERROR) 3452 ret &= VM_FAULT_ERROR; 3453 goto out; 3454 } 3455 3456 /* No need to invalidate - it was non-present before */ 3457 update_mmu_cache(vma, vmf->address, vmf->pte); 3458 unlock: 3459 pte_unmap_unlock(vmf->pte, vmf->ptl); 3460 out: 3461 return ret; 3462 out_nomap: 3463 pte_unmap_unlock(vmf->pte, vmf->ptl); 3464 out_page: 3465 unlock_page(page); 3466 out_release: 3467 put_page(page); 3468 if (page != swapcache && swapcache) { 3469 unlock_page(swapcache); 3470 put_page(swapcache); 3471 } 3472 return ret; 3473 } 3474 3475 /* 3476 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3477 * but allow concurrent faults), and pte mapped but not yet locked. 3478 * We return with mmap_lock still held, but pte unmapped and unlocked. 3479 */ 3480 static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 3481 { 3482 struct vm_area_struct *vma = vmf->vma; 3483 struct page *page; 3484 vm_fault_t ret = 0; 3485 pte_t entry; 3486 3487 /* File mapping without ->vm_ops ? */ 3488 if (vma->vm_flags & VM_SHARED) 3489 return VM_FAULT_SIGBUS; 3490 3491 /* 3492 * Use pte_alloc() instead of pte_alloc_map(). We can't run 3493 * pte_offset_map() on pmds where a huge pmd might be created 3494 * from a different thread. 3495 * 3496 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 3497 * parallel threads are excluded by other means. 3498 * 3499 * Here we only have mmap_read_lock(mm). 3500 */ 3501 if (pte_alloc(vma->vm_mm, vmf->pmd)) 3502 return VM_FAULT_OOM; 3503 3504 /* See the comment in pte_alloc_one_map() */ 3505 if (unlikely(pmd_trans_unstable(vmf->pmd))) 3506 return 0; 3507 3508 /* Use the zero-page for reads */ 3509 if (!(vmf->flags & FAULT_FLAG_WRITE) && 3510 !mm_forbids_zeropage(vma->vm_mm)) { 3511 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 3512 vma->vm_page_prot)); 3513 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3514 vmf->address, &vmf->ptl); 3515 if (!pte_none(*vmf->pte)) { 3516 update_mmu_tlb(vma, vmf->address, vmf->pte); 3517 goto unlock; 3518 } 3519 ret = check_stable_address_space(vma->vm_mm); 3520 if (ret) 3521 goto unlock; 3522 /* Deliver the page fault to userland, check inside PT lock */ 3523 if (userfaultfd_missing(vma)) { 3524 pte_unmap_unlock(vmf->pte, vmf->ptl); 3525 return handle_userfault(vmf, VM_UFFD_MISSING); 3526 } 3527 goto setpte; 3528 } 3529 3530 /* Allocate our own private page. */ 3531 if (unlikely(anon_vma_prepare(vma))) 3532 goto oom; 3533 page = alloc_zeroed_user_highpage_movable(vma, vmf->address); 3534 if (!page) 3535 goto oom; 3536 3537 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) 3538 goto oom_free_page; 3539 cgroup_throttle_swaprate(page, GFP_KERNEL); 3540 3541 /* 3542 * The memory barrier inside __SetPageUptodate makes sure that 3543 * preceding stores to the page contents become visible before 3544 * the set_pte_at() write. 3545 */ 3546 __SetPageUptodate(page); 3547 3548 entry = mk_pte(page, vma->vm_page_prot); 3549 entry = pte_sw_mkyoung(entry); 3550 if (vma->vm_flags & VM_WRITE) 3551 entry = pte_mkwrite(pte_mkdirty(entry)); 3552 3553 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3554 &vmf->ptl); 3555 if (!pte_none(*vmf->pte)) { 3556 update_mmu_cache(vma, vmf->address, vmf->pte); 3557 goto release; 3558 } 3559 3560 ret = check_stable_address_space(vma->vm_mm); 3561 if (ret) 3562 goto release; 3563 3564 /* Deliver the page fault to userland, check inside PT lock */ 3565 if (userfaultfd_missing(vma)) { 3566 pte_unmap_unlock(vmf->pte, vmf->ptl); 3567 put_page(page); 3568 return handle_userfault(vmf, VM_UFFD_MISSING); 3569 } 3570 3571 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3572 page_add_new_anon_rmap(page, vma, vmf->address, false); 3573 lru_cache_add_inactive_or_unevictable(page, vma); 3574 setpte: 3575 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3576 3577 /* No need to invalidate - it was non-present before */ 3578 update_mmu_cache(vma, vmf->address, vmf->pte); 3579 unlock: 3580 pte_unmap_unlock(vmf->pte, vmf->ptl); 3581 return ret; 3582 release: 3583 put_page(page); 3584 goto unlock; 3585 oom_free_page: 3586 put_page(page); 3587 oom: 3588 return VM_FAULT_OOM; 3589 } 3590 3591 /* 3592 * The mmap_lock must have been held on entry, and may have been 3593 * released depending on flags and vma->vm_ops->fault() return value. 3594 * See filemap_fault() and __lock_page_retry(). 3595 */ 3596 static vm_fault_t __do_fault(struct vm_fault *vmf) 3597 { 3598 struct vm_area_struct *vma = vmf->vma; 3599 vm_fault_t ret; 3600 3601 /* 3602 * Preallocate pte before we take page_lock because this might lead to 3603 * deadlocks for memcg reclaim which waits for pages under writeback: 3604 * lock_page(A) 3605 * SetPageWriteback(A) 3606 * unlock_page(A) 3607 * lock_page(B) 3608 * lock_page(B) 3609 * pte_alloc_one 3610 * shrink_page_list 3611 * wait_on_page_writeback(A) 3612 * SetPageWriteback(B) 3613 * unlock_page(B) 3614 * # flush A, B to clear the writeback 3615 */ 3616 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 3617 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3618 if (!vmf->prealloc_pte) 3619 return VM_FAULT_OOM; 3620 smp_wmb(); /* See comment in __pte_alloc() */ 3621 } 3622 3623 ret = vma->vm_ops->fault(vmf); 3624 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 3625 VM_FAULT_DONE_COW))) 3626 return ret; 3627 3628 if (unlikely(PageHWPoison(vmf->page))) { 3629 if (ret & VM_FAULT_LOCKED) 3630 unlock_page(vmf->page); 3631 put_page(vmf->page); 3632 vmf->page = NULL; 3633 return VM_FAULT_HWPOISON; 3634 } 3635 3636 if (unlikely(!(ret & VM_FAULT_LOCKED))) 3637 lock_page(vmf->page); 3638 else 3639 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page); 3640 3641 return ret; 3642 } 3643 3644 /* 3645 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set. 3646 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check 3647 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly 3648 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output. 3649 */ 3650 static int pmd_devmap_trans_unstable(pmd_t *pmd) 3651 { 3652 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd); 3653 } 3654 3655 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf) 3656 { 3657 struct vm_area_struct *vma = vmf->vma; 3658 3659 if (!pmd_none(*vmf->pmd)) 3660 goto map_pte; 3661 if (vmf->prealloc_pte) { 3662 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3663 if (unlikely(!pmd_none(*vmf->pmd))) { 3664 spin_unlock(vmf->ptl); 3665 goto map_pte; 3666 } 3667 3668 mm_inc_nr_ptes(vma->vm_mm); 3669 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3670 spin_unlock(vmf->ptl); 3671 vmf->prealloc_pte = NULL; 3672 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) { 3673 return VM_FAULT_OOM; 3674 } 3675 map_pte: 3676 /* 3677 * If a huge pmd materialized under us just retry later. Use 3678 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of 3679 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge 3680 * under us and then back to pmd_none, as a result of MADV_DONTNEED 3681 * running immediately after a huge pmd fault in a different thread of 3682 * this mm, in turn leading to a misleading pmd_trans_huge() retval. 3683 * All we have to ensure is that it is a regular pmd that we can walk 3684 * with pte_offset_map() and we can do that through an atomic read in 3685 * C, which is what pmd_trans_unstable() provides. 3686 */ 3687 if (pmd_devmap_trans_unstable(vmf->pmd)) 3688 return VM_FAULT_NOPAGE; 3689 3690 /* 3691 * At this point we know that our vmf->pmd points to a page of ptes 3692 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge() 3693 * for the duration of the fault. If a racing MADV_DONTNEED runs and 3694 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still 3695 * be valid and we will re-check to make sure the vmf->pte isn't 3696 * pte_none() under vmf->ptl protection when we return to 3697 * alloc_set_pte(). 3698 */ 3699 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3700 &vmf->ptl); 3701 return 0; 3702 } 3703 3704 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3705 static void deposit_prealloc_pte(struct vm_fault *vmf) 3706 { 3707 struct vm_area_struct *vma = vmf->vma; 3708 3709 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3710 /* 3711 * We are going to consume the prealloc table, 3712 * count that as nr_ptes. 3713 */ 3714 mm_inc_nr_ptes(vma->vm_mm); 3715 vmf->prealloc_pte = NULL; 3716 } 3717 3718 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3719 { 3720 struct vm_area_struct *vma = vmf->vma; 3721 bool write = vmf->flags & FAULT_FLAG_WRITE; 3722 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 3723 pmd_t entry; 3724 int i; 3725 vm_fault_t ret = VM_FAULT_FALLBACK; 3726 3727 if (!transhuge_vma_suitable(vma, haddr)) 3728 return ret; 3729 3730 page = compound_head(page); 3731 if (compound_order(page) != HPAGE_PMD_ORDER) 3732 return ret; 3733 3734 /* 3735 * Archs like ppc64 need additonal space to store information 3736 * related to pte entry. Use the preallocated table for that. 3737 */ 3738 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 3739 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3740 if (!vmf->prealloc_pte) 3741 return VM_FAULT_OOM; 3742 smp_wmb(); /* See comment in __pte_alloc() */ 3743 } 3744 3745 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3746 if (unlikely(!pmd_none(*vmf->pmd))) 3747 goto out; 3748 3749 for (i = 0; i < HPAGE_PMD_NR; i++) 3750 flush_icache_page(vma, page + i); 3751 3752 entry = mk_huge_pmd(page, vma->vm_page_prot); 3753 if (write) 3754 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 3755 3756 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR); 3757 page_add_file_rmap(page, true); 3758 /* 3759 * deposit and withdraw with pmd lock held 3760 */ 3761 if (arch_needs_pgtable_deposit()) 3762 deposit_prealloc_pte(vmf); 3763 3764 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 3765 3766 update_mmu_cache_pmd(vma, haddr, vmf->pmd); 3767 3768 /* fault is handled */ 3769 ret = 0; 3770 count_vm_event(THP_FILE_MAPPED); 3771 out: 3772 spin_unlock(vmf->ptl); 3773 return ret; 3774 } 3775 #else 3776 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3777 { 3778 BUILD_BUG(); 3779 return 0; 3780 } 3781 #endif 3782 3783 /** 3784 * alloc_set_pte - setup new PTE entry for given page and add reverse page 3785 * mapping. If needed, the function allocates page table or use pre-allocated. 3786 * 3787 * @vmf: fault environment 3788 * @page: page to map 3789 * 3790 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on 3791 * return. 3792 * 3793 * Target users are page handler itself and implementations of 3794 * vm_ops->map_pages. 3795 * 3796 * Return: %0 on success, %VM_FAULT_ code in case of error. 3797 */ 3798 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page) 3799 { 3800 struct vm_area_struct *vma = vmf->vma; 3801 bool write = vmf->flags & FAULT_FLAG_WRITE; 3802 pte_t entry; 3803 vm_fault_t ret; 3804 3805 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) { 3806 ret = do_set_pmd(vmf, page); 3807 if (ret != VM_FAULT_FALLBACK) 3808 return ret; 3809 } 3810 3811 if (!vmf->pte) { 3812 ret = pte_alloc_one_map(vmf); 3813 if (ret) 3814 return ret; 3815 } 3816 3817 /* Re-check under ptl */ 3818 if (unlikely(!pte_none(*vmf->pte))) { 3819 update_mmu_tlb(vma, vmf->address, vmf->pte); 3820 return VM_FAULT_NOPAGE; 3821 } 3822 3823 flush_icache_page(vma, page); 3824 entry = mk_pte(page, vma->vm_page_prot); 3825 entry = pte_sw_mkyoung(entry); 3826 if (write) 3827 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3828 /* copy-on-write page */ 3829 if (write && !(vma->vm_flags & VM_SHARED)) { 3830 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3831 page_add_new_anon_rmap(page, vma, vmf->address, false); 3832 lru_cache_add_inactive_or_unevictable(page, vma); 3833 } else { 3834 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 3835 page_add_file_rmap(page, false); 3836 } 3837 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3838 3839 /* no need to invalidate: a not-present page won't be cached */ 3840 update_mmu_cache(vma, vmf->address, vmf->pte); 3841 3842 return 0; 3843 } 3844 3845 3846 /** 3847 * finish_fault - finish page fault once we have prepared the page to fault 3848 * 3849 * @vmf: structure describing the fault 3850 * 3851 * This function handles all that is needed to finish a page fault once the 3852 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 3853 * given page, adds reverse page mapping, handles memcg charges and LRU 3854 * addition. 3855 * 3856 * The function expects the page to be locked and on success it consumes a 3857 * reference of a page being mapped (for the PTE which maps it). 3858 * 3859 * Return: %0 on success, %VM_FAULT_ code in case of error. 3860 */ 3861 vm_fault_t finish_fault(struct vm_fault *vmf) 3862 { 3863 struct page *page; 3864 vm_fault_t ret = 0; 3865 3866 /* Did we COW the page? */ 3867 if ((vmf->flags & FAULT_FLAG_WRITE) && 3868 !(vmf->vma->vm_flags & VM_SHARED)) 3869 page = vmf->cow_page; 3870 else 3871 page = vmf->page; 3872 3873 /* 3874 * check even for read faults because we might have lost our CoWed 3875 * page 3876 */ 3877 if (!(vmf->vma->vm_flags & VM_SHARED)) 3878 ret = check_stable_address_space(vmf->vma->vm_mm); 3879 if (!ret) 3880 ret = alloc_set_pte(vmf, page); 3881 if (vmf->pte) 3882 pte_unmap_unlock(vmf->pte, vmf->ptl); 3883 return ret; 3884 } 3885 3886 static unsigned long fault_around_bytes __read_mostly = 3887 rounddown_pow_of_two(65536); 3888 3889 #ifdef CONFIG_DEBUG_FS 3890 static int fault_around_bytes_get(void *data, u64 *val) 3891 { 3892 *val = fault_around_bytes; 3893 return 0; 3894 } 3895 3896 /* 3897 * fault_around_bytes must be rounded down to the nearest page order as it's 3898 * what do_fault_around() expects to see. 3899 */ 3900 static int fault_around_bytes_set(void *data, u64 val) 3901 { 3902 if (val / PAGE_SIZE > PTRS_PER_PTE) 3903 return -EINVAL; 3904 if (val > PAGE_SIZE) 3905 fault_around_bytes = rounddown_pow_of_two(val); 3906 else 3907 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 3908 return 0; 3909 } 3910 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 3911 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 3912 3913 static int __init fault_around_debugfs(void) 3914 { 3915 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 3916 &fault_around_bytes_fops); 3917 return 0; 3918 } 3919 late_initcall(fault_around_debugfs); 3920 #endif 3921 3922 /* 3923 * do_fault_around() tries to map few pages around the fault address. The hope 3924 * is that the pages will be needed soon and this will lower the number of 3925 * faults to handle. 3926 * 3927 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 3928 * not ready to be mapped: not up-to-date, locked, etc. 3929 * 3930 * This function is called with the page table lock taken. In the split ptlock 3931 * case the page table lock only protects only those entries which belong to 3932 * the page table corresponding to the fault address. 3933 * 3934 * This function doesn't cross the VMA boundaries, in order to call map_pages() 3935 * only once. 3936 * 3937 * fault_around_bytes defines how many bytes we'll try to map. 3938 * do_fault_around() expects it to be set to a power of two less than or equal 3939 * to PTRS_PER_PTE. 3940 * 3941 * The virtual address of the area that we map is naturally aligned to 3942 * fault_around_bytes rounded down to the machine page size 3943 * (and therefore to page order). This way it's easier to guarantee 3944 * that we don't cross page table boundaries. 3945 */ 3946 static vm_fault_t do_fault_around(struct vm_fault *vmf) 3947 { 3948 unsigned long address = vmf->address, nr_pages, mask; 3949 pgoff_t start_pgoff = vmf->pgoff; 3950 pgoff_t end_pgoff; 3951 int off; 3952 vm_fault_t ret = 0; 3953 3954 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 3955 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 3956 3957 vmf->address = max(address & mask, vmf->vma->vm_start); 3958 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 3959 start_pgoff -= off; 3960 3961 /* 3962 * end_pgoff is either the end of the page table, the end of 3963 * the vma or nr_pages from start_pgoff, depending what is nearest. 3964 */ 3965 end_pgoff = start_pgoff - 3966 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 3967 PTRS_PER_PTE - 1; 3968 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1, 3969 start_pgoff + nr_pages - 1); 3970 3971 if (pmd_none(*vmf->pmd)) { 3972 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 3973 if (!vmf->prealloc_pte) 3974 goto out; 3975 smp_wmb(); /* See comment in __pte_alloc() */ 3976 } 3977 3978 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff); 3979 3980 /* Huge page is mapped? Page fault is solved */ 3981 if (pmd_trans_huge(*vmf->pmd)) { 3982 ret = VM_FAULT_NOPAGE; 3983 goto out; 3984 } 3985 3986 /* ->map_pages() haven't done anything useful. Cold page cache? */ 3987 if (!vmf->pte) 3988 goto out; 3989 3990 /* check if the page fault is solved */ 3991 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT); 3992 if (!pte_none(*vmf->pte)) 3993 ret = VM_FAULT_NOPAGE; 3994 pte_unmap_unlock(vmf->pte, vmf->ptl); 3995 out: 3996 vmf->address = address; 3997 vmf->pte = NULL; 3998 return ret; 3999 } 4000 4001 static vm_fault_t do_read_fault(struct vm_fault *vmf) 4002 { 4003 struct vm_area_struct *vma = vmf->vma; 4004 vm_fault_t ret = 0; 4005 4006 /* 4007 * Let's call ->map_pages() first and use ->fault() as fallback 4008 * if page by the offset is not ready to be mapped (cold cache or 4009 * something). 4010 */ 4011 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 4012 ret = do_fault_around(vmf); 4013 if (ret) 4014 return ret; 4015 } 4016 4017 ret = __do_fault(vmf); 4018 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4019 return ret; 4020 4021 ret |= finish_fault(vmf); 4022 unlock_page(vmf->page); 4023 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4024 put_page(vmf->page); 4025 return ret; 4026 } 4027 4028 static vm_fault_t do_cow_fault(struct vm_fault *vmf) 4029 { 4030 struct vm_area_struct *vma = vmf->vma; 4031 vm_fault_t ret; 4032 4033 if (unlikely(anon_vma_prepare(vma))) 4034 return VM_FAULT_OOM; 4035 4036 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address); 4037 if (!vmf->cow_page) 4038 return VM_FAULT_OOM; 4039 4040 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) { 4041 put_page(vmf->cow_page); 4042 return VM_FAULT_OOM; 4043 } 4044 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL); 4045 4046 ret = __do_fault(vmf); 4047 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4048 goto uncharge_out; 4049 if (ret & VM_FAULT_DONE_COW) 4050 return ret; 4051 4052 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 4053 __SetPageUptodate(vmf->cow_page); 4054 4055 ret |= finish_fault(vmf); 4056 unlock_page(vmf->page); 4057 put_page(vmf->page); 4058 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4059 goto uncharge_out; 4060 return ret; 4061 uncharge_out: 4062 put_page(vmf->cow_page); 4063 return ret; 4064 } 4065 4066 static vm_fault_t do_shared_fault(struct vm_fault *vmf) 4067 { 4068 struct vm_area_struct *vma = vmf->vma; 4069 vm_fault_t ret, tmp; 4070 4071 ret = __do_fault(vmf); 4072 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4073 return ret; 4074 4075 /* 4076 * Check if the backing address space wants to know that the page is 4077 * about to become writable 4078 */ 4079 if (vma->vm_ops->page_mkwrite) { 4080 unlock_page(vmf->page); 4081 tmp = do_page_mkwrite(vmf); 4082 if (unlikely(!tmp || 4083 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 4084 put_page(vmf->page); 4085 return tmp; 4086 } 4087 } 4088 4089 ret |= finish_fault(vmf); 4090 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 4091 VM_FAULT_RETRY))) { 4092 unlock_page(vmf->page); 4093 put_page(vmf->page); 4094 return ret; 4095 } 4096 4097 ret |= fault_dirty_shared_page(vmf); 4098 return ret; 4099 } 4100 4101 /* 4102 * We enter with non-exclusive mmap_lock (to exclude vma changes, 4103 * but allow concurrent faults). 4104 * The mmap_lock may have been released depending on flags and our 4105 * return value. See filemap_fault() and __lock_page_or_retry(). 4106 * If mmap_lock is released, vma may become invalid (for example 4107 * by other thread calling munmap()). 4108 */ 4109 static vm_fault_t do_fault(struct vm_fault *vmf) 4110 { 4111 struct vm_area_struct *vma = vmf->vma; 4112 struct mm_struct *vm_mm = vma->vm_mm; 4113 vm_fault_t ret; 4114 4115 /* 4116 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 4117 */ 4118 if (!vma->vm_ops->fault) { 4119 /* 4120 * If we find a migration pmd entry or a none pmd entry, which 4121 * should never happen, return SIGBUS 4122 */ 4123 if (unlikely(!pmd_present(*vmf->pmd))) 4124 ret = VM_FAULT_SIGBUS; 4125 else { 4126 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, 4127 vmf->pmd, 4128 vmf->address, 4129 &vmf->ptl); 4130 /* 4131 * Make sure this is not a temporary clearing of pte 4132 * by holding ptl and checking again. A R/M/W update 4133 * of pte involves: take ptl, clearing the pte so that 4134 * we don't have concurrent modification by hardware 4135 * followed by an update. 4136 */ 4137 if (unlikely(pte_none(*vmf->pte))) 4138 ret = VM_FAULT_SIGBUS; 4139 else 4140 ret = VM_FAULT_NOPAGE; 4141 4142 pte_unmap_unlock(vmf->pte, vmf->ptl); 4143 } 4144 } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 4145 ret = do_read_fault(vmf); 4146 else if (!(vma->vm_flags & VM_SHARED)) 4147 ret = do_cow_fault(vmf); 4148 else 4149 ret = do_shared_fault(vmf); 4150 4151 /* preallocated pagetable is unused: free it */ 4152 if (vmf->prealloc_pte) { 4153 pte_free(vm_mm, vmf->prealloc_pte); 4154 vmf->prealloc_pte = NULL; 4155 } 4156 return ret; 4157 } 4158 4159 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 4160 unsigned long addr, int page_nid, 4161 int *flags) 4162 { 4163 get_page(page); 4164 4165 count_vm_numa_event(NUMA_HINT_FAULTS); 4166 if (page_nid == numa_node_id()) { 4167 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 4168 *flags |= TNF_FAULT_LOCAL; 4169 } 4170 4171 return mpol_misplaced(page, vma, addr); 4172 } 4173 4174 static vm_fault_t do_numa_page(struct vm_fault *vmf) 4175 { 4176 struct vm_area_struct *vma = vmf->vma; 4177 struct page *page = NULL; 4178 int page_nid = NUMA_NO_NODE; 4179 int last_cpupid; 4180 int target_nid; 4181 bool migrated = false; 4182 pte_t pte, old_pte; 4183 bool was_writable = pte_savedwrite(vmf->orig_pte); 4184 int flags = 0; 4185 4186 /* 4187 * The "pte" at this point cannot be used safely without 4188 * validation through pte_unmap_same(). It's of NUMA type but 4189 * the pfn may be screwed if the read is non atomic. 4190 */ 4191 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd); 4192 spin_lock(vmf->ptl); 4193 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4194 pte_unmap_unlock(vmf->pte, vmf->ptl); 4195 goto out; 4196 } 4197 4198 /* 4199 * Make it present again, Depending on how arch implementes non 4200 * accessible ptes, some can allow access by kernel mode. 4201 */ 4202 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 4203 pte = pte_modify(old_pte, vma->vm_page_prot); 4204 pte = pte_mkyoung(pte); 4205 if (was_writable) 4206 pte = pte_mkwrite(pte); 4207 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 4208 update_mmu_cache(vma, vmf->address, vmf->pte); 4209 4210 page = vm_normal_page(vma, vmf->address, pte); 4211 if (!page) { 4212 pte_unmap_unlock(vmf->pte, vmf->ptl); 4213 return 0; 4214 } 4215 4216 /* TODO: handle PTE-mapped THP */ 4217 if (PageCompound(page)) { 4218 pte_unmap_unlock(vmf->pte, vmf->ptl); 4219 return 0; 4220 } 4221 4222 /* 4223 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 4224 * much anyway since they can be in shared cache state. This misses 4225 * the case where a mapping is writable but the process never writes 4226 * to it but pte_write gets cleared during protection updates and 4227 * pte_dirty has unpredictable behaviour between PTE scan updates, 4228 * background writeback, dirty balancing and application behaviour. 4229 */ 4230 if (!pte_write(pte)) 4231 flags |= TNF_NO_GROUP; 4232 4233 /* 4234 * Flag if the page is shared between multiple address spaces. This 4235 * is later used when determining whether to group tasks together 4236 */ 4237 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 4238 flags |= TNF_SHARED; 4239 4240 last_cpupid = page_cpupid_last(page); 4241 page_nid = page_to_nid(page); 4242 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid, 4243 &flags); 4244 pte_unmap_unlock(vmf->pte, vmf->ptl); 4245 if (target_nid == NUMA_NO_NODE) { 4246 put_page(page); 4247 goto out; 4248 } 4249 4250 /* Migrate to the requested node */ 4251 migrated = migrate_misplaced_page(page, vma, target_nid); 4252 if (migrated) { 4253 page_nid = target_nid; 4254 flags |= TNF_MIGRATED; 4255 } else 4256 flags |= TNF_MIGRATE_FAIL; 4257 4258 out: 4259 if (page_nid != NUMA_NO_NODE) 4260 task_numa_fault(last_cpupid, page_nid, 1, flags); 4261 return 0; 4262 } 4263 4264 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 4265 { 4266 if (vma_is_anonymous(vmf->vma)) 4267 return do_huge_pmd_anonymous_page(vmf); 4268 if (vmf->vma->vm_ops->huge_fault) 4269 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4270 return VM_FAULT_FALLBACK; 4271 } 4272 4273 /* `inline' is required to avoid gcc 4.1.2 build error */ 4274 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd) 4275 { 4276 if (vma_is_anonymous(vmf->vma)) { 4277 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd)) 4278 return handle_userfault(vmf, VM_UFFD_WP); 4279 return do_huge_pmd_wp_page(vmf, orig_pmd); 4280 } 4281 if (vmf->vma->vm_ops->huge_fault) { 4282 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4283 4284 if (!(ret & VM_FAULT_FALLBACK)) 4285 return ret; 4286 } 4287 4288 /* COW or write-notify handled on pte level: split pmd. */ 4289 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL); 4290 4291 return VM_FAULT_FALLBACK; 4292 } 4293 4294 static vm_fault_t create_huge_pud(struct vm_fault *vmf) 4295 { 4296 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 4297 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 4298 /* No support for anonymous transparent PUD pages yet */ 4299 if (vma_is_anonymous(vmf->vma)) 4300 goto split; 4301 if (vmf->vma->vm_ops->huge_fault) { 4302 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4303 4304 if (!(ret & VM_FAULT_FALLBACK)) 4305 return ret; 4306 } 4307 split: 4308 /* COW or write-notify not handled on PUD level: split pud.*/ 4309 __split_huge_pud(vmf->vma, vmf->pud, vmf->address); 4310 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4311 return VM_FAULT_FALLBACK; 4312 } 4313 4314 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 4315 { 4316 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4317 /* No support for anonymous transparent PUD pages yet */ 4318 if (vma_is_anonymous(vmf->vma)) 4319 return VM_FAULT_FALLBACK; 4320 if (vmf->vma->vm_ops->huge_fault) 4321 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4322 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4323 return VM_FAULT_FALLBACK; 4324 } 4325 4326 /* 4327 * These routines also need to handle stuff like marking pages dirty 4328 * and/or accessed for architectures that don't do it in hardware (most 4329 * RISC architectures). The early dirtying is also good on the i386. 4330 * 4331 * There is also a hook called "update_mmu_cache()" that architectures 4332 * with external mmu caches can use to update those (ie the Sparc or 4333 * PowerPC hashed page tables that act as extended TLBs). 4334 * 4335 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow 4336 * concurrent faults). 4337 * 4338 * The mmap_lock may have been released depending on flags and our return value. 4339 * See filemap_fault() and __lock_page_or_retry(). 4340 */ 4341 static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 4342 { 4343 pte_t entry; 4344 4345 if (unlikely(pmd_none(*vmf->pmd))) { 4346 /* 4347 * Leave __pte_alloc() until later: because vm_ops->fault may 4348 * want to allocate huge page, and if we expose page table 4349 * for an instant, it will be difficult to retract from 4350 * concurrent faults and from rmap lookups. 4351 */ 4352 vmf->pte = NULL; 4353 } else { 4354 /* See comment in pte_alloc_one_map() */ 4355 if (pmd_devmap_trans_unstable(vmf->pmd)) 4356 return 0; 4357 /* 4358 * A regular pmd is established and it can't morph into a huge 4359 * pmd from under us anymore at this point because we hold the 4360 * mmap_lock read mode and khugepaged takes it in write mode. 4361 * So now it's safe to run pte_offset_map(). 4362 */ 4363 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4364 vmf->orig_pte = *vmf->pte; 4365 4366 /* 4367 * some architectures can have larger ptes than wordsize, 4368 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and 4369 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic 4370 * accesses. The code below just needs a consistent view 4371 * for the ifs and we later double check anyway with the 4372 * ptl lock held. So here a barrier will do. 4373 */ 4374 barrier(); 4375 if (pte_none(vmf->orig_pte)) { 4376 pte_unmap(vmf->pte); 4377 vmf->pte = NULL; 4378 } 4379 } 4380 4381 if (!vmf->pte) { 4382 if (vma_is_anonymous(vmf->vma)) 4383 return do_anonymous_page(vmf); 4384 else 4385 return do_fault(vmf); 4386 } 4387 4388 if (!pte_present(vmf->orig_pte)) 4389 return do_swap_page(vmf); 4390 4391 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 4392 return do_numa_page(vmf); 4393 4394 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 4395 spin_lock(vmf->ptl); 4396 entry = vmf->orig_pte; 4397 if (unlikely(!pte_same(*vmf->pte, entry))) { 4398 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 4399 goto unlock; 4400 } 4401 if (vmf->flags & FAULT_FLAG_WRITE) { 4402 if (!pte_write(entry)) 4403 return do_wp_page(vmf); 4404 entry = pte_mkdirty(entry); 4405 } 4406 entry = pte_mkyoung(entry); 4407 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 4408 vmf->flags & FAULT_FLAG_WRITE)) { 4409 update_mmu_cache(vmf->vma, vmf->address, vmf->pte); 4410 } else { 4411 /* Skip spurious TLB flush for retried page fault */ 4412 if (vmf->flags & FAULT_FLAG_TRIED) 4413 goto unlock; 4414 /* 4415 * This is needed only for protection faults but the arch code 4416 * is not yet telling us if this is a protection fault or not. 4417 * This still avoids useless tlb flushes for .text page faults 4418 * with threads. 4419 */ 4420 if (vmf->flags & FAULT_FLAG_WRITE) 4421 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address); 4422 } 4423 unlock: 4424 pte_unmap_unlock(vmf->pte, vmf->ptl); 4425 return 0; 4426 } 4427 4428 /* 4429 * By the time we get here, we already hold the mm semaphore 4430 * 4431 * The mmap_lock may have been released depending on flags and our 4432 * return value. See filemap_fault() and __lock_page_or_retry(). 4433 */ 4434 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 4435 unsigned long address, unsigned int flags) 4436 { 4437 struct vm_fault vmf = { 4438 .vma = vma, 4439 .address = address & PAGE_MASK, 4440 .flags = flags, 4441 .pgoff = linear_page_index(vma, address), 4442 .gfp_mask = __get_fault_gfp_mask(vma), 4443 }; 4444 unsigned int dirty = flags & FAULT_FLAG_WRITE; 4445 struct mm_struct *mm = vma->vm_mm; 4446 pgd_t *pgd; 4447 p4d_t *p4d; 4448 vm_fault_t ret; 4449 4450 pgd = pgd_offset(mm, address); 4451 p4d = p4d_alloc(mm, pgd, address); 4452 if (!p4d) 4453 return VM_FAULT_OOM; 4454 4455 vmf.pud = pud_alloc(mm, p4d, address); 4456 if (!vmf.pud) 4457 return VM_FAULT_OOM; 4458 retry_pud: 4459 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) { 4460 ret = create_huge_pud(&vmf); 4461 if (!(ret & VM_FAULT_FALLBACK)) 4462 return ret; 4463 } else { 4464 pud_t orig_pud = *vmf.pud; 4465 4466 barrier(); 4467 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 4468 4469 /* NUMA case for anonymous PUDs would go here */ 4470 4471 if (dirty && !pud_write(orig_pud)) { 4472 ret = wp_huge_pud(&vmf, orig_pud); 4473 if (!(ret & VM_FAULT_FALLBACK)) 4474 return ret; 4475 } else { 4476 huge_pud_set_accessed(&vmf, orig_pud); 4477 return 0; 4478 } 4479 } 4480 } 4481 4482 vmf.pmd = pmd_alloc(mm, vmf.pud, address); 4483 if (!vmf.pmd) 4484 return VM_FAULT_OOM; 4485 4486 /* Huge pud page fault raced with pmd_alloc? */ 4487 if (pud_trans_unstable(vmf.pud)) 4488 goto retry_pud; 4489 4490 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) { 4491 ret = create_huge_pmd(&vmf); 4492 if (!(ret & VM_FAULT_FALLBACK)) 4493 return ret; 4494 } else { 4495 pmd_t orig_pmd = *vmf.pmd; 4496 4497 barrier(); 4498 if (unlikely(is_swap_pmd(orig_pmd))) { 4499 VM_BUG_ON(thp_migration_supported() && 4500 !is_pmd_migration_entry(orig_pmd)); 4501 if (is_pmd_migration_entry(orig_pmd)) 4502 pmd_migration_entry_wait(mm, vmf.pmd); 4503 return 0; 4504 } 4505 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) { 4506 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma)) 4507 return do_huge_pmd_numa_page(&vmf, orig_pmd); 4508 4509 if (dirty && !pmd_write(orig_pmd)) { 4510 ret = wp_huge_pmd(&vmf, orig_pmd); 4511 if (!(ret & VM_FAULT_FALLBACK)) 4512 return ret; 4513 } else { 4514 huge_pmd_set_accessed(&vmf, orig_pmd); 4515 return 0; 4516 } 4517 } 4518 } 4519 4520 return handle_pte_fault(&vmf); 4521 } 4522 4523 /** 4524 * mm_account_fault - Do page fault accountings 4525 * 4526 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting 4527 * of perf event counters, but we'll still do the per-task accounting to 4528 * the task who triggered this page fault. 4529 * @address: the faulted address. 4530 * @flags: the fault flags. 4531 * @ret: the fault retcode. 4532 * 4533 * This will take care of most of the page fault accountings. Meanwhile, it 4534 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter 4535 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should 4536 * still be in per-arch page fault handlers at the entry of page fault. 4537 */ 4538 static inline void mm_account_fault(struct pt_regs *regs, 4539 unsigned long address, unsigned int flags, 4540 vm_fault_t ret) 4541 { 4542 bool major; 4543 4544 /* 4545 * We don't do accounting for some specific faults: 4546 * 4547 * - Unsuccessful faults (e.g. when the address wasn't valid). That 4548 * includes arch_vma_access_permitted() failing before reaching here. 4549 * So this is not a "this many hardware page faults" counter. We 4550 * should use the hw profiling for that. 4551 * 4552 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted 4553 * once they're completed. 4554 */ 4555 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY)) 4556 return; 4557 4558 /* 4559 * We define the fault as a major fault when the final successful fault 4560 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't 4561 * handle it immediately previously). 4562 */ 4563 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); 4564 4565 if (major) 4566 current->maj_flt++; 4567 else 4568 current->min_flt++; 4569 4570 /* 4571 * If the fault is done for GUP, regs will be NULL. We only do the 4572 * accounting for the per thread fault counters who triggered the 4573 * fault, and we skip the perf event updates. 4574 */ 4575 if (!regs) 4576 return; 4577 4578 if (major) 4579 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 4580 else 4581 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 4582 } 4583 4584 /* 4585 * By the time we get here, we already hold the mm semaphore 4586 * 4587 * The mmap_lock may have been released depending on flags and our 4588 * return value. See filemap_fault() and __lock_page_or_retry(). 4589 */ 4590 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 4591 unsigned int flags, struct pt_regs *regs) 4592 { 4593 vm_fault_t ret; 4594 4595 __set_current_state(TASK_RUNNING); 4596 4597 count_vm_event(PGFAULT); 4598 count_memcg_event_mm(vma->vm_mm, PGFAULT); 4599 4600 /* do counter updates before entering really critical section. */ 4601 check_sync_rss_stat(current); 4602 4603 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 4604 flags & FAULT_FLAG_INSTRUCTION, 4605 flags & FAULT_FLAG_REMOTE)) 4606 return VM_FAULT_SIGSEGV; 4607 4608 /* 4609 * Enable the memcg OOM handling for faults triggered in user 4610 * space. Kernel faults are handled more gracefully. 4611 */ 4612 if (flags & FAULT_FLAG_USER) 4613 mem_cgroup_enter_user_fault(); 4614 4615 if (unlikely(is_vm_hugetlb_page(vma))) 4616 ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 4617 else 4618 ret = __handle_mm_fault(vma, address, flags); 4619 4620 if (flags & FAULT_FLAG_USER) { 4621 mem_cgroup_exit_user_fault(); 4622 /* 4623 * The task may have entered a memcg OOM situation but 4624 * if the allocation error was handled gracefully (no 4625 * VM_FAULT_OOM), there is no need to kill anything. 4626 * Just clean up the OOM state peacefully. 4627 */ 4628 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 4629 mem_cgroup_oom_synchronize(false); 4630 } 4631 4632 mm_account_fault(regs, address, flags, ret); 4633 4634 return ret; 4635 } 4636 EXPORT_SYMBOL_GPL(handle_mm_fault); 4637 4638 #ifndef __PAGETABLE_P4D_FOLDED 4639 /* 4640 * Allocate p4d page table. 4641 * We've already handled the fast-path in-line. 4642 */ 4643 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 4644 { 4645 p4d_t *new = p4d_alloc_one(mm, address); 4646 if (!new) 4647 return -ENOMEM; 4648 4649 smp_wmb(); /* See comment in __pte_alloc */ 4650 4651 spin_lock(&mm->page_table_lock); 4652 if (pgd_present(*pgd)) /* Another has populated it */ 4653 p4d_free(mm, new); 4654 else 4655 pgd_populate(mm, pgd, new); 4656 spin_unlock(&mm->page_table_lock); 4657 return 0; 4658 } 4659 #endif /* __PAGETABLE_P4D_FOLDED */ 4660 4661 #ifndef __PAGETABLE_PUD_FOLDED 4662 /* 4663 * Allocate page upper directory. 4664 * We've already handled the fast-path in-line. 4665 */ 4666 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 4667 { 4668 pud_t *new = pud_alloc_one(mm, address); 4669 if (!new) 4670 return -ENOMEM; 4671 4672 smp_wmb(); /* See comment in __pte_alloc */ 4673 4674 spin_lock(&mm->page_table_lock); 4675 if (!p4d_present(*p4d)) { 4676 mm_inc_nr_puds(mm); 4677 p4d_populate(mm, p4d, new); 4678 } else /* Another has populated it */ 4679 pud_free(mm, new); 4680 spin_unlock(&mm->page_table_lock); 4681 return 0; 4682 } 4683 #endif /* __PAGETABLE_PUD_FOLDED */ 4684 4685 #ifndef __PAGETABLE_PMD_FOLDED 4686 /* 4687 * Allocate page middle directory. 4688 * We've already handled the fast-path in-line. 4689 */ 4690 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 4691 { 4692 spinlock_t *ptl; 4693 pmd_t *new = pmd_alloc_one(mm, address); 4694 if (!new) 4695 return -ENOMEM; 4696 4697 smp_wmb(); /* See comment in __pte_alloc */ 4698 4699 ptl = pud_lock(mm, pud); 4700 if (!pud_present(*pud)) { 4701 mm_inc_nr_pmds(mm); 4702 pud_populate(mm, pud, new); 4703 } else /* Another has populated it */ 4704 pmd_free(mm, new); 4705 spin_unlock(ptl); 4706 return 0; 4707 } 4708 #endif /* __PAGETABLE_PMD_FOLDED */ 4709 4710 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address, 4711 struct mmu_notifier_range *range, 4712 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp) 4713 { 4714 pgd_t *pgd; 4715 p4d_t *p4d; 4716 pud_t *pud; 4717 pmd_t *pmd; 4718 pte_t *ptep; 4719 4720 pgd = pgd_offset(mm, address); 4721 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 4722 goto out; 4723 4724 p4d = p4d_offset(pgd, address); 4725 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 4726 goto out; 4727 4728 pud = pud_offset(p4d, address); 4729 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 4730 goto out; 4731 4732 pmd = pmd_offset(pud, address); 4733 VM_BUG_ON(pmd_trans_huge(*pmd)); 4734 4735 if (pmd_huge(*pmd)) { 4736 if (!pmdpp) 4737 goto out; 4738 4739 if (range) { 4740 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, 4741 NULL, mm, address & PMD_MASK, 4742 (address & PMD_MASK) + PMD_SIZE); 4743 mmu_notifier_invalidate_range_start(range); 4744 } 4745 *ptlp = pmd_lock(mm, pmd); 4746 if (pmd_huge(*pmd)) { 4747 *pmdpp = pmd; 4748 return 0; 4749 } 4750 spin_unlock(*ptlp); 4751 if (range) 4752 mmu_notifier_invalidate_range_end(range); 4753 } 4754 4755 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 4756 goto out; 4757 4758 if (range) { 4759 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 4760 address & PAGE_MASK, 4761 (address & PAGE_MASK) + PAGE_SIZE); 4762 mmu_notifier_invalidate_range_start(range); 4763 } 4764 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 4765 if (!pte_present(*ptep)) 4766 goto unlock; 4767 *ptepp = ptep; 4768 return 0; 4769 unlock: 4770 pte_unmap_unlock(ptep, *ptlp); 4771 if (range) 4772 mmu_notifier_invalidate_range_end(range); 4773 out: 4774 return -EINVAL; 4775 } 4776 4777 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 4778 pte_t **ptepp, spinlock_t **ptlp) 4779 { 4780 int res; 4781 4782 /* (void) is needed to make gcc happy */ 4783 (void) __cond_lock(*ptlp, 4784 !(res = __follow_pte_pmd(mm, address, NULL, 4785 ptepp, NULL, ptlp))); 4786 return res; 4787 } 4788 4789 int follow_pte_pmd(struct mm_struct *mm, unsigned long address, 4790 struct mmu_notifier_range *range, 4791 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp) 4792 { 4793 int res; 4794 4795 /* (void) is needed to make gcc happy */ 4796 (void) __cond_lock(*ptlp, 4797 !(res = __follow_pte_pmd(mm, address, range, 4798 ptepp, pmdpp, ptlp))); 4799 return res; 4800 } 4801 EXPORT_SYMBOL(follow_pte_pmd); 4802 4803 /** 4804 * follow_pfn - look up PFN at a user virtual address 4805 * @vma: memory mapping 4806 * @address: user virtual address 4807 * @pfn: location to store found PFN 4808 * 4809 * Only IO mappings and raw PFN mappings are allowed. 4810 * 4811 * Return: zero and the pfn at @pfn on success, -ve otherwise. 4812 */ 4813 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 4814 unsigned long *pfn) 4815 { 4816 int ret = -EINVAL; 4817 spinlock_t *ptl; 4818 pte_t *ptep; 4819 4820 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4821 return ret; 4822 4823 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 4824 if (ret) 4825 return ret; 4826 *pfn = pte_pfn(*ptep); 4827 pte_unmap_unlock(ptep, ptl); 4828 return 0; 4829 } 4830 EXPORT_SYMBOL(follow_pfn); 4831 4832 #ifdef CONFIG_HAVE_IOREMAP_PROT 4833 int follow_phys(struct vm_area_struct *vma, 4834 unsigned long address, unsigned int flags, 4835 unsigned long *prot, resource_size_t *phys) 4836 { 4837 int ret = -EINVAL; 4838 pte_t *ptep, pte; 4839 spinlock_t *ptl; 4840 4841 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4842 goto out; 4843 4844 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 4845 goto out; 4846 pte = *ptep; 4847 4848 if ((flags & FOLL_WRITE) && !pte_write(pte)) 4849 goto unlock; 4850 4851 *prot = pgprot_val(pte_pgprot(pte)); 4852 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 4853 4854 ret = 0; 4855 unlock: 4856 pte_unmap_unlock(ptep, ptl); 4857 out: 4858 return ret; 4859 } 4860 4861 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 4862 void *buf, int len, int write) 4863 { 4864 resource_size_t phys_addr; 4865 unsigned long prot = 0; 4866 void __iomem *maddr; 4867 int offset = addr & (PAGE_SIZE-1); 4868 4869 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 4870 return -EINVAL; 4871 4872 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 4873 if (!maddr) 4874 return -ENOMEM; 4875 4876 if (write) 4877 memcpy_toio(maddr + offset, buf, len); 4878 else 4879 memcpy_fromio(buf, maddr + offset, len); 4880 iounmap(maddr); 4881 4882 return len; 4883 } 4884 EXPORT_SYMBOL_GPL(generic_access_phys); 4885 #endif 4886 4887 /* 4888 * Access another process' address space as given in mm. 4889 */ 4890 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf, 4891 int len, unsigned int gup_flags) 4892 { 4893 struct vm_area_struct *vma; 4894 void *old_buf = buf; 4895 int write = gup_flags & FOLL_WRITE; 4896 4897 if (mmap_read_lock_killable(mm)) 4898 return 0; 4899 4900 /* ignore errors, just check how much was successfully transferred */ 4901 while (len) { 4902 int bytes, ret, offset; 4903 void *maddr; 4904 struct page *page = NULL; 4905 4906 ret = get_user_pages_remote(mm, addr, 1, 4907 gup_flags, &page, &vma, NULL); 4908 if (ret <= 0) { 4909 #ifndef CONFIG_HAVE_IOREMAP_PROT 4910 break; 4911 #else 4912 /* 4913 * Check if this is a VM_IO | VM_PFNMAP VMA, which 4914 * we can access using slightly different code. 4915 */ 4916 vma = find_vma(mm, addr); 4917 if (!vma || vma->vm_start > addr) 4918 break; 4919 if (vma->vm_ops && vma->vm_ops->access) 4920 ret = vma->vm_ops->access(vma, addr, buf, 4921 len, write); 4922 if (ret <= 0) 4923 break; 4924 bytes = ret; 4925 #endif 4926 } else { 4927 bytes = len; 4928 offset = addr & (PAGE_SIZE-1); 4929 if (bytes > PAGE_SIZE-offset) 4930 bytes = PAGE_SIZE-offset; 4931 4932 maddr = kmap(page); 4933 if (write) { 4934 copy_to_user_page(vma, page, addr, 4935 maddr + offset, buf, bytes); 4936 set_page_dirty_lock(page); 4937 } else { 4938 copy_from_user_page(vma, page, addr, 4939 buf, maddr + offset, bytes); 4940 } 4941 kunmap(page); 4942 put_page(page); 4943 } 4944 len -= bytes; 4945 buf += bytes; 4946 addr += bytes; 4947 } 4948 mmap_read_unlock(mm); 4949 4950 return buf - old_buf; 4951 } 4952 4953 /** 4954 * access_remote_vm - access another process' address space 4955 * @mm: the mm_struct of the target address space 4956 * @addr: start address to access 4957 * @buf: source or destination buffer 4958 * @len: number of bytes to transfer 4959 * @gup_flags: flags modifying lookup behaviour 4960 * 4961 * The caller must hold a reference on @mm. 4962 * 4963 * Return: number of bytes copied from source to destination. 4964 */ 4965 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 4966 void *buf, int len, unsigned int gup_flags) 4967 { 4968 return __access_remote_vm(mm, addr, buf, len, gup_flags); 4969 } 4970 4971 /* 4972 * Access another process' address space. 4973 * Source/target buffer must be kernel space, 4974 * Do not walk the page table directly, use get_user_pages 4975 */ 4976 int access_process_vm(struct task_struct *tsk, unsigned long addr, 4977 void *buf, int len, unsigned int gup_flags) 4978 { 4979 struct mm_struct *mm; 4980 int ret; 4981 4982 mm = get_task_mm(tsk); 4983 if (!mm) 4984 return 0; 4985 4986 ret = __access_remote_vm(mm, addr, buf, len, gup_flags); 4987 4988 mmput(mm); 4989 4990 return ret; 4991 } 4992 EXPORT_SYMBOL_GPL(access_process_vm); 4993 4994 /* 4995 * Print the name of a VMA. 4996 */ 4997 void print_vma_addr(char *prefix, unsigned long ip) 4998 { 4999 struct mm_struct *mm = current->mm; 5000 struct vm_area_struct *vma; 5001 5002 /* 5003 * we might be running from an atomic context so we cannot sleep 5004 */ 5005 if (!mmap_read_trylock(mm)) 5006 return; 5007 5008 vma = find_vma(mm, ip); 5009 if (vma && vma->vm_file) { 5010 struct file *f = vma->vm_file; 5011 char *buf = (char *)__get_free_page(GFP_NOWAIT); 5012 if (buf) { 5013 char *p; 5014 5015 p = file_path(f, buf, PAGE_SIZE); 5016 if (IS_ERR(p)) 5017 p = "?"; 5018 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 5019 vma->vm_start, 5020 vma->vm_end - vma->vm_start); 5021 free_page((unsigned long)buf); 5022 } 5023 } 5024 mmap_read_unlock(mm); 5025 } 5026 5027 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5028 void __might_fault(const char *file, int line) 5029 { 5030 /* 5031 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 5032 * holding the mmap_lock, this is safe because kernel memory doesn't 5033 * get paged out, therefore we'll never actually fault, and the 5034 * below annotations will generate false positives. 5035 */ 5036 if (uaccess_kernel()) 5037 return; 5038 if (pagefault_disabled()) 5039 return; 5040 __might_sleep(file, line, 0); 5041 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5042 if (current->mm) 5043 might_lock_read(¤t->mm->mmap_lock); 5044 #endif 5045 } 5046 EXPORT_SYMBOL(__might_fault); 5047 #endif 5048 5049 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 5050 /* 5051 * Process all subpages of the specified huge page with the specified 5052 * operation. The target subpage will be processed last to keep its 5053 * cache lines hot. 5054 */ 5055 static inline void process_huge_page( 5056 unsigned long addr_hint, unsigned int pages_per_huge_page, 5057 void (*process_subpage)(unsigned long addr, int idx, void *arg), 5058 void *arg) 5059 { 5060 int i, n, base, l; 5061 unsigned long addr = addr_hint & 5062 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5063 5064 /* Process target subpage last to keep its cache lines hot */ 5065 might_sleep(); 5066 n = (addr_hint - addr) / PAGE_SIZE; 5067 if (2 * n <= pages_per_huge_page) { 5068 /* If target subpage in first half of huge page */ 5069 base = 0; 5070 l = n; 5071 /* Process subpages at the end of huge page */ 5072 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 5073 cond_resched(); 5074 process_subpage(addr + i * PAGE_SIZE, i, arg); 5075 } 5076 } else { 5077 /* If target subpage in second half of huge page */ 5078 base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 5079 l = pages_per_huge_page - n; 5080 /* Process subpages at the begin of huge page */ 5081 for (i = 0; i < base; i++) { 5082 cond_resched(); 5083 process_subpage(addr + i * PAGE_SIZE, i, arg); 5084 } 5085 } 5086 /* 5087 * Process remaining subpages in left-right-left-right pattern 5088 * towards the target subpage 5089 */ 5090 for (i = 0; i < l; i++) { 5091 int left_idx = base + i; 5092 int right_idx = base + 2 * l - 1 - i; 5093 5094 cond_resched(); 5095 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 5096 cond_resched(); 5097 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 5098 } 5099 } 5100 5101 static void clear_gigantic_page(struct page *page, 5102 unsigned long addr, 5103 unsigned int pages_per_huge_page) 5104 { 5105 int i; 5106 struct page *p = page; 5107 5108 might_sleep(); 5109 for (i = 0; i < pages_per_huge_page; 5110 i++, p = mem_map_next(p, page, i)) { 5111 cond_resched(); 5112 clear_user_highpage(p, addr + i * PAGE_SIZE); 5113 } 5114 } 5115 5116 static void clear_subpage(unsigned long addr, int idx, void *arg) 5117 { 5118 struct page *page = arg; 5119 5120 clear_user_highpage(page + idx, addr); 5121 } 5122 5123 void clear_huge_page(struct page *page, 5124 unsigned long addr_hint, unsigned int pages_per_huge_page) 5125 { 5126 unsigned long addr = addr_hint & 5127 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5128 5129 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5130 clear_gigantic_page(page, addr, pages_per_huge_page); 5131 return; 5132 } 5133 5134 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 5135 } 5136 5137 static void copy_user_gigantic_page(struct page *dst, struct page *src, 5138 unsigned long addr, 5139 struct vm_area_struct *vma, 5140 unsigned int pages_per_huge_page) 5141 { 5142 int i; 5143 struct page *dst_base = dst; 5144 struct page *src_base = src; 5145 5146 for (i = 0; i < pages_per_huge_page; ) { 5147 cond_resched(); 5148 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 5149 5150 i++; 5151 dst = mem_map_next(dst, dst_base, i); 5152 src = mem_map_next(src, src_base, i); 5153 } 5154 } 5155 5156 struct copy_subpage_arg { 5157 struct page *dst; 5158 struct page *src; 5159 struct vm_area_struct *vma; 5160 }; 5161 5162 static void copy_subpage(unsigned long addr, int idx, void *arg) 5163 { 5164 struct copy_subpage_arg *copy_arg = arg; 5165 5166 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx, 5167 addr, copy_arg->vma); 5168 } 5169 5170 void copy_user_huge_page(struct page *dst, struct page *src, 5171 unsigned long addr_hint, struct vm_area_struct *vma, 5172 unsigned int pages_per_huge_page) 5173 { 5174 unsigned long addr = addr_hint & 5175 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5176 struct copy_subpage_arg arg = { 5177 .dst = dst, 5178 .src = src, 5179 .vma = vma, 5180 }; 5181 5182 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5183 copy_user_gigantic_page(dst, src, addr, vma, 5184 pages_per_huge_page); 5185 return; 5186 } 5187 5188 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 5189 } 5190 5191 long copy_huge_page_from_user(struct page *dst_page, 5192 const void __user *usr_src, 5193 unsigned int pages_per_huge_page, 5194 bool allow_pagefault) 5195 { 5196 void *src = (void *)usr_src; 5197 void *page_kaddr; 5198 unsigned long i, rc = 0; 5199 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE; 5200 5201 for (i = 0; i < pages_per_huge_page; i++) { 5202 if (allow_pagefault) 5203 page_kaddr = kmap(dst_page + i); 5204 else 5205 page_kaddr = kmap_atomic(dst_page + i); 5206 rc = copy_from_user(page_kaddr, 5207 (const void __user *)(src + i * PAGE_SIZE), 5208 PAGE_SIZE); 5209 if (allow_pagefault) 5210 kunmap(dst_page + i); 5211 else 5212 kunmap_atomic(page_kaddr); 5213 5214 ret_val -= (PAGE_SIZE - rc); 5215 if (rc) 5216 break; 5217 5218 cond_resched(); 5219 } 5220 return ret_val; 5221 } 5222 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 5223 5224 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 5225 5226 static struct kmem_cache *page_ptl_cachep; 5227 5228 void __init ptlock_cache_init(void) 5229 { 5230 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 5231 SLAB_PANIC, NULL); 5232 } 5233 5234 bool ptlock_alloc(struct page *page) 5235 { 5236 spinlock_t *ptl; 5237 5238 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 5239 if (!ptl) 5240 return false; 5241 page->ptl = ptl; 5242 return true; 5243 } 5244 5245 void ptlock_free(struct page *page) 5246 { 5247 kmem_cache_free(page_ptl_cachep, page->ptl); 5248 } 5249 #endif 5250