1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/memory.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 */ 7 8 /* 9 * demand-loading started 01.12.91 - seems it is high on the list of 10 * things wanted, and it should be easy to implement. - Linus 11 */ 12 13 /* 14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared 15 * pages started 02.12.91, seems to work. - Linus. 16 * 17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it 18 * would have taken more than the 6M I have free, but it worked well as 19 * far as I could see. 20 * 21 * Also corrected some "invalidate()"s - I wasn't doing enough of them. 22 */ 23 24 /* 25 * Real VM (paging to/from disk) started 18.12.91. Much more work and 26 * thought has to go into this. Oh, well.. 27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why. 28 * Found it. Everything seems to work now. 29 * 20.12.91 - Ok, making the swap-device changeable like the root. 30 */ 31 32 /* 33 * 05.04.94 - Multi-page memory management added for v1.1. 34 * Idea by Alex Bligh (alex@cconcepts.co.uk) 35 * 36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG 37 * (Gerhard.Wichert@pdb.siemens.de) 38 * 39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen) 40 */ 41 42 #include <linux/kernel_stat.h> 43 #include <linux/mm.h> 44 #include <linux/sched/mm.h> 45 #include <linux/sched/coredump.h> 46 #include <linux/sched/numa_balancing.h> 47 #include <linux/sched/task.h> 48 #include <linux/hugetlb.h> 49 #include <linux/mman.h> 50 #include <linux/swap.h> 51 #include <linux/highmem.h> 52 #include <linux/pagemap.h> 53 #include <linux/memremap.h> 54 #include <linux/ksm.h> 55 #include <linux/rmap.h> 56 #include <linux/export.h> 57 #include <linux/delayacct.h> 58 #include <linux/init.h> 59 #include <linux/pfn_t.h> 60 #include <linux/writeback.h> 61 #include <linux/memcontrol.h> 62 #include <linux/mmu_notifier.h> 63 #include <linux/swapops.h> 64 #include <linux/elf.h> 65 #include <linux/gfp.h> 66 #include <linux/migrate.h> 67 #include <linux/string.h> 68 #include <linux/dma-debug.h> 69 #include <linux/debugfs.h> 70 #include <linux/userfaultfd_k.h> 71 #include <linux/dax.h> 72 #include <linux/oom.h> 73 #include <linux/numa.h> 74 75 #include <asm/io.h> 76 #include <asm/mmu_context.h> 77 #include <asm/pgalloc.h> 78 #include <linux/uaccess.h> 79 #include <asm/tlb.h> 80 #include <asm/tlbflush.h> 81 #include <asm/pgtable.h> 82 83 #include "internal.h" 84 85 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) 86 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 87 #endif 88 89 #ifndef CONFIG_NEED_MULTIPLE_NODES 90 /* use the per-pgdat data instead for discontigmem - mbligh */ 91 unsigned long max_mapnr; 92 EXPORT_SYMBOL(max_mapnr); 93 94 struct page *mem_map; 95 EXPORT_SYMBOL(mem_map); 96 #endif 97 98 /* 99 * A number of key systems in x86 including ioremap() rely on the assumption 100 * that high_memory defines the upper bound on direct map memory, then end 101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and 102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL 103 * and ZONE_HIGHMEM. 104 */ 105 void *high_memory; 106 EXPORT_SYMBOL(high_memory); 107 108 /* 109 * Randomize the address space (stacks, mmaps, brk, etc.). 110 * 111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 112 * as ancient (libc5 based) binaries can segfault. ) 113 */ 114 int randomize_va_space __read_mostly = 115 #ifdef CONFIG_COMPAT_BRK 116 1; 117 #else 118 2; 119 #endif 120 121 static int __init disable_randmaps(char *s) 122 { 123 randomize_va_space = 0; 124 return 1; 125 } 126 __setup("norandmaps", disable_randmaps); 127 128 unsigned long zero_pfn __read_mostly; 129 EXPORT_SYMBOL(zero_pfn); 130 131 unsigned long highest_memmap_pfn __read_mostly; 132 133 /* 134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 135 */ 136 static int __init init_zero_pfn(void) 137 { 138 zero_pfn = page_to_pfn(ZERO_PAGE(0)); 139 return 0; 140 } 141 core_initcall(init_zero_pfn); 142 143 144 #if defined(SPLIT_RSS_COUNTING) 145 146 void sync_mm_rss(struct mm_struct *mm) 147 { 148 int i; 149 150 for (i = 0; i < NR_MM_COUNTERS; i++) { 151 if (current->rss_stat.count[i]) { 152 add_mm_counter(mm, i, current->rss_stat.count[i]); 153 current->rss_stat.count[i] = 0; 154 } 155 } 156 current->rss_stat.events = 0; 157 } 158 159 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) 160 { 161 struct task_struct *task = current; 162 163 if (likely(task->mm == mm)) 164 task->rss_stat.count[member] += val; 165 else 166 add_mm_counter(mm, member, val); 167 } 168 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) 169 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) 170 171 /* sync counter once per 64 page faults */ 172 #define TASK_RSS_EVENTS_THRESH (64) 173 static void check_sync_rss_stat(struct task_struct *task) 174 { 175 if (unlikely(task != current)) 176 return; 177 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) 178 sync_mm_rss(task->mm); 179 } 180 #else /* SPLIT_RSS_COUNTING */ 181 182 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) 183 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) 184 185 static void check_sync_rss_stat(struct task_struct *task) 186 { 187 } 188 189 #endif /* SPLIT_RSS_COUNTING */ 190 191 /* 192 * Note: this doesn't free the actual pages themselves. That 193 * has been handled earlier when unmapping all the memory regions. 194 */ 195 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 196 unsigned long addr) 197 { 198 pgtable_t token = pmd_pgtable(*pmd); 199 pmd_clear(pmd); 200 pte_free_tlb(tlb, token, addr); 201 mm_dec_nr_ptes(tlb->mm); 202 } 203 204 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 205 unsigned long addr, unsigned long end, 206 unsigned long floor, unsigned long ceiling) 207 { 208 pmd_t *pmd; 209 unsigned long next; 210 unsigned long start; 211 212 start = addr; 213 pmd = pmd_offset(pud, addr); 214 do { 215 next = pmd_addr_end(addr, end); 216 if (pmd_none_or_clear_bad(pmd)) 217 continue; 218 free_pte_range(tlb, pmd, addr); 219 } while (pmd++, addr = next, addr != end); 220 221 start &= PUD_MASK; 222 if (start < floor) 223 return; 224 if (ceiling) { 225 ceiling &= PUD_MASK; 226 if (!ceiling) 227 return; 228 } 229 if (end - 1 > ceiling - 1) 230 return; 231 232 pmd = pmd_offset(pud, start); 233 pud_clear(pud); 234 pmd_free_tlb(tlb, pmd, start); 235 mm_dec_nr_pmds(tlb->mm); 236 } 237 238 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 239 unsigned long addr, unsigned long end, 240 unsigned long floor, unsigned long ceiling) 241 { 242 pud_t *pud; 243 unsigned long next; 244 unsigned long start; 245 246 start = addr; 247 pud = pud_offset(p4d, addr); 248 do { 249 next = pud_addr_end(addr, end); 250 if (pud_none_or_clear_bad(pud)) 251 continue; 252 free_pmd_range(tlb, pud, addr, next, floor, ceiling); 253 } while (pud++, addr = next, addr != end); 254 255 start &= P4D_MASK; 256 if (start < floor) 257 return; 258 if (ceiling) { 259 ceiling &= P4D_MASK; 260 if (!ceiling) 261 return; 262 } 263 if (end - 1 > ceiling - 1) 264 return; 265 266 pud = pud_offset(p4d, start); 267 p4d_clear(p4d); 268 pud_free_tlb(tlb, pud, start); 269 mm_dec_nr_puds(tlb->mm); 270 } 271 272 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 273 unsigned long addr, unsigned long end, 274 unsigned long floor, unsigned long ceiling) 275 { 276 p4d_t *p4d; 277 unsigned long next; 278 unsigned long start; 279 280 start = addr; 281 p4d = p4d_offset(pgd, addr); 282 do { 283 next = p4d_addr_end(addr, end); 284 if (p4d_none_or_clear_bad(p4d)) 285 continue; 286 free_pud_range(tlb, p4d, addr, next, floor, ceiling); 287 } while (p4d++, addr = next, addr != end); 288 289 start &= PGDIR_MASK; 290 if (start < floor) 291 return; 292 if (ceiling) { 293 ceiling &= PGDIR_MASK; 294 if (!ceiling) 295 return; 296 } 297 if (end - 1 > ceiling - 1) 298 return; 299 300 p4d = p4d_offset(pgd, start); 301 pgd_clear(pgd); 302 p4d_free_tlb(tlb, p4d, start); 303 } 304 305 /* 306 * This function frees user-level page tables of a process. 307 */ 308 void free_pgd_range(struct mmu_gather *tlb, 309 unsigned long addr, unsigned long end, 310 unsigned long floor, unsigned long ceiling) 311 { 312 pgd_t *pgd; 313 unsigned long next; 314 315 /* 316 * The next few lines have given us lots of grief... 317 * 318 * Why are we testing PMD* at this top level? Because often 319 * there will be no work to do at all, and we'd prefer not to 320 * go all the way down to the bottom just to discover that. 321 * 322 * Why all these "- 1"s? Because 0 represents both the bottom 323 * of the address space and the top of it (using -1 for the 324 * top wouldn't help much: the masks would do the wrong thing). 325 * The rule is that addr 0 and floor 0 refer to the bottom of 326 * the address space, but end 0 and ceiling 0 refer to the top 327 * Comparisons need to use "end - 1" and "ceiling - 1" (though 328 * that end 0 case should be mythical). 329 * 330 * Wherever addr is brought up or ceiling brought down, we must 331 * be careful to reject "the opposite 0" before it confuses the 332 * subsequent tests. But what about where end is brought down 333 * by PMD_SIZE below? no, end can't go down to 0 there. 334 * 335 * Whereas we round start (addr) and ceiling down, by different 336 * masks at different levels, in order to test whether a table 337 * now has no other vmas using it, so can be freed, we don't 338 * bother to round floor or end up - the tests don't need that. 339 */ 340 341 addr &= PMD_MASK; 342 if (addr < floor) { 343 addr += PMD_SIZE; 344 if (!addr) 345 return; 346 } 347 if (ceiling) { 348 ceiling &= PMD_MASK; 349 if (!ceiling) 350 return; 351 } 352 if (end - 1 > ceiling - 1) 353 end -= PMD_SIZE; 354 if (addr > end - 1) 355 return; 356 /* 357 * We add page table cache pages with PAGE_SIZE, 358 * (see pte_free_tlb()), flush the tlb if we need 359 */ 360 tlb_change_page_size(tlb, PAGE_SIZE); 361 pgd = pgd_offset(tlb->mm, addr); 362 do { 363 next = pgd_addr_end(addr, end); 364 if (pgd_none_or_clear_bad(pgd)) 365 continue; 366 free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 367 } while (pgd++, addr = next, addr != end); 368 } 369 370 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, 371 unsigned long floor, unsigned long ceiling) 372 { 373 while (vma) { 374 struct vm_area_struct *next = vma->vm_next; 375 unsigned long addr = vma->vm_start; 376 377 /* 378 * Hide vma from rmap and truncate_pagecache before freeing 379 * pgtables 380 */ 381 unlink_anon_vmas(vma); 382 unlink_file_vma(vma); 383 384 if (is_vm_hugetlb_page(vma)) { 385 hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 386 floor, next ? next->vm_start : ceiling); 387 } else { 388 /* 389 * Optimization: gather nearby vmas into one call down 390 */ 391 while (next && next->vm_start <= vma->vm_end + PMD_SIZE 392 && !is_vm_hugetlb_page(next)) { 393 vma = next; 394 next = vma->vm_next; 395 unlink_anon_vmas(vma); 396 unlink_file_vma(vma); 397 } 398 free_pgd_range(tlb, addr, vma->vm_end, 399 floor, next ? next->vm_start : ceiling); 400 } 401 vma = next; 402 } 403 } 404 405 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 406 { 407 spinlock_t *ptl; 408 pgtable_t new = pte_alloc_one(mm); 409 if (!new) 410 return -ENOMEM; 411 412 /* 413 * Ensure all pte setup (eg. pte page lock and page clearing) are 414 * visible before the pte is made visible to other CPUs by being 415 * put into page tables. 416 * 417 * The other side of the story is the pointer chasing in the page 418 * table walking code (when walking the page table without locking; 419 * ie. most of the time). Fortunately, these data accesses consist 420 * of a chain of data-dependent loads, meaning most CPUs (alpha 421 * being the notable exception) will already guarantee loads are 422 * seen in-order. See the alpha page table accessors for the 423 * smp_read_barrier_depends() barriers in page table walking code. 424 */ 425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 426 427 ptl = pmd_lock(mm, pmd); 428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 429 mm_inc_nr_ptes(mm); 430 pmd_populate(mm, pmd, new); 431 new = NULL; 432 } 433 spin_unlock(ptl); 434 if (new) 435 pte_free(mm, new); 436 return 0; 437 } 438 439 int __pte_alloc_kernel(pmd_t *pmd) 440 { 441 pte_t *new = pte_alloc_one_kernel(&init_mm); 442 if (!new) 443 return -ENOMEM; 444 445 smp_wmb(); /* See comment in __pte_alloc */ 446 447 spin_lock(&init_mm.page_table_lock); 448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 449 pmd_populate_kernel(&init_mm, pmd, new); 450 new = NULL; 451 } 452 spin_unlock(&init_mm.page_table_lock); 453 if (new) 454 pte_free_kernel(&init_mm, new); 455 return 0; 456 } 457 458 static inline void init_rss_vec(int *rss) 459 { 460 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 461 } 462 463 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 464 { 465 int i; 466 467 if (current->mm == mm) 468 sync_mm_rss(mm); 469 for (i = 0; i < NR_MM_COUNTERS; i++) 470 if (rss[i]) 471 add_mm_counter(mm, i, rss[i]); 472 } 473 474 /* 475 * This function is called to print an error when a bad pte 476 * is found. For example, we might have a PFN-mapped pte in 477 * a region that doesn't allow it. 478 * 479 * The calling function must still handle the error. 480 */ 481 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 482 pte_t pte, struct page *page) 483 { 484 pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 485 p4d_t *p4d = p4d_offset(pgd, addr); 486 pud_t *pud = pud_offset(p4d, addr); 487 pmd_t *pmd = pmd_offset(pud, addr); 488 struct address_space *mapping; 489 pgoff_t index; 490 static unsigned long resume; 491 static unsigned long nr_shown; 492 static unsigned long nr_unshown; 493 494 /* 495 * Allow a burst of 60 reports, then keep quiet for that minute; 496 * or allow a steady drip of one report per second. 497 */ 498 if (nr_shown == 60) { 499 if (time_before(jiffies, resume)) { 500 nr_unshown++; 501 return; 502 } 503 if (nr_unshown) { 504 pr_alert("BUG: Bad page map: %lu messages suppressed\n", 505 nr_unshown); 506 nr_unshown = 0; 507 } 508 nr_shown = 0; 509 } 510 if (nr_shown++ == 0) 511 resume = jiffies + 60 * HZ; 512 513 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 514 index = linear_page_index(vma, addr); 515 516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 517 current->comm, 518 (long long)pte_val(pte), (long long)pmd_val(*pmd)); 519 if (page) 520 dump_page(page, "bad pte"); 521 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", 522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n", 524 vma->vm_file, 525 vma->vm_ops ? vma->vm_ops->fault : NULL, 526 vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 527 mapping ? mapping->a_ops->readpage : NULL); 528 dump_stack(); 529 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 530 } 531 532 /* 533 * vm_normal_page -- This function gets the "struct page" associated with a pte. 534 * 535 * "Special" mappings do not wish to be associated with a "struct page" (either 536 * it doesn't exist, or it exists but they don't want to touch it). In this 537 * case, NULL is returned here. "Normal" mappings do have a struct page. 538 * 539 * There are 2 broad cases. Firstly, an architecture may define a pte_special() 540 * pte bit, in which case this function is trivial. Secondly, an architecture 541 * may not have a spare pte bit, which requires a more complicated scheme, 542 * described below. 543 * 544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 545 * special mapping (even if there are underlying and valid "struct pages"). 546 * COWed pages of a VM_PFNMAP are always normal. 547 * 548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the 549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special 551 * mapping will always honor the rule 552 * 553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 554 * 555 * And for normal mappings this is false. 556 * 557 * This restricts such mappings to be a linear translation from virtual address 558 * to pfn. To get around this restriction, we allow arbitrary mappings so long 559 * as the vma is not a COW mapping; in that case, we know that all ptes are 560 * special (because none can have been COWed). 561 * 562 * 563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 564 * 565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 566 * page" backing, however the difference is that _all_ pages with a struct 567 * page (that is, those where pfn_valid is true) are refcounted and considered 568 * normal pages by the VM. The disadvantage is that pages are refcounted 569 * (which can be slower and simply not an option for some PFNMAP users). The 570 * advantage is that we don't have to follow the strict linearity rule of 571 * PFNMAP mappings in order to support COWable mappings. 572 * 573 */ 574 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 575 pte_t pte) 576 { 577 unsigned long pfn = pte_pfn(pte); 578 579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 580 if (likely(!pte_special(pte))) 581 goto check_pfn; 582 if (vma->vm_ops && vma->vm_ops->find_special_page) 583 return vma->vm_ops->find_special_page(vma, addr); 584 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 585 return NULL; 586 if (is_zero_pfn(pfn)) 587 return NULL; 588 if (pte_devmap(pte)) 589 return NULL; 590 591 print_bad_pte(vma, addr, pte, NULL); 592 return NULL; 593 } 594 595 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 596 597 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 598 if (vma->vm_flags & VM_MIXEDMAP) { 599 if (!pfn_valid(pfn)) 600 return NULL; 601 goto out; 602 } else { 603 unsigned long off; 604 off = (addr - vma->vm_start) >> PAGE_SHIFT; 605 if (pfn == vma->vm_pgoff + off) 606 return NULL; 607 if (!is_cow_mapping(vma->vm_flags)) 608 return NULL; 609 } 610 } 611 612 if (is_zero_pfn(pfn)) 613 return NULL; 614 615 check_pfn: 616 if (unlikely(pfn > highest_memmap_pfn)) { 617 print_bad_pte(vma, addr, pte, NULL); 618 return NULL; 619 } 620 621 /* 622 * NOTE! We still have PageReserved() pages in the page tables. 623 * eg. VDSO mappings can cause them to exist. 624 */ 625 out: 626 return pfn_to_page(pfn); 627 } 628 629 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 630 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 631 pmd_t pmd) 632 { 633 unsigned long pfn = pmd_pfn(pmd); 634 635 /* 636 * There is no pmd_special() but there may be special pmds, e.g. 637 * in a direct-access (dax) mapping, so let's just replicate the 638 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 639 */ 640 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 641 if (vma->vm_flags & VM_MIXEDMAP) { 642 if (!pfn_valid(pfn)) 643 return NULL; 644 goto out; 645 } else { 646 unsigned long off; 647 off = (addr - vma->vm_start) >> PAGE_SHIFT; 648 if (pfn == vma->vm_pgoff + off) 649 return NULL; 650 if (!is_cow_mapping(vma->vm_flags)) 651 return NULL; 652 } 653 } 654 655 if (pmd_devmap(pmd)) 656 return NULL; 657 if (is_zero_pfn(pfn)) 658 return NULL; 659 if (unlikely(pfn > highest_memmap_pfn)) 660 return NULL; 661 662 /* 663 * NOTE! We still have PageReserved() pages in the page tables. 664 * eg. VDSO mappings can cause them to exist. 665 */ 666 out: 667 return pfn_to_page(pfn); 668 } 669 #endif 670 671 /* 672 * copy one vm_area from one task to the other. Assumes the page tables 673 * already present in the new task to be cleared in the whole range 674 * covered by this vma. 675 */ 676 677 static inline unsigned long 678 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 679 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, 680 unsigned long addr, int *rss) 681 { 682 unsigned long vm_flags = vma->vm_flags; 683 pte_t pte = *src_pte; 684 struct page *page; 685 686 /* pte contains position in swap or file, so copy. */ 687 if (unlikely(!pte_present(pte))) { 688 swp_entry_t entry = pte_to_swp_entry(pte); 689 690 if (likely(!non_swap_entry(entry))) { 691 if (swap_duplicate(entry) < 0) 692 return entry.val; 693 694 /* make sure dst_mm is on swapoff's mmlist. */ 695 if (unlikely(list_empty(&dst_mm->mmlist))) { 696 spin_lock(&mmlist_lock); 697 if (list_empty(&dst_mm->mmlist)) 698 list_add(&dst_mm->mmlist, 699 &src_mm->mmlist); 700 spin_unlock(&mmlist_lock); 701 } 702 rss[MM_SWAPENTS]++; 703 } else if (is_migration_entry(entry)) { 704 page = migration_entry_to_page(entry); 705 706 rss[mm_counter(page)]++; 707 708 if (is_write_migration_entry(entry) && 709 is_cow_mapping(vm_flags)) { 710 /* 711 * COW mappings require pages in both 712 * parent and child to be set to read. 713 */ 714 make_migration_entry_read(&entry); 715 pte = swp_entry_to_pte(entry); 716 if (pte_swp_soft_dirty(*src_pte)) 717 pte = pte_swp_mksoft_dirty(pte); 718 set_pte_at(src_mm, addr, src_pte, pte); 719 } 720 } else if (is_device_private_entry(entry)) { 721 page = device_private_entry_to_page(entry); 722 723 /* 724 * Update rss count even for unaddressable pages, as 725 * they should treated just like normal pages in this 726 * respect. 727 * 728 * We will likely want to have some new rss counters 729 * for unaddressable pages, at some point. But for now 730 * keep things as they are. 731 */ 732 get_page(page); 733 rss[mm_counter(page)]++; 734 page_dup_rmap(page, false); 735 736 /* 737 * We do not preserve soft-dirty information, because so 738 * far, checkpoint/restore is the only feature that 739 * requires that. And checkpoint/restore does not work 740 * when a device driver is involved (you cannot easily 741 * save and restore device driver state). 742 */ 743 if (is_write_device_private_entry(entry) && 744 is_cow_mapping(vm_flags)) { 745 make_device_private_entry_read(&entry); 746 pte = swp_entry_to_pte(entry); 747 set_pte_at(src_mm, addr, src_pte, pte); 748 } 749 } 750 goto out_set_pte; 751 } 752 753 /* 754 * If it's a COW mapping, write protect it both 755 * in the parent and the child 756 */ 757 if (is_cow_mapping(vm_flags) && pte_write(pte)) { 758 ptep_set_wrprotect(src_mm, addr, src_pte); 759 pte = pte_wrprotect(pte); 760 } 761 762 /* 763 * If it's a shared mapping, mark it clean in 764 * the child 765 */ 766 if (vm_flags & VM_SHARED) 767 pte = pte_mkclean(pte); 768 pte = pte_mkold(pte); 769 770 page = vm_normal_page(vma, addr, pte); 771 if (page) { 772 get_page(page); 773 page_dup_rmap(page, false); 774 rss[mm_counter(page)]++; 775 } else if (pte_devmap(pte)) { 776 page = pte_page(pte); 777 } 778 779 out_set_pte: 780 set_pte_at(dst_mm, addr, dst_pte, pte); 781 return 0; 782 } 783 784 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 785 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, 786 unsigned long addr, unsigned long end) 787 { 788 pte_t *orig_src_pte, *orig_dst_pte; 789 pte_t *src_pte, *dst_pte; 790 spinlock_t *src_ptl, *dst_ptl; 791 int progress = 0; 792 int rss[NR_MM_COUNTERS]; 793 swp_entry_t entry = (swp_entry_t){0}; 794 795 again: 796 init_rss_vec(rss); 797 798 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 799 if (!dst_pte) 800 return -ENOMEM; 801 src_pte = pte_offset_map(src_pmd, addr); 802 src_ptl = pte_lockptr(src_mm, src_pmd); 803 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 804 orig_src_pte = src_pte; 805 orig_dst_pte = dst_pte; 806 arch_enter_lazy_mmu_mode(); 807 808 do { 809 /* 810 * We are holding two locks at this point - either of them 811 * could generate latencies in another task on another CPU. 812 */ 813 if (progress >= 32) { 814 progress = 0; 815 if (need_resched() || 816 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 817 break; 818 } 819 if (pte_none(*src_pte)) { 820 progress++; 821 continue; 822 } 823 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, 824 vma, addr, rss); 825 if (entry.val) 826 break; 827 progress += 8; 828 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); 829 830 arch_leave_lazy_mmu_mode(); 831 spin_unlock(src_ptl); 832 pte_unmap(orig_src_pte); 833 add_mm_rss_vec(dst_mm, rss); 834 pte_unmap_unlock(orig_dst_pte, dst_ptl); 835 cond_resched(); 836 837 if (entry.val) { 838 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) 839 return -ENOMEM; 840 progress = 0; 841 } 842 if (addr != end) 843 goto again; 844 return 0; 845 } 846 847 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 848 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, 849 unsigned long addr, unsigned long end) 850 { 851 pmd_t *src_pmd, *dst_pmd; 852 unsigned long next; 853 854 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 855 if (!dst_pmd) 856 return -ENOMEM; 857 src_pmd = pmd_offset(src_pud, addr); 858 do { 859 next = pmd_addr_end(addr, end); 860 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 861 || pmd_devmap(*src_pmd)) { 862 int err; 863 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma); 864 err = copy_huge_pmd(dst_mm, src_mm, 865 dst_pmd, src_pmd, addr, vma); 866 if (err == -ENOMEM) 867 return -ENOMEM; 868 if (!err) 869 continue; 870 /* fall through */ 871 } 872 if (pmd_none_or_clear_bad(src_pmd)) 873 continue; 874 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, 875 vma, addr, next)) 876 return -ENOMEM; 877 } while (dst_pmd++, src_pmd++, addr = next, addr != end); 878 return 0; 879 } 880 881 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 882 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma, 883 unsigned long addr, unsigned long end) 884 { 885 pud_t *src_pud, *dst_pud; 886 unsigned long next; 887 888 dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 889 if (!dst_pud) 890 return -ENOMEM; 891 src_pud = pud_offset(src_p4d, addr); 892 do { 893 next = pud_addr_end(addr, end); 894 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 895 int err; 896 897 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma); 898 err = copy_huge_pud(dst_mm, src_mm, 899 dst_pud, src_pud, addr, vma); 900 if (err == -ENOMEM) 901 return -ENOMEM; 902 if (!err) 903 continue; 904 /* fall through */ 905 } 906 if (pud_none_or_clear_bad(src_pud)) 907 continue; 908 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, 909 vma, addr, next)) 910 return -ENOMEM; 911 } while (dst_pud++, src_pud++, addr = next, addr != end); 912 return 0; 913 } 914 915 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 916 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, 917 unsigned long addr, unsigned long end) 918 { 919 p4d_t *src_p4d, *dst_p4d; 920 unsigned long next; 921 922 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 923 if (!dst_p4d) 924 return -ENOMEM; 925 src_p4d = p4d_offset(src_pgd, addr); 926 do { 927 next = p4d_addr_end(addr, end); 928 if (p4d_none_or_clear_bad(src_p4d)) 929 continue; 930 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d, 931 vma, addr, next)) 932 return -ENOMEM; 933 } while (dst_p4d++, src_p4d++, addr = next, addr != end); 934 return 0; 935 } 936 937 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, 938 struct vm_area_struct *vma) 939 { 940 pgd_t *src_pgd, *dst_pgd; 941 unsigned long next; 942 unsigned long addr = vma->vm_start; 943 unsigned long end = vma->vm_end; 944 struct mmu_notifier_range range; 945 bool is_cow; 946 int ret; 947 948 /* 949 * Don't copy ptes where a page fault will fill them correctly. 950 * Fork becomes much lighter when there are big shared or private 951 * readonly mappings. The tradeoff is that copy_page_range is more 952 * efficient than faulting. 953 */ 954 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && 955 !vma->anon_vma) 956 return 0; 957 958 if (is_vm_hugetlb_page(vma)) 959 return copy_hugetlb_page_range(dst_mm, src_mm, vma); 960 961 if (unlikely(vma->vm_flags & VM_PFNMAP)) { 962 /* 963 * We do not free on error cases below as remove_vma 964 * gets called on error from higher level routine 965 */ 966 ret = track_pfn_copy(vma); 967 if (ret) 968 return ret; 969 } 970 971 /* 972 * We need to invalidate the secondary MMU mappings only when 973 * there could be a permission downgrade on the ptes of the 974 * parent mm. And a permission downgrade will only happen if 975 * is_cow_mapping() returns true. 976 */ 977 is_cow = is_cow_mapping(vma->vm_flags); 978 979 if (is_cow) { 980 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 981 0, vma, src_mm, addr, end); 982 mmu_notifier_invalidate_range_start(&range); 983 } 984 985 ret = 0; 986 dst_pgd = pgd_offset(dst_mm, addr); 987 src_pgd = pgd_offset(src_mm, addr); 988 do { 989 next = pgd_addr_end(addr, end); 990 if (pgd_none_or_clear_bad(src_pgd)) 991 continue; 992 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd, 993 vma, addr, next))) { 994 ret = -ENOMEM; 995 break; 996 } 997 } while (dst_pgd++, src_pgd++, addr = next, addr != end); 998 999 if (is_cow) 1000 mmu_notifier_invalidate_range_end(&range); 1001 return ret; 1002 } 1003 1004 static unsigned long zap_pte_range(struct mmu_gather *tlb, 1005 struct vm_area_struct *vma, pmd_t *pmd, 1006 unsigned long addr, unsigned long end, 1007 struct zap_details *details) 1008 { 1009 struct mm_struct *mm = tlb->mm; 1010 int force_flush = 0; 1011 int rss[NR_MM_COUNTERS]; 1012 spinlock_t *ptl; 1013 pte_t *start_pte; 1014 pte_t *pte; 1015 swp_entry_t entry; 1016 1017 tlb_change_page_size(tlb, PAGE_SIZE); 1018 again: 1019 init_rss_vec(rss); 1020 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 1021 pte = start_pte; 1022 flush_tlb_batched_pending(mm); 1023 arch_enter_lazy_mmu_mode(); 1024 do { 1025 pte_t ptent = *pte; 1026 if (pte_none(ptent)) 1027 continue; 1028 1029 if (pte_present(ptent)) { 1030 struct page *page; 1031 1032 page = vm_normal_page(vma, addr, ptent); 1033 if (unlikely(details) && page) { 1034 /* 1035 * unmap_shared_mapping_pages() wants to 1036 * invalidate cache without truncating: 1037 * unmap shared but keep private pages. 1038 */ 1039 if (details->check_mapping && 1040 details->check_mapping != page_rmapping(page)) 1041 continue; 1042 } 1043 ptent = ptep_get_and_clear_full(mm, addr, pte, 1044 tlb->fullmm); 1045 tlb_remove_tlb_entry(tlb, pte, addr); 1046 if (unlikely(!page)) 1047 continue; 1048 1049 if (!PageAnon(page)) { 1050 if (pte_dirty(ptent)) { 1051 force_flush = 1; 1052 set_page_dirty(page); 1053 } 1054 if (pte_young(ptent) && 1055 likely(!(vma->vm_flags & VM_SEQ_READ))) 1056 mark_page_accessed(page); 1057 } 1058 rss[mm_counter(page)]--; 1059 page_remove_rmap(page, false); 1060 if (unlikely(page_mapcount(page) < 0)) 1061 print_bad_pte(vma, addr, ptent, page); 1062 if (unlikely(__tlb_remove_page(tlb, page))) { 1063 force_flush = 1; 1064 addr += PAGE_SIZE; 1065 break; 1066 } 1067 continue; 1068 } 1069 1070 entry = pte_to_swp_entry(ptent); 1071 if (non_swap_entry(entry) && is_device_private_entry(entry)) { 1072 struct page *page = device_private_entry_to_page(entry); 1073 1074 if (unlikely(details && details->check_mapping)) { 1075 /* 1076 * unmap_shared_mapping_pages() wants to 1077 * invalidate cache without truncating: 1078 * unmap shared but keep private pages. 1079 */ 1080 if (details->check_mapping != 1081 page_rmapping(page)) 1082 continue; 1083 } 1084 1085 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1086 rss[mm_counter(page)]--; 1087 page_remove_rmap(page, false); 1088 put_page(page); 1089 continue; 1090 } 1091 1092 /* If details->check_mapping, we leave swap entries. */ 1093 if (unlikely(details)) 1094 continue; 1095 1096 entry = pte_to_swp_entry(ptent); 1097 if (!non_swap_entry(entry)) 1098 rss[MM_SWAPENTS]--; 1099 else if (is_migration_entry(entry)) { 1100 struct page *page; 1101 1102 page = migration_entry_to_page(entry); 1103 rss[mm_counter(page)]--; 1104 } 1105 if (unlikely(!free_swap_and_cache(entry))) 1106 print_bad_pte(vma, addr, ptent, NULL); 1107 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1108 } while (pte++, addr += PAGE_SIZE, addr != end); 1109 1110 add_mm_rss_vec(mm, rss); 1111 arch_leave_lazy_mmu_mode(); 1112 1113 /* Do the actual TLB flush before dropping ptl */ 1114 if (force_flush) 1115 tlb_flush_mmu_tlbonly(tlb); 1116 pte_unmap_unlock(start_pte, ptl); 1117 1118 /* 1119 * If we forced a TLB flush (either due to running out of 1120 * batch buffers or because we needed to flush dirty TLB 1121 * entries before releasing the ptl), free the batched 1122 * memory too. Restart if we didn't do everything. 1123 */ 1124 if (force_flush) { 1125 force_flush = 0; 1126 tlb_flush_mmu(tlb); 1127 if (addr != end) 1128 goto again; 1129 } 1130 1131 return addr; 1132 } 1133 1134 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1135 struct vm_area_struct *vma, pud_t *pud, 1136 unsigned long addr, unsigned long end, 1137 struct zap_details *details) 1138 { 1139 pmd_t *pmd; 1140 unsigned long next; 1141 1142 pmd = pmd_offset(pud, addr); 1143 do { 1144 next = pmd_addr_end(addr, end); 1145 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 1146 if (next - addr != HPAGE_PMD_SIZE) 1147 __split_huge_pmd(vma, pmd, addr, false, NULL); 1148 else if (zap_huge_pmd(tlb, vma, pmd, addr)) 1149 goto next; 1150 /* fall through */ 1151 } 1152 /* 1153 * Here there can be other concurrent MADV_DONTNEED or 1154 * trans huge page faults running, and if the pmd is 1155 * none or trans huge it can change under us. This is 1156 * because MADV_DONTNEED holds the mmap_sem in read 1157 * mode. 1158 */ 1159 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1160 goto next; 1161 next = zap_pte_range(tlb, vma, pmd, addr, next, details); 1162 next: 1163 cond_resched(); 1164 } while (pmd++, addr = next, addr != end); 1165 1166 return addr; 1167 } 1168 1169 static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1170 struct vm_area_struct *vma, p4d_t *p4d, 1171 unsigned long addr, unsigned long end, 1172 struct zap_details *details) 1173 { 1174 pud_t *pud; 1175 unsigned long next; 1176 1177 pud = pud_offset(p4d, addr); 1178 do { 1179 next = pud_addr_end(addr, end); 1180 if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1181 if (next - addr != HPAGE_PUD_SIZE) { 1182 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma); 1183 split_huge_pud(vma, pud, addr); 1184 } else if (zap_huge_pud(tlb, vma, pud, addr)) 1185 goto next; 1186 /* fall through */ 1187 } 1188 if (pud_none_or_clear_bad(pud)) 1189 continue; 1190 next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1191 next: 1192 cond_resched(); 1193 } while (pud++, addr = next, addr != end); 1194 1195 return addr; 1196 } 1197 1198 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1199 struct vm_area_struct *vma, pgd_t *pgd, 1200 unsigned long addr, unsigned long end, 1201 struct zap_details *details) 1202 { 1203 p4d_t *p4d; 1204 unsigned long next; 1205 1206 p4d = p4d_offset(pgd, addr); 1207 do { 1208 next = p4d_addr_end(addr, end); 1209 if (p4d_none_or_clear_bad(p4d)) 1210 continue; 1211 next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1212 } while (p4d++, addr = next, addr != end); 1213 1214 return addr; 1215 } 1216 1217 void unmap_page_range(struct mmu_gather *tlb, 1218 struct vm_area_struct *vma, 1219 unsigned long addr, unsigned long end, 1220 struct zap_details *details) 1221 { 1222 pgd_t *pgd; 1223 unsigned long next; 1224 1225 BUG_ON(addr >= end); 1226 tlb_start_vma(tlb, vma); 1227 pgd = pgd_offset(vma->vm_mm, addr); 1228 do { 1229 next = pgd_addr_end(addr, end); 1230 if (pgd_none_or_clear_bad(pgd)) 1231 continue; 1232 next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 1233 } while (pgd++, addr = next, addr != end); 1234 tlb_end_vma(tlb, vma); 1235 } 1236 1237 1238 static void unmap_single_vma(struct mmu_gather *tlb, 1239 struct vm_area_struct *vma, unsigned long start_addr, 1240 unsigned long end_addr, 1241 struct zap_details *details) 1242 { 1243 unsigned long start = max(vma->vm_start, start_addr); 1244 unsigned long end; 1245 1246 if (start >= vma->vm_end) 1247 return; 1248 end = min(vma->vm_end, end_addr); 1249 if (end <= vma->vm_start) 1250 return; 1251 1252 if (vma->vm_file) 1253 uprobe_munmap(vma, start, end); 1254 1255 if (unlikely(vma->vm_flags & VM_PFNMAP)) 1256 untrack_pfn(vma, 0, 0); 1257 1258 if (start != end) { 1259 if (unlikely(is_vm_hugetlb_page(vma))) { 1260 /* 1261 * It is undesirable to test vma->vm_file as it 1262 * should be non-null for valid hugetlb area. 1263 * However, vm_file will be NULL in the error 1264 * cleanup path of mmap_region. When 1265 * hugetlbfs ->mmap method fails, 1266 * mmap_region() nullifies vma->vm_file 1267 * before calling this function to clean up. 1268 * Since no pte has actually been setup, it is 1269 * safe to do nothing in this case. 1270 */ 1271 if (vma->vm_file) { 1272 i_mmap_lock_write(vma->vm_file->f_mapping); 1273 __unmap_hugepage_range_final(tlb, vma, start, end, NULL); 1274 i_mmap_unlock_write(vma->vm_file->f_mapping); 1275 } 1276 } else 1277 unmap_page_range(tlb, vma, start, end, details); 1278 } 1279 } 1280 1281 /** 1282 * unmap_vmas - unmap a range of memory covered by a list of vma's 1283 * @tlb: address of the caller's struct mmu_gather 1284 * @vma: the starting vma 1285 * @start_addr: virtual address at which to start unmapping 1286 * @end_addr: virtual address at which to end unmapping 1287 * 1288 * Unmap all pages in the vma list. 1289 * 1290 * Only addresses between `start' and `end' will be unmapped. 1291 * 1292 * The VMA list must be sorted in ascending virtual address order. 1293 * 1294 * unmap_vmas() assumes that the caller will flush the whole unmapped address 1295 * range after unmap_vmas() returns. So the only responsibility here is to 1296 * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1297 * drops the lock and schedules. 1298 */ 1299 void unmap_vmas(struct mmu_gather *tlb, 1300 struct vm_area_struct *vma, unsigned long start_addr, 1301 unsigned long end_addr) 1302 { 1303 struct mmu_notifier_range range; 1304 1305 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, 1306 start_addr, end_addr); 1307 mmu_notifier_invalidate_range_start(&range); 1308 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) 1309 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); 1310 mmu_notifier_invalidate_range_end(&range); 1311 } 1312 1313 /** 1314 * zap_page_range - remove user pages in a given range 1315 * @vma: vm_area_struct holding the applicable pages 1316 * @start: starting address of pages to zap 1317 * @size: number of bytes to zap 1318 * 1319 * Caller must protect the VMA list 1320 */ 1321 void zap_page_range(struct vm_area_struct *vma, unsigned long start, 1322 unsigned long size) 1323 { 1324 struct mmu_notifier_range range; 1325 struct mmu_gather tlb; 1326 1327 lru_add_drain(); 1328 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1329 start, start + size); 1330 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end); 1331 update_hiwater_rss(vma->vm_mm); 1332 mmu_notifier_invalidate_range_start(&range); 1333 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next) 1334 unmap_single_vma(&tlb, vma, start, range.end, NULL); 1335 mmu_notifier_invalidate_range_end(&range); 1336 tlb_finish_mmu(&tlb, start, range.end); 1337 } 1338 1339 /** 1340 * zap_page_range_single - remove user pages in a given range 1341 * @vma: vm_area_struct holding the applicable pages 1342 * @address: starting address of pages to zap 1343 * @size: number of bytes to zap 1344 * @details: details of shared cache invalidation 1345 * 1346 * The range must fit into one VMA. 1347 */ 1348 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1349 unsigned long size, struct zap_details *details) 1350 { 1351 struct mmu_notifier_range range; 1352 struct mmu_gather tlb; 1353 1354 lru_add_drain(); 1355 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1356 address, address + size); 1357 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end); 1358 update_hiwater_rss(vma->vm_mm); 1359 mmu_notifier_invalidate_range_start(&range); 1360 unmap_single_vma(&tlb, vma, address, range.end, details); 1361 mmu_notifier_invalidate_range_end(&range); 1362 tlb_finish_mmu(&tlb, address, range.end); 1363 } 1364 1365 /** 1366 * zap_vma_ptes - remove ptes mapping the vma 1367 * @vma: vm_area_struct holding ptes to be zapped 1368 * @address: starting address of pages to zap 1369 * @size: number of bytes to zap 1370 * 1371 * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1372 * 1373 * The entire address range must be fully contained within the vma. 1374 * 1375 */ 1376 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1377 unsigned long size) 1378 { 1379 if (address < vma->vm_start || address + size > vma->vm_end || 1380 !(vma->vm_flags & VM_PFNMAP)) 1381 return; 1382 1383 zap_page_range_single(vma, address, size, NULL); 1384 } 1385 EXPORT_SYMBOL_GPL(zap_vma_ptes); 1386 1387 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 1388 spinlock_t **ptl) 1389 { 1390 pgd_t *pgd; 1391 p4d_t *p4d; 1392 pud_t *pud; 1393 pmd_t *pmd; 1394 1395 pgd = pgd_offset(mm, addr); 1396 p4d = p4d_alloc(mm, pgd, addr); 1397 if (!p4d) 1398 return NULL; 1399 pud = pud_alloc(mm, p4d, addr); 1400 if (!pud) 1401 return NULL; 1402 pmd = pmd_alloc(mm, pud, addr); 1403 if (!pmd) 1404 return NULL; 1405 1406 VM_BUG_ON(pmd_trans_huge(*pmd)); 1407 return pte_alloc_map_lock(mm, pmd, addr, ptl); 1408 } 1409 1410 /* 1411 * This is the old fallback for page remapping. 1412 * 1413 * For historical reasons, it only allows reserved pages. Only 1414 * old drivers should use this, and they needed to mark their 1415 * pages reserved for the old functions anyway. 1416 */ 1417 static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1418 struct page *page, pgprot_t prot) 1419 { 1420 struct mm_struct *mm = vma->vm_mm; 1421 int retval; 1422 pte_t *pte; 1423 spinlock_t *ptl; 1424 1425 retval = -EINVAL; 1426 if (PageAnon(page) || PageSlab(page) || page_has_type(page)) 1427 goto out; 1428 retval = -ENOMEM; 1429 flush_dcache_page(page); 1430 pte = get_locked_pte(mm, addr, &ptl); 1431 if (!pte) 1432 goto out; 1433 retval = -EBUSY; 1434 if (!pte_none(*pte)) 1435 goto out_unlock; 1436 1437 /* Ok, finally just insert the thing.. */ 1438 get_page(page); 1439 inc_mm_counter_fast(mm, mm_counter_file(page)); 1440 page_add_file_rmap(page, false); 1441 set_pte_at(mm, addr, pte, mk_pte(page, prot)); 1442 1443 retval = 0; 1444 out_unlock: 1445 pte_unmap_unlock(pte, ptl); 1446 out: 1447 return retval; 1448 } 1449 1450 /** 1451 * vm_insert_page - insert single page into user vma 1452 * @vma: user vma to map to 1453 * @addr: target user address of this page 1454 * @page: source kernel page 1455 * 1456 * This allows drivers to insert individual pages they've allocated 1457 * into a user vma. 1458 * 1459 * The page has to be a nice clean _individual_ kernel allocation. 1460 * If you allocate a compound page, you need to have marked it as 1461 * such (__GFP_COMP), or manually just split the page up yourself 1462 * (see split_page()). 1463 * 1464 * NOTE! Traditionally this was done with "remap_pfn_range()" which 1465 * took an arbitrary page protection parameter. This doesn't allow 1466 * that. Your vma protection will have to be set up correctly, which 1467 * means that if you want a shared writable mapping, you'd better 1468 * ask for a shared writable mapping! 1469 * 1470 * The page does not need to be reserved. 1471 * 1472 * Usually this function is called from f_op->mmap() handler 1473 * under mm->mmap_sem write-lock, so it can change vma->vm_flags. 1474 * Caller must set VM_MIXEDMAP on vma if it wants to call this 1475 * function from other places, for example from page-fault handler. 1476 * 1477 * Return: %0 on success, negative error code otherwise. 1478 */ 1479 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 1480 struct page *page) 1481 { 1482 if (addr < vma->vm_start || addr >= vma->vm_end) 1483 return -EFAULT; 1484 if (!page_count(page)) 1485 return -EINVAL; 1486 if (!(vma->vm_flags & VM_MIXEDMAP)) { 1487 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem)); 1488 BUG_ON(vma->vm_flags & VM_PFNMAP); 1489 vma->vm_flags |= VM_MIXEDMAP; 1490 } 1491 return insert_page(vma, addr, page, vma->vm_page_prot); 1492 } 1493 EXPORT_SYMBOL(vm_insert_page); 1494 1495 /* 1496 * __vm_map_pages - maps range of kernel pages into user vma 1497 * @vma: user vma to map to 1498 * @pages: pointer to array of source kernel pages 1499 * @num: number of pages in page array 1500 * @offset: user's requested vm_pgoff 1501 * 1502 * This allows drivers to map range of kernel pages into a user vma. 1503 * 1504 * Return: 0 on success and error code otherwise. 1505 */ 1506 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1507 unsigned long num, unsigned long offset) 1508 { 1509 unsigned long count = vma_pages(vma); 1510 unsigned long uaddr = vma->vm_start; 1511 int ret, i; 1512 1513 /* Fail if the user requested offset is beyond the end of the object */ 1514 if (offset >= num) 1515 return -ENXIO; 1516 1517 /* Fail if the user requested size exceeds available object size */ 1518 if (count > num - offset) 1519 return -ENXIO; 1520 1521 for (i = 0; i < count; i++) { 1522 ret = vm_insert_page(vma, uaddr, pages[offset + i]); 1523 if (ret < 0) 1524 return ret; 1525 uaddr += PAGE_SIZE; 1526 } 1527 1528 return 0; 1529 } 1530 1531 /** 1532 * vm_map_pages - maps range of kernel pages starts with non zero offset 1533 * @vma: user vma to map to 1534 * @pages: pointer to array of source kernel pages 1535 * @num: number of pages in page array 1536 * 1537 * Maps an object consisting of @num pages, catering for the user's 1538 * requested vm_pgoff 1539 * 1540 * If we fail to insert any page into the vma, the function will return 1541 * immediately leaving any previously inserted pages present. Callers 1542 * from the mmap handler may immediately return the error as their caller 1543 * will destroy the vma, removing any successfully inserted pages. Other 1544 * callers should make their own arrangements for calling unmap_region(). 1545 * 1546 * Context: Process context. Called by mmap handlers. 1547 * Return: 0 on success and error code otherwise. 1548 */ 1549 int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 1550 unsigned long num) 1551 { 1552 return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 1553 } 1554 EXPORT_SYMBOL(vm_map_pages); 1555 1556 /** 1557 * vm_map_pages_zero - map range of kernel pages starts with zero offset 1558 * @vma: user vma to map to 1559 * @pages: pointer to array of source kernel pages 1560 * @num: number of pages in page array 1561 * 1562 * Similar to vm_map_pages(), except that it explicitly sets the offset 1563 * to 0. This function is intended for the drivers that did not consider 1564 * vm_pgoff. 1565 * 1566 * Context: Process context. Called by mmap handlers. 1567 * Return: 0 on success and error code otherwise. 1568 */ 1569 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 1570 unsigned long num) 1571 { 1572 return __vm_map_pages(vma, pages, num, 0); 1573 } 1574 EXPORT_SYMBOL(vm_map_pages_zero); 1575 1576 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1577 pfn_t pfn, pgprot_t prot, bool mkwrite) 1578 { 1579 struct mm_struct *mm = vma->vm_mm; 1580 pte_t *pte, entry; 1581 spinlock_t *ptl; 1582 1583 pte = get_locked_pte(mm, addr, &ptl); 1584 if (!pte) 1585 return VM_FAULT_OOM; 1586 if (!pte_none(*pte)) { 1587 if (mkwrite) { 1588 /* 1589 * For read faults on private mappings the PFN passed 1590 * in may not match the PFN we have mapped if the 1591 * mapped PFN is a writeable COW page. In the mkwrite 1592 * case we are creating a writable PTE for a shared 1593 * mapping and we expect the PFNs to match. If they 1594 * don't match, we are likely racing with block 1595 * allocation and mapping invalidation so just skip the 1596 * update. 1597 */ 1598 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) { 1599 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte))); 1600 goto out_unlock; 1601 } 1602 entry = pte_mkyoung(*pte); 1603 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1604 if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 1605 update_mmu_cache(vma, addr, pte); 1606 } 1607 goto out_unlock; 1608 } 1609 1610 /* Ok, finally just insert the thing.. */ 1611 if (pfn_t_devmap(pfn)) 1612 entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 1613 else 1614 entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 1615 1616 if (mkwrite) { 1617 entry = pte_mkyoung(entry); 1618 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1619 } 1620 1621 set_pte_at(mm, addr, pte, entry); 1622 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 1623 1624 out_unlock: 1625 pte_unmap_unlock(pte, ptl); 1626 return VM_FAULT_NOPAGE; 1627 } 1628 1629 /** 1630 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 1631 * @vma: user vma to map to 1632 * @addr: target user address of this page 1633 * @pfn: source kernel pfn 1634 * @pgprot: pgprot flags for the inserted page 1635 * 1636 * This is exactly like vmf_insert_pfn(), except that it allows drivers to 1637 * to override pgprot on a per-page basis. 1638 * 1639 * This only makes sense for IO mappings, and it makes no sense for 1640 * COW mappings. In general, using multiple vmas is preferable; 1641 * vmf_insert_pfn_prot should only be used if using multiple VMAs is 1642 * impractical. 1643 * 1644 * Context: Process context. May allocate using %GFP_KERNEL. 1645 * Return: vm_fault_t value. 1646 */ 1647 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 1648 unsigned long pfn, pgprot_t pgprot) 1649 { 1650 /* 1651 * Technically, architectures with pte_special can avoid all these 1652 * restrictions (same for remap_pfn_range). However we would like 1653 * consistency in testing and feature parity among all, so we should 1654 * try to keep these invariants in place for everybody. 1655 */ 1656 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 1657 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1658 (VM_PFNMAP|VM_MIXEDMAP)); 1659 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1660 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 1661 1662 if (addr < vma->vm_start || addr >= vma->vm_end) 1663 return VM_FAULT_SIGBUS; 1664 1665 if (!pfn_modify_allowed(pfn, pgprot)) 1666 return VM_FAULT_SIGBUS; 1667 1668 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 1669 1670 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 1671 false); 1672 } 1673 EXPORT_SYMBOL(vmf_insert_pfn_prot); 1674 1675 /** 1676 * vmf_insert_pfn - insert single pfn into user vma 1677 * @vma: user vma to map to 1678 * @addr: target user address of this page 1679 * @pfn: source kernel pfn 1680 * 1681 * Similar to vm_insert_page, this allows drivers to insert individual pages 1682 * they've allocated into a user vma. Same comments apply. 1683 * 1684 * This function should only be called from a vm_ops->fault handler, and 1685 * in that case the handler should return the result of this function. 1686 * 1687 * vma cannot be a COW mapping. 1688 * 1689 * As this is called only for pages that do not currently exist, we 1690 * do not need to flush old virtual caches or the TLB. 1691 * 1692 * Context: Process context. May allocate using %GFP_KERNEL. 1693 * Return: vm_fault_t value. 1694 */ 1695 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 1696 unsigned long pfn) 1697 { 1698 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 1699 } 1700 EXPORT_SYMBOL(vmf_insert_pfn); 1701 1702 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 1703 { 1704 /* these checks mirror the abort conditions in vm_normal_page */ 1705 if (vma->vm_flags & VM_MIXEDMAP) 1706 return true; 1707 if (pfn_t_devmap(pfn)) 1708 return true; 1709 if (pfn_t_special(pfn)) 1710 return true; 1711 if (is_zero_pfn(pfn_t_to_pfn(pfn))) 1712 return true; 1713 return false; 1714 } 1715 1716 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 1717 unsigned long addr, pfn_t pfn, bool mkwrite) 1718 { 1719 pgprot_t pgprot = vma->vm_page_prot; 1720 int err; 1721 1722 BUG_ON(!vm_mixed_ok(vma, pfn)); 1723 1724 if (addr < vma->vm_start || addr >= vma->vm_end) 1725 return VM_FAULT_SIGBUS; 1726 1727 track_pfn_insert(vma, &pgprot, pfn); 1728 1729 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 1730 return VM_FAULT_SIGBUS; 1731 1732 /* 1733 * If we don't have pte special, then we have to use the pfn_valid() 1734 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 1735 * refcount the page if pfn_valid is true (hence insert_page rather 1736 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 1737 * without pte special, it would there be refcounted as a normal page. 1738 */ 1739 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 1740 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 1741 struct page *page; 1742 1743 /* 1744 * At this point we are committed to insert_page() 1745 * regardless of whether the caller specified flags that 1746 * result in pfn_t_has_page() == false. 1747 */ 1748 page = pfn_to_page(pfn_t_to_pfn(pfn)); 1749 err = insert_page(vma, addr, page, pgprot); 1750 } else { 1751 return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 1752 } 1753 1754 if (err == -ENOMEM) 1755 return VM_FAULT_OOM; 1756 if (err < 0 && err != -EBUSY) 1757 return VM_FAULT_SIGBUS; 1758 1759 return VM_FAULT_NOPAGE; 1760 } 1761 1762 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 1763 pfn_t pfn) 1764 { 1765 return __vm_insert_mixed(vma, addr, pfn, false); 1766 } 1767 EXPORT_SYMBOL(vmf_insert_mixed); 1768 1769 /* 1770 * If the insertion of PTE failed because someone else already added a 1771 * different entry in the mean time, we treat that as success as we assume 1772 * the same entry was actually inserted. 1773 */ 1774 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 1775 unsigned long addr, pfn_t pfn) 1776 { 1777 return __vm_insert_mixed(vma, addr, pfn, true); 1778 } 1779 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 1780 1781 /* 1782 * maps a range of physical memory into the requested pages. the old 1783 * mappings are removed. any references to nonexistent pages results 1784 * in null mappings (currently treated as "copy-on-access") 1785 */ 1786 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 1787 unsigned long addr, unsigned long end, 1788 unsigned long pfn, pgprot_t prot) 1789 { 1790 pte_t *pte; 1791 spinlock_t *ptl; 1792 int err = 0; 1793 1794 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 1795 if (!pte) 1796 return -ENOMEM; 1797 arch_enter_lazy_mmu_mode(); 1798 do { 1799 BUG_ON(!pte_none(*pte)); 1800 if (!pfn_modify_allowed(pfn, prot)) { 1801 err = -EACCES; 1802 break; 1803 } 1804 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 1805 pfn++; 1806 } while (pte++, addr += PAGE_SIZE, addr != end); 1807 arch_leave_lazy_mmu_mode(); 1808 pte_unmap_unlock(pte - 1, ptl); 1809 return err; 1810 } 1811 1812 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 1813 unsigned long addr, unsigned long end, 1814 unsigned long pfn, pgprot_t prot) 1815 { 1816 pmd_t *pmd; 1817 unsigned long next; 1818 int err; 1819 1820 pfn -= addr >> PAGE_SHIFT; 1821 pmd = pmd_alloc(mm, pud, addr); 1822 if (!pmd) 1823 return -ENOMEM; 1824 VM_BUG_ON(pmd_trans_huge(*pmd)); 1825 do { 1826 next = pmd_addr_end(addr, end); 1827 err = remap_pte_range(mm, pmd, addr, next, 1828 pfn + (addr >> PAGE_SHIFT), prot); 1829 if (err) 1830 return err; 1831 } while (pmd++, addr = next, addr != end); 1832 return 0; 1833 } 1834 1835 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 1836 unsigned long addr, unsigned long end, 1837 unsigned long pfn, pgprot_t prot) 1838 { 1839 pud_t *pud; 1840 unsigned long next; 1841 int err; 1842 1843 pfn -= addr >> PAGE_SHIFT; 1844 pud = pud_alloc(mm, p4d, addr); 1845 if (!pud) 1846 return -ENOMEM; 1847 do { 1848 next = pud_addr_end(addr, end); 1849 err = remap_pmd_range(mm, pud, addr, next, 1850 pfn + (addr >> PAGE_SHIFT), prot); 1851 if (err) 1852 return err; 1853 } while (pud++, addr = next, addr != end); 1854 return 0; 1855 } 1856 1857 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 1858 unsigned long addr, unsigned long end, 1859 unsigned long pfn, pgprot_t prot) 1860 { 1861 p4d_t *p4d; 1862 unsigned long next; 1863 int err; 1864 1865 pfn -= addr >> PAGE_SHIFT; 1866 p4d = p4d_alloc(mm, pgd, addr); 1867 if (!p4d) 1868 return -ENOMEM; 1869 do { 1870 next = p4d_addr_end(addr, end); 1871 err = remap_pud_range(mm, p4d, addr, next, 1872 pfn + (addr >> PAGE_SHIFT), prot); 1873 if (err) 1874 return err; 1875 } while (p4d++, addr = next, addr != end); 1876 return 0; 1877 } 1878 1879 /** 1880 * remap_pfn_range - remap kernel memory to userspace 1881 * @vma: user vma to map to 1882 * @addr: target user address to start at 1883 * @pfn: physical address of kernel memory 1884 * @size: size of map area 1885 * @prot: page protection flags for this mapping 1886 * 1887 * Note: this is only safe if the mm semaphore is held when called. 1888 * 1889 * Return: %0 on success, negative error code otherwise. 1890 */ 1891 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 1892 unsigned long pfn, unsigned long size, pgprot_t prot) 1893 { 1894 pgd_t *pgd; 1895 unsigned long next; 1896 unsigned long end = addr + PAGE_ALIGN(size); 1897 struct mm_struct *mm = vma->vm_mm; 1898 unsigned long remap_pfn = pfn; 1899 int err; 1900 1901 /* 1902 * Physically remapped pages are special. Tell the 1903 * rest of the world about it: 1904 * VM_IO tells people not to look at these pages 1905 * (accesses can have side effects). 1906 * VM_PFNMAP tells the core MM that the base pages are just 1907 * raw PFN mappings, and do not have a "struct page" associated 1908 * with them. 1909 * VM_DONTEXPAND 1910 * Disable vma merging and expanding with mremap(). 1911 * VM_DONTDUMP 1912 * Omit vma from core dump, even when VM_IO turned off. 1913 * 1914 * There's a horrible special case to handle copy-on-write 1915 * behaviour that some programs depend on. We mark the "original" 1916 * un-COW'ed pages by matching them up with "vma->vm_pgoff". 1917 * See vm_normal_page() for details. 1918 */ 1919 if (is_cow_mapping(vma->vm_flags)) { 1920 if (addr != vma->vm_start || end != vma->vm_end) 1921 return -EINVAL; 1922 vma->vm_pgoff = pfn; 1923 } 1924 1925 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size)); 1926 if (err) 1927 return -EINVAL; 1928 1929 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 1930 1931 BUG_ON(addr >= end); 1932 pfn -= addr >> PAGE_SHIFT; 1933 pgd = pgd_offset(mm, addr); 1934 flush_cache_range(vma, addr, end); 1935 do { 1936 next = pgd_addr_end(addr, end); 1937 err = remap_p4d_range(mm, pgd, addr, next, 1938 pfn + (addr >> PAGE_SHIFT), prot); 1939 if (err) 1940 break; 1941 } while (pgd++, addr = next, addr != end); 1942 1943 if (err) 1944 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size)); 1945 1946 return err; 1947 } 1948 EXPORT_SYMBOL(remap_pfn_range); 1949 1950 /** 1951 * vm_iomap_memory - remap memory to userspace 1952 * @vma: user vma to map to 1953 * @start: start of area 1954 * @len: size of area 1955 * 1956 * This is a simplified io_remap_pfn_range() for common driver use. The 1957 * driver just needs to give us the physical memory range to be mapped, 1958 * we'll figure out the rest from the vma information. 1959 * 1960 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 1961 * whatever write-combining details or similar. 1962 * 1963 * Return: %0 on success, negative error code otherwise. 1964 */ 1965 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 1966 { 1967 unsigned long vm_len, pfn, pages; 1968 1969 /* Check that the physical memory area passed in looks valid */ 1970 if (start + len < start) 1971 return -EINVAL; 1972 /* 1973 * You *really* shouldn't map things that aren't page-aligned, 1974 * but we've historically allowed it because IO memory might 1975 * just have smaller alignment. 1976 */ 1977 len += start & ~PAGE_MASK; 1978 pfn = start >> PAGE_SHIFT; 1979 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 1980 if (pfn + pages < pfn) 1981 return -EINVAL; 1982 1983 /* We start the mapping 'vm_pgoff' pages into the area */ 1984 if (vma->vm_pgoff > pages) 1985 return -EINVAL; 1986 pfn += vma->vm_pgoff; 1987 pages -= vma->vm_pgoff; 1988 1989 /* Can we fit all of the mapping? */ 1990 vm_len = vma->vm_end - vma->vm_start; 1991 if (vm_len >> PAGE_SHIFT > pages) 1992 return -EINVAL; 1993 1994 /* Ok, let it rip */ 1995 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 1996 } 1997 EXPORT_SYMBOL(vm_iomap_memory); 1998 1999 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2000 unsigned long addr, unsigned long end, 2001 pte_fn_t fn, void *data) 2002 { 2003 pte_t *pte; 2004 int err; 2005 spinlock_t *uninitialized_var(ptl); 2006 2007 pte = (mm == &init_mm) ? 2008 pte_alloc_kernel(pmd, addr) : 2009 pte_alloc_map_lock(mm, pmd, addr, &ptl); 2010 if (!pte) 2011 return -ENOMEM; 2012 2013 BUG_ON(pmd_huge(*pmd)); 2014 2015 arch_enter_lazy_mmu_mode(); 2016 2017 do { 2018 err = fn(pte++, addr, data); 2019 if (err) 2020 break; 2021 } while (addr += PAGE_SIZE, addr != end); 2022 2023 arch_leave_lazy_mmu_mode(); 2024 2025 if (mm != &init_mm) 2026 pte_unmap_unlock(pte-1, ptl); 2027 return err; 2028 } 2029 2030 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2031 unsigned long addr, unsigned long end, 2032 pte_fn_t fn, void *data) 2033 { 2034 pmd_t *pmd; 2035 unsigned long next; 2036 int err; 2037 2038 BUG_ON(pud_huge(*pud)); 2039 2040 pmd = pmd_alloc(mm, pud, addr); 2041 if (!pmd) 2042 return -ENOMEM; 2043 do { 2044 next = pmd_addr_end(addr, end); 2045 err = apply_to_pte_range(mm, pmd, addr, next, fn, data); 2046 if (err) 2047 break; 2048 } while (pmd++, addr = next, addr != end); 2049 return err; 2050 } 2051 2052 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2053 unsigned long addr, unsigned long end, 2054 pte_fn_t fn, void *data) 2055 { 2056 pud_t *pud; 2057 unsigned long next; 2058 int err; 2059 2060 pud = pud_alloc(mm, p4d, addr); 2061 if (!pud) 2062 return -ENOMEM; 2063 do { 2064 next = pud_addr_end(addr, end); 2065 err = apply_to_pmd_range(mm, pud, addr, next, fn, data); 2066 if (err) 2067 break; 2068 } while (pud++, addr = next, addr != end); 2069 return err; 2070 } 2071 2072 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2073 unsigned long addr, unsigned long end, 2074 pte_fn_t fn, void *data) 2075 { 2076 p4d_t *p4d; 2077 unsigned long next; 2078 int err; 2079 2080 p4d = p4d_alloc(mm, pgd, addr); 2081 if (!p4d) 2082 return -ENOMEM; 2083 do { 2084 next = p4d_addr_end(addr, end); 2085 err = apply_to_pud_range(mm, p4d, addr, next, fn, data); 2086 if (err) 2087 break; 2088 } while (p4d++, addr = next, addr != end); 2089 return err; 2090 } 2091 2092 /* 2093 * Scan a region of virtual memory, filling in page tables as necessary 2094 * and calling a provided function on each leaf page table. 2095 */ 2096 int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2097 unsigned long size, pte_fn_t fn, void *data) 2098 { 2099 pgd_t *pgd; 2100 unsigned long next; 2101 unsigned long end = addr + size; 2102 int err; 2103 2104 if (WARN_ON(addr >= end)) 2105 return -EINVAL; 2106 2107 pgd = pgd_offset(mm, addr); 2108 do { 2109 next = pgd_addr_end(addr, end); 2110 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data); 2111 if (err) 2112 break; 2113 } while (pgd++, addr = next, addr != end); 2114 2115 return err; 2116 } 2117 EXPORT_SYMBOL_GPL(apply_to_page_range); 2118 2119 /* 2120 * handle_pte_fault chooses page fault handler according to an entry which was 2121 * read non-atomically. Before making any commitment, on those architectures 2122 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 2123 * parts, do_swap_page must check under lock before unmapping the pte and 2124 * proceeding (but do_wp_page is only called after already making such a check; 2125 * and do_anonymous_page can safely check later on). 2126 */ 2127 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, 2128 pte_t *page_table, pte_t orig_pte) 2129 { 2130 int same = 1; 2131 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) 2132 if (sizeof(pte_t) > sizeof(unsigned long)) { 2133 spinlock_t *ptl = pte_lockptr(mm, pmd); 2134 spin_lock(ptl); 2135 same = pte_same(*page_table, orig_pte); 2136 spin_unlock(ptl); 2137 } 2138 #endif 2139 pte_unmap(page_table); 2140 return same; 2141 } 2142 2143 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) 2144 { 2145 debug_dma_assert_idle(src); 2146 2147 /* 2148 * If the source page was a PFN mapping, we don't have 2149 * a "struct page" for it. We do a best-effort copy by 2150 * just copying from the original user address. If that 2151 * fails, we just zero-fill it. Live with it. 2152 */ 2153 if (unlikely(!src)) { 2154 void *kaddr = kmap_atomic(dst); 2155 void __user *uaddr = (void __user *)(va & PAGE_MASK); 2156 2157 /* 2158 * This really shouldn't fail, because the page is there 2159 * in the page tables. But it might just be unreadable, 2160 * in which case we just give up and fill the result with 2161 * zeroes. 2162 */ 2163 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) 2164 clear_page(kaddr); 2165 kunmap_atomic(kaddr); 2166 flush_dcache_page(dst); 2167 } else 2168 copy_user_highpage(dst, src, va, vma); 2169 } 2170 2171 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 2172 { 2173 struct file *vm_file = vma->vm_file; 2174 2175 if (vm_file) 2176 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 2177 2178 /* 2179 * Special mappings (e.g. VDSO) do not have any file so fake 2180 * a default GFP_KERNEL for them. 2181 */ 2182 return GFP_KERNEL; 2183 } 2184 2185 /* 2186 * Notify the address space that the page is about to become writable so that 2187 * it can prohibit this or wait for the page to get into an appropriate state. 2188 * 2189 * We do this without the lock held, so that it can sleep if it needs to. 2190 */ 2191 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf) 2192 { 2193 vm_fault_t ret; 2194 struct page *page = vmf->page; 2195 unsigned int old_flags = vmf->flags; 2196 2197 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 2198 2199 ret = vmf->vma->vm_ops->page_mkwrite(vmf); 2200 /* Restore original flags so that caller is not surprised */ 2201 vmf->flags = old_flags; 2202 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 2203 return ret; 2204 if (unlikely(!(ret & VM_FAULT_LOCKED))) { 2205 lock_page(page); 2206 if (!page->mapping) { 2207 unlock_page(page); 2208 return 0; /* retry */ 2209 } 2210 ret |= VM_FAULT_LOCKED; 2211 } else 2212 VM_BUG_ON_PAGE(!PageLocked(page), page); 2213 return ret; 2214 } 2215 2216 /* 2217 * Handle dirtying of a page in shared file mapping on a write fault. 2218 * 2219 * The function expects the page to be locked and unlocks it. 2220 */ 2221 static void fault_dirty_shared_page(struct vm_area_struct *vma, 2222 struct page *page) 2223 { 2224 struct address_space *mapping; 2225 bool dirtied; 2226 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 2227 2228 dirtied = set_page_dirty(page); 2229 VM_BUG_ON_PAGE(PageAnon(page), page); 2230 /* 2231 * Take a local copy of the address_space - page.mapping may be zeroed 2232 * by truncate after unlock_page(). The address_space itself remains 2233 * pinned by vma->vm_file's reference. We rely on unlock_page()'s 2234 * release semantics to prevent the compiler from undoing this copying. 2235 */ 2236 mapping = page_rmapping(page); 2237 unlock_page(page); 2238 2239 if ((dirtied || page_mkwrite) && mapping) { 2240 /* 2241 * Some device drivers do not set page.mapping 2242 * but still dirty their pages 2243 */ 2244 balance_dirty_pages_ratelimited(mapping); 2245 } 2246 2247 if (!page_mkwrite) 2248 file_update_time(vma->vm_file); 2249 } 2250 2251 /* 2252 * Handle write page faults for pages that can be reused in the current vma 2253 * 2254 * This can happen either due to the mapping being with the VM_SHARED flag, 2255 * or due to us being the last reference standing to the page. In either 2256 * case, all we need to do here is to mark the page as writable and update 2257 * any related book-keeping. 2258 */ 2259 static inline void wp_page_reuse(struct vm_fault *vmf) 2260 __releases(vmf->ptl) 2261 { 2262 struct vm_area_struct *vma = vmf->vma; 2263 struct page *page = vmf->page; 2264 pte_t entry; 2265 /* 2266 * Clear the pages cpupid information as the existing 2267 * information potentially belongs to a now completely 2268 * unrelated process. 2269 */ 2270 if (page) 2271 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1); 2272 2273 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2274 entry = pte_mkyoung(vmf->orig_pte); 2275 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2276 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 2277 update_mmu_cache(vma, vmf->address, vmf->pte); 2278 pte_unmap_unlock(vmf->pte, vmf->ptl); 2279 } 2280 2281 /* 2282 * Handle the case of a page which we actually need to copy to a new page. 2283 * 2284 * Called with mmap_sem locked and the old page referenced, but 2285 * without the ptl held. 2286 * 2287 * High level logic flow: 2288 * 2289 * - Allocate a page, copy the content of the old page to the new one. 2290 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 2291 * - Take the PTL. If the pte changed, bail out and release the allocated page 2292 * - If the pte is still the way we remember it, update the page table and all 2293 * relevant references. This includes dropping the reference the page-table 2294 * held to the old page, as well as updating the rmap. 2295 * - In any case, unlock the PTL and drop the reference we took to the old page. 2296 */ 2297 static vm_fault_t wp_page_copy(struct vm_fault *vmf) 2298 { 2299 struct vm_area_struct *vma = vmf->vma; 2300 struct mm_struct *mm = vma->vm_mm; 2301 struct page *old_page = vmf->page; 2302 struct page *new_page = NULL; 2303 pte_t entry; 2304 int page_copied = 0; 2305 struct mem_cgroup *memcg; 2306 struct mmu_notifier_range range; 2307 2308 if (unlikely(anon_vma_prepare(vma))) 2309 goto oom; 2310 2311 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) { 2312 new_page = alloc_zeroed_user_highpage_movable(vma, 2313 vmf->address); 2314 if (!new_page) 2315 goto oom; 2316 } else { 2317 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 2318 vmf->address); 2319 if (!new_page) 2320 goto oom; 2321 cow_user_page(new_page, old_page, vmf->address, vma); 2322 } 2323 2324 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false)) 2325 goto oom_free_new; 2326 2327 __SetPageUptodate(new_page); 2328 2329 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 2330 vmf->address & PAGE_MASK, 2331 (vmf->address & PAGE_MASK) + PAGE_SIZE); 2332 mmu_notifier_invalidate_range_start(&range); 2333 2334 /* 2335 * Re-check the pte - we dropped the lock 2336 */ 2337 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 2338 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) { 2339 if (old_page) { 2340 if (!PageAnon(old_page)) { 2341 dec_mm_counter_fast(mm, 2342 mm_counter_file(old_page)); 2343 inc_mm_counter_fast(mm, MM_ANONPAGES); 2344 } 2345 } else { 2346 inc_mm_counter_fast(mm, MM_ANONPAGES); 2347 } 2348 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 2349 entry = mk_pte(new_page, vma->vm_page_prot); 2350 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2351 /* 2352 * Clear the pte entry and flush it first, before updating the 2353 * pte with the new entry. This will avoid a race condition 2354 * seen in the presence of one thread doing SMC and another 2355 * thread doing COW. 2356 */ 2357 ptep_clear_flush_notify(vma, vmf->address, vmf->pte); 2358 page_add_new_anon_rmap(new_page, vma, vmf->address, false); 2359 mem_cgroup_commit_charge(new_page, memcg, false, false); 2360 lru_cache_add_active_or_unevictable(new_page, vma); 2361 /* 2362 * We call the notify macro here because, when using secondary 2363 * mmu page tables (such as kvm shadow page tables), we want the 2364 * new page to be mapped directly into the secondary page table. 2365 */ 2366 set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 2367 update_mmu_cache(vma, vmf->address, vmf->pte); 2368 if (old_page) { 2369 /* 2370 * Only after switching the pte to the new page may 2371 * we remove the mapcount here. Otherwise another 2372 * process may come and find the rmap count decremented 2373 * before the pte is switched to the new page, and 2374 * "reuse" the old page writing into it while our pte 2375 * here still points into it and can be read by other 2376 * threads. 2377 * 2378 * The critical issue is to order this 2379 * page_remove_rmap with the ptp_clear_flush above. 2380 * Those stores are ordered by (if nothing else,) 2381 * the barrier present in the atomic_add_negative 2382 * in page_remove_rmap. 2383 * 2384 * Then the TLB flush in ptep_clear_flush ensures that 2385 * no process can access the old page before the 2386 * decremented mapcount is visible. And the old page 2387 * cannot be reused until after the decremented 2388 * mapcount is visible. So transitively, TLBs to 2389 * old page will be flushed before it can be reused. 2390 */ 2391 page_remove_rmap(old_page, false); 2392 } 2393 2394 /* Free the old page.. */ 2395 new_page = old_page; 2396 page_copied = 1; 2397 } else { 2398 mem_cgroup_cancel_charge(new_page, memcg, false); 2399 } 2400 2401 if (new_page) 2402 put_page(new_page); 2403 2404 pte_unmap_unlock(vmf->pte, vmf->ptl); 2405 /* 2406 * No need to double call mmu_notifier->invalidate_range() callback as 2407 * the above ptep_clear_flush_notify() did already call it. 2408 */ 2409 mmu_notifier_invalidate_range_only_end(&range); 2410 if (old_page) { 2411 /* 2412 * Don't let another task, with possibly unlocked vma, 2413 * keep the mlocked page. 2414 */ 2415 if (page_copied && (vma->vm_flags & VM_LOCKED)) { 2416 lock_page(old_page); /* LRU manipulation */ 2417 if (PageMlocked(old_page)) 2418 munlock_vma_page(old_page); 2419 unlock_page(old_page); 2420 } 2421 put_page(old_page); 2422 } 2423 return page_copied ? VM_FAULT_WRITE : 0; 2424 oom_free_new: 2425 put_page(new_page); 2426 oom: 2427 if (old_page) 2428 put_page(old_page); 2429 return VM_FAULT_OOM; 2430 } 2431 2432 /** 2433 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 2434 * writeable once the page is prepared 2435 * 2436 * @vmf: structure describing the fault 2437 * 2438 * This function handles all that is needed to finish a write page fault in a 2439 * shared mapping due to PTE being read-only once the mapped page is prepared. 2440 * It handles locking of PTE and modifying it. 2441 * 2442 * The function expects the page to be locked or other protection against 2443 * concurrent faults / writeback (such as DAX radix tree locks). 2444 * 2445 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before 2446 * we acquired PTE lock. 2447 */ 2448 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf) 2449 { 2450 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 2451 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 2452 &vmf->ptl); 2453 /* 2454 * We might have raced with another page fault while we released the 2455 * pte_offset_map_lock. 2456 */ 2457 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 2458 pte_unmap_unlock(vmf->pte, vmf->ptl); 2459 return VM_FAULT_NOPAGE; 2460 } 2461 wp_page_reuse(vmf); 2462 return 0; 2463 } 2464 2465 /* 2466 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 2467 * mapping 2468 */ 2469 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 2470 { 2471 struct vm_area_struct *vma = vmf->vma; 2472 2473 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 2474 vm_fault_t ret; 2475 2476 pte_unmap_unlock(vmf->pte, vmf->ptl); 2477 vmf->flags |= FAULT_FLAG_MKWRITE; 2478 ret = vma->vm_ops->pfn_mkwrite(vmf); 2479 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 2480 return ret; 2481 return finish_mkwrite_fault(vmf); 2482 } 2483 wp_page_reuse(vmf); 2484 return VM_FAULT_WRITE; 2485 } 2486 2487 static vm_fault_t wp_page_shared(struct vm_fault *vmf) 2488 __releases(vmf->ptl) 2489 { 2490 struct vm_area_struct *vma = vmf->vma; 2491 2492 get_page(vmf->page); 2493 2494 if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 2495 vm_fault_t tmp; 2496 2497 pte_unmap_unlock(vmf->pte, vmf->ptl); 2498 tmp = do_page_mkwrite(vmf); 2499 if (unlikely(!tmp || (tmp & 2500 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 2501 put_page(vmf->page); 2502 return tmp; 2503 } 2504 tmp = finish_mkwrite_fault(vmf); 2505 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 2506 unlock_page(vmf->page); 2507 put_page(vmf->page); 2508 return tmp; 2509 } 2510 } else { 2511 wp_page_reuse(vmf); 2512 lock_page(vmf->page); 2513 } 2514 fault_dirty_shared_page(vma, vmf->page); 2515 put_page(vmf->page); 2516 2517 return VM_FAULT_WRITE; 2518 } 2519 2520 /* 2521 * This routine handles present pages, when users try to write 2522 * to a shared page. It is done by copying the page to a new address 2523 * and decrementing the shared-page counter for the old page. 2524 * 2525 * Note that this routine assumes that the protection checks have been 2526 * done by the caller (the low-level page fault routine in most cases). 2527 * Thus we can safely just mark it writable once we've done any necessary 2528 * COW. 2529 * 2530 * We also mark the page dirty at this point even though the page will 2531 * change only once the write actually happens. This avoids a few races, 2532 * and potentially makes it more efficient. 2533 * 2534 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2535 * but allow concurrent faults), with pte both mapped and locked. 2536 * We return with mmap_sem still held, but pte unmapped and unlocked. 2537 */ 2538 static vm_fault_t do_wp_page(struct vm_fault *vmf) 2539 __releases(vmf->ptl) 2540 { 2541 struct vm_area_struct *vma = vmf->vma; 2542 2543 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 2544 if (!vmf->page) { 2545 /* 2546 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 2547 * VM_PFNMAP VMA. 2548 * 2549 * We should not cow pages in a shared writeable mapping. 2550 * Just mark the pages writable and/or call ops->pfn_mkwrite. 2551 */ 2552 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2553 (VM_WRITE|VM_SHARED)) 2554 return wp_pfn_shared(vmf); 2555 2556 pte_unmap_unlock(vmf->pte, vmf->ptl); 2557 return wp_page_copy(vmf); 2558 } 2559 2560 /* 2561 * Take out anonymous pages first, anonymous shared vmas are 2562 * not dirty accountable. 2563 */ 2564 if (PageAnon(vmf->page)) { 2565 int total_map_swapcount; 2566 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) || 2567 page_count(vmf->page) != 1)) 2568 goto copy; 2569 if (!trylock_page(vmf->page)) { 2570 get_page(vmf->page); 2571 pte_unmap_unlock(vmf->pte, vmf->ptl); 2572 lock_page(vmf->page); 2573 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 2574 vmf->address, &vmf->ptl); 2575 if (!pte_same(*vmf->pte, vmf->orig_pte)) { 2576 unlock_page(vmf->page); 2577 pte_unmap_unlock(vmf->pte, vmf->ptl); 2578 put_page(vmf->page); 2579 return 0; 2580 } 2581 put_page(vmf->page); 2582 } 2583 if (PageKsm(vmf->page)) { 2584 bool reused = reuse_ksm_page(vmf->page, vmf->vma, 2585 vmf->address); 2586 unlock_page(vmf->page); 2587 if (!reused) 2588 goto copy; 2589 wp_page_reuse(vmf); 2590 return VM_FAULT_WRITE; 2591 } 2592 if (reuse_swap_page(vmf->page, &total_map_swapcount)) { 2593 if (total_map_swapcount == 1) { 2594 /* 2595 * The page is all ours. Move it to 2596 * our anon_vma so the rmap code will 2597 * not search our parent or siblings. 2598 * Protected against the rmap code by 2599 * the page lock. 2600 */ 2601 page_move_anon_rmap(vmf->page, vma); 2602 } 2603 unlock_page(vmf->page); 2604 wp_page_reuse(vmf); 2605 return VM_FAULT_WRITE; 2606 } 2607 unlock_page(vmf->page); 2608 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == 2609 (VM_WRITE|VM_SHARED))) { 2610 return wp_page_shared(vmf); 2611 } 2612 copy: 2613 /* 2614 * Ok, we need to copy. Oh, well.. 2615 */ 2616 get_page(vmf->page); 2617 2618 pte_unmap_unlock(vmf->pte, vmf->ptl); 2619 return wp_page_copy(vmf); 2620 } 2621 2622 static void unmap_mapping_range_vma(struct vm_area_struct *vma, 2623 unsigned long start_addr, unsigned long end_addr, 2624 struct zap_details *details) 2625 { 2626 zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 2627 } 2628 2629 static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 2630 struct zap_details *details) 2631 { 2632 struct vm_area_struct *vma; 2633 pgoff_t vba, vea, zba, zea; 2634 2635 vma_interval_tree_foreach(vma, root, 2636 details->first_index, details->last_index) { 2637 2638 vba = vma->vm_pgoff; 2639 vea = vba + vma_pages(vma) - 1; 2640 zba = details->first_index; 2641 if (zba < vba) 2642 zba = vba; 2643 zea = details->last_index; 2644 if (zea > vea) 2645 zea = vea; 2646 2647 unmap_mapping_range_vma(vma, 2648 ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 2649 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 2650 details); 2651 } 2652 } 2653 2654 /** 2655 * unmap_mapping_pages() - Unmap pages from processes. 2656 * @mapping: The address space containing pages to be unmapped. 2657 * @start: Index of first page to be unmapped. 2658 * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 2659 * @even_cows: Whether to unmap even private COWed pages. 2660 * 2661 * Unmap the pages in this address space from any userspace process which 2662 * has them mmaped. Generally, you want to remove COWed pages as well when 2663 * a file is being truncated, but not when invalidating pages from the page 2664 * cache. 2665 */ 2666 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 2667 pgoff_t nr, bool even_cows) 2668 { 2669 struct zap_details details = { }; 2670 2671 details.check_mapping = even_cows ? NULL : mapping; 2672 details.first_index = start; 2673 details.last_index = start + nr - 1; 2674 if (details.last_index < details.first_index) 2675 details.last_index = ULONG_MAX; 2676 2677 i_mmap_lock_write(mapping); 2678 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 2679 unmap_mapping_range_tree(&mapping->i_mmap, &details); 2680 i_mmap_unlock_write(mapping); 2681 } 2682 2683 /** 2684 * unmap_mapping_range - unmap the portion of all mmaps in the specified 2685 * address_space corresponding to the specified byte range in the underlying 2686 * file. 2687 * 2688 * @mapping: the address space containing mmaps to be unmapped. 2689 * @holebegin: byte in first page to unmap, relative to the start of 2690 * the underlying file. This will be rounded down to a PAGE_SIZE 2691 * boundary. Note that this is different from truncate_pagecache(), which 2692 * must keep the partial page. In contrast, we must get rid of 2693 * partial pages. 2694 * @holelen: size of prospective hole in bytes. This will be rounded 2695 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 2696 * end of the file. 2697 * @even_cows: 1 when truncating a file, unmap even private COWed pages; 2698 * but 0 when invalidating pagecache, don't throw away private data. 2699 */ 2700 void unmap_mapping_range(struct address_space *mapping, 2701 loff_t const holebegin, loff_t const holelen, int even_cows) 2702 { 2703 pgoff_t hba = holebegin >> PAGE_SHIFT; 2704 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2705 2706 /* Check for overflow. */ 2707 if (sizeof(holelen) > sizeof(hlen)) { 2708 long long holeend = 2709 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 2710 if (holeend & ~(long long)ULONG_MAX) 2711 hlen = ULONG_MAX - hba + 1; 2712 } 2713 2714 unmap_mapping_pages(mapping, hba, hlen, even_cows); 2715 } 2716 EXPORT_SYMBOL(unmap_mapping_range); 2717 2718 /* 2719 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2720 * but allow concurrent faults), and pte mapped but not yet locked. 2721 * We return with pte unmapped and unlocked. 2722 * 2723 * We return with the mmap_sem locked or unlocked in the same cases 2724 * as does filemap_fault(). 2725 */ 2726 vm_fault_t do_swap_page(struct vm_fault *vmf) 2727 { 2728 struct vm_area_struct *vma = vmf->vma; 2729 struct page *page = NULL, *swapcache; 2730 struct mem_cgroup *memcg; 2731 swp_entry_t entry; 2732 pte_t pte; 2733 int locked; 2734 int exclusive = 0; 2735 vm_fault_t ret = 0; 2736 2737 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) 2738 goto out; 2739 2740 entry = pte_to_swp_entry(vmf->orig_pte); 2741 if (unlikely(non_swap_entry(entry))) { 2742 if (is_migration_entry(entry)) { 2743 migration_entry_wait(vma->vm_mm, vmf->pmd, 2744 vmf->address); 2745 } else if (is_device_private_entry(entry)) { 2746 vmf->page = device_private_entry_to_page(entry); 2747 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 2748 } else if (is_hwpoison_entry(entry)) { 2749 ret = VM_FAULT_HWPOISON; 2750 } else { 2751 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 2752 ret = VM_FAULT_SIGBUS; 2753 } 2754 goto out; 2755 } 2756 2757 2758 delayacct_set_flag(DELAYACCT_PF_SWAPIN); 2759 page = lookup_swap_cache(entry, vma, vmf->address); 2760 swapcache = page; 2761 2762 if (!page) { 2763 struct swap_info_struct *si = swp_swap_info(entry); 2764 2765 if (si->flags & SWP_SYNCHRONOUS_IO && 2766 __swap_count(entry) == 1) { 2767 /* skip swapcache */ 2768 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, 2769 vmf->address); 2770 if (page) { 2771 __SetPageLocked(page); 2772 __SetPageSwapBacked(page); 2773 set_page_private(page, entry.val); 2774 lru_cache_add_anon(page); 2775 swap_readpage(page, true); 2776 } 2777 } else { 2778 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 2779 vmf); 2780 swapcache = page; 2781 } 2782 2783 if (!page) { 2784 /* 2785 * Back out if somebody else faulted in this pte 2786 * while we released the pte lock. 2787 */ 2788 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 2789 vmf->address, &vmf->ptl); 2790 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) 2791 ret = VM_FAULT_OOM; 2792 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2793 goto unlock; 2794 } 2795 2796 /* Had to read the page from swap area: Major fault */ 2797 ret = VM_FAULT_MAJOR; 2798 count_vm_event(PGMAJFAULT); 2799 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 2800 } else if (PageHWPoison(page)) { 2801 /* 2802 * hwpoisoned dirty swapcache pages are kept for killing 2803 * owner processes (which may be unknown at hwpoison time) 2804 */ 2805 ret = VM_FAULT_HWPOISON; 2806 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2807 goto out_release; 2808 } 2809 2810 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags); 2811 2812 delayacct_clear_flag(DELAYACCT_PF_SWAPIN); 2813 if (!locked) { 2814 ret |= VM_FAULT_RETRY; 2815 goto out_release; 2816 } 2817 2818 /* 2819 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not 2820 * release the swapcache from under us. The page pin, and pte_same 2821 * test below, are not enough to exclude that. Even if it is still 2822 * swapcache, we need to check that the page's swap has not changed. 2823 */ 2824 if (unlikely((!PageSwapCache(page) || 2825 page_private(page) != entry.val)) && swapcache) 2826 goto out_page; 2827 2828 page = ksm_might_need_to_copy(page, vma, vmf->address); 2829 if (unlikely(!page)) { 2830 ret = VM_FAULT_OOM; 2831 page = swapcache; 2832 goto out_page; 2833 } 2834 2835 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, 2836 &memcg, false)) { 2837 ret = VM_FAULT_OOM; 2838 goto out_page; 2839 } 2840 2841 /* 2842 * Back out if somebody else already faulted in this pte. 2843 */ 2844 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 2845 &vmf->ptl); 2846 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) 2847 goto out_nomap; 2848 2849 if (unlikely(!PageUptodate(page))) { 2850 ret = VM_FAULT_SIGBUS; 2851 goto out_nomap; 2852 } 2853 2854 /* 2855 * The page isn't present yet, go ahead with the fault. 2856 * 2857 * Be careful about the sequence of operations here. 2858 * To get its accounting right, reuse_swap_page() must be called 2859 * while the page is counted on swap but not yet in mapcount i.e. 2860 * before page_add_anon_rmap() and swap_free(); try_to_free_swap() 2861 * must be called after the swap_free(), or it will never succeed. 2862 */ 2863 2864 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 2865 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS); 2866 pte = mk_pte(page, vma->vm_page_prot); 2867 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) { 2868 pte = maybe_mkwrite(pte_mkdirty(pte), vma); 2869 vmf->flags &= ~FAULT_FLAG_WRITE; 2870 ret |= VM_FAULT_WRITE; 2871 exclusive = RMAP_EXCLUSIVE; 2872 } 2873 flush_icache_page(vma, page); 2874 if (pte_swp_soft_dirty(vmf->orig_pte)) 2875 pte = pte_mksoft_dirty(pte); 2876 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 2877 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 2878 vmf->orig_pte = pte; 2879 2880 /* ksm created a completely new copy */ 2881 if (unlikely(page != swapcache && swapcache)) { 2882 page_add_new_anon_rmap(page, vma, vmf->address, false); 2883 mem_cgroup_commit_charge(page, memcg, false, false); 2884 lru_cache_add_active_or_unevictable(page, vma); 2885 } else { 2886 do_page_add_anon_rmap(page, vma, vmf->address, exclusive); 2887 mem_cgroup_commit_charge(page, memcg, true, false); 2888 activate_page(page); 2889 } 2890 2891 swap_free(entry); 2892 if (mem_cgroup_swap_full(page) || 2893 (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) 2894 try_to_free_swap(page); 2895 unlock_page(page); 2896 if (page != swapcache && swapcache) { 2897 /* 2898 * Hold the lock to avoid the swap entry to be reused 2899 * until we take the PT lock for the pte_same() check 2900 * (to avoid false positives from pte_same). For 2901 * further safety release the lock after the swap_free 2902 * so that the swap count won't change under a 2903 * parallel locked swapcache. 2904 */ 2905 unlock_page(swapcache); 2906 put_page(swapcache); 2907 } 2908 2909 if (vmf->flags & FAULT_FLAG_WRITE) { 2910 ret |= do_wp_page(vmf); 2911 if (ret & VM_FAULT_ERROR) 2912 ret &= VM_FAULT_ERROR; 2913 goto out; 2914 } 2915 2916 /* No need to invalidate - it was non-present before */ 2917 update_mmu_cache(vma, vmf->address, vmf->pte); 2918 unlock: 2919 pte_unmap_unlock(vmf->pte, vmf->ptl); 2920 out: 2921 return ret; 2922 out_nomap: 2923 mem_cgroup_cancel_charge(page, memcg, false); 2924 pte_unmap_unlock(vmf->pte, vmf->ptl); 2925 out_page: 2926 unlock_page(page); 2927 out_release: 2928 put_page(page); 2929 if (page != swapcache && swapcache) { 2930 unlock_page(swapcache); 2931 put_page(swapcache); 2932 } 2933 return ret; 2934 } 2935 2936 /* 2937 * We enter with non-exclusive mmap_sem (to exclude vma changes, 2938 * but allow concurrent faults), and pte mapped but not yet locked. 2939 * We return with mmap_sem still held, but pte unmapped and unlocked. 2940 */ 2941 static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 2942 { 2943 struct vm_area_struct *vma = vmf->vma; 2944 struct mem_cgroup *memcg; 2945 struct page *page; 2946 vm_fault_t ret = 0; 2947 pte_t entry; 2948 2949 /* File mapping without ->vm_ops ? */ 2950 if (vma->vm_flags & VM_SHARED) 2951 return VM_FAULT_SIGBUS; 2952 2953 /* 2954 * Use pte_alloc() instead of pte_alloc_map(). We can't run 2955 * pte_offset_map() on pmds where a huge pmd might be created 2956 * from a different thread. 2957 * 2958 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when 2959 * parallel threads are excluded by other means. 2960 * 2961 * Here we only have down_read(mmap_sem). 2962 */ 2963 if (pte_alloc(vma->vm_mm, vmf->pmd)) 2964 return VM_FAULT_OOM; 2965 2966 /* See the comment in pte_alloc_one_map() */ 2967 if (unlikely(pmd_trans_unstable(vmf->pmd))) 2968 return 0; 2969 2970 /* Use the zero-page for reads */ 2971 if (!(vmf->flags & FAULT_FLAG_WRITE) && 2972 !mm_forbids_zeropage(vma->vm_mm)) { 2973 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 2974 vma->vm_page_prot)); 2975 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 2976 vmf->address, &vmf->ptl); 2977 if (!pte_none(*vmf->pte)) 2978 goto unlock; 2979 ret = check_stable_address_space(vma->vm_mm); 2980 if (ret) 2981 goto unlock; 2982 /* Deliver the page fault to userland, check inside PT lock */ 2983 if (userfaultfd_missing(vma)) { 2984 pte_unmap_unlock(vmf->pte, vmf->ptl); 2985 return handle_userfault(vmf, VM_UFFD_MISSING); 2986 } 2987 goto setpte; 2988 } 2989 2990 /* Allocate our own private page. */ 2991 if (unlikely(anon_vma_prepare(vma))) 2992 goto oom; 2993 page = alloc_zeroed_user_highpage_movable(vma, vmf->address); 2994 if (!page) 2995 goto oom; 2996 2997 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg, 2998 false)) 2999 goto oom_free_page; 3000 3001 /* 3002 * The memory barrier inside __SetPageUptodate makes sure that 3003 * preceeding stores to the page contents become visible before 3004 * the set_pte_at() write. 3005 */ 3006 __SetPageUptodate(page); 3007 3008 entry = mk_pte(page, vma->vm_page_prot); 3009 if (vma->vm_flags & VM_WRITE) 3010 entry = pte_mkwrite(pte_mkdirty(entry)); 3011 3012 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3013 &vmf->ptl); 3014 if (!pte_none(*vmf->pte)) 3015 goto release; 3016 3017 ret = check_stable_address_space(vma->vm_mm); 3018 if (ret) 3019 goto release; 3020 3021 /* Deliver the page fault to userland, check inside PT lock */ 3022 if (userfaultfd_missing(vma)) { 3023 pte_unmap_unlock(vmf->pte, vmf->ptl); 3024 mem_cgroup_cancel_charge(page, memcg, false); 3025 put_page(page); 3026 return handle_userfault(vmf, VM_UFFD_MISSING); 3027 } 3028 3029 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3030 page_add_new_anon_rmap(page, vma, vmf->address, false); 3031 mem_cgroup_commit_charge(page, memcg, false, false); 3032 lru_cache_add_active_or_unevictable(page, vma); 3033 setpte: 3034 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3035 3036 /* No need to invalidate - it was non-present before */ 3037 update_mmu_cache(vma, vmf->address, vmf->pte); 3038 unlock: 3039 pte_unmap_unlock(vmf->pte, vmf->ptl); 3040 return ret; 3041 release: 3042 mem_cgroup_cancel_charge(page, memcg, false); 3043 put_page(page); 3044 goto unlock; 3045 oom_free_page: 3046 put_page(page); 3047 oom: 3048 return VM_FAULT_OOM; 3049 } 3050 3051 /* 3052 * The mmap_sem must have been held on entry, and may have been 3053 * released depending on flags and vma->vm_ops->fault() return value. 3054 * See filemap_fault() and __lock_page_retry(). 3055 */ 3056 static vm_fault_t __do_fault(struct vm_fault *vmf) 3057 { 3058 struct vm_area_struct *vma = vmf->vma; 3059 vm_fault_t ret; 3060 3061 /* 3062 * Preallocate pte before we take page_lock because this might lead to 3063 * deadlocks for memcg reclaim which waits for pages under writeback: 3064 * lock_page(A) 3065 * SetPageWriteback(A) 3066 * unlock_page(A) 3067 * lock_page(B) 3068 * lock_page(B) 3069 * pte_alloc_pne 3070 * shrink_page_list 3071 * wait_on_page_writeback(A) 3072 * SetPageWriteback(B) 3073 * unlock_page(B) 3074 * # flush A, B to clear the writeback 3075 */ 3076 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 3077 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 3078 if (!vmf->prealloc_pte) 3079 return VM_FAULT_OOM; 3080 smp_wmb(); /* See comment in __pte_alloc() */ 3081 } 3082 3083 ret = vma->vm_ops->fault(vmf); 3084 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 3085 VM_FAULT_DONE_COW))) 3086 return ret; 3087 3088 if (unlikely(PageHWPoison(vmf->page))) { 3089 if (ret & VM_FAULT_LOCKED) 3090 unlock_page(vmf->page); 3091 put_page(vmf->page); 3092 vmf->page = NULL; 3093 return VM_FAULT_HWPOISON; 3094 } 3095 3096 if (unlikely(!(ret & VM_FAULT_LOCKED))) 3097 lock_page(vmf->page); 3098 else 3099 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page); 3100 3101 return ret; 3102 } 3103 3104 /* 3105 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set. 3106 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check 3107 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly 3108 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output. 3109 */ 3110 static int pmd_devmap_trans_unstable(pmd_t *pmd) 3111 { 3112 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd); 3113 } 3114 3115 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf) 3116 { 3117 struct vm_area_struct *vma = vmf->vma; 3118 3119 if (!pmd_none(*vmf->pmd)) 3120 goto map_pte; 3121 if (vmf->prealloc_pte) { 3122 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3123 if (unlikely(!pmd_none(*vmf->pmd))) { 3124 spin_unlock(vmf->ptl); 3125 goto map_pte; 3126 } 3127 3128 mm_inc_nr_ptes(vma->vm_mm); 3129 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3130 spin_unlock(vmf->ptl); 3131 vmf->prealloc_pte = NULL; 3132 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) { 3133 return VM_FAULT_OOM; 3134 } 3135 map_pte: 3136 /* 3137 * If a huge pmd materialized under us just retry later. Use 3138 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of 3139 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge 3140 * under us and then back to pmd_none, as a result of MADV_DONTNEED 3141 * running immediately after a huge pmd fault in a different thread of 3142 * this mm, in turn leading to a misleading pmd_trans_huge() retval. 3143 * All we have to ensure is that it is a regular pmd that we can walk 3144 * with pte_offset_map() and we can do that through an atomic read in 3145 * C, which is what pmd_trans_unstable() provides. 3146 */ 3147 if (pmd_devmap_trans_unstable(vmf->pmd)) 3148 return VM_FAULT_NOPAGE; 3149 3150 /* 3151 * At this point we know that our vmf->pmd points to a page of ptes 3152 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge() 3153 * for the duration of the fault. If a racing MADV_DONTNEED runs and 3154 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still 3155 * be valid and we will re-check to make sure the vmf->pte isn't 3156 * pte_none() under vmf->ptl protection when we return to 3157 * alloc_set_pte(). 3158 */ 3159 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3160 &vmf->ptl); 3161 return 0; 3162 } 3163 3164 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE 3165 3166 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1) 3167 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma, 3168 unsigned long haddr) 3169 { 3170 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) != 3171 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK)) 3172 return false; 3173 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 3174 return false; 3175 return true; 3176 } 3177 3178 static void deposit_prealloc_pte(struct vm_fault *vmf) 3179 { 3180 struct vm_area_struct *vma = vmf->vma; 3181 3182 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 3183 /* 3184 * We are going to consume the prealloc table, 3185 * count that as nr_ptes. 3186 */ 3187 mm_inc_nr_ptes(vma->vm_mm); 3188 vmf->prealloc_pte = NULL; 3189 } 3190 3191 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3192 { 3193 struct vm_area_struct *vma = vmf->vma; 3194 bool write = vmf->flags & FAULT_FLAG_WRITE; 3195 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 3196 pmd_t entry; 3197 int i; 3198 vm_fault_t ret; 3199 3200 if (!transhuge_vma_suitable(vma, haddr)) 3201 return VM_FAULT_FALLBACK; 3202 3203 ret = VM_FAULT_FALLBACK; 3204 page = compound_head(page); 3205 3206 /* 3207 * Archs like ppc64 need additonal space to store information 3208 * related to pte entry. Use the preallocated table for that. 3209 */ 3210 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 3211 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 3212 if (!vmf->prealloc_pte) 3213 return VM_FAULT_OOM; 3214 smp_wmb(); /* See comment in __pte_alloc() */ 3215 } 3216 3217 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 3218 if (unlikely(!pmd_none(*vmf->pmd))) 3219 goto out; 3220 3221 for (i = 0; i < HPAGE_PMD_NR; i++) 3222 flush_icache_page(vma, page + i); 3223 3224 entry = mk_huge_pmd(page, vma->vm_page_prot); 3225 if (write) 3226 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 3227 3228 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR); 3229 page_add_file_rmap(page, true); 3230 /* 3231 * deposit and withdraw with pmd lock held 3232 */ 3233 if (arch_needs_pgtable_deposit()) 3234 deposit_prealloc_pte(vmf); 3235 3236 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 3237 3238 update_mmu_cache_pmd(vma, haddr, vmf->pmd); 3239 3240 /* fault is handled */ 3241 ret = 0; 3242 count_vm_event(THP_FILE_MAPPED); 3243 out: 3244 spin_unlock(vmf->ptl); 3245 return ret; 3246 } 3247 #else 3248 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 3249 { 3250 BUILD_BUG(); 3251 return 0; 3252 } 3253 #endif 3254 3255 /** 3256 * alloc_set_pte - setup new PTE entry for given page and add reverse page 3257 * mapping. If needed, the fucntion allocates page table or use pre-allocated. 3258 * 3259 * @vmf: fault environment 3260 * @memcg: memcg to charge page (only for private mappings) 3261 * @page: page to map 3262 * 3263 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on 3264 * return. 3265 * 3266 * Target users are page handler itself and implementations of 3267 * vm_ops->map_pages. 3268 * 3269 * Return: %0 on success, %VM_FAULT_ code in case of error. 3270 */ 3271 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg, 3272 struct page *page) 3273 { 3274 struct vm_area_struct *vma = vmf->vma; 3275 bool write = vmf->flags & FAULT_FLAG_WRITE; 3276 pte_t entry; 3277 vm_fault_t ret; 3278 3279 if (pmd_none(*vmf->pmd) && PageTransCompound(page) && 3280 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) { 3281 /* THP on COW? */ 3282 VM_BUG_ON_PAGE(memcg, page); 3283 3284 ret = do_set_pmd(vmf, page); 3285 if (ret != VM_FAULT_FALLBACK) 3286 return ret; 3287 } 3288 3289 if (!vmf->pte) { 3290 ret = pte_alloc_one_map(vmf); 3291 if (ret) 3292 return ret; 3293 } 3294 3295 /* Re-check under ptl */ 3296 if (unlikely(!pte_none(*vmf->pte))) 3297 return VM_FAULT_NOPAGE; 3298 3299 flush_icache_page(vma, page); 3300 entry = mk_pte(page, vma->vm_page_prot); 3301 if (write) 3302 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3303 /* copy-on-write page */ 3304 if (write && !(vma->vm_flags & VM_SHARED)) { 3305 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); 3306 page_add_new_anon_rmap(page, vma, vmf->address, false); 3307 mem_cgroup_commit_charge(page, memcg, false, false); 3308 lru_cache_add_active_or_unevictable(page, vma); 3309 } else { 3310 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page)); 3311 page_add_file_rmap(page, false); 3312 } 3313 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry); 3314 3315 /* no need to invalidate: a not-present page won't be cached */ 3316 update_mmu_cache(vma, vmf->address, vmf->pte); 3317 3318 return 0; 3319 } 3320 3321 3322 /** 3323 * finish_fault - finish page fault once we have prepared the page to fault 3324 * 3325 * @vmf: structure describing the fault 3326 * 3327 * This function handles all that is needed to finish a page fault once the 3328 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 3329 * given page, adds reverse page mapping, handles memcg charges and LRU 3330 * addition. 3331 * 3332 * The function expects the page to be locked and on success it consumes a 3333 * reference of a page being mapped (for the PTE which maps it). 3334 * 3335 * Return: %0 on success, %VM_FAULT_ code in case of error. 3336 */ 3337 vm_fault_t finish_fault(struct vm_fault *vmf) 3338 { 3339 struct page *page; 3340 vm_fault_t ret = 0; 3341 3342 /* Did we COW the page? */ 3343 if ((vmf->flags & FAULT_FLAG_WRITE) && 3344 !(vmf->vma->vm_flags & VM_SHARED)) 3345 page = vmf->cow_page; 3346 else 3347 page = vmf->page; 3348 3349 /* 3350 * check even for read faults because we might have lost our CoWed 3351 * page 3352 */ 3353 if (!(vmf->vma->vm_flags & VM_SHARED)) 3354 ret = check_stable_address_space(vmf->vma->vm_mm); 3355 if (!ret) 3356 ret = alloc_set_pte(vmf, vmf->memcg, page); 3357 if (vmf->pte) 3358 pte_unmap_unlock(vmf->pte, vmf->ptl); 3359 return ret; 3360 } 3361 3362 static unsigned long fault_around_bytes __read_mostly = 3363 rounddown_pow_of_two(65536); 3364 3365 #ifdef CONFIG_DEBUG_FS 3366 static int fault_around_bytes_get(void *data, u64 *val) 3367 { 3368 *val = fault_around_bytes; 3369 return 0; 3370 } 3371 3372 /* 3373 * fault_around_bytes must be rounded down to the nearest page order as it's 3374 * what do_fault_around() expects to see. 3375 */ 3376 static int fault_around_bytes_set(void *data, u64 val) 3377 { 3378 if (val / PAGE_SIZE > PTRS_PER_PTE) 3379 return -EINVAL; 3380 if (val > PAGE_SIZE) 3381 fault_around_bytes = rounddown_pow_of_two(val); 3382 else 3383 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ 3384 return 0; 3385 } 3386 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 3387 fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 3388 3389 static int __init fault_around_debugfs(void) 3390 { 3391 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 3392 &fault_around_bytes_fops); 3393 return 0; 3394 } 3395 late_initcall(fault_around_debugfs); 3396 #endif 3397 3398 /* 3399 * do_fault_around() tries to map few pages around the fault address. The hope 3400 * is that the pages will be needed soon and this will lower the number of 3401 * faults to handle. 3402 * 3403 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 3404 * not ready to be mapped: not up-to-date, locked, etc. 3405 * 3406 * This function is called with the page table lock taken. In the split ptlock 3407 * case the page table lock only protects only those entries which belong to 3408 * the page table corresponding to the fault address. 3409 * 3410 * This function doesn't cross the VMA boundaries, in order to call map_pages() 3411 * only once. 3412 * 3413 * fault_around_bytes defines how many bytes we'll try to map. 3414 * do_fault_around() expects it to be set to a power of two less than or equal 3415 * to PTRS_PER_PTE. 3416 * 3417 * The virtual address of the area that we map is naturally aligned to 3418 * fault_around_bytes rounded down to the machine page size 3419 * (and therefore to page order). This way it's easier to guarantee 3420 * that we don't cross page table boundaries. 3421 */ 3422 static vm_fault_t do_fault_around(struct vm_fault *vmf) 3423 { 3424 unsigned long address = vmf->address, nr_pages, mask; 3425 pgoff_t start_pgoff = vmf->pgoff; 3426 pgoff_t end_pgoff; 3427 int off; 3428 vm_fault_t ret = 0; 3429 3430 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT; 3431 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; 3432 3433 vmf->address = max(address & mask, vmf->vma->vm_start); 3434 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); 3435 start_pgoff -= off; 3436 3437 /* 3438 * end_pgoff is either the end of the page table, the end of 3439 * the vma or nr_pages from start_pgoff, depending what is nearest. 3440 */ 3441 end_pgoff = start_pgoff - 3442 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + 3443 PTRS_PER_PTE - 1; 3444 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1, 3445 start_pgoff + nr_pages - 1); 3446 3447 if (pmd_none(*vmf->pmd)) { 3448 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 3449 if (!vmf->prealloc_pte) 3450 goto out; 3451 smp_wmb(); /* See comment in __pte_alloc() */ 3452 } 3453 3454 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff); 3455 3456 /* Huge page is mapped? Page fault is solved */ 3457 if (pmd_trans_huge(*vmf->pmd)) { 3458 ret = VM_FAULT_NOPAGE; 3459 goto out; 3460 } 3461 3462 /* ->map_pages() haven't done anything useful. Cold page cache? */ 3463 if (!vmf->pte) 3464 goto out; 3465 3466 /* check if the page fault is solved */ 3467 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT); 3468 if (!pte_none(*vmf->pte)) 3469 ret = VM_FAULT_NOPAGE; 3470 pte_unmap_unlock(vmf->pte, vmf->ptl); 3471 out: 3472 vmf->address = address; 3473 vmf->pte = NULL; 3474 return ret; 3475 } 3476 3477 static vm_fault_t do_read_fault(struct vm_fault *vmf) 3478 { 3479 struct vm_area_struct *vma = vmf->vma; 3480 vm_fault_t ret = 0; 3481 3482 /* 3483 * Let's call ->map_pages() first and use ->fault() as fallback 3484 * if page by the offset is not ready to be mapped (cold cache or 3485 * something). 3486 */ 3487 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { 3488 ret = do_fault_around(vmf); 3489 if (ret) 3490 return ret; 3491 } 3492 3493 ret = __do_fault(vmf); 3494 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3495 return ret; 3496 3497 ret |= finish_fault(vmf); 3498 unlock_page(vmf->page); 3499 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3500 put_page(vmf->page); 3501 return ret; 3502 } 3503 3504 static vm_fault_t do_cow_fault(struct vm_fault *vmf) 3505 { 3506 struct vm_area_struct *vma = vmf->vma; 3507 vm_fault_t ret; 3508 3509 if (unlikely(anon_vma_prepare(vma))) 3510 return VM_FAULT_OOM; 3511 3512 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address); 3513 if (!vmf->cow_page) 3514 return VM_FAULT_OOM; 3515 3516 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL, 3517 &vmf->memcg, false)) { 3518 put_page(vmf->cow_page); 3519 return VM_FAULT_OOM; 3520 } 3521 3522 ret = __do_fault(vmf); 3523 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3524 goto uncharge_out; 3525 if (ret & VM_FAULT_DONE_COW) 3526 return ret; 3527 3528 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 3529 __SetPageUptodate(vmf->cow_page); 3530 3531 ret |= finish_fault(vmf); 3532 unlock_page(vmf->page); 3533 put_page(vmf->page); 3534 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3535 goto uncharge_out; 3536 return ret; 3537 uncharge_out: 3538 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false); 3539 put_page(vmf->cow_page); 3540 return ret; 3541 } 3542 3543 static vm_fault_t do_shared_fault(struct vm_fault *vmf) 3544 { 3545 struct vm_area_struct *vma = vmf->vma; 3546 vm_fault_t ret, tmp; 3547 3548 ret = __do_fault(vmf); 3549 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 3550 return ret; 3551 3552 /* 3553 * Check if the backing address space wants to know that the page is 3554 * about to become writable 3555 */ 3556 if (vma->vm_ops->page_mkwrite) { 3557 unlock_page(vmf->page); 3558 tmp = do_page_mkwrite(vmf); 3559 if (unlikely(!tmp || 3560 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 3561 put_page(vmf->page); 3562 return tmp; 3563 } 3564 } 3565 3566 ret |= finish_fault(vmf); 3567 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 3568 VM_FAULT_RETRY))) { 3569 unlock_page(vmf->page); 3570 put_page(vmf->page); 3571 return ret; 3572 } 3573 3574 fault_dirty_shared_page(vma, vmf->page); 3575 return ret; 3576 } 3577 3578 /* 3579 * We enter with non-exclusive mmap_sem (to exclude vma changes, 3580 * but allow concurrent faults). 3581 * The mmap_sem may have been released depending on flags and our 3582 * return value. See filemap_fault() and __lock_page_or_retry(). 3583 * If mmap_sem is released, vma may become invalid (for example 3584 * by other thread calling munmap()). 3585 */ 3586 static vm_fault_t do_fault(struct vm_fault *vmf) 3587 { 3588 struct vm_area_struct *vma = vmf->vma; 3589 struct mm_struct *vm_mm = vma->vm_mm; 3590 vm_fault_t ret; 3591 3592 /* 3593 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 3594 */ 3595 if (!vma->vm_ops->fault) { 3596 /* 3597 * If we find a migration pmd entry or a none pmd entry, which 3598 * should never happen, return SIGBUS 3599 */ 3600 if (unlikely(!pmd_present(*vmf->pmd))) 3601 ret = VM_FAULT_SIGBUS; 3602 else { 3603 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, 3604 vmf->pmd, 3605 vmf->address, 3606 &vmf->ptl); 3607 /* 3608 * Make sure this is not a temporary clearing of pte 3609 * by holding ptl and checking again. A R/M/W update 3610 * of pte involves: take ptl, clearing the pte so that 3611 * we don't have concurrent modification by hardware 3612 * followed by an update. 3613 */ 3614 if (unlikely(pte_none(*vmf->pte))) 3615 ret = VM_FAULT_SIGBUS; 3616 else 3617 ret = VM_FAULT_NOPAGE; 3618 3619 pte_unmap_unlock(vmf->pte, vmf->ptl); 3620 } 3621 } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 3622 ret = do_read_fault(vmf); 3623 else if (!(vma->vm_flags & VM_SHARED)) 3624 ret = do_cow_fault(vmf); 3625 else 3626 ret = do_shared_fault(vmf); 3627 3628 /* preallocated pagetable is unused: free it */ 3629 if (vmf->prealloc_pte) { 3630 pte_free(vm_mm, vmf->prealloc_pte); 3631 vmf->prealloc_pte = NULL; 3632 } 3633 return ret; 3634 } 3635 3636 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, 3637 unsigned long addr, int page_nid, 3638 int *flags) 3639 { 3640 get_page(page); 3641 3642 count_vm_numa_event(NUMA_HINT_FAULTS); 3643 if (page_nid == numa_node_id()) { 3644 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 3645 *flags |= TNF_FAULT_LOCAL; 3646 } 3647 3648 return mpol_misplaced(page, vma, addr); 3649 } 3650 3651 static vm_fault_t do_numa_page(struct vm_fault *vmf) 3652 { 3653 struct vm_area_struct *vma = vmf->vma; 3654 struct page *page = NULL; 3655 int page_nid = NUMA_NO_NODE; 3656 int last_cpupid; 3657 int target_nid; 3658 bool migrated = false; 3659 pte_t pte, old_pte; 3660 bool was_writable = pte_savedwrite(vmf->orig_pte); 3661 int flags = 0; 3662 3663 /* 3664 * The "pte" at this point cannot be used safely without 3665 * validation through pte_unmap_same(). It's of NUMA type but 3666 * the pfn may be screwed if the read is non atomic. 3667 */ 3668 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd); 3669 spin_lock(vmf->ptl); 3670 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) { 3671 pte_unmap_unlock(vmf->pte, vmf->ptl); 3672 goto out; 3673 } 3674 3675 /* 3676 * Make it present again, Depending on how arch implementes non 3677 * accessible ptes, some can allow access by kernel mode. 3678 */ 3679 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 3680 pte = pte_modify(old_pte, vma->vm_page_prot); 3681 pte = pte_mkyoung(pte); 3682 if (was_writable) 3683 pte = pte_mkwrite(pte); 3684 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 3685 update_mmu_cache(vma, vmf->address, vmf->pte); 3686 3687 page = vm_normal_page(vma, vmf->address, pte); 3688 if (!page) { 3689 pte_unmap_unlock(vmf->pte, vmf->ptl); 3690 return 0; 3691 } 3692 3693 /* TODO: handle PTE-mapped THP */ 3694 if (PageCompound(page)) { 3695 pte_unmap_unlock(vmf->pte, vmf->ptl); 3696 return 0; 3697 } 3698 3699 /* 3700 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 3701 * much anyway since they can be in shared cache state. This misses 3702 * the case where a mapping is writable but the process never writes 3703 * to it but pte_write gets cleared during protection updates and 3704 * pte_dirty has unpredictable behaviour between PTE scan updates, 3705 * background writeback, dirty balancing and application behaviour. 3706 */ 3707 if (!pte_write(pte)) 3708 flags |= TNF_NO_GROUP; 3709 3710 /* 3711 * Flag if the page is shared between multiple address spaces. This 3712 * is later used when determining whether to group tasks together 3713 */ 3714 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) 3715 flags |= TNF_SHARED; 3716 3717 last_cpupid = page_cpupid_last(page); 3718 page_nid = page_to_nid(page); 3719 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid, 3720 &flags); 3721 pte_unmap_unlock(vmf->pte, vmf->ptl); 3722 if (target_nid == NUMA_NO_NODE) { 3723 put_page(page); 3724 goto out; 3725 } 3726 3727 /* Migrate to the requested node */ 3728 migrated = migrate_misplaced_page(page, vma, target_nid); 3729 if (migrated) { 3730 page_nid = target_nid; 3731 flags |= TNF_MIGRATED; 3732 } else 3733 flags |= TNF_MIGRATE_FAIL; 3734 3735 out: 3736 if (page_nid != NUMA_NO_NODE) 3737 task_numa_fault(last_cpupid, page_nid, 1, flags); 3738 return 0; 3739 } 3740 3741 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 3742 { 3743 if (vma_is_anonymous(vmf->vma)) 3744 return do_huge_pmd_anonymous_page(vmf); 3745 if (vmf->vma->vm_ops->huge_fault) 3746 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 3747 return VM_FAULT_FALLBACK; 3748 } 3749 3750 /* `inline' is required to avoid gcc 4.1.2 build error */ 3751 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd) 3752 { 3753 if (vma_is_anonymous(vmf->vma)) 3754 return do_huge_pmd_wp_page(vmf, orig_pmd); 3755 if (vmf->vma->vm_ops->huge_fault) 3756 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD); 3757 3758 /* COW handled on pte level: split pmd */ 3759 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma); 3760 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL); 3761 3762 return VM_FAULT_FALLBACK; 3763 } 3764 3765 static inline bool vma_is_accessible(struct vm_area_struct *vma) 3766 { 3767 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE); 3768 } 3769 3770 static vm_fault_t create_huge_pud(struct vm_fault *vmf) 3771 { 3772 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3773 /* No support for anonymous transparent PUD pages yet */ 3774 if (vma_is_anonymous(vmf->vma)) 3775 return VM_FAULT_FALLBACK; 3776 if (vmf->vma->vm_ops->huge_fault) 3777 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 3778 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 3779 return VM_FAULT_FALLBACK; 3780 } 3781 3782 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 3783 { 3784 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3785 /* No support for anonymous transparent PUD pages yet */ 3786 if (vma_is_anonymous(vmf->vma)) 3787 return VM_FAULT_FALLBACK; 3788 if (vmf->vma->vm_ops->huge_fault) 3789 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD); 3790 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 3791 return VM_FAULT_FALLBACK; 3792 } 3793 3794 /* 3795 * These routines also need to handle stuff like marking pages dirty 3796 * and/or accessed for architectures that don't do it in hardware (most 3797 * RISC architectures). The early dirtying is also good on the i386. 3798 * 3799 * There is also a hook called "update_mmu_cache()" that architectures 3800 * with external mmu caches can use to update those (ie the Sparc or 3801 * PowerPC hashed page tables that act as extended TLBs). 3802 * 3803 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow 3804 * concurrent faults). 3805 * 3806 * The mmap_sem may have been released depending on flags and our return value. 3807 * See filemap_fault() and __lock_page_or_retry(). 3808 */ 3809 static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 3810 { 3811 pte_t entry; 3812 3813 if (unlikely(pmd_none(*vmf->pmd))) { 3814 /* 3815 * Leave __pte_alloc() until later: because vm_ops->fault may 3816 * want to allocate huge page, and if we expose page table 3817 * for an instant, it will be difficult to retract from 3818 * concurrent faults and from rmap lookups. 3819 */ 3820 vmf->pte = NULL; 3821 } else { 3822 /* See comment in pte_alloc_one_map() */ 3823 if (pmd_devmap_trans_unstable(vmf->pmd)) 3824 return 0; 3825 /* 3826 * A regular pmd is established and it can't morph into a huge 3827 * pmd from under us anymore at this point because we hold the 3828 * mmap_sem read mode and khugepaged takes it in write mode. 3829 * So now it's safe to run pte_offset_map(). 3830 */ 3831 vmf->pte = pte_offset_map(vmf->pmd, vmf->address); 3832 vmf->orig_pte = *vmf->pte; 3833 3834 /* 3835 * some architectures can have larger ptes than wordsize, 3836 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and 3837 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic 3838 * accesses. The code below just needs a consistent view 3839 * for the ifs and we later double check anyway with the 3840 * ptl lock held. So here a barrier will do. 3841 */ 3842 barrier(); 3843 if (pte_none(vmf->orig_pte)) { 3844 pte_unmap(vmf->pte); 3845 vmf->pte = NULL; 3846 } 3847 } 3848 3849 if (!vmf->pte) { 3850 if (vma_is_anonymous(vmf->vma)) 3851 return do_anonymous_page(vmf); 3852 else 3853 return do_fault(vmf); 3854 } 3855 3856 if (!pte_present(vmf->orig_pte)) 3857 return do_swap_page(vmf); 3858 3859 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 3860 return do_numa_page(vmf); 3861 3862 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd); 3863 spin_lock(vmf->ptl); 3864 entry = vmf->orig_pte; 3865 if (unlikely(!pte_same(*vmf->pte, entry))) 3866 goto unlock; 3867 if (vmf->flags & FAULT_FLAG_WRITE) { 3868 if (!pte_write(entry)) 3869 return do_wp_page(vmf); 3870 entry = pte_mkdirty(entry); 3871 } 3872 entry = pte_mkyoung(entry); 3873 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 3874 vmf->flags & FAULT_FLAG_WRITE)) { 3875 update_mmu_cache(vmf->vma, vmf->address, vmf->pte); 3876 } else { 3877 /* 3878 * This is needed only for protection faults but the arch code 3879 * is not yet telling us if this is a protection fault or not. 3880 * This still avoids useless tlb flushes for .text page faults 3881 * with threads. 3882 */ 3883 if (vmf->flags & FAULT_FLAG_WRITE) 3884 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address); 3885 } 3886 unlock: 3887 pte_unmap_unlock(vmf->pte, vmf->ptl); 3888 return 0; 3889 } 3890 3891 /* 3892 * By the time we get here, we already hold the mm semaphore 3893 * 3894 * The mmap_sem may have been released depending on flags and our 3895 * return value. See filemap_fault() and __lock_page_or_retry(). 3896 */ 3897 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 3898 unsigned long address, unsigned int flags) 3899 { 3900 struct vm_fault vmf = { 3901 .vma = vma, 3902 .address = address & PAGE_MASK, 3903 .flags = flags, 3904 .pgoff = linear_page_index(vma, address), 3905 .gfp_mask = __get_fault_gfp_mask(vma), 3906 }; 3907 unsigned int dirty = flags & FAULT_FLAG_WRITE; 3908 struct mm_struct *mm = vma->vm_mm; 3909 pgd_t *pgd; 3910 p4d_t *p4d; 3911 vm_fault_t ret; 3912 3913 pgd = pgd_offset(mm, address); 3914 p4d = p4d_alloc(mm, pgd, address); 3915 if (!p4d) 3916 return VM_FAULT_OOM; 3917 3918 vmf.pud = pud_alloc(mm, p4d, address); 3919 if (!vmf.pud) 3920 return VM_FAULT_OOM; 3921 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) { 3922 ret = create_huge_pud(&vmf); 3923 if (!(ret & VM_FAULT_FALLBACK)) 3924 return ret; 3925 } else { 3926 pud_t orig_pud = *vmf.pud; 3927 3928 barrier(); 3929 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 3930 3931 /* NUMA case for anonymous PUDs would go here */ 3932 3933 if (dirty && !pud_write(orig_pud)) { 3934 ret = wp_huge_pud(&vmf, orig_pud); 3935 if (!(ret & VM_FAULT_FALLBACK)) 3936 return ret; 3937 } else { 3938 huge_pud_set_accessed(&vmf, orig_pud); 3939 return 0; 3940 } 3941 } 3942 } 3943 3944 vmf.pmd = pmd_alloc(mm, vmf.pud, address); 3945 if (!vmf.pmd) 3946 return VM_FAULT_OOM; 3947 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) { 3948 ret = create_huge_pmd(&vmf); 3949 if (!(ret & VM_FAULT_FALLBACK)) 3950 return ret; 3951 } else { 3952 pmd_t orig_pmd = *vmf.pmd; 3953 3954 barrier(); 3955 if (unlikely(is_swap_pmd(orig_pmd))) { 3956 VM_BUG_ON(thp_migration_supported() && 3957 !is_pmd_migration_entry(orig_pmd)); 3958 if (is_pmd_migration_entry(orig_pmd)) 3959 pmd_migration_entry_wait(mm, vmf.pmd); 3960 return 0; 3961 } 3962 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) { 3963 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma)) 3964 return do_huge_pmd_numa_page(&vmf, orig_pmd); 3965 3966 if (dirty && !pmd_write(orig_pmd)) { 3967 ret = wp_huge_pmd(&vmf, orig_pmd); 3968 if (!(ret & VM_FAULT_FALLBACK)) 3969 return ret; 3970 } else { 3971 huge_pmd_set_accessed(&vmf, orig_pmd); 3972 return 0; 3973 } 3974 } 3975 } 3976 3977 return handle_pte_fault(&vmf); 3978 } 3979 3980 /* 3981 * By the time we get here, we already hold the mm semaphore 3982 * 3983 * The mmap_sem may have been released depending on flags and our 3984 * return value. See filemap_fault() and __lock_page_or_retry(). 3985 */ 3986 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 3987 unsigned int flags) 3988 { 3989 vm_fault_t ret; 3990 3991 __set_current_state(TASK_RUNNING); 3992 3993 count_vm_event(PGFAULT); 3994 count_memcg_event_mm(vma->vm_mm, PGFAULT); 3995 3996 /* do counter updates before entering really critical section. */ 3997 check_sync_rss_stat(current); 3998 3999 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 4000 flags & FAULT_FLAG_INSTRUCTION, 4001 flags & FAULT_FLAG_REMOTE)) 4002 return VM_FAULT_SIGSEGV; 4003 4004 /* 4005 * Enable the memcg OOM handling for faults triggered in user 4006 * space. Kernel faults are handled more gracefully. 4007 */ 4008 if (flags & FAULT_FLAG_USER) 4009 mem_cgroup_enter_user_fault(); 4010 4011 if (unlikely(is_vm_hugetlb_page(vma))) 4012 ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 4013 else 4014 ret = __handle_mm_fault(vma, address, flags); 4015 4016 if (flags & FAULT_FLAG_USER) { 4017 mem_cgroup_exit_user_fault(); 4018 /* 4019 * The task may have entered a memcg OOM situation but 4020 * if the allocation error was handled gracefully (no 4021 * VM_FAULT_OOM), there is no need to kill anything. 4022 * Just clean up the OOM state peacefully. 4023 */ 4024 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 4025 mem_cgroup_oom_synchronize(false); 4026 } 4027 4028 return ret; 4029 } 4030 EXPORT_SYMBOL_GPL(handle_mm_fault); 4031 4032 #ifndef __PAGETABLE_P4D_FOLDED 4033 /* 4034 * Allocate p4d page table. 4035 * We've already handled the fast-path in-line. 4036 */ 4037 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 4038 { 4039 p4d_t *new = p4d_alloc_one(mm, address); 4040 if (!new) 4041 return -ENOMEM; 4042 4043 smp_wmb(); /* See comment in __pte_alloc */ 4044 4045 spin_lock(&mm->page_table_lock); 4046 if (pgd_present(*pgd)) /* Another has populated it */ 4047 p4d_free(mm, new); 4048 else 4049 pgd_populate(mm, pgd, new); 4050 spin_unlock(&mm->page_table_lock); 4051 return 0; 4052 } 4053 #endif /* __PAGETABLE_P4D_FOLDED */ 4054 4055 #ifndef __PAGETABLE_PUD_FOLDED 4056 /* 4057 * Allocate page upper directory. 4058 * We've already handled the fast-path in-line. 4059 */ 4060 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 4061 { 4062 pud_t *new = pud_alloc_one(mm, address); 4063 if (!new) 4064 return -ENOMEM; 4065 4066 smp_wmb(); /* See comment in __pte_alloc */ 4067 4068 spin_lock(&mm->page_table_lock); 4069 #ifndef __ARCH_HAS_5LEVEL_HACK 4070 if (!p4d_present(*p4d)) { 4071 mm_inc_nr_puds(mm); 4072 p4d_populate(mm, p4d, new); 4073 } else /* Another has populated it */ 4074 pud_free(mm, new); 4075 #else 4076 if (!pgd_present(*p4d)) { 4077 mm_inc_nr_puds(mm); 4078 pgd_populate(mm, p4d, new); 4079 } else /* Another has populated it */ 4080 pud_free(mm, new); 4081 #endif /* __ARCH_HAS_5LEVEL_HACK */ 4082 spin_unlock(&mm->page_table_lock); 4083 return 0; 4084 } 4085 #endif /* __PAGETABLE_PUD_FOLDED */ 4086 4087 #ifndef __PAGETABLE_PMD_FOLDED 4088 /* 4089 * Allocate page middle directory. 4090 * We've already handled the fast-path in-line. 4091 */ 4092 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 4093 { 4094 spinlock_t *ptl; 4095 pmd_t *new = pmd_alloc_one(mm, address); 4096 if (!new) 4097 return -ENOMEM; 4098 4099 smp_wmb(); /* See comment in __pte_alloc */ 4100 4101 ptl = pud_lock(mm, pud); 4102 #ifndef __ARCH_HAS_4LEVEL_HACK 4103 if (!pud_present(*pud)) { 4104 mm_inc_nr_pmds(mm); 4105 pud_populate(mm, pud, new); 4106 } else /* Another has populated it */ 4107 pmd_free(mm, new); 4108 #else 4109 if (!pgd_present(*pud)) { 4110 mm_inc_nr_pmds(mm); 4111 pgd_populate(mm, pud, new); 4112 } else /* Another has populated it */ 4113 pmd_free(mm, new); 4114 #endif /* __ARCH_HAS_4LEVEL_HACK */ 4115 spin_unlock(ptl); 4116 return 0; 4117 } 4118 #endif /* __PAGETABLE_PMD_FOLDED */ 4119 4120 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address, 4121 struct mmu_notifier_range *range, 4122 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp) 4123 { 4124 pgd_t *pgd; 4125 p4d_t *p4d; 4126 pud_t *pud; 4127 pmd_t *pmd; 4128 pte_t *ptep; 4129 4130 pgd = pgd_offset(mm, address); 4131 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 4132 goto out; 4133 4134 p4d = p4d_offset(pgd, address); 4135 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 4136 goto out; 4137 4138 pud = pud_offset(p4d, address); 4139 if (pud_none(*pud) || unlikely(pud_bad(*pud))) 4140 goto out; 4141 4142 pmd = pmd_offset(pud, address); 4143 VM_BUG_ON(pmd_trans_huge(*pmd)); 4144 4145 if (pmd_huge(*pmd)) { 4146 if (!pmdpp) 4147 goto out; 4148 4149 if (range) { 4150 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, 4151 NULL, mm, address & PMD_MASK, 4152 (address & PMD_MASK) + PMD_SIZE); 4153 mmu_notifier_invalidate_range_start(range); 4154 } 4155 *ptlp = pmd_lock(mm, pmd); 4156 if (pmd_huge(*pmd)) { 4157 *pmdpp = pmd; 4158 return 0; 4159 } 4160 spin_unlock(*ptlp); 4161 if (range) 4162 mmu_notifier_invalidate_range_end(range); 4163 } 4164 4165 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) 4166 goto out; 4167 4168 if (range) { 4169 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 4170 address & PAGE_MASK, 4171 (address & PAGE_MASK) + PAGE_SIZE); 4172 mmu_notifier_invalidate_range_start(range); 4173 } 4174 ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 4175 if (!pte_present(*ptep)) 4176 goto unlock; 4177 *ptepp = ptep; 4178 return 0; 4179 unlock: 4180 pte_unmap_unlock(ptep, *ptlp); 4181 if (range) 4182 mmu_notifier_invalidate_range_end(range); 4183 out: 4184 return -EINVAL; 4185 } 4186 4187 static inline int follow_pte(struct mm_struct *mm, unsigned long address, 4188 pte_t **ptepp, spinlock_t **ptlp) 4189 { 4190 int res; 4191 4192 /* (void) is needed to make gcc happy */ 4193 (void) __cond_lock(*ptlp, 4194 !(res = __follow_pte_pmd(mm, address, NULL, 4195 ptepp, NULL, ptlp))); 4196 return res; 4197 } 4198 4199 int follow_pte_pmd(struct mm_struct *mm, unsigned long address, 4200 struct mmu_notifier_range *range, 4201 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp) 4202 { 4203 int res; 4204 4205 /* (void) is needed to make gcc happy */ 4206 (void) __cond_lock(*ptlp, 4207 !(res = __follow_pte_pmd(mm, address, range, 4208 ptepp, pmdpp, ptlp))); 4209 return res; 4210 } 4211 EXPORT_SYMBOL(follow_pte_pmd); 4212 4213 /** 4214 * follow_pfn - look up PFN at a user virtual address 4215 * @vma: memory mapping 4216 * @address: user virtual address 4217 * @pfn: location to store found PFN 4218 * 4219 * Only IO mappings and raw PFN mappings are allowed. 4220 * 4221 * Return: zero and the pfn at @pfn on success, -ve otherwise. 4222 */ 4223 int follow_pfn(struct vm_area_struct *vma, unsigned long address, 4224 unsigned long *pfn) 4225 { 4226 int ret = -EINVAL; 4227 spinlock_t *ptl; 4228 pte_t *ptep; 4229 4230 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4231 return ret; 4232 4233 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 4234 if (ret) 4235 return ret; 4236 *pfn = pte_pfn(*ptep); 4237 pte_unmap_unlock(ptep, ptl); 4238 return 0; 4239 } 4240 EXPORT_SYMBOL(follow_pfn); 4241 4242 #ifdef CONFIG_HAVE_IOREMAP_PROT 4243 int follow_phys(struct vm_area_struct *vma, 4244 unsigned long address, unsigned int flags, 4245 unsigned long *prot, resource_size_t *phys) 4246 { 4247 int ret = -EINVAL; 4248 pte_t *ptep, pte; 4249 spinlock_t *ptl; 4250 4251 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 4252 goto out; 4253 4254 if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 4255 goto out; 4256 pte = *ptep; 4257 4258 if ((flags & FOLL_WRITE) && !pte_write(pte)) 4259 goto unlock; 4260 4261 *prot = pgprot_val(pte_pgprot(pte)); 4262 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 4263 4264 ret = 0; 4265 unlock: 4266 pte_unmap_unlock(ptep, ptl); 4267 out: 4268 return ret; 4269 } 4270 4271 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 4272 void *buf, int len, int write) 4273 { 4274 resource_size_t phys_addr; 4275 unsigned long prot = 0; 4276 void __iomem *maddr; 4277 int offset = addr & (PAGE_SIZE-1); 4278 4279 if (follow_phys(vma, addr, write, &prot, &phys_addr)) 4280 return -EINVAL; 4281 4282 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 4283 if (!maddr) 4284 return -ENOMEM; 4285 4286 if (write) 4287 memcpy_toio(maddr + offset, buf, len); 4288 else 4289 memcpy_fromio(buf, maddr + offset, len); 4290 iounmap(maddr); 4291 4292 return len; 4293 } 4294 EXPORT_SYMBOL_GPL(generic_access_phys); 4295 #endif 4296 4297 /* 4298 * Access another process' address space as given in mm. If non-NULL, use the 4299 * given task for page fault accounting. 4300 */ 4301 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, 4302 unsigned long addr, void *buf, int len, unsigned int gup_flags) 4303 { 4304 struct vm_area_struct *vma; 4305 void *old_buf = buf; 4306 int write = gup_flags & FOLL_WRITE; 4307 4308 if (down_read_killable(&mm->mmap_sem)) 4309 return 0; 4310 4311 /* ignore errors, just check how much was successfully transferred */ 4312 while (len) { 4313 int bytes, ret, offset; 4314 void *maddr; 4315 struct page *page = NULL; 4316 4317 ret = get_user_pages_remote(tsk, mm, addr, 1, 4318 gup_flags, &page, &vma, NULL); 4319 if (ret <= 0) { 4320 #ifndef CONFIG_HAVE_IOREMAP_PROT 4321 break; 4322 #else 4323 /* 4324 * Check if this is a VM_IO | VM_PFNMAP VMA, which 4325 * we can access using slightly different code. 4326 */ 4327 vma = find_vma(mm, addr); 4328 if (!vma || vma->vm_start > addr) 4329 break; 4330 if (vma->vm_ops && vma->vm_ops->access) 4331 ret = vma->vm_ops->access(vma, addr, buf, 4332 len, write); 4333 if (ret <= 0) 4334 break; 4335 bytes = ret; 4336 #endif 4337 } else { 4338 bytes = len; 4339 offset = addr & (PAGE_SIZE-1); 4340 if (bytes > PAGE_SIZE-offset) 4341 bytes = PAGE_SIZE-offset; 4342 4343 maddr = kmap(page); 4344 if (write) { 4345 copy_to_user_page(vma, page, addr, 4346 maddr + offset, buf, bytes); 4347 set_page_dirty_lock(page); 4348 } else { 4349 copy_from_user_page(vma, page, addr, 4350 buf, maddr + offset, bytes); 4351 } 4352 kunmap(page); 4353 put_page(page); 4354 } 4355 len -= bytes; 4356 buf += bytes; 4357 addr += bytes; 4358 } 4359 up_read(&mm->mmap_sem); 4360 4361 return buf - old_buf; 4362 } 4363 4364 /** 4365 * access_remote_vm - access another process' address space 4366 * @mm: the mm_struct of the target address space 4367 * @addr: start address to access 4368 * @buf: source or destination buffer 4369 * @len: number of bytes to transfer 4370 * @gup_flags: flags modifying lookup behaviour 4371 * 4372 * The caller must hold a reference on @mm. 4373 * 4374 * Return: number of bytes copied from source to destination. 4375 */ 4376 int access_remote_vm(struct mm_struct *mm, unsigned long addr, 4377 void *buf, int len, unsigned int gup_flags) 4378 { 4379 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags); 4380 } 4381 4382 /* 4383 * Access another process' address space. 4384 * Source/target buffer must be kernel space, 4385 * Do not walk the page table directly, use get_user_pages 4386 */ 4387 int access_process_vm(struct task_struct *tsk, unsigned long addr, 4388 void *buf, int len, unsigned int gup_flags) 4389 { 4390 struct mm_struct *mm; 4391 int ret; 4392 4393 mm = get_task_mm(tsk); 4394 if (!mm) 4395 return 0; 4396 4397 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags); 4398 4399 mmput(mm); 4400 4401 return ret; 4402 } 4403 EXPORT_SYMBOL_GPL(access_process_vm); 4404 4405 /* 4406 * Print the name of a VMA. 4407 */ 4408 void print_vma_addr(char *prefix, unsigned long ip) 4409 { 4410 struct mm_struct *mm = current->mm; 4411 struct vm_area_struct *vma; 4412 4413 /* 4414 * we might be running from an atomic context so we cannot sleep 4415 */ 4416 if (!down_read_trylock(&mm->mmap_sem)) 4417 return; 4418 4419 vma = find_vma(mm, ip); 4420 if (vma && vma->vm_file) { 4421 struct file *f = vma->vm_file; 4422 char *buf = (char *)__get_free_page(GFP_NOWAIT); 4423 if (buf) { 4424 char *p; 4425 4426 p = file_path(f, buf, PAGE_SIZE); 4427 if (IS_ERR(p)) 4428 p = "?"; 4429 printk("%s%s[%lx+%lx]", prefix, kbasename(p), 4430 vma->vm_start, 4431 vma->vm_end - vma->vm_start); 4432 free_page((unsigned long)buf); 4433 } 4434 } 4435 up_read(&mm->mmap_sem); 4436 } 4437 4438 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 4439 void __might_fault(const char *file, int line) 4440 { 4441 /* 4442 * Some code (nfs/sunrpc) uses socket ops on kernel memory while 4443 * holding the mmap_sem, this is safe because kernel memory doesn't 4444 * get paged out, therefore we'll never actually fault, and the 4445 * below annotations will generate false positives. 4446 */ 4447 if (uaccess_kernel()) 4448 return; 4449 if (pagefault_disabled()) 4450 return; 4451 __might_sleep(file, line, 0); 4452 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 4453 if (current->mm) 4454 might_lock_read(¤t->mm->mmap_sem); 4455 #endif 4456 } 4457 EXPORT_SYMBOL(__might_fault); 4458 #endif 4459 4460 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 4461 /* 4462 * Process all subpages of the specified huge page with the specified 4463 * operation. The target subpage will be processed last to keep its 4464 * cache lines hot. 4465 */ 4466 static inline void process_huge_page( 4467 unsigned long addr_hint, unsigned int pages_per_huge_page, 4468 void (*process_subpage)(unsigned long addr, int idx, void *arg), 4469 void *arg) 4470 { 4471 int i, n, base, l; 4472 unsigned long addr = addr_hint & 4473 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 4474 4475 /* Process target subpage last to keep its cache lines hot */ 4476 might_sleep(); 4477 n = (addr_hint - addr) / PAGE_SIZE; 4478 if (2 * n <= pages_per_huge_page) { 4479 /* If target subpage in first half of huge page */ 4480 base = 0; 4481 l = n; 4482 /* Process subpages at the end of huge page */ 4483 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 4484 cond_resched(); 4485 process_subpage(addr + i * PAGE_SIZE, i, arg); 4486 } 4487 } else { 4488 /* If target subpage in second half of huge page */ 4489 base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 4490 l = pages_per_huge_page - n; 4491 /* Process subpages at the begin of huge page */ 4492 for (i = 0; i < base; i++) { 4493 cond_resched(); 4494 process_subpage(addr + i * PAGE_SIZE, i, arg); 4495 } 4496 } 4497 /* 4498 * Process remaining subpages in left-right-left-right pattern 4499 * towards the target subpage 4500 */ 4501 for (i = 0; i < l; i++) { 4502 int left_idx = base + i; 4503 int right_idx = base + 2 * l - 1 - i; 4504 4505 cond_resched(); 4506 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 4507 cond_resched(); 4508 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 4509 } 4510 } 4511 4512 static void clear_gigantic_page(struct page *page, 4513 unsigned long addr, 4514 unsigned int pages_per_huge_page) 4515 { 4516 int i; 4517 struct page *p = page; 4518 4519 might_sleep(); 4520 for (i = 0; i < pages_per_huge_page; 4521 i++, p = mem_map_next(p, page, i)) { 4522 cond_resched(); 4523 clear_user_highpage(p, addr + i * PAGE_SIZE); 4524 } 4525 } 4526 4527 static void clear_subpage(unsigned long addr, int idx, void *arg) 4528 { 4529 struct page *page = arg; 4530 4531 clear_user_highpage(page + idx, addr); 4532 } 4533 4534 void clear_huge_page(struct page *page, 4535 unsigned long addr_hint, unsigned int pages_per_huge_page) 4536 { 4537 unsigned long addr = addr_hint & 4538 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 4539 4540 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 4541 clear_gigantic_page(page, addr, pages_per_huge_page); 4542 return; 4543 } 4544 4545 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 4546 } 4547 4548 static void copy_user_gigantic_page(struct page *dst, struct page *src, 4549 unsigned long addr, 4550 struct vm_area_struct *vma, 4551 unsigned int pages_per_huge_page) 4552 { 4553 int i; 4554 struct page *dst_base = dst; 4555 struct page *src_base = src; 4556 4557 for (i = 0; i < pages_per_huge_page; ) { 4558 cond_resched(); 4559 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 4560 4561 i++; 4562 dst = mem_map_next(dst, dst_base, i); 4563 src = mem_map_next(src, src_base, i); 4564 } 4565 } 4566 4567 struct copy_subpage_arg { 4568 struct page *dst; 4569 struct page *src; 4570 struct vm_area_struct *vma; 4571 }; 4572 4573 static void copy_subpage(unsigned long addr, int idx, void *arg) 4574 { 4575 struct copy_subpage_arg *copy_arg = arg; 4576 4577 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx, 4578 addr, copy_arg->vma); 4579 } 4580 4581 void copy_user_huge_page(struct page *dst, struct page *src, 4582 unsigned long addr_hint, struct vm_area_struct *vma, 4583 unsigned int pages_per_huge_page) 4584 { 4585 unsigned long addr = addr_hint & 4586 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 4587 struct copy_subpage_arg arg = { 4588 .dst = dst, 4589 .src = src, 4590 .vma = vma, 4591 }; 4592 4593 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 4594 copy_user_gigantic_page(dst, src, addr, vma, 4595 pages_per_huge_page); 4596 return; 4597 } 4598 4599 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 4600 } 4601 4602 long copy_huge_page_from_user(struct page *dst_page, 4603 const void __user *usr_src, 4604 unsigned int pages_per_huge_page, 4605 bool allow_pagefault) 4606 { 4607 void *src = (void *)usr_src; 4608 void *page_kaddr; 4609 unsigned long i, rc = 0; 4610 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE; 4611 4612 for (i = 0; i < pages_per_huge_page; i++) { 4613 if (allow_pagefault) 4614 page_kaddr = kmap(dst_page + i); 4615 else 4616 page_kaddr = kmap_atomic(dst_page + i); 4617 rc = copy_from_user(page_kaddr, 4618 (const void __user *)(src + i * PAGE_SIZE), 4619 PAGE_SIZE); 4620 if (allow_pagefault) 4621 kunmap(dst_page + i); 4622 else 4623 kunmap_atomic(page_kaddr); 4624 4625 ret_val -= (PAGE_SIZE - rc); 4626 if (rc) 4627 break; 4628 4629 cond_resched(); 4630 } 4631 return ret_val; 4632 } 4633 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 4634 4635 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 4636 4637 static struct kmem_cache *page_ptl_cachep; 4638 4639 void __init ptlock_cache_init(void) 4640 { 4641 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 4642 SLAB_PANIC, NULL); 4643 } 4644 4645 bool ptlock_alloc(struct page *page) 4646 { 4647 spinlock_t *ptl; 4648 4649 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 4650 if (!ptl) 4651 return false; 4652 page->ptl = ptl; 4653 return true; 4654 } 4655 4656 void ptlock_free(struct page *page) 4657 { 4658 kmem_cache_free(page_ptl_cachep, page->ptl); 4659 } 4660 #endif 4661