| 1 | |
| 2 | // SPDX-License-Identifier: GPL-2.0-only |
| 3 | /* |
| 4 | * linux/mm/memory.c |
| 5 | * |
| 6 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
| 7 | */ |
| 8 | |
| 9 | /* |
| 10 | * demand-loading started 01.12.91 - seems it is high on the list of |
| 11 | * things wanted, and it should be easy to implement. - Linus |
| 12 | */ |
| 13 | |
| 14 | /* |
| 15 | * Ok, demand-loading was easy, shared pages a little bit tricker. Shared |
| 16 | * pages started 02.12.91, seems to work. - Linus. |
| 17 | * |
| 18 | * Tested sharing by executing about 30 /bin/sh: under the old kernel it |
| 19 | * would have taken more than the 6M I have free, but it worked well as |
| 20 | * far as I could see. |
| 21 | * |
| 22 | * Also corrected some "invalidate()"s - I wasn't doing enough of them. |
| 23 | */ |
| 24 | |
| 25 | /* |
| 26 | * Real VM (paging to/from disk) started 18.12.91. Much more work and |
| 27 | * thought has to go into this. Oh, well.. |
| 28 | * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why. |
| 29 | * Found it. Everything seems to work now. |
| 30 | * 20.12.91 - Ok, making the swap-device changeable like the root. |
| 31 | */ |
| 32 | |
| 33 | /* |
| 34 | * 05.04.94 - Multi-page memory management added for v1.1. |
| 35 | * Idea by Alex Bligh (alex@cconcepts.co.uk) |
| 36 | * |
| 37 | * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG |
| 38 | * (Gerhard.Wichert@pdb.siemens.de) |
| 39 | * |
| 40 | * Aug/Sep 2004 Changed to four level page tables (Andi Kleen) |
| 41 | */ |
| 42 | |
| 43 | #include <linux/kernel_stat.h> |
| 44 | #include <linux/mm.h> |
| 45 | #include <linux/mm_inline.h> |
| 46 | #include <linux/sched/mm.h> |
| 47 | #include <linux/sched/coredump.h> |
| 48 | #include <linux/sched/numa_balancing.h> |
| 49 | #include <linux/sched/task.h> |
| 50 | #include <linux/hugetlb.h> |
| 51 | #include <linux/mman.h> |
| 52 | #include <linux/swap.h> |
| 53 | #include <linux/highmem.h> |
| 54 | #include <linux/pagemap.h> |
| 55 | #include <linux/memremap.h> |
| 56 | #include <linux/kmsan.h> |
| 57 | #include <linux/ksm.h> |
| 58 | #include <linux/rmap.h> |
| 59 | #include <linux/export.h> |
| 60 | #include <linux/delayacct.h> |
| 61 | #include <linux/init.h> |
| 62 | #include <linux/pfn_t.h> |
| 63 | #include <linux/writeback.h> |
| 64 | #include <linux/memcontrol.h> |
| 65 | #include <linux/mmu_notifier.h> |
| 66 | #include <linux/swapops.h> |
| 67 | #include <linux/elf.h> |
| 68 | #include <linux/gfp.h> |
| 69 | #include <linux/migrate.h> |
| 70 | #include <linux/string.h> |
| 71 | #include <linux/memory-tiers.h> |
| 72 | #include <linux/debugfs.h> |
| 73 | #include <linux/userfaultfd_k.h> |
| 74 | #include <linux/dax.h> |
| 75 | #include <linux/oom.h> |
| 76 | #include <linux/numa.h> |
| 77 | #include <linux/perf_event.h> |
| 78 | #include <linux/ptrace.h> |
| 79 | #include <linux/vmalloc.h> |
| 80 | #include <linux/sched/sysctl.h> |
| 81 | |
| 82 | #include <trace/events/kmem.h> |
| 83 | |
| 84 | #include <asm/io.h> |
| 85 | #include <asm/mmu_context.h> |
| 86 | #include <asm/pgalloc.h> |
| 87 | #include <linux/uaccess.h> |
| 88 | #include <asm/tlb.h> |
| 89 | #include <asm/tlbflush.h> |
| 90 | |
| 91 | #include "pgalloc-track.h" |
| 92 | #include "internal.h" |
| 93 | #include "swap.h" |
| 94 | |
| 95 | #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) |
| 96 | #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. |
| 97 | #endif |
| 98 | |
| 99 | #ifndef CONFIG_NUMA |
| 100 | unsigned long max_mapnr; |
| 101 | EXPORT_SYMBOL(max_mapnr); |
| 102 | |
| 103 | struct page *mem_map; |
| 104 | EXPORT_SYMBOL(mem_map); |
| 105 | #endif |
| 106 | |
| 107 | static vm_fault_t do_fault(struct vm_fault *vmf); |
| 108 | static vm_fault_t do_anonymous_page(struct vm_fault *vmf); |
| 109 | static bool vmf_pte_changed(struct vm_fault *vmf); |
| 110 | |
| 111 | /* |
| 112 | * Return true if the original pte was a uffd-wp pte marker (so the pte was |
| 113 | * wr-protected). |
| 114 | */ |
| 115 | static __always_inline bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf) |
| 116 | { |
| 117 | if (!userfaultfd_wp(vmf->vma)) |
| 118 | return false; |
| 119 | if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)) |
| 120 | return false; |
| 121 | |
| 122 | return pte_marker_uffd_wp(vmf->orig_pte); |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * A number of key systems in x86 including ioremap() rely on the assumption |
| 127 | * that high_memory defines the upper bound on direct map memory, then end |
| 128 | * of ZONE_NORMAL. |
| 129 | */ |
| 130 | void *high_memory; |
| 131 | EXPORT_SYMBOL(high_memory); |
| 132 | |
| 133 | /* |
| 134 | * Randomize the address space (stacks, mmaps, brk, etc.). |
| 135 | * |
| 136 | * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, |
| 137 | * as ancient (libc5 based) binaries can segfault. ) |
| 138 | */ |
| 139 | int randomize_va_space __read_mostly = |
| 140 | #ifdef CONFIG_COMPAT_BRK |
| 141 | 1; |
| 142 | #else |
| 143 | 2; |
| 144 | #endif |
| 145 | |
| 146 | #ifndef arch_wants_old_prefaulted_pte |
| 147 | static inline bool arch_wants_old_prefaulted_pte(void) |
| 148 | { |
| 149 | /* |
| 150 | * Transitioning a PTE from 'old' to 'young' can be expensive on |
| 151 | * some architectures, even if it's performed in hardware. By |
| 152 | * default, "false" means prefaulted entries will be 'young'. |
| 153 | */ |
| 154 | return false; |
| 155 | } |
| 156 | #endif |
| 157 | |
| 158 | static int __init disable_randmaps(char *s) |
| 159 | { |
| 160 | randomize_va_space = 0; |
| 161 | return 1; |
| 162 | } |
| 163 | __setup("norandmaps", disable_randmaps); |
| 164 | |
| 165 | unsigned long zero_pfn __read_mostly; |
| 166 | EXPORT_SYMBOL(zero_pfn); |
| 167 | |
| 168 | unsigned long highest_memmap_pfn __read_mostly; |
| 169 | |
| 170 | /* |
| 171 | * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() |
| 172 | */ |
| 173 | static int __init init_zero_pfn(void) |
| 174 | { |
| 175 | zero_pfn = page_to_pfn(ZERO_PAGE(0)); |
| 176 | return 0; |
| 177 | } |
| 178 | early_initcall(init_zero_pfn); |
| 179 | |
| 180 | void mm_trace_rss_stat(struct mm_struct *mm, int member) |
| 181 | { |
| 182 | trace_rss_stat(mm, member); |
| 183 | } |
| 184 | |
| 185 | /* |
| 186 | * Note: this doesn't free the actual pages themselves. That |
| 187 | * has been handled earlier when unmapping all the memory regions. |
| 188 | */ |
| 189 | static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, |
| 190 | unsigned long addr) |
| 191 | { |
| 192 | pgtable_t token = pmd_pgtable(*pmd); |
| 193 | pmd_clear(pmd); |
| 194 | pte_free_tlb(tlb, token, addr); |
| 195 | mm_dec_nr_ptes(tlb->mm); |
| 196 | } |
| 197 | |
| 198 | static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
| 199 | unsigned long addr, unsigned long end, |
| 200 | unsigned long floor, unsigned long ceiling) |
| 201 | { |
| 202 | pmd_t *pmd; |
| 203 | unsigned long next; |
| 204 | unsigned long start; |
| 205 | |
| 206 | start = addr; |
| 207 | pmd = pmd_offset(pud, addr); |
| 208 | do { |
| 209 | next = pmd_addr_end(addr, end); |
| 210 | if (pmd_none_or_clear_bad(pmd)) |
| 211 | continue; |
| 212 | free_pte_range(tlb, pmd, addr); |
| 213 | } while (pmd++, addr = next, addr != end); |
| 214 | |
| 215 | start &= PUD_MASK; |
| 216 | if (start < floor) |
| 217 | return; |
| 218 | if (ceiling) { |
| 219 | ceiling &= PUD_MASK; |
| 220 | if (!ceiling) |
| 221 | return; |
| 222 | } |
| 223 | if (end - 1 > ceiling - 1) |
| 224 | return; |
| 225 | |
| 226 | pmd = pmd_offset(pud, start); |
| 227 | pud_clear(pud); |
| 228 | pmd_free_tlb(tlb, pmd, start); |
| 229 | mm_dec_nr_pmds(tlb->mm); |
| 230 | } |
| 231 | |
| 232 | static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, |
| 233 | unsigned long addr, unsigned long end, |
| 234 | unsigned long floor, unsigned long ceiling) |
| 235 | { |
| 236 | pud_t *pud; |
| 237 | unsigned long next; |
| 238 | unsigned long start; |
| 239 | |
| 240 | start = addr; |
| 241 | pud = pud_offset(p4d, addr); |
| 242 | do { |
| 243 | next = pud_addr_end(addr, end); |
| 244 | if (pud_none_or_clear_bad(pud)) |
| 245 | continue; |
| 246 | free_pmd_range(tlb, pud, addr, next, floor, ceiling); |
| 247 | } while (pud++, addr = next, addr != end); |
| 248 | |
| 249 | start &= P4D_MASK; |
| 250 | if (start < floor) |
| 251 | return; |
| 252 | if (ceiling) { |
| 253 | ceiling &= P4D_MASK; |
| 254 | if (!ceiling) |
| 255 | return; |
| 256 | } |
| 257 | if (end - 1 > ceiling - 1) |
| 258 | return; |
| 259 | |
| 260 | pud = pud_offset(p4d, start); |
| 261 | p4d_clear(p4d); |
| 262 | pud_free_tlb(tlb, pud, start); |
| 263 | mm_dec_nr_puds(tlb->mm); |
| 264 | } |
| 265 | |
| 266 | static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, |
| 267 | unsigned long addr, unsigned long end, |
| 268 | unsigned long floor, unsigned long ceiling) |
| 269 | { |
| 270 | p4d_t *p4d; |
| 271 | unsigned long next; |
| 272 | unsigned long start; |
| 273 | |
| 274 | start = addr; |
| 275 | p4d = p4d_offset(pgd, addr); |
| 276 | do { |
| 277 | next = p4d_addr_end(addr, end); |
| 278 | if (p4d_none_or_clear_bad(p4d)) |
| 279 | continue; |
| 280 | free_pud_range(tlb, p4d, addr, next, floor, ceiling); |
| 281 | } while (p4d++, addr = next, addr != end); |
| 282 | |
| 283 | start &= PGDIR_MASK; |
| 284 | if (start < floor) |
| 285 | return; |
| 286 | if (ceiling) { |
| 287 | ceiling &= PGDIR_MASK; |
| 288 | if (!ceiling) |
| 289 | return; |
| 290 | } |
| 291 | if (end - 1 > ceiling - 1) |
| 292 | return; |
| 293 | |
| 294 | p4d = p4d_offset(pgd, start); |
| 295 | pgd_clear(pgd); |
| 296 | p4d_free_tlb(tlb, p4d, start); |
| 297 | } |
| 298 | |
| 299 | /* |
| 300 | * This function frees user-level page tables of a process. |
| 301 | */ |
| 302 | void free_pgd_range(struct mmu_gather *tlb, |
| 303 | unsigned long addr, unsigned long end, |
| 304 | unsigned long floor, unsigned long ceiling) |
| 305 | { |
| 306 | pgd_t *pgd; |
| 307 | unsigned long next; |
| 308 | |
| 309 | /* |
| 310 | * The next few lines have given us lots of grief... |
| 311 | * |
| 312 | * Why are we testing PMD* at this top level? Because often |
| 313 | * there will be no work to do at all, and we'd prefer not to |
| 314 | * go all the way down to the bottom just to discover that. |
| 315 | * |
| 316 | * Why all these "- 1"s? Because 0 represents both the bottom |
| 317 | * of the address space and the top of it (using -1 for the |
| 318 | * top wouldn't help much: the masks would do the wrong thing). |
| 319 | * The rule is that addr 0 and floor 0 refer to the bottom of |
| 320 | * the address space, but end 0 and ceiling 0 refer to the top |
| 321 | * Comparisons need to use "end - 1" and "ceiling - 1" (though |
| 322 | * that end 0 case should be mythical). |
| 323 | * |
| 324 | * Wherever addr is brought up or ceiling brought down, we must |
| 325 | * be careful to reject "the opposite 0" before it confuses the |
| 326 | * subsequent tests. But what about where end is brought down |
| 327 | * by PMD_SIZE below? no, end can't go down to 0 there. |
| 328 | * |
| 329 | * Whereas we round start (addr) and ceiling down, by different |
| 330 | * masks at different levels, in order to test whether a table |
| 331 | * now has no other vmas using it, so can be freed, we don't |
| 332 | * bother to round floor or end up - the tests don't need that. |
| 333 | */ |
| 334 | |
| 335 | addr &= PMD_MASK; |
| 336 | if (addr < floor) { |
| 337 | addr += PMD_SIZE; |
| 338 | if (!addr) |
| 339 | return; |
| 340 | } |
| 341 | if (ceiling) { |
| 342 | ceiling &= PMD_MASK; |
| 343 | if (!ceiling) |
| 344 | return; |
| 345 | } |
| 346 | if (end - 1 > ceiling - 1) |
| 347 | end -= PMD_SIZE; |
| 348 | if (addr > end - 1) |
| 349 | return; |
| 350 | /* |
| 351 | * We add page table cache pages with PAGE_SIZE, |
| 352 | * (see pte_free_tlb()), flush the tlb if we need |
| 353 | */ |
| 354 | tlb_change_page_size(tlb, PAGE_SIZE); |
| 355 | pgd = pgd_offset(tlb->mm, addr); |
| 356 | do { |
| 357 | next = pgd_addr_end(addr, end); |
| 358 | if (pgd_none_or_clear_bad(pgd)) |
| 359 | continue; |
| 360 | free_p4d_range(tlb, pgd, addr, next, floor, ceiling); |
| 361 | } while (pgd++, addr = next, addr != end); |
| 362 | } |
| 363 | |
| 364 | void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas, |
| 365 | struct vm_area_struct *vma, unsigned long floor, |
| 366 | unsigned long ceiling, bool mm_wr_locked) |
| 367 | { |
| 368 | struct unlink_vma_file_batch vb; |
| 369 | |
| 370 | do { |
| 371 | unsigned long addr = vma->vm_start; |
| 372 | struct vm_area_struct *next; |
| 373 | |
| 374 | /* |
| 375 | * Note: USER_PGTABLES_CEILING may be passed as ceiling and may |
| 376 | * be 0. This will underflow and is okay. |
| 377 | */ |
| 378 | next = mas_find(mas, ceiling - 1); |
| 379 | if (unlikely(xa_is_zero(next))) |
| 380 | next = NULL; |
| 381 | |
| 382 | /* |
| 383 | * Hide vma from rmap and truncate_pagecache before freeing |
| 384 | * pgtables |
| 385 | */ |
| 386 | if (mm_wr_locked) |
| 387 | vma_start_write(vma); |
| 388 | unlink_anon_vmas(vma); |
| 389 | |
| 390 | if (is_vm_hugetlb_page(vma)) { |
| 391 | unlink_file_vma(vma); |
| 392 | hugetlb_free_pgd_range(tlb, addr, vma->vm_end, |
| 393 | floor, next ? next->vm_start : ceiling); |
| 394 | } else { |
| 395 | unlink_file_vma_batch_init(&vb); |
| 396 | unlink_file_vma_batch_add(&vb, vma); |
| 397 | |
| 398 | /* |
| 399 | * Optimization: gather nearby vmas into one call down |
| 400 | */ |
| 401 | while (next && next->vm_start <= vma->vm_end + PMD_SIZE |
| 402 | && !is_vm_hugetlb_page(next)) { |
| 403 | vma = next; |
| 404 | next = mas_find(mas, ceiling - 1); |
| 405 | if (unlikely(xa_is_zero(next))) |
| 406 | next = NULL; |
| 407 | if (mm_wr_locked) |
| 408 | vma_start_write(vma); |
| 409 | unlink_anon_vmas(vma); |
| 410 | unlink_file_vma_batch_add(&vb, vma); |
| 411 | } |
| 412 | unlink_file_vma_batch_final(&vb); |
| 413 | free_pgd_range(tlb, addr, vma->vm_end, |
| 414 | floor, next ? next->vm_start : ceiling); |
| 415 | } |
| 416 | vma = next; |
| 417 | } while (vma); |
| 418 | } |
| 419 | |
| 420 | void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte) |
| 421 | { |
| 422 | spinlock_t *ptl = pmd_lock(mm, pmd); |
| 423 | |
| 424 | if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ |
| 425 | mm_inc_nr_ptes(mm); |
| 426 | /* |
| 427 | * Ensure all pte setup (eg. pte page lock and page clearing) are |
| 428 | * visible before the pte is made visible to other CPUs by being |
| 429 | * put into page tables. |
| 430 | * |
| 431 | * The other side of the story is the pointer chasing in the page |
| 432 | * table walking code (when walking the page table without locking; |
| 433 | * ie. most of the time). Fortunately, these data accesses consist |
| 434 | * of a chain of data-dependent loads, meaning most CPUs (alpha |
| 435 | * being the notable exception) will already guarantee loads are |
| 436 | * seen in-order. See the alpha page table accessors for the |
| 437 | * smp_rmb() barriers in page table walking code. |
| 438 | */ |
| 439 | smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ |
| 440 | pmd_populate(mm, pmd, *pte); |
| 441 | *pte = NULL; |
| 442 | } |
| 443 | spin_unlock(ptl); |
| 444 | } |
| 445 | |
| 446 | int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) |
| 447 | { |
| 448 | pgtable_t new = pte_alloc_one(mm); |
| 449 | if (!new) |
| 450 | return -ENOMEM; |
| 451 | |
| 452 | pmd_install(mm, pmd, &new); |
| 453 | if (new) |
| 454 | pte_free(mm, new); |
| 455 | return 0; |
| 456 | } |
| 457 | |
| 458 | int __pte_alloc_kernel(pmd_t *pmd) |
| 459 | { |
| 460 | pte_t *new = pte_alloc_one_kernel(&init_mm); |
| 461 | if (!new) |
| 462 | return -ENOMEM; |
| 463 | |
| 464 | spin_lock(&init_mm.page_table_lock); |
| 465 | if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ |
| 466 | smp_wmb(); /* See comment in pmd_install() */ |
| 467 | pmd_populate_kernel(&init_mm, pmd, new); |
| 468 | new = NULL; |
| 469 | } |
| 470 | spin_unlock(&init_mm.page_table_lock); |
| 471 | if (new) |
| 472 | pte_free_kernel(&init_mm, new); |
| 473 | return 0; |
| 474 | } |
| 475 | |
| 476 | static inline void init_rss_vec(int *rss) |
| 477 | { |
| 478 | memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); |
| 479 | } |
| 480 | |
| 481 | static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) |
| 482 | { |
| 483 | int i; |
| 484 | |
| 485 | for (i = 0; i < NR_MM_COUNTERS; i++) |
| 486 | if (rss[i]) |
| 487 | add_mm_counter(mm, i, rss[i]); |
| 488 | } |
| 489 | |
| 490 | /* |
| 491 | * This function is called to print an error when a bad pte |
| 492 | * is found. For example, we might have a PFN-mapped pte in |
| 493 | * a region that doesn't allow it. |
| 494 | * |
| 495 | * The calling function must still handle the error. |
| 496 | */ |
| 497 | static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, |
| 498 | pte_t pte, struct page *page) |
| 499 | { |
| 500 | pgd_t *pgd = pgd_offset(vma->vm_mm, addr); |
| 501 | p4d_t *p4d = p4d_offset(pgd, addr); |
| 502 | pud_t *pud = pud_offset(p4d, addr); |
| 503 | pmd_t *pmd = pmd_offset(pud, addr); |
| 504 | struct address_space *mapping; |
| 505 | pgoff_t index; |
| 506 | static unsigned long resume; |
| 507 | static unsigned long nr_shown; |
| 508 | static unsigned long nr_unshown; |
| 509 | |
| 510 | /* |
| 511 | * Allow a burst of 60 reports, then keep quiet for that minute; |
| 512 | * or allow a steady drip of one report per second. |
| 513 | */ |
| 514 | if (nr_shown == 60) { |
| 515 | if (time_before(jiffies, resume)) { |
| 516 | nr_unshown++; |
| 517 | return; |
| 518 | } |
| 519 | if (nr_unshown) { |
| 520 | pr_alert("BUG: Bad page map: %lu messages suppressed\n", |
| 521 | nr_unshown); |
| 522 | nr_unshown = 0; |
| 523 | } |
| 524 | nr_shown = 0; |
| 525 | } |
| 526 | if (nr_shown++ == 0) |
| 527 | resume = jiffies + 60 * HZ; |
| 528 | |
| 529 | mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; |
| 530 | index = linear_page_index(vma, addr); |
| 531 | |
| 532 | pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", |
| 533 | current->comm, |
| 534 | (long long)pte_val(pte), (long long)pmd_val(*pmd)); |
| 535 | if (page) |
| 536 | dump_page(page, "bad pte"); |
| 537 | pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", |
| 538 | (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); |
| 539 | pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n", |
| 540 | vma->vm_file, |
| 541 | vma->vm_ops ? vma->vm_ops->fault : NULL, |
| 542 | vma->vm_file ? vma->vm_file->f_op->mmap : NULL, |
| 543 | mapping ? mapping->a_ops->read_folio : NULL); |
| 544 | dump_stack(); |
| 545 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
| 546 | } |
| 547 | |
| 548 | /* |
| 549 | * vm_normal_page -- This function gets the "struct page" associated with a pte. |
| 550 | * |
| 551 | * "Special" mappings do not wish to be associated with a "struct page" (either |
| 552 | * it doesn't exist, or it exists but they don't want to touch it). In this |
| 553 | * case, NULL is returned here. "Normal" mappings do have a struct page. |
| 554 | * |
| 555 | * There are 2 broad cases. Firstly, an architecture may define a pte_special() |
| 556 | * pte bit, in which case this function is trivial. Secondly, an architecture |
| 557 | * may not have a spare pte bit, which requires a more complicated scheme, |
| 558 | * described below. |
| 559 | * |
| 560 | * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a |
| 561 | * special mapping (even if there are underlying and valid "struct pages"). |
| 562 | * COWed pages of a VM_PFNMAP are always normal. |
| 563 | * |
| 564 | * The way we recognize COWed pages within VM_PFNMAP mappings is through the |
| 565 | * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit |
| 566 | * set, and the vm_pgoff will point to the first PFN mapped: thus every special |
| 567 | * mapping will always honor the rule |
| 568 | * |
| 569 | * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) |
| 570 | * |
| 571 | * And for normal mappings this is false. |
| 572 | * |
| 573 | * This restricts such mappings to be a linear translation from virtual address |
| 574 | * to pfn. To get around this restriction, we allow arbitrary mappings so long |
| 575 | * as the vma is not a COW mapping; in that case, we know that all ptes are |
| 576 | * special (because none can have been COWed). |
| 577 | * |
| 578 | * |
| 579 | * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. |
| 580 | * |
| 581 | * VM_MIXEDMAP mappings can likewise contain memory with or without "struct |
| 582 | * page" backing, however the difference is that _all_ pages with a struct |
| 583 | * page (that is, those where pfn_valid is true) are refcounted and considered |
| 584 | * normal pages by the VM. The only exception are zeropages, which are |
| 585 | * *never* refcounted. |
| 586 | * |
| 587 | * The disadvantage is that pages are refcounted (which can be slower and |
| 588 | * simply not an option for some PFNMAP users). The advantage is that we |
| 589 | * don't have to follow the strict linearity rule of PFNMAP mappings in |
| 590 | * order to support COWable mappings. |
| 591 | * |
| 592 | */ |
| 593 | struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, |
| 594 | pte_t pte) |
| 595 | { |
| 596 | unsigned long pfn = pte_pfn(pte); |
| 597 | |
| 598 | if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { |
| 599 | if (likely(!pte_special(pte))) |
| 600 | goto check_pfn; |
| 601 | if (vma->vm_ops && vma->vm_ops->find_special_page) |
| 602 | return vma->vm_ops->find_special_page(vma, addr); |
| 603 | if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) |
| 604 | return NULL; |
| 605 | if (is_zero_pfn(pfn)) |
| 606 | return NULL; |
| 607 | if (pte_devmap(pte)) |
| 608 | /* |
| 609 | * NOTE: New users of ZONE_DEVICE will not set pte_devmap() |
| 610 | * and will have refcounts incremented on their struct pages |
| 611 | * when they are inserted into PTEs, thus they are safe to |
| 612 | * return here. Legacy ZONE_DEVICE pages that set pte_devmap() |
| 613 | * do not have refcounts. Example of legacy ZONE_DEVICE is |
| 614 | * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers. |
| 615 | */ |
| 616 | return NULL; |
| 617 | |
| 618 | print_bad_pte(vma, addr, pte, NULL); |
| 619 | return NULL; |
| 620 | } |
| 621 | |
| 622 | /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ |
| 623 | |
| 624 | if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { |
| 625 | if (vma->vm_flags & VM_MIXEDMAP) { |
| 626 | if (!pfn_valid(pfn)) |
| 627 | return NULL; |
| 628 | if (is_zero_pfn(pfn)) |
| 629 | return NULL; |
| 630 | goto out; |
| 631 | } else { |
| 632 | unsigned long off; |
| 633 | off = (addr - vma->vm_start) >> PAGE_SHIFT; |
| 634 | if (pfn == vma->vm_pgoff + off) |
| 635 | return NULL; |
| 636 | if (!is_cow_mapping(vma->vm_flags)) |
| 637 | return NULL; |
| 638 | } |
| 639 | } |
| 640 | |
| 641 | if (is_zero_pfn(pfn)) |
| 642 | return NULL; |
| 643 | |
| 644 | check_pfn: |
| 645 | if (unlikely(pfn > highest_memmap_pfn)) { |
| 646 | print_bad_pte(vma, addr, pte, NULL); |
| 647 | return NULL; |
| 648 | } |
| 649 | |
| 650 | /* |
| 651 | * NOTE! We still have PageReserved() pages in the page tables. |
| 652 | * eg. VDSO mappings can cause them to exist. |
| 653 | */ |
| 654 | out: |
| 655 | VM_WARN_ON_ONCE(is_zero_pfn(pfn)); |
| 656 | return pfn_to_page(pfn); |
| 657 | } |
| 658 | |
| 659 | struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr, |
| 660 | pte_t pte) |
| 661 | { |
| 662 | struct page *page = vm_normal_page(vma, addr, pte); |
| 663 | |
| 664 | if (page) |
| 665 | return page_folio(page); |
| 666 | return NULL; |
| 667 | } |
| 668 | |
| 669 | #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES |
| 670 | struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, |
| 671 | pmd_t pmd) |
| 672 | { |
| 673 | unsigned long pfn = pmd_pfn(pmd); |
| 674 | |
| 675 | /* Currently it's only used for huge pfnmaps */ |
| 676 | if (unlikely(pmd_special(pmd))) |
| 677 | return NULL; |
| 678 | |
| 679 | if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { |
| 680 | if (vma->vm_flags & VM_MIXEDMAP) { |
| 681 | if (!pfn_valid(pfn)) |
| 682 | return NULL; |
| 683 | goto out; |
| 684 | } else { |
| 685 | unsigned long off; |
| 686 | off = (addr - vma->vm_start) >> PAGE_SHIFT; |
| 687 | if (pfn == vma->vm_pgoff + off) |
| 688 | return NULL; |
| 689 | if (!is_cow_mapping(vma->vm_flags)) |
| 690 | return NULL; |
| 691 | } |
| 692 | } |
| 693 | |
| 694 | if (pmd_devmap(pmd)) |
| 695 | return NULL; |
| 696 | if (is_huge_zero_pmd(pmd)) |
| 697 | return NULL; |
| 698 | if (unlikely(pfn > highest_memmap_pfn)) |
| 699 | return NULL; |
| 700 | |
| 701 | /* |
| 702 | * NOTE! We still have PageReserved() pages in the page tables. |
| 703 | * eg. VDSO mappings can cause them to exist. |
| 704 | */ |
| 705 | out: |
| 706 | return pfn_to_page(pfn); |
| 707 | } |
| 708 | |
| 709 | struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma, |
| 710 | unsigned long addr, pmd_t pmd) |
| 711 | { |
| 712 | struct page *page = vm_normal_page_pmd(vma, addr, pmd); |
| 713 | |
| 714 | if (page) |
| 715 | return page_folio(page); |
| 716 | return NULL; |
| 717 | } |
| 718 | #endif |
| 719 | |
| 720 | static void restore_exclusive_pte(struct vm_area_struct *vma, |
| 721 | struct page *page, unsigned long address, |
| 722 | pte_t *ptep) |
| 723 | { |
| 724 | struct folio *folio = page_folio(page); |
| 725 | pte_t orig_pte; |
| 726 | pte_t pte; |
| 727 | swp_entry_t entry; |
| 728 | |
| 729 | orig_pte = ptep_get(ptep); |
| 730 | pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); |
| 731 | if (pte_swp_soft_dirty(orig_pte)) |
| 732 | pte = pte_mksoft_dirty(pte); |
| 733 | |
| 734 | entry = pte_to_swp_entry(orig_pte); |
| 735 | if (pte_swp_uffd_wp(orig_pte)) |
| 736 | pte = pte_mkuffd_wp(pte); |
| 737 | else if (is_writable_device_exclusive_entry(entry)) |
| 738 | pte = maybe_mkwrite(pte_mkdirty(pte), vma); |
| 739 | |
| 740 | VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) && |
| 741 | PageAnonExclusive(page)), folio); |
| 742 | |
| 743 | /* |
| 744 | * No need to take a page reference as one was already |
| 745 | * created when the swap entry was made. |
| 746 | */ |
| 747 | if (folio_test_anon(folio)) |
| 748 | folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE); |
| 749 | else |
| 750 | /* |
| 751 | * Currently device exclusive access only supports anonymous |
| 752 | * memory so the entry shouldn't point to a filebacked page. |
| 753 | */ |
| 754 | WARN_ON_ONCE(1); |
| 755 | |
| 756 | set_pte_at(vma->vm_mm, address, ptep, pte); |
| 757 | |
| 758 | /* |
| 759 | * No need to invalidate - it was non-present before. However |
| 760 | * secondary CPUs may have mappings that need invalidating. |
| 761 | */ |
| 762 | update_mmu_cache(vma, address, ptep); |
| 763 | } |
| 764 | |
| 765 | /* |
| 766 | * Tries to restore an exclusive pte if the page lock can be acquired without |
| 767 | * sleeping. |
| 768 | */ |
| 769 | static int |
| 770 | try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma, |
| 771 | unsigned long addr) |
| 772 | { |
| 773 | swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte)); |
| 774 | struct page *page = pfn_swap_entry_to_page(entry); |
| 775 | |
| 776 | if (trylock_page(page)) { |
| 777 | restore_exclusive_pte(vma, page, addr, src_pte); |
| 778 | unlock_page(page); |
| 779 | return 0; |
| 780 | } |
| 781 | |
| 782 | return -EBUSY; |
| 783 | } |
| 784 | |
| 785 | /* |
| 786 | * copy one vm_area from one task to the other. Assumes the page tables |
| 787 | * already present in the new task to be cleared in the whole range |
| 788 | * covered by this vma. |
| 789 | */ |
| 790 | |
| 791 | static unsigned long |
| 792 | copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
| 793 | pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma, |
| 794 | struct vm_area_struct *src_vma, unsigned long addr, int *rss) |
| 795 | { |
| 796 | unsigned long vm_flags = dst_vma->vm_flags; |
| 797 | pte_t orig_pte = ptep_get(src_pte); |
| 798 | pte_t pte = orig_pte; |
| 799 | struct folio *folio; |
| 800 | struct page *page; |
| 801 | swp_entry_t entry = pte_to_swp_entry(orig_pte); |
| 802 | |
| 803 | if (likely(!non_swap_entry(entry))) { |
| 804 | if (swap_duplicate(entry) < 0) |
| 805 | return -EIO; |
| 806 | |
| 807 | /* make sure dst_mm is on swapoff's mmlist. */ |
| 808 | if (unlikely(list_empty(&dst_mm->mmlist))) { |
| 809 | spin_lock(&mmlist_lock); |
| 810 | if (list_empty(&dst_mm->mmlist)) |
| 811 | list_add(&dst_mm->mmlist, |
| 812 | &src_mm->mmlist); |
| 813 | spin_unlock(&mmlist_lock); |
| 814 | } |
| 815 | /* Mark the swap entry as shared. */ |
| 816 | if (pte_swp_exclusive(orig_pte)) { |
| 817 | pte = pte_swp_clear_exclusive(orig_pte); |
| 818 | set_pte_at(src_mm, addr, src_pte, pte); |
| 819 | } |
| 820 | rss[MM_SWAPENTS]++; |
| 821 | } else if (is_migration_entry(entry)) { |
| 822 | folio = pfn_swap_entry_folio(entry); |
| 823 | |
| 824 | rss[mm_counter(folio)]++; |
| 825 | |
| 826 | if (!is_readable_migration_entry(entry) && |
| 827 | is_cow_mapping(vm_flags)) { |
| 828 | /* |
| 829 | * COW mappings require pages in both parent and child |
| 830 | * to be set to read. A previously exclusive entry is |
| 831 | * now shared. |
| 832 | */ |
| 833 | entry = make_readable_migration_entry( |
| 834 | swp_offset(entry)); |
| 835 | pte = swp_entry_to_pte(entry); |
| 836 | if (pte_swp_soft_dirty(orig_pte)) |
| 837 | pte = pte_swp_mksoft_dirty(pte); |
| 838 | if (pte_swp_uffd_wp(orig_pte)) |
| 839 | pte = pte_swp_mkuffd_wp(pte); |
| 840 | set_pte_at(src_mm, addr, src_pte, pte); |
| 841 | } |
| 842 | } else if (is_device_private_entry(entry)) { |
| 843 | page = pfn_swap_entry_to_page(entry); |
| 844 | folio = page_folio(page); |
| 845 | |
| 846 | /* |
| 847 | * Update rss count even for unaddressable pages, as |
| 848 | * they should treated just like normal pages in this |
| 849 | * respect. |
| 850 | * |
| 851 | * We will likely want to have some new rss counters |
| 852 | * for unaddressable pages, at some point. But for now |
| 853 | * keep things as they are. |
| 854 | */ |
| 855 | folio_get(folio); |
| 856 | rss[mm_counter(folio)]++; |
| 857 | /* Cannot fail as these pages cannot get pinned. */ |
| 858 | folio_try_dup_anon_rmap_pte(folio, page, src_vma); |
| 859 | |
| 860 | /* |
| 861 | * We do not preserve soft-dirty information, because so |
| 862 | * far, checkpoint/restore is the only feature that |
| 863 | * requires that. And checkpoint/restore does not work |
| 864 | * when a device driver is involved (you cannot easily |
| 865 | * save and restore device driver state). |
| 866 | */ |
| 867 | if (is_writable_device_private_entry(entry) && |
| 868 | is_cow_mapping(vm_flags)) { |
| 869 | entry = make_readable_device_private_entry( |
| 870 | swp_offset(entry)); |
| 871 | pte = swp_entry_to_pte(entry); |
| 872 | if (pte_swp_uffd_wp(orig_pte)) |
| 873 | pte = pte_swp_mkuffd_wp(pte); |
| 874 | set_pte_at(src_mm, addr, src_pte, pte); |
| 875 | } |
| 876 | } else if (is_device_exclusive_entry(entry)) { |
| 877 | /* |
| 878 | * Make device exclusive entries present by restoring the |
| 879 | * original entry then copying as for a present pte. Device |
| 880 | * exclusive entries currently only support private writable |
| 881 | * (ie. COW) mappings. |
| 882 | */ |
| 883 | VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags)); |
| 884 | if (try_restore_exclusive_pte(src_pte, src_vma, addr)) |
| 885 | return -EBUSY; |
| 886 | return -ENOENT; |
| 887 | } else if (is_pte_marker_entry(entry)) { |
| 888 | pte_marker marker = copy_pte_marker(entry, dst_vma); |
| 889 | |
| 890 | if (marker) |
| 891 | set_pte_at(dst_mm, addr, dst_pte, |
| 892 | make_pte_marker(marker)); |
| 893 | return 0; |
| 894 | } |
| 895 | if (!userfaultfd_wp(dst_vma)) |
| 896 | pte = pte_swp_clear_uffd_wp(pte); |
| 897 | set_pte_at(dst_mm, addr, dst_pte, pte); |
| 898 | return 0; |
| 899 | } |
| 900 | |
| 901 | /* |
| 902 | * Copy a present and normal page. |
| 903 | * |
| 904 | * NOTE! The usual case is that this isn't required; |
| 905 | * instead, the caller can just increase the page refcount |
| 906 | * and re-use the pte the traditional way. |
| 907 | * |
| 908 | * And if we need a pre-allocated page but don't yet have |
| 909 | * one, return a negative error to let the preallocation |
| 910 | * code know so that it can do so outside the page table |
| 911 | * lock. |
| 912 | */ |
| 913 | static inline int |
| 914 | copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 915 | pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, |
| 916 | struct folio **prealloc, struct page *page) |
| 917 | { |
| 918 | struct folio *new_folio; |
| 919 | pte_t pte; |
| 920 | |
| 921 | new_folio = *prealloc; |
| 922 | if (!new_folio) |
| 923 | return -EAGAIN; |
| 924 | |
| 925 | /* |
| 926 | * We have a prealloc page, all good! Take it |
| 927 | * over and copy the page & arm it. |
| 928 | */ |
| 929 | |
| 930 | if (copy_mc_user_highpage(&new_folio->page, page, addr, src_vma)) |
| 931 | return -EHWPOISON; |
| 932 | |
| 933 | *prealloc = NULL; |
| 934 | __folio_mark_uptodate(new_folio); |
| 935 | folio_add_new_anon_rmap(new_folio, dst_vma, addr, RMAP_EXCLUSIVE); |
| 936 | folio_add_lru_vma(new_folio, dst_vma); |
| 937 | rss[MM_ANONPAGES]++; |
| 938 | |
| 939 | /* All done, just insert the new page copy in the child */ |
| 940 | pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot); |
| 941 | pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); |
| 942 | if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte))) |
| 943 | /* Uffd-wp needs to be delivered to dest pte as well */ |
| 944 | pte = pte_mkuffd_wp(pte); |
| 945 | set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); |
| 946 | return 0; |
| 947 | } |
| 948 | |
| 949 | static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma, |
| 950 | struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte, |
| 951 | pte_t pte, unsigned long addr, int nr) |
| 952 | { |
| 953 | struct mm_struct *src_mm = src_vma->vm_mm; |
| 954 | |
| 955 | /* If it's a COW mapping, write protect it both processes. */ |
| 956 | if (is_cow_mapping(src_vma->vm_flags) && pte_write(pte)) { |
| 957 | wrprotect_ptes(src_mm, addr, src_pte, nr); |
| 958 | pte = pte_wrprotect(pte); |
| 959 | } |
| 960 | |
| 961 | /* If it's a shared mapping, mark it clean in the child. */ |
| 962 | if (src_vma->vm_flags & VM_SHARED) |
| 963 | pte = pte_mkclean(pte); |
| 964 | pte = pte_mkold(pte); |
| 965 | |
| 966 | if (!userfaultfd_wp(dst_vma)) |
| 967 | pte = pte_clear_uffd_wp(pte); |
| 968 | |
| 969 | set_ptes(dst_vma->vm_mm, addr, dst_pte, pte, nr); |
| 970 | } |
| 971 | |
| 972 | /* |
| 973 | * Copy one present PTE, trying to batch-process subsequent PTEs that map |
| 974 | * consecutive pages of the same folio by copying them as well. |
| 975 | * |
| 976 | * Returns -EAGAIN if one preallocated page is required to copy the next PTE. |
| 977 | * Otherwise, returns the number of copied PTEs (at least 1). |
| 978 | */ |
| 979 | static inline int |
| 980 | copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 981 | pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr, |
| 982 | int max_nr, int *rss, struct folio **prealloc) |
| 983 | { |
| 984 | struct page *page; |
| 985 | struct folio *folio; |
| 986 | bool any_writable; |
| 987 | fpb_t flags = 0; |
| 988 | int err, nr; |
| 989 | |
| 990 | page = vm_normal_page(src_vma, addr, pte); |
| 991 | if (unlikely(!page)) |
| 992 | goto copy_pte; |
| 993 | |
| 994 | folio = page_folio(page); |
| 995 | |
| 996 | /* |
| 997 | * If we likely have to copy, just don't bother with batching. Make |
| 998 | * sure that the common "small folio" case is as fast as possible |
| 999 | * by keeping the batching logic separate. |
| 1000 | */ |
| 1001 | if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) { |
| 1002 | if (src_vma->vm_flags & VM_SHARED) |
| 1003 | flags |= FPB_IGNORE_DIRTY; |
| 1004 | if (!vma_soft_dirty_enabled(src_vma)) |
| 1005 | flags |= FPB_IGNORE_SOFT_DIRTY; |
| 1006 | |
| 1007 | nr = folio_pte_batch(folio, addr, src_pte, pte, max_nr, flags, |
| 1008 | &any_writable, NULL, NULL); |
| 1009 | folio_ref_add(folio, nr); |
| 1010 | if (folio_test_anon(folio)) { |
| 1011 | if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page, |
| 1012 | nr, src_vma))) { |
| 1013 | folio_ref_sub(folio, nr); |
| 1014 | return -EAGAIN; |
| 1015 | } |
| 1016 | rss[MM_ANONPAGES] += nr; |
| 1017 | VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio); |
| 1018 | } else { |
| 1019 | folio_dup_file_rmap_ptes(folio, page, nr); |
| 1020 | rss[mm_counter_file(folio)] += nr; |
| 1021 | } |
| 1022 | if (any_writable) |
| 1023 | pte = pte_mkwrite(pte, src_vma); |
| 1024 | __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, |
| 1025 | addr, nr); |
| 1026 | return nr; |
| 1027 | } |
| 1028 | |
| 1029 | folio_get(folio); |
| 1030 | if (folio_test_anon(folio)) { |
| 1031 | /* |
| 1032 | * If this page may have been pinned by the parent process, |
| 1033 | * copy the page immediately for the child so that we'll always |
| 1034 | * guarantee the pinned page won't be randomly replaced in the |
| 1035 | * future. |
| 1036 | */ |
| 1037 | if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) { |
| 1038 | /* Page may be pinned, we have to copy. */ |
| 1039 | folio_put(folio); |
| 1040 | err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte, |
| 1041 | addr, rss, prealloc, page); |
| 1042 | return err ? err : 1; |
| 1043 | } |
| 1044 | rss[MM_ANONPAGES]++; |
| 1045 | VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio); |
| 1046 | } else { |
| 1047 | folio_dup_file_rmap_pte(folio, page); |
| 1048 | rss[mm_counter_file(folio)]++; |
| 1049 | } |
| 1050 | |
| 1051 | copy_pte: |
| 1052 | __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, 1); |
| 1053 | return 1; |
| 1054 | } |
| 1055 | |
| 1056 | static inline struct folio *folio_prealloc(struct mm_struct *src_mm, |
| 1057 | struct vm_area_struct *vma, unsigned long addr, bool need_zero) |
| 1058 | { |
| 1059 | struct folio *new_folio; |
| 1060 | |
| 1061 | if (need_zero) |
| 1062 | new_folio = vma_alloc_zeroed_movable_folio(vma, addr); |
| 1063 | else |
| 1064 | new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, |
| 1065 | addr, false); |
| 1066 | |
| 1067 | if (!new_folio) |
| 1068 | return NULL; |
| 1069 | |
| 1070 | if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) { |
| 1071 | folio_put(new_folio); |
| 1072 | return NULL; |
| 1073 | } |
| 1074 | folio_throttle_swaprate(new_folio, GFP_KERNEL); |
| 1075 | |
| 1076 | return new_folio; |
| 1077 | } |
| 1078 | |
| 1079 | static int |
| 1080 | copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 1081 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, |
| 1082 | unsigned long end) |
| 1083 | { |
| 1084 | struct mm_struct *dst_mm = dst_vma->vm_mm; |
| 1085 | struct mm_struct *src_mm = src_vma->vm_mm; |
| 1086 | pte_t *orig_src_pte, *orig_dst_pte; |
| 1087 | pte_t *src_pte, *dst_pte; |
| 1088 | pte_t ptent; |
| 1089 | spinlock_t *src_ptl, *dst_ptl; |
| 1090 | int progress, max_nr, ret = 0; |
| 1091 | int rss[NR_MM_COUNTERS]; |
| 1092 | swp_entry_t entry = (swp_entry_t){0}; |
| 1093 | struct folio *prealloc = NULL; |
| 1094 | int nr; |
| 1095 | |
| 1096 | again: |
| 1097 | progress = 0; |
| 1098 | init_rss_vec(rss); |
| 1099 | |
| 1100 | /* |
| 1101 | * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the |
| 1102 | * error handling here, assume that exclusive mmap_lock on dst and src |
| 1103 | * protects anon from unexpected THP transitions; with shmem and file |
| 1104 | * protected by mmap_lock-less collapse skipping areas with anon_vma |
| 1105 | * (whereas vma_needs_copy() skips areas without anon_vma). A rework |
| 1106 | * can remove such assumptions later, but this is good enough for now. |
| 1107 | */ |
| 1108 | dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); |
| 1109 | if (!dst_pte) { |
| 1110 | ret = -ENOMEM; |
| 1111 | goto out; |
| 1112 | } |
| 1113 | src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl); |
| 1114 | if (!src_pte) { |
| 1115 | pte_unmap_unlock(dst_pte, dst_ptl); |
| 1116 | /* ret == 0 */ |
| 1117 | goto out; |
| 1118 | } |
| 1119 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); |
| 1120 | orig_src_pte = src_pte; |
| 1121 | orig_dst_pte = dst_pte; |
| 1122 | arch_enter_lazy_mmu_mode(); |
| 1123 | |
| 1124 | do { |
| 1125 | nr = 1; |
| 1126 | |
| 1127 | /* |
| 1128 | * We are holding two locks at this point - either of them |
| 1129 | * could generate latencies in another task on another CPU. |
| 1130 | */ |
| 1131 | if (progress >= 32) { |
| 1132 | progress = 0; |
| 1133 | if (need_resched() || |
| 1134 | spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) |
| 1135 | break; |
| 1136 | } |
| 1137 | ptent = ptep_get(src_pte); |
| 1138 | if (pte_none(ptent)) { |
| 1139 | progress++; |
| 1140 | continue; |
| 1141 | } |
| 1142 | if (unlikely(!pte_present(ptent))) { |
| 1143 | ret = copy_nonpresent_pte(dst_mm, src_mm, |
| 1144 | dst_pte, src_pte, |
| 1145 | dst_vma, src_vma, |
| 1146 | addr, rss); |
| 1147 | if (ret == -EIO) { |
| 1148 | entry = pte_to_swp_entry(ptep_get(src_pte)); |
| 1149 | break; |
| 1150 | } else if (ret == -EBUSY) { |
| 1151 | break; |
| 1152 | } else if (!ret) { |
| 1153 | progress += 8; |
| 1154 | continue; |
| 1155 | } |
| 1156 | ptent = ptep_get(src_pte); |
| 1157 | VM_WARN_ON_ONCE(!pte_present(ptent)); |
| 1158 | |
| 1159 | /* |
| 1160 | * Device exclusive entry restored, continue by copying |
| 1161 | * the now present pte. |
| 1162 | */ |
| 1163 | WARN_ON_ONCE(ret != -ENOENT); |
| 1164 | } |
| 1165 | /* copy_present_ptes() will clear `*prealloc' if consumed */ |
| 1166 | max_nr = (end - addr) / PAGE_SIZE; |
| 1167 | ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, |
| 1168 | ptent, addr, max_nr, rss, &prealloc); |
| 1169 | /* |
| 1170 | * If we need a pre-allocated page for this pte, drop the |
| 1171 | * locks, allocate, and try again. |
| 1172 | * If copy failed due to hwpoison in source page, break out. |
| 1173 | */ |
| 1174 | if (unlikely(ret == -EAGAIN || ret == -EHWPOISON)) |
| 1175 | break; |
| 1176 | if (unlikely(prealloc)) { |
| 1177 | /* |
| 1178 | * pre-alloc page cannot be reused by next time so as |
| 1179 | * to strictly follow mempolicy (e.g., alloc_page_vma() |
| 1180 | * will allocate page according to address). This |
| 1181 | * could only happen if one pinned pte changed. |
| 1182 | */ |
| 1183 | folio_put(prealloc); |
| 1184 | prealloc = NULL; |
| 1185 | } |
| 1186 | nr = ret; |
| 1187 | progress += 8 * nr; |
| 1188 | } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr, |
| 1189 | addr != end); |
| 1190 | |
| 1191 | arch_leave_lazy_mmu_mode(); |
| 1192 | pte_unmap_unlock(orig_src_pte, src_ptl); |
| 1193 | add_mm_rss_vec(dst_mm, rss); |
| 1194 | pte_unmap_unlock(orig_dst_pte, dst_ptl); |
| 1195 | cond_resched(); |
| 1196 | |
| 1197 | if (ret == -EIO) { |
| 1198 | VM_WARN_ON_ONCE(!entry.val); |
| 1199 | if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { |
| 1200 | ret = -ENOMEM; |
| 1201 | goto out; |
| 1202 | } |
| 1203 | entry.val = 0; |
| 1204 | } else if (ret == -EBUSY || unlikely(ret == -EHWPOISON)) { |
| 1205 | goto out; |
| 1206 | } else if (ret == -EAGAIN) { |
| 1207 | prealloc = folio_prealloc(src_mm, src_vma, addr, false); |
| 1208 | if (!prealloc) |
| 1209 | return -ENOMEM; |
| 1210 | } else if (ret < 0) { |
| 1211 | VM_WARN_ON_ONCE(1); |
| 1212 | } |
| 1213 | |
| 1214 | /* We've captured and resolved the error. Reset, try again. */ |
| 1215 | ret = 0; |
| 1216 | |
| 1217 | if (addr != end) |
| 1218 | goto again; |
| 1219 | out: |
| 1220 | if (unlikely(prealloc)) |
| 1221 | folio_put(prealloc); |
| 1222 | return ret; |
| 1223 | } |
| 1224 | |
| 1225 | static inline int |
| 1226 | copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 1227 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, |
| 1228 | unsigned long end) |
| 1229 | { |
| 1230 | struct mm_struct *dst_mm = dst_vma->vm_mm; |
| 1231 | struct mm_struct *src_mm = src_vma->vm_mm; |
| 1232 | pmd_t *src_pmd, *dst_pmd; |
| 1233 | unsigned long next; |
| 1234 | |
| 1235 | dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); |
| 1236 | if (!dst_pmd) |
| 1237 | return -ENOMEM; |
| 1238 | src_pmd = pmd_offset(src_pud, addr); |
| 1239 | do { |
| 1240 | next = pmd_addr_end(addr, end); |
| 1241 | if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) |
| 1242 | || pmd_devmap(*src_pmd)) { |
| 1243 | int err; |
| 1244 | VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); |
| 1245 | err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd, |
| 1246 | addr, dst_vma, src_vma); |
| 1247 | if (err == -ENOMEM) |
| 1248 | return -ENOMEM; |
| 1249 | if (!err) |
| 1250 | continue; |
| 1251 | /* fall through */ |
| 1252 | } |
| 1253 | if (pmd_none_or_clear_bad(src_pmd)) |
| 1254 | continue; |
| 1255 | if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, |
| 1256 | addr, next)) |
| 1257 | return -ENOMEM; |
| 1258 | } while (dst_pmd++, src_pmd++, addr = next, addr != end); |
| 1259 | return 0; |
| 1260 | } |
| 1261 | |
| 1262 | static inline int |
| 1263 | copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 1264 | p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, |
| 1265 | unsigned long end) |
| 1266 | { |
| 1267 | struct mm_struct *dst_mm = dst_vma->vm_mm; |
| 1268 | struct mm_struct *src_mm = src_vma->vm_mm; |
| 1269 | pud_t *src_pud, *dst_pud; |
| 1270 | unsigned long next; |
| 1271 | |
| 1272 | dst_pud = pud_alloc(dst_mm, dst_p4d, addr); |
| 1273 | if (!dst_pud) |
| 1274 | return -ENOMEM; |
| 1275 | src_pud = pud_offset(src_p4d, addr); |
| 1276 | do { |
| 1277 | next = pud_addr_end(addr, end); |
| 1278 | if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { |
| 1279 | int err; |
| 1280 | |
| 1281 | VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); |
| 1282 | err = copy_huge_pud(dst_mm, src_mm, |
| 1283 | dst_pud, src_pud, addr, src_vma); |
| 1284 | if (err == -ENOMEM) |
| 1285 | return -ENOMEM; |
| 1286 | if (!err) |
| 1287 | continue; |
| 1288 | /* fall through */ |
| 1289 | } |
| 1290 | if (pud_none_or_clear_bad(src_pud)) |
| 1291 | continue; |
| 1292 | if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, |
| 1293 | addr, next)) |
| 1294 | return -ENOMEM; |
| 1295 | } while (dst_pud++, src_pud++, addr = next, addr != end); |
| 1296 | return 0; |
| 1297 | } |
| 1298 | |
| 1299 | static inline int |
| 1300 | copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, |
| 1301 | pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, |
| 1302 | unsigned long end) |
| 1303 | { |
| 1304 | struct mm_struct *dst_mm = dst_vma->vm_mm; |
| 1305 | p4d_t *src_p4d, *dst_p4d; |
| 1306 | unsigned long next; |
| 1307 | |
| 1308 | dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); |
| 1309 | if (!dst_p4d) |
| 1310 | return -ENOMEM; |
| 1311 | src_p4d = p4d_offset(src_pgd, addr); |
| 1312 | do { |
| 1313 | next = p4d_addr_end(addr, end); |
| 1314 | if (p4d_none_or_clear_bad(src_p4d)) |
| 1315 | continue; |
| 1316 | if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, |
| 1317 | addr, next)) |
| 1318 | return -ENOMEM; |
| 1319 | } while (dst_p4d++, src_p4d++, addr = next, addr != end); |
| 1320 | return 0; |
| 1321 | } |
| 1322 | |
| 1323 | /* |
| 1324 | * Return true if the vma needs to copy the pgtable during this fork(). Return |
| 1325 | * false when we can speed up fork() by allowing lazy page faults later until |
| 1326 | * when the child accesses the memory range. |
| 1327 | */ |
| 1328 | static bool |
| 1329 | vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) |
| 1330 | { |
| 1331 | /* |
| 1332 | * Always copy pgtables when dst_vma has uffd-wp enabled even if it's |
| 1333 | * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable |
| 1334 | * contains uffd-wp protection information, that's something we can't |
| 1335 | * retrieve from page cache, and skip copying will lose those info. |
| 1336 | */ |
| 1337 | if (userfaultfd_wp(dst_vma)) |
| 1338 | return true; |
| 1339 | |
| 1340 | if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) |
| 1341 | return true; |
| 1342 | |
| 1343 | if (src_vma->anon_vma) |
| 1344 | return true; |
| 1345 | |
| 1346 | /* |
| 1347 | * Don't copy ptes where a page fault will fill them correctly. Fork |
| 1348 | * becomes much lighter when there are big shared or private readonly |
| 1349 | * mappings. The tradeoff is that copy_page_range is more efficient |
| 1350 | * than faulting. |
| 1351 | */ |
| 1352 | return false; |
| 1353 | } |
| 1354 | |
| 1355 | int |
| 1356 | copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) |
| 1357 | { |
| 1358 | pgd_t *src_pgd, *dst_pgd; |
| 1359 | unsigned long next; |
| 1360 | unsigned long addr = src_vma->vm_start; |
| 1361 | unsigned long end = src_vma->vm_end; |
| 1362 | struct mm_struct *dst_mm = dst_vma->vm_mm; |
| 1363 | struct mm_struct *src_mm = src_vma->vm_mm; |
| 1364 | struct mmu_notifier_range range; |
| 1365 | bool is_cow; |
| 1366 | int ret; |
| 1367 | |
| 1368 | if (!vma_needs_copy(dst_vma, src_vma)) |
| 1369 | return 0; |
| 1370 | |
| 1371 | if (is_vm_hugetlb_page(src_vma)) |
| 1372 | return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma); |
| 1373 | |
| 1374 | if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { |
| 1375 | /* |
| 1376 | * We do not free on error cases below as remove_vma |
| 1377 | * gets called on error from higher level routine |
| 1378 | */ |
| 1379 | ret = track_pfn_copy(src_vma); |
| 1380 | if (ret) |
| 1381 | return ret; |
| 1382 | } |
| 1383 | |
| 1384 | /* |
| 1385 | * We need to invalidate the secondary MMU mappings only when |
| 1386 | * there could be a permission downgrade on the ptes of the |
| 1387 | * parent mm. And a permission downgrade will only happen if |
| 1388 | * is_cow_mapping() returns true. |
| 1389 | */ |
| 1390 | is_cow = is_cow_mapping(src_vma->vm_flags); |
| 1391 | |
| 1392 | if (is_cow) { |
| 1393 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, |
| 1394 | 0, src_mm, addr, end); |
| 1395 | mmu_notifier_invalidate_range_start(&range); |
| 1396 | /* |
| 1397 | * Disabling preemption is not needed for the write side, as |
| 1398 | * the read side doesn't spin, but goes to the mmap_lock. |
| 1399 | * |
| 1400 | * Use the raw variant of the seqcount_t write API to avoid |
| 1401 | * lockdep complaining about preemptibility. |
| 1402 | */ |
| 1403 | vma_assert_write_locked(src_vma); |
| 1404 | raw_write_seqcount_begin(&src_mm->write_protect_seq); |
| 1405 | } |
| 1406 | |
| 1407 | ret = 0; |
| 1408 | dst_pgd = pgd_offset(dst_mm, addr); |
| 1409 | src_pgd = pgd_offset(src_mm, addr); |
| 1410 | do { |
| 1411 | next = pgd_addr_end(addr, end); |
| 1412 | if (pgd_none_or_clear_bad(src_pgd)) |
| 1413 | continue; |
| 1414 | if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, |
| 1415 | addr, next))) { |
| 1416 | untrack_pfn_clear(dst_vma); |
| 1417 | ret = -ENOMEM; |
| 1418 | break; |
| 1419 | } |
| 1420 | } while (dst_pgd++, src_pgd++, addr = next, addr != end); |
| 1421 | |
| 1422 | if (is_cow) { |
| 1423 | raw_write_seqcount_end(&src_mm->write_protect_seq); |
| 1424 | mmu_notifier_invalidate_range_end(&range); |
| 1425 | } |
| 1426 | return ret; |
| 1427 | } |
| 1428 | |
| 1429 | /* Whether we should zap all COWed (private) pages too */ |
| 1430 | static inline bool should_zap_cows(struct zap_details *details) |
| 1431 | { |
| 1432 | /* By default, zap all pages */ |
| 1433 | if (!details) |
| 1434 | return true; |
| 1435 | |
| 1436 | /* Or, we zap COWed pages only if the caller wants to */ |
| 1437 | return details->even_cows; |
| 1438 | } |
| 1439 | |
| 1440 | /* Decides whether we should zap this folio with the folio pointer specified */ |
| 1441 | static inline bool should_zap_folio(struct zap_details *details, |
| 1442 | struct folio *folio) |
| 1443 | { |
| 1444 | /* If we can make a decision without *folio.. */ |
| 1445 | if (should_zap_cows(details)) |
| 1446 | return true; |
| 1447 | |
| 1448 | /* Otherwise we should only zap non-anon folios */ |
| 1449 | return !folio_test_anon(folio); |
| 1450 | } |
| 1451 | |
| 1452 | static inline bool zap_drop_file_uffd_wp(struct zap_details *details) |
| 1453 | { |
| 1454 | if (!details) |
| 1455 | return false; |
| 1456 | |
| 1457 | return details->zap_flags & ZAP_FLAG_DROP_MARKER; |
| 1458 | } |
| 1459 | |
| 1460 | /* |
| 1461 | * This function makes sure that we'll replace the none pte with an uffd-wp |
| 1462 | * swap special pte marker when necessary. Must be with the pgtable lock held. |
| 1463 | */ |
| 1464 | static inline void |
| 1465 | zap_install_uffd_wp_if_needed(struct vm_area_struct *vma, |
| 1466 | unsigned long addr, pte_t *pte, int nr, |
| 1467 | struct zap_details *details, pte_t pteval) |
| 1468 | { |
| 1469 | /* Zap on anonymous always means dropping everything */ |
| 1470 | if (vma_is_anonymous(vma)) |
| 1471 | return; |
| 1472 | |
| 1473 | if (zap_drop_file_uffd_wp(details)) |
| 1474 | return; |
| 1475 | |
| 1476 | for (;;) { |
| 1477 | /* the PFN in the PTE is irrelevant. */ |
| 1478 | pte_install_uffd_wp_if_needed(vma, addr, pte, pteval); |
| 1479 | if (--nr == 0) |
| 1480 | break; |
| 1481 | pte++; |
| 1482 | addr += PAGE_SIZE; |
| 1483 | } |
| 1484 | } |
| 1485 | |
| 1486 | static __always_inline void zap_present_folio_ptes(struct mmu_gather *tlb, |
| 1487 | struct vm_area_struct *vma, struct folio *folio, |
| 1488 | struct page *page, pte_t *pte, pte_t ptent, unsigned int nr, |
| 1489 | unsigned long addr, struct zap_details *details, int *rss, |
| 1490 | bool *force_flush, bool *force_break) |
| 1491 | { |
| 1492 | struct mm_struct *mm = tlb->mm; |
| 1493 | bool delay_rmap = false; |
| 1494 | |
| 1495 | if (!folio_test_anon(folio)) { |
| 1496 | ptent = get_and_clear_full_ptes(mm, addr, pte, nr, tlb->fullmm); |
| 1497 | if (pte_dirty(ptent)) { |
| 1498 | folio_mark_dirty(folio); |
| 1499 | if (tlb_delay_rmap(tlb)) { |
| 1500 | delay_rmap = true; |
| 1501 | *force_flush = true; |
| 1502 | } |
| 1503 | } |
| 1504 | if (pte_young(ptent) && likely(vma_has_recency(vma))) |
| 1505 | folio_mark_accessed(folio); |
| 1506 | rss[mm_counter(folio)] -= nr; |
| 1507 | } else { |
| 1508 | /* We don't need up-to-date accessed/dirty bits. */ |
| 1509 | clear_full_ptes(mm, addr, pte, nr, tlb->fullmm); |
| 1510 | rss[MM_ANONPAGES] -= nr; |
| 1511 | } |
| 1512 | /* Checking a single PTE in a batch is sufficient. */ |
| 1513 | arch_check_zapped_pte(vma, ptent); |
| 1514 | tlb_remove_tlb_entries(tlb, pte, nr, addr); |
| 1515 | if (unlikely(userfaultfd_pte_wp(vma, ptent))) |
| 1516 | zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, |
| 1517 | ptent); |
| 1518 | |
| 1519 | if (!delay_rmap) { |
| 1520 | folio_remove_rmap_ptes(folio, page, nr, vma); |
| 1521 | |
| 1522 | if (unlikely(folio_mapcount(folio) < 0)) |
| 1523 | print_bad_pte(vma, addr, ptent, page); |
| 1524 | } |
| 1525 | if (unlikely(__tlb_remove_folio_pages(tlb, page, nr, delay_rmap))) { |
| 1526 | *force_flush = true; |
| 1527 | *force_break = true; |
| 1528 | } |
| 1529 | } |
| 1530 | |
| 1531 | /* |
| 1532 | * Zap or skip at least one present PTE, trying to batch-process subsequent |
| 1533 | * PTEs that map consecutive pages of the same folio. |
| 1534 | * |
| 1535 | * Returns the number of processed (skipped or zapped) PTEs (at least 1). |
| 1536 | */ |
| 1537 | static inline int zap_present_ptes(struct mmu_gather *tlb, |
| 1538 | struct vm_area_struct *vma, pte_t *pte, pte_t ptent, |
| 1539 | unsigned int max_nr, unsigned long addr, |
| 1540 | struct zap_details *details, int *rss, bool *force_flush, |
| 1541 | bool *force_break) |
| 1542 | { |
| 1543 | const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY; |
| 1544 | struct mm_struct *mm = tlb->mm; |
| 1545 | struct folio *folio; |
| 1546 | struct page *page; |
| 1547 | int nr; |
| 1548 | |
| 1549 | page = vm_normal_page(vma, addr, ptent); |
| 1550 | if (!page) { |
| 1551 | /* We don't need up-to-date accessed/dirty bits. */ |
| 1552 | ptep_get_and_clear_full(mm, addr, pte, tlb->fullmm); |
| 1553 | arch_check_zapped_pte(vma, ptent); |
| 1554 | tlb_remove_tlb_entry(tlb, pte, addr); |
| 1555 | if (userfaultfd_pte_wp(vma, ptent)) |
| 1556 | zap_install_uffd_wp_if_needed(vma, addr, pte, 1, |
| 1557 | details, ptent); |
| 1558 | ksm_might_unmap_zero_page(mm, ptent); |
| 1559 | return 1; |
| 1560 | } |
| 1561 | |
| 1562 | folio = page_folio(page); |
| 1563 | if (unlikely(!should_zap_folio(details, folio))) |
| 1564 | return 1; |
| 1565 | |
| 1566 | /* |
| 1567 | * Make sure that the common "small folio" case is as fast as possible |
| 1568 | * by keeping the batching logic separate. |
| 1569 | */ |
| 1570 | if (unlikely(folio_test_large(folio) && max_nr != 1)) { |
| 1571 | nr = folio_pte_batch(folio, addr, pte, ptent, max_nr, fpb_flags, |
| 1572 | NULL, NULL, NULL); |
| 1573 | |
| 1574 | zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr, |
| 1575 | addr, details, rss, force_flush, |
| 1576 | force_break); |
| 1577 | return nr; |
| 1578 | } |
| 1579 | zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, 1, addr, |
| 1580 | details, rss, force_flush, force_break); |
| 1581 | return 1; |
| 1582 | } |
| 1583 | |
| 1584 | static unsigned long zap_pte_range(struct mmu_gather *tlb, |
| 1585 | struct vm_area_struct *vma, pmd_t *pmd, |
| 1586 | unsigned long addr, unsigned long end, |
| 1587 | struct zap_details *details) |
| 1588 | { |
| 1589 | bool force_flush = false, force_break = false; |
| 1590 | struct mm_struct *mm = tlb->mm; |
| 1591 | int rss[NR_MM_COUNTERS]; |
| 1592 | spinlock_t *ptl; |
| 1593 | pte_t *start_pte; |
| 1594 | pte_t *pte; |
| 1595 | swp_entry_t entry; |
| 1596 | int nr; |
| 1597 | |
| 1598 | tlb_change_page_size(tlb, PAGE_SIZE); |
| 1599 | init_rss_vec(rss); |
| 1600 | start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 1601 | if (!pte) |
| 1602 | return addr; |
| 1603 | |
| 1604 | flush_tlb_batched_pending(mm); |
| 1605 | arch_enter_lazy_mmu_mode(); |
| 1606 | do { |
| 1607 | pte_t ptent = ptep_get(pte); |
| 1608 | struct folio *folio; |
| 1609 | struct page *page; |
| 1610 | int max_nr; |
| 1611 | |
| 1612 | nr = 1; |
| 1613 | if (pte_none(ptent)) |
| 1614 | continue; |
| 1615 | |
| 1616 | if (need_resched()) |
| 1617 | break; |
| 1618 | |
| 1619 | if (pte_present(ptent)) { |
| 1620 | max_nr = (end - addr) / PAGE_SIZE; |
| 1621 | nr = zap_present_ptes(tlb, vma, pte, ptent, max_nr, |
| 1622 | addr, details, rss, &force_flush, |
| 1623 | &force_break); |
| 1624 | if (unlikely(force_break)) { |
| 1625 | addr += nr * PAGE_SIZE; |
| 1626 | break; |
| 1627 | } |
| 1628 | continue; |
| 1629 | } |
| 1630 | |
| 1631 | entry = pte_to_swp_entry(ptent); |
| 1632 | if (is_device_private_entry(entry) || |
| 1633 | is_device_exclusive_entry(entry)) { |
| 1634 | page = pfn_swap_entry_to_page(entry); |
| 1635 | folio = page_folio(page); |
| 1636 | if (unlikely(!should_zap_folio(details, folio))) |
| 1637 | continue; |
| 1638 | /* |
| 1639 | * Both device private/exclusive mappings should only |
| 1640 | * work with anonymous page so far, so we don't need to |
| 1641 | * consider uffd-wp bit when zap. For more information, |
| 1642 | * see zap_install_uffd_wp_if_needed(). |
| 1643 | */ |
| 1644 | WARN_ON_ONCE(!vma_is_anonymous(vma)); |
| 1645 | rss[mm_counter(folio)]--; |
| 1646 | if (is_device_private_entry(entry)) |
| 1647 | folio_remove_rmap_pte(folio, page, vma); |
| 1648 | folio_put(folio); |
| 1649 | } else if (!non_swap_entry(entry)) { |
| 1650 | max_nr = (end - addr) / PAGE_SIZE; |
| 1651 | nr = swap_pte_batch(pte, max_nr, ptent); |
| 1652 | /* Genuine swap entries, hence a private anon pages */ |
| 1653 | if (!should_zap_cows(details)) |
| 1654 | continue; |
| 1655 | rss[MM_SWAPENTS] -= nr; |
| 1656 | free_swap_and_cache_nr(entry, nr); |
| 1657 | } else if (is_migration_entry(entry)) { |
| 1658 | folio = pfn_swap_entry_folio(entry); |
| 1659 | if (!should_zap_folio(details, folio)) |
| 1660 | continue; |
| 1661 | rss[mm_counter(folio)]--; |
| 1662 | } else if (pte_marker_entry_uffd_wp(entry)) { |
| 1663 | /* |
| 1664 | * For anon: always drop the marker; for file: only |
| 1665 | * drop the marker if explicitly requested. |
| 1666 | */ |
| 1667 | if (!vma_is_anonymous(vma) && |
| 1668 | !zap_drop_file_uffd_wp(details)) |
| 1669 | continue; |
| 1670 | } else if (is_hwpoison_entry(entry) || |
| 1671 | is_poisoned_swp_entry(entry)) { |
| 1672 | if (!should_zap_cows(details)) |
| 1673 | continue; |
| 1674 | } else { |
| 1675 | /* We should have covered all the swap entry types */ |
| 1676 | pr_alert("unrecognized swap entry 0x%lx\n", entry.val); |
| 1677 | WARN_ON_ONCE(1); |
| 1678 | } |
| 1679 | clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm); |
| 1680 | zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, ptent); |
| 1681 | } while (pte += nr, addr += PAGE_SIZE * nr, addr != end); |
| 1682 | |
| 1683 | add_mm_rss_vec(mm, rss); |
| 1684 | arch_leave_lazy_mmu_mode(); |
| 1685 | |
| 1686 | /* Do the actual TLB flush before dropping ptl */ |
| 1687 | if (force_flush) { |
| 1688 | tlb_flush_mmu_tlbonly(tlb); |
| 1689 | tlb_flush_rmaps(tlb, vma); |
| 1690 | } |
| 1691 | pte_unmap_unlock(start_pte, ptl); |
| 1692 | |
| 1693 | /* |
| 1694 | * If we forced a TLB flush (either due to running out of |
| 1695 | * batch buffers or because we needed to flush dirty TLB |
| 1696 | * entries before releasing the ptl), free the batched |
| 1697 | * memory too. Come back again if we didn't do everything. |
| 1698 | */ |
| 1699 | if (force_flush) |
| 1700 | tlb_flush_mmu(tlb); |
| 1701 | |
| 1702 | return addr; |
| 1703 | } |
| 1704 | |
| 1705 | static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, |
| 1706 | struct vm_area_struct *vma, pud_t *pud, |
| 1707 | unsigned long addr, unsigned long end, |
| 1708 | struct zap_details *details) |
| 1709 | { |
| 1710 | pmd_t *pmd; |
| 1711 | unsigned long next; |
| 1712 | |
| 1713 | pmd = pmd_offset(pud, addr); |
| 1714 | do { |
| 1715 | next = pmd_addr_end(addr, end); |
| 1716 | if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { |
| 1717 | if (next - addr != HPAGE_PMD_SIZE) |
| 1718 | __split_huge_pmd(vma, pmd, addr, false, NULL); |
| 1719 | else if (zap_huge_pmd(tlb, vma, pmd, addr)) { |
| 1720 | addr = next; |
| 1721 | continue; |
| 1722 | } |
| 1723 | /* fall through */ |
| 1724 | } else if (details && details->single_folio && |
| 1725 | folio_test_pmd_mappable(details->single_folio) && |
| 1726 | next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) { |
| 1727 | spinlock_t *ptl = pmd_lock(tlb->mm, pmd); |
| 1728 | /* |
| 1729 | * Take and drop THP pmd lock so that we cannot return |
| 1730 | * prematurely, while zap_huge_pmd() has cleared *pmd, |
| 1731 | * but not yet decremented compound_mapcount(). |
| 1732 | */ |
| 1733 | spin_unlock(ptl); |
| 1734 | } |
| 1735 | if (pmd_none(*pmd)) { |
| 1736 | addr = next; |
| 1737 | continue; |
| 1738 | } |
| 1739 | addr = zap_pte_range(tlb, vma, pmd, addr, next, details); |
| 1740 | if (addr != next) |
| 1741 | pmd--; |
| 1742 | } while (pmd++, cond_resched(), addr != end); |
| 1743 | |
| 1744 | return addr; |
| 1745 | } |
| 1746 | |
| 1747 | static inline unsigned long zap_pud_range(struct mmu_gather *tlb, |
| 1748 | struct vm_area_struct *vma, p4d_t *p4d, |
| 1749 | unsigned long addr, unsigned long end, |
| 1750 | struct zap_details *details) |
| 1751 | { |
| 1752 | pud_t *pud; |
| 1753 | unsigned long next; |
| 1754 | |
| 1755 | pud = pud_offset(p4d, addr); |
| 1756 | do { |
| 1757 | next = pud_addr_end(addr, end); |
| 1758 | if (pud_trans_huge(*pud) || pud_devmap(*pud)) { |
| 1759 | if (next - addr != HPAGE_PUD_SIZE) { |
| 1760 | mmap_assert_locked(tlb->mm); |
| 1761 | split_huge_pud(vma, pud, addr); |
| 1762 | } else if (zap_huge_pud(tlb, vma, pud, addr)) |
| 1763 | goto next; |
| 1764 | /* fall through */ |
| 1765 | } |
| 1766 | if (pud_none_or_clear_bad(pud)) |
| 1767 | continue; |
| 1768 | next = zap_pmd_range(tlb, vma, pud, addr, next, details); |
| 1769 | next: |
| 1770 | cond_resched(); |
| 1771 | } while (pud++, addr = next, addr != end); |
| 1772 | |
| 1773 | return addr; |
| 1774 | } |
| 1775 | |
| 1776 | static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, |
| 1777 | struct vm_area_struct *vma, pgd_t *pgd, |
| 1778 | unsigned long addr, unsigned long end, |
| 1779 | struct zap_details *details) |
| 1780 | { |
| 1781 | p4d_t *p4d; |
| 1782 | unsigned long next; |
| 1783 | |
| 1784 | p4d = p4d_offset(pgd, addr); |
| 1785 | do { |
| 1786 | next = p4d_addr_end(addr, end); |
| 1787 | if (p4d_none_or_clear_bad(p4d)) |
| 1788 | continue; |
| 1789 | next = zap_pud_range(tlb, vma, p4d, addr, next, details); |
| 1790 | } while (p4d++, addr = next, addr != end); |
| 1791 | |
| 1792 | return addr; |
| 1793 | } |
| 1794 | |
| 1795 | void unmap_page_range(struct mmu_gather *tlb, |
| 1796 | struct vm_area_struct *vma, |
| 1797 | unsigned long addr, unsigned long end, |
| 1798 | struct zap_details *details) |
| 1799 | { |
| 1800 | pgd_t *pgd; |
| 1801 | unsigned long next; |
| 1802 | |
| 1803 | BUG_ON(addr >= end); |
| 1804 | tlb_start_vma(tlb, vma); |
| 1805 | pgd = pgd_offset(vma->vm_mm, addr); |
| 1806 | do { |
| 1807 | next = pgd_addr_end(addr, end); |
| 1808 | if (pgd_none_or_clear_bad(pgd)) |
| 1809 | continue; |
| 1810 | next = zap_p4d_range(tlb, vma, pgd, addr, next, details); |
| 1811 | } while (pgd++, addr = next, addr != end); |
| 1812 | tlb_end_vma(tlb, vma); |
| 1813 | } |
| 1814 | |
| 1815 | |
| 1816 | static void unmap_single_vma(struct mmu_gather *tlb, |
| 1817 | struct vm_area_struct *vma, unsigned long start_addr, |
| 1818 | unsigned long end_addr, |
| 1819 | struct zap_details *details, bool mm_wr_locked) |
| 1820 | { |
| 1821 | unsigned long start = max(vma->vm_start, start_addr); |
| 1822 | unsigned long end; |
| 1823 | |
| 1824 | if (start >= vma->vm_end) |
| 1825 | return; |
| 1826 | end = min(vma->vm_end, end_addr); |
| 1827 | if (end <= vma->vm_start) |
| 1828 | return; |
| 1829 | |
| 1830 | if (vma->vm_file) |
| 1831 | uprobe_munmap(vma, start, end); |
| 1832 | |
| 1833 | if (unlikely(vma->vm_flags & VM_PFNMAP)) |
| 1834 | untrack_pfn(vma, 0, 0, mm_wr_locked); |
| 1835 | |
| 1836 | if (start != end) { |
| 1837 | if (unlikely(is_vm_hugetlb_page(vma))) { |
| 1838 | /* |
| 1839 | * It is undesirable to test vma->vm_file as it |
| 1840 | * should be non-null for valid hugetlb area. |
| 1841 | * However, vm_file will be NULL in the error |
| 1842 | * cleanup path of mmap_region. When |
| 1843 | * hugetlbfs ->mmap method fails, |
| 1844 | * mmap_region() nullifies vma->vm_file |
| 1845 | * before calling this function to clean up. |
| 1846 | * Since no pte has actually been setup, it is |
| 1847 | * safe to do nothing in this case. |
| 1848 | */ |
| 1849 | if (vma->vm_file) { |
| 1850 | zap_flags_t zap_flags = details ? |
| 1851 | details->zap_flags : 0; |
| 1852 | __unmap_hugepage_range(tlb, vma, start, end, |
| 1853 | NULL, zap_flags); |
| 1854 | } |
| 1855 | } else |
| 1856 | unmap_page_range(tlb, vma, start, end, details); |
| 1857 | } |
| 1858 | } |
| 1859 | |
| 1860 | /** |
| 1861 | * unmap_vmas - unmap a range of memory covered by a list of vma's |
| 1862 | * @tlb: address of the caller's struct mmu_gather |
| 1863 | * @mas: the maple state |
| 1864 | * @vma: the starting vma |
| 1865 | * @start_addr: virtual address at which to start unmapping |
| 1866 | * @end_addr: virtual address at which to end unmapping |
| 1867 | * @tree_end: The maximum index to check |
| 1868 | * @mm_wr_locked: lock flag |
| 1869 | * |
| 1870 | * Unmap all pages in the vma list. |
| 1871 | * |
| 1872 | * Only addresses between `start' and `end' will be unmapped. |
| 1873 | * |
| 1874 | * The VMA list must be sorted in ascending virtual address order. |
| 1875 | * |
| 1876 | * unmap_vmas() assumes that the caller will flush the whole unmapped address |
| 1877 | * range after unmap_vmas() returns. So the only responsibility here is to |
| 1878 | * ensure that any thus-far unmapped pages are flushed before unmap_vmas() |
| 1879 | * drops the lock and schedules. |
| 1880 | */ |
| 1881 | void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas, |
| 1882 | struct vm_area_struct *vma, unsigned long start_addr, |
| 1883 | unsigned long end_addr, unsigned long tree_end, |
| 1884 | bool mm_wr_locked) |
| 1885 | { |
| 1886 | struct mmu_notifier_range range; |
| 1887 | struct zap_details details = { |
| 1888 | .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP, |
| 1889 | /* Careful - we need to zap private pages too! */ |
| 1890 | .even_cows = true, |
| 1891 | }; |
| 1892 | |
| 1893 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm, |
| 1894 | start_addr, end_addr); |
| 1895 | mmu_notifier_invalidate_range_start(&range); |
| 1896 | do { |
| 1897 | unsigned long start = start_addr; |
| 1898 | unsigned long end = end_addr; |
| 1899 | hugetlb_zap_begin(vma, &start, &end); |
| 1900 | unmap_single_vma(tlb, vma, start, end, &details, |
| 1901 | mm_wr_locked); |
| 1902 | hugetlb_zap_end(vma, &details); |
| 1903 | vma = mas_find(mas, tree_end - 1); |
| 1904 | } while (vma && likely(!xa_is_zero(vma))); |
| 1905 | mmu_notifier_invalidate_range_end(&range); |
| 1906 | } |
| 1907 | |
| 1908 | /** |
| 1909 | * zap_page_range_single - remove user pages in a given range |
| 1910 | * @vma: vm_area_struct holding the applicable pages |
| 1911 | * @address: starting address of pages to zap |
| 1912 | * @size: number of bytes to zap |
| 1913 | * @details: details of shared cache invalidation |
| 1914 | * |
| 1915 | * The range must fit into one VMA. |
| 1916 | */ |
| 1917 | void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, |
| 1918 | unsigned long size, struct zap_details *details) |
| 1919 | { |
| 1920 | const unsigned long end = address + size; |
| 1921 | struct mmu_notifier_range range; |
| 1922 | struct mmu_gather tlb; |
| 1923 | |
| 1924 | lru_add_drain(); |
| 1925 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, |
| 1926 | address, end); |
| 1927 | hugetlb_zap_begin(vma, &range.start, &range.end); |
| 1928 | tlb_gather_mmu(&tlb, vma->vm_mm); |
| 1929 | update_hiwater_rss(vma->vm_mm); |
| 1930 | mmu_notifier_invalidate_range_start(&range); |
| 1931 | /* |
| 1932 | * unmap 'address-end' not 'range.start-range.end' as range |
| 1933 | * could have been expanded for hugetlb pmd sharing. |
| 1934 | */ |
| 1935 | unmap_single_vma(&tlb, vma, address, end, details, false); |
| 1936 | mmu_notifier_invalidate_range_end(&range); |
| 1937 | tlb_finish_mmu(&tlb); |
| 1938 | hugetlb_zap_end(vma, details); |
| 1939 | } |
| 1940 | |
| 1941 | /** |
| 1942 | * zap_vma_ptes - remove ptes mapping the vma |
| 1943 | * @vma: vm_area_struct holding ptes to be zapped |
| 1944 | * @address: starting address of pages to zap |
| 1945 | * @size: number of bytes to zap |
| 1946 | * |
| 1947 | * This function only unmaps ptes assigned to VM_PFNMAP vmas. |
| 1948 | * |
| 1949 | * The entire address range must be fully contained within the vma. |
| 1950 | * |
| 1951 | */ |
| 1952 | void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, |
| 1953 | unsigned long size) |
| 1954 | { |
| 1955 | if (!range_in_vma(vma, address, address + size) || |
| 1956 | !(vma->vm_flags & VM_PFNMAP)) |
| 1957 | return; |
| 1958 | |
| 1959 | zap_page_range_single(vma, address, size, NULL); |
| 1960 | } |
| 1961 | EXPORT_SYMBOL_GPL(zap_vma_ptes); |
| 1962 | |
| 1963 | static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) |
| 1964 | { |
| 1965 | pgd_t *pgd; |
| 1966 | p4d_t *p4d; |
| 1967 | pud_t *pud; |
| 1968 | pmd_t *pmd; |
| 1969 | |
| 1970 | pgd = pgd_offset(mm, addr); |
| 1971 | p4d = p4d_alloc(mm, pgd, addr); |
| 1972 | if (!p4d) |
| 1973 | return NULL; |
| 1974 | pud = pud_alloc(mm, p4d, addr); |
| 1975 | if (!pud) |
| 1976 | return NULL; |
| 1977 | pmd = pmd_alloc(mm, pud, addr); |
| 1978 | if (!pmd) |
| 1979 | return NULL; |
| 1980 | |
| 1981 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
| 1982 | return pmd; |
| 1983 | } |
| 1984 | |
| 1985 | pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, |
| 1986 | spinlock_t **ptl) |
| 1987 | { |
| 1988 | pmd_t *pmd = walk_to_pmd(mm, addr); |
| 1989 | |
| 1990 | if (!pmd) |
| 1991 | return NULL; |
| 1992 | return pte_alloc_map_lock(mm, pmd, addr, ptl); |
| 1993 | } |
| 1994 | |
| 1995 | static bool vm_mixed_zeropage_allowed(struct vm_area_struct *vma) |
| 1996 | { |
| 1997 | VM_WARN_ON_ONCE(vma->vm_flags & VM_PFNMAP); |
| 1998 | /* |
| 1999 | * Whoever wants to forbid the zeropage after some zeropages |
| 2000 | * might already have been mapped has to scan the page tables and |
| 2001 | * bail out on any zeropages. Zeropages in COW mappings can |
| 2002 | * be unshared using FAULT_FLAG_UNSHARE faults. |
| 2003 | */ |
| 2004 | if (mm_forbids_zeropage(vma->vm_mm)) |
| 2005 | return false; |
| 2006 | /* zeropages in COW mappings are common and unproblematic. */ |
| 2007 | if (is_cow_mapping(vma->vm_flags)) |
| 2008 | return true; |
| 2009 | /* Mappings that do not allow for writable PTEs are unproblematic. */ |
| 2010 | if (!(vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) |
| 2011 | return true; |
| 2012 | /* |
| 2013 | * Why not allow any VMA that has vm_ops->pfn_mkwrite? GUP could |
| 2014 | * find the shared zeropage and longterm-pin it, which would |
| 2015 | * be problematic as soon as the zeropage gets replaced by a different |
| 2016 | * page due to vma->vm_ops->pfn_mkwrite, because what's mapped would |
| 2017 | * now differ to what GUP looked up. FSDAX is incompatible to |
| 2018 | * FOLL_LONGTERM and VM_IO is incompatible to GUP completely (see |
| 2019 | * check_vma_flags). |
| 2020 | */ |
| 2021 | return vma->vm_ops && vma->vm_ops->pfn_mkwrite && |
| 2022 | (vma_is_fsdax(vma) || vma->vm_flags & VM_IO); |
| 2023 | } |
| 2024 | |
| 2025 | static int validate_page_before_insert(struct vm_area_struct *vma, |
| 2026 | struct page *page) |
| 2027 | { |
| 2028 | struct folio *folio = page_folio(page); |
| 2029 | |
| 2030 | if (!folio_ref_count(folio)) |
| 2031 | return -EINVAL; |
| 2032 | if (unlikely(is_zero_folio(folio))) { |
| 2033 | if (!vm_mixed_zeropage_allowed(vma)) |
| 2034 | return -EINVAL; |
| 2035 | return 0; |
| 2036 | } |
| 2037 | if (folio_test_anon(folio) || folio_test_slab(folio) || |
| 2038 | page_has_type(page)) |
| 2039 | return -EINVAL; |
| 2040 | flush_dcache_folio(folio); |
| 2041 | return 0; |
| 2042 | } |
| 2043 | |
| 2044 | static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte, |
| 2045 | unsigned long addr, struct page *page, pgprot_t prot) |
| 2046 | { |
| 2047 | struct folio *folio = page_folio(page); |
| 2048 | pte_t pteval; |
| 2049 | |
| 2050 | if (!pte_none(ptep_get(pte))) |
| 2051 | return -EBUSY; |
| 2052 | /* Ok, finally just insert the thing.. */ |
| 2053 | pteval = mk_pte(page, prot); |
| 2054 | if (unlikely(is_zero_folio(folio))) { |
| 2055 | pteval = pte_mkspecial(pteval); |
| 2056 | } else { |
| 2057 | folio_get(folio); |
| 2058 | inc_mm_counter(vma->vm_mm, mm_counter_file(folio)); |
| 2059 | folio_add_file_rmap_pte(folio, page, vma); |
| 2060 | } |
| 2061 | set_pte_at(vma->vm_mm, addr, pte, pteval); |
| 2062 | return 0; |
| 2063 | } |
| 2064 | |
| 2065 | static int insert_page(struct vm_area_struct *vma, unsigned long addr, |
| 2066 | struct page *page, pgprot_t prot) |
| 2067 | { |
| 2068 | int retval; |
| 2069 | pte_t *pte; |
| 2070 | spinlock_t *ptl; |
| 2071 | |
| 2072 | retval = validate_page_before_insert(vma, page); |
| 2073 | if (retval) |
| 2074 | goto out; |
| 2075 | retval = -ENOMEM; |
| 2076 | pte = get_locked_pte(vma->vm_mm, addr, &ptl); |
| 2077 | if (!pte) |
| 2078 | goto out; |
| 2079 | retval = insert_page_into_pte_locked(vma, pte, addr, page, prot); |
| 2080 | pte_unmap_unlock(pte, ptl); |
| 2081 | out: |
| 2082 | return retval; |
| 2083 | } |
| 2084 | |
| 2085 | static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte, |
| 2086 | unsigned long addr, struct page *page, pgprot_t prot) |
| 2087 | { |
| 2088 | int err; |
| 2089 | |
| 2090 | err = validate_page_before_insert(vma, page); |
| 2091 | if (err) |
| 2092 | return err; |
| 2093 | return insert_page_into_pte_locked(vma, pte, addr, page, prot); |
| 2094 | } |
| 2095 | |
| 2096 | /* insert_pages() amortizes the cost of spinlock operations |
| 2097 | * when inserting pages in a loop. |
| 2098 | */ |
| 2099 | static int insert_pages(struct vm_area_struct *vma, unsigned long addr, |
| 2100 | struct page **pages, unsigned long *num, pgprot_t prot) |
| 2101 | { |
| 2102 | pmd_t *pmd = NULL; |
| 2103 | pte_t *start_pte, *pte; |
| 2104 | spinlock_t *pte_lock; |
| 2105 | struct mm_struct *const mm = vma->vm_mm; |
| 2106 | unsigned long curr_page_idx = 0; |
| 2107 | unsigned long remaining_pages_total = *num; |
| 2108 | unsigned long pages_to_write_in_pmd; |
| 2109 | int ret; |
| 2110 | more: |
| 2111 | ret = -EFAULT; |
| 2112 | pmd = walk_to_pmd(mm, addr); |
| 2113 | if (!pmd) |
| 2114 | goto out; |
| 2115 | |
| 2116 | pages_to_write_in_pmd = min_t(unsigned long, |
| 2117 | remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); |
| 2118 | |
| 2119 | /* Allocate the PTE if necessary; takes PMD lock once only. */ |
| 2120 | ret = -ENOMEM; |
| 2121 | if (pte_alloc(mm, pmd)) |
| 2122 | goto out; |
| 2123 | |
| 2124 | while (pages_to_write_in_pmd) { |
| 2125 | int pte_idx = 0; |
| 2126 | const int batch_size = min_t(int, pages_to_write_in_pmd, 8); |
| 2127 | |
| 2128 | start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); |
| 2129 | if (!start_pte) { |
| 2130 | ret = -EFAULT; |
| 2131 | goto out; |
| 2132 | } |
| 2133 | for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { |
| 2134 | int err = insert_page_in_batch_locked(vma, pte, |
| 2135 | addr, pages[curr_page_idx], prot); |
| 2136 | if (unlikely(err)) { |
| 2137 | pte_unmap_unlock(start_pte, pte_lock); |
| 2138 | ret = err; |
| 2139 | remaining_pages_total -= pte_idx; |
| 2140 | goto out; |
| 2141 | } |
| 2142 | addr += PAGE_SIZE; |
| 2143 | ++curr_page_idx; |
| 2144 | } |
| 2145 | pte_unmap_unlock(start_pte, pte_lock); |
| 2146 | pages_to_write_in_pmd -= batch_size; |
| 2147 | remaining_pages_total -= batch_size; |
| 2148 | } |
| 2149 | if (remaining_pages_total) |
| 2150 | goto more; |
| 2151 | ret = 0; |
| 2152 | out: |
| 2153 | *num = remaining_pages_total; |
| 2154 | return ret; |
| 2155 | } |
| 2156 | |
| 2157 | /** |
| 2158 | * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. |
| 2159 | * @vma: user vma to map to |
| 2160 | * @addr: target start user address of these pages |
| 2161 | * @pages: source kernel pages |
| 2162 | * @num: in: number of pages to map. out: number of pages that were *not* |
| 2163 | * mapped. (0 means all pages were successfully mapped). |
| 2164 | * |
| 2165 | * Preferred over vm_insert_page() when inserting multiple pages. |
| 2166 | * |
| 2167 | * In case of error, we may have mapped a subset of the provided |
| 2168 | * pages. It is the caller's responsibility to account for this case. |
| 2169 | * |
| 2170 | * The same restrictions apply as in vm_insert_page(). |
| 2171 | */ |
| 2172 | int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, |
| 2173 | struct page **pages, unsigned long *num) |
| 2174 | { |
| 2175 | const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; |
| 2176 | |
| 2177 | if (addr < vma->vm_start || end_addr >= vma->vm_end) |
| 2178 | return -EFAULT; |
| 2179 | if (!(vma->vm_flags & VM_MIXEDMAP)) { |
| 2180 | BUG_ON(mmap_read_trylock(vma->vm_mm)); |
| 2181 | BUG_ON(vma->vm_flags & VM_PFNMAP); |
| 2182 | vm_flags_set(vma, VM_MIXEDMAP); |
| 2183 | } |
| 2184 | /* Defer page refcount checking till we're about to map that page. */ |
| 2185 | return insert_pages(vma, addr, pages, num, vma->vm_page_prot); |
| 2186 | } |
| 2187 | EXPORT_SYMBOL(vm_insert_pages); |
| 2188 | |
| 2189 | /** |
| 2190 | * vm_insert_page - insert single page into user vma |
| 2191 | * @vma: user vma to map to |
| 2192 | * @addr: target user address of this page |
| 2193 | * @page: source kernel page |
| 2194 | * |
| 2195 | * This allows drivers to insert individual pages they've allocated |
| 2196 | * into a user vma. The zeropage is supported in some VMAs, |
| 2197 | * see vm_mixed_zeropage_allowed(). |
| 2198 | * |
| 2199 | * The page has to be a nice clean _individual_ kernel allocation. |
| 2200 | * If you allocate a compound page, you need to have marked it as |
| 2201 | * such (__GFP_COMP), or manually just split the page up yourself |
| 2202 | * (see split_page()). |
| 2203 | * |
| 2204 | * NOTE! Traditionally this was done with "remap_pfn_range()" which |
| 2205 | * took an arbitrary page protection parameter. This doesn't allow |
| 2206 | * that. Your vma protection will have to be set up correctly, which |
| 2207 | * means that if you want a shared writable mapping, you'd better |
| 2208 | * ask for a shared writable mapping! |
| 2209 | * |
| 2210 | * The page does not need to be reserved. |
| 2211 | * |
| 2212 | * Usually this function is called from f_op->mmap() handler |
| 2213 | * under mm->mmap_lock write-lock, so it can change vma->vm_flags. |
| 2214 | * Caller must set VM_MIXEDMAP on vma if it wants to call this |
| 2215 | * function from other places, for example from page-fault handler. |
| 2216 | * |
| 2217 | * Return: %0 on success, negative error code otherwise. |
| 2218 | */ |
| 2219 | int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, |
| 2220 | struct page *page) |
| 2221 | { |
| 2222 | if (addr < vma->vm_start || addr >= vma->vm_end) |
| 2223 | return -EFAULT; |
| 2224 | if (!(vma->vm_flags & VM_MIXEDMAP)) { |
| 2225 | BUG_ON(mmap_read_trylock(vma->vm_mm)); |
| 2226 | BUG_ON(vma->vm_flags & VM_PFNMAP); |
| 2227 | vm_flags_set(vma, VM_MIXEDMAP); |
| 2228 | } |
| 2229 | return insert_page(vma, addr, page, vma->vm_page_prot); |
| 2230 | } |
| 2231 | EXPORT_SYMBOL(vm_insert_page); |
| 2232 | |
| 2233 | /* |
| 2234 | * __vm_map_pages - maps range of kernel pages into user vma |
| 2235 | * @vma: user vma to map to |
| 2236 | * @pages: pointer to array of source kernel pages |
| 2237 | * @num: number of pages in page array |
| 2238 | * @offset: user's requested vm_pgoff |
| 2239 | * |
| 2240 | * This allows drivers to map range of kernel pages into a user vma. |
| 2241 | * The zeropage is supported in some VMAs, see |
| 2242 | * vm_mixed_zeropage_allowed(). |
| 2243 | * |
| 2244 | * Return: 0 on success and error code otherwise. |
| 2245 | */ |
| 2246 | static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, |
| 2247 | unsigned long num, unsigned long offset) |
| 2248 | { |
| 2249 | unsigned long count = vma_pages(vma); |
| 2250 | unsigned long uaddr = vma->vm_start; |
| 2251 | int ret, i; |
| 2252 | |
| 2253 | /* Fail if the user requested offset is beyond the end of the object */ |
| 2254 | if (offset >= num) |
| 2255 | return -ENXIO; |
| 2256 | |
| 2257 | /* Fail if the user requested size exceeds available object size */ |
| 2258 | if (count > num - offset) |
| 2259 | return -ENXIO; |
| 2260 | |
| 2261 | for (i = 0; i < count; i++) { |
| 2262 | ret = vm_insert_page(vma, uaddr, pages[offset + i]); |
| 2263 | if (ret < 0) |
| 2264 | return ret; |
| 2265 | uaddr += PAGE_SIZE; |
| 2266 | } |
| 2267 | |
| 2268 | return 0; |
| 2269 | } |
| 2270 | |
| 2271 | /** |
| 2272 | * vm_map_pages - maps range of kernel pages starts with non zero offset |
| 2273 | * @vma: user vma to map to |
| 2274 | * @pages: pointer to array of source kernel pages |
| 2275 | * @num: number of pages in page array |
| 2276 | * |
| 2277 | * Maps an object consisting of @num pages, catering for the user's |
| 2278 | * requested vm_pgoff |
| 2279 | * |
| 2280 | * If we fail to insert any page into the vma, the function will return |
| 2281 | * immediately leaving any previously inserted pages present. Callers |
| 2282 | * from the mmap handler may immediately return the error as their caller |
| 2283 | * will destroy the vma, removing any successfully inserted pages. Other |
| 2284 | * callers should make their own arrangements for calling unmap_region(). |
| 2285 | * |
| 2286 | * Context: Process context. Called by mmap handlers. |
| 2287 | * Return: 0 on success and error code otherwise. |
| 2288 | */ |
| 2289 | int vm_map_pages(struct vm_area_struct *vma, struct page **pages, |
| 2290 | unsigned long num) |
| 2291 | { |
| 2292 | return __vm_map_pages(vma, pages, num, vma->vm_pgoff); |
| 2293 | } |
| 2294 | EXPORT_SYMBOL(vm_map_pages); |
| 2295 | |
| 2296 | /** |
| 2297 | * vm_map_pages_zero - map range of kernel pages starts with zero offset |
| 2298 | * @vma: user vma to map to |
| 2299 | * @pages: pointer to array of source kernel pages |
| 2300 | * @num: number of pages in page array |
| 2301 | * |
| 2302 | * Similar to vm_map_pages(), except that it explicitly sets the offset |
| 2303 | * to 0. This function is intended for the drivers that did not consider |
| 2304 | * vm_pgoff. |
| 2305 | * |
| 2306 | * Context: Process context. Called by mmap handlers. |
| 2307 | * Return: 0 on success and error code otherwise. |
| 2308 | */ |
| 2309 | int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, |
| 2310 | unsigned long num) |
| 2311 | { |
| 2312 | return __vm_map_pages(vma, pages, num, 0); |
| 2313 | } |
| 2314 | EXPORT_SYMBOL(vm_map_pages_zero); |
| 2315 | |
| 2316 | static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, |
| 2317 | pfn_t pfn, pgprot_t prot, bool mkwrite) |
| 2318 | { |
| 2319 | struct mm_struct *mm = vma->vm_mm; |
| 2320 | pte_t *pte, entry; |
| 2321 | spinlock_t *ptl; |
| 2322 | |
| 2323 | pte = get_locked_pte(mm, addr, &ptl); |
| 2324 | if (!pte) |
| 2325 | return VM_FAULT_OOM; |
| 2326 | entry = ptep_get(pte); |
| 2327 | if (!pte_none(entry)) { |
| 2328 | if (mkwrite) { |
| 2329 | /* |
| 2330 | * For read faults on private mappings the PFN passed |
| 2331 | * in may not match the PFN we have mapped if the |
| 2332 | * mapped PFN is a writeable COW page. In the mkwrite |
| 2333 | * case we are creating a writable PTE for a shared |
| 2334 | * mapping and we expect the PFNs to match. If they |
| 2335 | * don't match, we are likely racing with block |
| 2336 | * allocation and mapping invalidation so just skip the |
| 2337 | * update. |
| 2338 | */ |
| 2339 | if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) { |
| 2340 | WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry))); |
| 2341 | goto out_unlock; |
| 2342 | } |
| 2343 | entry = pte_mkyoung(entry); |
| 2344 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 2345 | if (ptep_set_access_flags(vma, addr, pte, entry, 1)) |
| 2346 | update_mmu_cache(vma, addr, pte); |
| 2347 | } |
| 2348 | goto out_unlock; |
| 2349 | } |
| 2350 | |
| 2351 | /* Ok, finally just insert the thing.. */ |
| 2352 | if (pfn_t_devmap(pfn)) |
| 2353 | entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); |
| 2354 | else |
| 2355 | entry = pte_mkspecial(pfn_t_pte(pfn, prot)); |
| 2356 | |
| 2357 | if (mkwrite) { |
| 2358 | entry = pte_mkyoung(entry); |
| 2359 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 2360 | } |
| 2361 | |
| 2362 | set_pte_at(mm, addr, pte, entry); |
| 2363 | update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ |
| 2364 | |
| 2365 | out_unlock: |
| 2366 | pte_unmap_unlock(pte, ptl); |
| 2367 | return VM_FAULT_NOPAGE; |
| 2368 | } |
| 2369 | |
| 2370 | /** |
| 2371 | * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot |
| 2372 | * @vma: user vma to map to |
| 2373 | * @addr: target user address of this page |
| 2374 | * @pfn: source kernel pfn |
| 2375 | * @pgprot: pgprot flags for the inserted page |
| 2376 | * |
| 2377 | * This is exactly like vmf_insert_pfn(), except that it allows drivers |
| 2378 | * to override pgprot on a per-page basis. |
| 2379 | * |
| 2380 | * This only makes sense for IO mappings, and it makes no sense for |
| 2381 | * COW mappings. In general, using multiple vmas is preferable; |
| 2382 | * vmf_insert_pfn_prot should only be used if using multiple VMAs is |
| 2383 | * impractical. |
| 2384 | * |
| 2385 | * pgprot typically only differs from @vma->vm_page_prot when drivers set |
| 2386 | * caching- and encryption bits different than those of @vma->vm_page_prot, |
| 2387 | * because the caching- or encryption mode may not be known at mmap() time. |
| 2388 | * |
| 2389 | * This is ok as long as @vma->vm_page_prot is not used by the core vm |
| 2390 | * to set caching and encryption bits for those vmas (except for COW pages). |
| 2391 | * This is ensured by core vm only modifying these page table entries using |
| 2392 | * functions that don't touch caching- or encryption bits, using pte_modify() |
| 2393 | * if needed. (See for example mprotect()). |
| 2394 | * |
| 2395 | * Also when new page-table entries are created, this is only done using the |
| 2396 | * fault() callback, and never using the value of vma->vm_page_prot, |
| 2397 | * except for page-table entries that point to anonymous pages as the result |
| 2398 | * of COW. |
| 2399 | * |
| 2400 | * Context: Process context. May allocate using %GFP_KERNEL. |
| 2401 | * Return: vm_fault_t value. |
| 2402 | */ |
| 2403 | vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, |
| 2404 | unsigned long pfn, pgprot_t pgprot) |
| 2405 | { |
| 2406 | /* |
| 2407 | * Technically, architectures with pte_special can avoid all these |
| 2408 | * restrictions (same for remap_pfn_range). However we would like |
| 2409 | * consistency in testing and feature parity among all, so we should |
| 2410 | * try to keep these invariants in place for everybody. |
| 2411 | */ |
| 2412 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); |
| 2413 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
| 2414 | (VM_PFNMAP|VM_MIXEDMAP)); |
| 2415 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); |
| 2416 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); |
| 2417 | |
| 2418 | if (addr < vma->vm_start || addr >= vma->vm_end) |
| 2419 | return VM_FAULT_SIGBUS; |
| 2420 | |
| 2421 | if (!pfn_modify_allowed(pfn, pgprot)) |
| 2422 | return VM_FAULT_SIGBUS; |
| 2423 | |
| 2424 | track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); |
| 2425 | |
| 2426 | return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, |
| 2427 | false); |
| 2428 | } |
| 2429 | EXPORT_SYMBOL(vmf_insert_pfn_prot); |
| 2430 | |
| 2431 | /** |
| 2432 | * vmf_insert_pfn - insert single pfn into user vma |
| 2433 | * @vma: user vma to map to |
| 2434 | * @addr: target user address of this page |
| 2435 | * @pfn: source kernel pfn |
| 2436 | * |
| 2437 | * Similar to vm_insert_page, this allows drivers to insert individual pages |
| 2438 | * they've allocated into a user vma. Same comments apply. |
| 2439 | * |
| 2440 | * This function should only be called from a vm_ops->fault handler, and |
| 2441 | * in that case the handler should return the result of this function. |
| 2442 | * |
| 2443 | * vma cannot be a COW mapping. |
| 2444 | * |
| 2445 | * As this is called only for pages that do not currently exist, we |
| 2446 | * do not need to flush old virtual caches or the TLB. |
| 2447 | * |
| 2448 | * Context: Process context. May allocate using %GFP_KERNEL. |
| 2449 | * Return: vm_fault_t value. |
| 2450 | */ |
| 2451 | vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, |
| 2452 | unsigned long pfn) |
| 2453 | { |
| 2454 | return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); |
| 2455 | } |
| 2456 | EXPORT_SYMBOL(vmf_insert_pfn); |
| 2457 | |
| 2458 | static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn, bool mkwrite) |
| 2459 | { |
| 2460 | if (unlikely(is_zero_pfn(pfn_t_to_pfn(pfn))) && |
| 2461 | (mkwrite || !vm_mixed_zeropage_allowed(vma))) |
| 2462 | return false; |
| 2463 | /* these checks mirror the abort conditions in vm_normal_page */ |
| 2464 | if (vma->vm_flags & VM_MIXEDMAP) |
| 2465 | return true; |
| 2466 | if (pfn_t_devmap(pfn)) |
| 2467 | return true; |
| 2468 | if (pfn_t_special(pfn)) |
| 2469 | return true; |
| 2470 | if (is_zero_pfn(pfn_t_to_pfn(pfn))) |
| 2471 | return true; |
| 2472 | return false; |
| 2473 | } |
| 2474 | |
| 2475 | static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, |
| 2476 | unsigned long addr, pfn_t pfn, bool mkwrite) |
| 2477 | { |
| 2478 | pgprot_t pgprot = vma->vm_page_prot; |
| 2479 | int err; |
| 2480 | |
| 2481 | if (!vm_mixed_ok(vma, pfn, mkwrite)) |
| 2482 | return VM_FAULT_SIGBUS; |
| 2483 | |
| 2484 | if (addr < vma->vm_start || addr >= vma->vm_end) |
| 2485 | return VM_FAULT_SIGBUS; |
| 2486 | |
| 2487 | track_pfn_insert(vma, &pgprot, pfn); |
| 2488 | |
| 2489 | if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) |
| 2490 | return VM_FAULT_SIGBUS; |
| 2491 | |
| 2492 | /* |
| 2493 | * If we don't have pte special, then we have to use the pfn_valid() |
| 2494 | * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* |
| 2495 | * refcount the page if pfn_valid is true (hence insert_page rather |
| 2496 | * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP |
| 2497 | * without pte special, it would there be refcounted as a normal page. |
| 2498 | */ |
| 2499 | if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && |
| 2500 | !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { |
| 2501 | struct page *page; |
| 2502 | |
| 2503 | /* |
| 2504 | * At this point we are committed to insert_page() |
| 2505 | * regardless of whether the caller specified flags that |
| 2506 | * result in pfn_t_has_page() == false. |
| 2507 | */ |
| 2508 | page = pfn_to_page(pfn_t_to_pfn(pfn)); |
| 2509 | err = insert_page(vma, addr, page, pgprot); |
| 2510 | } else { |
| 2511 | return insert_pfn(vma, addr, pfn, pgprot, mkwrite); |
| 2512 | } |
| 2513 | |
| 2514 | if (err == -ENOMEM) |
| 2515 | return VM_FAULT_OOM; |
| 2516 | if (err < 0 && err != -EBUSY) |
| 2517 | return VM_FAULT_SIGBUS; |
| 2518 | |
| 2519 | return VM_FAULT_NOPAGE; |
| 2520 | } |
| 2521 | |
| 2522 | vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, |
| 2523 | pfn_t pfn) |
| 2524 | { |
| 2525 | return __vm_insert_mixed(vma, addr, pfn, false); |
| 2526 | } |
| 2527 | EXPORT_SYMBOL(vmf_insert_mixed); |
| 2528 | |
| 2529 | /* |
| 2530 | * If the insertion of PTE failed because someone else already added a |
| 2531 | * different entry in the mean time, we treat that as success as we assume |
| 2532 | * the same entry was actually inserted. |
| 2533 | */ |
| 2534 | vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, |
| 2535 | unsigned long addr, pfn_t pfn) |
| 2536 | { |
| 2537 | return __vm_insert_mixed(vma, addr, pfn, true); |
| 2538 | } |
| 2539 | |
| 2540 | /* |
| 2541 | * maps a range of physical memory into the requested pages. the old |
| 2542 | * mappings are removed. any references to nonexistent pages results |
| 2543 | * in null mappings (currently treated as "copy-on-access") |
| 2544 | */ |
| 2545 | static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, |
| 2546 | unsigned long addr, unsigned long end, |
| 2547 | unsigned long pfn, pgprot_t prot) |
| 2548 | { |
| 2549 | pte_t *pte, *mapped_pte; |
| 2550 | spinlock_t *ptl; |
| 2551 | int err = 0; |
| 2552 | |
| 2553 | mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); |
| 2554 | if (!pte) |
| 2555 | return -ENOMEM; |
| 2556 | arch_enter_lazy_mmu_mode(); |
| 2557 | do { |
| 2558 | BUG_ON(!pte_none(ptep_get(pte))); |
| 2559 | if (!pfn_modify_allowed(pfn, prot)) { |
| 2560 | err = -EACCES; |
| 2561 | break; |
| 2562 | } |
| 2563 | set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); |
| 2564 | pfn++; |
| 2565 | } while (pte++, addr += PAGE_SIZE, addr != end); |
| 2566 | arch_leave_lazy_mmu_mode(); |
| 2567 | pte_unmap_unlock(mapped_pte, ptl); |
| 2568 | return err; |
| 2569 | } |
| 2570 | |
| 2571 | static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, |
| 2572 | unsigned long addr, unsigned long end, |
| 2573 | unsigned long pfn, pgprot_t prot) |
| 2574 | { |
| 2575 | pmd_t *pmd; |
| 2576 | unsigned long next; |
| 2577 | int err; |
| 2578 | |
| 2579 | pfn -= addr >> PAGE_SHIFT; |
| 2580 | pmd = pmd_alloc(mm, pud, addr); |
| 2581 | if (!pmd) |
| 2582 | return -ENOMEM; |
| 2583 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
| 2584 | do { |
| 2585 | next = pmd_addr_end(addr, end); |
| 2586 | err = remap_pte_range(mm, pmd, addr, next, |
| 2587 | pfn + (addr >> PAGE_SHIFT), prot); |
| 2588 | if (err) |
| 2589 | return err; |
| 2590 | } while (pmd++, addr = next, addr != end); |
| 2591 | return 0; |
| 2592 | } |
| 2593 | |
| 2594 | static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, |
| 2595 | unsigned long addr, unsigned long end, |
| 2596 | unsigned long pfn, pgprot_t prot) |
| 2597 | { |
| 2598 | pud_t *pud; |
| 2599 | unsigned long next; |
| 2600 | int err; |
| 2601 | |
| 2602 | pfn -= addr >> PAGE_SHIFT; |
| 2603 | pud = pud_alloc(mm, p4d, addr); |
| 2604 | if (!pud) |
| 2605 | return -ENOMEM; |
| 2606 | do { |
| 2607 | next = pud_addr_end(addr, end); |
| 2608 | err = remap_pmd_range(mm, pud, addr, next, |
| 2609 | pfn + (addr >> PAGE_SHIFT), prot); |
| 2610 | if (err) |
| 2611 | return err; |
| 2612 | } while (pud++, addr = next, addr != end); |
| 2613 | return 0; |
| 2614 | } |
| 2615 | |
| 2616 | static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, |
| 2617 | unsigned long addr, unsigned long end, |
| 2618 | unsigned long pfn, pgprot_t prot) |
| 2619 | { |
| 2620 | p4d_t *p4d; |
| 2621 | unsigned long next; |
| 2622 | int err; |
| 2623 | |
| 2624 | pfn -= addr >> PAGE_SHIFT; |
| 2625 | p4d = p4d_alloc(mm, pgd, addr); |
| 2626 | if (!p4d) |
| 2627 | return -ENOMEM; |
| 2628 | do { |
| 2629 | next = p4d_addr_end(addr, end); |
| 2630 | err = remap_pud_range(mm, p4d, addr, next, |
| 2631 | pfn + (addr >> PAGE_SHIFT), prot); |
| 2632 | if (err) |
| 2633 | return err; |
| 2634 | } while (p4d++, addr = next, addr != end); |
| 2635 | return 0; |
| 2636 | } |
| 2637 | |
| 2638 | static int remap_pfn_range_internal(struct vm_area_struct *vma, unsigned long addr, |
| 2639 | unsigned long pfn, unsigned long size, pgprot_t prot) |
| 2640 | { |
| 2641 | pgd_t *pgd; |
| 2642 | unsigned long next; |
| 2643 | unsigned long end = addr + PAGE_ALIGN(size); |
| 2644 | struct mm_struct *mm = vma->vm_mm; |
| 2645 | int err; |
| 2646 | |
| 2647 | if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) |
| 2648 | return -EINVAL; |
| 2649 | |
| 2650 | /* |
| 2651 | * Physically remapped pages are special. Tell the |
| 2652 | * rest of the world about it: |
| 2653 | * VM_IO tells people not to look at these pages |
| 2654 | * (accesses can have side effects). |
| 2655 | * VM_PFNMAP tells the core MM that the base pages are just |
| 2656 | * raw PFN mappings, and do not have a "struct page" associated |
| 2657 | * with them. |
| 2658 | * VM_DONTEXPAND |
| 2659 | * Disable vma merging and expanding with mremap(). |
| 2660 | * VM_DONTDUMP |
| 2661 | * Omit vma from core dump, even when VM_IO turned off. |
| 2662 | * |
| 2663 | * There's a horrible special case to handle copy-on-write |
| 2664 | * behaviour that some programs depend on. We mark the "original" |
| 2665 | * un-COW'ed pages by matching them up with "vma->vm_pgoff". |
| 2666 | * See vm_normal_page() for details. |
| 2667 | */ |
| 2668 | if (is_cow_mapping(vma->vm_flags)) { |
| 2669 | if (addr != vma->vm_start || end != vma->vm_end) |
| 2670 | return -EINVAL; |
| 2671 | vma->vm_pgoff = pfn; |
| 2672 | } |
| 2673 | |
| 2674 | vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP); |
| 2675 | |
| 2676 | BUG_ON(addr >= end); |
| 2677 | pfn -= addr >> PAGE_SHIFT; |
| 2678 | pgd = pgd_offset(mm, addr); |
| 2679 | flush_cache_range(vma, addr, end); |
| 2680 | do { |
| 2681 | next = pgd_addr_end(addr, end); |
| 2682 | err = remap_p4d_range(mm, pgd, addr, next, |
| 2683 | pfn + (addr >> PAGE_SHIFT), prot); |
| 2684 | if (err) |
| 2685 | return err; |
| 2686 | } while (pgd++, addr = next, addr != end); |
| 2687 | |
| 2688 | return 0; |
| 2689 | } |
| 2690 | |
| 2691 | /* |
| 2692 | * Variant of remap_pfn_range that does not call track_pfn_remap. The caller |
| 2693 | * must have pre-validated the caching bits of the pgprot_t. |
| 2694 | */ |
| 2695 | int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr, |
| 2696 | unsigned long pfn, unsigned long size, pgprot_t prot) |
| 2697 | { |
| 2698 | int error = remap_pfn_range_internal(vma, addr, pfn, size, prot); |
| 2699 | |
| 2700 | if (!error) |
| 2701 | return 0; |
| 2702 | |
| 2703 | /* |
| 2704 | * A partial pfn range mapping is dangerous: it does not |
| 2705 | * maintain page reference counts, and callers may free |
| 2706 | * pages due to the error. So zap it early. |
| 2707 | */ |
| 2708 | zap_page_range_single(vma, addr, size, NULL); |
| 2709 | return error; |
| 2710 | } |
| 2711 | |
| 2712 | /** |
| 2713 | * remap_pfn_range - remap kernel memory to userspace |
| 2714 | * @vma: user vma to map to |
| 2715 | * @addr: target page aligned user address to start at |
| 2716 | * @pfn: page frame number of kernel physical memory address |
| 2717 | * @size: size of mapping area |
| 2718 | * @prot: page protection flags for this mapping |
| 2719 | * |
| 2720 | * Note: this is only safe if the mm semaphore is held when called. |
| 2721 | * |
| 2722 | * Return: %0 on success, negative error code otherwise. |
| 2723 | */ |
| 2724 | int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, |
| 2725 | unsigned long pfn, unsigned long size, pgprot_t prot) |
| 2726 | { |
| 2727 | int err; |
| 2728 | |
| 2729 | err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); |
| 2730 | if (err) |
| 2731 | return -EINVAL; |
| 2732 | |
| 2733 | err = remap_pfn_range_notrack(vma, addr, pfn, size, prot); |
| 2734 | if (err) |
| 2735 | untrack_pfn(vma, pfn, PAGE_ALIGN(size), true); |
| 2736 | return err; |
| 2737 | } |
| 2738 | EXPORT_SYMBOL(remap_pfn_range); |
| 2739 | |
| 2740 | /** |
| 2741 | * vm_iomap_memory - remap memory to userspace |
| 2742 | * @vma: user vma to map to |
| 2743 | * @start: start of the physical memory to be mapped |
| 2744 | * @len: size of area |
| 2745 | * |
| 2746 | * This is a simplified io_remap_pfn_range() for common driver use. The |
| 2747 | * driver just needs to give us the physical memory range to be mapped, |
| 2748 | * we'll figure out the rest from the vma information. |
| 2749 | * |
| 2750 | * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get |
| 2751 | * whatever write-combining details or similar. |
| 2752 | * |
| 2753 | * Return: %0 on success, negative error code otherwise. |
| 2754 | */ |
| 2755 | int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) |
| 2756 | { |
| 2757 | unsigned long vm_len, pfn, pages; |
| 2758 | |
| 2759 | /* Check that the physical memory area passed in looks valid */ |
| 2760 | if (start + len < start) |
| 2761 | return -EINVAL; |
| 2762 | /* |
| 2763 | * You *really* shouldn't map things that aren't page-aligned, |
| 2764 | * but we've historically allowed it because IO memory might |
| 2765 | * just have smaller alignment. |
| 2766 | */ |
| 2767 | len += start & ~PAGE_MASK; |
| 2768 | pfn = start >> PAGE_SHIFT; |
| 2769 | pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; |
| 2770 | if (pfn + pages < pfn) |
| 2771 | return -EINVAL; |
| 2772 | |
| 2773 | /* We start the mapping 'vm_pgoff' pages into the area */ |
| 2774 | if (vma->vm_pgoff > pages) |
| 2775 | return -EINVAL; |
| 2776 | pfn += vma->vm_pgoff; |
| 2777 | pages -= vma->vm_pgoff; |
| 2778 | |
| 2779 | /* Can we fit all of the mapping? */ |
| 2780 | vm_len = vma->vm_end - vma->vm_start; |
| 2781 | if (vm_len >> PAGE_SHIFT > pages) |
| 2782 | return -EINVAL; |
| 2783 | |
| 2784 | /* Ok, let it rip */ |
| 2785 | return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); |
| 2786 | } |
| 2787 | EXPORT_SYMBOL(vm_iomap_memory); |
| 2788 | |
| 2789 | static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, |
| 2790 | unsigned long addr, unsigned long end, |
| 2791 | pte_fn_t fn, void *data, bool create, |
| 2792 | pgtbl_mod_mask *mask) |
| 2793 | { |
| 2794 | pte_t *pte, *mapped_pte; |
| 2795 | int err = 0; |
| 2796 | spinlock_t *ptl; |
| 2797 | |
| 2798 | if (create) { |
| 2799 | mapped_pte = pte = (mm == &init_mm) ? |
| 2800 | pte_alloc_kernel_track(pmd, addr, mask) : |
| 2801 | pte_alloc_map_lock(mm, pmd, addr, &ptl); |
| 2802 | if (!pte) |
| 2803 | return -ENOMEM; |
| 2804 | } else { |
| 2805 | mapped_pte = pte = (mm == &init_mm) ? |
| 2806 | pte_offset_kernel(pmd, addr) : |
| 2807 | pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 2808 | if (!pte) |
| 2809 | return -EINVAL; |
| 2810 | } |
| 2811 | |
| 2812 | arch_enter_lazy_mmu_mode(); |
| 2813 | |
| 2814 | if (fn) { |
| 2815 | do { |
| 2816 | if (create || !pte_none(ptep_get(pte))) { |
| 2817 | err = fn(pte++, addr, data); |
| 2818 | if (err) |
| 2819 | break; |
| 2820 | } |
| 2821 | } while (addr += PAGE_SIZE, addr != end); |
| 2822 | } |
| 2823 | *mask |= PGTBL_PTE_MODIFIED; |
| 2824 | |
| 2825 | arch_leave_lazy_mmu_mode(); |
| 2826 | |
| 2827 | if (mm != &init_mm) |
| 2828 | pte_unmap_unlock(mapped_pte, ptl); |
| 2829 | return err; |
| 2830 | } |
| 2831 | |
| 2832 | static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, |
| 2833 | unsigned long addr, unsigned long end, |
| 2834 | pte_fn_t fn, void *data, bool create, |
| 2835 | pgtbl_mod_mask *mask) |
| 2836 | { |
| 2837 | pmd_t *pmd; |
| 2838 | unsigned long next; |
| 2839 | int err = 0; |
| 2840 | |
| 2841 | BUG_ON(pud_leaf(*pud)); |
| 2842 | |
| 2843 | if (create) { |
| 2844 | pmd = pmd_alloc_track(mm, pud, addr, mask); |
| 2845 | if (!pmd) |
| 2846 | return -ENOMEM; |
| 2847 | } else { |
| 2848 | pmd = pmd_offset(pud, addr); |
| 2849 | } |
| 2850 | do { |
| 2851 | next = pmd_addr_end(addr, end); |
| 2852 | if (pmd_none(*pmd) && !create) |
| 2853 | continue; |
| 2854 | if (WARN_ON_ONCE(pmd_leaf(*pmd))) |
| 2855 | return -EINVAL; |
| 2856 | if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) { |
| 2857 | if (!create) |
| 2858 | continue; |
| 2859 | pmd_clear_bad(pmd); |
| 2860 | } |
| 2861 | err = apply_to_pte_range(mm, pmd, addr, next, |
| 2862 | fn, data, create, mask); |
| 2863 | if (err) |
| 2864 | break; |
| 2865 | } while (pmd++, addr = next, addr != end); |
| 2866 | |
| 2867 | return err; |
| 2868 | } |
| 2869 | |
| 2870 | static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, |
| 2871 | unsigned long addr, unsigned long end, |
| 2872 | pte_fn_t fn, void *data, bool create, |
| 2873 | pgtbl_mod_mask *mask) |
| 2874 | { |
| 2875 | pud_t *pud; |
| 2876 | unsigned long next; |
| 2877 | int err = 0; |
| 2878 | |
| 2879 | if (create) { |
| 2880 | pud = pud_alloc_track(mm, p4d, addr, mask); |
| 2881 | if (!pud) |
| 2882 | return -ENOMEM; |
| 2883 | } else { |
| 2884 | pud = pud_offset(p4d, addr); |
| 2885 | } |
| 2886 | do { |
| 2887 | next = pud_addr_end(addr, end); |
| 2888 | if (pud_none(*pud) && !create) |
| 2889 | continue; |
| 2890 | if (WARN_ON_ONCE(pud_leaf(*pud))) |
| 2891 | return -EINVAL; |
| 2892 | if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) { |
| 2893 | if (!create) |
| 2894 | continue; |
| 2895 | pud_clear_bad(pud); |
| 2896 | } |
| 2897 | err = apply_to_pmd_range(mm, pud, addr, next, |
| 2898 | fn, data, create, mask); |
| 2899 | if (err) |
| 2900 | break; |
| 2901 | } while (pud++, addr = next, addr != end); |
| 2902 | |
| 2903 | return err; |
| 2904 | } |
| 2905 | |
| 2906 | static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, |
| 2907 | unsigned long addr, unsigned long end, |
| 2908 | pte_fn_t fn, void *data, bool create, |
| 2909 | pgtbl_mod_mask *mask) |
| 2910 | { |
| 2911 | p4d_t *p4d; |
| 2912 | unsigned long next; |
| 2913 | int err = 0; |
| 2914 | |
| 2915 | if (create) { |
| 2916 | p4d = p4d_alloc_track(mm, pgd, addr, mask); |
| 2917 | if (!p4d) |
| 2918 | return -ENOMEM; |
| 2919 | } else { |
| 2920 | p4d = p4d_offset(pgd, addr); |
| 2921 | } |
| 2922 | do { |
| 2923 | next = p4d_addr_end(addr, end); |
| 2924 | if (p4d_none(*p4d) && !create) |
| 2925 | continue; |
| 2926 | if (WARN_ON_ONCE(p4d_leaf(*p4d))) |
| 2927 | return -EINVAL; |
| 2928 | if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) { |
| 2929 | if (!create) |
| 2930 | continue; |
| 2931 | p4d_clear_bad(p4d); |
| 2932 | } |
| 2933 | err = apply_to_pud_range(mm, p4d, addr, next, |
| 2934 | fn, data, create, mask); |
| 2935 | if (err) |
| 2936 | break; |
| 2937 | } while (p4d++, addr = next, addr != end); |
| 2938 | |
| 2939 | return err; |
| 2940 | } |
| 2941 | |
| 2942 | static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, |
| 2943 | unsigned long size, pte_fn_t fn, |
| 2944 | void *data, bool create) |
| 2945 | { |
| 2946 | pgd_t *pgd; |
| 2947 | unsigned long start = addr, next; |
| 2948 | unsigned long end = addr + size; |
| 2949 | pgtbl_mod_mask mask = 0; |
| 2950 | int err = 0; |
| 2951 | |
| 2952 | if (WARN_ON(addr >= end)) |
| 2953 | return -EINVAL; |
| 2954 | |
| 2955 | pgd = pgd_offset(mm, addr); |
| 2956 | do { |
| 2957 | next = pgd_addr_end(addr, end); |
| 2958 | if (pgd_none(*pgd) && !create) |
| 2959 | continue; |
| 2960 | if (WARN_ON_ONCE(pgd_leaf(*pgd))) |
| 2961 | return -EINVAL; |
| 2962 | if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) { |
| 2963 | if (!create) |
| 2964 | continue; |
| 2965 | pgd_clear_bad(pgd); |
| 2966 | } |
| 2967 | err = apply_to_p4d_range(mm, pgd, addr, next, |
| 2968 | fn, data, create, &mask); |
| 2969 | if (err) |
| 2970 | break; |
| 2971 | } while (pgd++, addr = next, addr != end); |
| 2972 | |
| 2973 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) |
| 2974 | arch_sync_kernel_mappings(start, start + size); |
| 2975 | |
| 2976 | return err; |
| 2977 | } |
| 2978 | |
| 2979 | /* |
| 2980 | * Scan a region of virtual memory, filling in page tables as necessary |
| 2981 | * and calling a provided function on each leaf page table. |
| 2982 | */ |
| 2983 | int apply_to_page_range(struct mm_struct *mm, unsigned long addr, |
| 2984 | unsigned long size, pte_fn_t fn, void *data) |
| 2985 | { |
| 2986 | return __apply_to_page_range(mm, addr, size, fn, data, true); |
| 2987 | } |
| 2988 | EXPORT_SYMBOL_GPL(apply_to_page_range); |
| 2989 | |
| 2990 | /* |
| 2991 | * Scan a region of virtual memory, calling a provided function on |
| 2992 | * each leaf page table where it exists. |
| 2993 | * |
| 2994 | * Unlike apply_to_page_range, this does _not_ fill in page tables |
| 2995 | * where they are absent. |
| 2996 | */ |
| 2997 | int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, |
| 2998 | unsigned long size, pte_fn_t fn, void *data) |
| 2999 | { |
| 3000 | return __apply_to_page_range(mm, addr, size, fn, data, false); |
| 3001 | } |
| 3002 | EXPORT_SYMBOL_GPL(apply_to_existing_page_range); |
| 3003 | |
| 3004 | /* |
| 3005 | * handle_pte_fault chooses page fault handler according to an entry which was |
| 3006 | * read non-atomically. Before making any commitment, on those architectures |
| 3007 | * or configurations (e.g. i386 with PAE) which might give a mix of unmatched |
| 3008 | * parts, do_swap_page must check under lock before unmapping the pte and |
| 3009 | * proceeding (but do_wp_page is only called after already making such a check; |
| 3010 | * and do_anonymous_page can safely check later on). |
| 3011 | */ |
| 3012 | static inline int pte_unmap_same(struct vm_fault *vmf) |
| 3013 | { |
| 3014 | int same = 1; |
| 3015 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) |
| 3016 | if (sizeof(pte_t) > sizeof(unsigned long)) { |
| 3017 | spin_lock(vmf->ptl); |
| 3018 | same = pte_same(ptep_get(vmf->pte), vmf->orig_pte); |
| 3019 | spin_unlock(vmf->ptl); |
| 3020 | } |
| 3021 | #endif |
| 3022 | pte_unmap(vmf->pte); |
| 3023 | vmf->pte = NULL; |
| 3024 | return same; |
| 3025 | } |
| 3026 | |
| 3027 | /* |
| 3028 | * Return: |
| 3029 | * 0: copied succeeded |
| 3030 | * -EHWPOISON: copy failed due to hwpoison in source page |
| 3031 | * -EAGAIN: copied failed (some other reason) |
| 3032 | */ |
| 3033 | static inline int __wp_page_copy_user(struct page *dst, struct page *src, |
| 3034 | struct vm_fault *vmf) |
| 3035 | { |
| 3036 | int ret; |
| 3037 | void *kaddr; |
| 3038 | void __user *uaddr; |
| 3039 | struct vm_area_struct *vma = vmf->vma; |
| 3040 | struct mm_struct *mm = vma->vm_mm; |
| 3041 | unsigned long addr = vmf->address; |
| 3042 | |
| 3043 | if (likely(src)) { |
| 3044 | if (copy_mc_user_highpage(dst, src, addr, vma)) |
| 3045 | return -EHWPOISON; |
| 3046 | return 0; |
| 3047 | } |
| 3048 | |
| 3049 | /* |
| 3050 | * If the source page was a PFN mapping, we don't have |
| 3051 | * a "struct page" for it. We do a best-effort copy by |
| 3052 | * just copying from the original user address. If that |
| 3053 | * fails, we just zero-fill it. Live with it. |
| 3054 | */ |
| 3055 | kaddr = kmap_local_page(dst); |
| 3056 | pagefault_disable(); |
| 3057 | uaddr = (void __user *)(addr & PAGE_MASK); |
| 3058 | |
| 3059 | /* |
| 3060 | * On architectures with software "accessed" bits, we would |
| 3061 | * take a double page fault, so mark it accessed here. |
| 3062 | */ |
| 3063 | vmf->pte = NULL; |
| 3064 | if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) { |
| 3065 | pte_t entry; |
| 3066 | |
| 3067 | vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); |
| 3068 | if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { |
| 3069 | /* |
| 3070 | * Other thread has already handled the fault |
| 3071 | * and update local tlb only |
| 3072 | */ |
| 3073 | if (vmf->pte) |
| 3074 | update_mmu_tlb(vma, addr, vmf->pte); |
| 3075 | ret = -EAGAIN; |
| 3076 | goto pte_unlock; |
| 3077 | } |
| 3078 | |
| 3079 | entry = pte_mkyoung(vmf->orig_pte); |
| 3080 | if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) |
| 3081 | update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1); |
| 3082 | } |
| 3083 | |
| 3084 | /* |
| 3085 | * This really shouldn't fail, because the page is there |
| 3086 | * in the page tables. But it might just be unreadable, |
| 3087 | * in which case we just give up and fill the result with |
| 3088 | * zeroes. |
| 3089 | */ |
| 3090 | if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { |
| 3091 | if (vmf->pte) |
| 3092 | goto warn; |
| 3093 | |
| 3094 | /* Re-validate under PTL if the page is still mapped */ |
| 3095 | vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); |
| 3096 | if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { |
| 3097 | /* The PTE changed under us, update local tlb */ |
| 3098 | if (vmf->pte) |
| 3099 | update_mmu_tlb(vma, addr, vmf->pte); |
| 3100 | ret = -EAGAIN; |
| 3101 | goto pte_unlock; |
| 3102 | } |
| 3103 | |
| 3104 | /* |
| 3105 | * The same page can be mapped back since last copy attempt. |
| 3106 | * Try to copy again under PTL. |
| 3107 | */ |
| 3108 | if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { |
| 3109 | /* |
| 3110 | * Give a warn in case there can be some obscure |
| 3111 | * use-case |
| 3112 | */ |
| 3113 | warn: |
| 3114 | WARN_ON_ONCE(1); |
| 3115 | clear_page(kaddr); |
| 3116 | } |
| 3117 | } |
| 3118 | |
| 3119 | ret = 0; |
| 3120 | |
| 3121 | pte_unlock: |
| 3122 | if (vmf->pte) |
| 3123 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3124 | pagefault_enable(); |
| 3125 | kunmap_local(kaddr); |
| 3126 | flush_dcache_page(dst); |
| 3127 | |
| 3128 | return ret; |
| 3129 | } |
| 3130 | |
| 3131 | static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) |
| 3132 | { |
| 3133 | struct file *vm_file = vma->vm_file; |
| 3134 | |
| 3135 | if (vm_file) |
| 3136 | return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; |
| 3137 | |
| 3138 | /* |
| 3139 | * Special mappings (e.g. VDSO) do not have any file so fake |
| 3140 | * a default GFP_KERNEL for them. |
| 3141 | */ |
| 3142 | return GFP_KERNEL; |
| 3143 | } |
| 3144 | |
| 3145 | /* |
| 3146 | * Notify the address space that the page is about to become writable so that |
| 3147 | * it can prohibit this or wait for the page to get into an appropriate state. |
| 3148 | * |
| 3149 | * We do this without the lock held, so that it can sleep if it needs to. |
| 3150 | */ |
| 3151 | static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio) |
| 3152 | { |
| 3153 | vm_fault_t ret; |
| 3154 | unsigned int old_flags = vmf->flags; |
| 3155 | |
| 3156 | vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; |
| 3157 | |
| 3158 | if (vmf->vma->vm_file && |
| 3159 | IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) |
| 3160 | return VM_FAULT_SIGBUS; |
| 3161 | |
| 3162 | ret = vmf->vma->vm_ops->page_mkwrite(vmf); |
| 3163 | /* Restore original flags so that caller is not surprised */ |
| 3164 | vmf->flags = old_flags; |
| 3165 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) |
| 3166 | return ret; |
| 3167 | if (unlikely(!(ret & VM_FAULT_LOCKED))) { |
| 3168 | folio_lock(folio); |
| 3169 | if (!folio->mapping) { |
| 3170 | folio_unlock(folio); |
| 3171 | return 0; /* retry */ |
| 3172 | } |
| 3173 | ret |= VM_FAULT_LOCKED; |
| 3174 | } else |
| 3175 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
| 3176 | return ret; |
| 3177 | } |
| 3178 | |
| 3179 | /* |
| 3180 | * Handle dirtying of a page in shared file mapping on a write fault. |
| 3181 | * |
| 3182 | * The function expects the page to be locked and unlocks it. |
| 3183 | */ |
| 3184 | static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) |
| 3185 | { |
| 3186 | struct vm_area_struct *vma = vmf->vma; |
| 3187 | struct address_space *mapping; |
| 3188 | struct folio *folio = page_folio(vmf->page); |
| 3189 | bool dirtied; |
| 3190 | bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; |
| 3191 | |
| 3192 | dirtied = folio_mark_dirty(folio); |
| 3193 | VM_BUG_ON_FOLIO(folio_test_anon(folio), folio); |
| 3194 | /* |
| 3195 | * Take a local copy of the address_space - folio.mapping may be zeroed |
| 3196 | * by truncate after folio_unlock(). The address_space itself remains |
| 3197 | * pinned by vma->vm_file's reference. We rely on folio_unlock()'s |
| 3198 | * release semantics to prevent the compiler from undoing this copying. |
| 3199 | */ |
| 3200 | mapping = folio_raw_mapping(folio); |
| 3201 | folio_unlock(folio); |
| 3202 | |
| 3203 | if (!page_mkwrite) |
| 3204 | file_update_time(vma->vm_file); |
| 3205 | |
| 3206 | /* |
| 3207 | * Throttle page dirtying rate down to writeback speed. |
| 3208 | * |
| 3209 | * mapping may be NULL here because some device drivers do not |
| 3210 | * set page.mapping but still dirty their pages |
| 3211 | * |
| 3212 | * Drop the mmap_lock before waiting on IO, if we can. The file |
| 3213 | * is pinning the mapping, as per above. |
| 3214 | */ |
| 3215 | if ((dirtied || page_mkwrite) && mapping) { |
| 3216 | struct file *fpin; |
| 3217 | |
| 3218 | fpin = maybe_unlock_mmap_for_io(vmf, NULL); |
| 3219 | balance_dirty_pages_ratelimited(mapping); |
| 3220 | if (fpin) { |
| 3221 | fput(fpin); |
| 3222 | return VM_FAULT_COMPLETED; |
| 3223 | } |
| 3224 | } |
| 3225 | |
| 3226 | return 0; |
| 3227 | } |
| 3228 | |
| 3229 | /* |
| 3230 | * Handle write page faults for pages that can be reused in the current vma |
| 3231 | * |
| 3232 | * This can happen either due to the mapping being with the VM_SHARED flag, |
| 3233 | * or due to us being the last reference standing to the page. In either |
| 3234 | * case, all we need to do here is to mark the page as writable and update |
| 3235 | * any related book-keeping. |
| 3236 | */ |
| 3237 | static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio) |
| 3238 | __releases(vmf->ptl) |
| 3239 | { |
| 3240 | struct vm_area_struct *vma = vmf->vma; |
| 3241 | pte_t entry; |
| 3242 | |
| 3243 | VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE)); |
| 3244 | VM_WARN_ON(is_zero_pfn(pte_pfn(vmf->orig_pte))); |
| 3245 | |
| 3246 | if (folio) { |
| 3247 | VM_BUG_ON(folio_test_anon(folio) && |
| 3248 | !PageAnonExclusive(vmf->page)); |
| 3249 | /* |
| 3250 | * Clear the folio's cpupid information as the existing |
| 3251 | * information potentially belongs to a now completely |
| 3252 | * unrelated process. |
| 3253 | */ |
| 3254 | folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1); |
| 3255 | } |
| 3256 | |
| 3257 | flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); |
| 3258 | entry = pte_mkyoung(vmf->orig_pte); |
| 3259 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 3260 | if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) |
| 3261 | update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); |
| 3262 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3263 | count_vm_event(PGREUSE); |
| 3264 | } |
| 3265 | |
| 3266 | /* |
| 3267 | * We could add a bitflag somewhere, but for now, we know that all |
| 3268 | * vm_ops that have a ->map_pages have been audited and don't need |
| 3269 | * the mmap_lock to be held. |
| 3270 | */ |
| 3271 | static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf) |
| 3272 | { |
| 3273 | struct vm_area_struct *vma = vmf->vma; |
| 3274 | |
| 3275 | if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK)) |
| 3276 | return 0; |
| 3277 | vma_end_read(vma); |
| 3278 | return VM_FAULT_RETRY; |
| 3279 | } |
| 3280 | |
| 3281 | /** |
| 3282 | * __vmf_anon_prepare - Prepare to handle an anonymous fault. |
| 3283 | * @vmf: The vm_fault descriptor passed from the fault handler. |
| 3284 | * |
| 3285 | * When preparing to insert an anonymous page into a VMA from a |
| 3286 | * fault handler, call this function rather than anon_vma_prepare(). |
| 3287 | * If this vma does not already have an associated anon_vma and we are |
| 3288 | * only protected by the per-VMA lock, the caller must retry with the |
| 3289 | * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to |
| 3290 | * determine if this VMA can share its anon_vma, and that's not safe to |
| 3291 | * do with only the per-VMA lock held for this VMA. |
| 3292 | * |
| 3293 | * Return: 0 if fault handling can proceed. Any other value should be |
| 3294 | * returned to the caller. |
| 3295 | */ |
| 3296 | vm_fault_t __vmf_anon_prepare(struct vm_fault *vmf) |
| 3297 | { |
| 3298 | struct vm_area_struct *vma = vmf->vma; |
| 3299 | vm_fault_t ret = 0; |
| 3300 | |
| 3301 | if (likely(vma->anon_vma)) |
| 3302 | return 0; |
| 3303 | if (vmf->flags & FAULT_FLAG_VMA_LOCK) { |
| 3304 | if (!mmap_read_trylock(vma->vm_mm)) |
| 3305 | return VM_FAULT_RETRY; |
| 3306 | } |
| 3307 | if (__anon_vma_prepare(vma)) |
| 3308 | ret = VM_FAULT_OOM; |
| 3309 | if (vmf->flags & FAULT_FLAG_VMA_LOCK) |
| 3310 | mmap_read_unlock(vma->vm_mm); |
| 3311 | return ret; |
| 3312 | } |
| 3313 | |
| 3314 | /* |
| 3315 | * Handle the case of a page which we actually need to copy to a new page, |
| 3316 | * either due to COW or unsharing. |
| 3317 | * |
| 3318 | * Called with mmap_lock locked and the old page referenced, but |
| 3319 | * without the ptl held. |
| 3320 | * |
| 3321 | * High level logic flow: |
| 3322 | * |
| 3323 | * - Allocate a page, copy the content of the old page to the new one. |
| 3324 | * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. |
| 3325 | * - Take the PTL. If the pte changed, bail out and release the allocated page |
| 3326 | * - If the pte is still the way we remember it, update the page table and all |
| 3327 | * relevant references. This includes dropping the reference the page-table |
| 3328 | * held to the old page, as well as updating the rmap. |
| 3329 | * - In any case, unlock the PTL and drop the reference we took to the old page. |
| 3330 | */ |
| 3331 | static vm_fault_t wp_page_copy(struct vm_fault *vmf) |
| 3332 | { |
| 3333 | const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; |
| 3334 | struct vm_area_struct *vma = vmf->vma; |
| 3335 | struct mm_struct *mm = vma->vm_mm; |
| 3336 | struct folio *old_folio = NULL; |
| 3337 | struct folio *new_folio = NULL; |
| 3338 | pte_t entry; |
| 3339 | int page_copied = 0; |
| 3340 | struct mmu_notifier_range range; |
| 3341 | vm_fault_t ret; |
| 3342 | bool pfn_is_zero; |
| 3343 | |
| 3344 | delayacct_wpcopy_start(); |
| 3345 | |
| 3346 | if (vmf->page) |
| 3347 | old_folio = page_folio(vmf->page); |
| 3348 | ret = vmf_anon_prepare(vmf); |
| 3349 | if (unlikely(ret)) |
| 3350 | goto out; |
| 3351 | |
| 3352 | pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte)); |
| 3353 | new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero); |
| 3354 | if (!new_folio) |
| 3355 | goto oom; |
| 3356 | |
| 3357 | if (!pfn_is_zero) { |
| 3358 | int err; |
| 3359 | |
| 3360 | err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf); |
| 3361 | if (err) { |
| 3362 | /* |
| 3363 | * COW failed, if the fault was solved by other, |
| 3364 | * it's fine. If not, userspace would re-fault on |
| 3365 | * the same address and we will handle the fault |
| 3366 | * from the second attempt. |
| 3367 | * The -EHWPOISON case will not be retried. |
| 3368 | */ |
| 3369 | folio_put(new_folio); |
| 3370 | if (old_folio) |
| 3371 | folio_put(old_folio); |
| 3372 | |
| 3373 | delayacct_wpcopy_end(); |
| 3374 | return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0; |
| 3375 | } |
| 3376 | kmsan_copy_page_meta(&new_folio->page, vmf->page); |
| 3377 | } |
| 3378 | |
| 3379 | __folio_mark_uptodate(new_folio); |
| 3380 | |
| 3381 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, |
| 3382 | vmf->address & PAGE_MASK, |
| 3383 | (vmf->address & PAGE_MASK) + PAGE_SIZE); |
| 3384 | mmu_notifier_invalidate_range_start(&range); |
| 3385 | |
| 3386 | /* |
| 3387 | * Re-check the pte - we dropped the lock |
| 3388 | */ |
| 3389 | vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); |
| 3390 | if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { |
| 3391 | if (old_folio) { |
| 3392 | if (!folio_test_anon(old_folio)) { |
| 3393 | dec_mm_counter(mm, mm_counter_file(old_folio)); |
| 3394 | inc_mm_counter(mm, MM_ANONPAGES); |
| 3395 | } |
| 3396 | } else { |
| 3397 | ksm_might_unmap_zero_page(mm, vmf->orig_pte); |
| 3398 | inc_mm_counter(mm, MM_ANONPAGES); |
| 3399 | } |
| 3400 | flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); |
| 3401 | entry = mk_pte(&new_folio->page, vma->vm_page_prot); |
| 3402 | entry = pte_sw_mkyoung(entry); |
| 3403 | if (unlikely(unshare)) { |
| 3404 | if (pte_soft_dirty(vmf->orig_pte)) |
| 3405 | entry = pte_mksoft_dirty(entry); |
| 3406 | if (pte_uffd_wp(vmf->orig_pte)) |
| 3407 | entry = pte_mkuffd_wp(entry); |
| 3408 | } else { |
| 3409 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 3410 | } |
| 3411 | |
| 3412 | /* |
| 3413 | * Clear the pte entry and flush it first, before updating the |
| 3414 | * pte with the new entry, to keep TLBs on different CPUs in |
| 3415 | * sync. This code used to set the new PTE then flush TLBs, but |
| 3416 | * that left a window where the new PTE could be loaded into |
| 3417 | * some TLBs while the old PTE remains in others. |
| 3418 | */ |
| 3419 | ptep_clear_flush(vma, vmf->address, vmf->pte); |
| 3420 | folio_add_new_anon_rmap(new_folio, vma, vmf->address, RMAP_EXCLUSIVE); |
| 3421 | folio_add_lru_vma(new_folio, vma); |
| 3422 | BUG_ON(unshare && pte_write(entry)); |
| 3423 | set_pte_at(mm, vmf->address, vmf->pte, entry); |
| 3424 | update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); |
| 3425 | if (old_folio) { |
| 3426 | /* |
| 3427 | * Only after switching the pte to the new page may |
| 3428 | * we remove the mapcount here. Otherwise another |
| 3429 | * process may come and find the rmap count decremented |
| 3430 | * before the pte is switched to the new page, and |
| 3431 | * "reuse" the old page writing into it while our pte |
| 3432 | * here still points into it and can be read by other |
| 3433 | * threads. |
| 3434 | * |
| 3435 | * The critical issue is to order this |
| 3436 | * folio_remove_rmap_pte() with the ptp_clear_flush |
| 3437 | * above. Those stores are ordered by (if nothing else,) |
| 3438 | * the barrier present in the atomic_add_negative |
| 3439 | * in folio_remove_rmap_pte(); |
| 3440 | * |
| 3441 | * Then the TLB flush in ptep_clear_flush ensures that |
| 3442 | * no process can access the old page before the |
| 3443 | * decremented mapcount is visible. And the old page |
| 3444 | * cannot be reused until after the decremented |
| 3445 | * mapcount is visible. So transitively, TLBs to |
| 3446 | * old page will be flushed before it can be reused. |
| 3447 | */ |
| 3448 | folio_remove_rmap_pte(old_folio, vmf->page, vma); |
| 3449 | } |
| 3450 | |
| 3451 | /* Free the old page.. */ |
| 3452 | new_folio = old_folio; |
| 3453 | page_copied = 1; |
| 3454 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3455 | } else if (vmf->pte) { |
| 3456 | update_mmu_tlb(vma, vmf->address, vmf->pte); |
| 3457 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3458 | } |
| 3459 | |
| 3460 | mmu_notifier_invalidate_range_end(&range); |
| 3461 | |
| 3462 | if (new_folio) |
| 3463 | folio_put(new_folio); |
| 3464 | if (old_folio) { |
| 3465 | if (page_copied) |
| 3466 | free_swap_cache(old_folio); |
| 3467 | folio_put(old_folio); |
| 3468 | } |
| 3469 | |
| 3470 | delayacct_wpcopy_end(); |
| 3471 | return 0; |
| 3472 | oom: |
| 3473 | ret = VM_FAULT_OOM; |
| 3474 | out: |
| 3475 | if (old_folio) |
| 3476 | folio_put(old_folio); |
| 3477 | |
| 3478 | delayacct_wpcopy_end(); |
| 3479 | return ret; |
| 3480 | } |
| 3481 | |
| 3482 | /** |
| 3483 | * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE |
| 3484 | * writeable once the page is prepared |
| 3485 | * |
| 3486 | * @vmf: structure describing the fault |
| 3487 | * @folio: the folio of vmf->page |
| 3488 | * |
| 3489 | * This function handles all that is needed to finish a write page fault in a |
| 3490 | * shared mapping due to PTE being read-only once the mapped page is prepared. |
| 3491 | * It handles locking of PTE and modifying it. |
| 3492 | * |
| 3493 | * The function expects the page to be locked or other protection against |
| 3494 | * concurrent faults / writeback (such as DAX radix tree locks). |
| 3495 | * |
| 3496 | * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before |
| 3497 | * we acquired PTE lock. |
| 3498 | */ |
| 3499 | static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio) |
| 3500 | { |
| 3501 | WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); |
| 3502 | vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, |
| 3503 | &vmf->ptl); |
| 3504 | if (!vmf->pte) |
| 3505 | return VM_FAULT_NOPAGE; |
| 3506 | /* |
| 3507 | * We might have raced with another page fault while we released the |
| 3508 | * pte_offset_map_lock. |
| 3509 | */ |
| 3510 | if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) { |
| 3511 | update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); |
| 3512 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3513 | return VM_FAULT_NOPAGE; |
| 3514 | } |
| 3515 | wp_page_reuse(vmf, folio); |
| 3516 | return 0; |
| 3517 | } |
| 3518 | |
| 3519 | /* |
| 3520 | * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED |
| 3521 | * mapping |
| 3522 | */ |
| 3523 | static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) |
| 3524 | { |
| 3525 | struct vm_area_struct *vma = vmf->vma; |
| 3526 | |
| 3527 | if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { |
| 3528 | vm_fault_t ret; |
| 3529 | |
| 3530 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3531 | ret = vmf_can_call_fault(vmf); |
| 3532 | if (ret) |
| 3533 | return ret; |
| 3534 | |
| 3535 | vmf->flags |= FAULT_FLAG_MKWRITE; |
| 3536 | ret = vma->vm_ops->pfn_mkwrite(vmf); |
| 3537 | if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) |
| 3538 | return ret; |
| 3539 | return finish_mkwrite_fault(vmf, NULL); |
| 3540 | } |
| 3541 | wp_page_reuse(vmf, NULL); |
| 3542 | return 0; |
| 3543 | } |
| 3544 | |
| 3545 | static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio) |
| 3546 | __releases(vmf->ptl) |
| 3547 | { |
| 3548 | struct vm_area_struct *vma = vmf->vma; |
| 3549 | vm_fault_t ret = 0; |
| 3550 | |
| 3551 | folio_get(folio); |
| 3552 | |
| 3553 | if (vma->vm_ops && vma->vm_ops->page_mkwrite) { |
| 3554 | vm_fault_t tmp; |
| 3555 | |
| 3556 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3557 | tmp = vmf_can_call_fault(vmf); |
| 3558 | if (tmp) { |
| 3559 | folio_put(folio); |
| 3560 | return tmp; |
| 3561 | } |
| 3562 | |
| 3563 | tmp = do_page_mkwrite(vmf, folio); |
| 3564 | if (unlikely(!tmp || (tmp & |
| 3565 | (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { |
| 3566 | folio_put(folio); |
| 3567 | return tmp; |
| 3568 | } |
| 3569 | tmp = finish_mkwrite_fault(vmf, folio); |
| 3570 | if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { |
| 3571 | folio_unlock(folio); |
| 3572 | folio_put(folio); |
| 3573 | return tmp; |
| 3574 | } |
| 3575 | } else { |
| 3576 | wp_page_reuse(vmf, folio); |
| 3577 | folio_lock(folio); |
| 3578 | } |
| 3579 | ret |= fault_dirty_shared_page(vmf); |
| 3580 | folio_put(folio); |
| 3581 | |
| 3582 | return ret; |
| 3583 | } |
| 3584 | |
| 3585 | static bool wp_can_reuse_anon_folio(struct folio *folio, |
| 3586 | struct vm_area_struct *vma) |
| 3587 | { |
| 3588 | /* |
| 3589 | * We could currently only reuse a subpage of a large folio if no |
| 3590 | * other subpages of the large folios are still mapped. However, |
| 3591 | * let's just consistently not reuse subpages even if we could |
| 3592 | * reuse in that scenario, and give back a large folio a bit |
| 3593 | * sooner. |
| 3594 | */ |
| 3595 | if (folio_test_large(folio)) |
| 3596 | return false; |
| 3597 | |
| 3598 | /* |
| 3599 | * We have to verify under folio lock: these early checks are |
| 3600 | * just an optimization to avoid locking the folio and freeing |
| 3601 | * the swapcache if there is little hope that we can reuse. |
| 3602 | * |
| 3603 | * KSM doesn't necessarily raise the folio refcount. |
| 3604 | */ |
| 3605 | if (folio_test_ksm(folio) || folio_ref_count(folio) > 3) |
| 3606 | return false; |
| 3607 | if (!folio_test_lru(folio)) |
| 3608 | /* |
| 3609 | * We cannot easily detect+handle references from |
| 3610 | * remote LRU caches or references to LRU folios. |
| 3611 | */ |
| 3612 | lru_add_drain(); |
| 3613 | if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio)) |
| 3614 | return false; |
| 3615 | if (!folio_trylock(folio)) |
| 3616 | return false; |
| 3617 | if (folio_test_swapcache(folio)) |
| 3618 | folio_free_swap(folio); |
| 3619 | if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) { |
| 3620 | folio_unlock(folio); |
| 3621 | return false; |
| 3622 | } |
| 3623 | /* |
| 3624 | * Ok, we've got the only folio reference from our mapping |
| 3625 | * and the folio is locked, it's dark out, and we're wearing |
| 3626 | * sunglasses. Hit it. |
| 3627 | */ |
| 3628 | folio_move_anon_rmap(folio, vma); |
| 3629 | folio_unlock(folio); |
| 3630 | return true; |
| 3631 | } |
| 3632 | |
| 3633 | /* |
| 3634 | * This routine handles present pages, when |
| 3635 | * * users try to write to a shared page (FAULT_FLAG_WRITE) |
| 3636 | * * GUP wants to take a R/O pin on a possibly shared anonymous page |
| 3637 | * (FAULT_FLAG_UNSHARE) |
| 3638 | * |
| 3639 | * It is done by copying the page to a new address and decrementing the |
| 3640 | * shared-page counter for the old page. |
| 3641 | * |
| 3642 | * Note that this routine assumes that the protection checks have been |
| 3643 | * done by the caller (the low-level page fault routine in most cases). |
| 3644 | * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've |
| 3645 | * done any necessary COW. |
| 3646 | * |
| 3647 | * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even |
| 3648 | * though the page will change only once the write actually happens. This |
| 3649 | * avoids a few races, and potentially makes it more efficient. |
| 3650 | * |
| 3651 | * We enter with non-exclusive mmap_lock (to exclude vma changes, |
| 3652 | * but allow concurrent faults), with pte both mapped and locked. |
| 3653 | * We return with mmap_lock still held, but pte unmapped and unlocked. |
| 3654 | */ |
| 3655 | static vm_fault_t do_wp_page(struct vm_fault *vmf) |
| 3656 | __releases(vmf->ptl) |
| 3657 | { |
| 3658 | const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; |
| 3659 | struct vm_area_struct *vma = vmf->vma; |
| 3660 | struct folio *folio = NULL; |
| 3661 | pte_t pte; |
| 3662 | |
| 3663 | if (likely(!unshare)) { |
| 3664 | if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) { |
| 3665 | if (!userfaultfd_wp_async(vma)) { |
| 3666 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3667 | return handle_userfault(vmf, VM_UFFD_WP); |
| 3668 | } |
| 3669 | |
| 3670 | /* |
| 3671 | * Nothing needed (cache flush, TLB invalidations, |
| 3672 | * etc.) because we're only removing the uffd-wp bit, |
| 3673 | * which is completely invisible to the user. |
| 3674 | */ |
| 3675 | pte = pte_clear_uffd_wp(ptep_get(vmf->pte)); |
| 3676 | |
| 3677 | set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); |
| 3678 | /* |
| 3679 | * Update this to be prepared for following up CoW |
| 3680 | * handling |
| 3681 | */ |
| 3682 | vmf->orig_pte = pte; |
| 3683 | } |
| 3684 | |
| 3685 | /* |
| 3686 | * Userfaultfd write-protect can defer flushes. Ensure the TLB |
| 3687 | * is flushed in this case before copying. |
| 3688 | */ |
| 3689 | if (unlikely(userfaultfd_wp(vmf->vma) && |
| 3690 | mm_tlb_flush_pending(vmf->vma->vm_mm))) |
| 3691 | flush_tlb_page(vmf->vma, vmf->address); |
| 3692 | } |
| 3693 | |
| 3694 | vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); |
| 3695 | |
| 3696 | if (vmf->page) |
| 3697 | folio = page_folio(vmf->page); |
| 3698 | |
| 3699 | /* |
| 3700 | * Shared mapping: we are guaranteed to have VM_WRITE and |
| 3701 | * FAULT_FLAG_WRITE set at this point. |
| 3702 | */ |
| 3703 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { |
| 3704 | /* |
| 3705 | * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a |
| 3706 | * VM_PFNMAP VMA. |
| 3707 | * |
| 3708 | * We should not cow pages in a shared writeable mapping. |
| 3709 | * Just mark the pages writable and/or call ops->pfn_mkwrite. |
| 3710 | */ |
| 3711 | if (!vmf->page) |
| 3712 | return wp_pfn_shared(vmf); |
| 3713 | return wp_page_shared(vmf, folio); |
| 3714 | } |
| 3715 | |
| 3716 | /* |
| 3717 | * Private mapping: create an exclusive anonymous page copy if reuse |
| 3718 | * is impossible. We might miss VM_WRITE for FOLL_FORCE handling. |
| 3719 | * |
| 3720 | * If we encounter a page that is marked exclusive, we must reuse |
| 3721 | * the page without further checks. |
| 3722 | */ |
| 3723 | if (folio && folio_test_anon(folio) && |
| 3724 | (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) { |
| 3725 | if (!PageAnonExclusive(vmf->page)) |
| 3726 | SetPageAnonExclusive(vmf->page); |
| 3727 | if (unlikely(unshare)) { |
| 3728 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3729 | return 0; |
| 3730 | } |
| 3731 | wp_page_reuse(vmf, folio); |
| 3732 | return 0; |
| 3733 | } |
| 3734 | /* |
| 3735 | * Ok, we need to copy. Oh, well.. |
| 3736 | */ |
| 3737 | if (folio) |
| 3738 | folio_get(folio); |
| 3739 | |
| 3740 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3741 | #ifdef CONFIG_KSM |
| 3742 | if (folio && folio_test_ksm(folio)) |
| 3743 | count_vm_event(COW_KSM); |
| 3744 | #endif |
| 3745 | return wp_page_copy(vmf); |
| 3746 | } |
| 3747 | |
| 3748 | static void unmap_mapping_range_vma(struct vm_area_struct *vma, |
| 3749 | unsigned long start_addr, unsigned long end_addr, |
| 3750 | struct zap_details *details) |
| 3751 | { |
| 3752 | zap_page_range_single(vma, start_addr, end_addr - start_addr, details); |
| 3753 | } |
| 3754 | |
| 3755 | static inline void unmap_mapping_range_tree(struct rb_root_cached *root, |
| 3756 | pgoff_t first_index, |
| 3757 | pgoff_t last_index, |
| 3758 | struct zap_details *details) |
| 3759 | { |
| 3760 | struct vm_area_struct *vma; |
| 3761 | pgoff_t vba, vea, zba, zea; |
| 3762 | |
| 3763 | vma_interval_tree_foreach(vma, root, first_index, last_index) { |
| 3764 | vba = vma->vm_pgoff; |
| 3765 | vea = vba + vma_pages(vma) - 1; |
| 3766 | zba = max(first_index, vba); |
| 3767 | zea = min(last_index, vea); |
| 3768 | |
| 3769 | unmap_mapping_range_vma(vma, |
| 3770 | ((zba - vba) << PAGE_SHIFT) + vma->vm_start, |
| 3771 | ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, |
| 3772 | details); |
| 3773 | } |
| 3774 | } |
| 3775 | |
| 3776 | /** |
| 3777 | * unmap_mapping_folio() - Unmap single folio from processes. |
| 3778 | * @folio: The locked folio to be unmapped. |
| 3779 | * |
| 3780 | * Unmap this folio from any userspace process which still has it mmaped. |
| 3781 | * Typically, for efficiency, the range of nearby pages has already been |
| 3782 | * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once |
| 3783 | * truncation or invalidation holds the lock on a folio, it may find that |
| 3784 | * the page has been remapped again: and then uses unmap_mapping_folio() |
| 3785 | * to unmap it finally. |
| 3786 | */ |
| 3787 | void unmap_mapping_folio(struct folio *folio) |
| 3788 | { |
| 3789 | struct address_space *mapping = folio->mapping; |
| 3790 | struct zap_details details = { }; |
| 3791 | pgoff_t first_index; |
| 3792 | pgoff_t last_index; |
| 3793 | |
| 3794 | VM_BUG_ON(!folio_test_locked(folio)); |
| 3795 | |
| 3796 | first_index = folio->index; |
| 3797 | last_index = folio_next_index(folio) - 1; |
| 3798 | |
| 3799 | details.even_cows = false; |
| 3800 | details.single_folio = folio; |
| 3801 | details.zap_flags = ZAP_FLAG_DROP_MARKER; |
| 3802 | |
| 3803 | i_mmap_lock_read(mapping); |
| 3804 | if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) |
| 3805 | unmap_mapping_range_tree(&mapping->i_mmap, first_index, |
| 3806 | last_index, &details); |
| 3807 | i_mmap_unlock_read(mapping); |
| 3808 | } |
| 3809 | |
| 3810 | /** |
| 3811 | * unmap_mapping_pages() - Unmap pages from processes. |
| 3812 | * @mapping: The address space containing pages to be unmapped. |
| 3813 | * @start: Index of first page to be unmapped. |
| 3814 | * @nr: Number of pages to be unmapped. 0 to unmap to end of file. |
| 3815 | * @even_cows: Whether to unmap even private COWed pages. |
| 3816 | * |
| 3817 | * Unmap the pages in this address space from any userspace process which |
| 3818 | * has them mmaped. Generally, you want to remove COWed pages as well when |
| 3819 | * a file is being truncated, but not when invalidating pages from the page |
| 3820 | * cache. |
| 3821 | */ |
| 3822 | void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, |
| 3823 | pgoff_t nr, bool even_cows) |
| 3824 | { |
| 3825 | struct zap_details details = { }; |
| 3826 | pgoff_t first_index = start; |
| 3827 | pgoff_t last_index = start + nr - 1; |
| 3828 | |
| 3829 | details.even_cows = even_cows; |
| 3830 | if (last_index < first_index) |
| 3831 | last_index = ULONG_MAX; |
| 3832 | |
| 3833 | i_mmap_lock_read(mapping); |
| 3834 | if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) |
| 3835 | unmap_mapping_range_tree(&mapping->i_mmap, first_index, |
| 3836 | last_index, &details); |
| 3837 | i_mmap_unlock_read(mapping); |
| 3838 | } |
| 3839 | EXPORT_SYMBOL_GPL(unmap_mapping_pages); |
| 3840 | |
| 3841 | /** |
| 3842 | * unmap_mapping_range - unmap the portion of all mmaps in the specified |
| 3843 | * address_space corresponding to the specified byte range in the underlying |
| 3844 | * file. |
| 3845 | * |
| 3846 | * @mapping: the address space containing mmaps to be unmapped. |
| 3847 | * @holebegin: byte in first page to unmap, relative to the start of |
| 3848 | * the underlying file. This will be rounded down to a PAGE_SIZE |
| 3849 | * boundary. Note that this is different from truncate_pagecache(), which |
| 3850 | * must keep the partial page. In contrast, we must get rid of |
| 3851 | * partial pages. |
| 3852 | * @holelen: size of prospective hole in bytes. This will be rounded |
| 3853 | * up to a PAGE_SIZE boundary. A holelen of zero truncates to the |
| 3854 | * end of the file. |
| 3855 | * @even_cows: 1 when truncating a file, unmap even private COWed pages; |
| 3856 | * but 0 when invalidating pagecache, don't throw away private data. |
| 3857 | */ |
| 3858 | void unmap_mapping_range(struct address_space *mapping, |
| 3859 | loff_t const holebegin, loff_t const holelen, int even_cows) |
| 3860 | { |
| 3861 | pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT; |
| 3862 | pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT; |
| 3863 | |
| 3864 | /* Check for overflow. */ |
| 3865 | if (sizeof(holelen) > sizeof(hlen)) { |
| 3866 | long long holeend = |
| 3867 | (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; |
| 3868 | if (holeend & ~(long long)ULONG_MAX) |
| 3869 | hlen = ULONG_MAX - hba + 1; |
| 3870 | } |
| 3871 | |
| 3872 | unmap_mapping_pages(mapping, hba, hlen, even_cows); |
| 3873 | } |
| 3874 | EXPORT_SYMBOL(unmap_mapping_range); |
| 3875 | |
| 3876 | /* |
| 3877 | * Restore a potential device exclusive pte to a working pte entry |
| 3878 | */ |
| 3879 | static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf) |
| 3880 | { |
| 3881 | struct folio *folio = page_folio(vmf->page); |
| 3882 | struct vm_area_struct *vma = vmf->vma; |
| 3883 | struct mmu_notifier_range range; |
| 3884 | vm_fault_t ret; |
| 3885 | |
| 3886 | /* |
| 3887 | * We need a reference to lock the folio because we don't hold |
| 3888 | * the PTL so a racing thread can remove the device-exclusive |
| 3889 | * entry and unmap it. If the folio is free the entry must |
| 3890 | * have been removed already. If it happens to have already |
| 3891 | * been re-allocated after being freed all we do is lock and |
| 3892 | * unlock it. |
| 3893 | */ |
| 3894 | if (!folio_try_get(folio)) |
| 3895 | return 0; |
| 3896 | |
| 3897 | ret = folio_lock_or_retry(folio, vmf); |
| 3898 | if (ret) { |
| 3899 | folio_put(folio); |
| 3900 | return ret; |
| 3901 | } |
| 3902 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, |
| 3903 | vma->vm_mm, vmf->address & PAGE_MASK, |
| 3904 | (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL); |
| 3905 | mmu_notifier_invalidate_range_start(&range); |
| 3906 | |
| 3907 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, |
| 3908 | &vmf->ptl); |
| 3909 | if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) |
| 3910 | restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte); |
| 3911 | |
| 3912 | if (vmf->pte) |
| 3913 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3914 | folio_unlock(folio); |
| 3915 | folio_put(folio); |
| 3916 | |
| 3917 | mmu_notifier_invalidate_range_end(&range); |
| 3918 | return 0; |
| 3919 | } |
| 3920 | |
| 3921 | static inline bool should_try_to_free_swap(struct folio *folio, |
| 3922 | struct vm_area_struct *vma, |
| 3923 | unsigned int fault_flags) |
| 3924 | { |
| 3925 | if (!folio_test_swapcache(folio)) |
| 3926 | return false; |
| 3927 | if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) || |
| 3928 | folio_test_mlocked(folio)) |
| 3929 | return true; |
| 3930 | /* |
| 3931 | * If we want to map a page that's in the swapcache writable, we |
| 3932 | * have to detect via the refcount if we're really the exclusive |
| 3933 | * user. Try freeing the swapcache to get rid of the swapcache |
| 3934 | * reference only in case it's likely that we'll be the exlusive user. |
| 3935 | */ |
| 3936 | return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) && |
| 3937 | folio_ref_count(folio) == (1 + folio_nr_pages(folio)); |
| 3938 | } |
| 3939 | |
| 3940 | static vm_fault_t pte_marker_clear(struct vm_fault *vmf) |
| 3941 | { |
| 3942 | vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, |
| 3943 | vmf->address, &vmf->ptl); |
| 3944 | if (!vmf->pte) |
| 3945 | return 0; |
| 3946 | /* |
| 3947 | * Be careful so that we will only recover a special uffd-wp pte into a |
| 3948 | * none pte. Otherwise it means the pte could have changed, so retry. |
| 3949 | * |
| 3950 | * This should also cover the case where e.g. the pte changed |
| 3951 | * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED. |
| 3952 | * So is_pte_marker() check is not enough to safely drop the pte. |
| 3953 | */ |
| 3954 | if (pte_same(vmf->orig_pte, ptep_get(vmf->pte))) |
| 3955 | pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte); |
| 3956 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 3957 | return 0; |
| 3958 | } |
| 3959 | |
| 3960 | static vm_fault_t do_pte_missing(struct vm_fault *vmf) |
| 3961 | { |
| 3962 | if (vma_is_anonymous(vmf->vma)) |
| 3963 | return do_anonymous_page(vmf); |
| 3964 | else |
| 3965 | return do_fault(vmf); |
| 3966 | } |
| 3967 | |
| 3968 | /* |
| 3969 | * This is actually a page-missing access, but with uffd-wp special pte |
| 3970 | * installed. It means this pte was wr-protected before being unmapped. |
| 3971 | */ |
| 3972 | static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf) |
| 3973 | { |
| 3974 | /* |
| 3975 | * Just in case there're leftover special ptes even after the region |
| 3976 | * got unregistered - we can simply clear them. |
| 3977 | */ |
| 3978 | if (unlikely(!userfaultfd_wp(vmf->vma))) |
| 3979 | return pte_marker_clear(vmf); |
| 3980 | |
| 3981 | return do_pte_missing(vmf); |
| 3982 | } |
| 3983 | |
| 3984 | static vm_fault_t handle_pte_marker(struct vm_fault *vmf) |
| 3985 | { |
| 3986 | swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte); |
| 3987 | unsigned long marker = pte_marker_get(entry); |
| 3988 | |
| 3989 | /* |
| 3990 | * PTE markers should never be empty. If anything weird happened, |
| 3991 | * the best thing to do is to kill the process along with its mm. |
| 3992 | */ |
| 3993 | if (WARN_ON_ONCE(!marker)) |
| 3994 | return VM_FAULT_SIGBUS; |
| 3995 | |
| 3996 | /* Higher priority than uffd-wp when data corrupted */ |
| 3997 | if (marker & PTE_MARKER_POISONED) |
| 3998 | return VM_FAULT_HWPOISON; |
| 3999 | |
| 4000 | if (pte_marker_entry_uffd_wp(entry)) |
| 4001 | return pte_marker_handle_uffd_wp(vmf); |
| 4002 | |
| 4003 | /* This is an unknown pte marker */ |
| 4004 | return VM_FAULT_SIGBUS; |
| 4005 | } |
| 4006 | |
| 4007 | static struct folio *__alloc_swap_folio(struct vm_fault *vmf) |
| 4008 | { |
| 4009 | struct vm_area_struct *vma = vmf->vma; |
| 4010 | struct folio *folio; |
| 4011 | swp_entry_t entry; |
| 4012 | |
| 4013 | folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, |
| 4014 | vmf->address, false); |
| 4015 | if (!folio) |
| 4016 | return NULL; |
| 4017 | |
| 4018 | entry = pte_to_swp_entry(vmf->orig_pte); |
| 4019 | if (mem_cgroup_swapin_charge_folio(folio, vma->vm_mm, |
| 4020 | GFP_KERNEL, entry)) { |
| 4021 | folio_put(folio); |
| 4022 | return NULL; |
| 4023 | } |
| 4024 | |
| 4025 | return folio; |
| 4026 | } |
| 4027 | |
| 4028 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 4029 | static inline int non_swapcache_batch(swp_entry_t entry, int max_nr) |
| 4030 | { |
| 4031 | struct swap_info_struct *si = swp_swap_info(entry); |
| 4032 | pgoff_t offset = swp_offset(entry); |
| 4033 | int i; |
| 4034 | |
| 4035 | /* |
| 4036 | * While allocating a large folio and doing swap_read_folio, which is |
| 4037 | * the case the being faulted pte doesn't have swapcache. We need to |
| 4038 | * ensure all PTEs have no cache as well, otherwise, we might go to |
| 4039 | * swap devices while the content is in swapcache. |
| 4040 | */ |
| 4041 | for (i = 0; i < max_nr; i++) { |
| 4042 | if ((si->swap_map[offset + i] & SWAP_HAS_CACHE)) |
| 4043 | return i; |
| 4044 | } |
| 4045 | |
| 4046 | return i; |
| 4047 | } |
| 4048 | |
| 4049 | /* |
| 4050 | * Check if the PTEs within a range are contiguous swap entries |
| 4051 | * and have consistent swapcache, zeromap. |
| 4052 | */ |
| 4053 | static bool can_swapin_thp(struct vm_fault *vmf, pte_t *ptep, int nr_pages) |
| 4054 | { |
| 4055 | unsigned long addr; |
| 4056 | swp_entry_t entry; |
| 4057 | int idx; |
| 4058 | pte_t pte; |
| 4059 | |
| 4060 | addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE); |
| 4061 | idx = (vmf->address - addr) / PAGE_SIZE; |
| 4062 | pte = ptep_get(ptep); |
| 4063 | |
| 4064 | if (!pte_same(pte, pte_move_swp_offset(vmf->orig_pte, -idx))) |
| 4065 | return false; |
| 4066 | entry = pte_to_swp_entry(pte); |
| 4067 | if (swap_pte_batch(ptep, nr_pages, pte) != nr_pages) |
| 4068 | return false; |
| 4069 | |
| 4070 | /* |
| 4071 | * swap_read_folio() can't handle the case a large folio is hybridly |
| 4072 | * from different backends. And they are likely corner cases. Similar |
| 4073 | * things might be added once zswap support large folios. |
| 4074 | */ |
| 4075 | if (unlikely(swap_zeromap_batch(entry, nr_pages, NULL) != nr_pages)) |
| 4076 | return false; |
| 4077 | if (unlikely(non_swapcache_batch(entry, nr_pages) != nr_pages)) |
| 4078 | return false; |
| 4079 | |
| 4080 | return true; |
| 4081 | } |
| 4082 | |
| 4083 | static inline unsigned long thp_swap_suitable_orders(pgoff_t swp_offset, |
| 4084 | unsigned long addr, |
| 4085 | unsigned long orders) |
| 4086 | { |
| 4087 | int order, nr; |
| 4088 | |
| 4089 | order = highest_order(orders); |
| 4090 | |
| 4091 | /* |
| 4092 | * To swap in a THP with nr pages, we require that its first swap_offset |
| 4093 | * is aligned with that number, as it was when the THP was swapped out. |
| 4094 | * This helps filter out most invalid entries. |
| 4095 | */ |
| 4096 | while (orders) { |
| 4097 | nr = 1 << order; |
| 4098 | if ((addr >> PAGE_SHIFT) % nr == swp_offset % nr) |
| 4099 | break; |
| 4100 | order = next_order(&orders, order); |
| 4101 | } |
| 4102 | |
| 4103 | return orders; |
| 4104 | } |
| 4105 | |
| 4106 | static struct folio *alloc_swap_folio(struct vm_fault *vmf) |
| 4107 | { |
| 4108 | struct vm_area_struct *vma = vmf->vma; |
| 4109 | unsigned long orders; |
| 4110 | struct folio *folio; |
| 4111 | unsigned long addr; |
| 4112 | swp_entry_t entry; |
| 4113 | spinlock_t *ptl; |
| 4114 | pte_t *pte; |
| 4115 | gfp_t gfp; |
| 4116 | int order; |
| 4117 | |
| 4118 | /* |
| 4119 | * If uffd is active for the vma we need per-page fault fidelity to |
| 4120 | * maintain the uffd semantics. |
| 4121 | */ |
| 4122 | if (unlikely(userfaultfd_armed(vma))) |
| 4123 | goto fallback; |
| 4124 | |
| 4125 | /* |
| 4126 | * A large swapped out folio could be partially or fully in zswap. We |
| 4127 | * lack handling for such cases, so fallback to swapping in order-0 |
| 4128 | * folio. |
| 4129 | */ |
| 4130 | if (!zswap_never_enabled()) |
| 4131 | goto fallback; |
| 4132 | |
| 4133 | entry = pte_to_swp_entry(vmf->orig_pte); |
| 4134 | /* |
| 4135 | * Get a list of all the (large) orders below PMD_ORDER that are enabled |
| 4136 | * and suitable for swapping THP. |
| 4137 | */ |
| 4138 | orders = thp_vma_allowable_orders(vma, vma->vm_flags, |
| 4139 | TVA_IN_PF | TVA_ENFORCE_SYSFS, BIT(PMD_ORDER) - 1); |
| 4140 | orders = thp_vma_suitable_orders(vma, vmf->address, orders); |
| 4141 | orders = thp_swap_suitable_orders(swp_offset(entry), |
| 4142 | vmf->address, orders); |
| 4143 | |
| 4144 | if (!orders) |
| 4145 | goto fallback; |
| 4146 | |
| 4147 | pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, |
| 4148 | vmf->address & PMD_MASK, &ptl); |
| 4149 | if (unlikely(!pte)) |
| 4150 | goto fallback; |
| 4151 | |
| 4152 | /* |
| 4153 | * For do_swap_page, find the highest order where the aligned range is |
| 4154 | * completely swap entries with contiguous swap offsets. |
| 4155 | */ |
| 4156 | order = highest_order(orders); |
| 4157 | while (orders) { |
| 4158 | addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); |
| 4159 | if (can_swapin_thp(vmf, pte + pte_index(addr), 1 << order)) |
| 4160 | break; |
| 4161 | order = next_order(&orders, order); |
| 4162 | } |
| 4163 | |
| 4164 | pte_unmap_unlock(pte, ptl); |
| 4165 | |
| 4166 | /* Try allocating the highest of the remaining orders. */ |
| 4167 | gfp = vma_thp_gfp_mask(vma); |
| 4168 | while (orders) { |
| 4169 | addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); |
| 4170 | folio = vma_alloc_folio(gfp, order, vma, addr, true); |
| 4171 | if (folio) { |
| 4172 | if (!mem_cgroup_swapin_charge_folio(folio, vma->vm_mm, |
| 4173 | gfp, entry)) |
| 4174 | return folio; |
| 4175 | folio_put(folio); |
| 4176 | } |
| 4177 | order = next_order(&orders, order); |
| 4178 | } |
| 4179 | |
| 4180 | fallback: |
| 4181 | return __alloc_swap_folio(vmf); |
| 4182 | } |
| 4183 | #else /* !CONFIG_TRANSPARENT_HUGEPAGE */ |
| 4184 | static inline bool can_swapin_thp(struct vm_fault *vmf, pte_t *ptep, int nr_pages) |
| 4185 | { |
| 4186 | return false; |
| 4187 | } |
| 4188 | |
| 4189 | static struct folio *alloc_swap_folio(struct vm_fault *vmf) |
| 4190 | { |
| 4191 | return __alloc_swap_folio(vmf); |
| 4192 | } |
| 4193 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |
| 4194 | |
| 4195 | /* |
| 4196 | * We enter with non-exclusive mmap_lock (to exclude vma changes, |
| 4197 | * but allow concurrent faults), and pte mapped but not yet locked. |
| 4198 | * We return with pte unmapped and unlocked. |
| 4199 | * |
| 4200 | * We return with the mmap_lock locked or unlocked in the same cases |
| 4201 | * as does filemap_fault(). |
| 4202 | */ |
| 4203 | vm_fault_t do_swap_page(struct vm_fault *vmf) |
| 4204 | { |
| 4205 | struct vm_area_struct *vma = vmf->vma; |
| 4206 | struct folio *swapcache, *folio = NULL; |
| 4207 | struct page *page; |
| 4208 | struct swap_info_struct *si = NULL; |
| 4209 | rmap_t rmap_flags = RMAP_NONE; |
| 4210 | bool need_clear_cache = false; |
| 4211 | bool exclusive = false; |
| 4212 | swp_entry_t entry; |
| 4213 | pte_t pte; |
| 4214 | vm_fault_t ret = 0; |
| 4215 | void *shadow = NULL; |
| 4216 | int nr_pages; |
| 4217 | unsigned long page_idx; |
| 4218 | unsigned long address; |
| 4219 | pte_t *ptep; |
| 4220 | |
| 4221 | if (!pte_unmap_same(vmf)) |
| 4222 | goto out; |
| 4223 | |
| 4224 | entry = pte_to_swp_entry(vmf->orig_pte); |
| 4225 | if (unlikely(non_swap_entry(entry))) { |
| 4226 | if (is_migration_entry(entry)) { |
| 4227 | migration_entry_wait(vma->vm_mm, vmf->pmd, |
| 4228 | vmf->address); |
| 4229 | } else if (is_device_exclusive_entry(entry)) { |
| 4230 | vmf->page = pfn_swap_entry_to_page(entry); |
| 4231 | ret = remove_device_exclusive_entry(vmf); |
| 4232 | } else if (is_device_private_entry(entry)) { |
| 4233 | if (vmf->flags & FAULT_FLAG_VMA_LOCK) { |
| 4234 | /* |
| 4235 | * migrate_to_ram is not yet ready to operate |
| 4236 | * under VMA lock. |
| 4237 | */ |
| 4238 | vma_end_read(vma); |
| 4239 | ret = VM_FAULT_RETRY; |
| 4240 | goto out; |
| 4241 | } |
| 4242 | |
| 4243 | vmf->page = pfn_swap_entry_to_page(entry); |
| 4244 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, |
| 4245 | vmf->address, &vmf->ptl); |
| 4246 | if (unlikely(!vmf->pte || |
| 4247 | !pte_same(ptep_get(vmf->pte), |
| 4248 | vmf->orig_pte))) |
| 4249 | goto unlock; |
| 4250 | |
| 4251 | /* |
| 4252 | * Get a page reference while we know the page can't be |
| 4253 | * freed. |
| 4254 | */ |
| 4255 | get_page(vmf->page); |
| 4256 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4257 | ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); |
| 4258 | put_page(vmf->page); |
| 4259 | } else if (is_hwpoison_entry(entry)) { |
| 4260 | ret = VM_FAULT_HWPOISON; |
| 4261 | } else if (is_pte_marker_entry(entry)) { |
| 4262 | ret = handle_pte_marker(vmf); |
| 4263 | } else { |
| 4264 | print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); |
| 4265 | ret = VM_FAULT_SIGBUS; |
| 4266 | } |
| 4267 | goto out; |
| 4268 | } |
| 4269 | |
| 4270 | /* Prevent swapoff from happening to us. */ |
| 4271 | si = get_swap_device(entry); |
| 4272 | if (unlikely(!si)) |
| 4273 | goto out; |
| 4274 | |
| 4275 | folio = swap_cache_get_folio(entry, vma, vmf->address); |
| 4276 | if (folio) |
| 4277 | page = folio_file_page(folio, swp_offset(entry)); |
| 4278 | swapcache = folio; |
| 4279 | |
| 4280 | if (!folio) { |
| 4281 | if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && |
| 4282 | __swap_count(entry) == 1) { |
| 4283 | /* skip swapcache */ |
| 4284 | folio = alloc_swap_folio(vmf); |
| 4285 | if (folio) { |
| 4286 | __folio_set_locked(folio); |
| 4287 | __folio_set_swapbacked(folio); |
| 4288 | |
| 4289 | nr_pages = folio_nr_pages(folio); |
| 4290 | if (folio_test_large(folio)) |
| 4291 | entry.val = ALIGN_DOWN(entry.val, nr_pages); |
| 4292 | /* |
| 4293 | * Prevent parallel swapin from proceeding with |
| 4294 | * the cache flag. Otherwise, another thread |
| 4295 | * may finish swapin first, free the entry, and |
| 4296 | * swapout reusing the same entry. It's |
| 4297 | * undetectable as pte_same() returns true due |
| 4298 | * to entry reuse. |
| 4299 | */ |
| 4300 | if (swapcache_prepare(entry, nr_pages)) { |
| 4301 | /* |
| 4302 | * Relax a bit to prevent rapid |
| 4303 | * repeated page faults. |
| 4304 | */ |
| 4305 | schedule_timeout_uninterruptible(1); |
| 4306 | goto out_page; |
| 4307 | } |
| 4308 | need_clear_cache = true; |
| 4309 | |
| 4310 | mem_cgroup_swapin_uncharge_swap(entry, nr_pages); |
| 4311 | |
| 4312 | shadow = get_shadow_from_swap_cache(entry); |
| 4313 | if (shadow) |
| 4314 | workingset_refault(folio, shadow); |
| 4315 | |
| 4316 | folio_add_lru(folio); |
| 4317 | |
| 4318 | /* To provide entry to swap_read_folio() */ |
| 4319 | folio->swap = entry; |
| 4320 | swap_read_folio(folio, NULL); |
| 4321 | folio->private = NULL; |
| 4322 | } |
| 4323 | } else { |
| 4324 | folio = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, |
| 4325 | vmf); |
| 4326 | swapcache = folio; |
| 4327 | } |
| 4328 | |
| 4329 | if (!folio) { |
| 4330 | /* |
| 4331 | * Back out if somebody else faulted in this pte |
| 4332 | * while we released the pte lock. |
| 4333 | */ |
| 4334 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, |
| 4335 | vmf->address, &vmf->ptl); |
| 4336 | if (likely(vmf->pte && |
| 4337 | pte_same(ptep_get(vmf->pte), vmf->orig_pte))) |
| 4338 | ret = VM_FAULT_OOM; |
| 4339 | goto unlock; |
| 4340 | } |
| 4341 | |
| 4342 | /* Had to read the page from swap area: Major fault */ |
| 4343 | ret = VM_FAULT_MAJOR; |
| 4344 | count_vm_event(PGMAJFAULT); |
| 4345 | count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); |
| 4346 | page = folio_file_page(folio, swp_offset(entry)); |
| 4347 | } else if (PageHWPoison(page)) { |
| 4348 | /* |
| 4349 | * hwpoisoned dirty swapcache pages are kept for killing |
| 4350 | * owner processes (which may be unknown at hwpoison time) |
| 4351 | */ |
| 4352 | ret = VM_FAULT_HWPOISON; |
| 4353 | goto out_release; |
| 4354 | } |
| 4355 | |
| 4356 | ret |= folio_lock_or_retry(folio, vmf); |
| 4357 | if (ret & VM_FAULT_RETRY) |
| 4358 | goto out_release; |
| 4359 | |
| 4360 | if (swapcache) { |
| 4361 | /* |
| 4362 | * Make sure folio_free_swap() or swapoff did not release the |
| 4363 | * swapcache from under us. The page pin, and pte_same test |
| 4364 | * below, are not enough to exclude that. Even if it is still |
| 4365 | * swapcache, we need to check that the page's swap has not |
| 4366 | * changed. |
| 4367 | */ |
| 4368 | if (unlikely(!folio_test_swapcache(folio) || |
| 4369 | page_swap_entry(page).val != entry.val)) |
| 4370 | goto out_page; |
| 4371 | |
| 4372 | /* |
| 4373 | * KSM sometimes has to copy on read faults, for example, if |
| 4374 | * page->index of !PageKSM() pages would be nonlinear inside the |
| 4375 | * anon VMA -- PageKSM() is lost on actual swapout. |
| 4376 | */ |
| 4377 | folio = ksm_might_need_to_copy(folio, vma, vmf->address); |
| 4378 | if (unlikely(!folio)) { |
| 4379 | ret = VM_FAULT_OOM; |
| 4380 | folio = swapcache; |
| 4381 | goto out_page; |
| 4382 | } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) { |
| 4383 | ret = VM_FAULT_HWPOISON; |
| 4384 | folio = swapcache; |
| 4385 | goto out_page; |
| 4386 | } |
| 4387 | if (folio != swapcache) |
| 4388 | page = folio_page(folio, 0); |
| 4389 | |
| 4390 | /* |
| 4391 | * If we want to map a page that's in the swapcache writable, we |
| 4392 | * have to detect via the refcount if we're really the exclusive |
| 4393 | * owner. Try removing the extra reference from the local LRU |
| 4394 | * caches if required. |
| 4395 | */ |
| 4396 | if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache && |
| 4397 | !folio_test_ksm(folio) && !folio_test_lru(folio)) |
| 4398 | lru_add_drain(); |
| 4399 | } |
| 4400 | |
| 4401 | folio_throttle_swaprate(folio, GFP_KERNEL); |
| 4402 | |
| 4403 | /* |
| 4404 | * Back out if somebody else already faulted in this pte. |
| 4405 | */ |
| 4406 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, |
| 4407 | &vmf->ptl); |
| 4408 | if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) |
| 4409 | goto out_nomap; |
| 4410 | |
| 4411 | if (unlikely(!folio_test_uptodate(folio))) { |
| 4412 | ret = VM_FAULT_SIGBUS; |
| 4413 | goto out_nomap; |
| 4414 | } |
| 4415 | |
| 4416 | /* allocated large folios for SWP_SYNCHRONOUS_IO */ |
| 4417 | if (folio_test_large(folio) && !folio_test_swapcache(folio)) { |
| 4418 | unsigned long nr = folio_nr_pages(folio); |
| 4419 | unsigned long folio_start = ALIGN_DOWN(vmf->address, nr * PAGE_SIZE); |
| 4420 | unsigned long idx = (vmf->address - folio_start) / PAGE_SIZE; |
| 4421 | pte_t *folio_ptep = vmf->pte - idx; |
| 4422 | pte_t folio_pte = ptep_get(folio_ptep); |
| 4423 | |
| 4424 | if (!pte_same(folio_pte, pte_move_swp_offset(vmf->orig_pte, -idx)) || |
| 4425 | swap_pte_batch(folio_ptep, nr, folio_pte) != nr) |
| 4426 | goto out_nomap; |
| 4427 | |
| 4428 | page_idx = idx; |
| 4429 | address = folio_start; |
| 4430 | ptep = folio_ptep; |
| 4431 | goto check_folio; |
| 4432 | } |
| 4433 | |
| 4434 | nr_pages = 1; |
| 4435 | page_idx = 0; |
| 4436 | address = vmf->address; |
| 4437 | ptep = vmf->pte; |
| 4438 | if (folio_test_large(folio) && folio_test_swapcache(folio)) { |
| 4439 | int nr = folio_nr_pages(folio); |
| 4440 | unsigned long idx = folio_page_idx(folio, page); |
| 4441 | unsigned long folio_start = address - idx * PAGE_SIZE; |
| 4442 | unsigned long folio_end = folio_start + nr * PAGE_SIZE; |
| 4443 | pte_t *folio_ptep; |
| 4444 | pte_t folio_pte; |
| 4445 | |
| 4446 | if (unlikely(folio_start < max(address & PMD_MASK, vma->vm_start))) |
| 4447 | goto check_folio; |
| 4448 | if (unlikely(folio_end > pmd_addr_end(address, vma->vm_end))) |
| 4449 | goto check_folio; |
| 4450 | |
| 4451 | folio_ptep = vmf->pte - idx; |
| 4452 | folio_pte = ptep_get(folio_ptep); |
| 4453 | if (!pte_same(folio_pte, pte_move_swp_offset(vmf->orig_pte, -idx)) || |
| 4454 | swap_pte_batch(folio_ptep, nr, folio_pte) != nr) |
| 4455 | goto check_folio; |
| 4456 | |
| 4457 | page_idx = idx; |
| 4458 | address = folio_start; |
| 4459 | ptep = folio_ptep; |
| 4460 | nr_pages = nr; |
| 4461 | entry = folio->swap; |
| 4462 | page = &folio->page; |
| 4463 | } |
| 4464 | |
| 4465 | check_folio: |
| 4466 | /* |
| 4467 | * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte |
| 4468 | * must never point at an anonymous page in the swapcache that is |
| 4469 | * PG_anon_exclusive. Sanity check that this holds and especially, that |
| 4470 | * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity |
| 4471 | * check after taking the PT lock and making sure that nobody |
| 4472 | * concurrently faulted in this page and set PG_anon_exclusive. |
| 4473 | */ |
| 4474 | BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio)); |
| 4475 | BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page)); |
| 4476 | |
| 4477 | /* |
| 4478 | * Check under PT lock (to protect against concurrent fork() sharing |
| 4479 | * the swap entry concurrently) for certainly exclusive pages. |
| 4480 | */ |
| 4481 | if (!folio_test_ksm(folio)) { |
| 4482 | exclusive = pte_swp_exclusive(vmf->orig_pte); |
| 4483 | if (folio != swapcache) { |
| 4484 | /* |
| 4485 | * We have a fresh page that is not exposed to the |
| 4486 | * swapcache -> certainly exclusive. |
| 4487 | */ |
| 4488 | exclusive = true; |
| 4489 | } else if (exclusive && folio_test_writeback(folio) && |
| 4490 | data_race(si->flags & SWP_STABLE_WRITES)) { |
| 4491 | /* |
| 4492 | * This is tricky: not all swap backends support |
| 4493 | * concurrent page modifications while under writeback. |
| 4494 | * |
| 4495 | * So if we stumble over such a page in the swapcache |
| 4496 | * we must not set the page exclusive, otherwise we can |
| 4497 | * map it writable without further checks and modify it |
| 4498 | * while still under writeback. |
| 4499 | * |
| 4500 | * For these problematic swap backends, simply drop the |
| 4501 | * exclusive marker: this is perfectly fine as we start |
| 4502 | * writeback only if we fully unmapped the page and |
| 4503 | * there are no unexpected references on the page after |
| 4504 | * unmapping succeeded. After fully unmapped, no |
| 4505 | * further GUP references (FOLL_GET and FOLL_PIN) can |
| 4506 | * appear, so dropping the exclusive marker and mapping |
| 4507 | * it only R/O is fine. |
| 4508 | */ |
| 4509 | exclusive = false; |
| 4510 | } |
| 4511 | } |
| 4512 | |
| 4513 | /* |
| 4514 | * Some architectures may have to restore extra metadata to the page |
| 4515 | * when reading from swap. This metadata may be indexed by swap entry |
| 4516 | * so this must be called before swap_free(). |
| 4517 | */ |
| 4518 | arch_swap_restore(folio_swap(entry, folio), folio); |
| 4519 | |
| 4520 | /* |
| 4521 | * Remove the swap entry and conditionally try to free up the swapcache. |
| 4522 | * We're already holding a reference on the page but haven't mapped it |
| 4523 | * yet. |
| 4524 | */ |
| 4525 | swap_free_nr(entry, nr_pages); |
| 4526 | if (should_try_to_free_swap(folio, vma, vmf->flags)) |
| 4527 | folio_free_swap(folio); |
| 4528 | |
| 4529 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages); |
| 4530 | add_mm_counter(vma->vm_mm, MM_SWAPENTS, -nr_pages); |
| 4531 | pte = mk_pte(page, vma->vm_page_prot); |
| 4532 | if (pte_swp_soft_dirty(vmf->orig_pte)) |
| 4533 | pte = pte_mksoft_dirty(pte); |
| 4534 | if (pte_swp_uffd_wp(vmf->orig_pte)) |
| 4535 | pte = pte_mkuffd_wp(pte); |
| 4536 | |
| 4537 | /* |
| 4538 | * Same logic as in do_wp_page(); however, optimize for pages that are |
| 4539 | * certainly not shared either because we just allocated them without |
| 4540 | * exposing them to the swapcache or because the swap entry indicates |
| 4541 | * exclusivity. |
| 4542 | */ |
| 4543 | if (!folio_test_ksm(folio) && |
| 4544 | (exclusive || folio_ref_count(folio) == 1)) { |
| 4545 | if ((vma->vm_flags & VM_WRITE) && !userfaultfd_pte_wp(vma, pte) && |
| 4546 | !pte_needs_soft_dirty_wp(vma, pte)) { |
| 4547 | pte = pte_mkwrite(pte, vma); |
| 4548 | if (vmf->flags & FAULT_FLAG_WRITE) { |
| 4549 | pte = pte_mkdirty(pte); |
| 4550 | vmf->flags &= ~FAULT_FLAG_WRITE; |
| 4551 | } |
| 4552 | } |
| 4553 | rmap_flags |= RMAP_EXCLUSIVE; |
| 4554 | } |
| 4555 | folio_ref_add(folio, nr_pages - 1); |
| 4556 | flush_icache_pages(vma, page, nr_pages); |
| 4557 | vmf->orig_pte = pte_advance_pfn(pte, page_idx); |
| 4558 | |
| 4559 | /* ksm created a completely new copy */ |
| 4560 | if (unlikely(folio != swapcache && swapcache)) { |
| 4561 | folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE); |
| 4562 | folio_add_lru_vma(folio, vma); |
| 4563 | } else if (!folio_test_anon(folio)) { |
| 4564 | /* |
| 4565 | * We currently only expect small !anon folios which are either |
| 4566 | * fully exclusive or fully shared, or new allocated large |
| 4567 | * folios which are fully exclusive. If we ever get large |
| 4568 | * folios within swapcache here, we have to be careful. |
| 4569 | */ |
| 4570 | VM_WARN_ON_ONCE(folio_test_large(folio) && folio_test_swapcache(folio)); |
| 4571 | VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio); |
| 4572 | folio_add_new_anon_rmap(folio, vma, address, rmap_flags); |
| 4573 | } else { |
| 4574 | folio_add_anon_rmap_ptes(folio, page, nr_pages, vma, address, |
| 4575 | rmap_flags); |
| 4576 | } |
| 4577 | |
| 4578 | VM_BUG_ON(!folio_test_anon(folio) || |
| 4579 | (pte_write(pte) && !PageAnonExclusive(page))); |
| 4580 | set_ptes(vma->vm_mm, address, ptep, pte, nr_pages); |
| 4581 | arch_do_swap_page_nr(vma->vm_mm, vma, address, |
| 4582 | pte, pte, nr_pages); |
| 4583 | |
| 4584 | folio_unlock(folio); |
| 4585 | if (folio != swapcache && swapcache) { |
| 4586 | /* |
| 4587 | * Hold the lock to avoid the swap entry to be reused |
| 4588 | * until we take the PT lock for the pte_same() check |
| 4589 | * (to avoid false positives from pte_same). For |
| 4590 | * further safety release the lock after the swap_free |
| 4591 | * so that the swap count won't change under a |
| 4592 | * parallel locked swapcache. |
| 4593 | */ |
| 4594 | folio_unlock(swapcache); |
| 4595 | folio_put(swapcache); |
| 4596 | } |
| 4597 | |
| 4598 | if (vmf->flags & FAULT_FLAG_WRITE) { |
| 4599 | ret |= do_wp_page(vmf); |
| 4600 | if (ret & VM_FAULT_ERROR) |
| 4601 | ret &= VM_FAULT_ERROR; |
| 4602 | goto out; |
| 4603 | } |
| 4604 | |
| 4605 | /* No need to invalidate - it was non-present before */ |
| 4606 | update_mmu_cache_range(vmf, vma, address, ptep, nr_pages); |
| 4607 | unlock: |
| 4608 | if (vmf->pte) |
| 4609 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4610 | out: |
| 4611 | /* Clear the swap cache pin for direct swapin after PTL unlock */ |
| 4612 | if (need_clear_cache) |
| 4613 | swapcache_clear(si, entry, nr_pages); |
| 4614 | if (si) |
| 4615 | put_swap_device(si); |
| 4616 | return ret; |
| 4617 | out_nomap: |
| 4618 | if (vmf->pte) |
| 4619 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4620 | out_page: |
| 4621 | folio_unlock(folio); |
| 4622 | out_release: |
| 4623 | folio_put(folio); |
| 4624 | if (folio != swapcache && swapcache) { |
| 4625 | folio_unlock(swapcache); |
| 4626 | folio_put(swapcache); |
| 4627 | } |
| 4628 | if (need_clear_cache) |
| 4629 | swapcache_clear(si, entry, nr_pages); |
| 4630 | if (si) |
| 4631 | put_swap_device(si); |
| 4632 | return ret; |
| 4633 | } |
| 4634 | |
| 4635 | static bool pte_range_none(pte_t *pte, int nr_pages) |
| 4636 | { |
| 4637 | int i; |
| 4638 | |
| 4639 | for (i = 0; i < nr_pages; i++) { |
| 4640 | if (!pte_none(ptep_get_lockless(pte + i))) |
| 4641 | return false; |
| 4642 | } |
| 4643 | |
| 4644 | return true; |
| 4645 | } |
| 4646 | |
| 4647 | static struct folio *alloc_anon_folio(struct vm_fault *vmf) |
| 4648 | { |
| 4649 | struct vm_area_struct *vma = vmf->vma; |
| 4650 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 4651 | unsigned long orders; |
| 4652 | struct folio *folio; |
| 4653 | unsigned long addr; |
| 4654 | pte_t *pte; |
| 4655 | gfp_t gfp; |
| 4656 | int order; |
| 4657 | |
| 4658 | /* |
| 4659 | * If uffd is active for the vma we need per-page fault fidelity to |
| 4660 | * maintain the uffd semantics. |
| 4661 | */ |
| 4662 | if (unlikely(userfaultfd_armed(vma))) |
| 4663 | goto fallback; |
| 4664 | |
| 4665 | /* |
| 4666 | * Get a list of all the (large) orders below PMD_ORDER that are enabled |
| 4667 | * for this vma. Then filter out the orders that can't be allocated over |
| 4668 | * the faulting address and still be fully contained in the vma. |
| 4669 | */ |
| 4670 | orders = thp_vma_allowable_orders(vma, vma->vm_flags, |
| 4671 | TVA_IN_PF | TVA_ENFORCE_SYSFS, BIT(PMD_ORDER) - 1); |
| 4672 | orders = thp_vma_suitable_orders(vma, vmf->address, orders); |
| 4673 | |
| 4674 | if (!orders) |
| 4675 | goto fallback; |
| 4676 | |
| 4677 | pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK); |
| 4678 | if (!pte) |
| 4679 | return ERR_PTR(-EAGAIN); |
| 4680 | |
| 4681 | /* |
| 4682 | * Find the highest order where the aligned range is completely |
| 4683 | * pte_none(). Note that all remaining orders will be completely |
| 4684 | * pte_none(). |
| 4685 | */ |
| 4686 | order = highest_order(orders); |
| 4687 | while (orders) { |
| 4688 | addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); |
| 4689 | if (pte_range_none(pte + pte_index(addr), 1 << order)) |
| 4690 | break; |
| 4691 | order = next_order(&orders, order); |
| 4692 | } |
| 4693 | |
| 4694 | pte_unmap(pte); |
| 4695 | |
| 4696 | if (!orders) |
| 4697 | goto fallback; |
| 4698 | |
| 4699 | /* Try allocating the highest of the remaining orders. */ |
| 4700 | gfp = vma_thp_gfp_mask(vma); |
| 4701 | while (orders) { |
| 4702 | addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); |
| 4703 | folio = vma_alloc_folio(gfp, order, vma, addr, true); |
| 4704 | if (folio) { |
| 4705 | if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { |
| 4706 | count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE); |
| 4707 | folio_put(folio); |
| 4708 | goto next; |
| 4709 | } |
| 4710 | folio_throttle_swaprate(folio, gfp); |
| 4711 | folio_zero_user(folio, vmf->address); |
| 4712 | return folio; |
| 4713 | } |
| 4714 | next: |
| 4715 | count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK); |
| 4716 | order = next_order(&orders, order); |
| 4717 | } |
| 4718 | |
| 4719 | fallback: |
| 4720 | #endif |
| 4721 | return folio_prealloc(vma->vm_mm, vma, vmf->address, true); |
| 4722 | } |
| 4723 | |
| 4724 | /* |
| 4725 | * We enter with non-exclusive mmap_lock (to exclude vma changes, |
| 4726 | * but allow concurrent faults), and pte mapped but not yet locked. |
| 4727 | * We return with mmap_lock still held, but pte unmapped and unlocked. |
| 4728 | */ |
| 4729 | static vm_fault_t do_anonymous_page(struct vm_fault *vmf) |
| 4730 | { |
| 4731 | struct vm_area_struct *vma = vmf->vma; |
| 4732 | unsigned long addr = vmf->address; |
| 4733 | struct folio *folio; |
| 4734 | vm_fault_t ret = 0; |
| 4735 | int nr_pages = 1; |
| 4736 | pte_t entry; |
| 4737 | |
| 4738 | /* File mapping without ->vm_ops ? */ |
| 4739 | if (vma->vm_flags & VM_SHARED) |
| 4740 | return VM_FAULT_SIGBUS; |
| 4741 | |
| 4742 | /* |
| 4743 | * Use pte_alloc() instead of pte_alloc_map(), so that OOM can |
| 4744 | * be distinguished from a transient failure of pte_offset_map(). |
| 4745 | */ |
| 4746 | if (pte_alloc(vma->vm_mm, vmf->pmd)) |
| 4747 | return VM_FAULT_OOM; |
| 4748 | |
| 4749 | /* Use the zero-page for reads */ |
| 4750 | if (!(vmf->flags & FAULT_FLAG_WRITE) && |
| 4751 | !mm_forbids_zeropage(vma->vm_mm)) { |
| 4752 | entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), |
| 4753 | vma->vm_page_prot)); |
| 4754 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, |
| 4755 | vmf->address, &vmf->ptl); |
| 4756 | if (!vmf->pte) |
| 4757 | goto unlock; |
| 4758 | if (vmf_pte_changed(vmf)) { |
| 4759 | update_mmu_tlb(vma, vmf->address, vmf->pte); |
| 4760 | goto unlock; |
| 4761 | } |
| 4762 | ret = check_stable_address_space(vma->vm_mm); |
| 4763 | if (ret) |
| 4764 | goto unlock; |
| 4765 | /* Deliver the page fault to userland, check inside PT lock */ |
| 4766 | if (userfaultfd_missing(vma)) { |
| 4767 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4768 | return handle_userfault(vmf, VM_UFFD_MISSING); |
| 4769 | } |
| 4770 | goto setpte; |
| 4771 | } |
| 4772 | |
| 4773 | /* Allocate our own private page. */ |
| 4774 | ret = vmf_anon_prepare(vmf); |
| 4775 | if (ret) |
| 4776 | return ret; |
| 4777 | /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */ |
| 4778 | folio = alloc_anon_folio(vmf); |
| 4779 | if (IS_ERR(folio)) |
| 4780 | return 0; |
| 4781 | if (!folio) |
| 4782 | goto oom; |
| 4783 | |
| 4784 | nr_pages = folio_nr_pages(folio); |
| 4785 | addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE); |
| 4786 | |
| 4787 | /* |
| 4788 | * The memory barrier inside __folio_mark_uptodate makes sure that |
| 4789 | * preceding stores to the page contents become visible before |
| 4790 | * the set_pte_at() write. |
| 4791 | */ |
| 4792 | __folio_mark_uptodate(folio); |
| 4793 | |
| 4794 | entry = mk_pte(&folio->page, vma->vm_page_prot); |
| 4795 | entry = pte_sw_mkyoung(entry); |
| 4796 | if (vma->vm_flags & VM_WRITE) |
| 4797 | entry = pte_mkwrite(pte_mkdirty(entry), vma); |
| 4798 | |
| 4799 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl); |
| 4800 | if (!vmf->pte) |
| 4801 | goto release; |
| 4802 | if (nr_pages == 1 && vmf_pte_changed(vmf)) { |
| 4803 | update_mmu_tlb(vma, addr, vmf->pte); |
| 4804 | goto release; |
| 4805 | } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) { |
| 4806 | update_mmu_tlb_range(vma, addr, vmf->pte, nr_pages); |
| 4807 | goto release; |
| 4808 | } |
| 4809 | |
| 4810 | ret = check_stable_address_space(vma->vm_mm); |
| 4811 | if (ret) |
| 4812 | goto release; |
| 4813 | |
| 4814 | /* Deliver the page fault to userland, check inside PT lock */ |
| 4815 | if (userfaultfd_missing(vma)) { |
| 4816 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4817 | folio_put(folio); |
| 4818 | return handle_userfault(vmf, VM_UFFD_MISSING); |
| 4819 | } |
| 4820 | |
| 4821 | folio_ref_add(folio, nr_pages - 1); |
| 4822 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages); |
| 4823 | count_mthp_stat(folio_order(folio), MTHP_STAT_ANON_FAULT_ALLOC); |
| 4824 | folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE); |
| 4825 | folio_add_lru_vma(folio, vma); |
| 4826 | setpte: |
| 4827 | if (vmf_orig_pte_uffd_wp(vmf)) |
| 4828 | entry = pte_mkuffd_wp(entry); |
| 4829 | set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages); |
| 4830 | |
| 4831 | /* No need to invalidate - it was non-present before */ |
| 4832 | update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages); |
| 4833 | unlock: |
| 4834 | if (vmf->pte) |
| 4835 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 4836 | return ret; |
| 4837 | release: |
| 4838 | folio_put(folio); |
| 4839 | goto unlock; |
| 4840 | oom: |
| 4841 | return VM_FAULT_OOM; |
| 4842 | } |
| 4843 | |
| 4844 | /* |
| 4845 | * The mmap_lock must have been held on entry, and may have been |
| 4846 | * released depending on flags and vma->vm_ops->fault() return value. |
| 4847 | * See filemap_fault() and __lock_page_retry(). |
| 4848 | */ |
| 4849 | static vm_fault_t __do_fault(struct vm_fault *vmf) |
| 4850 | { |
| 4851 | struct vm_area_struct *vma = vmf->vma; |
| 4852 | struct folio *folio; |
| 4853 | vm_fault_t ret; |
| 4854 | |
| 4855 | /* |
| 4856 | * Preallocate pte before we take page_lock because this might lead to |
| 4857 | * deadlocks for memcg reclaim which waits for pages under writeback: |
| 4858 | * lock_page(A) |
| 4859 | * SetPageWriteback(A) |
| 4860 | * unlock_page(A) |
| 4861 | * lock_page(B) |
| 4862 | * lock_page(B) |
| 4863 | * pte_alloc_one |
| 4864 | * shrink_folio_list |
| 4865 | * wait_on_page_writeback(A) |
| 4866 | * SetPageWriteback(B) |
| 4867 | * unlock_page(B) |
| 4868 | * # flush A, B to clear the writeback |
| 4869 | */ |
| 4870 | if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { |
| 4871 | vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); |
| 4872 | if (!vmf->prealloc_pte) |
| 4873 | return VM_FAULT_OOM; |
| 4874 | } |
| 4875 | |
| 4876 | ret = vma->vm_ops->fault(vmf); |
| 4877 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | |
| 4878 | VM_FAULT_DONE_COW))) |
| 4879 | return ret; |
| 4880 | |
| 4881 | folio = page_folio(vmf->page); |
| 4882 | if (unlikely(PageHWPoison(vmf->page))) { |
| 4883 | vm_fault_t poisonret = VM_FAULT_HWPOISON; |
| 4884 | if (ret & VM_FAULT_LOCKED) { |
| 4885 | if (page_mapped(vmf->page)) |
| 4886 | unmap_mapping_folio(folio); |
| 4887 | /* Retry if a clean folio was removed from the cache. */ |
| 4888 | if (mapping_evict_folio(folio->mapping, folio)) |
| 4889 | poisonret = VM_FAULT_NOPAGE; |
| 4890 | folio_unlock(folio); |
| 4891 | } |
| 4892 | folio_put(folio); |
| 4893 | vmf->page = NULL; |
| 4894 | return poisonret; |
| 4895 | } |
| 4896 | |
| 4897 | if (unlikely(!(ret & VM_FAULT_LOCKED))) |
| 4898 | folio_lock(folio); |
| 4899 | else |
| 4900 | VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page); |
| 4901 | |
| 4902 | return ret; |
| 4903 | } |
| 4904 | |
| 4905 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 4906 | static void deposit_prealloc_pte(struct vm_fault *vmf) |
| 4907 | { |
| 4908 | struct vm_area_struct *vma = vmf->vma; |
| 4909 | |
| 4910 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); |
| 4911 | /* |
| 4912 | * We are going to consume the prealloc table, |
| 4913 | * count that as nr_ptes. |
| 4914 | */ |
| 4915 | mm_inc_nr_ptes(vma->vm_mm); |
| 4916 | vmf->prealloc_pte = NULL; |
| 4917 | } |
| 4918 | |
| 4919 | vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) |
| 4920 | { |
| 4921 | struct folio *folio = page_folio(page); |
| 4922 | struct vm_area_struct *vma = vmf->vma; |
| 4923 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
| 4924 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
| 4925 | pmd_t entry; |
| 4926 | vm_fault_t ret = VM_FAULT_FALLBACK; |
| 4927 | |
| 4928 | if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER)) |
| 4929 | return ret; |
| 4930 | |
| 4931 | if (folio_order(folio) != HPAGE_PMD_ORDER) |
| 4932 | return ret; |
| 4933 | page = &folio->page; |
| 4934 | |
| 4935 | /* |
| 4936 | * Just backoff if any subpage of a THP is corrupted otherwise |
| 4937 | * the corrupted page may mapped by PMD silently to escape the |
| 4938 | * check. This kind of THP just can be PTE mapped. Access to |
| 4939 | * the corrupted subpage should trigger SIGBUS as expected. |
| 4940 | */ |
| 4941 | if (unlikely(folio_test_has_hwpoisoned(folio))) |
| 4942 | return ret; |
| 4943 | |
| 4944 | /* |
| 4945 | * Archs like ppc64 need additional space to store information |
| 4946 | * related to pte entry. Use the preallocated table for that. |
| 4947 | */ |
| 4948 | if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { |
| 4949 | vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); |
| 4950 | if (!vmf->prealloc_pte) |
| 4951 | return VM_FAULT_OOM; |
| 4952 | } |
| 4953 | |
| 4954 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
| 4955 | if (unlikely(!pmd_none(*vmf->pmd))) |
| 4956 | goto out; |
| 4957 | |
| 4958 | flush_icache_pages(vma, page, HPAGE_PMD_NR); |
| 4959 | |
| 4960 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
| 4961 | if (write) |
| 4962 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
| 4963 | |
| 4964 | add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR); |
| 4965 | folio_add_file_rmap_pmd(folio, page, vma); |
| 4966 | |
| 4967 | /* |
| 4968 | * deposit and withdraw with pmd lock held |
| 4969 | */ |
| 4970 | if (arch_needs_pgtable_deposit()) |
| 4971 | deposit_prealloc_pte(vmf); |
| 4972 | |
| 4973 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); |
| 4974 | |
| 4975 | update_mmu_cache_pmd(vma, haddr, vmf->pmd); |
| 4976 | |
| 4977 | /* fault is handled */ |
| 4978 | ret = 0; |
| 4979 | count_vm_event(THP_FILE_MAPPED); |
| 4980 | out: |
| 4981 | spin_unlock(vmf->ptl); |
| 4982 | return ret; |
| 4983 | } |
| 4984 | #else |
| 4985 | vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) |
| 4986 | { |
| 4987 | return VM_FAULT_FALLBACK; |
| 4988 | } |
| 4989 | #endif |
| 4990 | |
| 4991 | /** |
| 4992 | * set_pte_range - Set a range of PTEs to point to pages in a folio. |
| 4993 | * @vmf: Fault decription. |
| 4994 | * @folio: The folio that contains @page. |
| 4995 | * @page: The first page to create a PTE for. |
| 4996 | * @nr: The number of PTEs to create. |
| 4997 | * @addr: The first address to create a PTE for. |
| 4998 | */ |
| 4999 | void set_pte_range(struct vm_fault *vmf, struct folio *folio, |
| 5000 | struct page *page, unsigned int nr, unsigned long addr) |
| 5001 | { |
| 5002 | struct vm_area_struct *vma = vmf->vma; |
| 5003 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
| 5004 | bool prefault = !in_range(vmf->address, addr, nr * PAGE_SIZE); |
| 5005 | pte_t entry; |
| 5006 | |
| 5007 | flush_icache_pages(vma, page, nr); |
| 5008 | entry = mk_pte(page, vma->vm_page_prot); |
| 5009 | |
| 5010 | if (prefault && arch_wants_old_prefaulted_pte()) |
| 5011 | entry = pte_mkold(entry); |
| 5012 | else |
| 5013 | entry = pte_sw_mkyoung(entry); |
| 5014 | |
| 5015 | if (write) |
| 5016 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
| 5017 | if (unlikely(vmf_orig_pte_uffd_wp(vmf))) |
| 5018 | entry = pte_mkuffd_wp(entry); |
| 5019 | /* copy-on-write page */ |
| 5020 | if (write && !(vma->vm_flags & VM_SHARED)) { |
| 5021 | VM_BUG_ON_FOLIO(nr != 1, folio); |
| 5022 | folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE); |
| 5023 | folio_add_lru_vma(folio, vma); |
| 5024 | } else { |
| 5025 | folio_add_file_rmap_ptes(folio, page, nr, vma); |
| 5026 | } |
| 5027 | set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr); |
| 5028 | |
| 5029 | /* no need to invalidate: a not-present page won't be cached */ |
| 5030 | update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr); |
| 5031 | } |
| 5032 | |
| 5033 | static bool vmf_pte_changed(struct vm_fault *vmf) |
| 5034 | { |
| 5035 | if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID) |
| 5036 | return !pte_same(ptep_get(vmf->pte), vmf->orig_pte); |
| 5037 | |
| 5038 | return !pte_none(ptep_get(vmf->pte)); |
| 5039 | } |
| 5040 | |
| 5041 | /** |
| 5042 | * finish_fault - finish page fault once we have prepared the page to fault |
| 5043 | * |
| 5044 | * @vmf: structure describing the fault |
| 5045 | * |
| 5046 | * This function handles all that is needed to finish a page fault once the |
| 5047 | * page to fault in is prepared. It handles locking of PTEs, inserts PTE for |
| 5048 | * given page, adds reverse page mapping, handles memcg charges and LRU |
| 5049 | * addition. |
| 5050 | * |
| 5051 | * The function expects the page to be locked and on success it consumes a |
| 5052 | * reference of a page being mapped (for the PTE which maps it). |
| 5053 | * |
| 5054 | * Return: %0 on success, %VM_FAULT_ code in case of error. |
| 5055 | */ |
| 5056 | vm_fault_t finish_fault(struct vm_fault *vmf) |
| 5057 | { |
| 5058 | struct vm_area_struct *vma = vmf->vma; |
| 5059 | struct page *page; |
| 5060 | struct folio *folio; |
| 5061 | vm_fault_t ret; |
| 5062 | bool is_cow = (vmf->flags & FAULT_FLAG_WRITE) && |
| 5063 | !(vma->vm_flags & VM_SHARED); |
| 5064 | int type, nr_pages; |
| 5065 | unsigned long addr = vmf->address; |
| 5066 | |
| 5067 | /* Did we COW the page? */ |
| 5068 | if (is_cow) |
| 5069 | page = vmf->cow_page; |
| 5070 | else |
| 5071 | page = vmf->page; |
| 5072 | |
| 5073 | /* |
| 5074 | * check even for read faults because we might have lost our CoWed |
| 5075 | * page |
| 5076 | */ |
| 5077 | if (!(vma->vm_flags & VM_SHARED)) { |
| 5078 | ret = check_stable_address_space(vma->vm_mm); |
| 5079 | if (ret) |
| 5080 | return ret; |
| 5081 | } |
| 5082 | |
| 5083 | if (pmd_none(*vmf->pmd)) { |
| 5084 | if (PageTransCompound(page)) { |
| 5085 | ret = do_set_pmd(vmf, page); |
| 5086 | if (ret != VM_FAULT_FALLBACK) |
| 5087 | return ret; |
| 5088 | } |
| 5089 | |
| 5090 | if (vmf->prealloc_pte) |
| 5091 | pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte); |
| 5092 | else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) |
| 5093 | return VM_FAULT_OOM; |
| 5094 | } |
| 5095 | |
| 5096 | folio = page_folio(page); |
| 5097 | nr_pages = folio_nr_pages(folio); |
| 5098 | |
| 5099 | /* |
| 5100 | * Using per-page fault to maintain the uffd semantics, and same |
| 5101 | * approach also applies to non-anonymous-shmem faults to avoid |
| 5102 | * inflating the RSS of the process. |
| 5103 | */ |
| 5104 | if (!vma_is_anon_shmem(vma) || unlikely(userfaultfd_armed(vma))) { |
| 5105 | nr_pages = 1; |
| 5106 | } else if (nr_pages > 1) { |
| 5107 | pgoff_t idx = folio_page_idx(folio, page); |
| 5108 | /* The page offset of vmf->address within the VMA. */ |
| 5109 | pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff; |
| 5110 | /* The index of the entry in the pagetable for fault page. */ |
| 5111 | pgoff_t pte_off = pte_index(vmf->address); |
| 5112 | |
| 5113 | /* |
| 5114 | * Fallback to per-page fault in case the folio size in page |
| 5115 | * cache beyond the VMA limits and PMD pagetable limits. |
| 5116 | */ |
| 5117 | if (unlikely(vma_off < idx || |
| 5118 | vma_off + (nr_pages - idx) > vma_pages(vma) || |
| 5119 | pte_off < idx || |
| 5120 | pte_off + (nr_pages - idx) > PTRS_PER_PTE)) { |
| 5121 | nr_pages = 1; |
| 5122 | } else { |
| 5123 | /* Now we can set mappings for the whole large folio. */ |
| 5124 | addr = vmf->address - idx * PAGE_SIZE; |
| 5125 | page = &folio->page; |
| 5126 | } |
| 5127 | } |
| 5128 | |
| 5129 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, |
| 5130 | addr, &vmf->ptl); |
| 5131 | if (!vmf->pte) |
| 5132 | return VM_FAULT_NOPAGE; |
| 5133 | |
| 5134 | /* Re-check under ptl */ |
| 5135 | if (nr_pages == 1 && unlikely(vmf_pte_changed(vmf))) { |
| 5136 | update_mmu_tlb(vma, addr, vmf->pte); |
| 5137 | ret = VM_FAULT_NOPAGE; |
| 5138 | goto unlock; |
| 5139 | } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) { |
| 5140 | update_mmu_tlb_range(vma, addr, vmf->pte, nr_pages); |
| 5141 | ret = VM_FAULT_NOPAGE; |
| 5142 | goto unlock; |
| 5143 | } |
| 5144 | |
| 5145 | folio_ref_add(folio, nr_pages - 1); |
| 5146 | set_pte_range(vmf, folio, page, nr_pages, addr); |
| 5147 | type = is_cow ? MM_ANONPAGES : mm_counter_file(folio); |
| 5148 | add_mm_counter(vma->vm_mm, type, nr_pages); |
| 5149 | ret = 0; |
| 5150 | |
| 5151 | unlock: |
| 5152 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5153 | return ret; |
| 5154 | } |
| 5155 | |
| 5156 | static unsigned long fault_around_pages __read_mostly = |
| 5157 | 65536 >> PAGE_SHIFT; |
| 5158 | |
| 5159 | #ifdef CONFIG_DEBUG_FS |
| 5160 | static int fault_around_bytes_get(void *data, u64 *val) |
| 5161 | { |
| 5162 | *val = fault_around_pages << PAGE_SHIFT; |
| 5163 | return 0; |
| 5164 | } |
| 5165 | |
| 5166 | /* |
| 5167 | * fault_around_bytes must be rounded down to the nearest page order as it's |
| 5168 | * what do_fault_around() expects to see. |
| 5169 | */ |
| 5170 | static int fault_around_bytes_set(void *data, u64 val) |
| 5171 | { |
| 5172 | if (val / PAGE_SIZE > PTRS_PER_PTE) |
| 5173 | return -EINVAL; |
| 5174 | |
| 5175 | /* |
| 5176 | * The minimum value is 1 page, however this results in no fault-around |
| 5177 | * at all. See should_fault_around(). |
| 5178 | */ |
| 5179 | val = max(val, PAGE_SIZE); |
| 5180 | fault_around_pages = rounddown_pow_of_two(val) >> PAGE_SHIFT; |
| 5181 | |
| 5182 | return 0; |
| 5183 | } |
| 5184 | DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, |
| 5185 | fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); |
| 5186 | |
| 5187 | static int __init fault_around_debugfs(void) |
| 5188 | { |
| 5189 | debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, |
| 5190 | &fault_around_bytes_fops); |
| 5191 | return 0; |
| 5192 | } |
| 5193 | late_initcall(fault_around_debugfs); |
| 5194 | #endif |
| 5195 | |
| 5196 | /* |
| 5197 | * do_fault_around() tries to map few pages around the fault address. The hope |
| 5198 | * is that the pages will be needed soon and this will lower the number of |
| 5199 | * faults to handle. |
| 5200 | * |
| 5201 | * It uses vm_ops->map_pages() to map the pages, which skips the page if it's |
| 5202 | * not ready to be mapped: not up-to-date, locked, etc. |
| 5203 | * |
| 5204 | * This function doesn't cross VMA or page table boundaries, in order to call |
| 5205 | * map_pages() and acquire a PTE lock only once. |
| 5206 | * |
| 5207 | * fault_around_pages defines how many pages we'll try to map. |
| 5208 | * do_fault_around() expects it to be set to a power of two less than or equal |
| 5209 | * to PTRS_PER_PTE. |
| 5210 | * |
| 5211 | * The virtual address of the area that we map is naturally aligned to |
| 5212 | * fault_around_pages * PAGE_SIZE rounded down to the machine page size |
| 5213 | * (and therefore to page order). This way it's easier to guarantee |
| 5214 | * that we don't cross page table boundaries. |
| 5215 | */ |
| 5216 | static vm_fault_t do_fault_around(struct vm_fault *vmf) |
| 5217 | { |
| 5218 | pgoff_t nr_pages = READ_ONCE(fault_around_pages); |
| 5219 | pgoff_t pte_off = pte_index(vmf->address); |
| 5220 | /* The page offset of vmf->address within the VMA. */ |
| 5221 | pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff; |
| 5222 | pgoff_t from_pte, to_pte; |
| 5223 | vm_fault_t ret; |
| 5224 | |
| 5225 | /* The PTE offset of the start address, clamped to the VMA. */ |
| 5226 | from_pte = max(ALIGN_DOWN(pte_off, nr_pages), |
| 5227 | pte_off - min(pte_off, vma_off)); |
| 5228 | |
| 5229 | /* The PTE offset of the end address, clamped to the VMA and PTE. */ |
| 5230 | to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE, |
| 5231 | pte_off + vma_pages(vmf->vma) - vma_off) - 1; |
| 5232 | |
| 5233 | if (pmd_none(*vmf->pmd)) { |
| 5234 | vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); |
| 5235 | if (!vmf->prealloc_pte) |
| 5236 | return VM_FAULT_OOM; |
| 5237 | } |
| 5238 | |
| 5239 | rcu_read_lock(); |
| 5240 | ret = vmf->vma->vm_ops->map_pages(vmf, |
| 5241 | vmf->pgoff + from_pte - pte_off, |
| 5242 | vmf->pgoff + to_pte - pte_off); |
| 5243 | rcu_read_unlock(); |
| 5244 | |
| 5245 | return ret; |
| 5246 | } |
| 5247 | |
| 5248 | /* Return true if we should do read fault-around, false otherwise */ |
| 5249 | static inline bool should_fault_around(struct vm_fault *vmf) |
| 5250 | { |
| 5251 | /* No ->map_pages? No way to fault around... */ |
| 5252 | if (!vmf->vma->vm_ops->map_pages) |
| 5253 | return false; |
| 5254 | |
| 5255 | if (uffd_disable_fault_around(vmf->vma)) |
| 5256 | return false; |
| 5257 | |
| 5258 | /* A single page implies no faulting 'around' at all. */ |
| 5259 | return fault_around_pages > 1; |
| 5260 | } |
| 5261 | |
| 5262 | static vm_fault_t do_read_fault(struct vm_fault *vmf) |
| 5263 | { |
| 5264 | vm_fault_t ret = 0; |
| 5265 | struct folio *folio; |
| 5266 | |
| 5267 | /* |
| 5268 | * Let's call ->map_pages() first and use ->fault() as fallback |
| 5269 | * if page by the offset is not ready to be mapped (cold cache or |
| 5270 | * something). |
| 5271 | */ |
| 5272 | if (should_fault_around(vmf)) { |
| 5273 | ret = do_fault_around(vmf); |
| 5274 | if (ret) |
| 5275 | return ret; |
| 5276 | } |
| 5277 | |
| 5278 | ret = vmf_can_call_fault(vmf); |
| 5279 | if (ret) |
| 5280 | return ret; |
| 5281 | |
| 5282 | ret = __do_fault(vmf); |
| 5283 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
| 5284 | return ret; |
| 5285 | |
| 5286 | ret |= finish_fault(vmf); |
| 5287 | folio = page_folio(vmf->page); |
| 5288 | folio_unlock(folio); |
| 5289 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
| 5290 | folio_put(folio); |
| 5291 | return ret; |
| 5292 | } |
| 5293 | |
| 5294 | static vm_fault_t do_cow_fault(struct vm_fault *vmf) |
| 5295 | { |
| 5296 | struct vm_area_struct *vma = vmf->vma; |
| 5297 | struct folio *folio; |
| 5298 | vm_fault_t ret; |
| 5299 | |
| 5300 | ret = vmf_can_call_fault(vmf); |
| 5301 | if (!ret) |
| 5302 | ret = vmf_anon_prepare(vmf); |
| 5303 | if (ret) |
| 5304 | return ret; |
| 5305 | |
| 5306 | folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false); |
| 5307 | if (!folio) |
| 5308 | return VM_FAULT_OOM; |
| 5309 | |
| 5310 | vmf->cow_page = &folio->page; |
| 5311 | |
| 5312 | ret = __do_fault(vmf); |
| 5313 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
| 5314 | goto uncharge_out; |
| 5315 | if (ret & VM_FAULT_DONE_COW) |
| 5316 | return ret; |
| 5317 | |
| 5318 | if (copy_mc_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma)) { |
| 5319 | ret = VM_FAULT_HWPOISON; |
| 5320 | goto unlock; |
| 5321 | } |
| 5322 | __folio_mark_uptodate(folio); |
| 5323 | |
| 5324 | ret |= finish_fault(vmf); |
| 5325 | unlock: |
| 5326 | unlock_page(vmf->page); |
| 5327 | put_page(vmf->page); |
| 5328 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
| 5329 | goto uncharge_out; |
| 5330 | return ret; |
| 5331 | uncharge_out: |
| 5332 | folio_put(folio); |
| 5333 | return ret; |
| 5334 | } |
| 5335 | |
| 5336 | static vm_fault_t do_shared_fault(struct vm_fault *vmf) |
| 5337 | { |
| 5338 | struct vm_area_struct *vma = vmf->vma; |
| 5339 | vm_fault_t ret, tmp; |
| 5340 | struct folio *folio; |
| 5341 | |
| 5342 | ret = vmf_can_call_fault(vmf); |
| 5343 | if (ret) |
| 5344 | return ret; |
| 5345 | |
| 5346 | ret = __do_fault(vmf); |
| 5347 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
| 5348 | return ret; |
| 5349 | |
| 5350 | folio = page_folio(vmf->page); |
| 5351 | |
| 5352 | /* |
| 5353 | * Check if the backing address space wants to know that the page is |
| 5354 | * about to become writable |
| 5355 | */ |
| 5356 | if (vma->vm_ops->page_mkwrite) { |
| 5357 | folio_unlock(folio); |
| 5358 | tmp = do_page_mkwrite(vmf, folio); |
| 5359 | if (unlikely(!tmp || |
| 5360 | (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { |
| 5361 | folio_put(folio); |
| 5362 | return tmp; |
| 5363 | } |
| 5364 | } |
| 5365 | |
| 5366 | ret |= finish_fault(vmf); |
| 5367 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | |
| 5368 | VM_FAULT_RETRY))) { |
| 5369 | folio_unlock(folio); |
| 5370 | folio_put(folio); |
| 5371 | return ret; |
| 5372 | } |
| 5373 | |
| 5374 | ret |= fault_dirty_shared_page(vmf); |
| 5375 | return ret; |
| 5376 | } |
| 5377 | |
| 5378 | /* |
| 5379 | * We enter with non-exclusive mmap_lock (to exclude vma changes, |
| 5380 | * but allow concurrent faults). |
| 5381 | * The mmap_lock may have been released depending on flags and our |
| 5382 | * return value. See filemap_fault() and __folio_lock_or_retry(). |
| 5383 | * If mmap_lock is released, vma may become invalid (for example |
| 5384 | * by other thread calling munmap()). |
| 5385 | */ |
| 5386 | static vm_fault_t do_fault(struct vm_fault *vmf) |
| 5387 | { |
| 5388 | struct vm_area_struct *vma = vmf->vma; |
| 5389 | struct mm_struct *vm_mm = vma->vm_mm; |
| 5390 | vm_fault_t ret; |
| 5391 | |
| 5392 | /* |
| 5393 | * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND |
| 5394 | */ |
| 5395 | if (!vma->vm_ops->fault) { |
| 5396 | vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, |
| 5397 | vmf->address, &vmf->ptl); |
| 5398 | if (unlikely(!vmf->pte)) |
| 5399 | ret = VM_FAULT_SIGBUS; |
| 5400 | else { |
| 5401 | /* |
| 5402 | * Make sure this is not a temporary clearing of pte |
| 5403 | * by holding ptl and checking again. A R/M/W update |
| 5404 | * of pte involves: take ptl, clearing the pte so that |
| 5405 | * we don't have concurrent modification by hardware |
| 5406 | * followed by an update. |
| 5407 | */ |
| 5408 | if (unlikely(pte_none(ptep_get(vmf->pte)))) |
| 5409 | ret = VM_FAULT_SIGBUS; |
| 5410 | else |
| 5411 | ret = VM_FAULT_NOPAGE; |
| 5412 | |
| 5413 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5414 | } |
| 5415 | } else if (!(vmf->flags & FAULT_FLAG_WRITE)) |
| 5416 | ret = do_read_fault(vmf); |
| 5417 | else if (!(vma->vm_flags & VM_SHARED)) |
| 5418 | ret = do_cow_fault(vmf); |
| 5419 | else |
| 5420 | ret = do_shared_fault(vmf); |
| 5421 | |
| 5422 | /* preallocated pagetable is unused: free it */ |
| 5423 | if (vmf->prealloc_pte) { |
| 5424 | pte_free(vm_mm, vmf->prealloc_pte); |
| 5425 | vmf->prealloc_pte = NULL; |
| 5426 | } |
| 5427 | return ret; |
| 5428 | } |
| 5429 | |
| 5430 | int numa_migrate_check(struct folio *folio, struct vm_fault *vmf, |
| 5431 | unsigned long addr, int *flags, |
| 5432 | bool writable, int *last_cpupid) |
| 5433 | { |
| 5434 | struct vm_area_struct *vma = vmf->vma; |
| 5435 | |
| 5436 | /* |
| 5437 | * Avoid grouping on RO pages in general. RO pages shouldn't hurt as |
| 5438 | * much anyway since they can be in shared cache state. This misses |
| 5439 | * the case where a mapping is writable but the process never writes |
| 5440 | * to it but pte_write gets cleared during protection updates and |
| 5441 | * pte_dirty has unpredictable behaviour between PTE scan updates, |
| 5442 | * background writeback, dirty balancing and application behaviour. |
| 5443 | */ |
| 5444 | if (!writable) |
| 5445 | *flags |= TNF_NO_GROUP; |
| 5446 | |
| 5447 | /* |
| 5448 | * Flag if the folio is shared between multiple address spaces. This |
| 5449 | * is later used when determining whether to group tasks together |
| 5450 | */ |
| 5451 | if (folio_likely_mapped_shared(folio) && (vma->vm_flags & VM_SHARED)) |
| 5452 | *flags |= TNF_SHARED; |
| 5453 | /* |
| 5454 | * For memory tiering mode, cpupid of slow memory page is used |
| 5455 | * to record page access time. So use default value. |
| 5456 | */ |
| 5457 | if (folio_use_access_time(folio)) |
| 5458 | *last_cpupid = (-1 & LAST_CPUPID_MASK); |
| 5459 | else |
| 5460 | *last_cpupid = folio_last_cpupid(folio); |
| 5461 | |
| 5462 | /* Record the current PID acceesing VMA */ |
| 5463 | vma_set_access_pid_bit(vma); |
| 5464 | |
| 5465 | count_vm_numa_event(NUMA_HINT_FAULTS); |
| 5466 | #ifdef CONFIG_NUMA_BALANCING |
| 5467 | count_memcg_folio_events(folio, NUMA_HINT_FAULTS, 1); |
| 5468 | #endif |
| 5469 | if (folio_nid(folio) == numa_node_id()) { |
| 5470 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
| 5471 | *flags |= TNF_FAULT_LOCAL; |
| 5472 | } |
| 5473 | |
| 5474 | return mpol_misplaced(folio, vmf, addr); |
| 5475 | } |
| 5476 | |
| 5477 | static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_struct *vma, |
| 5478 | unsigned long fault_addr, pte_t *fault_pte, |
| 5479 | bool writable) |
| 5480 | { |
| 5481 | pte_t pte, old_pte; |
| 5482 | |
| 5483 | old_pte = ptep_modify_prot_start(vma, fault_addr, fault_pte); |
| 5484 | pte = pte_modify(old_pte, vma->vm_page_prot); |
| 5485 | pte = pte_mkyoung(pte); |
| 5486 | if (writable) |
| 5487 | pte = pte_mkwrite(pte, vma); |
| 5488 | ptep_modify_prot_commit(vma, fault_addr, fault_pte, old_pte, pte); |
| 5489 | update_mmu_cache_range(vmf, vma, fault_addr, fault_pte, 1); |
| 5490 | } |
| 5491 | |
| 5492 | static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma, |
| 5493 | struct folio *folio, pte_t fault_pte, |
| 5494 | bool ignore_writable, bool pte_write_upgrade) |
| 5495 | { |
| 5496 | int nr = pte_pfn(fault_pte) - folio_pfn(folio); |
| 5497 | unsigned long start, end, addr = vmf->address; |
| 5498 | unsigned long addr_start = addr - (nr << PAGE_SHIFT); |
| 5499 | unsigned long pt_start = ALIGN_DOWN(addr, PMD_SIZE); |
| 5500 | pte_t *start_ptep; |
| 5501 | |
| 5502 | /* Stay within the VMA and within the page table. */ |
| 5503 | start = max3(addr_start, pt_start, vma->vm_start); |
| 5504 | end = min3(addr_start + folio_size(folio), pt_start + PMD_SIZE, |
| 5505 | vma->vm_end); |
| 5506 | start_ptep = vmf->pte - ((addr - start) >> PAGE_SHIFT); |
| 5507 | |
| 5508 | /* Restore all PTEs' mapping of the large folio */ |
| 5509 | for (addr = start; addr != end; start_ptep++, addr += PAGE_SIZE) { |
| 5510 | pte_t ptent = ptep_get(start_ptep); |
| 5511 | bool writable = false; |
| 5512 | |
| 5513 | if (!pte_present(ptent) || !pte_protnone(ptent)) |
| 5514 | continue; |
| 5515 | |
| 5516 | if (pfn_folio(pte_pfn(ptent)) != folio) |
| 5517 | continue; |
| 5518 | |
| 5519 | if (!ignore_writable) { |
| 5520 | ptent = pte_modify(ptent, vma->vm_page_prot); |
| 5521 | writable = pte_write(ptent); |
| 5522 | if (!writable && pte_write_upgrade && |
| 5523 | can_change_pte_writable(vma, addr, ptent)) |
| 5524 | writable = true; |
| 5525 | } |
| 5526 | |
| 5527 | numa_rebuild_single_mapping(vmf, vma, addr, start_ptep, writable); |
| 5528 | } |
| 5529 | } |
| 5530 | |
| 5531 | static vm_fault_t do_numa_page(struct vm_fault *vmf) |
| 5532 | { |
| 5533 | struct vm_area_struct *vma = vmf->vma; |
| 5534 | struct folio *folio = NULL; |
| 5535 | int nid = NUMA_NO_NODE; |
| 5536 | bool writable = false, ignore_writable = false; |
| 5537 | bool pte_write_upgrade = vma_wants_manual_pte_write_upgrade(vma); |
| 5538 | int last_cpupid; |
| 5539 | int target_nid; |
| 5540 | pte_t pte, old_pte; |
| 5541 | int flags = 0, nr_pages; |
| 5542 | |
| 5543 | /* |
| 5544 | * The pte cannot be used safely until we verify, while holding the page |
| 5545 | * table lock, that its contents have not changed during fault handling. |
| 5546 | */ |
| 5547 | spin_lock(vmf->ptl); |
| 5548 | /* Read the live PTE from the page tables: */ |
| 5549 | old_pte = ptep_get(vmf->pte); |
| 5550 | |
| 5551 | if (unlikely(!pte_same(old_pte, vmf->orig_pte))) { |
| 5552 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5553 | return 0; |
| 5554 | } |
| 5555 | |
| 5556 | pte = pte_modify(old_pte, vma->vm_page_prot); |
| 5557 | |
| 5558 | /* |
| 5559 | * Detect now whether the PTE could be writable; this information |
| 5560 | * is only valid while holding the PT lock. |
| 5561 | */ |
| 5562 | writable = pte_write(pte); |
| 5563 | if (!writable && pte_write_upgrade && |
| 5564 | can_change_pte_writable(vma, vmf->address, pte)) |
| 5565 | writable = true; |
| 5566 | |
| 5567 | folio = vm_normal_folio(vma, vmf->address, pte); |
| 5568 | if (!folio || folio_is_zone_device(folio)) |
| 5569 | goto out_map; |
| 5570 | |
| 5571 | nid = folio_nid(folio); |
| 5572 | nr_pages = folio_nr_pages(folio); |
| 5573 | |
| 5574 | target_nid = numa_migrate_check(folio, vmf, vmf->address, &flags, |
| 5575 | writable, &last_cpupid); |
| 5576 | if (target_nid == NUMA_NO_NODE) |
| 5577 | goto out_map; |
| 5578 | if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) { |
| 5579 | flags |= TNF_MIGRATE_FAIL; |
| 5580 | goto out_map; |
| 5581 | } |
| 5582 | /* The folio is isolated and isolation code holds a folio reference. */ |
| 5583 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5584 | writable = false; |
| 5585 | ignore_writable = true; |
| 5586 | |
| 5587 | /* Migrate to the requested node */ |
| 5588 | if (!migrate_misplaced_folio(folio, vma, target_nid)) { |
| 5589 | nid = target_nid; |
| 5590 | flags |= TNF_MIGRATED; |
| 5591 | task_numa_fault(last_cpupid, nid, nr_pages, flags); |
| 5592 | return 0; |
| 5593 | } |
| 5594 | |
| 5595 | flags |= TNF_MIGRATE_FAIL; |
| 5596 | vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, |
| 5597 | vmf->address, &vmf->ptl); |
| 5598 | if (unlikely(!vmf->pte)) |
| 5599 | return 0; |
| 5600 | if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { |
| 5601 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5602 | return 0; |
| 5603 | } |
| 5604 | out_map: |
| 5605 | /* |
| 5606 | * Make it present again, depending on how arch implements |
| 5607 | * non-accessible ptes, some can allow access by kernel mode. |
| 5608 | */ |
| 5609 | if (folio && folio_test_large(folio)) |
| 5610 | numa_rebuild_large_mapping(vmf, vma, folio, pte, ignore_writable, |
| 5611 | pte_write_upgrade); |
| 5612 | else |
| 5613 | numa_rebuild_single_mapping(vmf, vma, vmf->address, vmf->pte, |
| 5614 | writable); |
| 5615 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5616 | |
| 5617 | if (nid != NUMA_NO_NODE) |
| 5618 | task_numa_fault(last_cpupid, nid, nr_pages, flags); |
| 5619 | return 0; |
| 5620 | } |
| 5621 | |
| 5622 | static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) |
| 5623 | { |
| 5624 | struct vm_area_struct *vma = vmf->vma; |
| 5625 | if (vma_is_anonymous(vma)) |
| 5626 | return do_huge_pmd_anonymous_page(vmf); |
| 5627 | if (vma->vm_ops->huge_fault) |
| 5628 | return vma->vm_ops->huge_fault(vmf, PMD_ORDER); |
| 5629 | return VM_FAULT_FALLBACK; |
| 5630 | } |
| 5631 | |
| 5632 | /* `inline' is required to avoid gcc 4.1.2 build error */ |
| 5633 | static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf) |
| 5634 | { |
| 5635 | struct vm_area_struct *vma = vmf->vma; |
| 5636 | const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; |
| 5637 | vm_fault_t ret; |
| 5638 | |
| 5639 | if (vma_is_anonymous(vma)) { |
| 5640 | if (likely(!unshare) && |
| 5641 | userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) { |
| 5642 | if (userfaultfd_wp_async(vmf->vma)) |
| 5643 | goto split; |
| 5644 | return handle_userfault(vmf, VM_UFFD_WP); |
| 5645 | } |
| 5646 | return do_huge_pmd_wp_page(vmf); |
| 5647 | } |
| 5648 | |
| 5649 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { |
| 5650 | if (vma->vm_ops->huge_fault) { |
| 5651 | ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER); |
| 5652 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5653 | return ret; |
| 5654 | } |
| 5655 | } |
| 5656 | |
| 5657 | split: |
| 5658 | /* COW or write-notify handled on pte level: split pmd. */ |
| 5659 | __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); |
| 5660 | |
| 5661 | return VM_FAULT_FALLBACK; |
| 5662 | } |
| 5663 | |
| 5664 | static vm_fault_t create_huge_pud(struct vm_fault *vmf) |
| 5665 | { |
| 5666 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ |
| 5667 | defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) |
| 5668 | struct vm_area_struct *vma = vmf->vma; |
| 5669 | /* No support for anonymous transparent PUD pages yet */ |
| 5670 | if (vma_is_anonymous(vma)) |
| 5671 | return VM_FAULT_FALLBACK; |
| 5672 | if (vma->vm_ops->huge_fault) |
| 5673 | return vma->vm_ops->huge_fault(vmf, PUD_ORDER); |
| 5674 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |
| 5675 | return VM_FAULT_FALLBACK; |
| 5676 | } |
| 5677 | |
| 5678 | static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) |
| 5679 | { |
| 5680 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ |
| 5681 | defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) |
| 5682 | struct vm_area_struct *vma = vmf->vma; |
| 5683 | vm_fault_t ret; |
| 5684 | |
| 5685 | /* No support for anonymous transparent PUD pages yet */ |
| 5686 | if (vma_is_anonymous(vma)) |
| 5687 | goto split; |
| 5688 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { |
| 5689 | if (vma->vm_ops->huge_fault) { |
| 5690 | ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER); |
| 5691 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5692 | return ret; |
| 5693 | } |
| 5694 | } |
| 5695 | split: |
| 5696 | /* COW or write-notify not handled on PUD level: split pud.*/ |
| 5697 | __split_huge_pud(vma, vmf->pud, vmf->address); |
| 5698 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ |
| 5699 | return VM_FAULT_FALLBACK; |
| 5700 | } |
| 5701 | |
| 5702 | /* |
| 5703 | * These routines also need to handle stuff like marking pages dirty |
| 5704 | * and/or accessed for architectures that don't do it in hardware (most |
| 5705 | * RISC architectures). The early dirtying is also good on the i386. |
| 5706 | * |
| 5707 | * There is also a hook called "update_mmu_cache()" that architectures |
| 5708 | * with external mmu caches can use to update those (ie the Sparc or |
| 5709 | * PowerPC hashed page tables that act as extended TLBs). |
| 5710 | * |
| 5711 | * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow |
| 5712 | * concurrent faults). |
| 5713 | * |
| 5714 | * The mmap_lock may have been released depending on flags and our return value. |
| 5715 | * See filemap_fault() and __folio_lock_or_retry(). |
| 5716 | */ |
| 5717 | static vm_fault_t handle_pte_fault(struct vm_fault *vmf) |
| 5718 | { |
| 5719 | pte_t entry; |
| 5720 | |
| 5721 | if (unlikely(pmd_none(*vmf->pmd))) { |
| 5722 | /* |
| 5723 | * Leave __pte_alloc() until later: because vm_ops->fault may |
| 5724 | * want to allocate huge page, and if we expose page table |
| 5725 | * for an instant, it will be difficult to retract from |
| 5726 | * concurrent faults and from rmap lookups. |
| 5727 | */ |
| 5728 | vmf->pte = NULL; |
| 5729 | vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID; |
| 5730 | } else { |
| 5731 | /* |
| 5732 | * A regular pmd is established and it can't morph into a huge |
| 5733 | * pmd by anon khugepaged, since that takes mmap_lock in write |
| 5734 | * mode; but shmem or file collapse to THP could still morph |
| 5735 | * it into a huge pmd: just retry later if so. |
| 5736 | */ |
| 5737 | vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd, |
| 5738 | vmf->address, &vmf->ptl); |
| 5739 | if (unlikely(!vmf->pte)) |
| 5740 | return 0; |
| 5741 | vmf->orig_pte = ptep_get_lockless(vmf->pte); |
| 5742 | vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID; |
| 5743 | |
| 5744 | if (pte_none(vmf->orig_pte)) { |
| 5745 | pte_unmap(vmf->pte); |
| 5746 | vmf->pte = NULL; |
| 5747 | } |
| 5748 | } |
| 5749 | |
| 5750 | if (!vmf->pte) |
| 5751 | return do_pte_missing(vmf); |
| 5752 | |
| 5753 | if (!pte_present(vmf->orig_pte)) |
| 5754 | return do_swap_page(vmf); |
| 5755 | |
| 5756 | if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) |
| 5757 | return do_numa_page(vmf); |
| 5758 | |
| 5759 | spin_lock(vmf->ptl); |
| 5760 | entry = vmf->orig_pte; |
| 5761 | if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) { |
| 5762 | update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); |
| 5763 | goto unlock; |
| 5764 | } |
| 5765 | if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
| 5766 | if (!pte_write(entry)) |
| 5767 | return do_wp_page(vmf); |
| 5768 | else if (likely(vmf->flags & FAULT_FLAG_WRITE)) |
| 5769 | entry = pte_mkdirty(entry); |
| 5770 | } |
| 5771 | entry = pte_mkyoung(entry); |
| 5772 | if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, |
| 5773 | vmf->flags & FAULT_FLAG_WRITE)) { |
| 5774 | update_mmu_cache_range(vmf, vmf->vma, vmf->address, |
| 5775 | vmf->pte, 1); |
| 5776 | } else { |
| 5777 | /* Skip spurious TLB flush for retried page fault */ |
| 5778 | if (vmf->flags & FAULT_FLAG_TRIED) |
| 5779 | goto unlock; |
| 5780 | /* |
| 5781 | * This is needed only for protection faults but the arch code |
| 5782 | * is not yet telling us if this is a protection fault or not. |
| 5783 | * This still avoids useless tlb flushes for .text page faults |
| 5784 | * with threads. |
| 5785 | */ |
| 5786 | if (vmf->flags & FAULT_FLAG_WRITE) |
| 5787 | flush_tlb_fix_spurious_fault(vmf->vma, vmf->address, |
| 5788 | vmf->pte); |
| 5789 | } |
| 5790 | unlock: |
| 5791 | pte_unmap_unlock(vmf->pte, vmf->ptl); |
| 5792 | return 0; |
| 5793 | } |
| 5794 | |
| 5795 | /* |
| 5796 | * On entry, we hold either the VMA lock or the mmap_lock |
| 5797 | * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in |
| 5798 | * the result, the mmap_lock is not held on exit. See filemap_fault() |
| 5799 | * and __folio_lock_or_retry(). |
| 5800 | */ |
| 5801 | static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, |
| 5802 | unsigned long address, unsigned int flags) |
| 5803 | { |
| 5804 | struct vm_fault vmf = { |
| 5805 | .vma = vma, |
| 5806 | .address = address & PAGE_MASK, |
| 5807 | .real_address = address, |
| 5808 | .flags = flags, |
| 5809 | .pgoff = linear_page_index(vma, address), |
| 5810 | .gfp_mask = __get_fault_gfp_mask(vma), |
| 5811 | }; |
| 5812 | struct mm_struct *mm = vma->vm_mm; |
| 5813 | unsigned long vm_flags = vma->vm_flags; |
| 5814 | pgd_t *pgd; |
| 5815 | p4d_t *p4d; |
| 5816 | vm_fault_t ret; |
| 5817 | |
| 5818 | pgd = pgd_offset(mm, address); |
| 5819 | p4d = p4d_alloc(mm, pgd, address); |
| 5820 | if (!p4d) |
| 5821 | return VM_FAULT_OOM; |
| 5822 | |
| 5823 | vmf.pud = pud_alloc(mm, p4d, address); |
| 5824 | if (!vmf.pud) |
| 5825 | return VM_FAULT_OOM; |
| 5826 | retry_pud: |
| 5827 | if (pud_none(*vmf.pud) && |
| 5828 | thp_vma_allowable_order(vma, vm_flags, |
| 5829 | TVA_IN_PF | TVA_ENFORCE_SYSFS, PUD_ORDER)) { |
| 5830 | ret = create_huge_pud(&vmf); |
| 5831 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5832 | return ret; |
| 5833 | } else { |
| 5834 | pud_t orig_pud = *vmf.pud; |
| 5835 | |
| 5836 | barrier(); |
| 5837 | if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { |
| 5838 | |
| 5839 | /* |
| 5840 | * TODO once we support anonymous PUDs: NUMA case and |
| 5841 | * FAULT_FLAG_UNSHARE handling. |
| 5842 | */ |
| 5843 | if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) { |
| 5844 | ret = wp_huge_pud(&vmf, orig_pud); |
| 5845 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5846 | return ret; |
| 5847 | } else { |
| 5848 | huge_pud_set_accessed(&vmf, orig_pud); |
| 5849 | return 0; |
| 5850 | } |
| 5851 | } |
| 5852 | } |
| 5853 | |
| 5854 | vmf.pmd = pmd_alloc(mm, vmf.pud, address); |
| 5855 | if (!vmf.pmd) |
| 5856 | return VM_FAULT_OOM; |
| 5857 | |
| 5858 | /* Huge pud page fault raced with pmd_alloc? */ |
| 5859 | if (pud_trans_unstable(vmf.pud)) |
| 5860 | goto retry_pud; |
| 5861 | |
| 5862 | if (pmd_none(*vmf.pmd) && |
| 5863 | thp_vma_allowable_order(vma, vm_flags, |
| 5864 | TVA_IN_PF | TVA_ENFORCE_SYSFS, PMD_ORDER)) { |
| 5865 | ret = create_huge_pmd(&vmf); |
| 5866 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5867 | return ret; |
| 5868 | } else { |
| 5869 | vmf.orig_pmd = pmdp_get_lockless(vmf.pmd); |
| 5870 | |
| 5871 | if (unlikely(is_swap_pmd(vmf.orig_pmd))) { |
| 5872 | VM_BUG_ON(thp_migration_supported() && |
| 5873 | !is_pmd_migration_entry(vmf.orig_pmd)); |
| 5874 | if (is_pmd_migration_entry(vmf.orig_pmd)) |
| 5875 | pmd_migration_entry_wait(mm, vmf.pmd); |
| 5876 | return 0; |
| 5877 | } |
| 5878 | if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) { |
| 5879 | if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma)) |
| 5880 | return do_huge_pmd_numa_page(&vmf); |
| 5881 | |
| 5882 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
| 5883 | !pmd_write(vmf.orig_pmd)) { |
| 5884 | ret = wp_huge_pmd(&vmf); |
| 5885 | if (!(ret & VM_FAULT_FALLBACK)) |
| 5886 | return ret; |
| 5887 | } else { |
| 5888 | huge_pmd_set_accessed(&vmf); |
| 5889 | return 0; |
| 5890 | } |
| 5891 | } |
| 5892 | } |
| 5893 | |
| 5894 | return handle_pte_fault(&vmf); |
| 5895 | } |
| 5896 | |
| 5897 | /** |
| 5898 | * mm_account_fault - Do page fault accounting |
| 5899 | * @mm: mm from which memcg should be extracted. It can be NULL. |
| 5900 | * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting |
| 5901 | * of perf event counters, but we'll still do the per-task accounting to |
| 5902 | * the task who triggered this page fault. |
| 5903 | * @address: the faulted address. |
| 5904 | * @flags: the fault flags. |
| 5905 | * @ret: the fault retcode. |
| 5906 | * |
| 5907 | * This will take care of most of the page fault accounting. Meanwhile, it |
| 5908 | * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter |
| 5909 | * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should |
| 5910 | * still be in per-arch page fault handlers at the entry of page fault. |
| 5911 | */ |
| 5912 | static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs, |
| 5913 | unsigned long address, unsigned int flags, |
| 5914 | vm_fault_t ret) |
| 5915 | { |
| 5916 | bool major; |
| 5917 | |
| 5918 | /* Incomplete faults will be accounted upon completion. */ |
| 5919 | if (ret & VM_FAULT_RETRY) |
| 5920 | return; |
| 5921 | |
| 5922 | /* |
| 5923 | * To preserve the behavior of older kernels, PGFAULT counters record |
| 5924 | * both successful and failed faults, as opposed to perf counters, |
| 5925 | * which ignore failed cases. |
| 5926 | */ |
| 5927 | count_vm_event(PGFAULT); |
| 5928 | count_memcg_event_mm(mm, PGFAULT); |
| 5929 | |
| 5930 | /* |
| 5931 | * Do not account for unsuccessful faults (e.g. when the address wasn't |
| 5932 | * valid). That includes arch_vma_access_permitted() failing before |
| 5933 | * reaching here. So this is not a "this many hardware page faults" |
| 5934 | * counter. We should use the hw profiling for that. |
| 5935 | */ |
| 5936 | if (ret & VM_FAULT_ERROR) |
| 5937 | return; |
| 5938 | |
| 5939 | /* |
| 5940 | * We define the fault as a major fault when the final successful fault |
| 5941 | * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't |
| 5942 | * handle it immediately previously). |
| 5943 | */ |
| 5944 | major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); |
| 5945 | |
| 5946 | if (major) |
| 5947 | current->maj_flt++; |
| 5948 | else |
| 5949 | current->min_flt++; |
| 5950 | |
| 5951 | /* |
| 5952 | * If the fault is done for GUP, regs will be NULL. We only do the |
| 5953 | * accounting for the per thread fault counters who triggered the |
| 5954 | * fault, and we skip the perf event updates. |
| 5955 | */ |
| 5956 | if (!regs) |
| 5957 | return; |
| 5958 | |
| 5959 | if (major) |
| 5960 | perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); |
| 5961 | else |
| 5962 | perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); |
| 5963 | } |
| 5964 | |
| 5965 | #ifdef CONFIG_LRU_GEN |
| 5966 | static void lru_gen_enter_fault(struct vm_area_struct *vma) |
| 5967 | { |
| 5968 | /* the LRU algorithm only applies to accesses with recency */ |
| 5969 | current->in_lru_fault = vma_has_recency(vma); |
| 5970 | } |
| 5971 | |
| 5972 | static void lru_gen_exit_fault(void) |
| 5973 | { |
| 5974 | current->in_lru_fault = false; |
| 5975 | } |
| 5976 | #else |
| 5977 | static void lru_gen_enter_fault(struct vm_area_struct *vma) |
| 5978 | { |
| 5979 | } |
| 5980 | |
| 5981 | static void lru_gen_exit_fault(void) |
| 5982 | { |
| 5983 | } |
| 5984 | #endif /* CONFIG_LRU_GEN */ |
| 5985 | |
| 5986 | static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma, |
| 5987 | unsigned int *flags) |
| 5988 | { |
| 5989 | if (unlikely(*flags & FAULT_FLAG_UNSHARE)) { |
| 5990 | if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE)) |
| 5991 | return VM_FAULT_SIGSEGV; |
| 5992 | /* |
| 5993 | * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's |
| 5994 | * just treat it like an ordinary read-fault otherwise. |
| 5995 | */ |
| 5996 | if (!is_cow_mapping(vma->vm_flags)) |
| 5997 | *flags &= ~FAULT_FLAG_UNSHARE; |
| 5998 | } else if (*flags & FAULT_FLAG_WRITE) { |
| 5999 | /* Write faults on read-only mappings are impossible ... */ |
| 6000 | if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE))) |
| 6001 | return VM_FAULT_SIGSEGV; |
| 6002 | /* ... and FOLL_FORCE only applies to COW mappings. */ |
| 6003 | if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) && |
| 6004 | !is_cow_mapping(vma->vm_flags))) |
| 6005 | return VM_FAULT_SIGSEGV; |
| 6006 | } |
| 6007 | #ifdef CONFIG_PER_VMA_LOCK |
| 6008 | /* |
| 6009 | * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of |
| 6010 | * the assumption that lock is dropped on VM_FAULT_RETRY. |
| 6011 | */ |
| 6012 | if (WARN_ON_ONCE((*flags & |
| 6013 | (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) == |
| 6014 | (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT))) |
| 6015 | return VM_FAULT_SIGSEGV; |
| 6016 | #endif |
| 6017 | |
| 6018 | return 0; |
| 6019 | } |
| 6020 | |
| 6021 | /* |
| 6022 | * By the time we get here, we already hold the mm semaphore |
| 6023 | * |
| 6024 | * The mmap_lock may have been released depending on flags and our |
| 6025 | * return value. See filemap_fault() and __folio_lock_or_retry(). |
| 6026 | */ |
| 6027 | vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, |
| 6028 | unsigned int flags, struct pt_regs *regs) |
| 6029 | { |
| 6030 | /* If the fault handler drops the mmap_lock, vma may be freed */ |
| 6031 | struct mm_struct *mm = vma->vm_mm; |
| 6032 | vm_fault_t ret; |
| 6033 | bool is_droppable; |
| 6034 | |
| 6035 | __set_current_state(TASK_RUNNING); |
| 6036 | |
| 6037 | ret = sanitize_fault_flags(vma, &flags); |
| 6038 | if (ret) |
| 6039 | goto out; |
| 6040 | |
| 6041 | if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, |
| 6042 | flags & FAULT_FLAG_INSTRUCTION, |
| 6043 | flags & FAULT_FLAG_REMOTE)) { |
| 6044 | ret = VM_FAULT_SIGSEGV; |
| 6045 | goto out; |
| 6046 | } |
| 6047 | |
| 6048 | is_droppable = !!(vma->vm_flags & VM_DROPPABLE); |
| 6049 | |
| 6050 | /* |
| 6051 | * Enable the memcg OOM handling for faults triggered in user |
| 6052 | * space. Kernel faults are handled more gracefully. |
| 6053 | */ |
| 6054 | if (flags & FAULT_FLAG_USER) |
| 6055 | mem_cgroup_enter_user_fault(); |
| 6056 | |
| 6057 | lru_gen_enter_fault(vma); |
| 6058 | |
| 6059 | if (unlikely(is_vm_hugetlb_page(vma))) |
| 6060 | ret = hugetlb_fault(vma->vm_mm, vma, address, flags); |
| 6061 | else |
| 6062 | ret = __handle_mm_fault(vma, address, flags); |
| 6063 | |
| 6064 | /* |
| 6065 | * Warning: It is no longer safe to dereference vma-> after this point, |
| 6066 | * because mmap_lock might have been dropped by __handle_mm_fault(), so |
| 6067 | * vma might be destroyed from underneath us. |
| 6068 | */ |
| 6069 | |
| 6070 | lru_gen_exit_fault(); |
| 6071 | |
| 6072 | /* If the mapping is droppable, then errors due to OOM aren't fatal. */ |
| 6073 | if (is_droppable) |
| 6074 | ret &= ~VM_FAULT_OOM; |
| 6075 | |
| 6076 | if (flags & FAULT_FLAG_USER) { |
| 6077 | mem_cgroup_exit_user_fault(); |
| 6078 | /* |
| 6079 | * The task may have entered a memcg OOM situation but |
| 6080 | * if the allocation error was handled gracefully (no |
| 6081 | * VM_FAULT_OOM), there is no need to kill anything. |
| 6082 | * Just clean up the OOM state peacefully. |
| 6083 | */ |
| 6084 | if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) |
| 6085 | mem_cgroup_oom_synchronize(false); |
| 6086 | } |
| 6087 | out: |
| 6088 | mm_account_fault(mm, regs, address, flags, ret); |
| 6089 | |
| 6090 | return ret; |
| 6091 | } |
| 6092 | EXPORT_SYMBOL_GPL(handle_mm_fault); |
| 6093 | |
| 6094 | #ifdef CONFIG_LOCK_MM_AND_FIND_VMA |
| 6095 | #include <linux/extable.h> |
| 6096 | |
| 6097 | static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) |
| 6098 | { |
| 6099 | if (likely(mmap_read_trylock(mm))) |
| 6100 | return true; |
| 6101 | |
| 6102 | if (regs && !user_mode(regs)) { |
| 6103 | unsigned long ip = exception_ip(regs); |
| 6104 | if (!search_exception_tables(ip)) |
| 6105 | return false; |
| 6106 | } |
| 6107 | |
| 6108 | return !mmap_read_lock_killable(mm); |
| 6109 | } |
| 6110 | |
| 6111 | static inline bool mmap_upgrade_trylock(struct mm_struct *mm) |
| 6112 | { |
| 6113 | /* |
| 6114 | * We don't have this operation yet. |
| 6115 | * |
| 6116 | * It should be easy enough to do: it's basically a |
| 6117 | * atomic_long_try_cmpxchg_acquire() |
| 6118 | * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but |
| 6119 | * it also needs the proper lockdep magic etc. |
| 6120 | */ |
| 6121 | return false; |
| 6122 | } |
| 6123 | |
| 6124 | static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) |
| 6125 | { |
| 6126 | mmap_read_unlock(mm); |
| 6127 | if (regs && !user_mode(regs)) { |
| 6128 | unsigned long ip = exception_ip(regs); |
| 6129 | if (!search_exception_tables(ip)) |
| 6130 | return false; |
| 6131 | } |
| 6132 | return !mmap_write_lock_killable(mm); |
| 6133 | } |
| 6134 | |
| 6135 | /* |
| 6136 | * Helper for page fault handling. |
| 6137 | * |
| 6138 | * This is kind of equivalend to "mmap_read_lock()" followed |
| 6139 | * by "find_extend_vma()", except it's a lot more careful about |
| 6140 | * the locking (and will drop the lock on failure). |
| 6141 | * |
| 6142 | * For example, if we have a kernel bug that causes a page |
| 6143 | * fault, we don't want to just use mmap_read_lock() to get |
| 6144 | * the mm lock, because that would deadlock if the bug were |
| 6145 | * to happen while we're holding the mm lock for writing. |
| 6146 | * |
| 6147 | * So this checks the exception tables on kernel faults in |
| 6148 | * order to only do this all for instructions that are actually |
| 6149 | * expected to fault. |
| 6150 | * |
| 6151 | * We can also actually take the mm lock for writing if we |
| 6152 | * need to extend the vma, which helps the VM layer a lot. |
| 6153 | */ |
| 6154 | struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm, |
| 6155 | unsigned long addr, struct pt_regs *regs) |
| 6156 | { |
| 6157 | struct vm_area_struct *vma; |
| 6158 | |
| 6159 | if (!get_mmap_lock_carefully(mm, regs)) |
| 6160 | return NULL; |
| 6161 | |
| 6162 | vma = find_vma(mm, addr); |
| 6163 | if (likely(vma && (vma->vm_start <= addr))) |
| 6164 | return vma; |
| 6165 | |
| 6166 | /* |
| 6167 | * Well, dang. We might still be successful, but only |
| 6168 | * if we can extend a vma to do so. |
| 6169 | */ |
| 6170 | if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) { |
| 6171 | mmap_read_unlock(mm); |
| 6172 | return NULL; |
| 6173 | } |
| 6174 | |
| 6175 | /* |
| 6176 | * We can try to upgrade the mmap lock atomically, |
| 6177 | * in which case we can continue to use the vma |
| 6178 | * we already looked up. |
| 6179 | * |
| 6180 | * Otherwise we'll have to drop the mmap lock and |
| 6181 | * re-take it, and also look up the vma again, |
| 6182 | * re-checking it. |
| 6183 | */ |
| 6184 | if (!mmap_upgrade_trylock(mm)) { |
| 6185 | if (!upgrade_mmap_lock_carefully(mm, regs)) |
| 6186 | return NULL; |
| 6187 | |
| 6188 | vma = find_vma(mm, addr); |
| 6189 | if (!vma) |
| 6190 | goto fail; |
| 6191 | if (vma->vm_start <= addr) |
| 6192 | goto success; |
| 6193 | if (!(vma->vm_flags & VM_GROWSDOWN)) |
| 6194 | goto fail; |
| 6195 | } |
| 6196 | |
| 6197 | if (expand_stack_locked(vma, addr)) |
| 6198 | goto fail; |
| 6199 | |
| 6200 | success: |
| 6201 | mmap_write_downgrade(mm); |
| 6202 | return vma; |
| 6203 | |
| 6204 | fail: |
| 6205 | mmap_write_unlock(mm); |
| 6206 | return NULL; |
| 6207 | } |
| 6208 | #endif |
| 6209 | |
| 6210 | #ifdef CONFIG_PER_VMA_LOCK |
| 6211 | /* |
| 6212 | * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be |
| 6213 | * stable and not isolated. If the VMA is not found or is being modified the |
| 6214 | * function returns NULL. |
| 6215 | */ |
| 6216 | struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm, |
| 6217 | unsigned long address) |
| 6218 | { |
| 6219 | MA_STATE(mas, &mm->mm_mt, address, address); |
| 6220 | struct vm_area_struct *vma; |
| 6221 | |
| 6222 | rcu_read_lock(); |
| 6223 | retry: |
| 6224 | vma = mas_walk(&mas); |
| 6225 | if (!vma) |
| 6226 | goto inval; |
| 6227 | |
| 6228 | if (!vma_start_read(vma)) |
| 6229 | goto inval; |
| 6230 | |
| 6231 | /* Check if the VMA got isolated after we found it */ |
| 6232 | if (vma->detached) { |
| 6233 | vma_end_read(vma); |
| 6234 | count_vm_vma_lock_event(VMA_LOCK_MISS); |
| 6235 | /* The area was replaced with another one */ |
| 6236 | goto retry; |
| 6237 | } |
| 6238 | /* |
| 6239 | * At this point, we have a stable reference to a VMA: The VMA is |
| 6240 | * locked and we know it hasn't already been isolated. |
| 6241 | * From here on, we can access the VMA without worrying about which |
| 6242 | * fields are accessible for RCU readers. |
| 6243 | */ |
| 6244 | |
| 6245 | /* Check since vm_start/vm_end might change before we lock the VMA */ |
| 6246 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) |
| 6247 | goto inval_end_read; |
| 6248 | |
| 6249 | rcu_read_unlock(); |
| 6250 | return vma; |
| 6251 | |
| 6252 | inval_end_read: |
| 6253 | vma_end_read(vma); |
| 6254 | inval: |
| 6255 | rcu_read_unlock(); |
| 6256 | count_vm_vma_lock_event(VMA_LOCK_ABORT); |
| 6257 | return NULL; |
| 6258 | } |
| 6259 | #endif /* CONFIG_PER_VMA_LOCK */ |
| 6260 | |
| 6261 | #ifndef __PAGETABLE_P4D_FOLDED |
| 6262 | /* |
| 6263 | * Allocate p4d page table. |
| 6264 | * We've already handled the fast-path in-line. |
| 6265 | */ |
| 6266 | int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) |
| 6267 | { |
| 6268 | p4d_t *new = p4d_alloc_one(mm, address); |
| 6269 | if (!new) |
| 6270 | return -ENOMEM; |
| 6271 | |
| 6272 | spin_lock(&mm->page_table_lock); |
| 6273 | if (pgd_present(*pgd)) { /* Another has populated it */ |
| 6274 | p4d_free(mm, new); |
| 6275 | } else { |
| 6276 | smp_wmb(); /* See comment in pmd_install() */ |
| 6277 | pgd_populate(mm, pgd, new); |
| 6278 | } |
| 6279 | spin_unlock(&mm->page_table_lock); |
| 6280 | return 0; |
| 6281 | } |
| 6282 | #endif /* __PAGETABLE_P4D_FOLDED */ |
| 6283 | |
| 6284 | #ifndef __PAGETABLE_PUD_FOLDED |
| 6285 | /* |
| 6286 | * Allocate page upper directory. |
| 6287 | * We've already handled the fast-path in-line. |
| 6288 | */ |
| 6289 | int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) |
| 6290 | { |
| 6291 | pud_t *new = pud_alloc_one(mm, address); |
| 6292 | if (!new) |
| 6293 | return -ENOMEM; |
| 6294 | |
| 6295 | spin_lock(&mm->page_table_lock); |
| 6296 | if (!p4d_present(*p4d)) { |
| 6297 | mm_inc_nr_puds(mm); |
| 6298 | smp_wmb(); /* See comment in pmd_install() */ |
| 6299 | p4d_populate(mm, p4d, new); |
| 6300 | } else /* Another has populated it */ |
| 6301 | pud_free(mm, new); |
| 6302 | spin_unlock(&mm->page_table_lock); |
| 6303 | return 0; |
| 6304 | } |
| 6305 | #endif /* __PAGETABLE_PUD_FOLDED */ |
| 6306 | |
| 6307 | #ifndef __PAGETABLE_PMD_FOLDED |
| 6308 | /* |
| 6309 | * Allocate page middle directory. |
| 6310 | * We've already handled the fast-path in-line. |
| 6311 | */ |
| 6312 | int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) |
| 6313 | { |
| 6314 | spinlock_t *ptl; |
| 6315 | pmd_t *new = pmd_alloc_one(mm, address); |
| 6316 | if (!new) |
| 6317 | return -ENOMEM; |
| 6318 | |
| 6319 | ptl = pud_lock(mm, pud); |
| 6320 | if (!pud_present(*pud)) { |
| 6321 | mm_inc_nr_pmds(mm); |
| 6322 | smp_wmb(); /* See comment in pmd_install() */ |
| 6323 | pud_populate(mm, pud, new); |
| 6324 | } else { /* Another has populated it */ |
| 6325 | pmd_free(mm, new); |
| 6326 | } |
| 6327 | spin_unlock(ptl); |
| 6328 | return 0; |
| 6329 | } |
| 6330 | #endif /* __PAGETABLE_PMD_FOLDED */ |
| 6331 | |
| 6332 | static inline void pfnmap_args_setup(struct follow_pfnmap_args *args, |
| 6333 | spinlock_t *lock, pte_t *ptep, |
| 6334 | pgprot_t pgprot, unsigned long pfn_base, |
| 6335 | unsigned long addr_mask, bool writable, |
| 6336 | bool special) |
| 6337 | { |
| 6338 | args->lock = lock; |
| 6339 | args->ptep = ptep; |
| 6340 | args->pfn = pfn_base + ((args->address & ~addr_mask) >> PAGE_SHIFT); |
| 6341 | args->pgprot = pgprot; |
| 6342 | args->writable = writable; |
| 6343 | args->special = special; |
| 6344 | } |
| 6345 | |
| 6346 | static inline void pfnmap_lockdep_assert(struct vm_area_struct *vma) |
| 6347 | { |
| 6348 | #ifdef CONFIG_LOCKDEP |
| 6349 | struct address_space *mapping = vma->vm_file->f_mapping; |
| 6350 | |
| 6351 | if (mapping) |
| 6352 | lockdep_assert(lockdep_is_held(&vma->vm_file->f_mapping->i_mmap_rwsem) || |
| 6353 | lockdep_is_held(&vma->vm_mm->mmap_lock)); |
| 6354 | else |
| 6355 | lockdep_assert(lockdep_is_held(&vma->vm_mm->mmap_lock)); |
| 6356 | #endif |
| 6357 | } |
| 6358 | |
| 6359 | /** |
| 6360 | * follow_pfnmap_start() - Look up a pfn mapping at a user virtual address |
| 6361 | * @args: Pointer to struct @follow_pfnmap_args |
| 6362 | * |
| 6363 | * The caller needs to setup args->vma and args->address to point to the |
| 6364 | * virtual address as the target of such lookup. On a successful return, |
| 6365 | * the results will be put into other output fields. |
| 6366 | * |
| 6367 | * After the caller finished using the fields, the caller must invoke |
| 6368 | * another follow_pfnmap_end() to proper releases the locks and resources |
| 6369 | * of such look up request. |
| 6370 | * |
| 6371 | * During the start() and end() calls, the results in @args will be valid |
| 6372 | * as proper locks will be held. After the end() is called, all the fields |
| 6373 | * in @follow_pfnmap_args will be invalid to be further accessed. Further |
| 6374 | * use of such information after end() may require proper synchronizations |
| 6375 | * by the caller with page table updates, otherwise it can create a |
| 6376 | * security bug. |
| 6377 | * |
| 6378 | * If the PTE maps a refcounted page, callers are responsible to protect |
| 6379 | * against invalidation with MMU notifiers; otherwise access to the PFN at |
| 6380 | * a later point in time can trigger use-after-free. |
| 6381 | * |
| 6382 | * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore |
| 6383 | * should be taken for read, and the mmap semaphore cannot be released |
| 6384 | * before the end() is invoked. |
| 6385 | * |
| 6386 | * This function must not be used to modify PTE content. |
| 6387 | * |
| 6388 | * Return: zero on success, negative otherwise. |
| 6389 | */ |
| 6390 | int follow_pfnmap_start(struct follow_pfnmap_args *args) |
| 6391 | { |
| 6392 | struct vm_area_struct *vma = args->vma; |
| 6393 | unsigned long address = args->address; |
| 6394 | struct mm_struct *mm = vma->vm_mm; |
| 6395 | spinlock_t *lock; |
| 6396 | pgd_t *pgdp; |
| 6397 | p4d_t *p4dp, p4d; |
| 6398 | pud_t *pudp, pud; |
| 6399 | pmd_t *pmdp, pmd; |
| 6400 | pte_t *ptep, pte; |
| 6401 | |
| 6402 | pfnmap_lockdep_assert(vma); |
| 6403 | |
| 6404 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) |
| 6405 | goto out; |
| 6406 | |
| 6407 | if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) |
| 6408 | goto out; |
| 6409 | retry: |
| 6410 | pgdp = pgd_offset(mm, address); |
| 6411 | if (pgd_none(*pgdp) || unlikely(pgd_bad(*pgdp))) |
| 6412 | goto out; |
| 6413 | |
| 6414 | p4dp = p4d_offset(pgdp, address); |
| 6415 | p4d = READ_ONCE(*p4dp); |
| 6416 | if (p4d_none(p4d) || unlikely(p4d_bad(p4d))) |
| 6417 | goto out; |
| 6418 | |
| 6419 | pudp = pud_offset(p4dp, address); |
| 6420 | pud = READ_ONCE(*pudp); |
| 6421 | if (pud_none(pud)) |
| 6422 | goto out; |
| 6423 | if (pud_leaf(pud)) { |
| 6424 | lock = pud_lock(mm, pudp); |
| 6425 | if (!unlikely(pud_leaf(pud))) { |
| 6426 | spin_unlock(lock); |
| 6427 | goto retry; |
| 6428 | } |
| 6429 | pfnmap_args_setup(args, lock, NULL, pud_pgprot(pud), |
| 6430 | pud_pfn(pud), PUD_MASK, pud_write(pud), |
| 6431 | pud_special(pud)); |
| 6432 | return 0; |
| 6433 | } |
| 6434 | |
| 6435 | pmdp = pmd_offset(pudp, address); |
| 6436 | pmd = pmdp_get_lockless(pmdp); |
| 6437 | if (pmd_leaf(pmd)) { |
| 6438 | lock = pmd_lock(mm, pmdp); |
| 6439 | if (!unlikely(pmd_leaf(pmd))) { |
| 6440 | spin_unlock(lock); |
| 6441 | goto retry; |
| 6442 | } |
| 6443 | pfnmap_args_setup(args, lock, NULL, pmd_pgprot(pmd), |
| 6444 | pmd_pfn(pmd), PMD_MASK, pmd_write(pmd), |
| 6445 | pmd_special(pmd)); |
| 6446 | return 0; |
| 6447 | } |
| 6448 | |
| 6449 | ptep = pte_offset_map_lock(mm, pmdp, address, &lock); |
| 6450 | if (!ptep) |
| 6451 | goto out; |
| 6452 | pte = ptep_get(ptep); |
| 6453 | if (!pte_present(pte)) |
| 6454 | goto unlock; |
| 6455 | pfnmap_args_setup(args, lock, ptep, pte_pgprot(pte), |
| 6456 | pte_pfn(pte), PAGE_MASK, pte_write(pte), |
| 6457 | pte_special(pte)); |
| 6458 | return 0; |
| 6459 | unlock: |
| 6460 | pte_unmap_unlock(ptep, lock); |
| 6461 | out: |
| 6462 | return -EINVAL; |
| 6463 | } |
| 6464 | EXPORT_SYMBOL_GPL(follow_pfnmap_start); |
| 6465 | |
| 6466 | /** |
| 6467 | * follow_pfnmap_end(): End a follow_pfnmap_start() process |
| 6468 | * @args: Pointer to struct @follow_pfnmap_args |
| 6469 | * |
| 6470 | * Must be used in pair of follow_pfnmap_start(). See the start() function |
| 6471 | * above for more information. |
| 6472 | */ |
| 6473 | void follow_pfnmap_end(struct follow_pfnmap_args *args) |
| 6474 | { |
| 6475 | if (args->lock) |
| 6476 | spin_unlock(args->lock); |
| 6477 | if (args->ptep) |
| 6478 | pte_unmap(args->ptep); |
| 6479 | } |
| 6480 | EXPORT_SYMBOL_GPL(follow_pfnmap_end); |
| 6481 | |
| 6482 | #ifdef CONFIG_HAVE_IOREMAP_PROT |
| 6483 | /** |
| 6484 | * generic_access_phys - generic implementation for iomem mmap access |
| 6485 | * @vma: the vma to access |
| 6486 | * @addr: userspace address, not relative offset within @vma |
| 6487 | * @buf: buffer to read/write |
| 6488 | * @len: length of transfer |
| 6489 | * @write: set to FOLL_WRITE when writing, otherwise reading |
| 6490 | * |
| 6491 | * This is a generic implementation for &vm_operations_struct.access for an |
| 6492 | * iomem mapping. This callback is used by access_process_vm() when the @vma is |
| 6493 | * not page based. |
| 6494 | */ |
| 6495 | int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, |
| 6496 | void *buf, int len, int write) |
| 6497 | { |
| 6498 | resource_size_t phys_addr; |
| 6499 | unsigned long prot = 0; |
| 6500 | void __iomem *maddr; |
| 6501 | int offset = offset_in_page(addr); |
| 6502 | int ret = -EINVAL; |
| 6503 | bool writable; |
| 6504 | struct follow_pfnmap_args args = { .vma = vma, .address = addr }; |
| 6505 | |
| 6506 | retry: |
| 6507 | if (follow_pfnmap_start(&args)) |
| 6508 | return -EINVAL; |
| 6509 | prot = pgprot_val(args.pgprot); |
| 6510 | phys_addr = (resource_size_t)args.pfn << PAGE_SHIFT; |
| 6511 | writable = args.writable; |
| 6512 | follow_pfnmap_end(&args); |
| 6513 | |
| 6514 | if ((write & FOLL_WRITE) && !writable) |
| 6515 | return -EINVAL; |
| 6516 | |
| 6517 | maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); |
| 6518 | if (!maddr) |
| 6519 | return -ENOMEM; |
| 6520 | |
| 6521 | if (follow_pfnmap_start(&args)) |
| 6522 | goto out_unmap; |
| 6523 | |
| 6524 | if ((prot != pgprot_val(args.pgprot)) || |
| 6525 | (phys_addr != (args.pfn << PAGE_SHIFT)) || |
| 6526 | (writable != args.writable)) { |
| 6527 | follow_pfnmap_end(&args); |
| 6528 | iounmap(maddr); |
| 6529 | goto retry; |
| 6530 | } |
| 6531 | |
| 6532 | if (write) |
| 6533 | memcpy_toio(maddr + offset, buf, len); |
| 6534 | else |
| 6535 | memcpy_fromio(buf, maddr + offset, len); |
| 6536 | ret = len; |
| 6537 | follow_pfnmap_end(&args); |
| 6538 | out_unmap: |
| 6539 | iounmap(maddr); |
| 6540 | |
| 6541 | return ret; |
| 6542 | } |
| 6543 | EXPORT_SYMBOL_GPL(generic_access_phys); |
| 6544 | #endif |
| 6545 | |
| 6546 | /* |
| 6547 | * Access another process' address space as given in mm. |
| 6548 | */ |
| 6549 | static int __access_remote_vm(struct mm_struct *mm, unsigned long addr, |
| 6550 | void *buf, int len, unsigned int gup_flags) |
| 6551 | { |
| 6552 | void *old_buf = buf; |
| 6553 | int write = gup_flags & FOLL_WRITE; |
| 6554 | |
| 6555 | if (mmap_read_lock_killable(mm)) |
| 6556 | return 0; |
| 6557 | |
| 6558 | /* Untag the address before looking up the VMA */ |
| 6559 | addr = untagged_addr_remote(mm, addr); |
| 6560 | |
| 6561 | /* Avoid triggering the temporary warning in __get_user_pages */ |
| 6562 | if (!vma_lookup(mm, addr) && !expand_stack(mm, addr)) |
| 6563 | return 0; |
| 6564 | |
| 6565 | /* ignore errors, just check how much was successfully transferred */ |
| 6566 | while (len) { |
| 6567 | int bytes, offset; |
| 6568 | void *maddr; |
| 6569 | struct vm_area_struct *vma = NULL; |
| 6570 | struct page *page = get_user_page_vma_remote(mm, addr, |
| 6571 | gup_flags, &vma); |
| 6572 | |
| 6573 | if (IS_ERR(page)) { |
| 6574 | /* We might need to expand the stack to access it */ |
| 6575 | vma = vma_lookup(mm, addr); |
| 6576 | if (!vma) { |
| 6577 | vma = expand_stack(mm, addr); |
| 6578 | |
| 6579 | /* mmap_lock was dropped on failure */ |
| 6580 | if (!vma) |
| 6581 | return buf - old_buf; |
| 6582 | |
| 6583 | /* Try again if stack expansion worked */ |
| 6584 | continue; |
| 6585 | } |
| 6586 | |
| 6587 | /* |
| 6588 | * Check if this is a VM_IO | VM_PFNMAP VMA, which |
| 6589 | * we can access using slightly different code. |
| 6590 | */ |
| 6591 | bytes = 0; |
| 6592 | #ifdef CONFIG_HAVE_IOREMAP_PROT |
| 6593 | if (vma->vm_ops && vma->vm_ops->access) |
| 6594 | bytes = vma->vm_ops->access(vma, addr, buf, |
| 6595 | len, write); |
| 6596 | #endif |
| 6597 | if (bytes <= 0) |
| 6598 | break; |
| 6599 | } else { |
| 6600 | bytes = len; |
| 6601 | offset = addr & (PAGE_SIZE-1); |
| 6602 | if (bytes > PAGE_SIZE-offset) |
| 6603 | bytes = PAGE_SIZE-offset; |
| 6604 | |
| 6605 | maddr = kmap_local_page(page); |
| 6606 | if (write) { |
| 6607 | copy_to_user_page(vma, page, addr, |
| 6608 | maddr + offset, buf, bytes); |
| 6609 | set_page_dirty_lock(page); |
| 6610 | } else { |
| 6611 | copy_from_user_page(vma, page, addr, |
| 6612 | buf, maddr + offset, bytes); |
| 6613 | } |
| 6614 | unmap_and_put_page(page, maddr); |
| 6615 | } |
| 6616 | len -= bytes; |
| 6617 | buf += bytes; |
| 6618 | addr += bytes; |
| 6619 | } |
| 6620 | mmap_read_unlock(mm); |
| 6621 | |
| 6622 | return buf - old_buf; |
| 6623 | } |
| 6624 | |
| 6625 | /** |
| 6626 | * access_remote_vm - access another process' address space |
| 6627 | * @mm: the mm_struct of the target address space |
| 6628 | * @addr: start address to access |
| 6629 | * @buf: source or destination buffer |
| 6630 | * @len: number of bytes to transfer |
| 6631 | * @gup_flags: flags modifying lookup behaviour |
| 6632 | * |
| 6633 | * The caller must hold a reference on @mm. |
| 6634 | * |
| 6635 | * Return: number of bytes copied from source to destination. |
| 6636 | */ |
| 6637 | int access_remote_vm(struct mm_struct *mm, unsigned long addr, |
| 6638 | void *buf, int len, unsigned int gup_flags) |
| 6639 | { |
| 6640 | return __access_remote_vm(mm, addr, buf, len, gup_flags); |
| 6641 | } |
| 6642 | |
| 6643 | /* |
| 6644 | * Access another process' address space. |
| 6645 | * Source/target buffer must be kernel space, |
| 6646 | * Do not walk the page table directly, use get_user_pages |
| 6647 | */ |
| 6648 | int access_process_vm(struct task_struct *tsk, unsigned long addr, |
| 6649 | void *buf, int len, unsigned int gup_flags) |
| 6650 | { |
| 6651 | struct mm_struct *mm; |
| 6652 | int ret; |
| 6653 | |
| 6654 | mm = get_task_mm(tsk); |
| 6655 | if (!mm) |
| 6656 | return 0; |
| 6657 | |
| 6658 | ret = __access_remote_vm(mm, addr, buf, len, gup_flags); |
| 6659 | |
| 6660 | mmput(mm); |
| 6661 | |
| 6662 | return ret; |
| 6663 | } |
| 6664 | EXPORT_SYMBOL_GPL(access_process_vm); |
| 6665 | |
| 6666 | /* |
| 6667 | * Print the name of a VMA. |
| 6668 | */ |
| 6669 | void print_vma_addr(char *prefix, unsigned long ip) |
| 6670 | { |
| 6671 | struct mm_struct *mm = current->mm; |
| 6672 | struct vm_area_struct *vma; |
| 6673 | |
| 6674 | /* |
| 6675 | * we might be running from an atomic context so we cannot sleep |
| 6676 | */ |
| 6677 | if (!mmap_read_trylock(mm)) |
| 6678 | return; |
| 6679 | |
| 6680 | vma = vma_lookup(mm, ip); |
| 6681 | if (vma && vma->vm_file) { |
| 6682 | struct file *f = vma->vm_file; |
| 6683 | ip -= vma->vm_start; |
| 6684 | ip += vma->vm_pgoff << PAGE_SHIFT; |
| 6685 | printk("%s%pD[%lx,%lx+%lx]", prefix, f, ip, |
| 6686 | vma->vm_start, |
| 6687 | vma->vm_end - vma->vm_start); |
| 6688 | } |
| 6689 | mmap_read_unlock(mm); |
| 6690 | } |
| 6691 | |
| 6692 | #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) |
| 6693 | void __might_fault(const char *file, int line) |
| 6694 | { |
| 6695 | if (pagefault_disabled()) |
| 6696 | return; |
| 6697 | __might_sleep(file, line); |
| 6698 | #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) |
| 6699 | if (current->mm) |
| 6700 | might_lock_read(¤t->mm->mmap_lock); |
| 6701 | #endif |
| 6702 | } |
| 6703 | EXPORT_SYMBOL(__might_fault); |
| 6704 | #endif |
| 6705 | |
| 6706 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) |
| 6707 | /* |
| 6708 | * Process all subpages of the specified huge page with the specified |
| 6709 | * operation. The target subpage will be processed last to keep its |
| 6710 | * cache lines hot. |
| 6711 | */ |
| 6712 | static inline int process_huge_page( |
| 6713 | unsigned long addr_hint, unsigned int nr_pages, |
| 6714 | int (*process_subpage)(unsigned long addr, int idx, void *arg), |
| 6715 | void *arg) |
| 6716 | { |
| 6717 | int i, n, base, l, ret; |
| 6718 | unsigned long addr = addr_hint & |
| 6719 | ~(((unsigned long)nr_pages << PAGE_SHIFT) - 1); |
| 6720 | |
| 6721 | /* Process target subpage last to keep its cache lines hot */ |
| 6722 | might_sleep(); |
| 6723 | n = (addr_hint - addr) / PAGE_SIZE; |
| 6724 | if (2 * n <= nr_pages) { |
| 6725 | /* If target subpage in first half of huge page */ |
| 6726 | base = 0; |
| 6727 | l = n; |
| 6728 | /* Process subpages at the end of huge page */ |
| 6729 | for (i = nr_pages - 1; i >= 2 * n; i--) { |
| 6730 | cond_resched(); |
| 6731 | ret = process_subpage(addr + i * PAGE_SIZE, i, arg); |
| 6732 | if (ret) |
| 6733 | return ret; |
| 6734 | } |
| 6735 | } else { |
| 6736 | /* If target subpage in second half of huge page */ |
| 6737 | base = nr_pages - 2 * (nr_pages - n); |
| 6738 | l = nr_pages - n; |
| 6739 | /* Process subpages at the begin of huge page */ |
| 6740 | for (i = 0; i < base; i++) { |
| 6741 | cond_resched(); |
| 6742 | ret = process_subpage(addr + i * PAGE_SIZE, i, arg); |
| 6743 | if (ret) |
| 6744 | return ret; |
| 6745 | } |
| 6746 | } |
| 6747 | /* |
| 6748 | * Process remaining subpages in left-right-left-right pattern |
| 6749 | * towards the target subpage |
| 6750 | */ |
| 6751 | for (i = 0; i < l; i++) { |
| 6752 | int left_idx = base + i; |
| 6753 | int right_idx = base + 2 * l - 1 - i; |
| 6754 | |
| 6755 | cond_resched(); |
| 6756 | ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); |
| 6757 | if (ret) |
| 6758 | return ret; |
| 6759 | cond_resched(); |
| 6760 | ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); |
| 6761 | if (ret) |
| 6762 | return ret; |
| 6763 | } |
| 6764 | return 0; |
| 6765 | } |
| 6766 | |
| 6767 | static void clear_gigantic_page(struct folio *folio, unsigned long addr, |
| 6768 | unsigned int nr_pages) |
| 6769 | { |
| 6770 | int i; |
| 6771 | |
| 6772 | might_sleep(); |
| 6773 | for (i = 0; i < nr_pages; i++) { |
| 6774 | cond_resched(); |
| 6775 | clear_user_highpage(folio_page(folio, i), addr + i * PAGE_SIZE); |
| 6776 | } |
| 6777 | } |
| 6778 | |
| 6779 | static int clear_subpage(unsigned long addr, int idx, void *arg) |
| 6780 | { |
| 6781 | struct folio *folio = arg; |
| 6782 | |
| 6783 | clear_user_highpage(folio_page(folio, idx), addr); |
| 6784 | return 0; |
| 6785 | } |
| 6786 | |
| 6787 | /** |
| 6788 | * folio_zero_user - Zero a folio which will be mapped to userspace. |
| 6789 | * @folio: The folio to zero. |
| 6790 | * @addr_hint: The address will be accessed or the base address if uncelar. |
| 6791 | */ |
| 6792 | void folio_zero_user(struct folio *folio, unsigned long addr_hint) |
| 6793 | { |
| 6794 | unsigned int nr_pages = folio_nr_pages(folio); |
| 6795 | |
| 6796 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) |
| 6797 | clear_gigantic_page(folio, addr_hint, nr_pages); |
| 6798 | else |
| 6799 | process_huge_page(addr_hint, nr_pages, clear_subpage, folio); |
| 6800 | } |
| 6801 | |
| 6802 | static int copy_user_gigantic_page(struct folio *dst, struct folio *src, |
| 6803 | unsigned long addr, |
| 6804 | struct vm_area_struct *vma, |
| 6805 | unsigned int nr_pages) |
| 6806 | { |
| 6807 | int i; |
| 6808 | struct page *dst_page; |
| 6809 | struct page *src_page; |
| 6810 | |
| 6811 | for (i = 0; i < nr_pages; i++) { |
| 6812 | dst_page = folio_page(dst, i); |
| 6813 | src_page = folio_page(src, i); |
| 6814 | |
| 6815 | cond_resched(); |
| 6816 | if (copy_mc_user_highpage(dst_page, src_page, |
| 6817 | addr + i*PAGE_SIZE, vma)) |
| 6818 | return -EHWPOISON; |
| 6819 | } |
| 6820 | return 0; |
| 6821 | } |
| 6822 | |
| 6823 | struct copy_subpage_arg { |
| 6824 | struct folio *dst; |
| 6825 | struct folio *src; |
| 6826 | struct vm_area_struct *vma; |
| 6827 | }; |
| 6828 | |
| 6829 | static int copy_subpage(unsigned long addr, int idx, void *arg) |
| 6830 | { |
| 6831 | struct copy_subpage_arg *copy_arg = arg; |
| 6832 | struct page *dst = folio_page(copy_arg->dst, idx); |
| 6833 | struct page *src = folio_page(copy_arg->src, idx); |
| 6834 | |
| 6835 | if (copy_mc_user_highpage(dst, src, addr, copy_arg->vma)) |
| 6836 | return -EHWPOISON; |
| 6837 | return 0; |
| 6838 | } |
| 6839 | |
| 6840 | int copy_user_large_folio(struct folio *dst, struct folio *src, |
| 6841 | unsigned long addr_hint, struct vm_area_struct *vma) |
| 6842 | { |
| 6843 | unsigned int nr_pages = folio_nr_pages(dst); |
| 6844 | struct copy_subpage_arg arg = { |
| 6845 | .dst = dst, |
| 6846 | .src = src, |
| 6847 | .vma = vma, |
| 6848 | }; |
| 6849 | |
| 6850 | if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) |
| 6851 | return copy_user_gigantic_page(dst, src, addr_hint, vma, nr_pages); |
| 6852 | |
| 6853 | return process_huge_page(addr_hint, nr_pages, copy_subpage, &arg); |
| 6854 | } |
| 6855 | |
| 6856 | long copy_folio_from_user(struct folio *dst_folio, |
| 6857 | const void __user *usr_src, |
| 6858 | bool allow_pagefault) |
| 6859 | { |
| 6860 | void *kaddr; |
| 6861 | unsigned long i, rc = 0; |
| 6862 | unsigned int nr_pages = folio_nr_pages(dst_folio); |
| 6863 | unsigned long ret_val = nr_pages * PAGE_SIZE; |
| 6864 | struct page *subpage; |
| 6865 | |
| 6866 | for (i = 0; i < nr_pages; i++) { |
| 6867 | subpage = folio_page(dst_folio, i); |
| 6868 | kaddr = kmap_local_page(subpage); |
| 6869 | if (!allow_pagefault) |
| 6870 | pagefault_disable(); |
| 6871 | rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE); |
| 6872 | if (!allow_pagefault) |
| 6873 | pagefault_enable(); |
| 6874 | kunmap_local(kaddr); |
| 6875 | |
| 6876 | ret_val -= (PAGE_SIZE - rc); |
| 6877 | if (rc) |
| 6878 | break; |
| 6879 | |
| 6880 | flush_dcache_page(subpage); |
| 6881 | |
| 6882 | cond_resched(); |
| 6883 | } |
| 6884 | return ret_val; |
| 6885 | } |
| 6886 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ |
| 6887 | |
| 6888 | #if defined(CONFIG_SPLIT_PTE_PTLOCKS) && ALLOC_SPLIT_PTLOCKS |
| 6889 | |
| 6890 | static struct kmem_cache *page_ptl_cachep; |
| 6891 | |
| 6892 | void __init ptlock_cache_init(void) |
| 6893 | { |
| 6894 | page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, |
| 6895 | SLAB_PANIC, NULL); |
| 6896 | } |
| 6897 | |
| 6898 | bool ptlock_alloc(struct ptdesc *ptdesc) |
| 6899 | { |
| 6900 | spinlock_t *ptl; |
| 6901 | |
| 6902 | ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); |
| 6903 | if (!ptl) |
| 6904 | return false; |
| 6905 | ptdesc->ptl = ptl; |
| 6906 | return true; |
| 6907 | } |
| 6908 | |
| 6909 | void ptlock_free(struct ptdesc *ptdesc) |
| 6910 | { |
| 6911 | kmem_cache_free(page_ptl_cachep, ptdesc->ptl); |
| 6912 | } |
| 6913 | #endif |
| 6914 | |
| 6915 | void vma_pgtable_walk_begin(struct vm_area_struct *vma) |
| 6916 | { |
| 6917 | if (is_vm_hugetlb_page(vma)) |
| 6918 | hugetlb_vma_lock_read(vma); |
| 6919 | } |
| 6920 | |
| 6921 | void vma_pgtable_walk_end(struct vm_area_struct *vma) |
| 6922 | { |
| 6923 | if (is_vm_hugetlb_page(vma)) |
| 6924 | hugetlb_vma_unlock_read(vma); |
| 6925 | } |