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/mm_inline.h> 45 #include <linux/sched/mm.h> 46 #include <linux/sched/coredump.h> 47 #include <linux/sched/numa_balancing.h> 48 #include <linux/sched/task.h> 49 #include <linux/hugetlb.h> 50 #include <linux/mman.h> 51 #include <linux/swap.h> 52 #include <linux/highmem.h> 53 #include <linux/pagemap.h> 54 #include <linux/memremap.h> 55 #include <linux/ksm.h> 56 #include <linux/rmap.h> 57 #include <linux/export.h> 58 #include <linux/delayacct.h> 59 #include <linux/init.h> 60 #include <linux/pfn_t.h> 61 #include <linux/writeback.h> 62 #include <linux/memcontrol.h> 63 #include <linux/mmu_notifier.h> 64 #include <linux/swapops.h> 65 #include <linux/elf.h> 66 #include <linux/gfp.h> 67 #include <linux/migrate.h> 68 #include <linux/string.h> 69 #include <linux/debugfs.h> 70 #include <linux/userfaultfd_k.h> 71 #include <linux/dax.h> 72 #include <linux/oom.h> 73 #include <linux/numa.h> 74 #include <linux/perf_event.h> 75 #include <linux/ptrace.h> 76 #include <linux/vmalloc.h> 77 78 #include <trace/events/kmem.h> 79 80 #include <asm/io.h> 81 #include <asm/mmu_context.h> 82 #include <asm/pgalloc.h> 83 #include <linux/uaccess.h> 84 #include <asm/tlb.h> 85 #include <asm/tlbflush.h> 86 87 #include "pgalloc-track.h" 88 #include "internal.h" 89 90 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) 91 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 92 #endif 93 94 #ifndef CONFIG_NUMA 95 unsigned long max_mapnr; 96 EXPORT_SYMBOL(max_mapnr); 97 98 struct page *mem_map; 99 EXPORT_SYMBOL(mem_map); 100 #endif 101 102 /* 103 * A number of key systems in x86 including ioremap() rely on the assumption 104 * that high_memory defines the upper bound on direct map memory, then end 105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 107 * and ZONE_HIGHMEM. 108 */ 109 void *high_memory; 110 EXPORT_SYMBOL(high_memory); 111 112 /* 113 * Randomize the address space (stacks, mmaps, brk, etc.). 114 * 115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 116 * as ancient (libc5 based) binaries can segfault. ) 117 */ 118 int randomize_va_space __read_mostly = 119 #ifdef CONFIG_COMPAT_BRK 120 1; 121 #else 122 2; 123 #endif 124 125 #ifndef arch_faults_on_old_pte 126 static inline bool arch_faults_on_old_pte(void) 127 { 128 /* 129 * Those arches which don't have hw access flag feature need to 130 * implement their own helper. By default, "true" means pagefault 131 * will be hit on old pte. 132 */ 133 return true; 134 } 135 #endif 136 137 #ifndef arch_wants_old_prefaulted_pte 138 static inline bool arch_wants_old_prefaulted_pte(void) 139 { 140 /* 141 * Transitioning a PTE from 'old' to 'young' can be expensive on 142 * some architectures, even if it's performed in hardware. By 143 * default, "false" means prefaulted entries will be 'young'. 144 */ 145 return false; 146 } 147 #endif 148 149 static int __init disable_randmaps(char *s) 150 { 151 randomize_va_space = 0; 152 return 1; 153 } 154 __setup("norandmaps", disable_randmaps); 155 156 unsigned long zero_pfn __read_mostly; 157 EXPORT_SYMBOL(zero_pfn); 158 159 unsigned long highest_memmap_pfn __read_mostly; 160 161 /* 162 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 163 */ 164 static int __init init_zero_pfn(void) 165 { 166 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 167 return 0; 168 } 169 early_initcall(init_zero_pfn); 170 171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count) 172 { 173 trace_rss_stat(mm, member, count); 174 } 175 176 #if defined(SPLIT_RSS_COUNTING) 177 178 void sync_mm_rss(struct mm_struct *mm) 179 { 180 int i; 181 182 for (i = 0; i < NR_MM_COUNTERS; i++) { 183 if (current->rss_stat.count[i]) { 184 add_mm_counter(mm, i, current->rss_stat.count[i]); 185 current->rss_stat.count[i] = 0; 186 } 187 } 188 current->rss_stat.events = 0; 189 } 190 191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 192 { 193 struct task_struct *task = current; 194 195 if (likely(task->mm == mm)) 196 task->rss_stat.count[member] += val; 197 else 198 add_mm_counter(mm, member, val); 199 } 200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 202 203 /* sync counter once per 64 page faults */ 204 #define TASK_RSS_EVENTS_THRESH (64) 205 static void check_sync_rss_stat(struct task_struct *task) 206 { 207 if (unlikely(task != current)) 208 return; 209 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 210 sync_mm_rss(task->mm); 211 } 212 #else /* SPLIT_RSS_COUNTING */ 213 214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 216 217 static void check_sync_rss_stat(struct task_struct *task) 218 { 219 } 220 221 #endif /* SPLIT_RSS_COUNTING */ 222 223 /* 224 * Note: this doesn't free the actual pages themselves. That 225 * has been handled earlier when unmapping all the memory regions. 226 */ 227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 228 unsigned long addr) 229 { 230 pgtable_t token = pmd_pgtable(*pmd); 231 pmd_clear(pmd); 232 pte_free_tlb(tlb, token, addr); 233 mm_dec_nr_ptes(tlb->mm); 234 } 235 236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 237 unsigned long addr, unsigned long end, 238 unsigned long floor, unsigned long ceiling) 239 { 240 pmd_t *pmd; 241 unsigned long next; 242 unsigned long start; 243 244 start = addr; 245 pmd = pmd_offset(pud, addr); 246 do { 247 next = pmd_addr_end(addr, end); 248 if (pmd_none_or_clear_bad(pmd)) 249 continue; 250 free_pte_range(tlb, pmd, addr); 251 } while (pmd++, addr = next, addr != end); 252 253 start &= PUD_MASK; 254 if (start < floor) 255 return; 256 if (ceiling) { 257 ceiling &= PUD_MASK; 258 if (!ceiling) 259 return; 260 } 261 if (end - 1 > ceiling - 1) 262 return; 263 264 pmd = pmd_offset(pud, start); 265 pud_clear(pud); 266 pmd_free_tlb(tlb, pmd, start); 267 mm_dec_nr_pmds(tlb->mm); 268 } 269 270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 271 unsigned long addr, unsigned long end, 272 unsigned long floor, unsigned long ceiling) 273 { 274 pud_t *pud; 275 unsigned long next; 276 unsigned long start; 277 278 start = addr; 279 pud = pud_offset(p4d, addr); 280 do { 281 next = pud_addr_end(addr, end); 282 if (pud_none_or_clear_bad(pud)) 283 continue; 284 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 285 } while (pud++, addr = next, addr != end); 286 287 start &= P4D_MASK; 288 if (start < floor) 289 return; 290 if (ceiling) { 291 ceiling &= P4D_MASK; 292 if (!ceiling) 293 return; 294 } 295 if (end - 1 > ceiling - 1) 296 return; 297 298 pud = pud_offset(p4d, start); 299 p4d_clear(p4d); 300 pud_free_tlb(tlb, pud, start); 301 mm_dec_nr_puds(tlb->mm); 302 } 303 304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 305 unsigned long addr, unsigned long end, 306 unsigned long floor, unsigned long ceiling) 307 { 308 p4d_t *p4d; 309 unsigned long next; 310 unsigned long start; 311 312 start = addr; 313 p4d = p4d_offset(pgd, addr); 314 do { 315 next = p4d_addr_end(addr, end); 316 if (p4d_none_or_clear_bad(p4d)) 317 continue; 318 free_pud_range(tlb, p4d, addr, next, floor, ceiling); 319 } while (p4d++, addr = next, addr != end); 320 321 start &= PGDIR_MASK; 322 if (start < floor) 323 return; 324 if (ceiling) { 325 ceiling &= PGDIR_MASK; 326 if (!ceiling) 327 return; 328 } 329 if (end - 1 > ceiling - 1) 330 return; 331 332 p4d = p4d_offset(pgd, start); 333 pgd_clear(pgd); 334 p4d_free_tlb(tlb, p4d, start); 335 } 336 337 /* 338 * This function frees user-level page tables of a process. 339 */ 340 void free_pgd_range(struct mmu_gather *tlb, 341 unsigned long addr, unsigned long end, 342 unsigned long floor, unsigned long ceiling) 343 { 344 pgd_t *pgd; 345 unsigned long next; 346 347 /* 348 * The next few lines have given us lots of grief... 349 * 350 * Why are we testing PMD* at this top level? Because often 351 * there will be no work to do at all, and we'd prefer not to 352 * go all the way down to the bottom just to discover that. 353 * 354 * Why all these "- 1"s? Because 0 represents both the bottom 355 * of the address space and the top of it (using -1 for the 356 * top wouldn't help much: the masks would do the wrong thing). 357 * The rule is that addr 0 and floor 0 refer to the bottom of 358 * the address space, but end 0 and ceiling 0 refer to the top 359 * Comparisons need to use "end - 1" and "ceiling - 1" (though 360 * that end 0 case should be mythical). 361 * 362 * Wherever addr is brought up or ceiling brought down, we must 363 * be careful to reject "the opposite 0" before it confuses the 364 * subsequent tests. But what about where end is brought down 365 * by PMD_SIZE below? no, end can't go down to 0 there. 366 * 367 * Whereas we round start (addr) and ceiling down, by different 368 * masks at different levels, in order to test whether a table 369 * now has no other vmas using it, so can be freed, we don't 370 * bother to round floor or end up - the tests don't need that. 371 */ 372 373 addr &= PMD_MASK; 374 if (addr < floor) { 375 addr += PMD_SIZE; 376 if (!addr) 377 return; 378 } 379 if (ceiling) { 380 ceiling &= PMD_MASK; 381 if (!ceiling) 382 return; 383 } 384 if (end - 1 > ceiling - 1) 385 end -= PMD_SIZE; 386 if (addr > end - 1) 387 return; 388 /* 389 * We add page table cache pages with PAGE_SIZE, 390 * (see pte_free_tlb()), flush the tlb if we need 391 */ 392 tlb_change_page_size(tlb, PAGE_SIZE); 393 pgd = pgd_offset(tlb->mm, addr); 394 do { 395 next = pgd_addr_end(addr, end); 396 if (pgd_none_or_clear_bad(pgd)) 397 continue; 398 free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 399 } while (pgd++, addr = next, addr != end); 400 } 401 402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 403 unsigned long floor, unsigned long ceiling) 404 { 405 while (vma) { 406 struct vm_area_struct *next = vma->vm_next; 407 unsigned long addr = vma->vm_start; 408 409 /* 410 * Hide vma from rmap and truncate_pagecache before freeing 411 * pgtables 412 */ 413 unlink_anon_vmas(vma); 414 unlink_file_vma(vma); 415 416 if (is_vm_hugetlb_page(vma)) { 417 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 418 floor, next ? next->vm_start : ceiling); 419 } else { 420 /* 421 * Optimization: gather nearby vmas into one call down 422 */ 423 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 424 && !is_vm_hugetlb_page(next)) { 425 vma = next; 426 next = vma->vm_next; 427 unlink_anon_vmas(vma); 428 unlink_file_vma(vma); 429 } 430 free_pgd_range(tlb, addr, vma->vm_end, 431 floor, next ? next->vm_start : ceiling); 432 } 433 vma = next; 434 } 435 } 436 437 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte) 438 { 439 spinlock_t *ptl = pmd_lock(mm, pmd); 440 441 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 442 mm_inc_nr_ptes(mm); 443 /* 444 * Ensure all pte setup (eg. pte page lock and page clearing) are 445 * visible before the pte is made visible to other CPUs by being 446 * put into page tables. 447 * 448 * The other side of the story is the pointer chasing in the page 449 * table walking code (when walking the page table without locking; 450 * ie. most of the time). Fortunately, these data accesses consist 451 * of a chain of data-dependent loads, meaning most CPUs (alpha 452 * being the notable exception) will already guarantee loads are 453 * seen in-order. See the alpha page table accessors for the 454 * smp_rmb() barriers in page table walking code. 455 */ 456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 457 pmd_populate(mm, pmd, *pte); 458 *pte = NULL; 459 } 460 spin_unlock(ptl); 461 } 462 463 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 464 { 465 pgtable_t new = pte_alloc_one(mm); 466 if (!new) 467 return -ENOMEM; 468 469 pmd_install(mm, pmd, &new); 470 if (new) 471 pte_free(mm, new); 472 return 0; 473 } 474 475 int __pte_alloc_kernel(pmd_t *pmd) 476 { 477 pte_t *new = pte_alloc_one_kernel(&init_mm); 478 if (!new) 479 return -ENOMEM; 480 481 spin_lock(&init_mm.page_table_lock); 482 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 483 smp_wmb(); /* See comment in pmd_install() */ 484 pmd_populate_kernel(&init_mm, pmd, new); 485 new = NULL; 486 } 487 spin_unlock(&init_mm.page_table_lock); 488 if (new) 489 pte_free_kernel(&init_mm, new); 490 return 0; 491 } 492 493 static inline void init_rss_vec(int *rss) 494 { 495 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 496 } 497 498 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 499 { 500 int i; 501 502 if (current->mm == mm) 503 sync_mm_rss(mm); 504 for (i = 0; i < NR_MM_COUNTERS; i++) 505 if (rss[i]) 506 add_mm_counter(mm, i, rss[i]); 507 } 508 509 /* 510 * This function is called to print an error when a bad pte 511 * is found. For example, we might have a PFN-mapped pte in 512 * a region that doesn't allow it. 513 * 514 * The calling function must still handle the error. 515 */ 516 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 517 pte_t pte, struct page *page) 518 { 519 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 520 p4d_t *p4d = p4d_offset(pgd, addr); 521 pud_t *pud = pud_offset(p4d, addr); 522 pmd_t *pmd = pmd_offset(pud, addr); 523 struct address_space *mapping; 524 pgoff_t index; 525 static unsigned long resume; 526 static unsigned long nr_shown; 527 static unsigned long nr_unshown; 528 529 /* 530 * Allow a burst of 60 reports, then keep quiet for that minute; 531 * or allow a steady drip of one report per second. 532 */ 533 if (nr_shown == 60) { 534 if (time_before(jiffies, resume)) { 535 nr_unshown++; 536 return; 537 } 538 if (nr_unshown) { 539 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 540 nr_unshown); 541 nr_unshown = 0; 542 } 543 nr_shown = 0; 544 } 545 if (nr_shown++ == 0) 546 resume = jiffies + 60 * HZ; 547 548 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 549 index = linear_page_index(vma, addr); 550 551 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 552 current->comm, 553 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 554 if (page) 555 dump_page(page, "bad pte"); 556 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", 557 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 558 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n", 559 vma->vm_file, 560 vma->vm_ops ? vma->vm_ops->fault : NULL, 561 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 562 mapping ? mapping->a_ops->readpage : NULL); 563 dump_stack(); 564 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 565 } 566 567 /* 568 * vm_normal_page -- This function gets the "struct page" associated with a pte. 569 * 570 * "Special" mappings do not wish to be associated with a "struct page" (either 571 * it doesn't exist, or it exists but they don't want to touch it). In this 572 * case, NULL is returned here. "Normal" mappings do have a struct page. 573 * 574 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 575 * pte bit, in which case this function is trivial. Secondly, an architecture 576 * may not have a spare pte bit, which requires a more complicated scheme, 577 * described below. 578 * 579 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 580 * special mapping (even if there are underlying and valid "struct pages"). 581 * COWed pages of a VM_PFNMAP are always normal. 582 * 583 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 584 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 585 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 586 * mapping will always honor the rule 587 * 588 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 589 * 590 * And for normal mappings this is false. 591 * 592 * This restricts such mappings to be a linear translation from virtual address 593 * to pfn. To get around this restriction, we allow arbitrary mappings so long 594 * as the vma is not a COW mapping; in that case, we know that all ptes are 595 * special (because none can have been COWed). 596 * 597 * 598 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 599 * 600 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 601 * page" backing, however the difference is that _all_ pages with a struct 602 * page (that is, those where pfn_valid is true) are refcounted and considered 603 * normal pages by the VM. The disadvantage is that pages are refcounted 604 * (which can be slower and simply not an option for some PFNMAP users). The 605 * advantage is that we don't have to follow the strict linearity rule of 606 * PFNMAP mappings in order to support COWable mappings. 607 * 608 */ 609 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 610 pte_t pte) 611 { 612 unsigned long pfn = pte_pfn(pte); 613 614 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 615 if (likely(!pte_special(pte))) 616 goto check_pfn; 617 if (vma->vm_ops && vma->vm_ops->find_special_page) 618 return vma->vm_ops->find_special_page(vma, addr); 619 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 620 return NULL; 621 if (is_zero_pfn(pfn)) 622 return NULL; 623 if (pte_devmap(pte)) 624 return NULL; 625 626 print_bad_pte(vma, addr, pte, NULL); 627 return NULL; 628 } 629 630 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 631 632 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 633 if (vma->vm_flags & VM_MIXEDMAP) { 634 if (!pfn_valid(pfn)) 635 return NULL; 636 goto out; 637 } else { 638 unsigned long off; 639 off = (addr - vma->vm_start) >> PAGE_SHIFT; 640 if (pfn == vma->vm_pgoff + off) 641 return NULL; 642 if (!is_cow_mapping(vma->vm_flags)) 643 return NULL; 644 } 645 } 646 647 if (is_zero_pfn(pfn)) 648 return NULL; 649 650 check_pfn: 651 if (unlikely(pfn > highest_memmap_pfn)) { 652 print_bad_pte(vma, addr, pte, NULL); 653 return NULL; 654 } 655 656 /* 657 * NOTE! We still have PageReserved() pages in the page tables. 658 * eg. VDSO mappings can cause them to exist. 659 */ 660 out: 661 return pfn_to_page(pfn); 662 } 663 664 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 665 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 666 pmd_t pmd) 667 { 668 unsigned long pfn = pmd_pfn(pmd); 669 670 /* 671 * There is no pmd_special() but there may be special pmds, e.g. 672 * in a direct-access (dax) mapping, so let's just replicate the 673 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 674 */ 675 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 676 if (vma->vm_flags & VM_MIXEDMAP) { 677 if (!pfn_valid(pfn)) 678 return NULL; 679 goto out; 680 } else { 681 unsigned long off; 682 off = (addr - vma->vm_start) >> PAGE_SHIFT; 683 if (pfn == vma->vm_pgoff + off) 684 return NULL; 685 if (!is_cow_mapping(vma->vm_flags)) 686 return NULL; 687 } 688 } 689 690 if (pmd_devmap(pmd)) 691 return NULL; 692 if (is_huge_zero_pmd(pmd)) 693 return NULL; 694 if (unlikely(pfn > highest_memmap_pfn)) 695 return NULL; 696 697 /* 698 * NOTE! We still have PageReserved() pages in the page tables. 699 * eg. VDSO mappings can cause them to exist. 700 */ 701 out: 702 return pfn_to_page(pfn); 703 } 704 #endif 705 706 static void restore_exclusive_pte(struct vm_area_struct *vma, 707 struct page *page, unsigned long address, 708 pte_t *ptep) 709 { 710 pte_t pte; 711 swp_entry_t entry; 712 713 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); 714 if (pte_swp_soft_dirty(*ptep)) 715 pte = pte_mksoft_dirty(pte); 716 717 entry = pte_to_swp_entry(*ptep); 718 if (pte_swp_uffd_wp(*ptep)) 719 pte = pte_mkuffd_wp(pte); 720 else if (is_writable_device_exclusive_entry(entry)) 721 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 722 723 /* 724 * No need to take a page reference as one was already 725 * created when the swap entry was made. 726 */ 727 if (PageAnon(page)) 728 page_add_anon_rmap(page, vma, address, false); 729 else 730 /* 731 * Currently device exclusive access only supports anonymous 732 * memory so the entry shouldn't point to a filebacked page. 733 */ 734 WARN_ON_ONCE(!PageAnon(page)); 735 736 set_pte_at(vma->vm_mm, address, ptep, pte); 737 738 if (vma->vm_flags & VM_LOCKED) 739 mlock_vma_page(page); 740 741 /* 742 * No need to invalidate - it was non-present before. However 743 * secondary CPUs may have mappings that need invalidating. 744 */ 745 update_mmu_cache(vma, address, ptep); 746 } 747 748 /* 749 * Tries to restore an exclusive pte if the page lock can be acquired without 750 * sleeping. 751 */ 752 static int 753 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma, 754 unsigned long addr) 755 { 756 swp_entry_t entry = pte_to_swp_entry(*src_pte); 757 struct page *page = pfn_swap_entry_to_page(entry); 758 759 if (trylock_page(page)) { 760 restore_exclusive_pte(vma, page, addr, src_pte); 761 unlock_page(page); 762 return 0; 763 } 764 765 return -EBUSY; 766 } 767 768 /* 769 * copy one vm_area from one task to the other. Assumes the page tables 770 * already present in the new task to be cleared in the whole range 771 * covered by this vma. 772 */ 773 774 static unsigned long 775 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 776 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma, 777 struct vm_area_struct *src_vma, unsigned long addr, int *rss) 778 { 779 unsigned long vm_flags = dst_vma->vm_flags; 780 pte_t pte = *src_pte; 781 struct page *page; 782 swp_entry_t entry = pte_to_swp_entry(pte); 783 784 if (likely(!non_swap_entry(entry))) { 785 if (swap_duplicate(entry) < 0) 786 return -EIO; 787 788 /* make sure dst_mm is on swapoff's mmlist. */ 789 if (unlikely(list_empty(&dst_mm->mmlist))) { 790 spin_lock(&mmlist_lock); 791 if (list_empty(&dst_mm->mmlist)) 792 list_add(&dst_mm->mmlist, 793 &src_mm->mmlist); 794 spin_unlock(&mmlist_lock); 795 } 796 rss[MM_SWAPENTS]++; 797 } else if (is_migration_entry(entry)) { 798 page = pfn_swap_entry_to_page(entry); 799 800 rss[mm_counter(page)]++; 801 802 if (is_writable_migration_entry(entry) && 803 is_cow_mapping(vm_flags)) { 804 /* 805 * COW mappings require pages in both 806 * parent and child to be set to read. 807 */ 808 entry = make_readable_migration_entry( 809 swp_offset(entry)); 810 pte = swp_entry_to_pte(entry); 811 if (pte_swp_soft_dirty(*src_pte)) 812 pte = pte_swp_mksoft_dirty(pte); 813 if (pte_swp_uffd_wp(*src_pte)) 814 pte = pte_swp_mkuffd_wp(pte); 815 set_pte_at(src_mm, addr, src_pte, pte); 816 } 817 } else if (is_device_private_entry(entry)) { 818 page = pfn_swap_entry_to_page(entry); 819 820 /* 821 * Update rss count even for unaddressable pages, as 822 * they should treated just like normal pages in this 823 * respect. 824 * 825 * We will likely want to have some new rss counters 826 * for unaddressable pages, at some point. But for now 827 * keep things as they are. 828 */ 829 get_page(page); 830 rss[mm_counter(page)]++; 831 page_dup_rmap(page, false); 832 833 /* 834 * We do not preserve soft-dirty information, because so 835 * far, checkpoint/restore is the only feature that 836 * requires that. And checkpoint/restore does not work 837 * when a device driver is involved (you cannot easily 838 * save and restore device driver state). 839 */ 840 if (is_writable_device_private_entry(entry) && 841 is_cow_mapping(vm_flags)) { 842 entry = make_readable_device_private_entry( 843 swp_offset(entry)); 844 pte = swp_entry_to_pte(entry); 845 if (pte_swp_uffd_wp(*src_pte)) 846 pte = pte_swp_mkuffd_wp(pte); 847 set_pte_at(src_mm, addr, src_pte, pte); 848 } 849 } else if (is_device_exclusive_entry(entry)) { 850 /* 851 * Make device exclusive entries present by restoring the 852 * original entry then copying as for a present pte. Device 853 * exclusive entries currently only support private writable 854 * (ie. COW) mappings. 855 */ 856 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags)); 857 if (try_restore_exclusive_pte(src_pte, src_vma, addr)) 858 return -EBUSY; 859 return -ENOENT; 860 } 861 if (!userfaultfd_wp(dst_vma)) 862 pte = pte_swp_clear_uffd_wp(pte); 863 set_pte_at(dst_mm, addr, dst_pte, pte); 864 return 0; 865 } 866 867 /* 868 * Copy a present and normal page if necessary. 869 * 870 * NOTE! The usual case is that this doesn't need to do 871 * anything, and can just return a positive value. That 872 * will let the caller know that it can just increase 873 * the page refcount and re-use the pte the traditional 874 * way. 875 * 876 * But _if_ we need to copy it because it needs to be 877 * pinned in the parent (and the child should get its own 878 * copy rather than just a reference to the same page), 879 * we'll do that here and return zero to let the caller 880 * know we're done. 881 * 882 * And if we need a pre-allocated page but don't yet have 883 * one, return a negative error to let the preallocation 884 * code know so that it can do so outside the page table 885 * lock. 886 */ 887 static inline int 888 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 889 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 890 struct page **prealloc, pte_t pte, struct page *page) 891 { 892 struct page *new_page; 893 894 /* 895 * What we want to do is to check whether this page may 896 * have been pinned by the parent process. If so, 897 * instead of wrprotect the pte on both sides, we copy 898 * the page immediately so that we'll always guarantee 899 * the pinned page won't be randomly replaced in the 900 * future. 901 * 902 * The page pinning checks are just "has this mm ever 903 * seen pinning", along with the (inexact) check of 904 * the page count. That might give false positives for 905 * for pinning, but it will work correctly. 906 */ 907 if (likely(!page_needs_cow_for_dma(src_vma, page))) 908 return 1; 909 910 new_page = *prealloc; 911 if (!new_page) 912 return -EAGAIN; 913 914 /* 915 * We have a prealloc page, all good! Take it 916 * over and copy the page & arm it. 917 */ 918 *prealloc = NULL; 919 copy_user_highpage(new_page, page, addr, src_vma); 920 __SetPageUptodate(new_page); 921 page_add_new_anon_rmap(new_page, dst_vma, addr, false); 922 lru_cache_add_inactive_or_unevictable(new_page, dst_vma); 923 rss[mm_counter(new_page)]++; 924 925 /* All done, just insert the new page copy in the child */ 926 pte = mk_pte(new_page, dst_vma->vm_page_prot); 927 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); 928 if (userfaultfd_pte_wp(dst_vma, *src_pte)) 929 /* Uffd-wp needs to be delivered to dest pte as well */ 930 pte = pte_wrprotect(pte_mkuffd_wp(pte)); 931 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 932 return 0; 933 } 934 935 /* 936 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page 937 * is required to copy this pte. 938 */ 939 static inline int 940 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 941 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 942 struct page **prealloc) 943 { 944 struct mm_struct *src_mm = src_vma->vm_mm; 945 unsigned long vm_flags = src_vma->vm_flags; 946 pte_t pte = *src_pte; 947 struct page *page; 948 949 page = vm_normal_page(src_vma, addr, pte); 950 if (page) { 951 int retval; 952 953 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte, 954 addr, rss, prealloc, pte, page); 955 if (retval <= 0) 956 return retval; 957 958 get_page(page); 959 page_dup_rmap(page, false); 960 rss[mm_counter(page)]++; 961 } 962 963 /* 964 * If it's a COW mapping, write protect it both 965 * in the parent and the child 966 */ 967 if (is_cow_mapping(vm_flags) && pte_write(pte)) { 968 ptep_set_wrprotect(src_mm, addr, src_pte); 969 pte = pte_wrprotect(pte); 970 } 971 972 /* 973 * If it's a shared mapping, mark it clean in 974 * the child 975 */ 976 if (vm_flags & VM_SHARED) 977 pte = pte_mkclean(pte); 978 pte = pte_mkold(pte); 979 980 if (!userfaultfd_wp(dst_vma)) 981 pte = pte_clear_uffd_wp(pte); 982 983 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 984 return 0; 985 } 986 987 static inline struct page * 988 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma, 989 unsigned long addr) 990 { 991 struct page *new_page; 992 993 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr); 994 if (!new_page) 995 return NULL; 996 997 if (mem_cgroup_charge(page_folio(new_page), src_mm, GFP_KERNEL)) { 998 put_page(new_page); 999 return NULL; 1000 } 1001 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 1002 1003 return new_page; 1004 } 1005 1006 static int 1007 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1008 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1009 unsigned long end) 1010 { 1011 struct mm_struct *dst_mm = dst_vma->vm_mm; 1012 struct mm_struct *src_mm = src_vma->vm_mm; 1013 pte_t *orig_src_pte, *orig_dst_pte; 1014 pte_t *src_pte, *dst_pte; 1015 spinlock_t *src_ptl, *dst_ptl; 1016 int progress, ret = 0; 1017 int rss[NR_MM_COUNTERS]; 1018 swp_entry_t entry = (swp_entry_t){0}; 1019 struct page *prealloc = NULL; 1020 1021 again: 1022 progress = 0; 1023 init_rss_vec(rss); 1024 1025 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 1026 if (!dst_pte) { 1027 ret = -ENOMEM; 1028 goto out; 1029 } 1030 src_pte = pte_offset_map(src_pmd, addr); 1031 src_ptl = pte_lockptr(src_mm, src_pmd); 1032 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1033 orig_src_pte = src_pte; 1034 orig_dst_pte = dst_pte; 1035 arch_enter_lazy_mmu_mode(); 1036 1037 do { 1038 /* 1039 * We are holding two locks at this point - either of them 1040 * could generate latencies in another task on another CPU. 1041 */ 1042 if (progress >= 32) { 1043 progress = 0; 1044 if (need_resched() || 1045 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 1046 break; 1047 } 1048 if (pte_none(*src_pte)) { 1049 progress++; 1050 continue; 1051 } 1052 if (unlikely(!pte_present(*src_pte))) { 1053 ret = copy_nonpresent_pte(dst_mm, src_mm, 1054 dst_pte, src_pte, 1055 dst_vma, src_vma, 1056 addr, rss); 1057 if (ret == -EIO) { 1058 entry = pte_to_swp_entry(*src_pte); 1059 break; 1060 } else if (ret == -EBUSY) { 1061 break; 1062 } else if (!ret) { 1063 progress += 8; 1064 continue; 1065 } 1066 1067 /* 1068 * Device exclusive entry restored, continue by copying 1069 * the now present pte. 1070 */ 1071 WARN_ON_ONCE(ret != -ENOENT); 1072 } 1073 /* copy_present_pte() will clear `*prealloc' if consumed */ 1074 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte, 1075 addr, rss, &prealloc); 1076 /* 1077 * If we need a pre-allocated page for this pte, drop the 1078 * locks, allocate, and try again. 1079 */ 1080 if (unlikely(ret == -EAGAIN)) 1081 break; 1082 if (unlikely(prealloc)) { 1083 /* 1084 * pre-alloc page cannot be reused by next time so as 1085 * to strictly follow mempolicy (e.g., alloc_page_vma() 1086 * will allocate page according to address). This 1087 * could only happen if one pinned pte changed. 1088 */ 1089 put_page(prealloc); 1090 prealloc = NULL; 1091 } 1092 progress += 8; 1093 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 1094 1095 arch_leave_lazy_mmu_mode(); 1096 spin_unlock(src_ptl); 1097 pte_unmap(orig_src_pte); 1098 add_mm_rss_vec(dst_mm, rss); 1099 pte_unmap_unlock(orig_dst_pte, dst_ptl); 1100 cond_resched(); 1101 1102 if (ret == -EIO) { 1103 VM_WARN_ON_ONCE(!entry.val); 1104 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { 1105 ret = -ENOMEM; 1106 goto out; 1107 } 1108 entry.val = 0; 1109 } else if (ret == -EBUSY) { 1110 goto out; 1111 } else if (ret == -EAGAIN) { 1112 prealloc = page_copy_prealloc(src_mm, src_vma, addr); 1113 if (!prealloc) 1114 return -ENOMEM; 1115 } else if (ret) { 1116 VM_WARN_ON_ONCE(1); 1117 } 1118 1119 /* We've captured and resolved the error. Reset, try again. */ 1120 ret = 0; 1121 1122 if (addr != end) 1123 goto again; 1124 out: 1125 if (unlikely(prealloc)) 1126 put_page(prealloc); 1127 return ret; 1128 } 1129 1130 static inline int 1131 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1132 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1133 unsigned long end) 1134 { 1135 struct mm_struct *dst_mm = dst_vma->vm_mm; 1136 struct mm_struct *src_mm = src_vma->vm_mm; 1137 pmd_t *src_pmd, *dst_pmd; 1138 unsigned long next; 1139 1140 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 1141 if (!dst_pmd) 1142 return -ENOMEM; 1143 src_pmd = pmd_offset(src_pud, addr); 1144 do { 1145 next = pmd_addr_end(addr, end); 1146 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 1147 || pmd_devmap(*src_pmd)) { 1148 int err; 1149 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); 1150 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd, 1151 addr, dst_vma, src_vma); 1152 if (err == -ENOMEM) 1153 return -ENOMEM; 1154 if (!err) 1155 continue; 1156 /* fall through */ 1157 } 1158 if (pmd_none_or_clear_bad(src_pmd)) 1159 continue; 1160 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, 1161 addr, next)) 1162 return -ENOMEM; 1163 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 1164 return 0; 1165 } 1166 1167 static inline int 1168 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1169 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, 1170 unsigned long end) 1171 { 1172 struct mm_struct *dst_mm = dst_vma->vm_mm; 1173 struct mm_struct *src_mm = src_vma->vm_mm; 1174 pud_t *src_pud, *dst_pud; 1175 unsigned long next; 1176 1177 dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 1178 if (!dst_pud) 1179 return -ENOMEM; 1180 src_pud = pud_offset(src_p4d, addr); 1181 do { 1182 next = pud_addr_end(addr, end); 1183 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 1184 int err; 1185 1186 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); 1187 err = copy_huge_pud(dst_mm, src_mm, 1188 dst_pud, src_pud, addr, src_vma); 1189 if (err == -ENOMEM) 1190 return -ENOMEM; 1191 if (!err) 1192 continue; 1193 /* fall through */ 1194 } 1195 if (pud_none_or_clear_bad(src_pud)) 1196 continue; 1197 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, 1198 addr, next)) 1199 return -ENOMEM; 1200 } while (dst_pud++, src_pud++, addr = next, addr != end); 1201 return 0; 1202 } 1203 1204 static inline int 1205 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1206 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, 1207 unsigned long end) 1208 { 1209 struct mm_struct *dst_mm = dst_vma->vm_mm; 1210 p4d_t *src_p4d, *dst_p4d; 1211 unsigned long next; 1212 1213 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 1214 if (!dst_p4d) 1215 return -ENOMEM; 1216 src_p4d = p4d_offset(src_pgd, addr); 1217 do { 1218 next = p4d_addr_end(addr, end); 1219 if (p4d_none_or_clear_bad(src_p4d)) 1220 continue; 1221 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, 1222 addr, next)) 1223 return -ENOMEM; 1224 } while (dst_p4d++, src_p4d++, addr = next, addr != end); 1225 return 0; 1226 } 1227 1228 int 1229 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1230 { 1231 pgd_t *src_pgd, *dst_pgd; 1232 unsigned long next; 1233 unsigned long addr = src_vma->vm_start; 1234 unsigned long end = src_vma->vm_end; 1235 struct mm_struct *dst_mm = dst_vma->vm_mm; 1236 struct mm_struct *src_mm = src_vma->vm_mm; 1237 struct mmu_notifier_range range; 1238 bool is_cow; 1239 int ret; 1240 1241 /* 1242 * Don't copy ptes where a page fault will fill them correctly. 1243 * Fork becomes much lighter when there are big shared or private 1244 * readonly mappings. The tradeoff is that copy_page_range is more 1245 * efficient than faulting. 1246 */ 1247 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 1248 !src_vma->anon_vma) 1249 return 0; 1250 1251 if (is_vm_hugetlb_page(src_vma)) 1252 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma); 1253 1254 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { 1255 /* 1256 * We do not free on error cases below as remove_vma 1257 * gets called on error from higher level routine 1258 */ 1259 ret = track_pfn_copy(src_vma); 1260 if (ret) 1261 return ret; 1262 } 1263 1264 /* 1265 * We need to invalidate the secondary MMU mappings only when 1266 * there could be a permission downgrade on the ptes of the 1267 * parent mm. And a permission downgrade will only happen if 1268 * is_cow_mapping() returns true. 1269 */ 1270 is_cow = is_cow_mapping(src_vma->vm_flags); 1271 1272 if (is_cow) { 1273 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 1274 0, src_vma, src_mm, addr, end); 1275 mmu_notifier_invalidate_range_start(&range); 1276 /* 1277 * Disabling preemption is not needed for the write side, as 1278 * the read side doesn't spin, but goes to the mmap_lock. 1279 * 1280 * Use the raw variant of the seqcount_t write API to avoid 1281 * lockdep complaining about preemptibility. 1282 */ 1283 mmap_assert_write_locked(src_mm); 1284 raw_write_seqcount_begin(&src_mm->write_protect_seq); 1285 } 1286 1287 ret = 0; 1288 dst_pgd = pgd_offset(dst_mm, addr); 1289 src_pgd = pgd_offset(src_mm, addr); 1290 do { 1291 next = pgd_addr_end(addr, end); 1292 if (pgd_none_or_clear_bad(src_pgd)) 1293 continue; 1294 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, 1295 addr, next))) { 1296 ret = -ENOMEM; 1297 break; 1298 } 1299 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1300 1301 if (is_cow) { 1302 raw_write_seqcount_end(&src_mm->write_protect_seq); 1303 mmu_notifier_invalidate_range_end(&range); 1304 } 1305 return ret; 1306 } 1307 1308 /* 1309 * Parameter block passed down to zap_pte_range in exceptional cases. 1310 */ 1311 struct zap_details { 1312 struct address_space *zap_mapping; /* Check page->mapping if set */ 1313 struct folio *single_folio; /* Locked folio to be unmapped */ 1314 }; 1315 1316 /* 1317 * We set details->zap_mapping when we want to unmap shared but keep private 1318 * pages. Return true if skip zapping this page, false otherwise. 1319 */ 1320 static inline bool 1321 zap_skip_check_mapping(struct zap_details *details, struct page *page) 1322 { 1323 if (!details || !page) 1324 return false; 1325 1326 return details->zap_mapping && 1327 (details->zap_mapping != page_rmapping(page)); 1328 } 1329 1330 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1331 struct vm_area_struct *vma, pmd_t *pmd, 1332 unsigned long addr, unsigned long end, 1333 struct zap_details *details) 1334 { 1335 struct mm_struct *mm = tlb->mm; 1336 int force_flush = 0; 1337 int rss[NR_MM_COUNTERS]; 1338 spinlock_t *ptl; 1339 pte_t *start_pte; 1340 pte_t *pte; 1341 swp_entry_t entry; 1342 1343 tlb_change_page_size(tlb, PAGE_SIZE); 1344 again: 1345 init_rss_vec(rss); 1346 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1347 pte = start_pte; 1348 flush_tlb_batched_pending(mm); 1349 arch_enter_lazy_mmu_mode(); 1350 do { 1351 pte_t ptent = *pte; 1352 if (pte_none(ptent)) 1353 continue; 1354 1355 if (need_resched()) 1356 break; 1357 1358 if (pte_present(ptent)) { 1359 struct page *page; 1360 1361 page = vm_normal_page(vma, addr, ptent); 1362 if (unlikely(zap_skip_check_mapping(details, page))) 1363 continue; 1364 ptent = ptep_get_and_clear_full(mm, addr, pte, 1365 tlb->fullmm); 1366 tlb_remove_tlb_entry(tlb, pte, addr); 1367 if (unlikely(!page)) 1368 continue; 1369 1370 if (!PageAnon(page)) { 1371 if (pte_dirty(ptent)) { 1372 force_flush = 1; 1373 set_page_dirty(page); 1374 } 1375 if (pte_young(ptent) && 1376 likely(!(vma->vm_flags & VM_SEQ_READ))) 1377 mark_page_accessed(page); 1378 } 1379 rss[mm_counter(page)]--; 1380 page_remove_rmap(page, false); 1381 if (unlikely(page_mapcount(page) < 0)) 1382 print_bad_pte(vma, addr, ptent, page); 1383 if (unlikely(__tlb_remove_page(tlb, page))) { 1384 force_flush = 1; 1385 addr += PAGE_SIZE; 1386 break; 1387 } 1388 continue; 1389 } 1390 1391 entry = pte_to_swp_entry(ptent); 1392 if (is_device_private_entry(entry) || 1393 is_device_exclusive_entry(entry)) { 1394 struct page *page = pfn_swap_entry_to_page(entry); 1395 1396 if (unlikely(zap_skip_check_mapping(details, page))) 1397 continue; 1398 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1399 rss[mm_counter(page)]--; 1400 1401 if (is_device_private_entry(entry)) 1402 page_remove_rmap(page, false); 1403 1404 put_page(page); 1405 continue; 1406 } 1407 1408 /* If details->check_mapping, we leave swap entries. */ 1409 if (unlikely(details)) 1410 continue; 1411 1412 if (!non_swap_entry(entry)) 1413 rss[MM_SWAPENTS]--; 1414 else if (is_migration_entry(entry)) { 1415 struct page *page; 1416 1417 page = pfn_swap_entry_to_page(entry); 1418 rss[mm_counter(page)]--; 1419 } 1420 if (unlikely(!free_swap_and_cache(entry))) 1421 print_bad_pte(vma, addr, ptent, NULL); 1422 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1423 } while (pte++, addr += PAGE_SIZE, addr != end); 1424 1425 add_mm_rss_vec(mm, rss); 1426 arch_leave_lazy_mmu_mode(); 1427 1428 /* Do the actual TLB flush before dropping ptl */ 1429 if (force_flush) 1430 tlb_flush_mmu_tlbonly(tlb); 1431 pte_unmap_unlock(start_pte, ptl); 1432 1433 /* 1434 * If we forced a TLB flush (either due to running out of 1435 * batch buffers or because we needed to flush dirty TLB 1436 * entries before releasing the ptl), free the batched 1437 * memory too. Restart if we didn't do everything. 1438 */ 1439 if (force_flush) { 1440 force_flush = 0; 1441 tlb_flush_mmu(tlb); 1442 } 1443 1444 if (addr != end) { 1445 cond_resched(); 1446 goto again; 1447 } 1448 1449 return addr; 1450 } 1451 1452 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1453 struct vm_area_struct *vma, pud_t *pud, 1454 unsigned long addr, unsigned long end, 1455 struct zap_details *details) 1456 { 1457 pmd_t *pmd; 1458 unsigned long next; 1459 1460 pmd = pmd_offset(pud, addr); 1461 do { 1462 next = pmd_addr_end(addr, end); 1463 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1464 if (next - addr != HPAGE_PMD_SIZE) 1465 __split_huge_pmd(vma, pmd, addr, false, NULL); 1466 else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1467 goto next; 1468 /* fall through */ 1469 } else if (details && details->single_folio && 1470 folio_test_pmd_mappable(details->single_folio) && 1471 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) { 1472 spinlock_t *ptl = pmd_lock(tlb->mm, pmd); 1473 /* 1474 * Take and drop THP pmd lock so that we cannot return 1475 * prematurely, while zap_huge_pmd() has cleared *pmd, 1476 * but not yet decremented compound_mapcount(). 1477 */ 1478 spin_unlock(ptl); 1479 } 1480 1481 /* 1482 * Here there can be other concurrent MADV_DONTNEED or 1483 * trans huge page faults running, and if the pmd is 1484 * none or trans huge it can change under us. This is 1485 * because MADV_DONTNEED holds the mmap_lock in read 1486 * mode. 1487 */ 1488 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1489 goto next; 1490 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1491 next: 1492 cond_resched(); 1493 } while (pmd++, addr = next, addr != end); 1494 1495 return addr; 1496 } 1497 1498 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1499 struct vm_area_struct *vma, p4d_t *p4d, 1500 unsigned long addr, unsigned long end, 1501 struct zap_details *details) 1502 { 1503 pud_t *pud; 1504 unsigned long next; 1505 1506 pud = pud_offset(p4d, addr); 1507 do { 1508 next = pud_addr_end(addr, end); 1509 if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1510 if (next - addr != HPAGE_PUD_SIZE) { 1511 mmap_assert_locked(tlb->mm); 1512 split_huge_pud(vma, pud, addr); 1513 } else if (zap_huge_pud(tlb, vma, pud, addr)) 1514 goto next; 1515 /* fall through */ 1516 } 1517 if (pud_none_or_clear_bad(pud)) 1518 continue; 1519 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1520 next: 1521 cond_resched(); 1522 } while (pud++, addr = next, addr != end); 1523 1524 return addr; 1525 } 1526 1527 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1528 struct vm_area_struct *vma, pgd_t *pgd, 1529 unsigned long addr, unsigned long end, 1530 struct zap_details *details) 1531 { 1532 p4d_t *p4d; 1533 unsigned long next; 1534 1535 p4d = p4d_offset(pgd, addr); 1536 do { 1537 next = p4d_addr_end(addr, end); 1538 if (p4d_none_or_clear_bad(p4d)) 1539 continue; 1540 next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1541 } while (p4d++, addr = next, addr != end); 1542 1543 return addr; 1544 } 1545 1546 void unmap_page_range(struct mmu_gather *tlb, 1547 struct vm_area_struct *vma, 1548 unsigned long addr, unsigned long end, 1549 struct zap_details *details) 1550 { 1551 pgd_t *pgd; 1552 unsigned long next; 1553 1554 BUG_ON(addr >= end); 1555 tlb_start_vma(tlb, vma); 1556 pgd = pgd_offset(vma->vm_mm, addr); 1557 do { 1558 next = pgd_addr_end(addr, end); 1559 if (pgd_none_or_clear_bad(pgd)) 1560 continue; 1561 next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 1562 } while (pgd++, addr = next, addr != end); 1563 tlb_end_vma(tlb, vma); 1564 } 1565 1566 1567 static void unmap_single_vma(struct mmu_gather *tlb, 1568 struct vm_area_struct *vma, unsigned long start_addr, 1569 unsigned long end_addr, 1570 struct zap_details *details) 1571 { 1572 unsigned long start = max(vma->vm_start, start_addr); 1573 unsigned long end; 1574 1575 if (start >= vma->vm_end) 1576 return; 1577 end = min(vma->vm_end, end_addr); 1578 if (end <= vma->vm_start) 1579 return; 1580 1581 if (vma->vm_file) 1582 uprobe_munmap(vma, start, end); 1583 1584 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1585 untrack_pfn(vma, 0, 0); 1586 1587 if (start != end) { 1588 if (unlikely(is_vm_hugetlb_page(vma))) { 1589 /* 1590 * It is undesirable to test vma->vm_file as it 1591 * should be non-null for valid hugetlb area. 1592 * However, vm_file will be NULL in the error 1593 * cleanup path of mmap_region. When 1594 * hugetlbfs ->mmap method fails, 1595 * mmap_region() nullifies vma->vm_file 1596 * before calling this function to clean up. 1597 * Since no pte has actually been setup, it is 1598 * safe to do nothing in this case. 1599 */ 1600 if (vma->vm_file) { 1601 i_mmap_lock_write(vma->vm_file->f_mapping); 1602 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1603 i_mmap_unlock_write(vma->vm_file->f_mapping); 1604 } 1605 } else 1606 unmap_page_range(tlb, vma, start, end, details); 1607 } 1608 } 1609 1610 /** 1611 * unmap_vmas - unmap a range of memory covered by a list of vma's 1612 * @tlb: address of the caller's struct mmu_gather 1613 * @vma: the starting vma 1614 * @start_addr: virtual address at which to start unmapping 1615 * @end_addr: virtual address at which to end unmapping 1616 * 1617 * Unmap all pages in the vma list. 1618 * 1619 * Only addresses between `start' and `end' will be unmapped. 1620 * 1621 * The VMA list must be sorted in ascending virtual address order. 1622 * 1623 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1624 * range after unmap_vmas() returns. So the only responsibility here is to 1625 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1626 * drops the lock and schedules. 1627 */ 1628 void unmap_vmas(struct mmu_gather *tlb, 1629 struct vm_area_struct *vma, unsigned long start_addr, 1630 unsigned long end_addr) 1631 { 1632 struct mmu_notifier_range range; 1633 1634 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, 1635 start_addr, end_addr); 1636 mmu_notifier_invalidate_range_start(&range); 1637 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1638 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1639 mmu_notifier_invalidate_range_end(&range); 1640 } 1641 1642 /** 1643 * zap_page_range - remove user pages in a given range 1644 * @vma: vm_area_struct holding the applicable pages 1645 * @start: starting address of pages to zap 1646 * @size: number of bytes to zap 1647 * 1648 * Caller must protect the VMA list 1649 */ 1650 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1651 unsigned long size) 1652 { 1653 struct mmu_notifier_range range; 1654 struct mmu_gather tlb; 1655 1656 lru_add_drain(); 1657 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1658 start, start + size); 1659 tlb_gather_mmu(&tlb, vma->vm_mm); 1660 update_hiwater_rss(vma->vm_mm); 1661 mmu_notifier_invalidate_range_start(&range); 1662 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next) 1663 unmap_single_vma(&tlb, vma, start, range.end, NULL); 1664 mmu_notifier_invalidate_range_end(&range); 1665 tlb_finish_mmu(&tlb); 1666 } 1667 1668 /** 1669 * zap_page_range_single - remove user pages in a given range 1670 * @vma: vm_area_struct holding the applicable pages 1671 * @address: starting address of pages to zap 1672 * @size: number of bytes to zap 1673 * @details: details of shared cache invalidation 1674 * 1675 * The range must fit into one VMA. 1676 */ 1677 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1678 unsigned long size, struct zap_details *details) 1679 { 1680 struct mmu_notifier_range range; 1681 struct mmu_gather tlb; 1682 1683 lru_add_drain(); 1684 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1685 address, address + size); 1686 tlb_gather_mmu(&tlb, vma->vm_mm); 1687 update_hiwater_rss(vma->vm_mm); 1688 mmu_notifier_invalidate_range_start(&range); 1689 unmap_single_vma(&tlb, vma, address, range.end, details); 1690 mmu_notifier_invalidate_range_end(&range); 1691 tlb_finish_mmu(&tlb); 1692 } 1693 1694 /** 1695 * zap_vma_ptes - remove ptes mapping the vma 1696 * @vma: vm_area_struct holding ptes to be zapped 1697 * @address: starting address of pages to zap 1698 * @size: number of bytes to zap 1699 * 1700 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1701 * 1702 * The entire address range must be fully contained within the vma. 1703 * 1704 */ 1705 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1706 unsigned long size) 1707 { 1708 if (address < vma->vm_start || address + size > vma->vm_end || 1709 !(vma->vm_flags & VM_PFNMAP)) 1710 return; 1711 1712 zap_page_range_single(vma, address, size, NULL); 1713 } 1714 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1715 1716 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) 1717 { 1718 pgd_t *pgd; 1719 p4d_t *p4d; 1720 pud_t *pud; 1721 pmd_t *pmd; 1722 1723 pgd = pgd_offset(mm, addr); 1724 p4d = p4d_alloc(mm, pgd, addr); 1725 if (!p4d) 1726 return NULL; 1727 pud = pud_alloc(mm, p4d, addr); 1728 if (!pud) 1729 return NULL; 1730 pmd = pmd_alloc(mm, pud, addr); 1731 if (!pmd) 1732 return NULL; 1733 1734 VM_BUG_ON(pmd_trans_huge(*pmd)); 1735 return pmd; 1736 } 1737 1738 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1739 spinlock_t **ptl) 1740 { 1741 pmd_t *pmd = walk_to_pmd(mm, addr); 1742 1743 if (!pmd) 1744 return NULL; 1745 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1746 } 1747 1748 static int validate_page_before_insert(struct page *page) 1749 { 1750 if (PageAnon(page) || PageSlab(page) || page_has_type(page)) 1751 return -EINVAL; 1752 flush_dcache_page(page); 1753 return 0; 1754 } 1755 1756 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte, 1757 unsigned long addr, struct page *page, pgprot_t prot) 1758 { 1759 if (!pte_none(*pte)) 1760 return -EBUSY; 1761 /* Ok, finally just insert the thing.. */ 1762 get_page(page); 1763 inc_mm_counter_fast(mm, mm_counter_file(page)); 1764 page_add_file_rmap(page, false); 1765 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1766 return 0; 1767 } 1768 1769 /* 1770 * This is the old fallback for page remapping. 1771 * 1772 * For historical reasons, it only allows reserved pages. Only 1773 * old drivers should use this, and they needed to mark their 1774 * pages reserved for the old functions anyway. 1775 */ 1776 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1777 struct page *page, pgprot_t prot) 1778 { 1779 struct mm_struct *mm = vma->vm_mm; 1780 int retval; 1781 pte_t *pte; 1782 spinlock_t *ptl; 1783 1784 retval = validate_page_before_insert(page); 1785 if (retval) 1786 goto out; 1787 retval = -ENOMEM; 1788 pte = get_locked_pte(mm, addr, &ptl); 1789 if (!pte) 1790 goto out; 1791 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot); 1792 pte_unmap_unlock(pte, ptl); 1793 out: 1794 return retval; 1795 } 1796 1797 #ifdef pte_index 1798 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte, 1799 unsigned long addr, struct page *page, pgprot_t prot) 1800 { 1801 int err; 1802 1803 if (!page_count(page)) 1804 return -EINVAL; 1805 err = validate_page_before_insert(page); 1806 if (err) 1807 return err; 1808 return insert_page_into_pte_locked(mm, pte, addr, page, prot); 1809 } 1810 1811 /* insert_pages() amortizes the cost of spinlock operations 1812 * when inserting pages in a loop. Arch *must* define pte_index. 1813 */ 1814 static int insert_pages(struct vm_area_struct *vma, unsigned long addr, 1815 struct page **pages, unsigned long *num, pgprot_t prot) 1816 { 1817 pmd_t *pmd = NULL; 1818 pte_t *start_pte, *pte; 1819 spinlock_t *pte_lock; 1820 struct mm_struct *const mm = vma->vm_mm; 1821 unsigned long curr_page_idx = 0; 1822 unsigned long remaining_pages_total = *num; 1823 unsigned long pages_to_write_in_pmd; 1824 int ret; 1825 more: 1826 ret = -EFAULT; 1827 pmd = walk_to_pmd(mm, addr); 1828 if (!pmd) 1829 goto out; 1830 1831 pages_to_write_in_pmd = min_t(unsigned long, 1832 remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); 1833 1834 /* Allocate the PTE if necessary; takes PMD lock once only. */ 1835 ret = -ENOMEM; 1836 if (pte_alloc(mm, pmd)) 1837 goto out; 1838 1839 while (pages_to_write_in_pmd) { 1840 int pte_idx = 0; 1841 const int batch_size = min_t(int, pages_to_write_in_pmd, 8); 1842 1843 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); 1844 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { 1845 int err = insert_page_in_batch_locked(mm, pte, 1846 addr, pages[curr_page_idx], prot); 1847 if (unlikely(err)) { 1848 pte_unmap_unlock(start_pte, pte_lock); 1849 ret = err; 1850 remaining_pages_total -= pte_idx; 1851 goto out; 1852 } 1853 addr += PAGE_SIZE; 1854 ++curr_page_idx; 1855 } 1856 pte_unmap_unlock(start_pte, pte_lock); 1857 pages_to_write_in_pmd -= batch_size; 1858 remaining_pages_total -= batch_size; 1859 } 1860 if (remaining_pages_total) 1861 goto more; 1862 ret = 0; 1863 out: 1864 *num = remaining_pages_total; 1865 return ret; 1866 } 1867 #endif /* ifdef pte_index */ 1868 1869 /** 1870 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. 1871 * @vma: user vma to map to 1872 * @addr: target start user address of these pages 1873 * @pages: source kernel pages 1874 * @num: in: number of pages to map. out: number of pages that were *not* 1875 * mapped. (0 means all pages were successfully mapped). 1876 * 1877 * Preferred over vm_insert_page() when inserting multiple pages. 1878 * 1879 * In case of error, we may have mapped a subset of the provided 1880 * pages. It is the caller's responsibility to account for this case. 1881 * 1882 * The same restrictions apply as in vm_insert_page(). 1883 */ 1884 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, 1885 struct page **pages, unsigned long *num) 1886 { 1887 #ifdef pte_index 1888 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; 1889 1890 if (addr < vma->vm_start || end_addr >= vma->vm_end) 1891 return -EFAULT; 1892 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1893 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1894 BUG_ON(vma->vm_flags & VM_PFNMAP); 1895 vma->vm_flags |= VM_MIXEDMAP; 1896 } 1897 /* Defer page refcount checking till we're about to map that page. */ 1898 return insert_pages(vma, addr, pages, num, vma->vm_page_prot); 1899 #else 1900 unsigned long idx = 0, pgcount = *num; 1901 int err = -EINVAL; 1902 1903 for (; idx < pgcount; ++idx) { 1904 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]); 1905 if (err) 1906 break; 1907 } 1908 *num = pgcount - idx; 1909 return err; 1910 #endif /* ifdef pte_index */ 1911 } 1912 EXPORT_SYMBOL(vm_insert_pages); 1913 1914 /** 1915 * vm_insert_page - insert single page into user vma 1916 * @vma: user vma to map to 1917 * @addr: target user address of this page 1918 * @page: source kernel page 1919 * 1920 * This allows drivers to insert individual pages they've allocated 1921 * into a user vma. 1922 * 1923 * The page has to be a nice clean _individual_ kernel allocation. 1924 * If you allocate a compound page, you need to have marked it as 1925 * such (__GFP_COMP), or manually just split the page up yourself 1926 * (see split_page()). 1927 * 1928 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1929 * took an arbitrary page protection parameter. This doesn't allow 1930 * that. Your vma protection will have to be set up correctly, which 1931 * means that if you want a shared writable mapping, you'd better 1932 * ask for a shared writable mapping! 1933 * 1934 * The page does not need to be reserved. 1935 * 1936 * Usually this function is called from f_op->mmap() handler 1937 * under mm->mmap_lock write-lock, so it can change vma->vm_flags. 1938 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1939 * function from other places, for example from page-fault handler. 1940 * 1941 * Return: %0 on success, negative error code otherwise. 1942 */ 1943 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1944 struct page *page) 1945 { 1946 if (addr < vma->vm_start || addr >= vma->vm_end) 1947 return -EFAULT; 1948 if (!page_count(page)) 1949 return -EINVAL; 1950 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1951 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1952 BUG_ON(vma->vm_flags & VM_PFNMAP); 1953 vma->vm_flags |= VM_MIXEDMAP; 1954 } 1955 return insert_page(vma, addr, page, vma->vm_page_prot); 1956 } 1957 EXPORT_SYMBOL(vm_insert_page); 1958 1959 /* 1960 * __vm_map_pages - maps range of kernel pages into user vma 1961 * @vma: user vma to map to 1962 * @pages: pointer to array of source kernel pages 1963 * @num: number of pages in page array 1964 * @offset: user's requested vm_pgoff 1965 * 1966 * This allows drivers to map range of kernel pages into a user vma. 1967 * 1968 * Return: 0 on success and error code otherwise. 1969 */ 1970 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1971 unsigned long num, unsigned long offset) 1972 { 1973 unsigned long count = vma_pages(vma); 1974 unsigned long uaddr = vma->vm_start; 1975 int ret, i; 1976 1977 /* Fail if the user requested offset is beyond the end of the object */ 1978 if (offset >= num) 1979 return -ENXIO; 1980 1981 /* Fail if the user requested size exceeds available object size */ 1982 if (count > num - offset) 1983 return -ENXIO; 1984 1985 for (i = 0; i < count; i++) { 1986 ret = vm_insert_page(vma, uaddr, pages[offset + i]); 1987 if (ret < 0) 1988 return ret; 1989 uaddr += PAGE_SIZE; 1990 } 1991 1992 return 0; 1993 } 1994 1995 /** 1996 * vm_map_pages - maps range of kernel pages starts with non zero offset 1997 * @vma: user vma to map to 1998 * @pages: pointer to array of source kernel pages 1999 * @num: number of pages in page array 2000 * 2001 * Maps an object consisting of @num pages, catering for the user's 2002 * requested vm_pgoff 2003 * 2004 * If we fail to insert any page into the vma, the function will return 2005 * immediately leaving any previously inserted pages present. Callers 2006 * from the mmap handler may immediately return the error as their caller 2007 * will destroy the vma, removing any successfully inserted pages. Other 2008 * callers should make their own arrangements for calling unmap_region(). 2009 * 2010 * Context: Process context. Called by mmap handlers. 2011 * Return: 0 on success and error code otherwise. 2012 */ 2013 int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 2014 unsigned long num) 2015 { 2016 return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 2017 } 2018 EXPORT_SYMBOL(vm_map_pages); 2019 2020 /** 2021 * vm_map_pages_zero - map range of kernel pages starts with zero offset 2022 * @vma: user vma to map to 2023 * @pages: pointer to array of source kernel pages 2024 * @num: number of pages in page array 2025 * 2026 * Similar to vm_map_pages(), except that it explicitly sets the offset 2027 * to 0. This function is intended for the drivers that did not consider 2028 * vm_pgoff. 2029 * 2030 * Context: Process context. Called by mmap handlers. 2031 * Return: 0 on success and error code otherwise. 2032 */ 2033 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 2034 unsigned long num) 2035 { 2036 return __vm_map_pages(vma, pages, num, 0); 2037 } 2038 EXPORT_SYMBOL(vm_map_pages_zero); 2039 2040 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2041 pfn_t pfn, pgprot_t prot, bool mkwrite) 2042 { 2043 struct mm_struct *mm = vma->vm_mm; 2044 pte_t *pte, entry; 2045 spinlock_t *ptl; 2046 2047 pte = get_locked_pte(mm, addr, &ptl); 2048 if (!pte) 2049 return VM_FAULT_OOM; 2050 if (!pte_none(*pte)) { 2051 if (mkwrite) { 2052 /* 2053 * For read faults on private mappings the PFN passed 2054 * in may not match the PFN we have mapped if the 2055 * mapped PFN is a writeable COW page. In the mkwrite 2056 * case we are creating a writable PTE for a shared 2057 * mapping and we expect the PFNs to match. If they 2058 * don't match, we are likely racing with block 2059 * allocation and mapping invalidation so just skip the 2060 * update. 2061 */ 2062 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) { 2063 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte))); 2064 goto out_unlock; 2065 } 2066 entry = pte_mkyoung(*pte); 2067 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2068 if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 2069 update_mmu_cache(vma, addr, pte); 2070 } 2071 goto out_unlock; 2072 } 2073 2074 /* Ok, finally just insert the thing.. */ 2075 if (pfn_t_devmap(pfn)) 2076 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 2077 else 2078 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 2079 2080 if (mkwrite) { 2081 entry = pte_mkyoung(entry); 2082 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2083 } 2084 2085 set_pte_at(mm, addr, pte, entry); 2086 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 2087 2088 out_unlock: 2089 pte_unmap_unlock(pte, ptl); 2090 return VM_FAULT_NOPAGE; 2091 } 2092 2093 /** 2094 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 2095 * @vma: user vma to map to 2096 * @addr: target user address of this page 2097 * @pfn: source kernel pfn 2098 * @pgprot: pgprot flags for the inserted page 2099 * 2100 * This is exactly like vmf_insert_pfn(), except that it allows drivers 2101 * to override pgprot on a per-page basis. 2102 * 2103 * This only makes sense for IO mappings, and it makes no sense for 2104 * COW mappings. In general, using multiple vmas is preferable; 2105 * vmf_insert_pfn_prot should only be used if using multiple VMAs is 2106 * impractical. 2107 * 2108 * See vmf_insert_mixed_prot() for a discussion of the implication of using 2109 * a value of @pgprot different from that of @vma->vm_page_prot. 2110 * 2111 * Context: Process context. May allocate using %GFP_KERNEL. 2112 * Return: vm_fault_t value. 2113 */ 2114 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 2115 unsigned long pfn, pgprot_t pgprot) 2116 { 2117 /* 2118 * Technically, architectures with pte_special can avoid all these 2119 * restrictions (same for remap_pfn_range). However we would like 2120 * consistency in testing and feature parity among all, so we should 2121 * try to keep these invariants in place for everybody. 2122 */ 2123 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 2124 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 2125 (VM_PFNMAP|VM_MIXEDMAP)); 2126 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 2127 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 2128 2129 if (addr < vma->vm_start || addr >= vma->vm_end) 2130 return VM_FAULT_SIGBUS; 2131 2132 if (!pfn_modify_allowed(pfn, pgprot)) 2133 return VM_FAULT_SIGBUS; 2134 2135 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 2136 2137 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 2138 false); 2139 } 2140 EXPORT_SYMBOL(vmf_insert_pfn_prot); 2141 2142 /** 2143 * vmf_insert_pfn - insert single pfn into user vma 2144 * @vma: user vma to map to 2145 * @addr: target user address of this page 2146 * @pfn: source kernel pfn 2147 * 2148 * Similar to vm_insert_page, this allows drivers to insert individual pages 2149 * they've allocated into a user vma. Same comments apply. 2150 * 2151 * This function should only be called from a vm_ops->fault handler, and 2152 * in that case the handler should return the result of this function. 2153 * 2154 * vma cannot be a COW mapping. 2155 * 2156 * As this is called only for pages that do not currently exist, we 2157 * do not need to flush old virtual caches or the TLB. 2158 * 2159 * Context: Process context. May allocate using %GFP_KERNEL. 2160 * Return: vm_fault_t value. 2161 */ 2162 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2163 unsigned long pfn) 2164 { 2165 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 2166 } 2167 EXPORT_SYMBOL(vmf_insert_pfn); 2168 2169 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 2170 { 2171 /* these checks mirror the abort conditions in vm_normal_page */ 2172 if (vma->vm_flags & VM_MIXEDMAP) 2173 return true; 2174 if (pfn_t_devmap(pfn)) 2175 return true; 2176 if (pfn_t_special(pfn)) 2177 return true; 2178 if (is_zero_pfn(pfn_t_to_pfn(pfn))) 2179 return true; 2180 return false; 2181 } 2182 2183 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 2184 unsigned long addr, pfn_t pfn, pgprot_t pgprot, 2185 bool mkwrite) 2186 { 2187 int err; 2188 2189 BUG_ON(!vm_mixed_ok(vma, pfn)); 2190 2191 if (addr < vma->vm_start || addr >= vma->vm_end) 2192 return VM_FAULT_SIGBUS; 2193 2194 track_pfn_insert(vma, &pgprot, pfn); 2195 2196 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 2197 return VM_FAULT_SIGBUS; 2198 2199 /* 2200 * If we don't have pte special, then we have to use the pfn_valid() 2201 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2202 * refcount the page if pfn_valid is true (hence insert_page rather 2203 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 2204 * without pte special, it would there be refcounted as a normal page. 2205 */ 2206 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 2207 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 2208 struct page *page; 2209 2210 /* 2211 * At this point we are committed to insert_page() 2212 * regardless of whether the caller specified flags that 2213 * result in pfn_t_has_page() == false. 2214 */ 2215 page = pfn_to_page(pfn_t_to_pfn(pfn)); 2216 err = insert_page(vma, addr, page, pgprot); 2217 } else { 2218 return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 2219 } 2220 2221 if (err == -ENOMEM) 2222 return VM_FAULT_OOM; 2223 if (err < 0 && err != -EBUSY) 2224 return VM_FAULT_SIGBUS; 2225 2226 return VM_FAULT_NOPAGE; 2227 } 2228 2229 /** 2230 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot 2231 * @vma: user vma to map to 2232 * @addr: target user address of this page 2233 * @pfn: source kernel pfn 2234 * @pgprot: pgprot flags for the inserted page 2235 * 2236 * This is exactly like vmf_insert_mixed(), except that it allows drivers 2237 * to override pgprot on a per-page basis. 2238 * 2239 * Typically this function should be used by drivers to set caching- and 2240 * encryption bits different than those of @vma->vm_page_prot, because 2241 * the caching- or encryption mode may not be known at mmap() time. 2242 * This is ok as long as @vma->vm_page_prot is not used by the core vm 2243 * to set caching and encryption bits for those vmas (except for COW pages). 2244 * This is ensured by core vm only modifying these page table entries using 2245 * functions that don't touch caching- or encryption bits, using pte_modify() 2246 * if needed. (See for example mprotect()). 2247 * Also when new page-table entries are created, this is only done using the 2248 * fault() callback, and never using the value of vma->vm_page_prot, 2249 * except for page-table entries that point to anonymous pages as the result 2250 * of COW. 2251 * 2252 * Context: Process context. May allocate using %GFP_KERNEL. 2253 * Return: vm_fault_t value. 2254 */ 2255 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr, 2256 pfn_t pfn, pgprot_t pgprot) 2257 { 2258 return __vm_insert_mixed(vma, addr, pfn, pgprot, false); 2259 } 2260 EXPORT_SYMBOL(vmf_insert_mixed_prot); 2261 2262 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 2263 pfn_t pfn) 2264 { 2265 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false); 2266 } 2267 EXPORT_SYMBOL(vmf_insert_mixed); 2268 2269 /* 2270 * If the insertion of PTE failed because someone else already added a 2271 * different entry in the mean time, we treat that as success as we assume 2272 * the same entry was actually inserted. 2273 */ 2274 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 2275 unsigned long addr, pfn_t pfn) 2276 { 2277 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true); 2278 } 2279 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 2280 2281 /* 2282 * maps a range of physical memory into the requested pages. the old 2283 * mappings are removed. any references to nonexistent pages results 2284 * in null mappings (currently treated as "copy-on-access") 2285 */ 2286 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 2287 unsigned long addr, unsigned long end, 2288 unsigned long pfn, pgprot_t prot) 2289 { 2290 pte_t *pte, *mapped_pte; 2291 spinlock_t *ptl; 2292 int err = 0; 2293 2294 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 2295 if (!pte) 2296 return -ENOMEM; 2297 arch_enter_lazy_mmu_mode(); 2298 do { 2299 BUG_ON(!pte_none(*pte)); 2300 if (!pfn_modify_allowed(pfn, prot)) { 2301 err = -EACCES; 2302 break; 2303 } 2304 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 2305 pfn++; 2306 } while (pte++, addr += PAGE_SIZE, addr != end); 2307 arch_leave_lazy_mmu_mode(); 2308 pte_unmap_unlock(mapped_pte, ptl); 2309 return err; 2310 } 2311 2312 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 2313 unsigned long addr, unsigned long end, 2314 unsigned long pfn, pgprot_t prot) 2315 { 2316 pmd_t *pmd; 2317 unsigned long next; 2318 int err; 2319 2320 pfn -= addr >> PAGE_SHIFT; 2321 pmd = pmd_alloc(mm, pud, addr); 2322 if (!pmd) 2323 return -ENOMEM; 2324 VM_BUG_ON(pmd_trans_huge(*pmd)); 2325 do { 2326 next = pmd_addr_end(addr, end); 2327 err = remap_pte_range(mm, pmd, addr, next, 2328 pfn + (addr >> PAGE_SHIFT), prot); 2329 if (err) 2330 return err; 2331 } while (pmd++, addr = next, addr != end); 2332 return 0; 2333 } 2334 2335 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 2336 unsigned long addr, unsigned long end, 2337 unsigned long pfn, pgprot_t prot) 2338 { 2339 pud_t *pud; 2340 unsigned long next; 2341 int err; 2342 2343 pfn -= addr >> PAGE_SHIFT; 2344 pud = pud_alloc(mm, p4d, addr); 2345 if (!pud) 2346 return -ENOMEM; 2347 do { 2348 next = pud_addr_end(addr, end); 2349 err = remap_pmd_range(mm, pud, addr, next, 2350 pfn + (addr >> PAGE_SHIFT), prot); 2351 if (err) 2352 return err; 2353 } while (pud++, addr = next, addr != end); 2354 return 0; 2355 } 2356 2357 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2358 unsigned long addr, unsigned long end, 2359 unsigned long pfn, pgprot_t prot) 2360 { 2361 p4d_t *p4d; 2362 unsigned long next; 2363 int err; 2364 2365 pfn -= addr >> PAGE_SHIFT; 2366 p4d = p4d_alloc(mm, pgd, addr); 2367 if (!p4d) 2368 return -ENOMEM; 2369 do { 2370 next = p4d_addr_end(addr, end); 2371 err = remap_pud_range(mm, p4d, addr, next, 2372 pfn + (addr >> PAGE_SHIFT), prot); 2373 if (err) 2374 return err; 2375 } while (p4d++, addr = next, addr != end); 2376 return 0; 2377 } 2378 2379 /* 2380 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller 2381 * must have pre-validated the caching bits of the pgprot_t. 2382 */ 2383 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr, 2384 unsigned long pfn, unsigned long size, pgprot_t prot) 2385 { 2386 pgd_t *pgd; 2387 unsigned long next; 2388 unsigned long end = addr + PAGE_ALIGN(size); 2389 struct mm_struct *mm = vma->vm_mm; 2390 int err; 2391 2392 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) 2393 return -EINVAL; 2394 2395 /* 2396 * Physically remapped pages are special. Tell the 2397 * rest of the world about it: 2398 * VM_IO tells people not to look at these pages 2399 * (accesses can have side effects). 2400 * VM_PFNMAP tells the core MM that the base pages are just 2401 * raw PFN mappings, and do not have a "struct page" associated 2402 * with them. 2403 * VM_DONTEXPAND 2404 * Disable vma merging and expanding with mremap(). 2405 * VM_DONTDUMP 2406 * Omit vma from core dump, even when VM_IO turned off. 2407 * 2408 * There's a horrible special case to handle copy-on-write 2409 * behaviour that some programs depend on. We mark the "original" 2410 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2411 * See vm_normal_page() for details. 2412 */ 2413 if (is_cow_mapping(vma->vm_flags)) { 2414 if (addr != vma->vm_start || end != vma->vm_end) 2415 return -EINVAL; 2416 vma->vm_pgoff = pfn; 2417 } 2418 2419 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 2420 2421 BUG_ON(addr >= end); 2422 pfn -= addr >> PAGE_SHIFT; 2423 pgd = pgd_offset(mm, addr); 2424 flush_cache_range(vma, addr, end); 2425 do { 2426 next = pgd_addr_end(addr, end); 2427 err = remap_p4d_range(mm, pgd, addr, next, 2428 pfn + (addr >> PAGE_SHIFT), prot); 2429 if (err) 2430 return err; 2431 } while (pgd++, addr = next, addr != end); 2432 2433 return 0; 2434 } 2435 2436 /** 2437 * remap_pfn_range - remap kernel memory to userspace 2438 * @vma: user vma to map to 2439 * @addr: target page aligned user address to start at 2440 * @pfn: page frame number of kernel physical memory address 2441 * @size: size of mapping area 2442 * @prot: page protection flags for this mapping 2443 * 2444 * Note: this is only safe if the mm semaphore is held when called. 2445 * 2446 * Return: %0 on success, negative error code otherwise. 2447 */ 2448 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 2449 unsigned long pfn, unsigned long size, pgprot_t prot) 2450 { 2451 int err; 2452 2453 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 2454 if (err) 2455 return -EINVAL; 2456 2457 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot); 2458 if (err) 2459 untrack_pfn(vma, pfn, PAGE_ALIGN(size)); 2460 return err; 2461 } 2462 EXPORT_SYMBOL(remap_pfn_range); 2463 2464 /** 2465 * vm_iomap_memory - remap memory to userspace 2466 * @vma: user vma to map to 2467 * @start: start of the physical memory to be mapped 2468 * @len: size of area 2469 * 2470 * This is a simplified io_remap_pfn_range() for common driver use. The 2471 * driver just needs to give us the physical memory range to be mapped, 2472 * we'll figure out the rest from the vma information. 2473 * 2474 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 2475 * whatever write-combining details or similar. 2476 * 2477 * Return: %0 on success, negative error code otherwise. 2478 */ 2479 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 2480 { 2481 unsigned long vm_len, pfn, pages; 2482 2483 /* Check that the physical memory area passed in looks valid */ 2484 if (start + len < start) 2485 return -EINVAL; 2486 /* 2487 * You *really* shouldn't map things that aren't page-aligned, 2488 * but we've historically allowed it because IO memory might 2489 * just have smaller alignment. 2490 */ 2491 len += start & ~PAGE_MASK; 2492 pfn = start >> PAGE_SHIFT; 2493 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 2494 if (pfn + pages < pfn) 2495 return -EINVAL; 2496 2497 /* We start the mapping 'vm_pgoff' pages into the area */ 2498 if (vma->vm_pgoff > pages) 2499 return -EINVAL; 2500 pfn += vma->vm_pgoff; 2501 pages -= vma->vm_pgoff; 2502 2503 /* Can we fit all of the mapping? */ 2504 vm_len = vma->vm_end - vma->vm_start; 2505 if (vm_len >> PAGE_SHIFT > pages) 2506 return -EINVAL; 2507 2508 /* Ok, let it rip */ 2509 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 2510 } 2511 EXPORT_SYMBOL(vm_iomap_memory); 2512 2513 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2514 unsigned long addr, unsigned long end, 2515 pte_fn_t fn, void *data, bool create, 2516 pgtbl_mod_mask *mask) 2517 { 2518 pte_t *pte, *mapped_pte; 2519 int err = 0; 2520 spinlock_t *ptl; 2521 2522 if (create) { 2523 mapped_pte = pte = (mm == &init_mm) ? 2524 pte_alloc_kernel_track(pmd, addr, mask) : 2525 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2526 if (!pte) 2527 return -ENOMEM; 2528 } else { 2529 mapped_pte = pte = (mm == &init_mm) ? 2530 pte_offset_kernel(pmd, addr) : 2531 pte_offset_map_lock(mm, pmd, addr, &ptl); 2532 } 2533 2534 BUG_ON(pmd_huge(*pmd)); 2535 2536 arch_enter_lazy_mmu_mode(); 2537 2538 if (fn) { 2539 do { 2540 if (create || !pte_none(*pte)) { 2541 err = fn(pte++, addr, data); 2542 if (err) 2543 break; 2544 } 2545 } while (addr += PAGE_SIZE, addr != end); 2546 } 2547 *mask |= PGTBL_PTE_MODIFIED; 2548 2549 arch_leave_lazy_mmu_mode(); 2550 2551 if (mm != &init_mm) 2552 pte_unmap_unlock(mapped_pte, ptl); 2553 return err; 2554 } 2555 2556 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2557 unsigned long addr, unsigned long end, 2558 pte_fn_t fn, void *data, bool create, 2559 pgtbl_mod_mask *mask) 2560 { 2561 pmd_t *pmd; 2562 unsigned long next; 2563 int err = 0; 2564 2565 BUG_ON(pud_huge(*pud)); 2566 2567 if (create) { 2568 pmd = pmd_alloc_track(mm, pud, addr, mask); 2569 if (!pmd) 2570 return -ENOMEM; 2571 } else { 2572 pmd = pmd_offset(pud, addr); 2573 } 2574 do { 2575 next = pmd_addr_end(addr, end); 2576 if (pmd_none(*pmd) && !create) 2577 continue; 2578 if (WARN_ON_ONCE(pmd_leaf(*pmd))) 2579 return -EINVAL; 2580 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) { 2581 if (!create) 2582 continue; 2583 pmd_clear_bad(pmd); 2584 } 2585 err = apply_to_pte_range(mm, pmd, addr, next, 2586 fn, data, create, mask); 2587 if (err) 2588 break; 2589 } while (pmd++, addr = next, addr != end); 2590 2591 return err; 2592 } 2593 2594 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2595 unsigned long addr, unsigned long end, 2596 pte_fn_t fn, void *data, bool create, 2597 pgtbl_mod_mask *mask) 2598 { 2599 pud_t *pud; 2600 unsigned long next; 2601 int err = 0; 2602 2603 if (create) { 2604 pud = pud_alloc_track(mm, p4d, addr, mask); 2605 if (!pud) 2606 return -ENOMEM; 2607 } else { 2608 pud = pud_offset(p4d, addr); 2609 } 2610 do { 2611 next = pud_addr_end(addr, end); 2612 if (pud_none(*pud) && !create) 2613 continue; 2614 if (WARN_ON_ONCE(pud_leaf(*pud))) 2615 return -EINVAL; 2616 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) { 2617 if (!create) 2618 continue; 2619 pud_clear_bad(pud); 2620 } 2621 err = apply_to_pmd_range(mm, pud, addr, next, 2622 fn, data, create, mask); 2623 if (err) 2624 break; 2625 } while (pud++, addr = next, addr != end); 2626 2627 return err; 2628 } 2629 2630 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2631 unsigned long addr, unsigned long end, 2632 pte_fn_t fn, void *data, bool create, 2633 pgtbl_mod_mask *mask) 2634 { 2635 p4d_t *p4d; 2636 unsigned long next; 2637 int err = 0; 2638 2639 if (create) { 2640 p4d = p4d_alloc_track(mm, pgd, addr, mask); 2641 if (!p4d) 2642 return -ENOMEM; 2643 } else { 2644 p4d = p4d_offset(pgd, addr); 2645 } 2646 do { 2647 next = p4d_addr_end(addr, end); 2648 if (p4d_none(*p4d) && !create) 2649 continue; 2650 if (WARN_ON_ONCE(p4d_leaf(*p4d))) 2651 return -EINVAL; 2652 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) { 2653 if (!create) 2654 continue; 2655 p4d_clear_bad(p4d); 2656 } 2657 err = apply_to_pud_range(mm, p4d, addr, next, 2658 fn, data, create, mask); 2659 if (err) 2660 break; 2661 } while (p4d++, addr = next, addr != end); 2662 2663 return err; 2664 } 2665 2666 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2667 unsigned long size, pte_fn_t fn, 2668 void *data, bool create) 2669 { 2670 pgd_t *pgd; 2671 unsigned long start = addr, next; 2672 unsigned long end = addr + size; 2673 pgtbl_mod_mask mask = 0; 2674 int err = 0; 2675 2676 if (WARN_ON(addr >= end)) 2677 return -EINVAL; 2678 2679 pgd = pgd_offset(mm, addr); 2680 do { 2681 next = pgd_addr_end(addr, end); 2682 if (pgd_none(*pgd) && !create) 2683 continue; 2684 if (WARN_ON_ONCE(pgd_leaf(*pgd))) 2685 return -EINVAL; 2686 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) { 2687 if (!create) 2688 continue; 2689 pgd_clear_bad(pgd); 2690 } 2691 err = apply_to_p4d_range(mm, pgd, addr, next, 2692 fn, data, create, &mask); 2693 if (err) 2694 break; 2695 } while (pgd++, addr = next, addr != end); 2696 2697 if (mask & ARCH_PAGE_TABLE_SYNC_MASK) 2698 arch_sync_kernel_mappings(start, start + size); 2699 2700 return err; 2701 } 2702 2703 /* 2704 * Scan a region of virtual memory, filling in page tables as necessary 2705 * and calling a provided function on each leaf page table. 2706 */ 2707 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2708 unsigned long size, pte_fn_t fn, void *data) 2709 { 2710 return __apply_to_page_range(mm, addr, size, fn, data, true); 2711 } 2712 EXPORT_SYMBOL_GPL(apply_to_page_range); 2713 2714 /* 2715 * Scan a region of virtual memory, calling a provided function on 2716 * each leaf page table where it exists. 2717 * 2718 * Unlike apply_to_page_range, this does _not_ fill in page tables 2719 * where they are absent. 2720 */ 2721 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, 2722 unsigned long size, pte_fn_t fn, void *data) 2723 { 2724 return __apply_to_page_range(mm, addr, size, fn, data, false); 2725 } 2726 EXPORT_SYMBOL_GPL(apply_to_existing_page_range); 2727 2728 /* 2729 * handle_pte_fault chooses page fault handler according to an entry which was 2730 * read non-atomically. Before making any commitment, on those architectures 2731 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 2732 * parts, do_swap_page must check under lock before unmapping the pte and 2733 * proceeding (but do_wp_page is only called after already making such a check; 2734 * and do_anonymous_page can safely check later on). 2735 */ 2736 static inline int pte_unmap_same(struct vm_fault *vmf) 2737 { 2738 int same = 1; 2739 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) 2740 if (sizeof(pte_t) > sizeof(unsigned long)) { 2741 spinlock_t *ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 2742 spin_lock(ptl); 2743 same = pte_same(*vmf->pte, vmf->orig_pte); 2744 spin_unlock(ptl); 2745 } 2746 #endif 2747 pte_unmap(vmf->pte); 2748 vmf->pte = NULL; 2749 return same; 2750 } 2751 2752 static inline bool cow_user_page(struct page *dst, struct page *src, 2753 struct vm_fault *vmf) 2754 { 2755 bool ret; 2756 void *kaddr; 2757 void __user *uaddr; 2758 bool locked = false; 2759 struct vm_area_struct *vma = vmf->vma; 2760 struct mm_struct *mm = vma->vm_mm; 2761 unsigned long addr = vmf->address; 2762 2763 if (likely(src)) { 2764 copy_user_highpage(dst, src, addr, vma); 2765 return true; 2766 } 2767 2768 /* 2769 * If the source page was a PFN mapping, we don't have 2770 * a "struct page" for it. We do a best-effort copy by 2771 * just copying from the original user address. If that 2772 * fails, we just zero-fill it. Live with it. 2773 */ 2774 kaddr = kmap_atomic(dst); 2775 uaddr = (void __user *)(addr & PAGE_MASK); 2776 2777 /* 2778 * On architectures with software "accessed" bits, we would 2779 * take a double page fault, so mark it accessed here. 2780 */ 2781 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) { 2782 pte_t entry; 2783 2784 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2785 locked = true; 2786 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2787 /* 2788 * Other thread has already handled the fault 2789 * and update local tlb only 2790 */ 2791 update_mmu_tlb(vma, addr, vmf->pte); 2792 ret = false; 2793 goto pte_unlock; 2794 } 2795 2796 entry = pte_mkyoung(vmf->orig_pte); 2797 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) 2798 update_mmu_cache(vma, addr, vmf->pte); 2799 } 2800 2801 /* 2802 * This really shouldn't fail, because the page is there 2803 * in the page tables. But it might just be unreadable, 2804 * in which case we just give up and fill the result with 2805 * zeroes. 2806 */ 2807 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2808 if (locked) 2809 goto warn; 2810 2811 /* Re-validate under PTL if the page is still mapped */ 2812 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2813 locked = true; 2814 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2815 /* The PTE changed under us, update local tlb */ 2816 update_mmu_tlb(vma, addr, vmf->pte); 2817 ret = false; 2818 goto pte_unlock; 2819 } 2820 2821 /* 2822 * The same page can be mapped back since last copy attempt. 2823 * Try to copy again under PTL. 2824 */ 2825 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2826 /* 2827 * Give a warn in case there can be some obscure 2828 * use-case 2829 */ 2830 warn: 2831 WARN_ON_ONCE(1); 2832 clear_page(kaddr); 2833 } 2834 } 2835 2836 ret = true; 2837 2838 pte_unlock: 2839 if (locked) 2840 pte_unmap_unlock(vmf->pte, vmf->ptl); 2841 kunmap_atomic(kaddr); 2842 flush_dcache_page(dst); 2843 2844 return ret; 2845 } 2846 2847 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 2848 { 2849 struct file *vm_file = vma->vm_file; 2850 2851 if (vm_file) 2852 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 2853 2854 /* 2855 * Special mappings (e.g. VDSO) do not have any file so fake 2856 * a default GFP_KERNEL for them. 2857 */ 2858 return GFP_KERNEL; 2859 } 2860 2861 /* 2862 * Notify the address space that the page is about to become writable so that 2863 * it can prohibit this or wait for the page to get into an appropriate state. 2864 * 2865 * We do this without the lock held, so that it can sleep if it needs to. 2866 */ 2867 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf) 2868 { 2869 vm_fault_t ret; 2870 struct page *page = vmf->page; 2871 unsigned int old_flags = vmf->flags; 2872 2873 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2874 2875 if (vmf->vma->vm_file && 2876 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) 2877 return VM_FAULT_SIGBUS; 2878 2879 ret = vmf->vma->vm_ops->page_mkwrite(vmf); 2880 /* Restore original flags so that caller is not surprised */ 2881 vmf->flags = old_flags; 2882 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2883 return ret; 2884 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2885 lock_page(page); 2886 if (!page->mapping) { 2887 unlock_page(page); 2888 return 0; /* retry */ 2889 } 2890 ret |= VM_FAULT_LOCKED; 2891 } else 2892 VM_BUG_ON_PAGE(!PageLocked(page), page); 2893 return ret; 2894 } 2895 2896 /* 2897 * Handle dirtying of a page in shared file mapping on a write fault. 2898 * 2899 * The function expects the page to be locked and unlocks it. 2900 */ 2901 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) 2902 { 2903 struct vm_area_struct *vma = vmf->vma; 2904 struct address_space *mapping; 2905 struct page *page = vmf->page; 2906 bool dirtied; 2907 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 2908 2909 dirtied = set_page_dirty(page); 2910 VM_BUG_ON_PAGE(PageAnon(page), page); 2911 /* 2912 * Take a local copy of the address_space - page.mapping may be zeroed 2913 * by truncate after unlock_page(). The address_space itself remains 2914 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 2915 * release semantics to prevent the compiler from undoing this copying. 2916 */ 2917 mapping = page_rmapping(page); 2918 unlock_page(page); 2919 2920 if (!page_mkwrite) 2921 file_update_time(vma->vm_file); 2922 2923 /* 2924 * Throttle page dirtying rate down to writeback speed. 2925 * 2926 * mapping may be NULL here because some device drivers do not 2927 * set page.mapping but still dirty their pages 2928 * 2929 * Drop the mmap_lock before waiting on IO, if we can. The file 2930 * is pinning the mapping, as per above. 2931 */ 2932 if ((dirtied || page_mkwrite) && mapping) { 2933 struct file *fpin; 2934 2935 fpin = maybe_unlock_mmap_for_io(vmf, NULL); 2936 balance_dirty_pages_ratelimited(mapping); 2937 if (fpin) { 2938 fput(fpin); 2939 return VM_FAULT_RETRY; 2940 } 2941 } 2942 2943 return 0; 2944 } 2945 2946 /* 2947 * Handle write page faults for pages that can be reused in the current vma 2948 * 2949 * This can happen either due to the mapping being with the VM_SHARED flag, 2950 * or due to us being the last reference standing to the page. In either 2951 * case, all we need to do here is to mark the page as writable and update 2952 * any related book-keeping. 2953 */ 2954 static inline void wp_page_reuse(struct vm_fault *vmf) 2955 __releases(vmf->ptl) 2956 { 2957 struct vm_area_struct *vma = vmf->vma; 2958 struct page *page = vmf->page; 2959 pte_t entry; 2960 /* 2961 * Clear the pages cpupid information as the existing 2962 * information potentially belongs to a now completely 2963 * unrelated process. 2964 */ 2965 if (page) 2966 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2967 2968 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2969 entry = pte_mkyoung(vmf->orig_pte); 2970 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2971 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 2972 update_mmu_cache(vma, vmf->address, vmf->pte); 2973 pte_unmap_unlock(vmf->pte, vmf->ptl); 2974 count_vm_event(PGREUSE); 2975 } 2976 2977 /* 2978 * Handle the case of a page which we actually need to copy to a new page. 2979 * 2980 * Called with mmap_lock locked and the old page referenced, but 2981 * without the ptl held. 2982 * 2983 * High level logic flow: 2984 * 2985 * - Allocate a page, copy the content of the old page to the new one. 2986 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2987 * - Take the PTL. If the pte changed, bail out and release the allocated page 2988 * - If the pte is still the way we remember it, update the page table and all 2989 * relevant references. This includes dropping the reference the page-table 2990 * held to the old page, as well as updating the rmap. 2991 * - In any case, unlock the PTL and drop the reference we took to the old page. 2992 */ 2993 static vm_fault_t wp_page_copy(struct vm_fault *vmf) 2994 { 2995 struct vm_area_struct *vma = vmf->vma; 2996 struct mm_struct *mm = vma->vm_mm; 2997 struct page *old_page = vmf->page; 2998 struct page *new_page = NULL; 2999 pte_t entry; 3000 int page_copied = 0; 3001 struct mmu_notifier_range range; 3002 3003 if (unlikely(anon_vma_prepare(vma))) 3004 goto oom; 3005 3006 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) { 3007 new_page = alloc_zeroed_user_highpage_movable(vma, 3008 vmf->address); 3009 if (!new_page) 3010 goto oom; 3011 } else { 3012 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 3013 vmf->address); 3014 if (!new_page) 3015 goto oom; 3016 3017 if (!cow_user_page(new_page, old_page, vmf)) { 3018 /* 3019 * COW failed, if the fault was solved by other, 3020 * it's fine. If not, userspace would re-fault on 3021 * the same address and we will handle the fault 3022 * from the second attempt. 3023 */ 3024 put_page(new_page); 3025 if (old_page) 3026 put_page(old_page); 3027 return 0; 3028 } 3029 } 3030 3031 if (mem_cgroup_charge(page_folio(new_page), mm, GFP_KERNEL)) 3032 goto oom_free_new; 3033 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 3034 3035 __SetPageUptodate(new_page); 3036 3037 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 3038 vmf->address & PAGE_MASK, 3039 (vmf->address & PAGE_MASK) + PAGE_SIZE); 3040 mmu_notifier_invalidate_range_start(&range); 3041 3042 /* 3043 * Re-check the pte - we dropped the lock 3044 */ 3045 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 3046 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) { 3047 if (old_page) { 3048 if (!PageAnon(old_page)) { 3049 dec_mm_counter_fast(mm, 3050 mm_counter_file(old_page)); 3051 inc_mm_counter_fast(mm, MM_ANONPAGES); 3052 } 3053 } else { 3054 inc_mm_counter_fast(mm, MM_ANONPAGES); 3055 } 3056 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 3057 entry = mk_pte(new_page, vma->vm_page_prot); 3058 entry = pte_sw_mkyoung(entry); 3059 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3060 3061 /* 3062 * Clear the pte entry and flush it first, before updating the 3063 * pte with the new entry, to keep TLBs on different CPUs in 3064 * sync. This code used to set the new PTE then flush TLBs, but 3065 * that left a window where the new PTE could be loaded into 3066 * some TLBs while the old PTE remains in others. 3067 */ 3068 ptep_clear_flush_notify(vma, vmf->address, vmf->pte); 3069 page_add_new_anon_rmap(new_page, vma, vmf->address, false); 3070 lru_cache_add_inactive_or_unevictable(new_page, vma); 3071 /* 3072 * We call the notify macro here because, when using secondary 3073 * mmu page tables (such as kvm shadow page tables), we want the 3074 * new page to be mapped directly into the secondary page table. 3075 */ 3076 set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 3077 update_mmu_cache(vma, vmf->address, vmf->pte); 3078 if (old_page) { 3079 /* 3080 * Only after switching the pte to the new page may 3081 * we remove the mapcount here. Otherwise another 3082 * process may come and find the rmap count decremented 3083 * before the pte is switched to the new page, and 3084 * "reuse" the old page writing into it while our pte 3085 * here still points into it and can be read by other 3086 * threads. 3087 * 3088 * The critical issue is to order this 3089 * page_remove_rmap with the ptp_clear_flush above. 3090 * Those stores are ordered by (if nothing else,) 3091 * the barrier present in the atomic_add_negative 3092 * in page_remove_rmap. 3093 * 3094 * Then the TLB flush in ptep_clear_flush ensures that 3095 * no process can access the old page before the 3096 * decremented mapcount is visible. And the old page 3097 * cannot be reused until after the decremented 3098 * mapcount is visible. So transitively, TLBs to 3099 * old page will be flushed before it can be reused. 3100 */ 3101 page_remove_rmap(old_page, false); 3102 } 3103 3104 /* Free the old page.. */ 3105 new_page = old_page; 3106 page_copied = 1; 3107 } else { 3108 update_mmu_tlb(vma, vmf->address, vmf->pte); 3109 } 3110 3111 if (new_page) 3112 put_page(new_page); 3113 3114 pte_unmap_unlock(vmf->pte, vmf->ptl); 3115 /* 3116 * No need to double call mmu_notifier->invalidate_range() callback as 3117 * the above ptep_clear_flush_notify() did already call it. 3118 */ 3119 mmu_notifier_invalidate_range_only_end(&range); 3120 if (old_page) { 3121 /* 3122 * Don't let another task, with possibly unlocked vma, 3123 * keep the mlocked page. 3124 */ 3125 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 3126 lock_page(old_page); /* LRU manipulation */ 3127 if (PageMlocked(old_page)) 3128 munlock_vma_page(old_page); 3129 unlock_page(old_page); 3130 } 3131 if (page_copied) 3132 free_swap_cache(old_page); 3133 put_page(old_page); 3134 } 3135 return page_copied ? VM_FAULT_WRITE : 0; 3136 oom_free_new: 3137 put_page(new_page); 3138 oom: 3139 if (old_page) 3140 put_page(old_page); 3141 return VM_FAULT_OOM; 3142 } 3143 3144 /** 3145 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 3146 * writeable once the page is prepared 3147 * 3148 * @vmf: structure describing the fault 3149 * 3150 * This function handles all that is needed to finish a write page fault in a 3151 * shared mapping due to PTE being read-only once the mapped page is prepared. 3152 * It handles locking of PTE and modifying it. 3153 * 3154 * The function expects the page to be locked or other protection against 3155 * concurrent faults / writeback (such as DAX radix tree locks). 3156 * 3157 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before 3158 * we acquired PTE lock. 3159 */ 3160 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf) 3161 { 3162 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 3163 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 3164 &vmf->ptl); 3165 /* 3166 * We might have raced with another page fault while we released the 3167 * pte_offset_map_lock. 3168 */ 3169 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 3170 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 3171 pte_unmap_unlock(vmf->pte, vmf->ptl); 3172 return VM_FAULT_NOPAGE; 3173 } 3174 wp_page_reuse(vmf); 3175 return 0; 3176 } 3177 3178 /* 3179 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 3180 * mapping 3181 */ 3182 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 3183 { 3184 struct vm_area_struct *vma = vmf->vma; 3185 3186 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 3187 vm_fault_t ret; 3188 3189 pte_unmap_unlock(vmf->pte, vmf->ptl); 3190 vmf->flags |= FAULT_FLAG_MKWRITE; 3191 ret = vma->vm_ops->pfn_mkwrite(vmf); 3192 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 3193 return ret; 3194 return finish_mkwrite_fault(vmf); 3195 } 3196 wp_page_reuse(vmf); 3197 return VM_FAULT_WRITE; 3198 } 3199 3200 static vm_fault_t wp_page_shared(struct vm_fault *vmf) 3201 __releases(vmf->ptl) 3202 { 3203 struct vm_area_struct *vma = vmf->vma; 3204 vm_fault_t ret = VM_FAULT_WRITE; 3205 3206 get_page(vmf->page); 3207 3208 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 3209 vm_fault_t tmp; 3210 3211 pte_unmap_unlock(vmf->pte, vmf->ptl); 3212 tmp = do_page_mkwrite(vmf); 3213 if (unlikely(!tmp || (tmp & 3214 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3215 put_page(vmf->page); 3216 return tmp; 3217 } 3218 tmp = finish_mkwrite_fault(vmf); 3219 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 3220 unlock_page(vmf->page); 3221 put_page(vmf->page); 3222 return tmp; 3223 } 3224 } else { 3225 wp_page_reuse(vmf); 3226 lock_page(vmf->page); 3227 } 3228 ret |= fault_dirty_shared_page(vmf); 3229 put_page(vmf->page); 3230 3231 return ret; 3232 } 3233 3234 /* 3235 * This routine handles present pages, when users try to write 3236 * to a shared page. It is done by copying the page to a new address 3237 * and decrementing the shared-page counter for the old page. 3238 * 3239 * Note that this routine assumes that the protection checks have been 3240 * done by the caller (the low-level page fault routine in most cases). 3241 * Thus we can safely just mark it writable once we've done any necessary 3242 * COW. 3243 * 3244 * We also mark the page dirty at this point even though the page will 3245 * change only once the write actually happens. This avoids a few races, 3246 * and potentially makes it more efficient. 3247 * 3248 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3249 * but allow concurrent faults), with pte both mapped and locked. 3250 * We return with mmap_lock still held, but pte unmapped and unlocked. 3251 */ 3252 static vm_fault_t do_wp_page(struct vm_fault *vmf) 3253 __releases(vmf->ptl) 3254 { 3255 struct vm_area_struct *vma = vmf->vma; 3256 3257 if (userfaultfd_pte_wp(vma, *vmf->pte)) { 3258 pte_unmap_unlock(vmf->pte, vmf->ptl); 3259 return handle_userfault(vmf, VM_UFFD_WP); 3260 } 3261 3262 /* 3263 * Userfaultfd write-protect can defer flushes. Ensure the TLB 3264 * is flushed in this case before copying. 3265 */ 3266 if (unlikely(userfaultfd_wp(vmf->vma) && 3267 mm_tlb_flush_pending(vmf->vma->vm_mm))) 3268 flush_tlb_page(vmf->vma, vmf->address); 3269 3270 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 3271 if (!vmf->page) { 3272 /* 3273 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 3274 * VM_PFNMAP VMA. 3275 * 3276 * We should not cow pages in a shared writeable mapping. 3277 * Just mark the pages writable and/or call ops->pfn_mkwrite. 3278 */ 3279 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3280 (VM_WRITE|VM_SHARED)) 3281 return wp_pfn_shared(vmf); 3282 3283 pte_unmap_unlock(vmf->pte, vmf->ptl); 3284 return wp_page_copy(vmf); 3285 } 3286 3287 /* 3288 * Take out anonymous pages first, anonymous shared vmas are 3289 * not dirty accountable. 3290 */ 3291 if (PageAnon(vmf->page)) { 3292 struct page *page = vmf->page; 3293 3294 /* PageKsm() doesn't necessarily raise the page refcount */ 3295 if (PageKsm(page) || page_count(page) != 1) 3296 goto copy; 3297 if (!trylock_page(page)) 3298 goto copy; 3299 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) { 3300 unlock_page(page); 3301 goto copy; 3302 } 3303 /* 3304 * Ok, we've got the only map reference, and the only 3305 * page count reference, and the page is locked, 3306 * it's dark out, and we're wearing sunglasses. Hit it. 3307 */ 3308 unlock_page(page); 3309 wp_page_reuse(vmf); 3310 return VM_FAULT_WRITE; 3311 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3312 (VM_WRITE|VM_SHARED))) { 3313 return wp_page_shared(vmf); 3314 } 3315 copy: 3316 /* 3317 * Ok, we need to copy. Oh, well.. 3318 */ 3319 get_page(vmf->page); 3320 3321 pte_unmap_unlock(vmf->pte, vmf->ptl); 3322 return wp_page_copy(vmf); 3323 } 3324 3325 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 3326 unsigned long start_addr, unsigned long end_addr, 3327 struct zap_details *details) 3328 { 3329 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 3330 } 3331 3332 static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 3333 pgoff_t first_index, 3334 pgoff_t last_index, 3335 struct zap_details *details) 3336 { 3337 struct vm_area_struct *vma; 3338 pgoff_t vba, vea, zba, zea; 3339 3340 vma_interval_tree_foreach(vma, root, first_index, last_index) { 3341 vba = vma->vm_pgoff; 3342 vea = vba + vma_pages(vma) - 1; 3343 zba = first_index; 3344 if (zba < vba) 3345 zba = vba; 3346 zea = last_index; 3347 if (zea > vea) 3348 zea = vea; 3349 3350 unmap_mapping_range_vma(vma, 3351 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 3352 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 3353 details); 3354 } 3355 } 3356 3357 /** 3358 * unmap_mapping_folio() - Unmap single folio from processes. 3359 * @folio: The locked folio to be unmapped. 3360 * 3361 * Unmap this folio from any userspace process which still has it mmaped. 3362 * Typically, for efficiency, the range of nearby pages has already been 3363 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once 3364 * truncation or invalidation holds the lock on a folio, it may find that 3365 * the page has been remapped again: and then uses unmap_mapping_folio() 3366 * to unmap it finally. 3367 */ 3368 void unmap_mapping_folio(struct folio *folio) 3369 { 3370 struct address_space *mapping = folio->mapping; 3371 struct zap_details details = { }; 3372 pgoff_t first_index; 3373 pgoff_t last_index; 3374 3375 VM_BUG_ON(!folio_test_locked(folio)); 3376 3377 first_index = folio->index; 3378 last_index = folio->index + folio_nr_pages(folio) - 1; 3379 3380 details.zap_mapping = mapping; 3381 details.single_folio = folio; 3382 3383 i_mmap_lock_write(mapping); 3384 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3385 unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3386 last_index, &details); 3387 i_mmap_unlock_write(mapping); 3388 } 3389 3390 /** 3391 * unmap_mapping_pages() - Unmap pages from processes. 3392 * @mapping: The address space containing pages to be unmapped. 3393 * @start: Index of first page to be unmapped. 3394 * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 3395 * @even_cows: Whether to unmap even private COWed pages. 3396 * 3397 * Unmap the pages in this address space from any userspace process which 3398 * has them mmaped. Generally, you want to remove COWed pages as well when 3399 * a file is being truncated, but not when invalidating pages from the page 3400 * cache. 3401 */ 3402 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 3403 pgoff_t nr, bool even_cows) 3404 { 3405 struct zap_details details = { }; 3406 pgoff_t first_index = start; 3407 pgoff_t last_index = start + nr - 1; 3408 3409 details.zap_mapping = even_cows ? NULL : mapping; 3410 if (last_index < first_index) 3411 last_index = ULONG_MAX; 3412 3413 i_mmap_lock_write(mapping); 3414 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3415 unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3416 last_index, &details); 3417 i_mmap_unlock_write(mapping); 3418 } 3419 EXPORT_SYMBOL_GPL(unmap_mapping_pages); 3420 3421 /** 3422 * unmap_mapping_range - unmap the portion of all mmaps in the specified 3423 * address_space corresponding to the specified byte range in the underlying 3424 * file. 3425 * 3426 * @mapping: the address space containing mmaps to be unmapped. 3427 * @holebegin: byte in first page to unmap, relative to the start of 3428 * the underlying file. This will be rounded down to a PAGE_SIZE 3429 * boundary. Note that this is different from truncate_pagecache(), which 3430 * must keep the partial page. In contrast, we must get rid of 3431 * partial pages. 3432 * @holelen: size of prospective hole in bytes. This will be rounded 3433 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 3434 * end of the file. 3435 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 3436 * but 0 when invalidating pagecache, don't throw away private data. 3437 */ 3438 void unmap_mapping_range(struct address_space *mapping, 3439 loff_t const holebegin, loff_t const holelen, int even_cows) 3440 { 3441 pgoff_t hba = holebegin >> PAGE_SHIFT; 3442 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3443 3444 /* Check for overflow. */ 3445 if (sizeof(holelen) > sizeof(hlen)) { 3446 long long holeend = 3447 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3448 if (holeend & ~(long long)ULONG_MAX) 3449 hlen = ULONG_MAX - hba + 1; 3450 } 3451 3452 unmap_mapping_pages(mapping, hba, hlen, even_cows); 3453 } 3454 EXPORT_SYMBOL(unmap_mapping_range); 3455 3456 /* 3457 * Restore a potential device exclusive pte to a working pte entry 3458 */ 3459 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf) 3460 { 3461 struct page *page = vmf->page; 3462 struct vm_area_struct *vma = vmf->vma; 3463 struct mmu_notifier_range range; 3464 3465 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) 3466 return VM_FAULT_RETRY; 3467 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma, 3468 vma->vm_mm, vmf->address & PAGE_MASK, 3469 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL); 3470 mmu_notifier_invalidate_range_start(&range); 3471 3472 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3473 &vmf->ptl); 3474 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3475 restore_exclusive_pte(vma, page, vmf->address, vmf->pte); 3476 3477 pte_unmap_unlock(vmf->pte, vmf->ptl); 3478 unlock_page(page); 3479 3480 mmu_notifier_invalidate_range_end(&range); 3481 return 0; 3482 } 3483 3484 /* 3485 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3486 * but allow concurrent faults), and pte mapped but not yet locked. 3487 * We return with pte unmapped and unlocked. 3488 * 3489 * We return with the mmap_lock locked or unlocked in the same cases 3490 * as does filemap_fault(). 3491 */ 3492 vm_fault_t do_swap_page(struct vm_fault *vmf) 3493 { 3494 struct vm_area_struct *vma = vmf->vma; 3495 struct page *page = NULL, *swapcache; 3496 struct swap_info_struct *si = NULL; 3497 swp_entry_t entry; 3498 pte_t pte; 3499 int locked; 3500 int exclusive = 0; 3501 vm_fault_t ret = 0; 3502 void *shadow = NULL; 3503 3504 if (!pte_unmap_same(vmf)) 3505 goto out; 3506 3507 entry = pte_to_swp_entry(vmf->orig_pte); 3508 if (unlikely(non_swap_entry(entry))) { 3509 if (is_migration_entry(entry)) { 3510 migration_entry_wait(vma->vm_mm, vmf->pmd, 3511 vmf->address); 3512 } else if (is_device_exclusive_entry(entry)) { 3513 vmf->page = pfn_swap_entry_to_page(entry); 3514 ret = remove_device_exclusive_entry(vmf); 3515 } else if (is_device_private_entry(entry)) { 3516 vmf->page = pfn_swap_entry_to_page(entry); 3517 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 3518 } else if (is_hwpoison_entry(entry)) { 3519 ret = VM_FAULT_HWPOISON; 3520 } else { 3521 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 3522 ret = VM_FAULT_SIGBUS; 3523 } 3524 goto out; 3525 } 3526 3527 /* Prevent swapoff from happening to us. */ 3528 si = get_swap_device(entry); 3529 if (unlikely(!si)) 3530 goto out; 3531 3532 delayacct_set_flag(current, DELAYACCT_PF_SWAPIN); 3533 page = lookup_swap_cache(entry, vma, vmf->address); 3534 swapcache = page; 3535 3536 if (!page) { 3537 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && 3538 __swap_count(entry) == 1) { 3539 /* skip swapcache */ 3540 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 3541 vmf->address); 3542 if (page) { 3543 __SetPageLocked(page); 3544 __SetPageSwapBacked(page); 3545 3546 if (mem_cgroup_swapin_charge_page(page, 3547 vma->vm_mm, GFP_KERNEL, entry)) { 3548 ret = VM_FAULT_OOM; 3549 goto out_page; 3550 } 3551 mem_cgroup_swapin_uncharge_swap(entry); 3552 3553 shadow = get_shadow_from_swap_cache(entry); 3554 if (shadow) 3555 workingset_refault(page_folio(page), 3556 shadow); 3557 3558 lru_cache_add(page); 3559 3560 /* To provide entry to swap_readpage() */ 3561 set_page_private(page, entry.val); 3562 swap_readpage(page, true); 3563 set_page_private(page, 0); 3564 } 3565 } else { 3566 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 3567 vmf); 3568 swapcache = page; 3569 } 3570 3571 if (!page) { 3572 /* 3573 * Back out if somebody else faulted in this pte 3574 * while we released the pte lock. 3575 */ 3576 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3577 vmf->address, &vmf->ptl); 3578 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3579 ret = VM_FAULT_OOM; 3580 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN); 3581 goto unlock; 3582 } 3583 3584 /* Had to read the page from swap area: Major fault */ 3585 ret = VM_FAULT_MAJOR; 3586 count_vm_event(PGMAJFAULT); 3587 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 3588 } else if (PageHWPoison(page)) { 3589 /* 3590 * hwpoisoned dirty swapcache pages are kept for killing 3591 * owner processes (which may be unknown at hwpoison time) 3592 */ 3593 ret = VM_FAULT_HWPOISON; 3594 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN); 3595 goto out_release; 3596 } 3597 3598 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags); 3599 3600 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN); 3601 if (!locked) { 3602 ret |= VM_FAULT_RETRY; 3603 goto out_release; 3604 } 3605 3606 /* 3607 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 3608 * release the swapcache from under us. The page pin, and pte_same 3609 * test below, are not enough to exclude that. Even if it is still 3610 * swapcache, we need to check that the page's swap has not changed. 3611 */ 3612 if (unlikely((!PageSwapCache(page) || 3613 page_private(page) != entry.val)) && swapcache) 3614 goto out_page; 3615 3616 page = ksm_might_need_to_copy(page, vma, vmf->address); 3617 if (unlikely(!page)) { 3618 ret = VM_FAULT_OOM; 3619 page = swapcache; 3620 goto out_page; 3621 } 3622 3623 cgroup_throttle_swaprate(page, GFP_KERNEL); 3624 3625 /* 3626 * Back out if somebody else already faulted in this pte. 3627 */ 3628 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3629 &vmf->ptl); 3630 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) 3631 goto out_nomap; 3632 3633 if (unlikely(!PageUptodate(page))) { 3634 ret = VM_FAULT_SIGBUS; 3635 goto out_nomap; 3636 } 3637 3638 /* 3639 * The page isn't present yet, go ahead with the fault. 3640 * 3641 * Be careful about the sequence of operations here. 3642 * To get its accounting right, reuse_swap_page() must be called 3643 * while the page is counted on swap but not yet in mapcount i.e. 3644 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 3645 * must be called after the swap_free(), or it will never succeed. 3646 */ 3647 3648 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3649 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS); 3650 pte = mk_pte(page, vma->vm_page_prot); 3651 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { 3652 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 3653 vmf->flags &= ~FAULT_FLAG_WRITE; 3654 ret |= VM_FAULT_WRITE; 3655 exclusive = RMAP_EXCLUSIVE; 3656 } 3657 flush_icache_page(vma, page); 3658 if (pte_swp_soft_dirty(vmf->orig_pte)) 3659 pte = pte_mksoft_dirty(pte); 3660 if (pte_swp_uffd_wp(vmf->orig_pte)) { 3661 pte = pte_mkuffd_wp(pte); 3662 pte = pte_wrprotect(pte); 3663 } 3664 vmf->orig_pte = pte; 3665 3666 /* ksm created a completely new copy */ 3667 if (unlikely(page != swapcache && swapcache)) { 3668 page_add_new_anon_rmap(page, vma, vmf->address, false); 3669 lru_cache_add_inactive_or_unevictable(page, vma); 3670 } else { 3671 do_page_add_anon_rmap(page, vma, vmf->address, exclusive); 3672 } 3673 3674 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 3675 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 3676 3677 swap_free(entry); 3678 if (mem_cgroup_swap_full(page) || 3679 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 3680 try_to_free_swap(page); 3681 unlock_page(page); 3682 if (page != swapcache && swapcache) { 3683 /* 3684 * Hold the lock to avoid the swap entry to be reused 3685 * until we take the PT lock for the pte_same() check 3686 * (to avoid false positives from pte_same). For 3687 * further safety release the lock after the swap_free 3688 * so that the swap count won't change under a 3689 * parallel locked swapcache. 3690 */ 3691 unlock_page(swapcache); 3692 put_page(swapcache); 3693 } 3694 3695 if (vmf->flags & FAULT_FLAG_WRITE) { 3696 ret |= do_wp_page(vmf); 3697 if (ret & VM_FAULT_ERROR) 3698 ret &= VM_FAULT_ERROR; 3699 goto out; 3700 } 3701 3702 /* No need to invalidate - it was non-present before */ 3703 update_mmu_cache(vma, vmf->address, vmf->pte); 3704 unlock: 3705 pte_unmap_unlock(vmf->pte, vmf->ptl); 3706 out: 3707 if (si) 3708 put_swap_device(si); 3709 return ret; 3710 out_nomap: 3711 pte_unmap_unlock(vmf->pte, vmf->ptl); 3712 out_page: 3713 unlock_page(page); 3714 out_release: 3715 put_page(page); 3716 if (page != swapcache && swapcache) { 3717 unlock_page(swapcache); 3718 put_page(swapcache); 3719 } 3720 if (si) 3721 put_swap_device(si); 3722 return ret; 3723 } 3724 3725 /* 3726 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3727 * but allow concurrent faults), and pte mapped but not yet locked. 3728 * We return with mmap_lock still held, but pte unmapped and unlocked. 3729 */ 3730 static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 3731 { 3732 struct vm_area_struct *vma = vmf->vma; 3733 struct page *page; 3734 vm_fault_t ret = 0; 3735 pte_t entry; 3736 3737 /* File mapping without ->vm_ops ? */ 3738 if (vma->vm_flags & VM_SHARED) 3739 return VM_FAULT_SIGBUS; 3740 3741 /* 3742 * Use pte_alloc() instead of pte_alloc_map(). We can't run 3743 * pte_offset_map() on pmds where a huge pmd might be created 3744 * from a different thread. 3745 * 3746 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 3747 * parallel threads are excluded by other means. 3748 * 3749 * Here we only have mmap_read_lock(mm). 3750 */ 3751 if (pte_alloc(vma->vm_mm, vmf->pmd)) 3752 return VM_FAULT_OOM; 3753 3754 /* See comment in handle_pte_fault() */ 3755 if (unlikely(pmd_trans_unstable(vmf->pmd))) 3756 return 0; 3757 3758 /* Use the zero-page for reads */ 3759 if (!(vmf->flags & FAULT_FLAG_WRITE) && 3760 !mm_forbids_zeropage(vma->vm_mm)) { 3761 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 3762 vma->vm_page_prot)); 3763 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3764 vmf->address, &vmf->ptl); 3765 if (!pte_none(*vmf->pte)) { 3766 update_mmu_tlb(vma, vmf->address, vmf->pte); 3767 goto unlock; 3768 } 3769 ret = check_stable_address_space(vma->vm_mm); 3770 if (ret) 3771 goto unlock; 3772 /* Deliver the page fault to userland, check inside PT lock */ 3773 if (userfaultfd_missing(vma)) { 3774 pte_unmap_unlock(vmf->pte, vmf->ptl); 3775 return handle_userfault(vmf, VM_UFFD_MISSING); 3776 } 3777 goto setpte; 3778 } 3779 3780 /* Allocate our own private page. */ 3781 if (unlikely(anon_vma_prepare(vma))) 3782 goto oom; 3783 page = alloc_zeroed_user_highpage_movable(vma, vmf->address); 3784 if (!page) 3785 goto oom; 3786 3787 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL)) 3788 goto oom_free_page; 3789 cgroup_throttle_swaprate(page, GFP_KERNEL); 3790 3791 /* 3792 * The memory barrier inside __SetPageUptodate makes sure that 3793 * preceding stores to the page contents become visible before 3794 * the set_pte_at() write. 3795 */ 3796 __SetPageUptodate(page); 3797 3798 entry = mk_pte(page, vma->vm_page_prot); 3799 entry = pte_sw_mkyoung(entry); 3800 if (vma->vm_flags & VM_WRITE) 3801 entry = pte_mkwrite(pte_mkdirty(entry)); 3802 3803 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3804 &vmf->ptl); 3805 if (!pte_none(*vmf->pte)) { 3806 update_mmu_cache(vma, vmf->address, vmf->pte); 3807 goto release; 3808 } 3809 3810 ret = check_stable_address_space(vma->vm_mm); 3811 if (ret) 3812 goto release; 3813 3814 /* Deliver the page fault to userland, check inside PT lock */ 3815 if (userfaultfd_missing(vma)) { 3816 pte_unmap_unlock(vmf->pte, vmf->ptl); 3817 put_page(page); 3818 return handle_userfault(vmf, VM_UFFD_MISSING); 3819 } 3820 3821 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3822 page_add_new_anon_rmap(page, vma, vmf->address, false); 3823 lru_cache_add_inactive_or_unevictable(page, vma); 3824 setpte: 3825 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3826 3827 /* No need to invalidate - it was non-present before */ 3828 update_mmu_cache(vma, vmf->address, vmf->pte); 3829 unlock: 3830 pte_unmap_unlock(vmf->pte, vmf->ptl); 3831 return ret; 3832 release: 3833 put_page(page); 3834 goto unlock; 3835 oom_free_page: 3836 put_page(page); 3837 oom: 3838 return VM_FAULT_OOM; 3839 } 3840 3841 /* 3842 * The mmap_lock must have been held on entry, and may have been 3843 * released depending on flags and vma->vm_ops->fault() return value. 3844 * See filemap_fault() and __lock_page_retry(). 3845 */ 3846 static vm_fault_t __do_fault(struct vm_fault *vmf) 3847 { 3848 struct vm_area_struct *vma = vmf->vma; 3849 vm_fault_t ret; 3850 3851 /* 3852 * Preallocate pte before we take page_lock because this might lead to 3853 * deadlocks for memcg reclaim which waits for pages under writeback: 3854 * lock_page(A) 3855 * SetPageWriteback(A) 3856 * unlock_page(A) 3857 * lock_page(B) 3858 * lock_page(B) 3859 * pte_alloc_one 3860 * shrink_page_list 3861 * wait_on_page_writeback(A) 3862 * SetPageWriteback(B) 3863 * unlock_page(B) 3864 * # flush A, B to clear the writeback 3865 */ 3866 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 3867 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3868 if (!vmf->prealloc_pte) 3869 return VM_FAULT_OOM; 3870 } 3871 3872 ret = vma->vm_ops->fault(vmf); 3873 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 3874 VM_FAULT_DONE_COW))) 3875 return ret; 3876 3877 if (unlikely(PageHWPoison(vmf->page))) { 3878 if (ret & VM_FAULT_LOCKED) 3879 unlock_page(vmf->page); 3880 put_page(vmf->page); 3881 vmf->page = NULL; 3882 return VM_FAULT_HWPOISON; 3883 } 3884 3885 if (unlikely(!(ret & VM_FAULT_LOCKED))) 3886 lock_page(vmf->page); 3887 else 3888 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page); 3889 3890 return ret; 3891 } 3892 3893 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3894 static void deposit_prealloc_pte(struct vm_fault *vmf) 3895 { 3896 struct vm_area_struct *vma = vmf->vma; 3897 3898 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3899 /* 3900 * We are going to consume the prealloc table, 3901 * count that as nr_ptes. 3902 */ 3903 mm_inc_nr_ptes(vma->vm_mm); 3904 vmf->prealloc_pte = NULL; 3905 } 3906 3907 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3908 { 3909 struct vm_area_struct *vma = vmf->vma; 3910 bool write = vmf->flags & FAULT_FLAG_WRITE; 3911 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 3912 pmd_t entry; 3913 int i; 3914 vm_fault_t ret = VM_FAULT_FALLBACK; 3915 3916 if (!transhuge_vma_suitable(vma, haddr)) 3917 return ret; 3918 3919 page = compound_head(page); 3920 if (compound_order(page) != HPAGE_PMD_ORDER) 3921 return ret; 3922 3923 /* 3924 * Just backoff if any subpage of a THP is corrupted otherwise 3925 * the corrupted page may mapped by PMD silently to escape the 3926 * check. This kind of THP just can be PTE mapped. Access to 3927 * the corrupted subpage should trigger SIGBUS as expected. 3928 */ 3929 if (unlikely(PageHasHWPoisoned(page))) 3930 return ret; 3931 3932 /* 3933 * Archs like ppc64 need additional space to store information 3934 * related to pte entry. Use the preallocated table for that. 3935 */ 3936 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 3937 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3938 if (!vmf->prealloc_pte) 3939 return VM_FAULT_OOM; 3940 } 3941 3942 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3943 if (unlikely(!pmd_none(*vmf->pmd))) 3944 goto out; 3945 3946 for (i = 0; i < HPAGE_PMD_NR; i++) 3947 flush_icache_page(vma, page + i); 3948 3949 entry = mk_huge_pmd(page, vma->vm_page_prot); 3950 if (write) 3951 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 3952 3953 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR); 3954 page_add_file_rmap(page, true); 3955 /* 3956 * deposit and withdraw with pmd lock held 3957 */ 3958 if (arch_needs_pgtable_deposit()) 3959 deposit_prealloc_pte(vmf); 3960 3961 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 3962 3963 update_mmu_cache_pmd(vma, haddr, vmf->pmd); 3964 3965 /* fault is handled */ 3966 ret = 0; 3967 count_vm_event(THP_FILE_MAPPED); 3968 out: 3969 spin_unlock(vmf->ptl); 3970 return ret; 3971 } 3972 #else 3973 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3974 { 3975 return VM_FAULT_FALLBACK; 3976 } 3977 #endif 3978 3979 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr) 3980 { 3981 struct vm_area_struct *vma = vmf->vma; 3982 bool write = vmf->flags & FAULT_FLAG_WRITE; 3983 bool prefault = vmf->address != addr; 3984 pte_t entry; 3985 3986 flush_icache_page(vma, page); 3987 entry = mk_pte(page, vma->vm_page_prot); 3988 3989 if (prefault && arch_wants_old_prefaulted_pte()) 3990 entry = pte_mkold(entry); 3991 else 3992 entry = pte_sw_mkyoung(entry); 3993 3994 if (write) 3995 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3996 /* copy-on-write page */ 3997 if (write && !(vma->vm_flags & VM_SHARED)) { 3998 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3999 page_add_new_anon_rmap(page, vma, addr, false); 4000 lru_cache_add_inactive_or_unevictable(page, vma); 4001 } else { 4002 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 4003 page_add_file_rmap(page, false); 4004 } 4005 set_pte_at(vma->vm_mm, addr, vmf->pte, entry); 4006 } 4007 4008 /** 4009 * finish_fault - finish page fault once we have prepared the page to fault 4010 * 4011 * @vmf: structure describing the fault 4012 * 4013 * This function handles all that is needed to finish a page fault once the 4014 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 4015 * given page, adds reverse page mapping, handles memcg charges and LRU 4016 * addition. 4017 * 4018 * The function expects the page to be locked and on success it consumes a 4019 * reference of a page being mapped (for the PTE which maps it). 4020 * 4021 * Return: %0 on success, %VM_FAULT_ code in case of error. 4022 */ 4023 vm_fault_t finish_fault(struct vm_fault *vmf) 4024 { 4025 struct vm_area_struct *vma = vmf->vma; 4026 struct page *page; 4027 vm_fault_t ret; 4028 4029 /* Did we COW the page? */ 4030 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 4031 page = vmf->cow_page; 4032 else 4033 page = vmf->page; 4034 4035 /* 4036 * check even for read faults because we might have lost our CoWed 4037 * page 4038 */ 4039 if (!(vma->vm_flags & VM_SHARED)) { 4040 ret = check_stable_address_space(vma->vm_mm); 4041 if (ret) 4042 return ret; 4043 } 4044 4045 if (pmd_none(*vmf->pmd)) { 4046 if (PageTransCompound(page)) { 4047 ret = do_set_pmd(vmf, page); 4048 if (ret != VM_FAULT_FALLBACK) 4049 return ret; 4050 } 4051 4052 if (vmf->prealloc_pte) 4053 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte); 4054 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) 4055 return VM_FAULT_OOM; 4056 } 4057 4058 /* See comment in handle_pte_fault() */ 4059 if (pmd_devmap_trans_unstable(vmf->pmd)) 4060 return 0; 4061 4062 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 4063 vmf->address, &vmf->ptl); 4064 ret = 0; 4065 /* Re-check under ptl */ 4066 if (likely(pte_none(*vmf->pte))) 4067 do_set_pte(vmf, page, vmf->address); 4068 else 4069 ret = VM_FAULT_NOPAGE; 4070 4071 update_mmu_tlb(vma, vmf->address, vmf->pte); 4072 pte_unmap_unlock(vmf->pte, vmf->ptl); 4073 return ret; 4074 } 4075 4076 static unsigned long fault_around_bytes __read_mostly = 4077 rounddown_pow_of_two(65536); 4078 4079 #ifdef CONFIG_DEBUG_FS 4080 static int fault_around_bytes_get(void *data, u64 *val) 4081 { 4082 *val = fault_around_bytes; 4083 return 0; 4084 } 4085 4086 /* 4087 * fault_around_bytes must be rounded down to the nearest page order as it's 4088 * what do_fault_around() expects to see. 4089 */ 4090 static int fault_around_bytes_set(void *data, u64 val) 4091 { 4092 if (val / PAGE_SIZE > PTRS_PER_PTE) 4093 return -EINVAL; 4094 if (val > PAGE_SIZE) 4095 fault_around_bytes = rounddown_pow_of_two(val); 4096 else 4097 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 4098 return 0; 4099 } 4100 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 4101 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 4102 4103 static int __init fault_around_debugfs(void) 4104 { 4105 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 4106 &fault_around_bytes_fops); 4107 return 0; 4108 } 4109 late_initcall(fault_around_debugfs); 4110 #endif 4111 4112 /* 4113 * do_fault_around() tries to map few pages around the fault address. The hope 4114 * is that the pages will be needed soon and this will lower the number of 4115 * faults to handle. 4116 * 4117 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 4118 * not ready to be mapped: not up-to-date, locked, etc. 4119 * 4120 * This function is called with the page table lock taken. In the split ptlock 4121 * case the page table lock only protects only those entries which belong to 4122 * the page table corresponding to the fault address. 4123 * 4124 * This function doesn't cross the VMA boundaries, in order to call map_pages() 4125 * only once. 4126 * 4127 * fault_around_bytes defines how many bytes we'll try to map. 4128 * do_fault_around() expects it to be set to a power of two less than or equal 4129 * to PTRS_PER_PTE. 4130 * 4131 * The virtual address of the area that we map is naturally aligned to 4132 * fault_around_bytes rounded down to the machine page size 4133 * (and therefore to page order). This way it's easier to guarantee 4134 * that we don't cross page table boundaries. 4135 */ 4136 static vm_fault_t do_fault_around(struct vm_fault *vmf) 4137 { 4138 unsigned long address = vmf->address, nr_pages, mask; 4139 pgoff_t start_pgoff = vmf->pgoff; 4140 pgoff_t end_pgoff; 4141 int off; 4142 4143 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 4144 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 4145 4146 address = max(address & mask, vmf->vma->vm_start); 4147 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 4148 start_pgoff -= off; 4149 4150 /* 4151 * end_pgoff is either the end of the page table, the end of 4152 * the vma or nr_pages from start_pgoff, depending what is nearest. 4153 */ 4154 end_pgoff = start_pgoff - 4155 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 4156 PTRS_PER_PTE - 1; 4157 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1, 4158 start_pgoff + nr_pages - 1); 4159 4160 if (pmd_none(*vmf->pmd)) { 4161 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 4162 if (!vmf->prealloc_pte) 4163 return VM_FAULT_OOM; 4164 } 4165 4166 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff); 4167 } 4168 4169 static vm_fault_t do_read_fault(struct vm_fault *vmf) 4170 { 4171 struct vm_area_struct *vma = vmf->vma; 4172 vm_fault_t ret = 0; 4173 4174 /* 4175 * Let's call ->map_pages() first and use ->fault() as fallback 4176 * if page by the offset is not ready to be mapped (cold cache or 4177 * something). 4178 */ 4179 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 4180 if (likely(!userfaultfd_minor(vmf->vma))) { 4181 ret = do_fault_around(vmf); 4182 if (ret) 4183 return ret; 4184 } 4185 } 4186 4187 ret = __do_fault(vmf); 4188 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4189 return ret; 4190 4191 ret |= finish_fault(vmf); 4192 unlock_page(vmf->page); 4193 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4194 put_page(vmf->page); 4195 return ret; 4196 } 4197 4198 static vm_fault_t do_cow_fault(struct vm_fault *vmf) 4199 { 4200 struct vm_area_struct *vma = vmf->vma; 4201 vm_fault_t ret; 4202 4203 if (unlikely(anon_vma_prepare(vma))) 4204 return VM_FAULT_OOM; 4205 4206 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address); 4207 if (!vmf->cow_page) 4208 return VM_FAULT_OOM; 4209 4210 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm, 4211 GFP_KERNEL)) { 4212 put_page(vmf->cow_page); 4213 return VM_FAULT_OOM; 4214 } 4215 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL); 4216 4217 ret = __do_fault(vmf); 4218 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4219 goto uncharge_out; 4220 if (ret & VM_FAULT_DONE_COW) 4221 return ret; 4222 4223 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 4224 __SetPageUptodate(vmf->cow_page); 4225 4226 ret |= finish_fault(vmf); 4227 unlock_page(vmf->page); 4228 put_page(vmf->page); 4229 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4230 goto uncharge_out; 4231 return ret; 4232 uncharge_out: 4233 put_page(vmf->cow_page); 4234 return ret; 4235 } 4236 4237 static vm_fault_t do_shared_fault(struct vm_fault *vmf) 4238 { 4239 struct vm_area_struct *vma = vmf->vma; 4240 vm_fault_t ret, tmp; 4241 4242 ret = __do_fault(vmf); 4243 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4244 return ret; 4245 4246 /* 4247 * Check if the backing address space wants to know that the page is 4248 * about to become writable 4249 */ 4250 if (vma->vm_ops->page_mkwrite) { 4251 unlock_page(vmf->page); 4252 tmp = do_page_mkwrite(vmf); 4253 if (unlikely(!tmp || 4254 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 4255 put_page(vmf->page); 4256 return tmp; 4257 } 4258 } 4259 4260 ret |= finish_fault(vmf); 4261 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 4262 VM_FAULT_RETRY))) { 4263 unlock_page(vmf->page); 4264 put_page(vmf->page); 4265 return ret; 4266 } 4267 4268 ret |= fault_dirty_shared_page(vmf); 4269 return ret; 4270 } 4271 4272 /* 4273 * We enter with non-exclusive mmap_lock (to exclude vma changes, 4274 * but allow concurrent faults). 4275 * The mmap_lock may have been released depending on flags and our 4276 * return value. See filemap_fault() and __folio_lock_or_retry(). 4277 * If mmap_lock is released, vma may become invalid (for example 4278 * by other thread calling munmap()). 4279 */ 4280 static vm_fault_t do_fault(struct vm_fault *vmf) 4281 { 4282 struct vm_area_struct *vma = vmf->vma; 4283 struct mm_struct *vm_mm = vma->vm_mm; 4284 vm_fault_t ret; 4285 4286 /* 4287 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 4288 */ 4289 if (!vma->vm_ops->fault) { 4290 /* 4291 * If we find a migration pmd entry or a none pmd entry, which 4292 * should never happen, return SIGBUS 4293 */ 4294 if (unlikely(!pmd_present(*vmf->pmd))) 4295 ret = VM_FAULT_SIGBUS; 4296 else { 4297 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, 4298 vmf->pmd, 4299 vmf->address, 4300 &vmf->ptl); 4301 /* 4302 * Make sure this is not a temporary clearing of pte 4303 * by holding ptl and checking again. A R/M/W update 4304 * of pte involves: take ptl, clearing the pte so that 4305 * we don't have concurrent modification by hardware 4306 * followed by an update. 4307 */ 4308 if (unlikely(pte_none(*vmf->pte))) 4309 ret = VM_FAULT_SIGBUS; 4310 else 4311 ret = VM_FAULT_NOPAGE; 4312 4313 pte_unmap_unlock(vmf->pte, vmf->ptl); 4314 } 4315 } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 4316 ret = do_read_fault(vmf); 4317 else if (!(vma->vm_flags & VM_SHARED)) 4318 ret = do_cow_fault(vmf); 4319 else 4320 ret = do_shared_fault(vmf); 4321 4322 /* preallocated pagetable is unused: free it */ 4323 if (vmf->prealloc_pte) { 4324 pte_free(vm_mm, vmf->prealloc_pte); 4325 vmf->prealloc_pte = NULL; 4326 } 4327 return ret; 4328 } 4329 4330 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 4331 unsigned long addr, int page_nid, int *flags) 4332 { 4333 get_page(page); 4334 4335 count_vm_numa_event(NUMA_HINT_FAULTS); 4336 if (page_nid == numa_node_id()) { 4337 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 4338 *flags |= TNF_FAULT_LOCAL; 4339 } 4340 4341 return mpol_misplaced(page, vma, addr); 4342 } 4343 4344 static vm_fault_t do_numa_page(struct vm_fault *vmf) 4345 { 4346 struct vm_area_struct *vma = vmf->vma; 4347 struct page *page = NULL; 4348 int page_nid = NUMA_NO_NODE; 4349 int last_cpupid; 4350 int target_nid; 4351 pte_t pte, old_pte; 4352 bool was_writable = pte_savedwrite(vmf->orig_pte); 4353 int flags = 0; 4354 4355 /* 4356 * The "pte" at this point cannot be used safely without 4357 * validation through pte_unmap_same(). It's of NUMA type but 4358 * the pfn may be screwed if the read is non atomic. 4359 */ 4360 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd); 4361 spin_lock(vmf->ptl); 4362 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4363 pte_unmap_unlock(vmf->pte, vmf->ptl); 4364 goto out; 4365 } 4366 4367 /* Get the normal PTE */ 4368 old_pte = ptep_get(vmf->pte); 4369 pte = pte_modify(old_pte, vma->vm_page_prot); 4370 4371 page = vm_normal_page(vma, vmf->address, pte); 4372 if (!page) 4373 goto out_map; 4374 4375 /* TODO: handle PTE-mapped THP */ 4376 if (PageCompound(page)) 4377 goto out_map; 4378 4379 /* 4380 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 4381 * much anyway since they can be in shared cache state. This misses 4382 * the case where a mapping is writable but the process never writes 4383 * to it but pte_write gets cleared during protection updates and 4384 * pte_dirty has unpredictable behaviour between PTE scan updates, 4385 * background writeback, dirty balancing and application behaviour. 4386 */ 4387 if (!was_writable) 4388 flags |= TNF_NO_GROUP; 4389 4390 /* 4391 * Flag if the page is shared between multiple address spaces. This 4392 * is later used when determining whether to group tasks together 4393 */ 4394 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 4395 flags |= TNF_SHARED; 4396 4397 last_cpupid = page_cpupid_last(page); 4398 page_nid = page_to_nid(page); 4399 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid, 4400 &flags); 4401 if (target_nid == NUMA_NO_NODE) { 4402 put_page(page); 4403 goto out_map; 4404 } 4405 pte_unmap_unlock(vmf->pte, vmf->ptl); 4406 4407 /* Migrate to the requested node */ 4408 if (migrate_misplaced_page(page, vma, target_nid)) { 4409 page_nid = target_nid; 4410 flags |= TNF_MIGRATED; 4411 } else { 4412 flags |= TNF_MIGRATE_FAIL; 4413 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4414 spin_lock(vmf->ptl); 4415 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4416 pte_unmap_unlock(vmf->pte, vmf->ptl); 4417 goto out; 4418 } 4419 goto out_map; 4420 } 4421 4422 out: 4423 if (page_nid != NUMA_NO_NODE) 4424 task_numa_fault(last_cpupid, page_nid, 1, flags); 4425 return 0; 4426 out_map: 4427 /* 4428 * Make it present again, depending on how arch implements 4429 * non-accessible ptes, some can allow access by kernel mode. 4430 */ 4431 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 4432 pte = pte_modify(old_pte, vma->vm_page_prot); 4433 pte = pte_mkyoung(pte); 4434 if (was_writable) 4435 pte = pte_mkwrite(pte); 4436 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 4437 update_mmu_cache(vma, vmf->address, vmf->pte); 4438 pte_unmap_unlock(vmf->pte, vmf->ptl); 4439 goto out; 4440 } 4441 4442 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 4443 { 4444 if (vma_is_anonymous(vmf->vma)) 4445 return do_huge_pmd_anonymous_page(vmf); 4446 if (vmf->vma->vm_ops->huge_fault) 4447 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4448 return VM_FAULT_FALLBACK; 4449 } 4450 4451 /* `inline' is required to avoid gcc 4.1.2 build error */ 4452 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf) 4453 { 4454 if (vma_is_anonymous(vmf->vma)) { 4455 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd)) 4456 return handle_userfault(vmf, VM_UFFD_WP); 4457 return do_huge_pmd_wp_page(vmf); 4458 } 4459 if (vmf->vma->vm_ops->huge_fault) { 4460 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4461 4462 if (!(ret & VM_FAULT_FALLBACK)) 4463 return ret; 4464 } 4465 4466 /* COW or write-notify handled on pte level: split pmd. */ 4467 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL); 4468 4469 return VM_FAULT_FALLBACK; 4470 } 4471 4472 static vm_fault_t create_huge_pud(struct vm_fault *vmf) 4473 { 4474 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 4475 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 4476 /* No support for anonymous transparent PUD pages yet */ 4477 if (vma_is_anonymous(vmf->vma)) 4478 goto split; 4479 if (vmf->vma->vm_ops->huge_fault) { 4480 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4481 4482 if (!(ret & VM_FAULT_FALLBACK)) 4483 return ret; 4484 } 4485 split: 4486 /* COW or write-notify not handled on PUD level: split pud.*/ 4487 __split_huge_pud(vmf->vma, vmf->pud, vmf->address); 4488 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4489 return VM_FAULT_FALLBACK; 4490 } 4491 4492 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 4493 { 4494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4495 /* No support for anonymous transparent PUD pages yet */ 4496 if (vma_is_anonymous(vmf->vma)) 4497 return VM_FAULT_FALLBACK; 4498 if (vmf->vma->vm_ops->huge_fault) 4499 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4500 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4501 return VM_FAULT_FALLBACK; 4502 } 4503 4504 /* 4505 * These routines also need to handle stuff like marking pages dirty 4506 * and/or accessed for architectures that don't do it in hardware (most 4507 * RISC architectures). The early dirtying is also good on the i386. 4508 * 4509 * There is also a hook called "update_mmu_cache()" that architectures 4510 * with external mmu caches can use to update those (ie the Sparc or 4511 * PowerPC hashed page tables that act as extended TLBs). 4512 * 4513 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow 4514 * concurrent faults). 4515 * 4516 * The mmap_lock may have been released depending on flags and our return value. 4517 * See filemap_fault() and __folio_lock_or_retry(). 4518 */ 4519 static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 4520 { 4521 pte_t entry; 4522 4523 if (unlikely(pmd_none(*vmf->pmd))) { 4524 /* 4525 * Leave __pte_alloc() until later: because vm_ops->fault may 4526 * want to allocate huge page, and if we expose page table 4527 * for an instant, it will be difficult to retract from 4528 * concurrent faults and from rmap lookups. 4529 */ 4530 vmf->pte = NULL; 4531 } else { 4532 /* 4533 * If a huge pmd materialized under us just retry later. Use 4534 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead 4535 * of pmd_trans_huge() to ensure the pmd didn't become 4536 * pmd_trans_huge under us and then back to pmd_none, as a 4537 * result of MADV_DONTNEED running immediately after a huge pmd 4538 * fault in a different thread of this mm, in turn leading to a 4539 * misleading pmd_trans_huge() retval. All we have to ensure is 4540 * that it is a regular pmd that we can walk with 4541 * pte_offset_map() and we can do that through an atomic read 4542 * in C, which is what pmd_trans_unstable() provides. 4543 */ 4544 if (pmd_devmap_trans_unstable(vmf->pmd)) 4545 return 0; 4546 /* 4547 * A regular pmd is established and it can't morph into a huge 4548 * pmd from under us anymore at this point because we hold the 4549 * mmap_lock read mode and khugepaged takes it in write mode. 4550 * So now it's safe to run pte_offset_map(). 4551 */ 4552 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4553 vmf->orig_pte = *vmf->pte; 4554 4555 /* 4556 * some architectures can have larger ptes than wordsize, 4557 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and 4558 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic 4559 * accesses. The code below just needs a consistent view 4560 * for the ifs and we later double check anyway with the 4561 * ptl lock held. So here a barrier will do. 4562 */ 4563 barrier(); 4564 if (pte_none(vmf->orig_pte)) { 4565 pte_unmap(vmf->pte); 4566 vmf->pte = NULL; 4567 } 4568 } 4569 4570 if (!vmf->pte) { 4571 if (vma_is_anonymous(vmf->vma)) 4572 return do_anonymous_page(vmf); 4573 else 4574 return do_fault(vmf); 4575 } 4576 4577 if (!pte_present(vmf->orig_pte)) 4578 return do_swap_page(vmf); 4579 4580 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 4581 return do_numa_page(vmf); 4582 4583 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 4584 spin_lock(vmf->ptl); 4585 entry = vmf->orig_pte; 4586 if (unlikely(!pte_same(*vmf->pte, entry))) { 4587 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 4588 goto unlock; 4589 } 4590 if (vmf->flags & FAULT_FLAG_WRITE) { 4591 if (!pte_write(entry)) 4592 return do_wp_page(vmf); 4593 entry = pte_mkdirty(entry); 4594 } 4595 entry = pte_mkyoung(entry); 4596 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 4597 vmf->flags & FAULT_FLAG_WRITE)) { 4598 update_mmu_cache(vmf->vma, vmf->address, vmf->pte); 4599 } else { 4600 /* Skip spurious TLB flush for retried page fault */ 4601 if (vmf->flags & FAULT_FLAG_TRIED) 4602 goto unlock; 4603 /* 4604 * This is needed only for protection faults but the arch code 4605 * is not yet telling us if this is a protection fault or not. 4606 * This still avoids useless tlb flushes for .text page faults 4607 * with threads. 4608 */ 4609 if (vmf->flags & FAULT_FLAG_WRITE) 4610 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address); 4611 } 4612 unlock: 4613 pte_unmap_unlock(vmf->pte, vmf->ptl); 4614 return 0; 4615 } 4616 4617 /* 4618 * By the time we get here, we already hold the mm semaphore 4619 * 4620 * The mmap_lock may have been released depending on flags and our 4621 * return value. See filemap_fault() and __folio_lock_or_retry(). 4622 */ 4623 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 4624 unsigned long address, unsigned int flags) 4625 { 4626 struct vm_fault vmf = { 4627 .vma = vma, 4628 .address = address & PAGE_MASK, 4629 .flags = flags, 4630 .pgoff = linear_page_index(vma, address), 4631 .gfp_mask = __get_fault_gfp_mask(vma), 4632 }; 4633 unsigned int dirty = flags & FAULT_FLAG_WRITE; 4634 struct mm_struct *mm = vma->vm_mm; 4635 pgd_t *pgd; 4636 p4d_t *p4d; 4637 vm_fault_t ret; 4638 4639 pgd = pgd_offset(mm, address); 4640 p4d = p4d_alloc(mm, pgd, address); 4641 if (!p4d) 4642 return VM_FAULT_OOM; 4643 4644 vmf.pud = pud_alloc(mm, p4d, address); 4645 if (!vmf.pud) 4646 return VM_FAULT_OOM; 4647 retry_pud: 4648 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) { 4649 ret = create_huge_pud(&vmf); 4650 if (!(ret & VM_FAULT_FALLBACK)) 4651 return ret; 4652 } else { 4653 pud_t orig_pud = *vmf.pud; 4654 4655 barrier(); 4656 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 4657 4658 /* NUMA case for anonymous PUDs would go here */ 4659 4660 if (dirty && !pud_write(orig_pud)) { 4661 ret = wp_huge_pud(&vmf, orig_pud); 4662 if (!(ret & VM_FAULT_FALLBACK)) 4663 return ret; 4664 } else { 4665 huge_pud_set_accessed(&vmf, orig_pud); 4666 return 0; 4667 } 4668 } 4669 } 4670 4671 vmf.pmd = pmd_alloc(mm, vmf.pud, address); 4672 if (!vmf.pmd) 4673 return VM_FAULT_OOM; 4674 4675 /* Huge pud page fault raced with pmd_alloc? */ 4676 if (pud_trans_unstable(vmf.pud)) 4677 goto retry_pud; 4678 4679 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) { 4680 ret = create_huge_pmd(&vmf); 4681 if (!(ret & VM_FAULT_FALLBACK)) 4682 return ret; 4683 } else { 4684 vmf.orig_pmd = *vmf.pmd; 4685 4686 barrier(); 4687 if (unlikely(is_swap_pmd(vmf.orig_pmd))) { 4688 VM_BUG_ON(thp_migration_supported() && 4689 !is_pmd_migration_entry(vmf.orig_pmd)); 4690 if (is_pmd_migration_entry(vmf.orig_pmd)) 4691 pmd_migration_entry_wait(mm, vmf.pmd); 4692 return 0; 4693 } 4694 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) { 4695 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma)) 4696 return do_huge_pmd_numa_page(&vmf); 4697 4698 if (dirty && !pmd_write(vmf.orig_pmd)) { 4699 ret = wp_huge_pmd(&vmf); 4700 if (!(ret & VM_FAULT_FALLBACK)) 4701 return ret; 4702 } else { 4703 huge_pmd_set_accessed(&vmf); 4704 return 0; 4705 } 4706 } 4707 } 4708 4709 return handle_pte_fault(&vmf); 4710 } 4711 4712 /** 4713 * mm_account_fault - Do page fault accounting 4714 * 4715 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting 4716 * of perf event counters, but we'll still do the per-task accounting to 4717 * the task who triggered this page fault. 4718 * @address: the faulted address. 4719 * @flags: the fault flags. 4720 * @ret: the fault retcode. 4721 * 4722 * This will take care of most of the page fault accounting. Meanwhile, it 4723 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter 4724 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should 4725 * still be in per-arch page fault handlers at the entry of page fault. 4726 */ 4727 static inline void mm_account_fault(struct pt_regs *regs, 4728 unsigned long address, unsigned int flags, 4729 vm_fault_t ret) 4730 { 4731 bool major; 4732 4733 /* 4734 * We don't do accounting for some specific faults: 4735 * 4736 * - Unsuccessful faults (e.g. when the address wasn't valid). That 4737 * includes arch_vma_access_permitted() failing before reaching here. 4738 * So this is not a "this many hardware page faults" counter. We 4739 * should use the hw profiling for that. 4740 * 4741 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted 4742 * once they're completed. 4743 */ 4744 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY)) 4745 return; 4746 4747 /* 4748 * We define the fault as a major fault when the final successful fault 4749 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't 4750 * handle it immediately previously). 4751 */ 4752 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); 4753 4754 if (major) 4755 current->maj_flt++; 4756 else 4757 current->min_flt++; 4758 4759 /* 4760 * If the fault is done for GUP, regs will be NULL. We only do the 4761 * accounting for the per thread fault counters who triggered the 4762 * fault, and we skip the perf event updates. 4763 */ 4764 if (!regs) 4765 return; 4766 4767 if (major) 4768 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 4769 else 4770 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 4771 } 4772 4773 /* 4774 * By the time we get here, we already hold the mm semaphore 4775 * 4776 * The mmap_lock may have been released depending on flags and our 4777 * return value. See filemap_fault() and __folio_lock_or_retry(). 4778 */ 4779 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 4780 unsigned int flags, struct pt_regs *regs) 4781 { 4782 vm_fault_t ret; 4783 4784 __set_current_state(TASK_RUNNING); 4785 4786 count_vm_event(PGFAULT); 4787 count_memcg_event_mm(vma->vm_mm, PGFAULT); 4788 4789 /* do counter updates before entering really critical section. */ 4790 check_sync_rss_stat(current); 4791 4792 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 4793 flags & FAULT_FLAG_INSTRUCTION, 4794 flags & FAULT_FLAG_REMOTE)) 4795 return VM_FAULT_SIGSEGV; 4796 4797 /* 4798 * Enable the memcg OOM handling for faults triggered in user 4799 * space. Kernel faults are handled more gracefully. 4800 */ 4801 if (flags & FAULT_FLAG_USER) 4802 mem_cgroup_enter_user_fault(); 4803 4804 if (unlikely(is_vm_hugetlb_page(vma))) 4805 ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 4806 else 4807 ret = __handle_mm_fault(vma, address, flags); 4808 4809 if (flags & FAULT_FLAG_USER) { 4810 mem_cgroup_exit_user_fault(); 4811 /* 4812 * The task may have entered a memcg OOM situation but 4813 * if the allocation error was handled gracefully (no 4814 * VM_FAULT_OOM), there is no need to kill anything. 4815 * Just clean up the OOM state peacefully. 4816 */ 4817 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 4818 mem_cgroup_oom_synchronize(false); 4819 } 4820 4821 mm_account_fault(regs, address, flags, ret); 4822 4823 return ret; 4824 } 4825 EXPORT_SYMBOL_GPL(handle_mm_fault); 4826 4827 #ifndef __PAGETABLE_P4D_FOLDED 4828 /* 4829 * Allocate p4d page table. 4830 * We've already handled the fast-path in-line. 4831 */ 4832 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 4833 { 4834 p4d_t *new = p4d_alloc_one(mm, address); 4835 if (!new) 4836 return -ENOMEM; 4837 4838 spin_lock(&mm->page_table_lock); 4839 if (pgd_present(*pgd)) { /* Another has populated it */ 4840 p4d_free(mm, new); 4841 } else { 4842 smp_wmb(); /* See comment in pmd_install() */ 4843 pgd_populate(mm, pgd, new); 4844 } 4845 spin_unlock(&mm->page_table_lock); 4846 return 0; 4847 } 4848 #endif /* __PAGETABLE_P4D_FOLDED */ 4849 4850 #ifndef __PAGETABLE_PUD_FOLDED 4851 /* 4852 * Allocate page upper directory. 4853 * We've already handled the fast-path in-line. 4854 */ 4855 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 4856 { 4857 pud_t *new = pud_alloc_one(mm, address); 4858 if (!new) 4859 return -ENOMEM; 4860 4861 spin_lock(&mm->page_table_lock); 4862 if (!p4d_present(*p4d)) { 4863 mm_inc_nr_puds(mm); 4864 smp_wmb(); /* See comment in pmd_install() */ 4865 p4d_populate(mm, p4d, new); 4866 } else /* Another has populated it */ 4867 pud_free(mm, new); 4868 spin_unlock(&mm->page_table_lock); 4869 return 0; 4870 } 4871 #endif /* __PAGETABLE_PUD_FOLDED */ 4872 4873 #ifndef __PAGETABLE_PMD_FOLDED 4874 /* 4875 * Allocate page middle directory. 4876 * We've already handled the fast-path in-line. 4877 */ 4878 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 4879 { 4880 spinlock_t *ptl; 4881 pmd_t *new = pmd_alloc_one(mm, address); 4882 if (!new) 4883 return -ENOMEM; 4884 4885 ptl = pud_lock(mm, pud); 4886 if (!pud_present(*pud)) { 4887 mm_inc_nr_pmds(mm); 4888 smp_wmb(); /* See comment in pmd_install() */ 4889 pud_populate(mm, pud, new); 4890 } else { /* Another has populated it */ 4891 pmd_free(mm, new); 4892 } 4893 spin_unlock(ptl); 4894 return 0; 4895 } 4896 #endif /* __PAGETABLE_PMD_FOLDED */ 4897 4898 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address, 4899 struct mmu_notifier_range *range, pte_t **ptepp, 4900 pmd_t **pmdpp, spinlock_t **ptlp) 4901 { 4902 pgd_t *pgd; 4903 p4d_t *p4d; 4904 pud_t *pud; 4905 pmd_t *pmd; 4906 pte_t *ptep; 4907 4908 pgd = pgd_offset(mm, address); 4909 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 4910 goto out; 4911 4912 p4d = p4d_offset(pgd, address); 4913 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 4914 goto out; 4915 4916 pud = pud_offset(p4d, address); 4917 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 4918 goto out; 4919 4920 pmd = pmd_offset(pud, address); 4921 VM_BUG_ON(pmd_trans_huge(*pmd)); 4922 4923 if (pmd_huge(*pmd)) { 4924 if (!pmdpp) 4925 goto out; 4926 4927 if (range) { 4928 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, 4929 NULL, mm, address & PMD_MASK, 4930 (address & PMD_MASK) + PMD_SIZE); 4931 mmu_notifier_invalidate_range_start(range); 4932 } 4933 *ptlp = pmd_lock(mm, pmd); 4934 if (pmd_huge(*pmd)) { 4935 *pmdpp = pmd; 4936 return 0; 4937 } 4938 spin_unlock(*ptlp); 4939 if (range) 4940 mmu_notifier_invalidate_range_end(range); 4941 } 4942 4943 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 4944 goto out; 4945 4946 if (range) { 4947 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 4948 address & PAGE_MASK, 4949 (address & PAGE_MASK) + PAGE_SIZE); 4950 mmu_notifier_invalidate_range_start(range); 4951 } 4952 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 4953 if (!pte_present(*ptep)) 4954 goto unlock; 4955 *ptepp = ptep; 4956 return 0; 4957 unlock: 4958 pte_unmap_unlock(ptep, *ptlp); 4959 if (range) 4960 mmu_notifier_invalidate_range_end(range); 4961 out: 4962 return -EINVAL; 4963 } 4964 4965 /** 4966 * follow_pte - look up PTE at a user virtual address 4967 * @mm: the mm_struct of the target address space 4968 * @address: user virtual address 4969 * @ptepp: location to store found PTE 4970 * @ptlp: location to store the lock for the PTE 4971 * 4972 * On a successful return, the pointer to the PTE is stored in @ptepp; 4973 * the corresponding lock is taken and its location is stored in @ptlp. 4974 * The contents of the PTE are only stable until @ptlp is released; 4975 * any further use, if any, must be protected against invalidation 4976 * with MMU notifiers. 4977 * 4978 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore 4979 * should be taken for read. 4980 * 4981 * KVM uses this function. While it is arguably less bad than ``follow_pfn``, 4982 * it is not a good general-purpose API. 4983 * 4984 * Return: zero on success, -ve otherwise. 4985 */ 4986 int follow_pte(struct mm_struct *mm, unsigned long address, 4987 pte_t **ptepp, spinlock_t **ptlp) 4988 { 4989 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp); 4990 } 4991 EXPORT_SYMBOL_GPL(follow_pte); 4992 4993 /** 4994 * follow_pfn - look up PFN at a user virtual address 4995 * @vma: memory mapping 4996 * @address: user virtual address 4997 * @pfn: location to store found PFN 4998 * 4999 * Only IO mappings and raw PFN mappings are allowed. 5000 * 5001 * This function does not allow the caller to read the permissions 5002 * of the PTE. Do not use it. 5003 * 5004 * Return: zero and the pfn at @pfn on success, -ve otherwise. 5005 */ 5006 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 5007 unsigned long *pfn) 5008 { 5009 int ret = -EINVAL; 5010 spinlock_t *ptl; 5011 pte_t *ptep; 5012 5013 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5014 return ret; 5015 5016 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 5017 if (ret) 5018 return ret; 5019 *pfn = pte_pfn(*ptep); 5020 pte_unmap_unlock(ptep, ptl); 5021 return 0; 5022 } 5023 EXPORT_SYMBOL(follow_pfn); 5024 5025 #ifdef CONFIG_HAVE_IOREMAP_PROT 5026 int follow_phys(struct vm_area_struct *vma, 5027 unsigned long address, unsigned int flags, 5028 unsigned long *prot, resource_size_t *phys) 5029 { 5030 int ret = -EINVAL; 5031 pte_t *ptep, pte; 5032 spinlock_t *ptl; 5033 5034 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5035 goto out; 5036 5037 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 5038 goto out; 5039 pte = *ptep; 5040 5041 if ((flags & FOLL_WRITE) && !pte_write(pte)) 5042 goto unlock; 5043 5044 *prot = pgprot_val(pte_pgprot(pte)); 5045 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 5046 5047 ret = 0; 5048 unlock: 5049 pte_unmap_unlock(ptep, ptl); 5050 out: 5051 return ret; 5052 } 5053 5054 /** 5055 * generic_access_phys - generic implementation for iomem mmap access 5056 * @vma: the vma to access 5057 * @addr: userspace address, not relative offset within @vma 5058 * @buf: buffer to read/write 5059 * @len: length of transfer 5060 * @write: set to FOLL_WRITE when writing, otherwise reading 5061 * 5062 * This is a generic implementation for &vm_operations_struct.access for an 5063 * iomem mapping. This callback is used by access_process_vm() when the @vma is 5064 * not page based. 5065 */ 5066 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 5067 void *buf, int len, int write) 5068 { 5069 resource_size_t phys_addr; 5070 unsigned long prot = 0; 5071 void __iomem *maddr; 5072 pte_t *ptep, pte; 5073 spinlock_t *ptl; 5074 int offset = offset_in_page(addr); 5075 int ret = -EINVAL; 5076 5077 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5078 return -EINVAL; 5079 5080 retry: 5081 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 5082 return -EINVAL; 5083 pte = *ptep; 5084 pte_unmap_unlock(ptep, ptl); 5085 5086 prot = pgprot_val(pte_pgprot(pte)); 5087 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 5088 5089 if ((write & FOLL_WRITE) && !pte_write(pte)) 5090 return -EINVAL; 5091 5092 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 5093 if (!maddr) 5094 return -ENOMEM; 5095 5096 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 5097 goto out_unmap; 5098 5099 if (!pte_same(pte, *ptep)) { 5100 pte_unmap_unlock(ptep, ptl); 5101 iounmap(maddr); 5102 5103 goto retry; 5104 } 5105 5106 if (write) 5107 memcpy_toio(maddr + offset, buf, len); 5108 else 5109 memcpy_fromio(buf, maddr + offset, len); 5110 ret = len; 5111 pte_unmap_unlock(ptep, ptl); 5112 out_unmap: 5113 iounmap(maddr); 5114 5115 return ret; 5116 } 5117 EXPORT_SYMBOL_GPL(generic_access_phys); 5118 #endif 5119 5120 /* 5121 * Access another process' address space as given in mm. 5122 */ 5123 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf, 5124 int len, unsigned int gup_flags) 5125 { 5126 struct vm_area_struct *vma; 5127 void *old_buf = buf; 5128 int write = gup_flags & FOLL_WRITE; 5129 5130 if (mmap_read_lock_killable(mm)) 5131 return 0; 5132 5133 /* ignore errors, just check how much was successfully transferred */ 5134 while (len) { 5135 int bytes, ret, offset; 5136 void *maddr; 5137 struct page *page = NULL; 5138 5139 ret = get_user_pages_remote(mm, addr, 1, 5140 gup_flags, &page, &vma, NULL); 5141 if (ret <= 0) { 5142 #ifndef CONFIG_HAVE_IOREMAP_PROT 5143 break; 5144 #else 5145 /* 5146 * Check if this is a VM_IO | VM_PFNMAP VMA, which 5147 * we can access using slightly different code. 5148 */ 5149 vma = vma_lookup(mm, addr); 5150 if (!vma) 5151 break; 5152 if (vma->vm_ops && vma->vm_ops->access) 5153 ret = vma->vm_ops->access(vma, addr, buf, 5154 len, write); 5155 if (ret <= 0) 5156 break; 5157 bytes = ret; 5158 #endif 5159 } else { 5160 bytes = len; 5161 offset = addr & (PAGE_SIZE-1); 5162 if (bytes > PAGE_SIZE-offset) 5163 bytes = PAGE_SIZE-offset; 5164 5165 maddr = kmap(page); 5166 if (write) { 5167 copy_to_user_page(vma, page, addr, 5168 maddr + offset, buf, bytes); 5169 set_page_dirty_lock(page); 5170 } else { 5171 copy_from_user_page(vma, page, addr, 5172 buf, maddr + offset, bytes); 5173 } 5174 kunmap(page); 5175 put_page(page); 5176 } 5177 len -= bytes; 5178 buf += bytes; 5179 addr += bytes; 5180 } 5181 mmap_read_unlock(mm); 5182 5183 return buf - old_buf; 5184 } 5185 5186 /** 5187 * access_remote_vm - access another process' address space 5188 * @mm: the mm_struct of the target address space 5189 * @addr: start address to access 5190 * @buf: source or destination buffer 5191 * @len: number of bytes to transfer 5192 * @gup_flags: flags modifying lookup behaviour 5193 * 5194 * The caller must hold a reference on @mm. 5195 * 5196 * Return: number of bytes copied from source to destination. 5197 */ 5198 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 5199 void *buf, int len, unsigned int gup_flags) 5200 { 5201 return __access_remote_vm(mm, addr, buf, len, gup_flags); 5202 } 5203 5204 /* 5205 * Access another process' address space. 5206 * Source/target buffer must be kernel space, 5207 * Do not walk the page table directly, use get_user_pages 5208 */ 5209 int access_process_vm(struct task_struct *tsk, unsigned long addr, 5210 void *buf, int len, unsigned int gup_flags) 5211 { 5212 struct mm_struct *mm; 5213 int ret; 5214 5215 mm = get_task_mm(tsk); 5216 if (!mm) 5217 return 0; 5218 5219 ret = __access_remote_vm(mm, addr, buf, len, gup_flags); 5220 5221 mmput(mm); 5222 5223 return ret; 5224 } 5225 EXPORT_SYMBOL_GPL(access_process_vm); 5226 5227 /* 5228 * Print the name of a VMA. 5229 */ 5230 void print_vma_addr(char *prefix, unsigned long ip) 5231 { 5232 struct mm_struct *mm = current->mm; 5233 struct vm_area_struct *vma; 5234 5235 /* 5236 * we might be running from an atomic context so we cannot sleep 5237 */ 5238 if (!mmap_read_trylock(mm)) 5239 return; 5240 5241 vma = find_vma(mm, ip); 5242 if (vma && vma->vm_file) { 5243 struct file *f = vma->vm_file; 5244 char *buf = (char *)__get_free_page(GFP_NOWAIT); 5245 if (buf) { 5246 char *p; 5247 5248 p = file_path(f, buf, PAGE_SIZE); 5249 if (IS_ERR(p)) 5250 p = "?"; 5251 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 5252 vma->vm_start, 5253 vma->vm_end - vma->vm_start); 5254 free_page((unsigned long)buf); 5255 } 5256 } 5257 mmap_read_unlock(mm); 5258 } 5259 5260 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5261 void __might_fault(const char *file, int line) 5262 { 5263 /* 5264 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 5265 * holding the mmap_lock, this is safe because kernel memory doesn't 5266 * get paged out, therefore we'll never actually fault, and the 5267 * below annotations will generate false positives. 5268 */ 5269 if (uaccess_kernel()) 5270 return; 5271 if (pagefault_disabled()) 5272 return; 5273 __might_sleep(file, line); 5274 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5275 if (current->mm) 5276 might_lock_read(¤t->mm->mmap_lock); 5277 #endif 5278 } 5279 EXPORT_SYMBOL(__might_fault); 5280 #endif 5281 5282 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 5283 /* 5284 * Process all subpages of the specified huge page with the specified 5285 * operation. The target subpage will be processed last to keep its 5286 * cache lines hot. 5287 */ 5288 static inline void process_huge_page( 5289 unsigned long addr_hint, unsigned int pages_per_huge_page, 5290 void (*process_subpage)(unsigned long addr, int idx, void *arg), 5291 void *arg) 5292 { 5293 int i, n, base, l; 5294 unsigned long addr = addr_hint & 5295 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5296 5297 /* Process target subpage last to keep its cache lines hot */ 5298 might_sleep(); 5299 n = (addr_hint - addr) / PAGE_SIZE; 5300 if (2 * n <= pages_per_huge_page) { 5301 /* If target subpage in first half of huge page */ 5302 base = 0; 5303 l = n; 5304 /* Process subpages at the end of huge page */ 5305 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 5306 cond_resched(); 5307 process_subpage(addr + i * PAGE_SIZE, i, arg); 5308 } 5309 } else { 5310 /* If target subpage in second half of huge page */ 5311 base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 5312 l = pages_per_huge_page - n; 5313 /* Process subpages at the begin of huge page */ 5314 for (i = 0; i < base; i++) { 5315 cond_resched(); 5316 process_subpage(addr + i * PAGE_SIZE, i, arg); 5317 } 5318 } 5319 /* 5320 * Process remaining subpages in left-right-left-right pattern 5321 * towards the target subpage 5322 */ 5323 for (i = 0; i < l; i++) { 5324 int left_idx = base + i; 5325 int right_idx = base + 2 * l - 1 - i; 5326 5327 cond_resched(); 5328 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 5329 cond_resched(); 5330 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 5331 } 5332 } 5333 5334 static void clear_gigantic_page(struct page *page, 5335 unsigned long addr, 5336 unsigned int pages_per_huge_page) 5337 { 5338 int i; 5339 struct page *p = page; 5340 5341 might_sleep(); 5342 for (i = 0; i < pages_per_huge_page; 5343 i++, p = mem_map_next(p, page, i)) { 5344 cond_resched(); 5345 clear_user_highpage(p, addr + i * PAGE_SIZE); 5346 } 5347 } 5348 5349 static void clear_subpage(unsigned long addr, int idx, void *arg) 5350 { 5351 struct page *page = arg; 5352 5353 clear_user_highpage(page + idx, addr); 5354 } 5355 5356 void clear_huge_page(struct page *page, 5357 unsigned long addr_hint, unsigned int pages_per_huge_page) 5358 { 5359 unsigned long addr = addr_hint & 5360 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5361 5362 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5363 clear_gigantic_page(page, addr, pages_per_huge_page); 5364 return; 5365 } 5366 5367 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 5368 } 5369 5370 static void copy_user_gigantic_page(struct page *dst, struct page *src, 5371 unsigned long addr, 5372 struct vm_area_struct *vma, 5373 unsigned int pages_per_huge_page) 5374 { 5375 int i; 5376 struct page *dst_base = dst; 5377 struct page *src_base = src; 5378 5379 for (i = 0; i < pages_per_huge_page; ) { 5380 cond_resched(); 5381 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 5382 5383 i++; 5384 dst = mem_map_next(dst, dst_base, i); 5385 src = mem_map_next(src, src_base, i); 5386 } 5387 } 5388 5389 struct copy_subpage_arg { 5390 struct page *dst; 5391 struct page *src; 5392 struct vm_area_struct *vma; 5393 }; 5394 5395 static void copy_subpage(unsigned long addr, int idx, void *arg) 5396 { 5397 struct copy_subpage_arg *copy_arg = arg; 5398 5399 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx, 5400 addr, copy_arg->vma); 5401 } 5402 5403 void copy_user_huge_page(struct page *dst, struct page *src, 5404 unsigned long addr_hint, struct vm_area_struct *vma, 5405 unsigned int pages_per_huge_page) 5406 { 5407 unsigned long addr = addr_hint & 5408 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5409 struct copy_subpage_arg arg = { 5410 .dst = dst, 5411 .src = src, 5412 .vma = vma, 5413 }; 5414 5415 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5416 copy_user_gigantic_page(dst, src, addr, vma, 5417 pages_per_huge_page); 5418 return; 5419 } 5420 5421 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 5422 } 5423 5424 long copy_huge_page_from_user(struct page *dst_page, 5425 const void __user *usr_src, 5426 unsigned int pages_per_huge_page, 5427 bool allow_pagefault) 5428 { 5429 void *page_kaddr; 5430 unsigned long i, rc = 0; 5431 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE; 5432 struct page *subpage = dst_page; 5433 5434 for (i = 0; i < pages_per_huge_page; 5435 i++, subpage = mem_map_next(subpage, dst_page, i)) { 5436 if (allow_pagefault) 5437 page_kaddr = kmap(subpage); 5438 else 5439 page_kaddr = kmap_atomic(subpage); 5440 rc = copy_from_user(page_kaddr, 5441 usr_src + i * PAGE_SIZE, PAGE_SIZE); 5442 if (allow_pagefault) 5443 kunmap(subpage); 5444 else 5445 kunmap_atomic(page_kaddr); 5446 5447 ret_val -= (PAGE_SIZE - rc); 5448 if (rc) 5449 break; 5450 5451 cond_resched(); 5452 } 5453 return ret_val; 5454 } 5455 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 5456 5457 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 5458 5459 static struct kmem_cache *page_ptl_cachep; 5460 5461 void __init ptlock_cache_init(void) 5462 { 5463 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 5464 SLAB_PANIC, NULL); 5465 } 5466 5467 bool ptlock_alloc(struct page *page) 5468 { 5469 spinlock_t *ptl; 5470 5471 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 5472 if (!ptl) 5473 return false; 5474 page->ptl = ptl; 5475 return true; 5476 } 5477 5478 void ptlock_free(struct page *page) 5479 { 5480 kmem_cache_free(page_ptl_cachep, page->ptl); 5481 } 5482 #endif 5483