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/kmsan.h> 56 #include <linux/ksm.h> 57 #include <linux/rmap.h> 58 #include <linux/export.h> 59 #include <linux/delayacct.h> 60 #include <linux/init.h> 61 #include <linux/pfn_t.h> 62 #include <linux/writeback.h> 63 #include <linux/memcontrol.h> 64 #include <linux/mmu_notifier.h> 65 #include <linux/swapops.h> 66 #include <linux/elf.h> 67 #include <linux/gfp.h> 68 #include <linux/migrate.h> 69 #include <linux/string.h> 70 #include <linux/memory-tiers.h> 71 #include <linux/debugfs.h> 72 #include <linux/userfaultfd_k.h> 73 #include <linux/dax.h> 74 #include <linux/oom.h> 75 #include <linux/numa.h> 76 #include <linux/perf_event.h> 77 #include <linux/ptrace.h> 78 #include <linux/vmalloc.h> 79 #include <linux/sched/sysctl.h> 80 81 #include <trace/events/kmem.h> 82 83 #include <asm/io.h> 84 #include <asm/mmu_context.h> 85 #include <asm/pgalloc.h> 86 #include <linux/uaccess.h> 87 #include <asm/tlb.h> 88 #include <asm/tlbflush.h> 89 90 #include "pgalloc-track.h" 91 #include "internal.h" 92 #include "swap.h" 93 94 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) 95 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 96 #endif 97 98 #ifndef CONFIG_NUMA 99 unsigned long max_mapnr; 100 EXPORT_SYMBOL(max_mapnr); 101 102 struct page *mem_map; 103 EXPORT_SYMBOL(mem_map); 104 #endif 105 106 static vm_fault_t do_fault(struct vm_fault *vmf); 107 108 /* 109 * A number of key systems in x86 including ioremap() rely on the assumption 110 * that high_memory defines the upper bound on direct map memory, then end 111 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 112 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 113 * and ZONE_HIGHMEM. 114 */ 115 void *high_memory; 116 EXPORT_SYMBOL(high_memory); 117 118 /* 119 * Randomize the address space (stacks, mmaps, brk, etc.). 120 * 121 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 122 * as ancient (libc5 based) binaries can segfault. ) 123 */ 124 int randomize_va_space __read_mostly = 125 #ifdef CONFIG_COMPAT_BRK 126 1; 127 #else 128 2; 129 #endif 130 131 #ifndef arch_wants_old_prefaulted_pte 132 static inline bool arch_wants_old_prefaulted_pte(void) 133 { 134 /* 135 * Transitioning a PTE from 'old' to 'young' can be expensive on 136 * some architectures, even if it's performed in hardware. By 137 * default, "false" means prefaulted entries will be 'young'. 138 */ 139 return false; 140 } 141 #endif 142 143 static int __init disable_randmaps(char *s) 144 { 145 randomize_va_space = 0; 146 return 1; 147 } 148 __setup("norandmaps", disable_randmaps); 149 150 unsigned long zero_pfn __read_mostly; 151 EXPORT_SYMBOL(zero_pfn); 152 153 unsigned long highest_memmap_pfn __read_mostly; 154 155 /* 156 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 157 */ 158 static int __init init_zero_pfn(void) 159 { 160 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 161 return 0; 162 } 163 early_initcall(init_zero_pfn); 164 165 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count) 166 { 167 trace_rss_stat(mm, member, count); 168 } 169 170 #if defined(SPLIT_RSS_COUNTING) 171 172 void sync_mm_rss(struct mm_struct *mm) 173 { 174 int i; 175 176 for (i = 0; i < NR_MM_COUNTERS; i++) { 177 if (current->rss_stat.count[i]) { 178 add_mm_counter(mm, i, current->rss_stat.count[i]); 179 current->rss_stat.count[i] = 0; 180 } 181 } 182 current->rss_stat.events = 0; 183 } 184 185 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 186 { 187 struct task_struct *task = current; 188 189 if (likely(task->mm == mm)) 190 task->rss_stat.count[member] += val; 191 else 192 add_mm_counter(mm, member, val); 193 } 194 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 195 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 196 197 /* sync counter once per 64 page faults */ 198 #define TASK_RSS_EVENTS_THRESH (64) 199 static void check_sync_rss_stat(struct task_struct *task) 200 { 201 if (unlikely(task != current)) 202 return; 203 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 204 sync_mm_rss(task->mm); 205 } 206 #else /* SPLIT_RSS_COUNTING */ 207 208 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 209 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 210 211 static void check_sync_rss_stat(struct task_struct *task) 212 { 213 } 214 215 #endif /* SPLIT_RSS_COUNTING */ 216 217 /* 218 * Note: this doesn't free the actual pages themselves. That 219 * has been handled earlier when unmapping all the memory regions. 220 */ 221 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 222 unsigned long addr) 223 { 224 pgtable_t token = pmd_pgtable(*pmd); 225 pmd_clear(pmd); 226 pte_free_tlb(tlb, token, addr); 227 mm_dec_nr_ptes(tlb->mm); 228 } 229 230 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 231 unsigned long addr, unsigned long end, 232 unsigned long floor, unsigned long ceiling) 233 { 234 pmd_t *pmd; 235 unsigned long next; 236 unsigned long start; 237 238 start = addr; 239 pmd = pmd_offset(pud, addr); 240 do { 241 next = pmd_addr_end(addr, end); 242 if (pmd_none_or_clear_bad(pmd)) 243 continue; 244 free_pte_range(tlb, pmd, addr); 245 } while (pmd++, addr = next, addr != end); 246 247 start &= PUD_MASK; 248 if (start < floor) 249 return; 250 if (ceiling) { 251 ceiling &= PUD_MASK; 252 if (!ceiling) 253 return; 254 } 255 if (end - 1 > ceiling - 1) 256 return; 257 258 pmd = pmd_offset(pud, start); 259 pud_clear(pud); 260 pmd_free_tlb(tlb, pmd, start); 261 mm_dec_nr_pmds(tlb->mm); 262 } 263 264 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 265 unsigned long addr, unsigned long end, 266 unsigned long floor, unsigned long ceiling) 267 { 268 pud_t *pud; 269 unsigned long next; 270 unsigned long start; 271 272 start = addr; 273 pud = pud_offset(p4d, addr); 274 do { 275 next = pud_addr_end(addr, end); 276 if (pud_none_or_clear_bad(pud)) 277 continue; 278 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 279 } while (pud++, addr = next, addr != end); 280 281 start &= P4D_MASK; 282 if (start < floor) 283 return; 284 if (ceiling) { 285 ceiling &= P4D_MASK; 286 if (!ceiling) 287 return; 288 } 289 if (end - 1 > ceiling - 1) 290 return; 291 292 pud = pud_offset(p4d, start); 293 p4d_clear(p4d); 294 pud_free_tlb(tlb, pud, start); 295 mm_dec_nr_puds(tlb->mm); 296 } 297 298 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 299 unsigned long addr, unsigned long end, 300 unsigned long floor, unsigned long ceiling) 301 { 302 p4d_t *p4d; 303 unsigned long next; 304 unsigned long start; 305 306 start = addr; 307 p4d = p4d_offset(pgd, addr); 308 do { 309 next = p4d_addr_end(addr, end); 310 if (p4d_none_or_clear_bad(p4d)) 311 continue; 312 free_pud_range(tlb, p4d, addr, next, floor, ceiling); 313 } while (p4d++, addr = next, addr != end); 314 315 start &= PGDIR_MASK; 316 if (start < floor) 317 return; 318 if (ceiling) { 319 ceiling &= PGDIR_MASK; 320 if (!ceiling) 321 return; 322 } 323 if (end - 1 > ceiling - 1) 324 return; 325 326 p4d = p4d_offset(pgd, start); 327 pgd_clear(pgd); 328 p4d_free_tlb(tlb, p4d, start); 329 } 330 331 /* 332 * This function frees user-level page tables of a process. 333 */ 334 void free_pgd_range(struct mmu_gather *tlb, 335 unsigned long addr, unsigned long end, 336 unsigned long floor, unsigned long ceiling) 337 { 338 pgd_t *pgd; 339 unsigned long next; 340 341 /* 342 * The next few lines have given us lots of grief... 343 * 344 * Why are we testing PMD* at this top level? Because often 345 * there will be no work to do at all, and we'd prefer not to 346 * go all the way down to the bottom just to discover that. 347 * 348 * Why all these "- 1"s? Because 0 represents both the bottom 349 * of the address space and the top of it (using -1 for the 350 * top wouldn't help much: the masks would do the wrong thing). 351 * The rule is that addr 0 and floor 0 refer to the bottom of 352 * the address space, but end 0 and ceiling 0 refer to the top 353 * Comparisons need to use "end - 1" and "ceiling - 1" (though 354 * that end 0 case should be mythical). 355 * 356 * Wherever addr is brought up or ceiling brought down, we must 357 * be careful to reject "the opposite 0" before it confuses the 358 * subsequent tests. But what about where end is brought down 359 * by PMD_SIZE below? no, end can't go down to 0 there. 360 * 361 * Whereas we round start (addr) and ceiling down, by different 362 * masks at different levels, in order to test whether a table 363 * now has no other vmas using it, so can be freed, we don't 364 * bother to round floor or end up - the tests don't need that. 365 */ 366 367 addr &= PMD_MASK; 368 if (addr < floor) { 369 addr += PMD_SIZE; 370 if (!addr) 371 return; 372 } 373 if (ceiling) { 374 ceiling &= PMD_MASK; 375 if (!ceiling) 376 return; 377 } 378 if (end - 1 > ceiling - 1) 379 end -= PMD_SIZE; 380 if (addr > end - 1) 381 return; 382 /* 383 * We add page table cache pages with PAGE_SIZE, 384 * (see pte_free_tlb()), flush the tlb if we need 385 */ 386 tlb_change_page_size(tlb, PAGE_SIZE); 387 pgd = pgd_offset(tlb->mm, addr); 388 do { 389 next = pgd_addr_end(addr, end); 390 if (pgd_none_or_clear_bad(pgd)) 391 continue; 392 free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 393 } while (pgd++, addr = next, addr != end); 394 } 395 396 void free_pgtables(struct mmu_gather *tlb, struct maple_tree *mt, 397 struct vm_area_struct *vma, unsigned long floor, 398 unsigned long ceiling) 399 { 400 MA_STATE(mas, mt, vma->vm_end, vma->vm_end); 401 402 do { 403 unsigned long addr = vma->vm_start; 404 struct vm_area_struct *next; 405 406 /* 407 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may 408 * be 0. This will underflow and is okay. 409 */ 410 next = mas_find(&mas, ceiling - 1); 411 412 /* 413 * Hide vma from rmap and truncate_pagecache before freeing 414 * pgtables 415 */ 416 unlink_anon_vmas(vma); 417 unlink_file_vma(vma); 418 419 if (is_vm_hugetlb_page(vma)) { 420 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 421 floor, next ? next->vm_start : ceiling); 422 } else { 423 /* 424 * Optimization: gather nearby vmas into one call down 425 */ 426 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 427 && !is_vm_hugetlb_page(next)) { 428 vma = next; 429 next = mas_find(&mas, ceiling - 1); 430 unlink_anon_vmas(vma); 431 unlink_file_vma(vma); 432 } 433 free_pgd_range(tlb, addr, vma->vm_end, 434 floor, next ? next->vm_start : ceiling); 435 } 436 vma = next; 437 } while (vma); 438 } 439 440 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte) 441 { 442 spinlock_t *ptl = pmd_lock(mm, pmd); 443 444 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 445 mm_inc_nr_ptes(mm); 446 /* 447 * Ensure all pte setup (eg. pte page lock and page clearing) are 448 * visible before the pte is made visible to other CPUs by being 449 * put into page tables. 450 * 451 * The other side of the story is the pointer chasing in the page 452 * table walking code (when walking the page table without locking; 453 * ie. most of the time). Fortunately, these data accesses consist 454 * of a chain of data-dependent loads, meaning most CPUs (alpha 455 * being the notable exception) will already guarantee loads are 456 * seen in-order. See the alpha page table accessors for the 457 * smp_rmb() barriers in page table walking code. 458 */ 459 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 460 pmd_populate(mm, pmd, *pte); 461 *pte = NULL; 462 } 463 spin_unlock(ptl); 464 } 465 466 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 467 { 468 pgtable_t new = pte_alloc_one(mm); 469 if (!new) 470 return -ENOMEM; 471 472 pmd_install(mm, pmd, &new); 473 if (new) 474 pte_free(mm, new); 475 return 0; 476 } 477 478 int __pte_alloc_kernel(pmd_t *pmd) 479 { 480 pte_t *new = pte_alloc_one_kernel(&init_mm); 481 if (!new) 482 return -ENOMEM; 483 484 spin_lock(&init_mm.page_table_lock); 485 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 486 smp_wmb(); /* See comment in pmd_install() */ 487 pmd_populate_kernel(&init_mm, pmd, new); 488 new = NULL; 489 } 490 spin_unlock(&init_mm.page_table_lock); 491 if (new) 492 pte_free_kernel(&init_mm, new); 493 return 0; 494 } 495 496 static inline void init_rss_vec(int *rss) 497 { 498 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 499 } 500 501 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 502 { 503 int i; 504 505 if (current->mm == mm) 506 sync_mm_rss(mm); 507 for (i = 0; i < NR_MM_COUNTERS; i++) 508 if (rss[i]) 509 add_mm_counter(mm, i, rss[i]); 510 } 511 512 /* 513 * This function is called to print an error when a bad pte 514 * is found. For example, we might have a PFN-mapped pte in 515 * a region that doesn't allow it. 516 * 517 * The calling function must still handle the error. 518 */ 519 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 520 pte_t pte, struct page *page) 521 { 522 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 523 p4d_t *p4d = p4d_offset(pgd, addr); 524 pud_t *pud = pud_offset(p4d, addr); 525 pmd_t *pmd = pmd_offset(pud, addr); 526 struct address_space *mapping; 527 pgoff_t index; 528 static unsigned long resume; 529 static unsigned long nr_shown; 530 static unsigned long nr_unshown; 531 532 /* 533 * Allow a burst of 60 reports, then keep quiet for that minute; 534 * or allow a steady drip of one report per second. 535 */ 536 if (nr_shown == 60) { 537 if (time_before(jiffies, resume)) { 538 nr_unshown++; 539 return; 540 } 541 if (nr_unshown) { 542 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 543 nr_unshown); 544 nr_unshown = 0; 545 } 546 nr_shown = 0; 547 } 548 if (nr_shown++ == 0) 549 resume = jiffies + 60 * HZ; 550 551 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 552 index = linear_page_index(vma, addr); 553 554 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 555 current->comm, 556 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 557 if (page) 558 dump_page(page, "bad pte"); 559 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", 560 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 561 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n", 562 vma->vm_file, 563 vma->vm_ops ? vma->vm_ops->fault : NULL, 564 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 565 mapping ? mapping->a_ops->read_folio : NULL); 566 dump_stack(); 567 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 568 } 569 570 /* 571 * vm_normal_page -- This function gets the "struct page" associated with a pte. 572 * 573 * "Special" mappings do not wish to be associated with a "struct page" (either 574 * it doesn't exist, or it exists but they don't want to touch it). In this 575 * case, NULL is returned here. "Normal" mappings do have a struct page. 576 * 577 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 578 * pte bit, in which case this function is trivial. Secondly, an architecture 579 * may not have a spare pte bit, which requires a more complicated scheme, 580 * described below. 581 * 582 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 583 * special mapping (even if there are underlying and valid "struct pages"). 584 * COWed pages of a VM_PFNMAP are always normal. 585 * 586 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 587 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 588 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 589 * mapping will always honor the rule 590 * 591 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 592 * 593 * And for normal mappings this is false. 594 * 595 * This restricts such mappings to be a linear translation from virtual address 596 * to pfn. To get around this restriction, we allow arbitrary mappings so long 597 * as the vma is not a COW mapping; in that case, we know that all ptes are 598 * special (because none can have been COWed). 599 * 600 * 601 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 602 * 603 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 604 * page" backing, however the difference is that _all_ pages with a struct 605 * page (that is, those where pfn_valid is true) are refcounted and considered 606 * normal pages by the VM. The disadvantage is that pages are refcounted 607 * (which can be slower and simply not an option for some PFNMAP users). The 608 * advantage is that we don't have to follow the strict linearity rule of 609 * PFNMAP mappings in order to support COWable mappings. 610 * 611 */ 612 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 613 pte_t pte) 614 { 615 unsigned long pfn = pte_pfn(pte); 616 617 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 618 if (likely(!pte_special(pte))) 619 goto check_pfn; 620 if (vma->vm_ops && vma->vm_ops->find_special_page) 621 return vma->vm_ops->find_special_page(vma, addr); 622 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 623 return NULL; 624 if (is_zero_pfn(pfn)) 625 return NULL; 626 if (pte_devmap(pte)) 627 /* 628 * NOTE: New users of ZONE_DEVICE will not set pte_devmap() 629 * and will have refcounts incremented on their struct pages 630 * when they are inserted into PTEs, thus they are safe to 631 * return here. Legacy ZONE_DEVICE pages that set pte_devmap() 632 * do not have refcounts. Example of legacy ZONE_DEVICE is 633 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers. 634 */ 635 return NULL; 636 637 print_bad_pte(vma, addr, pte, NULL); 638 return NULL; 639 } 640 641 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 642 643 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 644 if (vma->vm_flags & VM_MIXEDMAP) { 645 if (!pfn_valid(pfn)) 646 return NULL; 647 goto out; 648 } else { 649 unsigned long off; 650 off = (addr - vma->vm_start) >> PAGE_SHIFT; 651 if (pfn == vma->vm_pgoff + off) 652 return NULL; 653 if (!is_cow_mapping(vma->vm_flags)) 654 return NULL; 655 } 656 } 657 658 if (is_zero_pfn(pfn)) 659 return NULL; 660 661 check_pfn: 662 if (unlikely(pfn > highest_memmap_pfn)) { 663 print_bad_pte(vma, addr, pte, NULL); 664 return NULL; 665 } 666 667 /* 668 * NOTE! We still have PageReserved() pages in the page tables. 669 * eg. VDSO mappings can cause them to exist. 670 */ 671 out: 672 return pfn_to_page(pfn); 673 } 674 675 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 676 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 677 pmd_t pmd) 678 { 679 unsigned long pfn = pmd_pfn(pmd); 680 681 /* 682 * There is no pmd_special() but there may be special pmds, e.g. 683 * in a direct-access (dax) mapping, so let's just replicate the 684 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 685 */ 686 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 687 if (vma->vm_flags & VM_MIXEDMAP) { 688 if (!pfn_valid(pfn)) 689 return NULL; 690 goto out; 691 } else { 692 unsigned long off; 693 off = (addr - vma->vm_start) >> PAGE_SHIFT; 694 if (pfn == vma->vm_pgoff + off) 695 return NULL; 696 if (!is_cow_mapping(vma->vm_flags)) 697 return NULL; 698 } 699 } 700 701 if (pmd_devmap(pmd)) 702 return NULL; 703 if (is_huge_zero_pmd(pmd)) 704 return NULL; 705 if (unlikely(pfn > highest_memmap_pfn)) 706 return NULL; 707 708 /* 709 * NOTE! We still have PageReserved() pages in the page tables. 710 * eg. VDSO mappings can cause them to exist. 711 */ 712 out: 713 return pfn_to_page(pfn); 714 } 715 #endif 716 717 static void restore_exclusive_pte(struct vm_area_struct *vma, 718 struct page *page, unsigned long address, 719 pte_t *ptep) 720 { 721 pte_t pte; 722 swp_entry_t entry; 723 724 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); 725 if (pte_swp_soft_dirty(*ptep)) 726 pte = pte_mksoft_dirty(pte); 727 728 entry = pte_to_swp_entry(*ptep); 729 if (pte_swp_uffd_wp(*ptep)) 730 pte = pte_mkuffd_wp(pte); 731 else if (is_writable_device_exclusive_entry(entry)) 732 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 733 734 VM_BUG_ON(pte_write(pte) && !(PageAnon(page) && PageAnonExclusive(page))); 735 736 /* 737 * No need to take a page reference as one was already 738 * created when the swap entry was made. 739 */ 740 if (PageAnon(page)) 741 page_add_anon_rmap(page, vma, address, RMAP_NONE); 742 else 743 /* 744 * Currently device exclusive access only supports anonymous 745 * memory so the entry shouldn't point to a filebacked page. 746 */ 747 WARN_ON_ONCE(1); 748 749 set_pte_at(vma->vm_mm, address, ptep, pte); 750 751 /* 752 * No need to invalidate - it was non-present before. However 753 * secondary CPUs may have mappings that need invalidating. 754 */ 755 update_mmu_cache(vma, address, ptep); 756 } 757 758 /* 759 * Tries to restore an exclusive pte if the page lock can be acquired without 760 * sleeping. 761 */ 762 static int 763 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma, 764 unsigned long addr) 765 { 766 swp_entry_t entry = pte_to_swp_entry(*src_pte); 767 struct page *page = pfn_swap_entry_to_page(entry); 768 769 if (trylock_page(page)) { 770 restore_exclusive_pte(vma, page, addr, src_pte); 771 unlock_page(page); 772 return 0; 773 } 774 775 return -EBUSY; 776 } 777 778 /* 779 * copy one vm_area from one task to the other. Assumes the page tables 780 * already present in the new task to be cleared in the whole range 781 * covered by this vma. 782 */ 783 784 static unsigned long 785 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 786 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma, 787 struct vm_area_struct *src_vma, unsigned long addr, int *rss) 788 { 789 unsigned long vm_flags = dst_vma->vm_flags; 790 pte_t pte = *src_pte; 791 struct page *page; 792 swp_entry_t entry = pte_to_swp_entry(pte); 793 794 if (likely(!non_swap_entry(entry))) { 795 if (swap_duplicate(entry) < 0) 796 return -EIO; 797 798 /* make sure dst_mm is on swapoff's mmlist. */ 799 if (unlikely(list_empty(&dst_mm->mmlist))) { 800 spin_lock(&mmlist_lock); 801 if (list_empty(&dst_mm->mmlist)) 802 list_add(&dst_mm->mmlist, 803 &src_mm->mmlist); 804 spin_unlock(&mmlist_lock); 805 } 806 /* Mark the swap entry as shared. */ 807 if (pte_swp_exclusive(*src_pte)) { 808 pte = pte_swp_clear_exclusive(*src_pte); 809 set_pte_at(src_mm, addr, src_pte, pte); 810 } 811 rss[MM_SWAPENTS]++; 812 } else if (is_migration_entry(entry)) { 813 page = pfn_swap_entry_to_page(entry); 814 815 rss[mm_counter(page)]++; 816 817 if (!is_readable_migration_entry(entry) && 818 is_cow_mapping(vm_flags)) { 819 /* 820 * COW mappings require pages in both parent and child 821 * to be set to read. A previously exclusive entry is 822 * now shared. 823 */ 824 entry = make_readable_migration_entry( 825 swp_offset(entry)); 826 pte = swp_entry_to_pte(entry); 827 if (pte_swp_soft_dirty(*src_pte)) 828 pte = pte_swp_mksoft_dirty(pte); 829 if (pte_swp_uffd_wp(*src_pte)) 830 pte = pte_swp_mkuffd_wp(pte); 831 set_pte_at(src_mm, addr, src_pte, pte); 832 } 833 } else if (is_device_private_entry(entry)) { 834 page = pfn_swap_entry_to_page(entry); 835 836 /* 837 * Update rss count even for unaddressable pages, as 838 * they should treated just like normal pages in this 839 * respect. 840 * 841 * We will likely want to have some new rss counters 842 * for unaddressable pages, at some point. But for now 843 * keep things as they are. 844 */ 845 get_page(page); 846 rss[mm_counter(page)]++; 847 /* Cannot fail as these pages cannot get pinned. */ 848 BUG_ON(page_try_dup_anon_rmap(page, false, src_vma)); 849 850 /* 851 * We do not preserve soft-dirty information, because so 852 * far, checkpoint/restore is the only feature that 853 * requires that. And checkpoint/restore does not work 854 * when a device driver is involved (you cannot easily 855 * save and restore device driver state). 856 */ 857 if (is_writable_device_private_entry(entry) && 858 is_cow_mapping(vm_flags)) { 859 entry = make_readable_device_private_entry( 860 swp_offset(entry)); 861 pte = swp_entry_to_pte(entry); 862 if (pte_swp_uffd_wp(*src_pte)) 863 pte = pte_swp_mkuffd_wp(pte); 864 set_pte_at(src_mm, addr, src_pte, pte); 865 } 866 } else if (is_device_exclusive_entry(entry)) { 867 /* 868 * Make device exclusive entries present by restoring the 869 * original entry then copying as for a present pte. Device 870 * exclusive entries currently only support private writable 871 * (ie. COW) mappings. 872 */ 873 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags)); 874 if (try_restore_exclusive_pte(src_pte, src_vma, addr)) 875 return -EBUSY; 876 return -ENOENT; 877 } else if (is_pte_marker_entry(entry)) { 878 /* 879 * We're copying the pgtable should only because dst_vma has 880 * uffd-wp enabled, do sanity check. 881 */ 882 WARN_ON_ONCE(!userfaultfd_wp(dst_vma)); 883 set_pte_at(dst_mm, addr, dst_pte, pte); 884 return 0; 885 } 886 if (!userfaultfd_wp(dst_vma)) 887 pte = pte_swp_clear_uffd_wp(pte); 888 set_pte_at(dst_mm, addr, dst_pte, pte); 889 return 0; 890 } 891 892 /* 893 * Copy a present and normal page. 894 * 895 * NOTE! The usual case is that this isn't required; 896 * instead, the caller can just increase the page refcount 897 * and re-use the pte the traditional way. 898 * 899 * And if we need a pre-allocated page but don't yet have 900 * one, return a negative error to let the preallocation 901 * code know so that it can do so outside the page table 902 * lock. 903 */ 904 static inline int 905 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 906 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 907 struct page **prealloc, struct page *page) 908 { 909 struct page *new_page; 910 pte_t pte; 911 912 new_page = *prealloc; 913 if (!new_page) 914 return -EAGAIN; 915 916 /* 917 * We have a prealloc page, all good! Take it 918 * over and copy the page & arm it. 919 */ 920 *prealloc = NULL; 921 copy_user_highpage(new_page, page, addr, src_vma); 922 __SetPageUptodate(new_page); 923 page_add_new_anon_rmap(new_page, dst_vma, addr); 924 lru_cache_add_inactive_or_unevictable(new_page, dst_vma); 925 rss[mm_counter(new_page)]++; 926 927 /* All done, just insert the new page copy in the child */ 928 pte = mk_pte(new_page, dst_vma->vm_page_prot); 929 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); 930 if (userfaultfd_pte_wp(dst_vma, *src_pte)) 931 /* Uffd-wp needs to be delivered to dest pte as well */ 932 pte = pte_wrprotect(pte_mkuffd_wp(pte)); 933 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 934 return 0; 935 } 936 937 /* 938 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page 939 * is required to copy this pte. 940 */ 941 static inline int 942 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 943 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 944 struct page **prealloc) 945 { 946 struct mm_struct *src_mm = src_vma->vm_mm; 947 unsigned long vm_flags = src_vma->vm_flags; 948 pte_t pte = *src_pte; 949 struct page *page; 950 951 page = vm_normal_page(src_vma, addr, pte); 952 if (page && PageAnon(page)) { 953 /* 954 * If this page may have been pinned by the parent process, 955 * copy the page immediately for the child so that we'll always 956 * guarantee the pinned page won't be randomly replaced in the 957 * future. 958 */ 959 get_page(page); 960 if (unlikely(page_try_dup_anon_rmap(page, false, src_vma))) { 961 /* Page maybe pinned, we have to copy. */ 962 put_page(page); 963 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte, 964 addr, rss, prealloc, page); 965 } 966 rss[mm_counter(page)]++; 967 } else if (page) { 968 get_page(page); 969 page_dup_file_rmap(page, false); 970 rss[mm_counter(page)]++; 971 } 972 973 /* 974 * If it's a COW mapping, write protect it both 975 * in the parent and the child 976 */ 977 if (is_cow_mapping(vm_flags) && pte_write(pte)) { 978 ptep_set_wrprotect(src_mm, addr, src_pte); 979 pte = pte_wrprotect(pte); 980 } 981 VM_BUG_ON(page && PageAnon(page) && PageAnonExclusive(page)); 982 983 /* 984 * If it's a shared mapping, mark it clean in 985 * the child 986 */ 987 if (vm_flags & VM_SHARED) 988 pte = pte_mkclean(pte); 989 pte = pte_mkold(pte); 990 991 if (!userfaultfd_wp(dst_vma)) 992 pte = pte_clear_uffd_wp(pte); 993 994 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 995 return 0; 996 } 997 998 static inline struct page * 999 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma, 1000 unsigned long addr) 1001 { 1002 struct page *new_page; 1003 1004 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr); 1005 if (!new_page) 1006 return NULL; 1007 1008 if (mem_cgroup_charge(page_folio(new_page), src_mm, GFP_KERNEL)) { 1009 put_page(new_page); 1010 return NULL; 1011 } 1012 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 1013 1014 return new_page; 1015 } 1016 1017 static int 1018 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1019 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1020 unsigned long end) 1021 { 1022 struct mm_struct *dst_mm = dst_vma->vm_mm; 1023 struct mm_struct *src_mm = src_vma->vm_mm; 1024 pte_t *orig_src_pte, *orig_dst_pte; 1025 pte_t *src_pte, *dst_pte; 1026 spinlock_t *src_ptl, *dst_ptl; 1027 int progress, ret = 0; 1028 int rss[NR_MM_COUNTERS]; 1029 swp_entry_t entry = (swp_entry_t){0}; 1030 struct page *prealloc = NULL; 1031 1032 again: 1033 progress = 0; 1034 init_rss_vec(rss); 1035 1036 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 1037 if (!dst_pte) { 1038 ret = -ENOMEM; 1039 goto out; 1040 } 1041 src_pte = pte_offset_map(src_pmd, addr); 1042 src_ptl = pte_lockptr(src_mm, src_pmd); 1043 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1044 orig_src_pte = src_pte; 1045 orig_dst_pte = dst_pte; 1046 arch_enter_lazy_mmu_mode(); 1047 1048 do { 1049 /* 1050 * We are holding two locks at this point - either of them 1051 * could generate latencies in another task on another CPU. 1052 */ 1053 if (progress >= 32) { 1054 progress = 0; 1055 if (need_resched() || 1056 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 1057 break; 1058 } 1059 if (pte_none(*src_pte)) { 1060 progress++; 1061 continue; 1062 } 1063 if (unlikely(!pte_present(*src_pte))) { 1064 ret = copy_nonpresent_pte(dst_mm, src_mm, 1065 dst_pte, src_pte, 1066 dst_vma, src_vma, 1067 addr, rss); 1068 if (ret == -EIO) { 1069 entry = pte_to_swp_entry(*src_pte); 1070 break; 1071 } else if (ret == -EBUSY) { 1072 break; 1073 } else if (!ret) { 1074 progress += 8; 1075 continue; 1076 } 1077 1078 /* 1079 * Device exclusive entry restored, continue by copying 1080 * the now present pte. 1081 */ 1082 WARN_ON_ONCE(ret != -ENOENT); 1083 } 1084 /* copy_present_pte() will clear `*prealloc' if consumed */ 1085 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte, 1086 addr, rss, &prealloc); 1087 /* 1088 * If we need a pre-allocated page for this pte, drop the 1089 * locks, allocate, and try again. 1090 */ 1091 if (unlikely(ret == -EAGAIN)) 1092 break; 1093 if (unlikely(prealloc)) { 1094 /* 1095 * pre-alloc page cannot be reused by next time so as 1096 * to strictly follow mempolicy (e.g., alloc_page_vma() 1097 * will allocate page according to address). This 1098 * could only happen if one pinned pte changed. 1099 */ 1100 put_page(prealloc); 1101 prealloc = NULL; 1102 } 1103 progress += 8; 1104 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 1105 1106 arch_leave_lazy_mmu_mode(); 1107 spin_unlock(src_ptl); 1108 pte_unmap(orig_src_pte); 1109 add_mm_rss_vec(dst_mm, rss); 1110 pte_unmap_unlock(orig_dst_pte, dst_ptl); 1111 cond_resched(); 1112 1113 if (ret == -EIO) { 1114 VM_WARN_ON_ONCE(!entry.val); 1115 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { 1116 ret = -ENOMEM; 1117 goto out; 1118 } 1119 entry.val = 0; 1120 } else if (ret == -EBUSY) { 1121 goto out; 1122 } else if (ret == -EAGAIN) { 1123 prealloc = page_copy_prealloc(src_mm, src_vma, addr); 1124 if (!prealloc) 1125 return -ENOMEM; 1126 } else if (ret) { 1127 VM_WARN_ON_ONCE(1); 1128 } 1129 1130 /* We've captured and resolved the error. Reset, try again. */ 1131 ret = 0; 1132 1133 if (addr != end) 1134 goto again; 1135 out: 1136 if (unlikely(prealloc)) 1137 put_page(prealloc); 1138 return ret; 1139 } 1140 1141 static inline int 1142 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1143 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1144 unsigned long end) 1145 { 1146 struct mm_struct *dst_mm = dst_vma->vm_mm; 1147 struct mm_struct *src_mm = src_vma->vm_mm; 1148 pmd_t *src_pmd, *dst_pmd; 1149 unsigned long next; 1150 1151 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 1152 if (!dst_pmd) 1153 return -ENOMEM; 1154 src_pmd = pmd_offset(src_pud, addr); 1155 do { 1156 next = pmd_addr_end(addr, end); 1157 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 1158 || pmd_devmap(*src_pmd)) { 1159 int err; 1160 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); 1161 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd, 1162 addr, dst_vma, src_vma); 1163 if (err == -ENOMEM) 1164 return -ENOMEM; 1165 if (!err) 1166 continue; 1167 /* fall through */ 1168 } 1169 if (pmd_none_or_clear_bad(src_pmd)) 1170 continue; 1171 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, 1172 addr, next)) 1173 return -ENOMEM; 1174 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 1175 return 0; 1176 } 1177 1178 static inline int 1179 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1180 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, 1181 unsigned long end) 1182 { 1183 struct mm_struct *dst_mm = dst_vma->vm_mm; 1184 struct mm_struct *src_mm = src_vma->vm_mm; 1185 pud_t *src_pud, *dst_pud; 1186 unsigned long next; 1187 1188 dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 1189 if (!dst_pud) 1190 return -ENOMEM; 1191 src_pud = pud_offset(src_p4d, addr); 1192 do { 1193 next = pud_addr_end(addr, end); 1194 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 1195 int err; 1196 1197 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); 1198 err = copy_huge_pud(dst_mm, src_mm, 1199 dst_pud, src_pud, addr, src_vma); 1200 if (err == -ENOMEM) 1201 return -ENOMEM; 1202 if (!err) 1203 continue; 1204 /* fall through */ 1205 } 1206 if (pud_none_or_clear_bad(src_pud)) 1207 continue; 1208 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, 1209 addr, next)) 1210 return -ENOMEM; 1211 } while (dst_pud++, src_pud++, addr = next, addr != end); 1212 return 0; 1213 } 1214 1215 static inline int 1216 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1217 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, 1218 unsigned long end) 1219 { 1220 struct mm_struct *dst_mm = dst_vma->vm_mm; 1221 p4d_t *src_p4d, *dst_p4d; 1222 unsigned long next; 1223 1224 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 1225 if (!dst_p4d) 1226 return -ENOMEM; 1227 src_p4d = p4d_offset(src_pgd, addr); 1228 do { 1229 next = p4d_addr_end(addr, end); 1230 if (p4d_none_or_clear_bad(src_p4d)) 1231 continue; 1232 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, 1233 addr, next)) 1234 return -ENOMEM; 1235 } while (dst_p4d++, src_p4d++, addr = next, addr != end); 1236 return 0; 1237 } 1238 1239 /* 1240 * Return true if the vma needs to copy the pgtable during this fork(). Return 1241 * false when we can speed up fork() by allowing lazy page faults later until 1242 * when the child accesses the memory range. 1243 */ 1244 static bool 1245 vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1246 { 1247 /* 1248 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's 1249 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable 1250 * contains uffd-wp protection information, that's something we can't 1251 * retrieve from page cache, and skip copying will lose those info. 1252 */ 1253 if (userfaultfd_wp(dst_vma)) 1254 return true; 1255 1256 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 1257 return true; 1258 1259 if (src_vma->anon_vma) 1260 return true; 1261 1262 /* 1263 * Don't copy ptes where a page fault will fill them correctly. Fork 1264 * becomes much lighter when there are big shared or private readonly 1265 * mappings. The tradeoff is that copy_page_range is more efficient 1266 * than faulting. 1267 */ 1268 return false; 1269 } 1270 1271 int 1272 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1273 { 1274 pgd_t *src_pgd, *dst_pgd; 1275 unsigned long next; 1276 unsigned long addr = src_vma->vm_start; 1277 unsigned long end = src_vma->vm_end; 1278 struct mm_struct *dst_mm = dst_vma->vm_mm; 1279 struct mm_struct *src_mm = src_vma->vm_mm; 1280 struct mmu_notifier_range range; 1281 bool is_cow; 1282 int ret; 1283 1284 if (!vma_needs_copy(dst_vma, src_vma)) 1285 return 0; 1286 1287 if (is_vm_hugetlb_page(src_vma)) 1288 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma); 1289 1290 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { 1291 /* 1292 * We do not free on error cases below as remove_vma 1293 * gets called on error from higher level routine 1294 */ 1295 ret = track_pfn_copy(src_vma); 1296 if (ret) 1297 return ret; 1298 } 1299 1300 /* 1301 * We need to invalidate the secondary MMU mappings only when 1302 * there could be a permission downgrade on the ptes of the 1303 * parent mm. And a permission downgrade will only happen if 1304 * is_cow_mapping() returns true. 1305 */ 1306 is_cow = is_cow_mapping(src_vma->vm_flags); 1307 1308 if (is_cow) { 1309 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 1310 0, src_vma, src_mm, addr, end); 1311 mmu_notifier_invalidate_range_start(&range); 1312 /* 1313 * Disabling preemption is not needed for the write side, as 1314 * the read side doesn't spin, but goes to the mmap_lock. 1315 * 1316 * Use the raw variant of the seqcount_t write API to avoid 1317 * lockdep complaining about preemptibility. 1318 */ 1319 mmap_assert_write_locked(src_mm); 1320 raw_write_seqcount_begin(&src_mm->write_protect_seq); 1321 } 1322 1323 ret = 0; 1324 dst_pgd = pgd_offset(dst_mm, addr); 1325 src_pgd = pgd_offset(src_mm, addr); 1326 do { 1327 next = pgd_addr_end(addr, end); 1328 if (pgd_none_or_clear_bad(src_pgd)) 1329 continue; 1330 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, 1331 addr, next))) { 1332 ret = -ENOMEM; 1333 break; 1334 } 1335 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1336 1337 if (is_cow) { 1338 raw_write_seqcount_end(&src_mm->write_protect_seq); 1339 mmu_notifier_invalidate_range_end(&range); 1340 } 1341 return ret; 1342 } 1343 1344 /* Whether we should zap all COWed (private) pages too */ 1345 static inline bool should_zap_cows(struct zap_details *details) 1346 { 1347 /* By default, zap all pages */ 1348 if (!details) 1349 return true; 1350 1351 /* Or, we zap COWed pages only if the caller wants to */ 1352 return details->even_cows; 1353 } 1354 1355 /* Decides whether we should zap this page with the page pointer specified */ 1356 static inline bool should_zap_page(struct zap_details *details, struct page *page) 1357 { 1358 /* If we can make a decision without *page.. */ 1359 if (should_zap_cows(details)) 1360 return true; 1361 1362 /* E.g. the caller passes NULL for the case of a zero page */ 1363 if (!page) 1364 return true; 1365 1366 /* Otherwise we should only zap non-anon pages */ 1367 return !PageAnon(page); 1368 } 1369 1370 static inline bool zap_drop_file_uffd_wp(struct zap_details *details) 1371 { 1372 if (!details) 1373 return false; 1374 1375 return details->zap_flags & ZAP_FLAG_DROP_MARKER; 1376 } 1377 1378 /* 1379 * This function makes sure that we'll replace the none pte with an uffd-wp 1380 * swap special pte marker when necessary. Must be with the pgtable lock held. 1381 */ 1382 static inline void 1383 zap_install_uffd_wp_if_needed(struct vm_area_struct *vma, 1384 unsigned long addr, pte_t *pte, 1385 struct zap_details *details, pte_t pteval) 1386 { 1387 #ifdef CONFIG_PTE_MARKER_UFFD_WP 1388 if (zap_drop_file_uffd_wp(details)) 1389 return; 1390 1391 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval); 1392 #endif 1393 } 1394 1395 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1396 struct vm_area_struct *vma, pmd_t *pmd, 1397 unsigned long addr, unsigned long end, 1398 struct zap_details *details) 1399 { 1400 struct mm_struct *mm = tlb->mm; 1401 int force_flush = 0; 1402 int rss[NR_MM_COUNTERS]; 1403 spinlock_t *ptl; 1404 pte_t *start_pte; 1405 pte_t *pte; 1406 swp_entry_t entry; 1407 1408 tlb_change_page_size(tlb, PAGE_SIZE); 1409 again: 1410 init_rss_vec(rss); 1411 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1412 pte = start_pte; 1413 flush_tlb_batched_pending(mm); 1414 arch_enter_lazy_mmu_mode(); 1415 do { 1416 pte_t ptent = *pte; 1417 struct page *page; 1418 1419 if (pte_none(ptent)) 1420 continue; 1421 1422 if (need_resched()) 1423 break; 1424 1425 if (pte_present(ptent)) { 1426 page = vm_normal_page(vma, addr, ptent); 1427 if (unlikely(!should_zap_page(details, page))) 1428 continue; 1429 ptent = ptep_get_and_clear_full(mm, addr, pte, 1430 tlb->fullmm); 1431 tlb_remove_tlb_entry(tlb, pte, addr); 1432 zap_install_uffd_wp_if_needed(vma, addr, pte, details, 1433 ptent); 1434 if (unlikely(!page)) 1435 continue; 1436 1437 if (!PageAnon(page)) { 1438 if (pte_dirty(ptent)) { 1439 force_flush = 1; 1440 set_page_dirty(page); 1441 } 1442 if (pte_young(ptent) && 1443 likely(!(vma->vm_flags & VM_SEQ_READ))) 1444 mark_page_accessed(page); 1445 } 1446 rss[mm_counter(page)]--; 1447 page_remove_rmap(page, vma, false); 1448 if (unlikely(page_mapcount(page) < 0)) 1449 print_bad_pte(vma, addr, ptent, page); 1450 if (unlikely(__tlb_remove_page(tlb, page))) { 1451 force_flush = 1; 1452 addr += PAGE_SIZE; 1453 break; 1454 } 1455 continue; 1456 } 1457 1458 entry = pte_to_swp_entry(ptent); 1459 if (is_device_private_entry(entry) || 1460 is_device_exclusive_entry(entry)) { 1461 page = pfn_swap_entry_to_page(entry); 1462 if (unlikely(!should_zap_page(details, page))) 1463 continue; 1464 /* 1465 * Both device private/exclusive mappings should only 1466 * work with anonymous page so far, so we don't need to 1467 * consider uffd-wp bit when zap. For more information, 1468 * see zap_install_uffd_wp_if_needed(). 1469 */ 1470 WARN_ON_ONCE(!vma_is_anonymous(vma)); 1471 rss[mm_counter(page)]--; 1472 if (is_device_private_entry(entry)) 1473 page_remove_rmap(page, vma, false); 1474 put_page(page); 1475 } else if (!non_swap_entry(entry)) { 1476 /* Genuine swap entry, hence a private anon page */ 1477 if (!should_zap_cows(details)) 1478 continue; 1479 rss[MM_SWAPENTS]--; 1480 if (unlikely(!free_swap_and_cache(entry))) 1481 print_bad_pte(vma, addr, ptent, NULL); 1482 } else if (is_migration_entry(entry)) { 1483 page = pfn_swap_entry_to_page(entry); 1484 if (!should_zap_page(details, page)) 1485 continue; 1486 rss[mm_counter(page)]--; 1487 } else if (pte_marker_entry_uffd_wp(entry)) { 1488 /* Only drop the uffd-wp marker if explicitly requested */ 1489 if (!zap_drop_file_uffd_wp(details)) 1490 continue; 1491 } else if (is_hwpoison_entry(entry) || 1492 is_swapin_error_entry(entry)) { 1493 if (!should_zap_cows(details)) 1494 continue; 1495 } else { 1496 /* We should have covered all the swap entry types */ 1497 WARN_ON_ONCE(1); 1498 } 1499 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1500 zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent); 1501 } while (pte++, addr += PAGE_SIZE, addr != end); 1502 1503 add_mm_rss_vec(mm, rss); 1504 arch_leave_lazy_mmu_mode(); 1505 1506 /* Do the actual TLB flush before dropping ptl */ 1507 if (force_flush) 1508 tlb_flush_mmu_tlbonly(tlb); 1509 pte_unmap_unlock(start_pte, ptl); 1510 1511 /* 1512 * If we forced a TLB flush (either due to running out of 1513 * batch buffers or because we needed to flush dirty TLB 1514 * entries before releasing the ptl), free the batched 1515 * memory too. Restart if we didn't do everything. 1516 */ 1517 if (force_flush) { 1518 force_flush = 0; 1519 tlb_flush_mmu(tlb); 1520 } 1521 1522 if (addr != end) { 1523 cond_resched(); 1524 goto again; 1525 } 1526 1527 return addr; 1528 } 1529 1530 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1531 struct vm_area_struct *vma, pud_t *pud, 1532 unsigned long addr, unsigned long end, 1533 struct zap_details *details) 1534 { 1535 pmd_t *pmd; 1536 unsigned long next; 1537 1538 pmd = pmd_offset(pud, addr); 1539 do { 1540 next = pmd_addr_end(addr, end); 1541 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1542 if (next - addr != HPAGE_PMD_SIZE) 1543 __split_huge_pmd(vma, pmd, addr, false, NULL); 1544 else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1545 goto next; 1546 /* fall through */ 1547 } else if (details && details->single_folio && 1548 folio_test_pmd_mappable(details->single_folio) && 1549 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) { 1550 spinlock_t *ptl = pmd_lock(tlb->mm, pmd); 1551 /* 1552 * Take and drop THP pmd lock so that we cannot return 1553 * prematurely, while zap_huge_pmd() has cleared *pmd, 1554 * but not yet decremented compound_mapcount(). 1555 */ 1556 spin_unlock(ptl); 1557 } 1558 1559 /* 1560 * Here there can be other concurrent MADV_DONTNEED or 1561 * trans huge page faults running, and if the pmd is 1562 * none or trans huge it can change under us. This is 1563 * because MADV_DONTNEED holds the mmap_lock in read 1564 * mode. 1565 */ 1566 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1567 goto next; 1568 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1569 next: 1570 cond_resched(); 1571 } while (pmd++, addr = next, addr != end); 1572 1573 return addr; 1574 } 1575 1576 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1577 struct vm_area_struct *vma, p4d_t *p4d, 1578 unsigned long addr, unsigned long end, 1579 struct zap_details *details) 1580 { 1581 pud_t *pud; 1582 unsigned long next; 1583 1584 pud = pud_offset(p4d, addr); 1585 do { 1586 next = pud_addr_end(addr, end); 1587 if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1588 if (next - addr != HPAGE_PUD_SIZE) { 1589 mmap_assert_locked(tlb->mm); 1590 split_huge_pud(vma, pud, addr); 1591 } else if (zap_huge_pud(tlb, vma, pud, addr)) 1592 goto next; 1593 /* fall through */ 1594 } 1595 if (pud_none_or_clear_bad(pud)) 1596 continue; 1597 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1598 next: 1599 cond_resched(); 1600 } while (pud++, addr = next, addr != end); 1601 1602 return addr; 1603 } 1604 1605 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1606 struct vm_area_struct *vma, pgd_t *pgd, 1607 unsigned long addr, unsigned long end, 1608 struct zap_details *details) 1609 { 1610 p4d_t *p4d; 1611 unsigned long next; 1612 1613 p4d = p4d_offset(pgd, addr); 1614 do { 1615 next = p4d_addr_end(addr, end); 1616 if (p4d_none_or_clear_bad(p4d)) 1617 continue; 1618 next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1619 } while (p4d++, addr = next, addr != end); 1620 1621 return addr; 1622 } 1623 1624 void unmap_page_range(struct mmu_gather *tlb, 1625 struct vm_area_struct *vma, 1626 unsigned long addr, unsigned long end, 1627 struct zap_details *details) 1628 { 1629 pgd_t *pgd; 1630 unsigned long next; 1631 1632 BUG_ON(addr >= end); 1633 tlb_start_vma(tlb, vma); 1634 pgd = pgd_offset(vma->vm_mm, addr); 1635 do { 1636 next = pgd_addr_end(addr, end); 1637 if (pgd_none_or_clear_bad(pgd)) 1638 continue; 1639 next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 1640 } while (pgd++, addr = next, addr != end); 1641 tlb_end_vma(tlb, vma); 1642 } 1643 1644 1645 static void unmap_single_vma(struct mmu_gather *tlb, 1646 struct vm_area_struct *vma, unsigned long start_addr, 1647 unsigned long end_addr, 1648 struct zap_details *details) 1649 { 1650 unsigned long start = max(vma->vm_start, start_addr); 1651 unsigned long end; 1652 1653 if (start >= vma->vm_end) 1654 return; 1655 end = min(vma->vm_end, end_addr); 1656 if (end <= vma->vm_start) 1657 return; 1658 1659 if (vma->vm_file) 1660 uprobe_munmap(vma, start, end); 1661 1662 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1663 untrack_pfn(vma, 0, 0); 1664 1665 if (start != end) { 1666 if (unlikely(is_vm_hugetlb_page(vma))) { 1667 /* 1668 * It is undesirable to test vma->vm_file as it 1669 * should be non-null for valid hugetlb area. 1670 * However, vm_file will be NULL in the error 1671 * cleanup path of mmap_region. When 1672 * hugetlbfs ->mmap method fails, 1673 * mmap_region() nullifies vma->vm_file 1674 * before calling this function to clean up. 1675 * Since no pte has actually been setup, it is 1676 * safe to do nothing in this case. 1677 */ 1678 if (vma->vm_file) { 1679 zap_flags_t zap_flags = details ? 1680 details->zap_flags : 0; 1681 __unmap_hugepage_range_final(tlb, vma, start, end, 1682 NULL, zap_flags); 1683 } 1684 } else 1685 unmap_page_range(tlb, vma, start, end, details); 1686 } 1687 } 1688 1689 /** 1690 * unmap_vmas - unmap a range of memory covered by a list of vma's 1691 * @tlb: address of the caller's struct mmu_gather 1692 * @mt: the maple tree 1693 * @vma: the starting vma 1694 * @start_addr: virtual address at which to start unmapping 1695 * @end_addr: virtual address at which to end unmapping 1696 * 1697 * Unmap all pages in the vma list. 1698 * 1699 * Only addresses between `start' and `end' will be unmapped. 1700 * 1701 * The VMA list must be sorted in ascending virtual address order. 1702 * 1703 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1704 * range after unmap_vmas() returns. So the only responsibility here is to 1705 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1706 * drops the lock and schedules. 1707 */ 1708 void unmap_vmas(struct mmu_gather *tlb, struct maple_tree *mt, 1709 struct vm_area_struct *vma, unsigned long start_addr, 1710 unsigned long end_addr) 1711 { 1712 struct mmu_notifier_range range; 1713 struct zap_details details = { 1714 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP, 1715 /* Careful - we need to zap private pages too! */ 1716 .even_cows = true, 1717 }; 1718 MA_STATE(mas, mt, vma->vm_end, vma->vm_end); 1719 1720 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, 1721 start_addr, end_addr); 1722 mmu_notifier_invalidate_range_start(&range); 1723 do { 1724 unmap_single_vma(tlb, vma, start_addr, end_addr, &details); 1725 } while ((vma = mas_find(&mas, end_addr - 1)) != NULL); 1726 mmu_notifier_invalidate_range_end(&range); 1727 } 1728 1729 /** 1730 * zap_page_range - remove user pages in a given range 1731 * @vma: vm_area_struct holding the applicable pages 1732 * @start: starting address of pages to zap 1733 * @size: number of bytes to zap 1734 * 1735 * Caller must protect the VMA list 1736 */ 1737 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1738 unsigned long size) 1739 { 1740 struct maple_tree *mt = &vma->vm_mm->mm_mt; 1741 unsigned long end = start + size; 1742 struct mmu_notifier_range range; 1743 struct mmu_gather tlb; 1744 MA_STATE(mas, mt, vma->vm_end, vma->vm_end); 1745 1746 lru_add_drain(); 1747 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1748 start, start + size); 1749 tlb_gather_mmu(&tlb, vma->vm_mm); 1750 update_hiwater_rss(vma->vm_mm); 1751 mmu_notifier_invalidate_range_start(&range); 1752 do { 1753 unmap_single_vma(&tlb, vma, start, range.end, NULL); 1754 } while ((vma = mas_find(&mas, end - 1)) != NULL); 1755 mmu_notifier_invalidate_range_end(&range); 1756 tlb_finish_mmu(&tlb); 1757 } 1758 1759 /** 1760 * zap_page_range_single - remove user pages in a given range 1761 * @vma: vm_area_struct holding the applicable pages 1762 * @address: starting address of pages to zap 1763 * @size: number of bytes to zap 1764 * @details: details of shared cache invalidation 1765 * 1766 * The range must fit into one VMA. 1767 */ 1768 void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1769 unsigned long size, struct zap_details *details) 1770 { 1771 const unsigned long end = address + size; 1772 struct mmu_notifier_range range; 1773 struct mmu_gather tlb; 1774 1775 lru_add_drain(); 1776 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1777 address, end); 1778 if (is_vm_hugetlb_page(vma)) 1779 adjust_range_if_pmd_sharing_possible(vma, &range.start, 1780 &range.end); 1781 tlb_gather_mmu(&tlb, vma->vm_mm); 1782 update_hiwater_rss(vma->vm_mm); 1783 mmu_notifier_invalidate_range_start(&range); 1784 /* 1785 * unmap 'address-end' not 'range.start-range.end' as range 1786 * could have been expanded for hugetlb pmd sharing. 1787 */ 1788 unmap_single_vma(&tlb, vma, address, end, details); 1789 mmu_notifier_invalidate_range_end(&range); 1790 tlb_finish_mmu(&tlb); 1791 } 1792 1793 /** 1794 * zap_vma_ptes - remove ptes mapping the vma 1795 * @vma: vm_area_struct holding ptes to be zapped 1796 * @address: starting address of pages to zap 1797 * @size: number of bytes to zap 1798 * 1799 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1800 * 1801 * The entire address range must be fully contained within the vma. 1802 * 1803 */ 1804 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1805 unsigned long size) 1806 { 1807 if (!range_in_vma(vma, address, address + size) || 1808 !(vma->vm_flags & VM_PFNMAP)) 1809 return; 1810 1811 zap_page_range_single(vma, address, size, NULL); 1812 } 1813 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1814 1815 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) 1816 { 1817 pgd_t *pgd; 1818 p4d_t *p4d; 1819 pud_t *pud; 1820 pmd_t *pmd; 1821 1822 pgd = pgd_offset(mm, addr); 1823 p4d = p4d_alloc(mm, pgd, addr); 1824 if (!p4d) 1825 return NULL; 1826 pud = pud_alloc(mm, p4d, addr); 1827 if (!pud) 1828 return NULL; 1829 pmd = pmd_alloc(mm, pud, addr); 1830 if (!pmd) 1831 return NULL; 1832 1833 VM_BUG_ON(pmd_trans_huge(*pmd)); 1834 return pmd; 1835 } 1836 1837 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1838 spinlock_t **ptl) 1839 { 1840 pmd_t *pmd = walk_to_pmd(mm, addr); 1841 1842 if (!pmd) 1843 return NULL; 1844 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1845 } 1846 1847 static int validate_page_before_insert(struct page *page) 1848 { 1849 if (PageAnon(page) || PageSlab(page) || page_has_type(page)) 1850 return -EINVAL; 1851 flush_dcache_page(page); 1852 return 0; 1853 } 1854 1855 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte, 1856 unsigned long addr, struct page *page, pgprot_t prot) 1857 { 1858 if (!pte_none(*pte)) 1859 return -EBUSY; 1860 /* Ok, finally just insert the thing.. */ 1861 get_page(page); 1862 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 1863 page_add_file_rmap(page, vma, false); 1864 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot)); 1865 return 0; 1866 } 1867 1868 /* 1869 * This is the old fallback for page remapping. 1870 * 1871 * For historical reasons, it only allows reserved pages. Only 1872 * old drivers should use this, and they needed to mark their 1873 * pages reserved for the old functions anyway. 1874 */ 1875 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1876 struct page *page, pgprot_t prot) 1877 { 1878 int retval; 1879 pte_t *pte; 1880 spinlock_t *ptl; 1881 1882 retval = validate_page_before_insert(page); 1883 if (retval) 1884 goto out; 1885 retval = -ENOMEM; 1886 pte = get_locked_pte(vma->vm_mm, addr, &ptl); 1887 if (!pte) 1888 goto out; 1889 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot); 1890 pte_unmap_unlock(pte, ptl); 1891 out: 1892 return retval; 1893 } 1894 1895 #ifdef pte_index 1896 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte, 1897 unsigned long addr, struct page *page, pgprot_t prot) 1898 { 1899 int err; 1900 1901 if (!page_count(page)) 1902 return -EINVAL; 1903 err = validate_page_before_insert(page); 1904 if (err) 1905 return err; 1906 return insert_page_into_pte_locked(vma, pte, addr, page, prot); 1907 } 1908 1909 /* insert_pages() amortizes the cost of spinlock operations 1910 * when inserting pages in a loop. Arch *must* define pte_index. 1911 */ 1912 static int insert_pages(struct vm_area_struct *vma, unsigned long addr, 1913 struct page **pages, unsigned long *num, pgprot_t prot) 1914 { 1915 pmd_t *pmd = NULL; 1916 pte_t *start_pte, *pte; 1917 spinlock_t *pte_lock; 1918 struct mm_struct *const mm = vma->vm_mm; 1919 unsigned long curr_page_idx = 0; 1920 unsigned long remaining_pages_total = *num; 1921 unsigned long pages_to_write_in_pmd; 1922 int ret; 1923 more: 1924 ret = -EFAULT; 1925 pmd = walk_to_pmd(mm, addr); 1926 if (!pmd) 1927 goto out; 1928 1929 pages_to_write_in_pmd = min_t(unsigned long, 1930 remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); 1931 1932 /* Allocate the PTE if necessary; takes PMD lock once only. */ 1933 ret = -ENOMEM; 1934 if (pte_alloc(mm, pmd)) 1935 goto out; 1936 1937 while (pages_to_write_in_pmd) { 1938 int pte_idx = 0; 1939 const int batch_size = min_t(int, pages_to_write_in_pmd, 8); 1940 1941 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); 1942 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { 1943 int err = insert_page_in_batch_locked(vma, pte, 1944 addr, pages[curr_page_idx], prot); 1945 if (unlikely(err)) { 1946 pte_unmap_unlock(start_pte, pte_lock); 1947 ret = err; 1948 remaining_pages_total -= pte_idx; 1949 goto out; 1950 } 1951 addr += PAGE_SIZE; 1952 ++curr_page_idx; 1953 } 1954 pte_unmap_unlock(start_pte, pte_lock); 1955 pages_to_write_in_pmd -= batch_size; 1956 remaining_pages_total -= batch_size; 1957 } 1958 if (remaining_pages_total) 1959 goto more; 1960 ret = 0; 1961 out: 1962 *num = remaining_pages_total; 1963 return ret; 1964 } 1965 #endif /* ifdef pte_index */ 1966 1967 /** 1968 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. 1969 * @vma: user vma to map to 1970 * @addr: target start user address of these pages 1971 * @pages: source kernel pages 1972 * @num: in: number of pages to map. out: number of pages that were *not* 1973 * mapped. (0 means all pages were successfully mapped). 1974 * 1975 * Preferred over vm_insert_page() when inserting multiple pages. 1976 * 1977 * In case of error, we may have mapped a subset of the provided 1978 * pages. It is the caller's responsibility to account for this case. 1979 * 1980 * The same restrictions apply as in vm_insert_page(). 1981 */ 1982 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, 1983 struct page **pages, unsigned long *num) 1984 { 1985 #ifdef pte_index 1986 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; 1987 1988 if (addr < vma->vm_start || end_addr >= vma->vm_end) 1989 return -EFAULT; 1990 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1991 BUG_ON(mmap_read_trylock(vma->vm_mm)); 1992 BUG_ON(vma->vm_flags & VM_PFNMAP); 1993 vma->vm_flags |= VM_MIXEDMAP; 1994 } 1995 /* Defer page refcount checking till we're about to map that page. */ 1996 return insert_pages(vma, addr, pages, num, vma->vm_page_prot); 1997 #else 1998 unsigned long idx = 0, pgcount = *num; 1999 int err = -EINVAL; 2000 2001 for (; idx < pgcount; ++idx) { 2002 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]); 2003 if (err) 2004 break; 2005 } 2006 *num = pgcount - idx; 2007 return err; 2008 #endif /* ifdef pte_index */ 2009 } 2010 EXPORT_SYMBOL(vm_insert_pages); 2011 2012 /** 2013 * vm_insert_page - insert single page into user vma 2014 * @vma: user vma to map to 2015 * @addr: target user address of this page 2016 * @page: source kernel page 2017 * 2018 * This allows drivers to insert individual pages they've allocated 2019 * into a user vma. 2020 * 2021 * The page has to be a nice clean _individual_ kernel allocation. 2022 * If you allocate a compound page, you need to have marked it as 2023 * such (__GFP_COMP), or manually just split the page up yourself 2024 * (see split_page()). 2025 * 2026 * NOTE! Traditionally this was done with "remap_pfn_range()" which 2027 * took an arbitrary page protection parameter. This doesn't allow 2028 * that. Your vma protection will have to be set up correctly, which 2029 * means that if you want a shared writable mapping, you'd better 2030 * ask for a shared writable mapping! 2031 * 2032 * The page does not need to be reserved. 2033 * 2034 * Usually this function is called from f_op->mmap() handler 2035 * under mm->mmap_lock write-lock, so it can change vma->vm_flags. 2036 * Caller must set VM_MIXEDMAP on vma if it wants to call this 2037 * function from other places, for example from page-fault handler. 2038 * 2039 * Return: %0 on success, negative error code otherwise. 2040 */ 2041 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 2042 struct page *page) 2043 { 2044 if (addr < vma->vm_start || addr >= vma->vm_end) 2045 return -EFAULT; 2046 if (!page_count(page)) 2047 return -EINVAL; 2048 if (!(vma->vm_flags & VM_MIXEDMAP)) { 2049 BUG_ON(mmap_read_trylock(vma->vm_mm)); 2050 BUG_ON(vma->vm_flags & VM_PFNMAP); 2051 vma->vm_flags |= VM_MIXEDMAP; 2052 } 2053 return insert_page(vma, addr, page, vma->vm_page_prot); 2054 } 2055 EXPORT_SYMBOL(vm_insert_page); 2056 2057 /* 2058 * __vm_map_pages - maps range of kernel pages into user vma 2059 * @vma: user vma to map to 2060 * @pages: pointer to array of source kernel pages 2061 * @num: number of pages in page array 2062 * @offset: user's requested vm_pgoff 2063 * 2064 * This allows drivers to map range of kernel pages into a user vma. 2065 * 2066 * Return: 0 on success and error code otherwise. 2067 */ 2068 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 2069 unsigned long num, unsigned long offset) 2070 { 2071 unsigned long count = vma_pages(vma); 2072 unsigned long uaddr = vma->vm_start; 2073 int ret, i; 2074 2075 /* Fail if the user requested offset is beyond the end of the object */ 2076 if (offset >= num) 2077 return -ENXIO; 2078 2079 /* Fail if the user requested size exceeds available object size */ 2080 if (count > num - offset) 2081 return -ENXIO; 2082 2083 for (i = 0; i < count; i++) { 2084 ret = vm_insert_page(vma, uaddr, pages[offset + i]); 2085 if (ret < 0) 2086 return ret; 2087 uaddr += PAGE_SIZE; 2088 } 2089 2090 return 0; 2091 } 2092 2093 /** 2094 * vm_map_pages - maps range of kernel pages starts with non zero offset 2095 * @vma: user vma to map to 2096 * @pages: pointer to array of source kernel pages 2097 * @num: number of pages in page array 2098 * 2099 * Maps an object consisting of @num pages, catering for the user's 2100 * requested vm_pgoff 2101 * 2102 * If we fail to insert any page into the vma, the function will return 2103 * immediately leaving any previously inserted pages present. Callers 2104 * from the mmap handler may immediately return the error as their caller 2105 * will destroy the vma, removing any successfully inserted pages. Other 2106 * callers should make their own arrangements for calling unmap_region(). 2107 * 2108 * Context: Process context. Called by mmap handlers. 2109 * Return: 0 on success and error code otherwise. 2110 */ 2111 int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 2112 unsigned long num) 2113 { 2114 return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 2115 } 2116 EXPORT_SYMBOL(vm_map_pages); 2117 2118 /** 2119 * vm_map_pages_zero - map range of kernel pages starts with zero offset 2120 * @vma: user vma to map to 2121 * @pages: pointer to array of source kernel pages 2122 * @num: number of pages in page array 2123 * 2124 * Similar to vm_map_pages(), except that it explicitly sets the offset 2125 * to 0. This function is intended for the drivers that did not consider 2126 * vm_pgoff. 2127 * 2128 * Context: Process context. Called by mmap handlers. 2129 * Return: 0 on success and error code otherwise. 2130 */ 2131 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 2132 unsigned long num) 2133 { 2134 return __vm_map_pages(vma, pages, num, 0); 2135 } 2136 EXPORT_SYMBOL(vm_map_pages_zero); 2137 2138 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2139 pfn_t pfn, pgprot_t prot, bool mkwrite) 2140 { 2141 struct mm_struct *mm = vma->vm_mm; 2142 pte_t *pte, entry; 2143 spinlock_t *ptl; 2144 2145 pte = get_locked_pte(mm, addr, &ptl); 2146 if (!pte) 2147 return VM_FAULT_OOM; 2148 if (!pte_none(*pte)) { 2149 if (mkwrite) { 2150 /* 2151 * For read faults on private mappings the PFN passed 2152 * in may not match the PFN we have mapped if the 2153 * mapped PFN is a writeable COW page. In the mkwrite 2154 * case we are creating a writable PTE for a shared 2155 * mapping and we expect the PFNs to match. If they 2156 * don't match, we are likely racing with block 2157 * allocation and mapping invalidation so just skip the 2158 * update. 2159 */ 2160 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) { 2161 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte))); 2162 goto out_unlock; 2163 } 2164 entry = pte_mkyoung(*pte); 2165 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2166 if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 2167 update_mmu_cache(vma, addr, pte); 2168 } 2169 goto out_unlock; 2170 } 2171 2172 /* Ok, finally just insert the thing.. */ 2173 if (pfn_t_devmap(pfn)) 2174 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 2175 else 2176 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 2177 2178 if (mkwrite) { 2179 entry = pte_mkyoung(entry); 2180 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2181 } 2182 2183 set_pte_at(mm, addr, pte, entry); 2184 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 2185 2186 out_unlock: 2187 pte_unmap_unlock(pte, ptl); 2188 return VM_FAULT_NOPAGE; 2189 } 2190 2191 /** 2192 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 2193 * @vma: user vma to map to 2194 * @addr: target user address of this page 2195 * @pfn: source kernel pfn 2196 * @pgprot: pgprot flags for the inserted page 2197 * 2198 * This is exactly like vmf_insert_pfn(), except that it allows drivers 2199 * to override pgprot on a per-page basis. 2200 * 2201 * This only makes sense for IO mappings, and it makes no sense for 2202 * COW mappings. In general, using multiple vmas is preferable; 2203 * vmf_insert_pfn_prot should only be used if using multiple VMAs is 2204 * impractical. 2205 * 2206 * See vmf_insert_mixed_prot() for a discussion of the implication of using 2207 * a value of @pgprot different from that of @vma->vm_page_prot. 2208 * 2209 * Context: Process context. May allocate using %GFP_KERNEL. 2210 * Return: vm_fault_t value. 2211 */ 2212 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 2213 unsigned long pfn, pgprot_t pgprot) 2214 { 2215 /* 2216 * Technically, architectures with pte_special can avoid all these 2217 * restrictions (same for remap_pfn_range). However we would like 2218 * consistency in testing and feature parity among all, so we should 2219 * try to keep these invariants in place for everybody. 2220 */ 2221 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 2222 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 2223 (VM_PFNMAP|VM_MIXEDMAP)); 2224 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 2225 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 2226 2227 if (addr < vma->vm_start || addr >= vma->vm_end) 2228 return VM_FAULT_SIGBUS; 2229 2230 if (!pfn_modify_allowed(pfn, pgprot)) 2231 return VM_FAULT_SIGBUS; 2232 2233 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 2234 2235 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 2236 false); 2237 } 2238 EXPORT_SYMBOL(vmf_insert_pfn_prot); 2239 2240 /** 2241 * vmf_insert_pfn - insert single pfn into user vma 2242 * @vma: user vma to map to 2243 * @addr: target user address of this page 2244 * @pfn: source kernel pfn 2245 * 2246 * Similar to vm_insert_page, this allows drivers to insert individual pages 2247 * they've allocated into a user vma. Same comments apply. 2248 * 2249 * This function should only be called from a vm_ops->fault handler, and 2250 * in that case the handler should return the result of this function. 2251 * 2252 * vma cannot be a COW mapping. 2253 * 2254 * As this is called only for pages that do not currently exist, we 2255 * do not need to flush old virtual caches or the TLB. 2256 * 2257 * Context: Process context. May allocate using %GFP_KERNEL. 2258 * Return: vm_fault_t value. 2259 */ 2260 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2261 unsigned long pfn) 2262 { 2263 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 2264 } 2265 EXPORT_SYMBOL(vmf_insert_pfn); 2266 2267 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 2268 { 2269 /* these checks mirror the abort conditions in vm_normal_page */ 2270 if (vma->vm_flags & VM_MIXEDMAP) 2271 return true; 2272 if (pfn_t_devmap(pfn)) 2273 return true; 2274 if (pfn_t_special(pfn)) 2275 return true; 2276 if (is_zero_pfn(pfn_t_to_pfn(pfn))) 2277 return true; 2278 return false; 2279 } 2280 2281 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 2282 unsigned long addr, pfn_t pfn, pgprot_t pgprot, 2283 bool mkwrite) 2284 { 2285 int err; 2286 2287 BUG_ON(!vm_mixed_ok(vma, pfn)); 2288 2289 if (addr < vma->vm_start || addr >= vma->vm_end) 2290 return VM_FAULT_SIGBUS; 2291 2292 track_pfn_insert(vma, &pgprot, pfn); 2293 2294 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 2295 return VM_FAULT_SIGBUS; 2296 2297 /* 2298 * If we don't have pte special, then we have to use the pfn_valid() 2299 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2300 * refcount the page if pfn_valid is true (hence insert_page rather 2301 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 2302 * without pte special, it would there be refcounted as a normal page. 2303 */ 2304 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 2305 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 2306 struct page *page; 2307 2308 /* 2309 * At this point we are committed to insert_page() 2310 * regardless of whether the caller specified flags that 2311 * result in pfn_t_has_page() == false. 2312 */ 2313 page = pfn_to_page(pfn_t_to_pfn(pfn)); 2314 err = insert_page(vma, addr, page, pgprot); 2315 } else { 2316 return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 2317 } 2318 2319 if (err == -ENOMEM) 2320 return VM_FAULT_OOM; 2321 if (err < 0 && err != -EBUSY) 2322 return VM_FAULT_SIGBUS; 2323 2324 return VM_FAULT_NOPAGE; 2325 } 2326 2327 /** 2328 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot 2329 * @vma: user vma to map to 2330 * @addr: target user address of this page 2331 * @pfn: source kernel pfn 2332 * @pgprot: pgprot flags for the inserted page 2333 * 2334 * This is exactly like vmf_insert_mixed(), except that it allows drivers 2335 * to override pgprot on a per-page basis. 2336 * 2337 * Typically this function should be used by drivers to set caching- and 2338 * encryption bits different than those of @vma->vm_page_prot, because 2339 * the caching- or encryption mode may not be known at mmap() time. 2340 * This is ok as long as @vma->vm_page_prot is not used by the core vm 2341 * to set caching and encryption bits for those vmas (except for COW pages). 2342 * This is ensured by core vm only modifying these page table entries using 2343 * functions that don't touch caching- or encryption bits, using pte_modify() 2344 * if needed. (See for example mprotect()). 2345 * Also when new page-table entries are created, this is only done using the 2346 * fault() callback, and never using the value of vma->vm_page_prot, 2347 * except for page-table entries that point to anonymous pages as the result 2348 * of COW. 2349 * 2350 * Context: Process context. May allocate using %GFP_KERNEL. 2351 * Return: vm_fault_t value. 2352 */ 2353 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr, 2354 pfn_t pfn, pgprot_t pgprot) 2355 { 2356 return __vm_insert_mixed(vma, addr, pfn, pgprot, false); 2357 } 2358 EXPORT_SYMBOL(vmf_insert_mixed_prot); 2359 2360 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 2361 pfn_t pfn) 2362 { 2363 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false); 2364 } 2365 EXPORT_SYMBOL(vmf_insert_mixed); 2366 2367 /* 2368 * If the insertion of PTE failed because someone else already added a 2369 * different entry in the mean time, we treat that as success as we assume 2370 * the same entry was actually inserted. 2371 */ 2372 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 2373 unsigned long addr, pfn_t pfn) 2374 { 2375 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true); 2376 } 2377 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 2378 2379 /* 2380 * maps a range of physical memory into the requested pages. the old 2381 * mappings are removed. any references to nonexistent pages results 2382 * in null mappings (currently treated as "copy-on-access") 2383 */ 2384 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 2385 unsigned long addr, unsigned long end, 2386 unsigned long pfn, pgprot_t prot) 2387 { 2388 pte_t *pte, *mapped_pte; 2389 spinlock_t *ptl; 2390 int err = 0; 2391 2392 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 2393 if (!pte) 2394 return -ENOMEM; 2395 arch_enter_lazy_mmu_mode(); 2396 do { 2397 BUG_ON(!pte_none(*pte)); 2398 if (!pfn_modify_allowed(pfn, prot)) { 2399 err = -EACCES; 2400 break; 2401 } 2402 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 2403 pfn++; 2404 } while (pte++, addr += PAGE_SIZE, addr != end); 2405 arch_leave_lazy_mmu_mode(); 2406 pte_unmap_unlock(mapped_pte, ptl); 2407 return err; 2408 } 2409 2410 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 2411 unsigned long addr, unsigned long end, 2412 unsigned long pfn, pgprot_t prot) 2413 { 2414 pmd_t *pmd; 2415 unsigned long next; 2416 int err; 2417 2418 pfn -= addr >> PAGE_SHIFT; 2419 pmd = pmd_alloc(mm, pud, addr); 2420 if (!pmd) 2421 return -ENOMEM; 2422 VM_BUG_ON(pmd_trans_huge(*pmd)); 2423 do { 2424 next = pmd_addr_end(addr, end); 2425 err = remap_pte_range(mm, pmd, addr, next, 2426 pfn + (addr >> PAGE_SHIFT), prot); 2427 if (err) 2428 return err; 2429 } while (pmd++, addr = next, addr != end); 2430 return 0; 2431 } 2432 2433 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 2434 unsigned long addr, unsigned long end, 2435 unsigned long pfn, pgprot_t prot) 2436 { 2437 pud_t *pud; 2438 unsigned long next; 2439 int err; 2440 2441 pfn -= addr >> PAGE_SHIFT; 2442 pud = pud_alloc(mm, p4d, addr); 2443 if (!pud) 2444 return -ENOMEM; 2445 do { 2446 next = pud_addr_end(addr, end); 2447 err = remap_pmd_range(mm, pud, addr, next, 2448 pfn + (addr >> PAGE_SHIFT), prot); 2449 if (err) 2450 return err; 2451 } while (pud++, addr = next, addr != end); 2452 return 0; 2453 } 2454 2455 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2456 unsigned long addr, unsigned long end, 2457 unsigned long pfn, pgprot_t prot) 2458 { 2459 p4d_t *p4d; 2460 unsigned long next; 2461 int err; 2462 2463 pfn -= addr >> PAGE_SHIFT; 2464 p4d = p4d_alloc(mm, pgd, addr); 2465 if (!p4d) 2466 return -ENOMEM; 2467 do { 2468 next = p4d_addr_end(addr, end); 2469 err = remap_pud_range(mm, p4d, addr, next, 2470 pfn + (addr >> PAGE_SHIFT), prot); 2471 if (err) 2472 return err; 2473 } while (p4d++, addr = next, addr != end); 2474 return 0; 2475 } 2476 2477 /* 2478 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller 2479 * must have pre-validated the caching bits of the pgprot_t. 2480 */ 2481 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr, 2482 unsigned long pfn, unsigned long size, pgprot_t prot) 2483 { 2484 pgd_t *pgd; 2485 unsigned long next; 2486 unsigned long end = addr + PAGE_ALIGN(size); 2487 struct mm_struct *mm = vma->vm_mm; 2488 int err; 2489 2490 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) 2491 return -EINVAL; 2492 2493 /* 2494 * Physically remapped pages are special. Tell the 2495 * rest of the world about it: 2496 * VM_IO tells people not to look at these pages 2497 * (accesses can have side effects). 2498 * VM_PFNMAP tells the core MM that the base pages are just 2499 * raw PFN mappings, and do not have a "struct page" associated 2500 * with them. 2501 * VM_DONTEXPAND 2502 * Disable vma merging and expanding with mremap(). 2503 * VM_DONTDUMP 2504 * Omit vma from core dump, even when VM_IO turned off. 2505 * 2506 * There's a horrible special case to handle copy-on-write 2507 * behaviour that some programs depend on. We mark the "original" 2508 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2509 * See vm_normal_page() for details. 2510 */ 2511 if (is_cow_mapping(vma->vm_flags)) { 2512 if (addr != vma->vm_start || end != vma->vm_end) 2513 return -EINVAL; 2514 vma->vm_pgoff = pfn; 2515 } 2516 2517 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 2518 2519 BUG_ON(addr >= end); 2520 pfn -= addr >> PAGE_SHIFT; 2521 pgd = pgd_offset(mm, addr); 2522 flush_cache_range(vma, addr, end); 2523 do { 2524 next = pgd_addr_end(addr, end); 2525 err = remap_p4d_range(mm, pgd, addr, next, 2526 pfn + (addr >> PAGE_SHIFT), prot); 2527 if (err) 2528 return err; 2529 } while (pgd++, addr = next, addr != end); 2530 2531 return 0; 2532 } 2533 2534 /** 2535 * remap_pfn_range - remap kernel memory to userspace 2536 * @vma: user vma to map to 2537 * @addr: target page aligned user address to start at 2538 * @pfn: page frame number of kernel physical memory address 2539 * @size: size of mapping area 2540 * @prot: page protection flags for this mapping 2541 * 2542 * Note: this is only safe if the mm semaphore is held when called. 2543 * 2544 * Return: %0 on success, negative error code otherwise. 2545 */ 2546 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 2547 unsigned long pfn, unsigned long size, pgprot_t prot) 2548 { 2549 int err; 2550 2551 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 2552 if (err) 2553 return -EINVAL; 2554 2555 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot); 2556 if (err) 2557 untrack_pfn(vma, pfn, PAGE_ALIGN(size)); 2558 return err; 2559 } 2560 EXPORT_SYMBOL(remap_pfn_range); 2561 2562 /** 2563 * vm_iomap_memory - remap memory to userspace 2564 * @vma: user vma to map to 2565 * @start: start of the physical memory to be mapped 2566 * @len: size of area 2567 * 2568 * This is a simplified io_remap_pfn_range() for common driver use. The 2569 * driver just needs to give us the physical memory range to be mapped, 2570 * we'll figure out the rest from the vma information. 2571 * 2572 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 2573 * whatever write-combining details or similar. 2574 * 2575 * Return: %0 on success, negative error code otherwise. 2576 */ 2577 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 2578 { 2579 unsigned long vm_len, pfn, pages; 2580 2581 /* Check that the physical memory area passed in looks valid */ 2582 if (start + len < start) 2583 return -EINVAL; 2584 /* 2585 * You *really* shouldn't map things that aren't page-aligned, 2586 * but we've historically allowed it because IO memory might 2587 * just have smaller alignment. 2588 */ 2589 len += start & ~PAGE_MASK; 2590 pfn = start >> PAGE_SHIFT; 2591 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 2592 if (pfn + pages < pfn) 2593 return -EINVAL; 2594 2595 /* We start the mapping 'vm_pgoff' pages into the area */ 2596 if (vma->vm_pgoff > pages) 2597 return -EINVAL; 2598 pfn += vma->vm_pgoff; 2599 pages -= vma->vm_pgoff; 2600 2601 /* Can we fit all of the mapping? */ 2602 vm_len = vma->vm_end - vma->vm_start; 2603 if (vm_len >> PAGE_SHIFT > pages) 2604 return -EINVAL; 2605 2606 /* Ok, let it rip */ 2607 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 2608 } 2609 EXPORT_SYMBOL(vm_iomap_memory); 2610 2611 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2612 unsigned long addr, unsigned long end, 2613 pte_fn_t fn, void *data, bool create, 2614 pgtbl_mod_mask *mask) 2615 { 2616 pte_t *pte, *mapped_pte; 2617 int err = 0; 2618 spinlock_t *ptl; 2619 2620 if (create) { 2621 mapped_pte = pte = (mm == &init_mm) ? 2622 pte_alloc_kernel_track(pmd, addr, mask) : 2623 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2624 if (!pte) 2625 return -ENOMEM; 2626 } else { 2627 mapped_pte = pte = (mm == &init_mm) ? 2628 pte_offset_kernel(pmd, addr) : 2629 pte_offset_map_lock(mm, pmd, addr, &ptl); 2630 } 2631 2632 BUG_ON(pmd_huge(*pmd)); 2633 2634 arch_enter_lazy_mmu_mode(); 2635 2636 if (fn) { 2637 do { 2638 if (create || !pte_none(*pte)) { 2639 err = fn(pte++, addr, data); 2640 if (err) 2641 break; 2642 } 2643 } while (addr += PAGE_SIZE, addr != end); 2644 } 2645 *mask |= PGTBL_PTE_MODIFIED; 2646 2647 arch_leave_lazy_mmu_mode(); 2648 2649 if (mm != &init_mm) 2650 pte_unmap_unlock(mapped_pte, ptl); 2651 return err; 2652 } 2653 2654 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2655 unsigned long addr, unsigned long end, 2656 pte_fn_t fn, void *data, bool create, 2657 pgtbl_mod_mask *mask) 2658 { 2659 pmd_t *pmd; 2660 unsigned long next; 2661 int err = 0; 2662 2663 BUG_ON(pud_huge(*pud)); 2664 2665 if (create) { 2666 pmd = pmd_alloc_track(mm, pud, addr, mask); 2667 if (!pmd) 2668 return -ENOMEM; 2669 } else { 2670 pmd = pmd_offset(pud, addr); 2671 } 2672 do { 2673 next = pmd_addr_end(addr, end); 2674 if (pmd_none(*pmd) && !create) 2675 continue; 2676 if (WARN_ON_ONCE(pmd_leaf(*pmd))) 2677 return -EINVAL; 2678 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) { 2679 if (!create) 2680 continue; 2681 pmd_clear_bad(pmd); 2682 } 2683 err = apply_to_pte_range(mm, pmd, addr, next, 2684 fn, data, create, mask); 2685 if (err) 2686 break; 2687 } while (pmd++, addr = next, addr != end); 2688 2689 return err; 2690 } 2691 2692 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2693 unsigned long addr, unsigned long end, 2694 pte_fn_t fn, void *data, bool create, 2695 pgtbl_mod_mask *mask) 2696 { 2697 pud_t *pud; 2698 unsigned long next; 2699 int err = 0; 2700 2701 if (create) { 2702 pud = pud_alloc_track(mm, p4d, addr, mask); 2703 if (!pud) 2704 return -ENOMEM; 2705 } else { 2706 pud = pud_offset(p4d, addr); 2707 } 2708 do { 2709 next = pud_addr_end(addr, end); 2710 if (pud_none(*pud) && !create) 2711 continue; 2712 if (WARN_ON_ONCE(pud_leaf(*pud))) 2713 return -EINVAL; 2714 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) { 2715 if (!create) 2716 continue; 2717 pud_clear_bad(pud); 2718 } 2719 err = apply_to_pmd_range(mm, pud, addr, next, 2720 fn, data, create, mask); 2721 if (err) 2722 break; 2723 } while (pud++, addr = next, addr != end); 2724 2725 return err; 2726 } 2727 2728 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2729 unsigned long addr, unsigned long end, 2730 pte_fn_t fn, void *data, bool create, 2731 pgtbl_mod_mask *mask) 2732 { 2733 p4d_t *p4d; 2734 unsigned long next; 2735 int err = 0; 2736 2737 if (create) { 2738 p4d = p4d_alloc_track(mm, pgd, addr, mask); 2739 if (!p4d) 2740 return -ENOMEM; 2741 } else { 2742 p4d = p4d_offset(pgd, addr); 2743 } 2744 do { 2745 next = p4d_addr_end(addr, end); 2746 if (p4d_none(*p4d) && !create) 2747 continue; 2748 if (WARN_ON_ONCE(p4d_leaf(*p4d))) 2749 return -EINVAL; 2750 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) { 2751 if (!create) 2752 continue; 2753 p4d_clear_bad(p4d); 2754 } 2755 err = apply_to_pud_range(mm, p4d, addr, next, 2756 fn, data, create, mask); 2757 if (err) 2758 break; 2759 } while (p4d++, addr = next, addr != end); 2760 2761 return err; 2762 } 2763 2764 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2765 unsigned long size, pte_fn_t fn, 2766 void *data, bool create) 2767 { 2768 pgd_t *pgd; 2769 unsigned long start = addr, next; 2770 unsigned long end = addr + size; 2771 pgtbl_mod_mask mask = 0; 2772 int err = 0; 2773 2774 if (WARN_ON(addr >= end)) 2775 return -EINVAL; 2776 2777 pgd = pgd_offset(mm, addr); 2778 do { 2779 next = pgd_addr_end(addr, end); 2780 if (pgd_none(*pgd) && !create) 2781 continue; 2782 if (WARN_ON_ONCE(pgd_leaf(*pgd))) 2783 return -EINVAL; 2784 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) { 2785 if (!create) 2786 continue; 2787 pgd_clear_bad(pgd); 2788 } 2789 err = apply_to_p4d_range(mm, pgd, addr, next, 2790 fn, data, create, &mask); 2791 if (err) 2792 break; 2793 } while (pgd++, addr = next, addr != end); 2794 2795 if (mask & ARCH_PAGE_TABLE_SYNC_MASK) 2796 arch_sync_kernel_mappings(start, start + size); 2797 2798 return err; 2799 } 2800 2801 /* 2802 * Scan a region of virtual memory, filling in page tables as necessary 2803 * and calling a provided function on each leaf page table. 2804 */ 2805 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2806 unsigned long size, pte_fn_t fn, void *data) 2807 { 2808 return __apply_to_page_range(mm, addr, size, fn, data, true); 2809 } 2810 EXPORT_SYMBOL_GPL(apply_to_page_range); 2811 2812 /* 2813 * Scan a region of virtual memory, calling a provided function on 2814 * each leaf page table where it exists. 2815 * 2816 * Unlike apply_to_page_range, this does _not_ fill in page tables 2817 * where they are absent. 2818 */ 2819 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, 2820 unsigned long size, pte_fn_t fn, void *data) 2821 { 2822 return __apply_to_page_range(mm, addr, size, fn, data, false); 2823 } 2824 EXPORT_SYMBOL_GPL(apply_to_existing_page_range); 2825 2826 /* 2827 * handle_pte_fault chooses page fault handler according to an entry which was 2828 * read non-atomically. Before making any commitment, on those architectures 2829 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 2830 * parts, do_swap_page must check under lock before unmapping the pte and 2831 * proceeding (but do_wp_page is only called after already making such a check; 2832 * and do_anonymous_page can safely check later on). 2833 */ 2834 static inline int pte_unmap_same(struct vm_fault *vmf) 2835 { 2836 int same = 1; 2837 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) 2838 if (sizeof(pte_t) > sizeof(unsigned long)) { 2839 spinlock_t *ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 2840 spin_lock(ptl); 2841 same = pte_same(*vmf->pte, vmf->orig_pte); 2842 spin_unlock(ptl); 2843 } 2844 #endif 2845 pte_unmap(vmf->pte); 2846 vmf->pte = NULL; 2847 return same; 2848 } 2849 2850 static inline bool __wp_page_copy_user(struct page *dst, struct page *src, 2851 struct vm_fault *vmf) 2852 { 2853 bool ret; 2854 void *kaddr; 2855 void __user *uaddr; 2856 bool locked = false; 2857 struct vm_area_struct *vma = vmf->vma; 2858 struct mm_struct *mm = vma->vm_mm; 2859 unsigned long addr = vmf->address; 2860 2861 if (likely(src)) { 2862 copy_user_highpage(dst, src, addr, vma); 2863 return true; 2864 } 2865 2866 /* 2867 * If the source page was a PFN mapping, we don't have 2868 * a "struct page" for it. We do a best-effort copy by 2869 * just copying from the original user address. If that 2870 * fails, we just zero-fill it. Live with it. 2871 */ 2872 kaddr = kmap_atomic(dst); 2873 uaddr = (void __user *)(addr & PAGE_MASK); 2874 2875 /* 2876 * On architectures with software "accessed" bits, we would 2877 * take a double page fault, so mark it accessed here. 2878 */ 2879 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) { 2880 pte_t entry; 2881 2882 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2883 locked = true; 2884 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2885 /* 2886 * Other thread has already handled the fault 2887 * and update local tlb only 2888 */ 2889 update_mmu_tlb(vma, addr, vmf->pte); 2890 ret = false; 2891 goto pte_unlock; 2892 } 2893 2894 entry = pte_mkyoung(vmf->orig_pte); 2895 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) 2896 update_mmu_cache(vma, addr, vmf->pte); 2897 } 2898 2899 /* 2900 * This really shouldn't fail, because the page is there 2901 * in the page tables. But it might just be unreadable, 2902 * in which case we just give up and fill the result with 2903 * zeroes. 2904 */ 2905 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2906 if (locked) 2907 goto warn; 2908 2909 /* Re-validate under PTL if the page is still mapped */ 2910 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2911 locked = true; 2912 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2913 /* The PTE changed under us, update local tlb */ 2914 update_mmu_tlb(vma, addr, vmf->pte); 2915 ret = false; 2916 goto pte_unlock; 2917 } 2918 2919 /* 2920 * The same page can be mapped back since last copy attempt. 2921 * Try to copy again under PTL. 2922 */ 2923 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 2924 /* 2925 * Give a warn in case there can be some obscure 2926 * use-case 2927 */ 2928 warn: 2929 WARN_ON_ONCE(1); 2930 clear_page(kaddr); 2931 } 2932 } 2933 2934 ret = true; 2935 2936 pte_unlock: 2937 if (locked) 2938 pte_unmap_unlock(vmf->pte, vmf->ptl); 2939 kunmap_atomic(kaddr); 2940 flush_dcache_page(dst); 2941 2942 return ret; 2943 } 2944 2945 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 2946 { 2947 struct file *vm_file = vma->vm_file; 2948 2949 if (vm_file) 2950 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 2951 2952 /* 2953 * Special mappings (e.g. VDSO) do not have any file so fake 2954 * a default GFP_KERNEL for them. 2955 */ 2956 return GFP_KERNEL; 2957 } 2958 2959 /* 2960 * Notify the address space that the page is about to become writable so that 2961 * it can prohibit this or wait for the page to get into an appropriate state. 2962 * 2963 * We do this without the lock held, so that it can sleep if it needs to. 2964 */ 2965 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf) 2966 { 2967 vm_fault_t ret; 2968 struct page *page = vmf->page; 2969 unsigned int old_flags = vmf->flags; 2970 2971 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2972 2973 if (vmf->vma->vm_file && 2974 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) 2975 return VM_FAULT_SIGBUS; 2976 2977 ret = vmf->vma->vm_ops->page_mkwrite(vmf); 2978 /* Restore original flags so that caller is not surprised */ 2979 vmf->flags = old_flags; 2980 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2981 return ret; 2982 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2983 lock_page(page); 2984 if (!page->mapping) { 2985 unlock_page(page); 2986 return 0; /* retry */ 2987 } 2988 ret |= VM_FAULT_LOCKED; 2989 } else 2990 VM_BUG_ON_PAGE(!PageLocked(page), page); 2991 return ret; 2992 } 2993 2994 /* 2995 * Handle dirtying of a page in shared file mapping on a write fault. 2996 * 2997 * The function expects the page to be locked and unlocks it. 2998 */ 2999 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) 3000 { 3001 struct vm_area_struct *vma = vmf->vma; 3002 struct address_space *mapping; 3003 struct page *page = vmf->page; 3004 bool dirtied; 3005 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 3006 3007 dirtied = set_page_dirty(page); 3008 VM_BUG_ON_PAGE(PageAnon(page), page); 3009 /* 3010 * Take a local copy of the address_space - page.mapping may be zeroed 3011 * by truncate after unlock_page(). The address_space itself remains 3012 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 3013 * release semantics to prevent the compiler from undoing this copying. 3014 */ 3015 mapping = page_rmapping(page); 3016 unlock_page(page); 3017 3018 if (!page_mkwrite) 3019 file_update_time(vma->vm_file); 3020 3021 /* 3022 * Throttle page dirtying rate down to writeback speed. 3023 * 3024 * mapping may be NULL here because some device drivers do not 3025 * set page.mapping but still dirty their pages 3026 * 3027 * Drop the mmap_lock before waiting on IO, if we can. The file 3028 * is pinning the mapping, as per above. 3029 */ 3030 if ((dirtied || page_mkwrite) && mapping) { 3031 struct file *fpin; 3032 3033 fpin = maybe_unlock_mmap_for_io(vmf, NULL); 3034 balance_dirty_pages_ratelimited(mapping); 3035 if (fpin) { 3036 fput(fpin); 3037 return VM_FAULT_COMPLETED; 3038 } 3039 } 3040 3041 return 0; 3042 } 3043 3044 /* 3045 * Handle write page faults for pages that can be reused in the current vma 3046 * 3047 * This can happen either due to the mapping being with the VM_SHARED flag, 3048 * or due to us being the last reference standing to the page. In either 3049 * case, all we need to do here is to mark the page as writable and update 3050 * any related book-keeping. 3051 */ 3052 static inline void wp_page_reuse(struct vm_fault *vmf) 3053 __releases(vmf->ptl) 3054 { 3055 struct vm_area_struct *vma = vmf->vma; 3056 struct page *page = vmf->page; 3057 pte_t entry; 3058 3059 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE)); 3060 VM_BUG_ON(page && PageAnon(page) && !PageAnonExclusive(page)); 3061 3062 /* 3063 * Clear the pages cpupid information as the existing 3064 * information potentially belongs to a now completely 3065 * unrelated process. 3066 */ 3067 if (page) 3068 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 3069 3070 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 3071 entry = pte_mkyoung(vmf->orig_pte); 3072 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3073 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 3074 update_mmu_cache(vma, vmf->address, vmf->pte); 3075 pte_unmap_unlock(vmf->pte, vmf->ptl); 3076 count_vm_event(PGREUSE); 3077 } 3078 3079 /* 3080 * Handle the case of a page which we actually need to copy to a new page, 3081 * either due to COW or unsharing. 3082 * 3083 * Called with mmap_lock locked and the old page referenced, but 3084 * without the ptl held. 3085 * 3086 * High level logic flow: 3087 * 3088 * - Allocate a page, copy the content of the old page to the new one. 3089 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 3090 * - Take the PTL. If the pte changed, bail out and release the allocated page 3091 * - If the pte is still the way we remember it, update the page table and all 3092 * relevant references. This includes dropping the reference the page-table 3093 * held to the old page, as well as updating the rmap. 3094 * - In any case, unlock the PTL and drop the reference we took to the old page. 3095 */ 3096 static vm_fault_t wp_page_copy(struct vm_fault *vmf) 3097 { 3098 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 3099 struct vm_area_struct *vma = vmf->vma; 3100 struct mm_struct *mm = vma->vm_mm; 3101 struct page *old_page = vmf->page; 3102 struct page *new_page = NULL; 3103 pte_t entry; 3104 int page_copied = 0; 3105 struct mmu_notifier_range range; 3106 3107 delayacct_wpcopy_start(); 3108 3109 if (unlikely(anon_vma_prepare(vma))) 3110 goto oom; 3111 3112 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) { 3113 new_page = alloc_zeroed_user_highpage_movable(vma, 3114 vmf->address); 3115 if (!new_page) 3116 goto oom; 3117 } else { 3118 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 3119 vmf->address); 3120 if (!new_page) 3121 goto oom; 3122 3123 if (!__wp_page_copy_user(new_page, old_page, vmf)) { 3124 /* 3125 * COW failed, if the fault was solved by other, 3126 * it's fine. If not, userspace would re-fault on 3127 * the same address and we will handle the fault 3128 * from the second attempt. 3129 */ 3130 put_page(new_page); 3131 if (old_page) 3132 put_page(old_page); 3133 3134 delayacct_wpcopy_end(); 3135 return 0; 3136 } 3137 kmsan_copy_page_meta(new_page, old_page); 3138 } 3139 3140 if (mem_cgroup_charge(page_folio(new_page), mm, GFP_KERNEL)) 3141 goto oom_free_new; 3142 cgroup_throttle_swaprate(new_page, GFP_KERNEL); 3143 3144 __SetPageUptodate(new_page); 3145 3146 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 3147 vmf->address & PAGE_MASK, 3148 (vmf->address & PAGE_MASK) + PAGE_SIZE); 3149 mmu_notifier_invalidate_range_start(&range); 3150 3151 /* 3152 * Re-check the pte - we dropped the lock 3153 */ 3154 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 3155 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) { 3156 if (old_page) { 3157 if (!PageAnon(old_page)) { 3158 dec_mm_counter_fast(mm, 3159 mm_counter_file(old_page)); 3160 inc_mm_counter_fast(mm, MM_ANONPAGES); 3161 } 3162 } else { 3163 inc_mm_counter_fast(mm, MM_ANONPAGES); 3164 } 3165 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 3166 entry = mk_pte(new_page, vma->vm_page_prot); 3167 entry = pte_sw_mkyoung(entry); 3168 if (unlikely(unshare)) { 3169 if (pte_soft_dirty(vmf->orig_pte)) 3170 entry = pte_mksoft_dirty(entry); 3171 if (pte_uffd_wp(vmf->orig_pte)) 3172 entry = pte_mkuffd_wp(entry); 3173 } else { 3174 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3175 } 3176 3177 /* 3178 * Clear the pte entry and flush it first, before updating the 3179 * pte with the new entry, to keep TLBs on different CPUs in 3180 * sync. This code used to set the new PTE then flush TLBs, but 3181 * that left a window where the new PTE could be loaded into 3182 * some TLBs while the old PTE remains in others. 3183 */ 3184 ptep_clear_flush_notify(vma, vmf->address, vmf->pte); 3185 page_add_new_anon_rmap(new_page, vma, vmf->address); 3186 lru_cache_add_inactive_or_unevictable(new_page, vma); 3187 /* 3188 * We call the notify macro here because, when using secondary 3189 * mmu page tables (such as kvm shadow page tables), we want the 3190 * new page to be mapped directly into the secondary page table. 3191 */ 3192 BUG_ON(unshare && pte_write(entry)); 3193 set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 3194 update_mmu_cache(vma, vmf->address, vmf->pte); 3195 if (old_page) { 3196 /* 3197 * Only after switching the pte to the new page may 3198 * we remove the mapcount here. Otherwise another 3199 * process may come and find the rmap count decremented 3200 * before the pte is switched to the new page, and 3201 * "reuse" the old page writing into it while our pte 3202 * here still points into it and can be read by other 3203 * threads. 3204 * 3205 * The critical issue is to order this 3206 * page_remove_rmap with the ptp_clear_flush above. 3207 * Those stores are ordered by (if nothing else,) 3208 * the barrier present in the atomic_add_negative 3209 * in page_remove_rmap. 3210 * 3211 * Then the TLB flush in ptep_clear_flush ensures that 3212 * no process can access the old page before the 3213 * decremented mapcount is visible. And the old page 3214 * cannot be reused until after the decremented 3215 * mapcount is visible. So transitively, TLBs to 3216 * old page will be flushed before it can be reused. 3217 */ 3218 page_remove_rmap(old_page, vma, false); 3219 } 3220 3221 /* Free the old page.. */ 3222 new_page = old_page; 3223 page_copied = 1; 3224 } else { 3225 update_mmu_tlb(vma, vmf->address, vmf->pte); 3226 } 3227 3228 if (new_page) 3229 put_page(new_page); 3230 3231 pte_unmap_unlock(vmf->pte, vmf->ptl); 3232 /* 3233 * No need to double call mmu_notifier->invalidate_range() callback as 3234 * the above ptep_clear_flush_notify() did already call it. 3235 */ 3236 mmu_notifier_invalidate_range_only_end(&range); 3237 if (old_page) { 3238 if (page_copied) 3239 free_swap_cache(old_page); 3240 put_page(old_page); 3241 } 3242 3243 delayacct_wpcopy_end(); 3244 return (page_copied && !unshare) ? VM_FAULT_WRITE : 0; 3245 oom_free_new: 3246 put_page(new_page); 3247 oom: 3248 if (old_page) 3249 put_page(old_page); 3250 3251 delayacct_wpcopy_end(); 3252 return VM_FAULT_OOM; 3253 } 3254 3255 /** 3256 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 3257 * writeable once the page is prepared 3258 * 3259 * @vmf: structure describing the fault 3260 * 3261 * This function handles all that is needed to finish a write page fault in a 3262 * shared mapping due to PTE being read-only once the mapped page is prepared. 3263 * It handles locking of PTE and modifying it. 3264 * 3265 * The function expects the page to be locked or other protection against 3266 * concurrent faults / writeback (such as DAX radix tree locks). 3267 * 3268 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before 3269 * we acquired PTE lock. 3270 */ 3271 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf) 3272 { 3273 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 3274 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 3275 &vmf->ptl); 3276 /* 3277 * We might have raced with another page fault while we released the 3278 * pte_offset_map_lock. 3279 */ 3280 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 3281 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 3282 pte_unmap_unlock(vmf->pte, vmf->ptl); 3283 return VM_FAULT_NOPAGE; 3284 } 3285 wp_page_reuse(vmf); 3286 return 0; 3287 } 3288 3289 /* 3290 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 3291 * mapping 3292 */ 3293 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 3294 { 3295 struct vm_area_struct *vma = vmf->vma; 3296 3297 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 3298 vm_fault_t ret; 3299 3300 pte_unmap_unlock(vmf->pte, vmf->ptl); 3301 vmf->flags |= FAULT_FLAG_MKWRITE; 3302 ret = vma->vm_ops->pfn_mkwrite(vmf); 3303 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 3304 return ret; 3305 return finish_mkwrite_fault(vmf); 3306 } 3307 wp_page_reuse(vmf); 3308 return VM_FAULT_WRITE; 3309 } 3310 3311 static vm_fault_t wp_page_shared(struct vm_fault *vmf) 3312 __releases(vmf->ptl) 3313 { 3314 struct vm_area_struct *vma = vmf->vma; 3315 vm_fault_t ret = VM_FAULT_WRITE; 3316 3317 get_page(vmf->page); 3318 3319 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 3320 vm_fault_t tmp; 3321 3322 pte_unmap_unlock(vmf->pte, vmf->ptl); 3323 tmp = do_page_mkwrite(vmf); 3324 if (unlikely(!tmp || (tmp & 3325 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3326 put_page(vmf->page); 3327 return tmp; 3328 } 3329 tmp = finish_mkwrite_fault(vmf); 3330 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 3331 unlock_page(vmf->page); 3332 put_page(vmf->page); 3333 return tmp; 3334 } 3335 } else { 3336 wp_page_reuse(vmf); 3337 lock_page(vmf->page); 3338 } 3339 ret |= fault_dirty_shared_page(vmf); 3340 put_page(vmf->page); 3341 3342 return ret; 3343 } 3344 3345 /* 3346 * This routine handles present pages, when 3347 * * users try to write to a shared page (FAULT_FLAG_WRITE) 3348 * * GUP wants to take a R/O pin on a possibly shared anonymous page 3349 * (FAULT_FLAG_UNSHARE) 3350 * 3351 * It is done by copying the page to a new address and decrementing the 3352 * shared-page counter for the old page. 3353 * 3354 * Note that this routine assumes that the protection checks have been 3355 * done by the caller (the low-level page fault routine in most cases). 3356 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've 3357 * done any necessary COW. 3358 * 3359 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even 3360 * though the page will change only once the write actually happens. This 3361 * avoids a few races, and potentially makes it more efficient. 3362 * 3363 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3364 * but allow concurrent faults), with pte both mapped and locked. 3365 * We return with mmap_lock still held, but pte unmapped and unlocked. 3366 */ 3367 static vm_fault_t do_wp_page(struct vm_fault *vmf) 3368 __releases(vmf->ptl) 3369 { 3370 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 3371 struct vm_area_struct *vma = vmf->vma; 3372 struct folio *folio; 3373 3374 VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE)); 3375 VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE)); 3376 3377 if (likely(!unshare)) { 3378 if (userfaultfd_pte_wp(vma, *vmf->pte)) { 3379 pte_unmap_unlock(vmf->pte, vmf->ptl); 3380 return handle_userfault(vmf, VM_UFFD_WP); 3381 } 3382 3383 /* 3384 * Userfaultfd write-protect can defer flushes. Ensure the TLB 3385 * is flushed in this case before copying. 3386 */ 3387 if (unlikely(userfaultfd_wp(vmf->vma) && 3388 mm_tlb_flush_pending(vmf->vma->vm_mm))) 3389 flush_tlb_page(vmf->vma, vmf->address); 3390 } 3391 3392 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 3393 if (!vmf->page) { 3394 if (unlikely(unshare)) { 3395 /* No anonymous page -> nothing to do. */ 3396 pte_unmap_unlock(vmf->pte, vmf->ptl); 3397 return 0; 3398 } 3399 3400 /* 3401 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 3402 * VM_PFNMAP VMA. 3403 * 3404 * We should not cow pages in a shared writeable mapping. 3405 * Just mark the pages writable and/or call ops->pfn_mkwrite. 3406 */ 3407 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3408 (VM_WRITE|VM_SHARED)) 3409 return wp_pfn_shared(vmf); 3410 3411 pte_unmap_unlock(vmf->pte, vmf->ptl); 3412 return wp_page_copy(vmf); 3413 } 3414 3415 /* 3416 * Take out anonymous pages first, anonymous shared vmas are 3417 * not dirty accountable. 3418 */ 3419 folio = page_folio(vmf->page); 3420 if (folio_test_anon(folio)) { 3421 /* 3422 * If the page is exclusive to this process we must reuse the 3423 * page without further checks. 3424 */ 3425 if (PageAnonExclusive(vmf->page)) 3426 goto reuse; 3427 3428 /* 3429 * We have to verify under folio lock: these early checks are 3430 * just an optimization to avoid locking the folio and freeing 3431 * the swapcache if there is little hope that we can reuse. 3432 * 3433 * KSM doesn't necessarily raise the folio refcount. 3434 */ 3435 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3) 3436 goto copy; 3437 if (!folio_test_lru(folio)) 3438 /* 3439 * Note: We cannot easily detect+handle references from 3440 * remote LRU pagevecs or references to LRU folios. 3441 */ 3442 lru_add_drain(); 3443 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio)) 3444 goto copy; 3445 if (!folio_trylock(folio)) 3446 goto copy; 3447 if (folio_test_swapcache(folio)) 3448 folio_free_swap(folio); 3449 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) { 3450 folio_unlock(folio); 3451 goto copy; 3452 } 3453 /* 3454 * Ok, we've got the only folio reference from our mapping 3455 * and the folio is locked, it's dark out, and we're wearing 3456 * sunglasses. Hit it. 3457 */ 3458 page_move_anon_rmap(vmf->page, vma); 3459 folio_unlock(folio); 3460 reuse: 3461 if (unlikely(unshare)) { 3462 pte_unmap_unlock(vmf->pte, vmf->ptl); 3463 return 0; 3464 } 3465 wp_page_reuse(vmf); 3466 return VM_FAULT_WRITE; 3467 } else if (unshare) { 3468 /* No anonymous page -> nothing to do. */ 3469 pte_unmap_unlock(vmf->pte, vmf->ptl); 3470 return 0; 3471 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 3472 (VM_WRITE|VM_SHARED))) { 3473 return wp_page_shared(vmf); 3474 } 3475 copy: 3476 /* 3477 * Ok, we need to copy. Oh, well.. 3478 */ 3479 get_page(vmf->page); 3480 3481 pte_unmap_unlock(vmf->pte, vmf->ptl); 3482 #ifdef CONFIG_KSM 3483 if (PageKsm(vmf->page)) 3484 count_vm_event(COW_KSM); 3485 #endif 3486 return wp_page_copy(vmf); 3487 } 3488 3489 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 3490 unsigned long start_addr, unsigned long end_addr, 3491 struct zap_details *details) 3492 { 3493 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 3494 } 3495 3496 static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 3497 pgoff_t first_index, 3498 pgoff_t last_index, 3499 struct zap_details *details) 3500 { 3501 struct vm_area_struct *vma; 3502 pgoff_t vba, vea, zba, zea; 3503 3504 vma_interval_tree_foreach(vma, root, first_index, last_index) { 3505 vba = vma->vm_pgoff; 3506 vea = vba + vma_pages(vma) - 1; 3507 zba = max(first_index, vba); 3508 zea = min(last_index, vea); 3509 3510 unmap_mapping_range_vma(vma, 3511 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 3512 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 3513 details); 3514 } 3515 } 3516 3517 /** 3518 * unmap_mapping_folio() - Unmap single folio from processes. 3519 * @folio: The locked folio to be unmapped. 3520 * 3521 * Unmap this folio from any userspace process which still has it mmaped. 3522 * Typically, for efficiency, the range of nearby pages has already been 3523 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once 3524 * truncation or invalidation holds the lock on a folio, it may find that 3525 * the page has been remapped again: and then uses unmap_mapping_folio() 3526 * to unmap it finally. 3527 */ 3528 void unmap_mapping_folio(struct folio *folio) 3529 { 3530 struct address_space *mapping = folio->mapping; 3531 struct zap_details details = { }; 3532 pgoff_t first_index; 3533 pgoff_t last_index; 3534 3535 VM_BUG_ON(!folio_test_locked(folio)); 3536 3537 first_index = folio->index; 3538 last_index = folio->index + folio_nr_pages(folio) - 1; 3539 3540 details.even_cows = false; 3541 details.single_folio = folio; 3542 details.zap_flags = ZAP_FLAG_DROP_MARKER; 3543 3544 i_mmap_lock_read(mapping); 3545 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3546 unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3547 last_index, &details); 3548 i_mmap_unlock_read(mapping); 3549 } 3550 3551 /** 3552 * unmap_mapping_pages() - Unmap pages from processes. 3553 * @mapping: The address space containing pages to be unmapped. 3554 * @start: Index of first page to be unmapped. 3555 * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 3556 * @even_cows: Whether to unmap even private COWed pages. 3557 * 3558 * Unmap the pages in this address space from any userspace process which 3559 * has them mmaped. Generally, you want to remove COWed pages as well when 3560 * a file is being truncated, but not when invalidating pages from the page 3561 * cache. 3562 */ 3563 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 3564 pgoff_t nr, bool even_cows) 3565 { 3566 struct zap_details details = { }; 3567 pgoff_t first_index = start; 3568 pgoff_t last_index = start + nr - 1; 3569 3570 details.even_cows = even_cows; 3571 if (last_index < first_index) 3572 last_index = ULONG_MAX; 3573 3574 i_mmap_lock_read(mapping); 3575 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3576 unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3577 last_index, &details); 3578 i_mmap_unlock_read(mapping); 3579 } 3580 EXPORT_SYMBOL_GPL(unmap_mapping_pages); 3581 3582 /** 3583 * unmap_mapping_range - unmap the portion of all mmaps in the specified 3584 * address_space corresponding to the specified byte range in the underlying 3585 * file. 3586 * 3587 * @mapping: the address space containing mmaps to be unmapped. 3588 * @holebegin: byte in first page to unmap, relative to the start of 3589 * the underlying file. This will be rounded down to a PAGE_SIZE 3590 * boundary. Note that this is different from truncate_pagecache(), which 3591 * must keep the partial page. In contrast, we must get rid of 3592 * partial pages. 3593 * @holelen: size of prospective hole in bytes. This will be rounded 3594 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 3595 * end of the file. 3596 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 3597 * but 0 when invalidating pagecache, don't throw away private data. 3598 */ 3599 void unmap_mapping_range(struct address_space *mapping, 3600 loff_t const holebegin, loff_t const holelen, int even_cows) 3601 { 3602 pgoff_t hba = holebegin >> PAGE_SHIFT; 3603 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3604 3605 /* Check for overflow. */ 3606 if (sizeof(holelen) > sizeof(hlen)) { 3607 long long holeend = 3608 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 3609 if (holeend & ~(long long)ULONG_MAX) 3610 hlen = ULONG_MAX - hba + 1; 3611 } 3612 3613 unmap_mapping_pages(mapping, hba, hlen, even_cows); 3614 } 3615 EXPORT_SYMBOL(unmap_mapping_range); 3616 3617 /* 3618 * Restore a potential device exclusive pte to a working pte entry 3619 */ 3620 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf) 3621 { 3622 struct folio *folio = page_folio(vmf->page); 3623 struct vm_area_struct *vma = vmf->vma; 3624 struct mmu_notifier_range range; 3625 3626 if (!folio_lock_or_retry(folio, vma->vm_mm, vmf->flags)) 3627 return VM_FAULT_RETRY; 3628 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma, 3629 vma->vm_mm, vmf->address & PAGE_MASK, 3630 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL); 3631 mmu_notifier_invalidate_range_start(&range); 3632 3633 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3634 &vmf->ptl); 3635 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3636 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte); 3637 3638 pte_unmap_unlock(vmf->pte, vmf->ptl); 3639 folio_unlock(folio); 3640 3641 mmu_notifier_invalidate_range_end(&range); 3642 return 0; 3643 } 3644 3645 static inline bool should_try_to_free_swap(struct folio *folio, 3646 struct vm_area_struct *vma, 3647 unsigned int fault_flags) 3648 { 3649 if (!folio_test_swapcache(folio)) 3650 return false; 3651 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) || 3652 folio_test_mlocked(folio)) 3653 return true; 3654 /* 3655 * If we want to map a page that's in the swapcache writable, we 3656 * have to detect via the refcount if we're really the exclusive 3657 * user. Try freeing the swapcache to get rid of the swapcache 3658 * reference only in case it's likely that we'll be the exlusive user. 3659 */ 3660 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) && 3661 folio_ref_count(folio) == 2; 3662 } 3663 3664 static vm_fault_t pte_marker_clear(struct vm_fault *vmf) 3665 { 3666 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, 3667 vmf->address, &vmf->ptl); 3668 /* 3669 * Be careful so that we will only recover a special uffd-wp pte into a 3670 * none pte. Otherwise it means the pte could have changed, so retry. 3671 */ 3672 if (is_pte_marker(*vmf->pte)) 3673 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte); 3674 pte_unmap_unlock(vmf->pte, vmf->ptl); 3675 return 0; 3676 } 3677 3678 /* 3679 * This is actually a page-missing access, but with uffd-wp special pte 3680 * installed. It means this pte was wr-protected before being unmapped. 3681 */ 3682 static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf) 3683 { 3684 /* 3685 * Just in case there're leftover special ptes even after the region 3686 * got unregistered - we can simply clear them. We can also do that 3687 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp 3688 * ranges, but it should be more efficient to be done lazily here. 3689 */ 3690 if (unlikely(!userfaultfd_wp(vmf->vma) || vma_is_anonymous(vmf->vma))) 3691 return pte_marker_clear(vmf); 3692 3693 /* do_fault() can handle pte markers too like none pte */ 3694 return do_fault(vmf); 3695 } 3696 3697 static vm_fault_t handle_pte_marker(struct vm_fault *vmf) 3698 { 3699 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte); 3700 unsigned long marker = pte_marker_get(entry); 3701 3702 /* 3703 * PTE markers should always be with file-backed memories, and the 3704 * marker should never be empty. If anything weird happened, the best 3705 * thing to do is to kill the process along with its mm. 3706 */ 3707 if (WARN_ON_ONCE(vma_is_anonymous(vmf->vma) || !marker)) 3708 return VM_FAULT_SIGBUS; 3709 3710 if (pte_marker_entry_uffd_wp(entry)) 3711 return pte_marker_handle_uffd_wp(vmf); 3712 3713 /* This is an unknown pte marker */ 3714 return VM_FAULT_SIGBUS; 3715 } 3716 3717 /* 3718 * We enter with non-exclusive mmap_lock (to exclude vma changes, 3719 * but allow concurrent faults), and pte mapped but not yet locked. 3720 * We return with pte unmapped and unlocked. 3721 * 3722 * We return with the mmap_lock locked or unlocked in the same cases 3723 * as does filemap_fault(). 3724 */ 3725 vm_fault_t do_swap_page(struct vm_fault *vmf) 3726 { 3727 struct vm_area_struct *vma = vmf->vma; 3728 struct folio *swapcache, *folio = NULL; 3729 struct page *page; 3730 struct swap_info_struct *si = NULL; 3731 rmap_t rmap_flags = RMAP_NONE; 3732 bool exclusive = false; 3733 swp_entry_t entry; 3734 pte_t pte; 3735 int locked; 3736 vm_fault_t ret = 0; 3737 void *shadow = NULL; 3738 3739 if (!pte_unmap_same(vmf)) 3740 goto out; 3741 3742 entry = pte_to_swp_entry(vmf->orig_pte); 3743 if (unlikely(non_swap_entry(entry))) { 3744 if (is_migration_entry(entry)) { 3745 migration_entry_wait(vma->vm_mm, vmf->pmd, 3746 vmf->address); 3747 } else if (is_device_exclusive_entry(entry)) { 3748 vmf->page = pfn_swap_entry_to_page(entry); 3749 ret = remove_device_exclusive_entry(vmf); 3750 } else if (is_device_private_entry(entry)) { 3751 vmf->page = pfn_swap_entry_to_page(entry); 3752 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3753 vmf->address, &vmf->ptl); 3754 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 3755 spin_unlock(vmf->ptl); 3756 goto out; 3757 } 3758 3759 /* 3760 * Get a page reference while we know the page can't be 3761 * freed. 3762 */ 3763 get_page(vmf->page); 3764 pte_unmap_unlock(vmf->pte, vmf->ptl); 3765 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 3766 put_page(vmf->page); 3767 } else if (is_hwpoison_entry(entry)) { 3768 ret = VM_FAULT_HWPOISON; 3769 } else if (is_swapin_error_entry(entry)) { 3770 ret = VM_FAULT_SIGBUS; 3771 } else if (is_pte_marker_entry(entry)) { 3772 ret = handle_pte_marker(vmf); 3773 } else { 3774 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 3775 ret = VM_FAULT_SIGBUS; 3776 } 3777 goto out; 3778 } 3779 3780 /* Prevent swapoff from happening to us. */ 3781 si = get_swap_device(entry); 3782 if (unlikely(!si)) 3783 goto out; 3784 3785 folio = swap_cache_get_folio(entry, vma, vmf->address); 3786 if (folio) 3787 page = folio_file_page(folio, swp_offset(entry)); 3788 swapcache = folio; 3789 3790 if (!folio) { 3791 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && 3792 __swap_count(entry) == 1) { 3793 /* skip swapcache */ 3794 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, 3795 vma, vmf->address, false); 3796 page = &folio->page; 3797 if (folio) { 3798 __folio_set_locked(folio); 3799 __folio_set_swapbacked(folio); 3800 3801 if (mem_cgroup_swapin_charge_folio(folio, 3802 vma->vm_mm, GFP_KERNEL, 3803 entry)) { 3804 ret = VM_FAULT_OOM; 3805 goto out_page; 3806 } 3807 mem_cgroup_swapin_uncharge_swap(entry); 3808 3809 shadow = get_shadow_from_swap_cache(entry); 3810 if (shadow) 3811 workingset_refault(folio, shadow); 3812 3813 folio_add_lru(folio); 3814 3815 /* To provide entry to swap_readpage() */ 3816 folio_set_swap_entry(folio, entry); 3817 swap_readpage(page, true, NULL); 3818 folio->private = NULL; 3819 } 3820 } else { 3821 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 3822 vmf); 3823 if (page) 3824 folio = page_folio(page); 3825 swapcache = folio; 3826 } 3827 3828 if (!folio) { 3829 /* 3830 * Back out if somebody else faulted in this pte 3831 * while we released the pte lock. 3832 */ 3833 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 3834 vmf->address, &vmf->ptl); 3835 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 3836 ret = VM_FAULT_OOM; 3837 goto unlock; 3838 } 3839 3840 /* Had to read the page from swap area: Major fault */ 3841 ret = VM_FAULT_MAJOR; 3842 count_vm_event(PGMAJFAULT); 3843 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 3844 } else if (PageHWPoison(page)) { 3845 /* 3846 * hwpoisoned dirty swapcache pages are kept for killing 3847 * owner processes (which may be unknown at hwpoison time) 3848 */ 3849 ret = VM_FAULT_HWPOISON; 3850 goto out_release; 3851 } 3852 3853 locked = folio_lock_or_retry(folio, vma->vm_mm, vmf->flags); 3854 3855 if (!locked) { 3856 ret |= VM_FAULT_RETRY; 3857 goto out_release; 3858 } 3859 3860 if (swapcache) { 3861 /* 3862 * Make sure folio_free_swap() or swapoff did not release the 3863 * swapcache from under us. The page pin, and pte_same test 3864 * below, are not enough to exclude that. Even if it is still 3865 * swapcache, we need to check that the page's swap has not 3866 * changed. 3867 */ 3868 if (unlikely(!folio_test_swapcache(folio) || 3869 page_private(page) != entry.val)) 3870 goto out_page; 3871 3872 /* 3873 * KSM sometimes has to copy on read faults, for example, if 3874 * page->index of !PageKSM() pages would be nonlinear inside the 3875 * anon VMA -- PageKSM() is lost on actual swapout. 3876 */ 3877 page = ksm_might_need_to_copy(page, vma, vmf->address); 3878 if (unlikely(!page)) { 3879 ret = VM_FAULT_OOM; 3880 goto out_page; 3881 } 3882 folio = page_folio(page); 3883 3884 /* 3885 * If we want to map a page that's in the swapcache writable, we 3886 * have to detect via the refcount if we're really the exclusive 3887 * owner. Try removing the extra reference from the local LRU 3888 * pagevecs if required. 3889 */ 3890 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache && 3891 !folio_test_ksm(folio) && !folio_test_lru(folio)) 3892 lru_add_drain(); 3893 } 3894 3895 cgroup_throttle_swaprate(page, GFP_KERNEL); 3896 3897 /* 3898 * Back out if somebody else already faulted in this pte. 3899 */ 3900 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3901 &vmf->ptl); 3902 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) 3903 goto out_nomap; 3904 3905 if (unlikely(!folio_test_uptodate(folio))) { 3906 ret = VM_FAULT_SIGBUS; 3907 goto out_nomap; 3908 } 3909 3910 /* 3911 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte 3912 * must never point at an anonymous page in the swapcache that is 3913 * PG_anon_exclusive. Sanity check that this holds and especially, that 3914 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity 3915 * check after taking the PT lock and making sure that nobody 3916 * concurrently faulted in this page and set PG_anon_exclusive. 3917 */ 3918 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio)); 3919 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page)); 3920 3921 /* 3922 * Check under PT lock (to protect against concurrent fork() sharing 3923 * the swap entry concurrently) for certainly exclusive pages. 3924 */ 3925 if (!folio_test_ksm(folio)) { 3926 /* 3927 * Note that pte_swp_exclusive() == false for architectures 3928 * without __HAVE_ARCH_PTE_SWP_EXCLUSIVE. 3929 */ 3930 exclusive = pte_swp_exclusive(vmf->orig_pte); 3931 if (folio != swapcache) { 3932 /* 3933 * We have a fresh page that is not exposed to the 3934 * swapcache -> certainly exclusive. 3935 */ 3936 exclusive = true; 3937 } else if (exclusive && folio_test_writeback(folio) && 3938 data_race(si->flags & SWP_STABLE_WRITES)) { 3939 /* 3940 * This is tricky: not all swap backends support 3941 * concurrent page modifications while under writeback. 3942 * 3943 * So if we stumble over such a page in the swapcache 3944 * we must not set the page exclusive, otherwise we can 3945 * map it writable without further checks and modify it 3946 * while still under writeback. 3947 * 3948 * For these problematic swap backends, simply drop the 3949 * exclusive marker: this is perfectly fine as we start 3950 * writeback only if we fully unmapped the page and 3951 * there are no unexpected references on the page after 3952 * unmapping succeeded. After fully unmapped, no 3953 * further GUP references (FOLL_GET and FOLL_PIN) can 3954 * appear, so dropping the exclusive marker and mapping 3955 * it only R/O is fine. 3956 */ 3957 exclusive = false; 3958 } 3959 } 3960 3961 /* 3962 * Remove the swap entry and conditionally try to free up the swapcache. 3963 * We're already holding a reference on the page but haven't mapped it 3964 * yet. 3965 */ 3966 swap_free(entry); 3967 if (should_try_to_free_swap(folio, vma, vmf->flags)) 3968 folio_free_swap(folio); 3969 3970 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3971 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS); 3972 pte = mk_pte(page, vma->vm_page_prot); 3973 3974 /* 3975 * Same logic as in do_wp_page(); however, optimize for pages that are 3976 * certainly not shared either because we just allocated them without 3977 * exposing them to the swapcache or because the swap entry indicates 3978 * exclusivity. 3979 */ 3980 if (!folio_test_ksm(folio) && 3981 (exclusive || folio_ref_count(folio) == 1)) { 3982 if (vmf->flags & FAULT_FLAG_WRITE) { 3983 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 3984 vmf->flags &= ~FAULT_FLAG_WRITE; 3985 ret |= VM_FAULT_WRITE; 3986 } 3987 rmap_flags |= RMAP_EXCLUSIVE; 3988 } 3989 flush_icache_page(vma, page); 3990 if (pte_swp_soft_dirty(vmf->orig_pte)) 3991 pte = pte_mksoft_dirty(pte); 3992 if (pte_swp_uffd_wp(vmf->orig_pte)) { 3993 pte = pte_mkuffd_wp(pte); 3994 pte = pte_wrprotect(pte); 3995 } 3996 vmf->orig_pte = pte; 3997 3998 /* ksm created a completely new copy */ 3999 if (unlikely(folio != swapcache && swapcache)) { 4000 page_add_new_anon_rmap(page, vma, vmf->address); 4001 folio_add_lru_vma(folio, vma); 4002 } else { 4003 page_add_anon_rmap(page, vma, vmf->address, rmap_flags); 4004 } 4005 4006 VM_BUG_ON(!folio_test_anon(folio) || 4007 (pte_write(pte) && !PageAnonExclusive(page))); 4008 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 4009 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 4010 4011 folio_unlock(folio); 4012 if (folio != swapcache && swapcache) { 4013 /* 4014 * Hold the lock to avoid the swap entry to be reused 4015 * until we take the PT lock for the pte_same() check 4016 * (to avoid false positives from pte_same). For 4017 * further safety release the lock after the swap_free 4018 * so that the swap count won't change under a 4019 * parallel locked swapcache. 4020 */ 4021 folio_unlock(swapcache); 4022 folio_put(swapcache); 4023 } 4024 4025 if (vmf->flags & FAULT_FLAG_WRITE) { 4026 ret |= do_wp_page(vmf); 4027 if (ret & VM_FAULT_ERROR) 4028 ret &= VM_FAULT_ERROR; 4029 goto out; 4030 } 4031 4032 /* No need to invalidate - it was non-present before */ 4033 update_mmu_cache(vma, vmf->address, vmf->pte); 4034 unlock: 4035 pte_unmap_unlock(vmf->pte, vmf->ptl); 4036 out: 4037 if (si) 4038 put_swap_device(si); 4039 return ret; 4040 out_nomap: 4041 pte_unmap_unlock(vmf->pte, vmf->ptl); 4042 out_page: 4043 folio_unlock(folio); 4044 out_release: 4045 folio_put(folio); 4046 if (folio != swapcache && swapcache) { 4047 folio_unlock(swapcache); 4048 folio_put(swapcache); 4049 } 4050 if (si) 4051 put_swap_device(si); 4052 return ret; 4053 } 4054 4055 /* 4056 * We enter with non-exclusive mmap_lock (to exclude vma changes, 4057 * but allow concurrent faults), and pte mapped but not yet locked. 4058 * We return with mmap_lock still held, but pte unmapped and unlocked. 4059 */ 4060 static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 4061 { 4062 struct vm_area_struct *vma = vmf->vma; 4063 struct page *page; 4064 vm_fault_t ret = 0; 4065 pte_t entry; 4066 4067 /* File mapping without ->vm_ops ? */ 4068 if (vma->vm_flags & VM_SHARED) 4069 return VM_FAULT_SIGBUS; 4070 4071 /* 4072 * Use pte_alloc() instead of pte_alloc_map(). We can't run 4073 * pte_offset_map() on pmds where a huge pmd might be created 4074 * from a different thread. 4075 * 4076 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 4077 * parallel threads are excluded by other means. 4078 * 4079 * Here we only have mmap_read_lock(mm). 4080 */ 4081 if (pte_alloc(vma->vm_mm, vmf->pmd)) 4082 return VM_FAULT_OOM; 4083 4084 /* See comment in handle_pte_fault() */ 4085 if (unlikely(pmd_trans_unstable(vmf->pmd))) 4086 return 0; 4087 4088 /* Use the zero-page for reads */ 4089 if (!(vmf->flags & FAULT_FLAG_WRITE) && 4090 !mm_forbids_zeropage(vma->vm_mm)) { 4091 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 4092 vma->vm_page_prot)); 4093 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 4094 vmf->address, &vmf->ptl); 4095 if (!pte_none(*vmf->pte)) { 4096 update_mmu_tlb(vma, vmf->address, vmf->pte); 4097 goto unlock; 4098 } 4099 ret = check_stable_address_space(vma->vm_mm); 4100 if (ret) 4101 goto unlock; 4102 /* Deliver the page fault to userland, check inside PT lock */ 4103 if (userfaultfd_missing(vma)) { 4104 pte_unmap_unlock(vmf->pte, vmf->ptl); 4105 return handle_userfault(vmf, VM_UFFD_MISSING); 4106 } 4107 goto setpte; 4108 } 4109 4110 /* Allocate our own private page. */ 4111 if (unlikely(anon_vma_prepare(vma))) 4112 goto oom; 4113 page = alloc_zeroed_user_highpage_movable(vma, vmf->address); 4114 if (!page) 4115 goto oom; 4116 4117 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL)) 4118 goto oom_free_page; 4119 cgroup_throttle_swaprate(page, GFP_KERNEL); 4120 4121 /* 4122 * The memory barrier inside __SetPageUptodate makes sure that 4123 * preceding stores to the page contents become visible before 4124 * the set_pte_at() write. 4125 */ 4126 __SetPageUptodate(page); 4127 4128 entry = mk_pte(page, vma->vm_page_prot); 4129 entry = pte_sw_mkyoung(entry); 4130 if (vma->vm_flags & VM_WRITE) 4131 entry = pte_mkwrite(pte_mkdirty(entry)); 4132 4133 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 4134 &vmf->ptl); 4135 if (!pte_none(*vmf->pte)) { 4136 update_mmu_tlb(vma, vmf->address, vmf->pte); 4137 goto release; 4138 } 4139 4140 ret = check_stable_address_space(vma->vm_mm); 4141 if (ret) 4142 goto release; 4143 4144 /* Deliver the page fault to userland, check inside PT lock */ 4145 if (userfaultfd_missing(vma)) { 4146 pte_unmap_unlock(vmf->pte, vmf->ptl); 4147 put_page(page); 4148 return handle_userfault(vmf, VM_UFFD_MISSING); 4149 } 4150 4151 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 4152 page_add_new_anon_rmap(page, vma, vmf->address); 4153 lru_cache_add_inactive_or_unevictable(page, vma); 4154 setpte: 4155 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 4156 4157 /* No need to invalidate - it was non-present before */ 4158 update_mmu_cache(vma, vmf->address, vmf->pte); 4159 unlock: 4160 pte_unmap_unlock(vmf->pte, vmf->ptl); 4161 return ret; 4162 release: 4163 put_page(page); 4164 goto unlock; 4165 oom_free_page: 4166 put_page(page); 4167 oom: 4168 return VM_FAULT_OOM; 4169 } 4170 4171 /* 4172 * The mmap_lock must have been held on entry, and may have been 4173 * released depending on flags and vma->vm_ops->fault() return value. 4174 * See filemap_fault() and __lock_page_retry(). 4175 */ 4176 static vm_fault_t __do_fault(struct vm_fault *vmf) 4177 { 4178 struct vm_area_struct *vma = vmf->vma; 4179 vm_fault_t ret; 4180 4181 /* 4182 * Preallocate pte before we take page_lock because this might lead to 4183 * deadlocks for memcg reclaim which waits for pages under writeback: 4184 * lock_page(A) 4185 * SetPageWriteback(A) 4186 * unlock_page(A) 4187 * lock_page(B) 4188 * lock_page(B) 4189 * pte_alloc_one 4190 * shrink_page_list 4191 * wait_on_page_writeback(A) 4192 * SetPageWriteback(B) 4193 * unlock_page(B) 4194 * # flush A, B to clear the writeback 4195 */ 4196 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 4197 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 4198 if (!vmf->prealloc_pte) 4199 return VM_FAULT_OOM; 4200 } 4201 4202 ret = vma->vm_ops->fault(vmf); 4203 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 4204 VM_FAULT_DONE_COW))) 4205 return ret; 4206 4207 if (unlikely(PageHWPoison(vmf->page))) { 4208 struct page *page = vmf->page; 4209 vm_fault_t poisonret = VM_FAULT_HWPOISON; 4210 if (ret & VM_FAULT_LOCKED) { 4211 if (page_mapped(page)) 4212 unmap_mapping_pages(page_mapping(page), 4213 page->index, 1, false); 4214 /* Retry if a clean page was removed from the cache. */ 4215 if (invalidate_inode_page(page)) 4216 poisonret = VM_FAULT_NOPAGE; 4217 unlock_page(page); 4218 } 4219 put_page(page); 4220 vmf->page = NULL; 4221 return poisonret; 4222 } 4223 4224 if (unlikely(!(ret & VM_FAULT_LOCKED))) 4225 lock_page(vmf->page); 4226 else 4227 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page); 4228 4229 return ret; 4230 } 4231 4232 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4233 static void deposit_prealloc_pte(struct vm_fault *vmf) 4234 { 4235 struct vm_area_struct *vma = vmf->vma; 4236 4237 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 4238 /* 4239 * We are going to consume the prealloc table, 4240 * count that as nr_ptes. 4241 */ 4242 mm_inc_nr_ptes(vma->vm_mm); 4243 vmf->prealloc_pte = NULL; 4244 } 4245 4246 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 4247 { 4248 struct vm_area_struct *vma = vmf->vma; 4249 bool write = vmf->flags & FAULT_FLAG_WRITE; 4250 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 4251 pmd_t entry; 4252 int i; 4253 vm_fault_t ret = VM_FAULT_FALLBACK; 4254 4255 if (!transhuge_vma_suitable(vma, haddr)) 4256 return ret; 4257 4258 page = compound_head(page); 4259 if (compound_order(page) != HPAGE_PMD_ORDER) 4260 return ret; 4261 4262 /* 4263 * Just backoff if any subpage of a THP is corrupted otherwise 4264 * the corrupted page may mapped by PMD silently to escape the 4265 * check. This kind of THP just can be PTE mapped. Access to 4266 * the corrupted subpage should trigger SIGBUS as expected. 4267 */ 4268 if (unlikely(PageHasHWPoisoned(page))) 4269 return ret; 4270 4271 /* 4272 * Archs like ppc64 need additional space to store information 4273 * related to pte entry. Use the preallocated table for that. 4274 */ 4275 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 4276 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 4277 if (!vmf->prealloc_pte) 4278 return VM_FAULT_OOM; 4279 } 4280 4281 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 4282 if (unlikely(!pmd_none(*vmf->pmd))) 4283 goto out; 4284 4285 for (i = 0; i < HPAGE_PMD_NR; i++) 4286 flush_icache_page(vma, page + i); 4287 4288 entry = mk_huge_pmd(page, vma->vm_page_prot); 4289 if (write) 4290 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 4291 4292 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR); 4293 page_add_file_rmap(page, vma, true); 4294 4295 /* 4296 * deposit and withdraw with pmd lock held 4297 */ 4298 if (arch_needs_pgtable_deposit()) 4299 deposit_prealloc_pte(vmf); 4300 4301 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 4302 4303 update_mmu_cache_pmd(vma, haddr, vmf->pmd); 4304 4305 /* fault is handled */ 4306 ret = 0; 4307 count_vm_event(THP_FILE_MAPPED); 4308 out: 4309 spin_unlock(vmf->ptl); 4310 return ret; 4311 } 4312 #else 4313 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 4314 { 4315 return VM_FAULT_FALLBACK; 4316 } 4317 #endif 4318 4319 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr) 4320 { 4321 struct vm_area_struct *vma = vmf->vma; 4322 bool uffd_wp = pte_marker_uffd_wp(vmf->orig_pte); 4323 bool write = vmf->flags & FAULT_FLAG_WRITE; 4324 bool prefault = vmf->address != addr; 4325 pte_t entry; 4326 4327 flush_icache_page(vma, page); 4328 entry = mk_pte(page, vma->vm_page_prot); 4329 4330 if (prefault && arch_wants_old_prefaulted_pte()) 4331 entry = pte_mkold(entry); 4332 else 4333 entry = pte_sw_mkyoung(entry); 4334 4335 if (write) 4336 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 4337 if (unlikely(uffd_wp)) 4338 entry = pte_mkuffd_wp(pte_wrprotect(entry)); 4339 /* copy-on-write page */ 4340 if (write && !(vma->vm_flags & VM_SHARED)) { 4341 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 4342 page_add_new_anon_rmap(page, vma, addr); 4343 lru_cache_add_inactive_or_unevictable(page, vma); 4344 } else { 4345 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 4346 page_add_file_rmap(page, vma, false); 4347 } 4348 set_pte_at(vma->vm_mm, addr, vmf->pte, entry); 4349 } 4350 4351 static bool vmf_pte_changed(struct vm_fault *vmf) 4352 { 4353 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID) 4354 return !pte_same(*vmf->pte, vmf->orig_pte); 4355 4356 return !pte_none(*vmf->pte); 4357 } 4358 4359 /** 4360 * finish_fault - finish page fault once we have prepared the page to fault 4361 * 4362 * @vmf: structure describing the fault 4363 * 4364 * This function handles all that is needed to finish a page fault once the 4365 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 4366 * given page, adds reverse page mapping, handles memcg charges and LRU 4367 * addition. 4368 * 4369 * The function expects the page to be locked and on success it consumes a 4370 * reference of a page being mapped (for the PTE which maps it). 4371 * 4372 * Return: %0 on success, %VM_FAULT_ code in case of error. 4373 */ 4374 vm_fault_t finish_fault(struct vm_fault *vmf) 4375 { 4376 struct vm_area_struct *vma = vmf->vma; 4377 struct page *page; 4378 vm_fault_t ret; 4379 4380 /* Did we COW the page? */ 4381 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 4382 page = vmf->cow_page; 4383 else 4384 page = vmf->page; 4385 4386 /* 4387 * check even for read faults because we might have lost our CoWed 4388 * page 4389 */ 4390 if (!(vma->vm_flags & VM_SHARED)) { 4391 ret = check_stable_address_space(vma->vm_mm); 4392 if (ret) 4393 return ret; 4394 } 4395 4396 if (pmd_none(*vmf->pmd)) { 4397 if (PageTransCompound(page)) { 4398 ret = do_set_pmd(vmf, page); 4399 if (ret != VM_FAULT_FALLBACK) 4400 return ret; 4401 } 4402 4403 if (vmf->prealloc_pte) 4404 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte); 4405 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) 4406 return VM_FAULT_OOM; 4407 } 4408 4409 /* 4410 * See comment in handle_pte_fault() for how this scenario happens, we 4411 * need to return NOPAGE so that we drop this page. 4412 */ 4413 if (pmd_devmap_trans_unstable(vmf->pmd)) 4414 return VM_FAULT_NOPAGE; 4415 4416 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 4417 vmf->address, &vmf->ptl); 4418 4419 /* Re-check under ptl */ 4420 if (likely(!vmf_pte_changed(vmf))) { 4421 do_set_pte(vmf, page, vmf->address); 4422 4423 /* no need to invalidate: a not-present page won't be cached */ 4424 update_mmu_cache(vma, vmf->address, vmf->pte); 4425 4426 ret = 0; 4427 } else { 4428 update_mmu_tlb(vma, vmf->address, vmf->pte); 4429 ret = VM_FAULT_NOPAGE; 4430 } 4431 4432 pte_unmap_unlock(vmf->pte, vmf->ptl); 4433 return ret; 4434 } 4435 4436 static unsigned long fault_around_bytes __read_mostly = 4437 rounddown_pow_of_two(65536); 4438 4439 #ifdef CONFIG_DEBUG_FS 4440 static int fault_around_bytes_get(void *data, u64 *val) 4441 { 4442 *val = fault_around_bytes; 4443 return 0; 4444 } 4445 4446 /* 4447 * fault_around_bytes must be rounded down to the nearest page order as it's 4448 * what do_fault_around() expects to see. 4449 */ 4450 static int fault_around_bytes_set(void *data, u64 val) 4451 { 4452 if (val / PAGE_SIZE > PTRS_PER_PTE) 4453 return -EINVAL; 4454 if (val > PAGE_SIZE) 4455 fault_around_bytes = rounddown_pow_of_two(val); 4456 else 4457 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 4458 return 0; 4459 } 4460 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 4461 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 4462 4463 static int __init fault_around_debugfs(void) 4464 { 4465 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 4466 &fault_around_bytes_fops); 4467 return 0; 4468 } 4469 late_initcall(fault_around_debugfs); 4470 #endif 4471 4472 /* 4473 * do_fault_around() tries to map few pages around the fault address. The hope 4474 * is that the pages will be needed soon and this will lower the number of 4475 * faults to handle. 4476 * 4477 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 4478 * not ready to be mapped: not up-to-date, locked, etc. 4479 * 4480 * This function doesn't cross the VMA boundaries, in order to call map_pages() 4481 * only once. 4482 * 4483 * fault_around_bytes defines how many bytes we'll try to map. 4484 * do_fault_around() expects it to be set to a power of two less than or equal 4485 * to PTRS_PER_PTE. 4486 * 4487 * The virtual address of the area that we map is naturally aligned to 4488 * fault_around_bytes rounded down to the machine page size 4489 * (and therefore to page order). This way it's easier to guarantee 4490 * that we don't cross page table boundaries. 4491 */ 4492 static vm_fault_t do_fault_around(struct vm_fault *vmf) 4493 { 4494 unsigned long address = vmf->address, nr_pages, mask; 4495 pgoff_t start_pgoff = vmf->pgoff; 4496 pgoff_t end_pgoff; 4497 int off; 4498 4499 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 4500 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 4501 4502 address = max(address & mask, vmf->vma->vm_start); 4503 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 4504 start_pgoff -= off; 4505 4506 /* 4507 * end_pgoff is either the end of the page table, the end of 4508 * the vma or nr_pages from start_pgoff, depending what is nearest. 4509 */ 4510 end_pgoff = start_pgoff - 4511 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 4512 PTRS_PER_PTE - 1; 4513 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1, 4514 start_pgoff + nr_pages - 1); 4515 4516 if (pmd_none(*vmf->pmd)) { 4517 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 4518 if (!vmf->prealloc_pte) 4519 return VM_FAULT_OOM; 4520 } 4521 4522 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff); 4523 } 4524 4525 /* Return true if we should do read fault-around, false otherwise */ 4526 static inline bool should_fault_around(struct vm_fault *vmf) 4527 { 4528 /* No ->map_pages? No way to fault around... */ 4529 if (!vmf->vma->vm_ops->map_pages) 4530 return false; 4531 4532 if (uffd_disable_fault_around(vmf->vma)) 4533 return false; 4534 4535 return fault_around_bytes >> PAGE_SHIFT > 1; 4536 } 4537 4538 static vm_fault_t do_read_fault(struct vm_fault *vmf) 4539 { 4540 vm_fault_t ret = 0; 4541 4542 /* 4543 * Let's call ->map_pages() first and use ->fault() as fallback 4544 * if page by the offset is not ready to be mapped (cold cache or 4545 * something). 4546 */ 4547 if (should_fault_around(vmf)) { 4548 ret = do_fault_around(vmf); 4549 if (ret) 4550 return ret; 4551 } 4552 4553 ret = __do_fault(vmf); 4554 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4555 return ret; 4556 4557 ret |= finish_fault(vmf); 4558 unlock_page(vmf->page); 4559 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4560 put_page(vmf->page); 4561 return ret; 4562 } 4563 4564 static vm_fault_t do_cow_fault(struct vm_fault *vmf) 4565 { 4566 struct vm_area_struct *vma = vmf->vma; 4567 vm_fault_t ret; 4568 4569 if (unlikely(anon_vma_prepare(vma))) 4570 return VM_FAULT_OOM; 4571 4572 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address); 4573 if (!vmf->cow_page) 4574 return VM_FAULT_OOM; 4575 4576 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm, 4577 GFP_KERNEL)) { 4578 put_page(vmf->cow_page); 4579 return VM_FAULT_OOM; 4580 } 4581 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL); 4582 4583 ret = __do_fault(vmf); 4584 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4585 goto uncharge_out; 4586 if (ret & VM_FAULT_DONE_COW) 4587 return ret; 4588 4589 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 4590 __SetPageUptodate(vmf->cow_page); 4591 4592 ret |= finish_fault(vmf); 4593 unlock_page(vmf->page); 4594 put_page(vmf->page); 4595 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4596 goto uncharge_out; 4597 return ret; 4598 uncharge_out: 4599 put_page(vmf->cow_page); 4600 return ret; 4601 } 4602 4603 static vm_fault_t do_shared_fault(struct vm_fault *vmf) 4604 { 4605 struct vm_area_struct *vma = vmf->vma; 4606 vm_fault_t ret, tmp; 4607 4608 ret = __do_fault(vmf); 4609 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4610 return ret; 4611 4612 /* 4613 * Check if the backing address space wants to know that the page is 4614 * about to become writable 4615 */ 4616 if (vma->vm_ops->page_mkwrite) { 4617 unlock_page(vmf->page); 4618 tmp = do_page_mkwrite(vmf); 4619 if (unlikely(!tmp || 4620 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 4621 put_page(vmf->page); 4622 return tmp; 4623 } 4624 } 4625 4626 ret |= finish_fault(vmf); 4627 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 4628 VM_FAULT_RETRY))) { 4629 unlock_page(vmf->page); 4630 put_page(vmf->page); 4631 return ret; 4632 } 4633 4634 ret |= fault_dirty_shared_page(vmf); 4635 return ret; 4636 } 4637 4638 /* 4639 * We enter with non-exclusive mmap_lock (to exclude vma changes, 4640 * but allow concurrent faults). 4641 * The mmap_lock may have been released depending on flags and our 4642 * return value. See filemap_fault() and __folio_lock_or_retry(). 4643 * If mmap_lock is released, vma may become invalid (for example 4644 * by other thread calling munmap()). 4645 */ 4646 static vm_fault_t do_fault(struct vm_fault *vmf) 4647 { 4648 struct vm_area_struct *vma = vmf->vma; 4649 struct mm_struct *vm_mm = vma->vm_mm; 4650 vm_fault_t ret; 4651 4652 /* 4653 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 4654 */ 4655 if (!vma->vm_ops->fault) { 4656 /* 4657 * If we find a migration pmd entry or a none pmd entry, which 4658 * should never happen, return SIGBUS 4659 */ 4660 if (unlikely(!pmd_present(*vmf->pmd))) 4661 ret = VM_FAULT_SIGBUS; 4662 else { 4663 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, 4664 vmf->pmd, 4665 vmf->address, 4666 &vmf->ptl); 4667 /* 4668 * Make sure this is not a temporary clearing of pte 4669 * by holding ptl and checking again. A R/M/W update 4670 * of pte involves: take ptl, clearing the pte so that 4671 * we don't have concurrent modification by hardware 4672 * followed by an update. 4673 */ 4674 if (unlikely(pte_none(*vmf->pte))) 4675 ret = VM_FAULT_SIGBUS; 4676 else 4677 ret = VM_FAULT_NOPAGE; 4678 4679 pte_unmap_unlock(vmf->pte, vmf->ptl); 4680 } 4681 } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 4682 ret = do_read_fault(vmf); 4683 else if (!(vma->vm_flags & VM_SHARED)) 4684 ret = do_cow_fault(vmf); 4685 else 4686 ret = do_shared_fault(vmf); 4687 4688 /* preallocated pagetable is unused: free it */ 4689 if (vmf->prealloc_pte) { 4690 pte_free(vm_mm, vmf->prealloc_pte); 4691 vmf->prealloc_pte = NULL; 4692 } 4693 return ret; 4694 } 4695 4696 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 4697 unsigned long addr, int page_nid, int *flags) 4698 { 4699 get_page(page); 4700 4701 count_vm_numa_event(NUMA_HINT_FAULTS); 4702 if (page_nid == numa_node_id()) { 4703 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 4704 *flags |= TNF_FAULT_LOCAL; 4705 } 4706 4707 return mpol_misplaced(page, vma, addr); 4708 } 4709 4710 static vm_fault_t do_numa_page(struct vm_fault *vmf) 4711 { 4712 struct vm_area_struct *vma = vmf->vma; 4713 struct page *page = NULL; 4714 int page_nid = NUMA_NO_NODE; 4715 int last_cpupid; 4716 int target_nid; 4717 pte_t pte, old_pte; 4718 bool was_writable = pte_savedwrite(vmf->orig_pte); 4719 int flags = 0; 4720 4721 /* 4722 * The "pte" at this point cannot be used safely without 4723 * validation through pte_unmap_same(). It's of NUMA type but 4724 * the pfn may be screwed if the read is non atomic. 4725 */ 4726 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd); 4727 spin_lock(vmf->ptl); 4728 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4729 pte_unmap_unlock(vmf->pte, vmf->ptl); 4730 goto out; 4731 } 4732 4733 /* Get the normal PTE */ 4734 old_pte = ptep_get(vmf->pte); 4735 pte = pte_modify(old_pte, vma->vm_page_prot); 4736 4737 page = vm_normal_page(vma, vmf->address, pte); 4738 if (!page || is_zone_device_page(page)) 4739 goto out_map; 4740 4741 /* TODO: handle PTE-mapped THP */ 4742 if (PageCompound(page)) 4743 goto out_map; 4744 4745 /* 4746 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 4747 * much anyway since they can be in shared cache state. This misses 4748 * the case where a mapping is writable but the process never writes 4749 * to it but pte_write gets cleared during protection updates and 4750 * pte_dirty has unpredictable behaviour between PTE scan updates, 4751 * background writeback, dirty balancing and application behaviour. 4752 */ 4753 if (!was_writable) 4754 flags |= TNF_NO_GROUP; 4755 4756 /* 4757 * Flag if the page is shared between multiple address spaces. This 4758 * is later used when determining whether to group tasks together 4759 */ 4760 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 4761 flags |= TNF_SHARED; 4762 4763 page_nid = page_to_nid(page); 4764 /* 4765 * For memory tiering mode, cpupid of slow memory page is used 4766 * to record page access time. So use default value. 4767 */ 4768 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && 4769 !node_is_toptier(page_nid)) 4770 last_cpupid = (-1 & LAST_CPUPID_MASK); 4771 else 4772 last_cpupid = page_cpupid_last(page); 4773 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid, 4774 &flags); 4775 if (target_nid == NUMA_NO_NODE) { 4776 put_page(page); 4777 goto out_map; 4778 } 4779 pte_unmap_unlock(vmf->pte, vmf->ptl); 4780 4781 /* Migrate to the requested node */ 4782 if (migrate_misplaced_page(page, vma, target_nid)) { 4783 page_nid = target_nid; 4784 flags |= TNF_MIGRATED; 4785 } else { 4786 flags |= TNF_MIGRATE_FAIL; 4787 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4788 spin_lock(vmf->ptl); 4789 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 4790 pte_unmap_unlock(vmf->pte, vmf->ptl); 4791 goto out; 4792 } 4793 goto out_map; 4794 } 4795 4796 out: 4797 if (page_nid != NUMA_NO_NODE) 4798 task_numa_fault(last_cpupid, page_nid, 1, flags); 4799 return 0; 4800 out_map: 4801 /* 4802 * Make it present again, depending on how arch implements 4803 * non-accessible ptes, some can allow access by kernel mode. 4804 */ 4805 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 4806 pte = pte_modify(old_pte, vma->vm_page_prot); 4807 pte = pte_mkyoung(pte); 4808 if (was_writable) 4809 pte = pte_mkwrite(pte); 4810 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 4811 update_mmu_cache(vma, vmf->address, vmf->pte); 4812 pte_unmap_unlock(vmf->pte, vmf->ptl); 4813 goto out; 4814 } 4815 4816 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 4817 { 4818 if (vma_is_anonymous(vmf->vma)) 4819 return do_huge_pmd_anonymous_page(vmf); 4820 if (vmf->vma->vm_ops->huge_fault) 4821 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4822 return VM_FAULT_FALLBACK; 4823 } 4824 4825 /* `inline' is required to avoid gcc 4.1.2 build error */ 4826 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf) 4827 { 4828 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 4829 4830 if (vma_is_anonymous(vmf->vma)) { 4831 if (likely(!unshare) && 4832 userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd)) 4833 return handle_userfault(vmf, VM_UFFD_WP); 4834 return do_huge_pmd_wp_page(vmf); 4835 } 4836 if (vmf->vma->vm_ops->huge_fault) { 4837 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 4838 4839 if (!(ret & VM_FAULT_FALLBACK)) 4840 return ret; 4841 } 4842 4843 /* COW or write-notify handled on pte level: split pmd. */ 4844 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL); 4845 4846 return VM_FAULT_FALLBACK; 4847 } 4848 4849 static vm_fault_t create_huge_pud(struct vm_fault *vmf) 4850 { 4851 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 4852 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 4853 /* No support for anonymous transparent PUD pages yet */ 4854 if (vma_is_anonymous(vmf->vma)) 4855 return VM_FAULT_FALLBACK; 4856 if (vmf->vma->vm_ops->huge_fault) 4857 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4858 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 4859 return VM_FAULT_FALLBACK; 4860 } 4861 4862 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 4863 { 4864 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 4865 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 4866 /* No support for anonymous transparent PUD pages yet */ 4867 if (vma_is_anonymous(vmf->vma)) 4868 goto split; 4869 if (vmf->vma->vm_ops->huge_fault) { 4870 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 4871 4872 if (!(ret & VM_FAULT_FALLBACK)) 4873 return ret; 4874 } 4875 split: 4876 /* COW or write-notify not handled on PUD level: split pud.*/ 4877 __split_huge_pud(vmf->vma, vmf->pud, vmf->address); 4878 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 4879 return VM_FAULT_FALLBACK; 4880 } 4881 4882 /* 4883 * These routines also need to handle stuff like marking pages dirty 4884 * and/or accessed for architectures that don't do it in hardware (most 4885 * RISC architectures). The early dirtying is also good on the i386. 4886 * 4887 * There is also a hook called "update_mmu_cache()" that architectures 4888 * with external mmu caches can use to update those (ie the Sparc or 4889 * PowerPC hashed page tables that act as extended TLBs). 4890 * 4891 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow 4892 * concurrent faults). 4893 * 4894 * The mmap_lock may have been released depending on flags and our return value. 4895 * See filemap_fault() and __folio_lock_or_retry(). 4896 */ 4897 static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 4898 { 4899 pte_t entry; 4900 4901 if (unlikely(pmd_none(*vmf->pmd))) { 4902 /* 4903 * Leave __pte_alloc() until later: because vm_ops->fault may 4904 * want to allocate huge page, and if we expose page table 4905 * for an instant, it will be difficult to retract from 4906 * concurrent faults and from rmap lookups. 4907 */ 4908 vmf->pte = NULL; 4909 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID; 4910 } else { 4911 /* 4912 * If a huge pmd materialized under us just retry later. Use 4913 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead 4914 * of pmd_trans_huge() to ensure the pmd didn't become 4915 * pmd_trans_huge under us and then back to pmd_none, as a 4916 * result of MADV_DONTNEED running immediately after a huge pmd 4917 * fault in a different thread of this mm, in turn leading to a 4918 * misleading pmd_trans_huge() retval. All we have to ensure is 4919 * that it is a regular pmd that we can walk with 4920 * pte_offset_map() and we can do that through an atomic read 4921 * in C, which is what pmd_trans_unstable() provides. 4922 */ 4923 if (pmd_devmap_trans_unstable(vmf->pmd)) 4924 return 0; 4925 /* 4926 * A regular pmd is established and it can't morph into a huge 4927 * pmd from under us anymore at this point because we hold the 4928 * mmap_lock read mode and khugepaged takes it in write mode. 4929 * So now it's safe to run pte_offset_map(). 4930 */ 4931 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 4932 vmf->orig_pte = *vmf->pte; 4933 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID; 4934 4935 /* 4936 * some architectures can have larger ptes than wordsize, 4937 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and 4938 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic 4939 * accesses. The code below just needs a consistent view 4940 * for the ifs and we later double check anyway with the 4941 * ptl lock held. So here a barrier will do. 4942 */ 4943 barrier(); 4944 if (pte_none(vmf->orig_pte)) { 4945 pte_unmap(vmf->pte); 4946 vmf->pte = NULL; 4947 } 4948 } 4949 4950 if (!vmf->pte) { 4951 if (vma_is_anonymous(vmf->vma)) 4952 return do_anonymous_page(vmf); 4953 else 4954 return do_fault(vmf); 4955 } 4956 4957 if (!pte_present(vmf->orig_pte)) 4958 return do_swap_page(vmf); 4959 4960 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 4961 return do_numa_page(vmf); 4962 4963 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 4964 spin_lock(vmf->ptl); 4965 entry = vmf->orig_pte; 4966 if (unlikely(!pte_same(*vmf->pte, entry))) { 4967 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 4968 goto unlock; 4969 } 4970 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 4971 if (!pte_write(entry)) 4972 return do_wp_page(vmf); 4973 else if (likely(vmf->flags & FAULT_FLAG_WRITE)) 4974 entry = pte_mkdirty(entry); 4975 } 4976 entry = pte_mkyoung(entry); 4977 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 4978 vmf->flags & FAULT_FLAG_WRITE)) { 4979 update_mmu_cache(vmf->vma, vmf->address, vmf->pte); 4980 } else { 4981 /* Skip spurious TLB flush for retried page fault */ 4982 if (vmf->flags & FAULT_FLAG_TRIED) 4983 goto unlock; 4984 /* 4985 * This is needed only for protection faults but the arch code 4986 * is not yet telling us if this is a protection fault or not. 4987 * This still avoids useless tlb flushes for .text page faults 4988 * with threads. 4989 */ 4990 if (vmf->flags & FAULT_FLAG_WRITE) 4991 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address); 4992 } 4993 unlock: 4994 pte_unmap_unlock(vmf->pte, vmf->ptl); 4995 return 0; 4996 } 4997 4998 /* 4999 * By the time we get here, we already hold the mm semaphore 5000 * 5001 * The mmap_lock may have been released depending on flags and our 5002 * return value. See filemap_fault() and __folio_lock_or_retry(). 5003 */ 5004 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 5005 unsigned long address, unsigned int flags) 5006 { 5007 struct vm_fault vmf = { 5008 .vma = vma, 5009 .address = address & PAGE_MASK, 5010 .real_address = address, 5011 .flags = flags, 5012 .pgoff = linear_page_index(vma, address), 5013 .gfp_mask = __get_fault_gfp_mask(vma), 5014 }; 5015 struct mm_struct *mm = vma->vm_mm; 5016 unsigned long vm_flags = vma->vm_flags; 5017 pgd_t *pgd; 5018 p4d_t *p4d; 5019 vm_fault_t ret; 5020 5021 pgd = pgd_offset(mm, address); 5022 p4d = p4d_alloc(mm, pgd, address); 5023 if (!p4d) 5024 return VM_FAULT_OOM; 5025 5026 vmf.pud = pud_alloc(mm, p4d, address); 5027 if (!vmf.pud) 5028 return VM_FAULT_OOM; 5029 retry_pud: 5030 if (pud_none(*vmf.pud) && 5031 hugepage_vma_check(vma, vm_flags, false, true, true)) { 5032 ret = create_huge_pud(&vmf); 5033 if (!(ret & VM_FAULT_FALLBACK)) 5034 return ret; 5035 } else { 5036 pud_t orig_pud = *vmf.pud; 5037 5038 barrier(); 5039 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 5040 5041 /* 5042 * TODO once we support anonymous PUDs: NUMA case and 5043 * FAULT_FLAG_UNSHARE handling. 5044 */ 5045 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) { 5046 ret = wp_huge_pud(&vmf, orig_pud); 5047 if (!(ret & VM_FAULT_FALLBACK)) 5048 return ret; 5049 } else { 5050 huge_pud_set_accessed(&vmf, orig_pud); 5051 return 0; 5052 } 5053 } 5054 } 5055 5056 vmf.pmd = pmd_alloc(mm, vmf.pud, address); 5057 if (!vmf.pmd) 5058 return VM_FAULT_OOM; 5059 5060 /* Huge pud page fault raced with pmd_alloc? */ 5061 if (pud_trans_unstable(vmf.pud)) 5062 goto retry_pud; 5063 5064 if (pmd_none(*vmf.pmd) && 5065 hugepage_vma_check(vma, vm_flags, false, true, true)) { 5066 ret = create_huge_pmd(&vmf); 5067 if (!(ret & VM_FAULT_FALLBACK)) 5068 return ret; 5069 } else { 5070 vmf.orig_pmd = *vmf.pmd; 5071 5072 barrier(); 5073 if (unlikely(is_swap_pmd(vmf.orig_pmd))) { 5074 VM_BUG_ON(thp_migration_supported() && 5075 !is_pmd_migration_entry(vmf.orig_pmd)); 5076 if (is_pmd_migration_entry(vmf.orig_pmd)) 5077 pmd_migration_entry_wait(mm, vmf.pmd); 5078 return 0; 5079 } 5080 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) { 5081 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma)) 5082 return do_huge_pmd_numa_page(&vmf); 5083 5084 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 5085 !pmd_write(vmf.orig_pmd)) { 5086 ret = wp_huge_pmd(&vmf); 5087 if (!(ret & VM_FAULT_FALLBACK)) 5088 return ret; 5089 } else { 5090 huge_pmd_set_accessed(&vmf); 5091 return 0; 5092 } 5093 } 5094 } 5095 5096 return handle_pte_fault(&vmf); 5097 } 5098 5099 /** 5100 * mm_account_fault - Do page fault accounting 5101 * 5102 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting 5103 * of perf event counters, but we'll still do the per-task accounting to 5104 * the task who triggered this page fault. 5105 * @address: the faulted address. 5106 * @flags: the fault flags. 5107 * @ret: the fault retcode. 5108 * 5109 * This will take care of most of the page fault accounting. Meanwhile, it 5110 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter 5111 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should 5112 * still be in per-arch page fault handlers at the entry of page fault. 5113 */ 5114 static inline void mm_account_fault(struct pt_regs *regs, 5115 unsigned long address, unsigned int flags, 5116 vm_fault_t ret) 5117 { 5118 bool major; 5119 5120 /* 5121 * We don't do accounting for some specific faults: 5122 * 5123 * - Unsuccessful faults (e.g. when the address wasn't valid). That 5124 * includes arch_vma_access_permitted() failing before reaching here. 5125 * So this is not a "this many hardware page faults" counter. We 5126 * should use the hw profiling for that. 5127 * 5128 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted 5129 * once they're completed. 5130 */ 5131 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY)) 5132 return; 5133 5134 /* 5135 * We define the fault as a major fault when the final successful fault 5136 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't 5137 * handle it immediately previously). 5138 */ 5139 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); 5140 5141 if (major) 5142 current->maj_flt++; 5143 else 5144 current->min_flt++; 5145 5146 /* 5147 * If the fault is done for GUP, regs will be NULL. We only do the 5148 * accounting for the per thread fault counters who triggered the 5149 * fault, and we skip the perf event updates. 5150 */ 5151 if (!regs) 5152 return; 5153 5154 if (major) 5155 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 5156 else 5157 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 5158 } 5159 5160 #ifdef CONFIG_LRU_GEN 5161 static void lru_gen_enter_fault(struct vm_area_struct *vma) 5162 { 5163 /* the LRU algorithm doesn't apply to sequential or random reads */ 5164 current->in_lru_fault = !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ)); 5165 } 5166 5167 static void lru_gen_exit_fault(void) 5168 { 5169 current->in_lru_fault = false; 5170 } 5171 #else 5172 static void lru_gen_enter_fault(struct vm_area_struct *vma) 5173 { 5174 } 5175 5176 static void lru_gen_exit_fault(void) 5177 { 5178 } 5179 #endif /* CONFIG_LRU_GEN */ 5180 5181 /* 5182 * By the time we get here, we already hold the mm semaphore 5183 * 5184 * The mmap_lock may have been released depending on flags and our 5185 * return value. See filemap_fault() and __folio_lock_or_retry(). 5186 */ 5187 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 5188 unsigned int flags, struct pt_regs *regs) 5189 { 5190 vm_fault_t ret; 5191 5192 __set_current_state(TASK_RUNNING); 5193 5194 count_vm_event(PGFAULT); 5195 count_memcg_event_mm(vma->vm_mm, PGFAULT); 5196 5197 /* do counter updates before entering really critical section. */ 5198 check_sync_rss_stat(current); 5199 5200 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 5201 flags & FAULT_FLAG_INSTRUCTION, 5202 flags & FAULT_FLAG_REMOTE)) 5203 return VM_FAULT_SIGSEGV; 5204 5205 /* 5206 * Enable the memcg OOM handling for faults triggered in user 5207 * space. Kernel faults are handled more gracefully. 5208 */ 5209 if (flags & FAULT_FLAG_USER) 5210 mem_cgroup_enter_user_fault(); 5211 5212 lru_gen_enter_fault(vma); 5213 5214 if (unlikely(is_vm_hugetlb_page(vma))) 5215 ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 5216 else 5217 ret = __handle_mm_fault(vma, address, flags); 5218 5219 lru_gen_exit_fault(); 5220 5221 if (flags & FAULT_FLAG_USER) { 5222 mem_cgroup_exit_user_fault(); 5223 /* 5224 * The task may have entered a memcg OOM situation but 5225 * if the allocation error was handled gracefully (no 5226 * VM_FAULT_OOM), there is no need to kill anything. 5227 * Just clean up the OOM state peacefully. 5228 */ 5229 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 5230 mem_cgroup_oom_synchronize(false); 5231 } 5232 5233 mm_account_fault(regs, address, flags, ret); 5234 5235 return ret; 5236 } 5237 EXPORT_SYMBOL_GPL(handle_mm_fault); 5238 5239 #ifndef __PAGETABLE_P4D_FOLDED 5240 /* 5241 * Allocate p4d page table. 5242 * We've already handled the fast-path in-line. 5243 */ 5244 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 5245 { 5246 p4d_t *new = p4d_alloc_one(mm, address); 5247 if (!new) 5248 return -ENOMEM; 5249 5250 spin_lock(&mm->page_table_lock); 5251 if (pgd_present(*pgd)) { /* Another has populated it */ 5252 p4d_free(mm, new); 5253 } else { 5254 smp_wmb(); /* See comment in pmd_install() */ 5255 pgd_populate(mm, pgd, new); 5256 } 5257 spin_unlock(&mm->page_table_lock); 5258 return 0; 5259 } 5260 #endif /* __PAGETABLE_P4D_FOLDED */ 5261 5262 #ifndef __PAGETABLE_PUD_FOLDED 5263 /* 5264 * Allocate page upper directory. 5265 * We've already handled the fast-path in-line. 5266 */ 5267 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 5268 { 5269 pud_t *new = pud_alloc_one(mm, address); 5270 if (!new) 5271 return -ENOMEM; 5272 5273 spin_lock(&mm->page_table_lock); 5274 if (!p4d_present(*p4d)) { 5275 mm_inc_nr_puds(mm); 5276 smp_wmb(); /* See comment in pmd_install() */ 5277 p4d_populate(mm, p4d, new); 5278 } else /* Another has populated it */ 5279 pud_free(mm, new); 5280 spin_unlock(&mm->page_table_lock); 5281 return 0; 5282 } 5283 #endif /* __PAGETABLE_PUD_FOLDED */ 5284 5285 #ifndef __PAGETABLE_PMD_FOLDED 5286 /* 5287 * Allocate page middle directory. 5288 * We've already handled the fast-path in-line. 5289 */ 5290 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 5291 { 5292 spinlock_t *ptl; 5293 pmd_t *new = pmd_alloc_one(mm, address); 5294 if (!new) 5295 return -ENOMEM; 5296 5297 ptl = pud_lock(mm, pud); 5298 if (!pud_present(*pud)) { 5299 mm_inc_nr_pmds(mm); 5300 smp_wmb(); /* See comment in pmd_install() */ 5301 pud_populate(mm, pud, new); 5302 } else { /* Another has populated it */ 5303 pmd_free(mm, new); 5304 } 5305 spin_unlock(ptl); 5306 return 0; 5307 } 5308 #endif /* __PAGETABLE_PMD_FOLDED */ 5309 5310 /** 5311 * follow_pte - look up PTE at a user virtual address 5312 * @mm: the mm_struct of the target address space 5313 * @address: user virtual address 5314 * @ptepp: location to store found PTE 5315 * @ptlp: location to store the lock for the PTE 5316 * 5317 * On a successful return, the pointer to the PTE is stored in @ptepp; 5318 * the corresponding lock is taken and its location is stored in @ptlp. 5319 * The contents of the PTE are only stable until @ptlp is released; 5320 * any further use, if any, must be protected against invalidation 5321 * with MMU notifiers. 5322 * 5323 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore 5324 * should be taken for read. 5325 * 5326 * KVM uses this function. While it is arguably less bad than ``follow_pfn``, 5327 * it is not a good general-purpose API. 5328 * 5329 * Return: zero on success, -ve otherwise. 5330 */ 5331 int follow_pte(struct mm_struct *mm, unsigned long address, 5332 pte_t **ptepp, spinlock_t **ptlp) 5333 { 5334 pgd_t *pgd; 5335 p4d_t *p4d; 5336 pud_t *pud; 5337 pmd_t *pmd; 5338 pte_t *ptep; 5339 5340 pgd = pgd_offset(mm, address); 5341 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 5342 goto out; 5343 5344 p4d = p4d_offset(pgd, address); 5345 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 5346 goto out; 5347 5348 pud = pud_offset(p4d, address); 5349 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 5350 goto out; 5351 5352 pmd = pmd_offset(pud, address); 5353 VM_BUG_ON(pmd_trans_huge(*pmd)); 5354 5355 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 5356 goto out; 5357 5358 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 5359 if (!pte_present(*ptep)) 5360 goto unlock; 5361 *ptepp = ptep; 5362 return 0; 5363 unlock: 5364 pte_unmap_unlock(ptep, *ptlp); 5365 out: 5366 return -EINVAL; 5367 } 5368 EXPORT_SYMBOL_GPL(follow_pte); 5369 5370 /** 5371 * follow_pfn - look up PFN at a user virtual address 5372 * @vma: memory mapping 5373 * @address: user virtual address 5374 * @pfn: location to store found PFN 5375 * 5376 * Only IO mappings and raw PFN mappings are allowed. 5377 * 5378 * This function does not allow the caller to read the permissions 5379 * of the PTE. Do not use it. 5380 * 5381 * Return: zero and the pfn at @pfn on success, -ve otherwise. 5382 */ 5383 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 5384 unsigned long *pfn) 5385 { 5386 int ret = -EINVAL; 5387 spinlock_t *ptl; 5388 pte_t *ptep; 5389 5390 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5391 return ret; 5392 5393 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 5394 if (ret) 5395 return ret; 5396 *pfn = pte_pfn(*ptep); 5397 pte_unmap_unlock(ptep, ptl); 5398 return 0; 5399 } 5400 EXPORT_SYMBOL(follow_pfn); 5401 5402 #ifdef CONFIG_HAVE_IOREMAP_PROT 5403 int follow_phys(struct vm_area_struct *vma, 5404 unsigned long address, unsigned int flags, 5405 unsigned long *prot, resource_size_t *phys) 5406 { 5407 int ret = -EINVAL; 5408 pte_t *ptep, pte; 5409 spinlock_t *ptl; 5410 5411 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5412 goto out; 5413 5414 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 5415 goto out; 5416 pte = *ptep; 5417 5418 if ((flags & FOLL_WRITE) && !pte_write(pte)) 5419 goto unlock; 5420 5421 *prot = pgprot_val(pte_pgprot(pte)); 5422 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 5423 5424 ret = 0; 5425 unlock: 5426 pte_unmap_unlock(ptep, ptl); 5427 out: 5428 return ret; 5429 } 5430 5431 /** 5432 * generic_access_phys - generic implementation for iomem mmap access 5433 * @vma: the vma to access 5434 * @addr: userspace address, not relative offset within @vma 5435 * @buf: buffer to read/write 5436 * @len: length of transfer 5437 * @write: set to FOLL_WRITE when writing, otherwise reading 5438 * 5439 * This is a generic implementation for &vm_operations_struct.access for an 5440 * iomem mapping. This callback is used by access_process_vm() when the @vma is 5441 * not page based. 5442 */ 5443 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 5444 void *buf, int len, int write) 5445 { 5446 resource_size_t phys_addr; 5447 unsigned long prot = 0; 5448 void __iomem *maddr; 5449 pte_t *ptep, pte; 5450 spinlock_t *ptl; 5451 int offset = offset_in_page(addr); 5452 int ret = -EINVAL; 5453 5454 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5455 return -EINVAL; 5456 5457 retry: 5458 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 5459 return -EINVAL; 5460 pte = *ptep; 5461 pte_unmap_unlock(ptep, ptl); 5462 5463 prot = pgprot_val(pte_pgprot(pte)); 5464 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 5465 5466 if ((write & FOLL_WRITE) && !pte_write(pte)) 5467 return -EINVAL; 5468 5469 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 5470 if (!maddr) 5471 return -ENOMEM; 5472 5473 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 5474 goto out_unmap; 5475 5476 if (!pte_same(pte, *ptep)) { 5477 pte_unmap_unlock(ptep, ptl); 5478 iounmap(maddr); 5479 5480 goto retry; 5481 } 5482 5483 if (write) 5484 memcpy_toio(maddr + offset, buf, len); 5485 else 5486 memcpy_fromio(buf, maddr + offset, len); 5487 ret = len; 5488 pte_unmap_unlock(ptep, ptl); 5489 out_unmap: 5490 iounmap(maddr); 5491 5492 return ret; 5493 } 5494 EXPORT_SYMBOL_GPL(generic_access_phys); 5495 #endif 5496 5497 /* 5498 * Access another process' address space as given in mm. 5499 */ 5500 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf, 5501 int len, unsigned int gup_flags) 5502 { 5503 struct vm_area_struct *vma; 5504 void *old_buf = buf; 5505 int write = gup_flags & FOLL_WRITE; 5506 5507 if (mmap_read_lock_killable(mm)) 5508 return 0; 5509 5510 /* ignore errors, just check how much was successfully transferred */ 5511 while (len) { 5512 int bytes, ret, offset; 5513 void *maddr; 5514 struct page *page = NULL; 5515 5516 ret = get_user_pages_remote(mm, addr, 1, 5517 gup_flags, &page, &vma, NULL); 5518 if (ret <= 0) { 5519 #ifndef CONFIG_HAVE_IOREMAP_PROT 5520 break; 5521 #else 5522 /* 5523 * Check if this is a VM_IO | VM_PFNMAP VMA, which 5524 * we can access using slightly different code. 5525 */ 5526 vma = vma_lookup(mm, addr); 5527 if (!vma) 5528 break; 5529 if (vma->vm_ops && vma->vm_ops->access) 5530 ret = vma->vm_ops->access(vma, addr, buf, 5531 len, write); 5532 if (ret <= 0) 5533 break; 5534 bytes = ret; 5535 #endif 5536 } else { 5537 bytes = len; 5538 offset = addr & (PAGE_SIZE-1); 5539 if (bytes > PAGE_SIZE-offset) 5540 bytes = PAGE_SIZE-offset; 5541 5542 maddr = kmap(page); 5543 if (write) { 5544 copy_to_user_page(vma, page, addr, 5545 maddr + offset, buf, bytes); 5546 set_page_dirty_lock(page); 5547 } else { 5548 copy_from_user_page(vma, page, addr, 5549 buf, maddr + offset, bytes); 5550 } 5551 kunmap(page); 5552 put_page(page); 5553 } 5554 len -= bytes; 5555 buf += bytes; 5556 addr += bytes; 5557 } 5558 mmap_read_unlock(mm); 5559 5560 return buf - old_buf; 5561 } 5562 5563 /** 5564 * access_remote_vm - access another process' address space 5565 * @mm: the mm_struct of the target address space 5566 * @addr: start address to access 5567 * @buf: source or destination buffer 5568 * @len: number of bytes to transfer 5569 * @gup_flags: flags modifying lookup behaviour 5570 * 5571 * The caller must hold a reference on @mm. 5572 * 5573 * Return: number of bytes copied from source to destination. 5574 */ 5575 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 5576 void *buf, int len, unsigned int gup_flags) 5577 { 5578 return __access_remote_vm(mm, addr, buf, len, gup_flags); 5579 } 5580 5581 /* 5582 * Access another process' address space. 5583 * Source/target buffer must be kernel space, 5584 * Do not walk the page table directly, use get_user_pages 5585 */ 5586 int access_process_vm(struct task_struct *tsk, unsigned long addr, 5587 void *buf, int len, unsigned int gup_flags) 5588 { 5589 struct mm_struct *mm; 5590 int ret; 5591 5592 mm = get_task_mm(tsk); 5593 if (!mm) 5594 return 0; 5595 5596 ret = __access_remote_vm(mm, addr, buf, len, gup_flags); 5597 5598 mmput(mm); 5599 5600 return ret; 5601 } 5602 EXPORT_SYMBOL_GPL(access_process_vm); 5603 5604 /* 5605 * Print the name of a VMA. 5606 */ 5607 void print_vma_addr(char *prefix, unsigned long ip) 5608 { 5609 struct mm_struct *mm = current->mm; 5610 struct vm_area_struct *vma; 5611 5612 /* 5613 * we might be running from an atomic context so we cannot sleep 5614 */ 5615 if (!mmap_read_trylock(mm)) 5616 return; 5617 5618 vma = find_vma(mm, ip); 5619 if (vma && vma->vm_file) { 5620 struct file *f = vma->vm_file; 5621 char *buf = (char *)__get_free_page(GFP_NOWAIT); 5622 if (buf) { 5623 char *p; 5624 5625 p = file_path(f, buf, PAGE_SIZE); 5626 if (IS_ERR(p)) 5627 p = "?"; 5628 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 5629 vma->vm_start, 5630 vma->vm_end - vma->vm_start); 5631 free_page((unsigned long)buf); 5632 } 5633 } 5634 mmap_read_unlock(mm); 5635 } 5636 5637 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5638 void __might_fault(const char *file, int line) 5639 { 5640 if (pagefault_disabled()) 5641 return; 5642 __might_sleep(file, line); 5643 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 5644 if (current->mm) 5645 might_lock_read(¤t->mm->mmap_lock); 5646 #endif 5647 } 5648 EXPORT_SYMBOL(__might_fault); 5649 #endif 5650 5651 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 5652 /* 5653 * Process all subpages of the specified huge page with the specified 5654 * operation. The target subpage will be processed last to keep its 5655 * cache lines hot. 5656 */ 5657 static inline void process_huge_page( 5658 unsigned long addr_hint, unsigned int pages_per_huge_page, 5659 void (*process_subpage)(unsigned long addr, int idx, void *arg), 5660 void *arg) 5661 { 5662 int i, n, base, l; 5663 unsigned long addr = addr_hint & 5664 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5665 5666 /* Process target subpage last to keep its cache lines hot */ 5667 might_sleep(); 5668 n = (addr_hint - addr) / PAGE_SIZE; 5669 if (2 * n <= pages_per_huge_page) { 5670 /* If target subpage in first half of huge page */ 5671 base = 0; 5672 l = n; 5673 /* Process subpages at the end of huge page */ 5674 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 5675 cond_resched(); 5676 process_subpage(addr + i * PAGE_SIZE, i, arg); 5677 } 5678 } else { 5679 /* If target subpage in second half of huge page */ 5680 base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 5681 l = pages_per_huge_page - n; 5682 /* Process subpages at the begin of huge page */ 5683 for (i = 0; i < base; i++) { 5684 cond_resched(); 5685 process_subpage(addr + i * PAGE_SIZE, i, arg); 5686 } 5687 } 5688 /* 5689 * Process remaining subpages in left-right-left-right pattern 5690 * towards the target subpage 5691 */ 5692 for (i = 0; i < l; i++) { 5693 int left_idx = base + i; 5694 int right_idx = base + 2 * l - 1 - i; 5695 5696 cond_resched(); 5697 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 5698 cond_resched(); 5699 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 5700 } 5701 } 5702 5703 static void clear_gigantic_page(struct page *page, 5704 unsigned long addr, 5705 unsigned int pages_per_huge_page) 5706 { 5707 int i; 5708 struct page *p; 5709 5710 might_sleep(); 5711 for (i = 0; i < pages_per_huge_page; i++) { 5712 p = nth_page(page, i); 5713 cond_resched(); 5714 clear_user_highpage(p, addr + i * PAGE_SIZE); 5715 } 5716 } 5717 5718 static void clear_subpage(unsigned long addr, int idx, void *arg) 5719 { 5720 struct page *page = arg; 5721 5722 clear_user_highpage(page + idx, addr); 5723 } 5724 5725 void clear_huge_page(struct page *page, 5726 unsigned long addr_hint, unsigned int pages_per_huge_page) 5727 { 5728 unsigned long addr = addr_hint & 5729 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5730 5731 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5732 clear_gigantic_page(page, addr, pages_per_huge_page); 5733 return; 5734 } 5735 5736 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 5737 } 5738 5739 static void copy_user_gigantic_page(struct page *dst, struct page *src, 5740 unsigned long addr, 5741 struct vm_area_struct *vma, 5742 unsigned int pages_per_huge_page) 5743 { 5744 int i; 5745 struct page *dst_base = dst; 5746 struct page *src_base = src; 5747 5748 for (i = 0; i < pages_per_huge_page; i++) { 5749 dst = nth_page(dst_base, i); 5750 src = nth_page(src_base, i); 5751 5752 cond_resched(); 5753 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 5754 } 5755 } 5756 5757 struct copy_subpage_arg { 5758 struct page *dst; 5759 struct page *src; 5760 struct vm_area_struct *vma; 5761 }; 5762 5763 static void copy_subpage(unsigned long addr, int idx, void *arg) 5764 { 5765 struct copy_subpage_arg *copy_arg = arg; 5766 5767 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx, 5768 addr, copy_arg->vma); 5769 } 5770 5771 void copy_user_huge_page(struct page *dst, struct page *src, 5772 unsigned long addr_hint, struct vm_area_struct *vma, 5773 unsigned int pages_per_huge_page) 5774 { 5775 unsigned long addr = addr_hint & 5776 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 5777 struct copy_subpage_arg arg = { 5778 .dst = dst, 5779 .src = src, 5780 .vma = vma, 5781 }; 5782 5783 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 5784 copy_user_gigantic_page(dst, src, addr, vma, 5785 pages_per_huge_page); 5786 return; 5787 } 5788 5789 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 5790 } 5791 5792 long copy_huge_page_from_user(struct page *dst_page, 5793 const void __user *usr_src, 5794 unsigned int pages_per_huge_page, 5795 bool allow_pagefault) 5796 { 5797 void *page_kaddr; 5798 unsigned long i, rc = 0; 5799 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE; 5800 struct page *subpage; 5801 5802 for (i = 0; i < pages_per_huge_page; i++) { 5803 subpage = nth_page(dst_page, i); 5804 if (allow_pagefault) 5805 page_kaddr = kmap(subpage); 5806 else 5807 page_kaddr = kmap_atomic(subpage); 5808 rc = copy_from_user(page_kaddr, 5809 usr_src + i * PAGE_SIZE, PAGE_SIZE); 5810 if (allow_pagefault) 5811 kunmap(subpage); 5812 else 5813 kunmap_atomic(page_kaddr); 5814 5815 ret_val -= (PAGE_SIZE - rc); 5816 if (rc) 5817 break; 5818 5819 flush_dcache_page(subpage); 5820 5821 cond_resched(); 5822 } 5823 return ret_val; 5824 } 5825 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 5826 5827 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 5828 5829 static struct kmem_cache *page_ptl_cachep; 5830 5831 void __init ptlock_cache_init(void) 5832 { 5833 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 5834 SLAB_PANIC, NULL); 5835 } 5836 5837 bool ptlock_alloc(struct page *page) 5838 { 5839 spinlock_t *ptl; 5840 5841 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 5842 if (!ptl) 5843 return false; 5844 page->ptl = ptl; 5845 return true; 5846 } 5847 5848 void ptlock_free(struct page *page) 5849 { 5850 kmem_cache_free(page_ptl_cachep, page->ptl); 5851 } 5852 #endif 5853