1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2009 Red Hat, Inc. 4 */ 5 6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 7 8 #include <linux/mm.h> 9 #include <linux/sched.h> 10 #include <linux/sched/mm.h> 11 #include <linux/sched/coredump.h> 12 #include <linux/sched/numa_balancing.h> 13 #include <linux/highmem.h> 14 #include <linux/hugetlb.h> 15 #include <linux/mmu_notifier.h> 16 #include <linux/rmap.h> 17 #include <linux/swap.h> 18 #include <linux/shrinker.h> 19 #include <linux/mm_inline.h> 20 #include <linux/swapops.h> 21 #include <linux/backing-dev.h> 22 #include <linux/dax.h> 23 #include <linux/khugepaged.h> 24 #include <linux/freezer.h> 25 #include <linux/pfn_t.h> 26 #include <linux/mman.h> 27 #include <linux/memremap.h> 28 #include <linux/pagemap.h> 29 #include <linux/debugfs.h> 30 #include <linux/migrate.h> 31 #include <linux/hashtable.h> 32 #include <linux/userfaultfd_k.h> 33 #include <linux/page_idle.h> 34 #include <linux/shmem_fs.h> 35 #include <linux/oom.h> 36 #include <linux/numa.h> 37 #include <linux/page_owner.h> 38 #include <linux/sched/sysctl.h> 39 #include <linux/memory-tiers.h> 40 #include <linux/compat.h> 41 42 #include <asm/tlb.h> 43 #include <asm/pgalloc.h> 44 #include "internal.h" 45 #include "swap.h" 46 47 #define CREATE_TRACE_POINTS 48 #include <trace/events/thp.h> 49 50 /* 51 * By default, transparent hugepage support is disabled in order to avoid 52 * risking an increased memory footprint for applications that are not 53 * guaranteed to benefit from it. When transparent hugepage support is 54 * enabled, it is for all mappings, and khugepaged scans all mappings. 55 * Defrag is invoked by khugepaged hugepage allocations and by page faults 56 * for all hugepage allocations. 57 */ 58 unsigned long transparent_hugepage_flags __read_mostly = 59 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 60 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 61 #endif 62 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 63 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 64 #endif 65 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 66 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 67 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 68 69 static struct shrinker deferred_split_shrinker; 70 71 static atomic_t huge_zero_refcount; 72 struct page *huge_zero_page __read_mostly; 73 unsigned long huge_zero_pfn __read_mostly = ~0UL; 74 75 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags, 76 bool smaps, bool in_pf, bool enforce_sysfs) 77 { 78 if (!vma->vm_mm) /* vdso */ 79 return false; 80 81 if (thp_disabled_by_hw() || vma_thp_disabled(vma, vm_flags)) 82 return false; 83 84 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */ 85 if (vma_is_dax(vma)) 86 return in_pf; 87 88 /* 89 * Special VMA and hugetlb VMA. 90 * Must be checked after dax since some dax mappings may have 91 * VM_MIXEDMAP set. 92 */ 93 if (vm_flags & VM_NO_KHUGEPAGED) 94 return false; 95 96 /* 97 * Check alignment for file vma and size for both file and anon vma. 98 * 99 * Skip the check for page fault. Huge fault does the check in fault 100 * handlers. And this check is not suitable for huge PUD fault. 101 */ 102 if (!in_pf && 103 !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE))) 104 return false; 105 106 /* 107 * Enabled via shmem mount options or sysfs settings. 108 * Must be done before hugepage flags check since shmem has its 109 * own flags. 110 */ 111 if (!in_pf && shmem_file(vma->vm_file)) 112 return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff, 113 !enforce_sysfs, vma->vm_mm, vm_flags); 114 115 /* Enforce sysfs THP requirements as necessary */ 116 if (enforce_sysfs && 117 (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) && 118 !hugepage_flags_always()))) 119 return false; 120 121 /* Only regular file is valid */ 122 if (!in_pf && file_thp_enabled(vma)) 123 return true; 124 125 if (!vma_is_anonymous(vma)) 126 return false; 127 128 if (vma_is_temporary_stack(vma)) 129 return false; 130 131 /* 132 * THPeligible bit of smaps should show 1 for proper VMAs even 133 * though anon_vma is not initialized yet. 134 * 135 * Allow page fault since anon_vma may be not initialized until 136 * the first page fault. 137 */ 138 if (!vma->anon_vma) 139 return (smaps || in_pf); 140 141 return true; 142 } 143 144 static bool get_huge_zero_page(void) 145 { 146 struct page *zero_page; 147 retry: 148 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 149 return true; 150 151 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 152 HPAGE_PMD_ORDER); 153 if (!zero_page) { 154 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 155 return false; 156 } 157 preempt_disable(); 158 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 159 preempt_enable(); 160 __free_pages(zero_page, compound_order(zero_page)); 161 goto retry; 162 } 163 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); 164 165 /* We take additional reference here. It will be put back by shrinker */ 166 atomic_set(&huge_zero_refcount, 2); 167 preempt_enable(); 168 count_vm_event(THP_ZERO_PAGE_ALLOC); 169 return true; 170 } 171 172 static void put_huge_zero_page(void) 173 { 174 /* 175 * Counter should never go to zero here. Only shrinker can put 176 * last reference. 177 */ 178 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 179 } 180 181 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 182 { 183 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 184 return READ_ONCE(huge_zero_page); 185 186 if (!get_huge_zero_page()) 187 return NULL; 188 189 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 190 put_huge_zero_page(); 191 192 return READ_ONCE(huge_zero_page); 193 } 194 195 void mm_put_huge_zero_page(struct mm_struct *mm) 196 { 197 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 198 put_huge_zero_page(); 199 } 200 201 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 202 struct shrink_control *sc) 203 { 204 /* we can free zero page only if last reference remains */ 205 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 206 } 207 208 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 209 struct shrink_control *sc) 210 { 211 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 212 struct page *zero_page = xchg(&huge_zero_page, NULL); 213 BUG_ON(zero_page == NULL); 214 WRITE_ONCE(huge_zero_pfn, ~0UL); 215 __free_pages(zero_page, compound_order(zero_page)); 216 return HPAGE_PMD_NR; 217 } 218 219 return 0; 220 } 221 222 static struct shrinker huge_zero_page_shrinker = { 223 .count_objects = shrink_huge_zero_page_count, 224 .scan_objects = shrink_huge_zero_page_scan, 225 .seeks = DEFAULT_SEEKS, 226 }; 227 228 #ifdef CONFIG_SYSFS 229 static ssize_t enabled_show(struct kobject *kobj, 230 struct kobj_attribute *attr, char *buf) 231 { 232 const char *output; 233 234 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 235 output = "[always] madvise never"; 236 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 237 &transparent_hugepage_flags)) 238 output = "always [madvise] never"; 239 else 240 output = "always madvise [never]"; 241 242 return sysfs_emit(buf, "%s\n", output); 243 } 244 245 static ssize_t enabled_store(struct kobject *kobj, 246 struct kobj_attribute *attr, 247 const char *buf, size_t count) 248 { 249 ssize_t ret = count; 250 251 if (sysfs_streq(buf, "always")) { 252 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 253 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 254 } else if (sysfs_streq(buf, "madvise")) { 255 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 256 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 257 } else if (sysfs_streq(buf, "never")) { 258 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 259 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 260 } else 261 ret = -EINVAL; 262 263 if (ret > 0) { 264 int err = start_stop_khugepaged(); 265 if (err) 266 ret = err; 267 } 268 return ret; 269 } 270 271 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled); 272 273 ssize_t single_hugepage_flag_show(struct kobject *kobj, 274 struct kobj_attribute *attr, char *buf, 275 enum transparent_hugepage_flag flag) 276 { 277 return sysfs_emit(buf, "%d\n", 278 !!test_bit(flag, &transparent_hugepage_flags)); 279 } 280 281 ssize_t single_hugepage_flag_store(struct kobject *kobj, 282 struct kobj_attribute *attr, 283 const char *buf, size_t count, 284 enum transparent_hugepage_flag flag) 285 { 286 unsigned long value; 287 int ret; 288 289 ret = kstrtoul(buf, 10, &value); 290 if (ret < 0) 291 return ret; 292 if (value > 1) 293 return -EINVAL; 294 295 if (value) 296 set_bit(flag, &transparent_hugepage_flags); 297 else 298 clear_bit(flag, &transparent_hugepage_flags); 299 300 return count; 301 } 302 303 static ssize_t defrag_show(struct kobject *kobj, 304 struct kobj_attribute *attr, char *buf) 305 { 306 const char *output; 307 308 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 309 &transparent_hugepage_flags)) 310 output = "[always] defer defer+madvise madvise never"; 311 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 312 &transparent_hugepage_flags)) 313 output = "always [defer] defer+madvise madvise never"; 314 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, 315 &transparent_hugepage_flags)) 316 output = "always defer [defer+madvise] madvise never"; 317 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, 318 &transparent_hugepage_flags)) 319 output = "always defer defer+madvise [madvise] never"; 320 else 321 output = "always defer defer+madvise madvise [never]"; 322 323 return sysfs_emit(buf, "%s\n", output); 324 } 325 326 static ssize_t defrag_store(struct kobject *kobj, 327 struct kobj_attribute *attr, 328 const char *buf, size_t count) 329 { 330 if (sysfs_streq(buf, "always")) { 331 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 332 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 333 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 334 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 335 } else if (sysfs_streq(buf, "defer+madvise")) { 336 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 337 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 338 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 339 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 340 } else if (sysfs_streq(buf, "defer")) { 341 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 342 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 343 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 344 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 345 } else if (sysfs_streq(buf, "madvise")) { 346 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 347 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 348 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 349 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 350 } else if (sysfs_streq(buf, "never")) { 351 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 352 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 353 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 354 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 355 } else 356 return -EINVAL; 357 358 return count; 359 } 360 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag); 361 362 static ssize_t use_zero_page_show(struct kobject *kobj, 363 struct kobj_attribute *attr, char *buf) 364 { 365 return single_hugepage_flag_show(kobj, attr, buf, 366 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 367 } 368 static ssize_t use_zero_page_store(struct kobject *kobj, 369 struct kobj_attribute *attr, const char *buf, size_t count) 370 { 371 return single_hugepage_flag_store(kobj, attr, buf, count, 372 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 373 } 374 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page); 375 376 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 377 struct kobj_attribute *attr, char *buf) 378 { 379 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); 380 } 381 static struct kobj_attribute hpage_pmd_size_attr = 382 __ATTR_RO(hpage_pmd_size); 383 384 static struct attribute *hugepage_attr[] = { 385 &enabled_attr.attr, 386 &defrag_attr.attr, 387 &use_zero_page_attr.attr, 388 &hpage_pmd_size_attr.attr, 389 #ifdef CONFIG_SHMEM 390 &shmem_enabled_attr.attr, 391 #endif 392 NULL, 393 }; 394 395 static const struct attribute_group hugepage_attr_group = { 396 .attrs = hugepage_attr, 397 }; 398 399 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 400 { 401 int err; 402 403 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 404 if (unlikely(!*hugepage_kobj)) { 405 pr_err("failed to create transparent hugepage kobject\n"); 406 return -ENOMEM; 407 } 408 409 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 410 if (err) { 411 pr_err("failed to register transparent hugepage group\n"); 412 goto delete_obj; 413 } 414 415 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 416 if (err) { 417 pr_err("failed to register transparent hugepage group\n"); 418 goto remove_hp_group; 419 } 420 421 return 0; 422 423 remove_hp_group: 424 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 425 delete_obj: 426 kobject_put(*hugepage_kobj); 427 return err; 428 } 429 430 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 431 { 432 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 433 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 434 kobject_put(hugepage_kobj); 435 } 436 #else 437 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 438 { 439 return 0; 440 } 441 442 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 443 { 444 } 445 #endif /* CONFIG_SYSFS */ 446 447 static int __init hugepage_init(void) 448 { 449 int err; 450 struct kobject *hugepage_kobj; 451 452 if (!has_transparent_hugepage()) { 453 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED; 454 return -EINVAL; 455 } 456 457 /* 458 * hugepages can't be allocated by the buddy allocator 459 */ 460 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_ORDER); 461 /* 462 * we use page->mapping and page->index in second tail page 463 * as list_head: assuming THP order >= 2 464 */ 465 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 466 467 err = hugepage_init_sysfs(&hugepage_kobj); 468 if (err) 469 goto err_sysfs; 470 471 err = khugepaged_init(); 472 if (err) 473 goto err_slab; 474 475 err = register_shrinker(&huge_zero_page_shrinker, "thp-zero"); 476 if (err) 477 goto err_hzp_shrinker; 478 err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split"); 479 if (err) 480 goto err_split_shrinker; 481 482 /* 483 * By default disable transparent hugepages on smaller systems, 484 * where the extra memory used could hurt more than TLB overhead 485 * is likely to save. The admin can still enable it through /sys. 486 */ 487 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 488 transparent_hugepage_flags = 0; 489 return 0; 490 } 491 492 err = start_stop_khugepaged(); 493 if (err) 494 goto err_khugepaged; 495 496 return 0; 497 err_khugepaged: 498 unregister_shrinker(&deferred_split_shrinker); 499 err_split_shrinker: 500 unregister_shrinker(&huge_zero_page_shrinker); 501 err_hzp_shrinker: 502 khugepaged_destroy(); 503 err_slab: 504 hugepage_exit_sysfs(hugepage_kobj); 505 err_sysfs: 506 return err; 507 } 508 subsys_initcall(hugepage_init); 509 510 static int __init setup_transparent_hugepage(char *str) 511 { 512 int ret = 0; 513 if (!str) 514 goto out; 515 if (!strcmp(str, "always")) { 516 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 517 &transparent_hugepage_flags); 518 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 519 &transparent_hugepage_flags); 520 ret = 1; 521 } else if (!strcmp(str, "madvise")) { 522 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 523 &transparent_hugepage_flags); 524 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 525 &transparent_hugepage_flags); 526 ret = 1; 527 } else if (!strcmp(str, "never")) { 528 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 529 &transparent_hugepage_flags); 530 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 531 &transparent_hugepage_flags); 532 ret = 1; 533 } 534 out: 535 if (!ret) 536 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 537 return ret; 538 } 539 __setup("transparent_hugepage=", setup_transparent_hugepage); 540 541 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 542 { 543 if (likely(vma->vm_flags & VM_WRITE)) 544 pmd = pmd_mkwrite(pmd, vma); 545 return pmd; 546 } 547 548 #ifdef CONFIG_MEMCG 549 static inline 550 struct deferred_split *get_deferred_split_queue(struct folio *folio) 551 { 552 struct mem_cgroup *memcg = folio_memcg(folio); 553 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 554 555 if (memcg) 556 return &memcg->deferred_split_queue; 557 else 558 return &pgdat->deferred_split_queue; 559 } 560 #else 561 static inline 562 struct deferred_split *get_deferred_split_queue(struct folio *folio) 563 { 564 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 565 566 return &pgdat->deferred_split_queue; 567 } 568 #endif 569 570 void folio_prep_large_rmappable(struct folio *folio) 571 { 572 if (!folio || !folio_test_large(folio)) 573 return; 574 folio_set_large_rmappable(folio); 575 } 576 577 static inline bool is_transparent_hugepage(struct folio *folio) 578 { 579 if (!folio_test_large(folio)) 580 return false; 581 582 return is_huge_zero_page(&folio->page) || 583 folio_test_large_rmappable(folio); 584 } 585 586 static unsigned long __thp_get_unmapped_area(struct file *filp, 587 unsigned long addr, unsigned long len, 588 loff_t off, unsigned long flags, unsigned long size) 589 { 590 loff_t off_end = off + len; 591 loff_t off_align = round_up(off, size); 592 unsigned long len_pad, ret; 593 594 if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall()) 595 return 0; 596 597 if (off_end <= off_align || (off_end - off_align) < size) 598 return 0; 599 600 len_pad = len + size; 601 if (len_pad < len || (off + len_pad) < off) 602 return 0; 603 604 ret = current->mm->get_unmapped_area(filp, addr, len_pad, 605 off >> PAGE_SHIFT, flags); 606 607 /* 608 * The failure might be due to length padding. The caller will retry 609 * without the padding. 610 */ 611 if (IS_ERR_VALUE(ret)) 612 return 0; 613 614 /* 615 * Do not try to align to THP boundary if allocation at the address 616 * hint succeeds. 617 */ 618 if (ret == addr) 619 return addr; 620 621 ret += (off - ret) & (size - 1); 622 return ret; 623 } 624 625 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 626 unsigned long len, unsigned long pgoff, unsigned long flags) 627 { 628 unsigned long ret; 629 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 630 631 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 632 if (ret) 633 return ret; 634 635 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 636 } 637 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 638 639 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 640 struct page *page, gfp_t gfp) 641 { 642 struct vm_area_struct *vma = vmf->vma; 643 struct folio *folio = page_folio(page); 644 pgtable_t pgtable; 645 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 646 vm_fault_t ret = 0; 647 648 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 649 650 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { 651 folio_put(folio); 652 count_vm_event(THP_FAULT_FALLBACK); 653 count_vm_event(THP_FAULT_FALLBACK_CHARGE); 654 return VM_FAULT_FALLBACK; 655 } 656 folio_throttle_swaprate(folio, gfp); 657 658 pgtable = pte_alloc_one(vma->vm_mm); 659 if (unlikely(!pgtable)) { 660 ret = VM_FAULT_OOM; 661 goto release; 662 } 663 664 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 665 /* 666 * The memory barrier inside __folio_mark_uptodate makes sure that 667 * clear_huge_page writes become visible before the set_pmd_at() 668 * write. 669 */ 670 __folio_mark_uptodate(folio); 671 672 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 673 if (unlikely(!pmd_none(*vmf->pmd))) { 674 goto unlock_release; 675 } else { 676 pmd_t entry; 677 678 ret = check_stable_address_space(vma->vm_mm); 679 if (ret) 680 goto unlock_release; 681 682 /* Deliver the page fault to userland */ 683 if (userfaultfd_missing(vma)) { 684 spin_unlock(vmf->ptl); 685 folio_put(folio); 686 pte_free(vma->vm_mm, pgtable); 687 ret = handle_userfault(vmf, VM_UFFD_MISSING); 688 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 689 return ret; 690 } 691 692 entry = mk_huge_pmd(page, vma->vm_page_prot); 693 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 694 folio_add_new_anon_rmap(folio, vma, haddr); 695 folio_add_lru_vma(folio, vma); 696 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 697 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 698 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 699 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 700 mm_inc_nr_ptes(vma->vm_mm); 701 spin_unlock(vmf->ptl); 702 count_vm_event(THP_FAULT_ALLOC); 703 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 704 } 705 706 return 0; 707 unlock_release: 708 spin_unlock(vmf->ptl); 709 release: 710 if (pgtable) 711 pte_free(vma->vm_mm, pgtable); 712 folio_put(folio); 713 return ret; 714 715 } 716 717 /* 718 * always: directly stall for all thp allocations 719 * defer: wake kswapd and fail if not immediately available 720 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 721 * fail if not immediately available 722 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 723 * available 724 * never: never stall for any thp allocation 725 */ 726 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 727 { 728 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 729 730 /* Always do synchronous compaction */ 731 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 732 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 733 734 /* Kick kcompactd and fail quickly */ 735 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 736 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 737 738 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 739 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 740 return GFP_TRANSHUGE_LIGHT | 741 (vma_madvised ? __GFP_DIRECT_RECLAIM : 742 __GFP_KSWAPD_RECLAIM); 743 744 /* Only do synchronous compaction if madvised */ 745 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 746 return GFP_TRANSHUGE_LIGHT | 747 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 748 749 return GFP_TRANSHUGE_LIGHT; 750 } 751 752 /* Caller must hold page table lock. */ 753 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 754 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 755 struct page *zero_page) 756 { 757 pmd_t entry; 758 if (!pmd_none(*pmd)) 759 return; 760 entry = mk_pmd(zero_page, vma->vm_page_prot); 761 entry = pmd_mkhuge(entry); 762 pgtable_trans_huge_deposit(mm, pmd, pgtable); 763 set_pmd_at(mm, haddr, pmd, entry); 764 mm_inc_nr_ptes(mm); 765 } 766 767 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 768 { 769 struct vm_area_struct *vma = vmf->vma; 770 gfp_t gfp; 771 struct folio *folio; 772 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 773 774 if (!transhuge_vma_suitable(vma, haddr)) 775 return VM_FAULT_FALLBACK; 776 if (unlikely(anon_vma_prepare(vma))) 777 return VM_FAULT_OOM; 778 khugepaged_enter_vma(vma, vma->vm_flags); 779 780 if (!(vmf->flags & FAULT_FLAG_WRITE) && 781 !mm_forbids_zeropage(vma->vm_mm) && 782 transparent_hugepage_use_zero_page()) { 783 pgtable_t pgtable; 784 struct page *zero_page; 785 vm_fault_t ret; 786 pgtable = pte_alloc_one(vma->vm_mm); 787 if (unlikely(!pgtable)) 788 return VM_FAULT_OOM; 789 zero_page = mm_get_huge_zero_page(vma->vm_mm); 790 if (unlikely(!zero_page)) { 791 pte_free(vma->vm_mm, pgtable); 792 count_vm_event(THP_FAULT_FALLBACK); 793 return VM_FAULT_FALLBACK; 794 } 795 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 796 ret = 0; 797 if (pmd_none(*vmf->pmd)) { 798 ret = check_stable_address_space(vma->vm_mm); 799 if (ret) { 800 spin_unlock(vmf->ptl); 801 pte_free(vma->vm_mm, pgtable); 802 } else if (userfaultfd_missing(vma)) { 803 spin_unlock(vmf->ptl); 804 pte_free(vma->vm_mm, pgtable); 805 ret = handle_userfault(vmf, VM_UFFD_MISSING); 806 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 807 } else { 808 set_huge_zero_page(pgtable, vma->vm_mm, vma, 809 haddr, vmf->pmd, zero_page); 810 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 811 spin_unlock(vmf->ptl); 812 } 813 } else { 814 spin_unlock(vmf->ptl); 815 pte_free(vma->vm_mm, pgtable); 816 } 817 return ret; 818 } 819 gfp = vma_thp_gfp_mask(vma); 820 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true); 821 if (unlikely(!folio)) { 822 count_vm_event(THP_FAULT_FALLBACK); 823 return VM_FAULT_FALLBACK; 824 } 825 return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp); 826 } 827 828 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 829 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 830 pgtable_t pgtable) 831 { 832 struct mm_struct *mm = vma->vm_mm; 833 pmd_t entry; 834 spinlock_t *ptl; 835 836 ptl = pmd_lock(mm, pmd); 837 if (!pmd_none(*pmd)) { 838 if (write) { 839 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 840 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 841 goto out_unlock; 842 } 843 entry = pmd_mkyoung(*pmd); 844 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 845 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 846 update_mmu_cache_pmd(vma, addr, pmd); 847 } 848 849 goto out_unlock; 850 } 851 852 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 853 if (pfn_t_devmap(pfn)) 854 entry = pmd_mkdevmap(entry); 855 if (write) { 856 entry = pmd_mkyoung(pmd_mkdirty(entry)); 857 entry = maybe_pmd_mkwrite(entry, vma); 858 } 859 860 if (pgtable) { 861 pgtable_trans_huge_deposit(mm, pmd, pgtable); 862 mm_inc_nr_ptes(mm); 863 pgtable = NULL; 864 } 865 866 set_pmd_at(mm, addr, pmd, entry); 867 update_mmu_cache_pmd(vma, addr, pmd); 868 869 out_unlock: 870 spin_unlock(ptl); 871 if (pgtable) 872 pte_free(mm, pgtable); 873 } 874 875 /** 876 * vmf_insert_pfn_pmd - insert a pmd size pfn 877 * @vmf: Structure describing the fault 878 * @pfn: pfn to insert 879 * @write: whether it's a write fault 880 * 881 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info. 882 * 883 * Return: vm_fault_t value. 884 */ 885 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write) 886 { 887 unsigned long addr = vmf->address & PMD_MASK; 888 struct vm_area_struct *vma = vmf->vma; 889 pgprot_t pgprot = vma->vm_page_prot; 890 pgtable_t pgtable = NULL; 891 892 /* 893 * If we had pmd_special, we could avoid all these restrictions, 894 * but we need to be consistent with PTEs and architectures that 895 * can't support a 'special' bit. 896 */ 897 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 898 !pfn_t_devmap(pfn)); 899 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 900 (VM_PFNMAP|VM_MIXEDMAP)); 901 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 902 903 if (addr < vma->vm_start || addr >= vma->vm_end) 904 return VM_FAULT_SIGBUS; 905 906 if (arch_needs_pgtable_deposit()) { 907 pgtable = pte_alloc_one(vma->vm_mm); 908 if (!pgtable) 909 return VM_FAULT_OOM; 910 } 911 912 track_pfn_insert(vma, &pgprot, pfn); 913 914 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 915 return VM_FAULT_NOPAGE; 916 } 917 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 918 919 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 920 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 921 { 922 if (likely(vma->vm_flags & VM_WRITE)) 923 pud = pud_mkwrite(pud); 924 return pud; 925 } 926 927 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 928 pud_t *pud, pfn_t pfn, bool write) 929 { 930 struct mm_struct *mm = vma->vm_mm; 931 pgprot_t prot = vma->vm_page_prot; 932 pud_t entry; 933 spinlock_t *ptl; 934 935 ptl = pud_lock(mm, pud); 936 if (!pud_none(*pud)) { 937 if (write) { 938 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 939 WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 940 goto out_unlock; 941 } 942 entry = pud_mkyoung(*pud); 943 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 944 if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 945 update_mmu_cache_pud(vma, addr, pud); 946 } 947 goto out_unlock; 948 } 949 950 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 951 if (pfn_t_devmap(pfn)) 952 entry = pud_mkdevmap(entry); 953 if (write) { 954 entry = pud_mkyoung(pud_mkdirty(entry)); 955 entry = maybe_pud_mkwrite(entry, vma); 956 } 957 set_pud_at(mm, addr, pud, entry); 958 update_mmu_cache_pud(vma, addr, pud); 959 960 out_unlock: 961 spin_unlock(ptl); 962 } 963 964 /** 965 * vmf_insert_pfn_pud - insert a pud size pfn 966 * @vmf: Structure describing the fault 967 * @pfn: pfn to insert 968 * @write: whether it's a write fault 969 * 970 * Insert a pud size pfn. See vmf_insert_pfn() for additional info. 971 * 972 * Return: vm_fault_t value. 973 */ 974 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write) 975 { 976 unsigned long addr = vmf->address & PUD_MASK; 977 struct vm_area_struct *vma = vmf->vma; 978 pgprot_t pgprot = vma->vm_page_prot; 979 980 /* 981 * If we had pud_special, we could avoid all these restrictions, 982 * but we need to be consistent with PTEs and architectures that 983 * can't support a 'special' bit. 984 */ 985 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 986 !pfn_t_devmap(pfn)); 987 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 988 (VM_PFNMAP|VM_MIXEDMAP)); 989 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 990 991 if (addr < vma->vm_start || addr >= vma->vm_end) 992 return VM_FAULT_SIGBUS; 993 994 track_pfn_insert(vma, &pgprot, pfn); 995 996 insert_pfn_pud(vma, addr, vmf->pud, pfn, write); 997 return VM_FAULT_NOPAGE; 998 } 999 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 1000 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1001 1002 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 1003 pmd_t *pmd, bool write) 1004 { 1005 pmd_t _pmd; 1006 1007 _pmd = pmd_mkyoung(*pmd); 1008 if (write) 1009 _pmd = pmd_mkdirty(_pmd); 1010 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1011 pmd, _pmd, write)) 1012 update_mmu_cache_pmd(vma, addr, pmd); 1013 } 1014 1015 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 1016 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 1017 { 1018 unsigned long pfn = pmd_pfn(*pmd); 1019 struct mm_struct *mm = vma->vm_mm; 1020 struct page *page; 1021 int ret; 1022 1023 assert_spin_locked(pmd_lockptr(mm, pmd)); 1024 1025 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 1026 return NULL; 1027 1028 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 1029 /* pass */; 1030 else 1031 return NULL; 1032 1033 if (flags & FOLL_TOUCH) 1034 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 1035 1036 /* 1037 * device mapped pages can only be returned if the 1038 * caller will manage the page reference count. 1039 */ 1040 if (!(flags & (FOLL_GET | FOLL_PIN))) 1041 return ERR_PTR(-EEXIST); 1042 1043 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1044 *pgmap = get_dev_pagemap(pfn, *pgmap); 1045 if (!*pgmap) 1046 return ERR_PTR(-EFAULT); 1047 page = pfn_to_page(pfn); 1048 ret = try_grab_folio(page_folio(page), 1, flags); 1049 if (ret) 1050 page = ERR_PTR(ret); 1051 1052 return page; 1053 } 1054 1055 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1056 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1057 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1058 { 1059 spinlock_t *dst_ptl, *src_ptl; 1060 struct page *src_page; 1061 pmd_t pmd; 1062 pgtable_t pgtable = NULL; 1063 int ret = -ENOMEM; 1064 1065 /* Skip if can be re-fill on fault */ 1066 if (!vma_is_anonymous(dst_vma)) 1067 return 0; 1068 1069 pgtable = pte_alloc_one(dst_mm); 1070 if (unlikely(!pgtable)) 1071 goto out; 1072 1073 dst_ptl = pmd_lock(dst_mm, dst_pmd); 1074 src_ptl = pmd_lockptr(src_mm, src_pmd); 1075 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1076 1077 ret = -EAGAIN; 1078 pmd = *src_pmd; 1079 1080 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1081 if (unlikely(is_swap_pmd(pmd))) { 1082 swp_entry_t entry = pmd_to_swp_entry(pmd); 1083 1084 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 1085 if (!is_readable_migration_entry(entry)) { 1086 entry = make_readable_migration_entry( 1087 swp_offset(entry)); 1088 pmd = swp_entry_to_pmd(entry); 1089 if (pmd_swp_soft_dirty(*src_pmd)) 1090 pmd = pmd_swp_mksoft_dirty(pmd); 1091 if (pmd_swp_uffd_wp(*src_pmd)) 1092 pmd = pmd_swp_mkuffd_wp(pmd); 1093 set_pmd_at(src_mm, addr, src_pmd, pmd); 1094 } 1095 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1096 mm_inc_nr_ptes(dst_mm); 1097 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1098 if (!userfaultfd_wp(dst_vma)) 1099 pmd = pmd_swp_clear_uffd_wp(pmd); 1100 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1101 ret = 0; 1102 goto out_unlock; 1103 } 1104 #endif 1105 1106 if (unlikely(!pmd_trans_huge(pmd))) { 1107 pte_free(dst_mm, pgtable); 1108 goto out_unlock; 1109 } 1110 /* 1111 * When page table lock is held, the huge zero pmd should not be 1112 * under splitting since we don't split the page itself, only pmd to 1113 * a page table. 1114 */ 1115 if (is_huge_zero_pmd(pmd)) { 1116 /* 1117 * get_huge_zero_page() will never allocate a new page here, 1118 * since we already have a zero page to copy. It just takes a 1119 * reference. 1120 */ 1121 mm_get_huge_zero_page(dst_mm); 1122 goto out_zero_page; 1123 } 1124 1125 src_page = pmd_page(pmd); 1126 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 1127 1128 get_page(src_page); 1129 if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) { 1130 /* Page maybe pinned: split and retry the fault on PTEs. */ 1131 put_page(src_page); 1132 pte_free(dst_mm, pgtable); 1133 spin_unlock(src_ptl); 1134 spin_unlock(dst_ptl); 1135 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL); 1136 return -EAGAIN; 1137 } 1138 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1139 out_zero_page: 1140 mm_inc_nr_ptes(dst_mm); 1141 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1142 pmdp_set_wrprotect(src_mm, addr, src_pmd); 1143 if (!userfaultfd_wp(dst_vma)) 1144 pmd = pmd_clear_uffd_wp(pmd); 1145 pmd = pmd_mkold(pmd_wrprotect(pmd)); 1146 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1147 1148 ret = 0; 1149 out_unlock: 1150 spin_unlock(src_ptl); 1151 spin_unlock(dst_ptl); 1152 out: 1153 return ret; 1154 } 1155 1156 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1157 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1158 pud_t *pud, bool write) 1159 { 1160 pud_t _pud; 1161 1162 _pud = pud_mkyoung(*pud); 1163 if (write) 1164 _pud = pud_mkdirty(_pud); 1165 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1166 pud, _pud, write)) 1167 update_mmu_cache_pud(vma, addr, pud); 1168 } 1169 1170 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1171 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1172 { 1173 unsigned long pfn = pud_pfn(*pud); 1174 struct mm_struct *mm = vma->vm_mm; 1175 struct page *page; 1176 int ret; 1177 1178 assert_spin_locked(pud_lockptr(mm, pud)); 1179 1180 if (flags & FOLL_WRITE && !pud_write(*pud)) 1181 return NULL; 1182 1183 if (pud_present(*pud) && pud_devmap(*pud)) 1184 /* pass */; 1185 else 1186 return NULL; 1187 1188 if (flags & FOLL_TOUCH) 1189 touch_pud(vma, addr, pud, flags & FOLL_WRITE); 1190 1191 /* 1192 * device mapped pages can only be returned if the 1193 * caller will manage the page reference count. 1194 * 1195 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1196 */ 1197 if (!(flags & (FOLL_GET | FOLL_PIN))) 1198 return ERR_PTR(-EEXIST); 1199 1200 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1201 *pgmap = get_dev_pagemap(pfn, *pgmap); 1202 if (!*pgmap) 1203 return ERR_PTR(-EFAULT); 1204 page = pfn_to_page(pfn); 1205 1206 ret = try_grab_folio(page_folio(page), 1, flags); 1207 if (ret) 1208 page = ERR_PTR(ret); 1209 1210 return page; 1211 } 1212 1213 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1214 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1215 struct vm_area_struct *vma) 1216 { 1217 spinlock_t *dst_ptl, *src_ptl; 1218 pud_t pud; 1219 int ret; 1220 1221 dst_ptl = pud_lock(dst_mm, dst_pud); 1222 src_ptl = pud_lockptr(src_mm, src_pud); 1223 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1224 1225 ret = -EAGAIN; 1226 pud = *src_pud; 1227 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1228 goto out_unlock; 1229 1230 /* 1231 * When page table lock is held, the huge zero pud should not be 1232 * under splitting since we don't split the page itself, only pud to 1233 * a page table. 1234 */ 1235 if (is_huge_zero_pud(pud)) { 1236 /* No huge zero pud yet */ 1237 } 1238 1239 /* 1240 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap() 1241 * and split if duplicating fails. 1242 */ 1243 pudp_set_wrprotect(src_mm, addr, src_pud); 1244 pud = pud_mkold(pud_wrprotect(pud)); 1245 set_pud_at(dst_mm, addr, dst_pud, pud); 1246 1247 ret = 0; 1248 out_unlock: 1249 spin_unlock(src_ptl); 1250 spin_unlock(dst_ptl); 1251 return ret; 1252 } 1253 1254 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1255 { 1256 bool write = vmf->flags & FAULT_FLAG_WRITE; 1257 1258 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1259 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1260 goto unlock; 1261 1262 touch_pud(vmf->vma, vmf->address, vmf->pud, write); 1263 unlock: 1264 spin_unlock(vmf->ptl); 1265 } 1266 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1267 1268 void huge_pmd_set_accessed(struct vm_fault *vmf) 1269 { 1270 bool write = vmf->flags & FAULT_FLAG_WRITE; 1271 1272 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1273 if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd))) 1274 goto unlock; 1275 1276 touch_pmd(vmf->vma, vmf->address, vmf->pmd, write); 1277 1278 unlock: 1279 spin_unlock(vmf->ptl); 1280 } 1281 1282 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf) 1283 { 1284 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 1285 struct vm_area_struct *vma = vmf->vma; 1286 struct folio *folio; 1287 struct page *page; 1288 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1289 pmd_t orig_pmd = vmf->orig_pmd; 1290 1291 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1292 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1293 1294 if (is_huge_zero_pmd(orig_pmd)) 1295 goto fallback; 1296 1297 spin_lock(vmf->ptl); 1298 1299 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1300 spin_unlock(vmf->ptl); 1301 return 0; 1302 } 1303 1304 page = pmd_page(orig_pmd); 1305 folio = page_folio(page); 1306 VM_BUG_ON_PAGE(!PageHead(page), page); 1307 1308 /* Early check when only holding the PT lock. */ 1309 if (PageAnonExclusive(page)) 1310 goto reuse; 1311 1312 if (!folio_trylock(folio)) { 1313 folio_get(folio); 1314 spin_unlock(vmf->ptl); 1315 folio_lock(folio); 1316 spin_lock(vmf->ptl); 1317 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1318 spin_unlock(vmf->ptl); 1319 folio_unlock(folio); 1320 folio_put(folio); 1321 return 0; 1322 } 1323 folio_put(folio); 1324 } 1325 1326 /* Recheck after temporarily dropping the PT lock. */ 1327 if (PageAnonExclusive(page)) { 1328 folio_unlock(folio); 1329 goto reuse; 1330 } 1331 1332 /* 1333 * See do_wp_page(): we can only reuse the folio exclusively if 1334 * there are no additional references. Note that we always drain 1335 * the LRU cache immediately after adding a THP. 1336 */ 1337 if (folio_ref_count(folio) > 1338 1 + folio_test_swapcache(folio) * folio_nr_pages(folio)) 1339 goto unlock_fallback; 1340 if (folio_test_swapcache(folio)) 1341 folio_free_swap(folio); 1342 if (folio_ref_count(folio) == 1) { 1343 pmd_t entry; 1344 1345 page_move_anon_rmap(page, vma); 1346 folio_unlock(folio); 1347 reuse: 1348 if (unlikely(unshare)) { 1349 spin_unlock(vmf->ptl); 1350 return 0; 1351 } 1352 entry = pmd_mkyoung(orig_pmd); 1353 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1354 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1355 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1356 spin_unlock(vmf->ptl); 1357 return 0; 1358 } 1359 1360 unlock_fallback: 1361 folio_unlock(folio); 1362 spin_unlock(vmf->ptl); 1363 fallback: 1364 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 1365 return VM_FAULT_FALLBACK; 1366 } 1367 1368 static inline bool can_change_pmd_writable(struct vm_area_struct *vma, 1369 unsigned long addr, pmd_t pmd) 1370 { 1371 struct page *page; 1372 1373 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE))) 1374 return false; 1375 1376 /* Don't touch entries that are not even readable (NUMA hinting). */ 1377 if (pmd_protnone(pmd)) 1378 return false; 1379 1380 /* Do we need write faults for softdirty tracking? */ 1381 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 1382 return false; 1383 1384 /* Do we need write faults for uffd-wp tracking? */ 1385 if (userfaultfd_huge_pmd_wp(vma, pmd)) 1386 return false; 1387 1388 if (!(vma->vm_flags & VM_SHARED)) { 1389 /* See can_change_pte_writable(). */ 1390 page = vm_normal_page_pmd(vma, addr, pmd); 1391 return page && PageAnon(page) && PageAnonExclusive(page); 1392 } 1393 1394 /* See can_change_pte_writable(). */ 1395 return pmd_dirty(pmd); 1396 } 1397 1398 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ 1399 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, 1400 struct vm_area_struct *vma, 1401 unsigned int flags) 1402 { 1403 /* If the pmd is writable, we can write to the page. */ 1404 if (pmd_write(pmd)) 1405 return true; 1406 1407 /* Maybe FOLL_FORCE is set to override it? */ 1408 if (!(flags & FOLL_FORCE)) 1409 return false; 1410 1411 /* But FOLL_FORCE has no effect on shared mappings */ 1412 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) 1413 return false; 1414 1415 /* ... or read-only private ones */ 1416 if (!(vma->vm_flags & VM_MAYWRITE)) 1417 return false; 1418 1419 /* ... or already writable ones that just need to take a write fault */ 1420 if (vma->vm_flags & VM_WRITE) 1421 return false; 1422 1423 /* 1424 * See can_change_pte_writable(): we broke COW and could map the page 1425 * writable if we have an exclusive anonymous page ... 1426 */ 1427 if (!page || !PageAnon(page) || !PageAnonExclusive(page)) 1428 return false; 1429 1430 /* ... and a write-fault isn't required for other reasons. */ 1431 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 1432 return false; 1433 return !userfaultfd_huge_pmd_wp(vma, pmd); 1434 } 1435 1436 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1437 unsigned long addr, 1438 pmd_t *pmd, 1439 unsigned int flags) 1440 { 1441 struct mm_struct *mm = vma->vm_mm; 1442 struct page *page; 1443 int ret; 1444 1445 assert_spin_locked(pmd_lockptr(mm, pmd)); 1446 1447 page = pmd_page(*pmd); 1448 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1449 1450 if ((flags & FOLL_WRITE) && 1451 !can_follow_write_pmd(*pmd, page, vma, flags)) 1452 return NULL; 1453 1454 /* Avoid dumping huge zero page */ 1455 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1456 return ERR_PTR(-EFAULT); 1457 1458 if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags)) 1459 return NULL; 1460 1461 if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page)) 1462 return ERR_PTR(-EMLINK); 1463 1464 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 1465 !PageAnonExclusive(page), page); 1466 1467 ret = try_grab_folio(page_folio(page), 1, flags); 1468 if (ret) 1469 return ERR_PTR(ret); 1470 1471 if (flags & FOLL_TOUCH) 1472 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 1473 1474 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1475 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1476 1477 return page; 1478 } 1479 1480 /* NUMA hinting page fault entry point for trans huge pmds */ 1481 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf) 1482 { 1483 struct vm_area_struct *vma = vmf->vma; 1484 pmd_t oldpmd = vmf->orig_pmd; 1485 pmd_t pmd; 1486 struct page *page; 1487 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1488 int page_nid = NUMA_NO_NODE; 1489 int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK); 1490 bool migrated = false, writable = false; 1491 int flags = 0; 1492 1493 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1494 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 1495 spin_unlock(vmf->ptl); 1496 return 0; 1497 } 1498 1499 pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1500 1501 /* 1502 * Detect now whether the PMD could be writable; this information 1503 * is only valid while holding the PT lock. 1504 */ 1505 writable = pmd_write(pmd); 1506 if (!writable && vma_wants_manual_pte_write_upgrade(vma) && 1507 can_change_pmd_writable(vma, vmf->address, pmd)) 1508 writable = true; 1509 1510 page = vm_normal_page_pmd(vma, haddr, pmd); 1511 if (!page) 1512 goto out_map; 1513 1514 /* See similar comment in do_numa_page for explanation */ 1515 if (!writable) 1516 flags |= TNF_NO_GROUP; 1517 1518 page_nid = page_to_nid(page); 1519 /* 1520 * For memory tiering mode, cpupid of slow memory page is used 1521 * to record page access time. So use default value. 1522 */ 1523 if (node_is_toptier(page_nid)) 1524 last_cpupid = page_cpupid_last(page); 1525 target_nid = numa_migrate_prep(page, vma, haddr, page_nid, 1526 &flags); 1527 1528 if (target_nid == NUMA_NO_NODE) { 1529 put_page(page); 1530 goto out_map; 1531 } 1532 1533 spin_unlock(vmf->ptl); 1534 writable = false; 1535 1536 migrated = migrate_misplaced_page(page, vma, target_nid); 1537 if (migrated) { 1538 flags |= TNF_MIGRATED; 1539 page_nid = target_nid; 1540 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); 1541 return 0; 1542 } 1543 1544 flags |= TNF_MIGRATE_FAIL; 1545 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1546 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 1547 spin_unlock(vmf->ptl); 1548 return 0; 1549 } 1550 out_map: 1551 /* Restore the PMD */ 1552 pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1553 pmd = pmd_mkyoung(pmd); 1554 if (writable) 1555 pmd = pmd_mkwrite(pmd, vma); 1556 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1557 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1558 spin_unlock(vmf->ptl); 1559 1560 if (page_nid != NUMA_NO_NODE) 1561 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); 1562 return 0; 1563 } 1564 1565 /* 1566 * Return true if we do MADV_FREE successfully on entire pmd page. 1567 * Otherwise, return false. 1568 */ 1569 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1570 pmd_t *pmd, unsigned long addr, unsigned long next) 1571 { 1572 spinlock_t *ptl; 1573 pmd_t orig_pmd; 1574 struct folio *folio; 1575 struct mm_struct *mm = tlb->mm; 1576 bool ret = false; 1577 1578 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1579 1580 ptl = pmd_trans_huge_lock(pmd, vma); 1581 if (!ptl) 1582 goto out_unlocked; 1583 1584 orig_pmd = *pmd; 1585 if (is_huge_zero_pmd(orig_pmd)) 1586 goto out; 1587 1588 if (unlikely(!pmd_present(orig_pmd))) { 1589 VM_BUG_ON(thp_migration_supported() && 1590 !is_pmd_migration_entry(orig_pmd)); 1591 goto out; 1592 } 1593 1594 folio = pfn_folio(pmd_pfn(orig_pmd)); 1595 /* 1596 * If other processes are mapping this folio, we couldn't discard 1597 * the folio unless they all do MADV_FREE so let's skip the folio. 1598 */ 1599 if (folio_estimated_sharers(folio) != 1) 1600 goto out; 1601 1602 if (!folio_trylock(folio)) 1603 goto out; 1604 1605 /* 1606 * If user want to discard part-pages of THP, split it so MADV_FREE 1607 * will deactivate only them. 1608 */ 1609 if (next - addr != HPAGE_PMD_SIZE) { 1610 folio_get(folio); 1611 spin_unlock(ptl); 1612 split_folio(folio); 1613 folio_unlock(folio); 1614 folio_put(folio); 1615 goto out_unlocked; 1616 } 1617 1618 if (folio_test_dirty(folio)) 1619 folio_clear_dirty(folio); 1620 folio_unlock(folio); 1621 1622 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1623 pmdp_invalidate(vma, addr, pmd); 1624 orig_pmd = pmd_mkold(orig_pmd); 1625 orig_pmd = pmd_mkclean(orig_pmd); 1626 1627 set_pmd_at(mm, addr, pmd, orig_pmd); 1628 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1629 } 1630 1631 folio_mark_lazyfree(folio); 1632 ret = true; 1633 out: 1634 spin_unlock(ptl); 1635 out_unlocked: 1636 return ret; 1637 } 1638 1639 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1640 { 1641 pgtable_t pgtable; 1642 1643 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1644 pte_free(mm, pgtable); 1645 mm_dec_nr_ptes(mm); 1646 } 1647 1648 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1649 pmd_t *pmd, unsigned long addr) 1650 { 1651 pmd_t orig_pmd; 1652 spinlock_t *ptl; 1653 1654 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1655 1656 ptl = __pmd_trans_huge_lock(pmd, vma); 1657 if (!ptl) 1658 return 0; 1659 /* 1660 * For architectures like ppc64 we look at deposited pgtable 1661 * when calling pmdp_huge_get_and_clear. So do the 1662 * pgtable_trans_huge_withdraw after finishing pmdp related 1663 * operations. 1664 */ 1665 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1666 tlb->fullmm); 1667 arch_check_zapped_pmd(vma, orig_pmd); 1668 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1669 if (vma_is_special_huge(vma)) { 1670 if (arch_needs_pgtable_deposit()) 1671 zap_deposited_table(tlb->mm, pmd); 1672 spin_unlock(ptl); 1673 } else if (is_huge_zero_pmd(orig_pmd)) { 1674 zap_deposited_table(tlb->mm, pmd); 1675 spin_unlock(ptl); 1676 } else { 1677 struct page *page = NULL; 1678 int flush_needed = 1; 1679 1680 if (pmd_present(orig_pmd)) { 1681 page = pmd_page(orig_pmd); 1682 page_remove_rmap(page, vma, true); 1683 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1684 VM_BUG_ON_PAGE(!PageHead(page), page); 1685 } else if (thp_migration_supported()) { 1686 swp_entry_t entry; 1687 1688 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1689 entry = pmd_to_swp_entry(orig_pmd); 1690 page = pfn_swap_entry_to_page(entry); 1691 flush_needed = 0; 1692 } else 1693 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1694 1695 if (PageAnon(page)) { 1696 zap_deposited_table(tlb->mm, pmd); 1697 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1698 } else { 1699 if (arch_needs_pgtable_deposit()) 1700 zap_deposited_table(tlb->mm, pmd); 1701 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1702 } 1703 1704 spin_unlock(ptl); 1705 if (flush_needed) 1706 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1707 } 1708 return 1; 1709 } 1710 1711 #ifndef pmd_move_must_withdraw 1712 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1713 spinlock_t *old_pmd_ptl, 1714 struct vm_area_struct *vma) 1715 { 1716 /* 1717 * With split pmd lock we also need to move preallocated 1718 * PTE page table if new_pmd is on different PMD page table. 1719 * 1720 * We also don't deposit and withdraw tables for file pages. 1721 */ 1722 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1723 } 1724 #endif 1725 1726 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1727 { 1728 #ifdef CONFIG_MEM_SOFT_DIRTY 1729 if (unlikely(is_pmd_migration_entry(pmd))) 1730 pmd = pmd_swp_mksoft_dirty(pmd); 1731 else if (pmd_present(pmd)) 1732 pmd = pmd_mksoft_dirty(pmd); 1733 #endif 1734 return pmd; 1735 } 1736 1737 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1738 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 1739 { 1740 spinlock_t *old_ptl, *new_ptl; 1741 pmd_t pmd; 1742 struct mm_struct *mm = vma->vm_mm; 1743 bool force_flush = false; 1744 1745 /* 1746 * The destination pmd shouldn't be established, free_pgtables() 1747 * should have released it; but move_page_tables() might have already 1748 * inserted a page table, if racing against shmem/file collapse. 1749 */ 1750 if (!pmd_none(*new_pmd)) { 1751 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1752 return false; 1753 } 1754 1755 /* 1756 * We don't have to worry about the ordering of src and dst 1757 * ptlocks because exclusive mmap_lock prevents deadlock. 1758 */ 1759 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1760 if (old_ptl) { 1761 new_ptl = pmd_lockptr(mm, new_pmd); 1762 if (new_ptl != old_ptl) 1763 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1764 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1765 if (pmd_present(pmd)) 1766 force_flush = true; 1767 VM_BUG_ON(!pmd_none(*new_pmd)); 1768 1769 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1770 pgtable_t pgtable; 1771 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1772 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1773 } 1774 pmd = move_soft_dirty_pmd(pmd); 1775 set_pmd_at(mm, new_addr, new_pmd, pmd); 1776 if (force_flush) 1777 flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1778 if (new_ptl != old_ptl) 1779 spin_unlock(new_ptl); 1780 spin_unlock(old_ptl); 1781 return true; 1782 } 1783 return false; 1784 } 1785 1786 /* 1787 * Returns 1788 * - 0 if PMD could not be locked 1789 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 1790 * or if prot_numa but THP migration is not supported 1791 * - HPAGE_PMD_NR if protections changed and TLB flush necessary 1792 */ 1793 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1794 pmd_t *pmd, unsigned long addr, pgprot_t newprot, 1795 unsigned long cp_flags) 1796 { 1797 struct mm_struct *mm = vma->vm_mm; 1798 spinlock_t *ptl; 1799 pmd_t oldpmd, entry; 1800 bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 1801 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 1802 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 1803 int ret = 1; 1804 1805 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1806 1807 if (prot_numa && !thp_migration_supported()) 1808 return 1; 1809 1810 ptl = __pmd_trans_huge_lock(pmd, vma); 1811 if (!ptl) 1812 return 0; 1813 1814 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1815 if (is_swap_pmd(*pmd)) { 1816 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1817 struct page *page = pfn_swap_entry_to_page(entry); 1818 pmd_t newpmd; 1819 1820 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1821 if (is_writable_migration_entry(entry)) { 1822 /* 1823 * A protection check is difficult so 1824 * just be safe and disable write 1825 */ 1826 if (PageAnon(page)) 1827 entry = make_readable_exclusive_migration_entry(swp_offset(entry)); 1828 else 1829 entry = make_readable_migration_entry(swp_offset(entry)); 1830 newpmd = swp_entry_to_pmd(entry); 1831 if (pmd_swp_soft_dirty(*pmd)) 1832 newpmd = pmd_swp_mksoft_dirty(newpmd); 1833 } else { 1834 newpmd = *pmd; 1835 } 1836 1837 if (uffd_wp) 1838 newpmd = pmd_swp_mkuffd_wp(newpmd); 1839 else if (uffd_wp_resolve) 1840 newpmd = pmd_swp_clear_uffd_wp(newpmd); 1841 if (!pmd_same(*pmd, newpmd)) 1842 set_pmd_at(mm, addr, pmd, newpmd); 1843 goto unlock; 1844 } 1845 #endif 1846 1847 if (prot_numa) { 1848 struct page *page; 1849 bool toptier; 1850 /* 1851 * Avoid trapping faults against the zero page. The read-only 1852 * data is likely to be read-cached on the local CPU and 1853 * local/remote hits to the zero page are not interesting. 1854 */ 1855 if (is_huge_zero_pmd(*pmd)) 1856 goto unlock; 1857 1858 if (pmd_protnone(*pmd)) 1859 goto unlock; 1860 1861 page = pmd_page(*pmd); 1862 toptier = node_is_toptier(page_to_nid(page)); 1863 /* 1864 * Skip scanning top tier node if normal numa 1865 * balancing is disabled 1866 */ 1867 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 1868 toptier) 1869 goto unlock; 1870 1871 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && 1872 !toptier) 1873 xchg_page_access_time(page, jiffies_to_msecs(jiffies)); 1874 } 1875 /* 1876 * In case prot_numa, we are under mmap_read_lock(mm). It's critical 1877 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1878 * which is also under mmap_read_lock(mm): 1879 * 1880 * CPU0: CPU1: 1881 * change_huge_pmd(prot_numa=1) 1882 * pmdp_huge_get_and_clear_notify() 1883 * madvise_dontneed() 1884 * zap_pmd_range() 1885 * pmd_trans_huge(*pmd) == 0 (without ptl) 1886 * // skip the pmd 1887 * set_pmd_at(); 1888 * // pmd is re-established 1889 * 1890 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1891 * which may break userspace. 1892 * 1893 * pmdp_invalidate_ad() is required to make sure we don't miss 1894 * dirty/young flags set by hardware. 1895 */ 1896 oldpmd = pmdp_invalidate_ad(vma, addr, pmd); 1897 1898 entry = pmd_modify(oldpmd, newprot); 1899 if (uffd_wp) 1900 entry = pmd_mkuffd_wp(entry); 1901 else if (uffd_wp_resolve) 1902 /* 1903 * Leave the write bit to be handled by PF interrupt 1904 * handler, then things like COW could be properly 1905 * handled. 1906 */ 1907 entry = pmd_clear_uffd_wp(entry); 1908 1909 /* See change_pte_range(). */ 1910 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) && 1911 can_change_pmd_writable(vma, addr, entry)) 1912 entry = pmd_mkwrite(entry, vma); 1913 1914 ret = HPAGE_PMD_NR; 1915 set_pmd_at(mm, addr, pmd, entry); 1916 1917 if (huge_pmd_needs_flush(oldpmd, entry)) 1918 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE); 1919 unlock: 1920 spin_unlock(ptl); 1921 return ret; 1922 } 1923 1924 /* 1925 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1926 * 1927 * Note that if it returns page table lock pointer, this routine returns without 1928 * unlocking page table lock. So callers must unlock it. 1929 */ 1930 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1931 { 1932 spinlock_t *ptl; 1933 ptl = pmd_lock(vma->vm_mm, pmd); 1934 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1935 pmd_devmap(*pmd))) 1936 return ptl; 1937 spin_unlock(ptl); 1938 return NULL; 1939 } 1940 1941 /* 1942 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise. 1943 * 1944 * Note that if it returns page table lock pointer, this routine returns without 1945 * unlocking page table lock. So callers must unlock it. 1946 */ 1947 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1948 { 1949 spinlock_t *ptl; 1950 1951 ptl = pud_lock(vma->vm_mm, pud); 1952 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1953 return ptl; 1954 spin_unlock(ptl); 1955 return NULL; 1956 } 1957 1958 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1959 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1960 pud_t *pud, unsigned long addr) 1961 { 1962 spinlock_t *ptl; 1963 1964 ptl = __pud_trans_huge_lock(pud, vma); 1965 if (!ptl) 1966 return 0; 1967 1968 pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm); 1969 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1970 if (vma_is_special_huge(vma)) { 1971 spin_unlock(ptl); 1972 /* No zero page support yet */ 1973 } else { 1974 /* No support for anonymous PUD pages yet */ 1975 BUG(); 1976 } 1977 return 1; 1978 } 1979 1980 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1981 unsigned long haddr) 1982 { 1983 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1984 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1985 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1986 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1987 1988 count_vm_event(THP_SPLIT_PUD); 1989 1990 pudp_huge_clear_flush(vma, haddr, pud); 1991 } 1992 1993 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 1994 unsigned long address) 1995 { 1996 spinlock_t *ptl; 1997 struct mmu_notifier_range range; 1998 1999 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 2000 address & HPAGE_PUD_MASK, 2001 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 2002 mmu_notifier_invalidate_range_start(&range); 2003 ptl = pud_lock(vma->vm_mm, pud); 2004 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2005 goto out; 2006 __split_huge_pud_locked(vma, pud, range.start); 2007 2008 out: 2009 spin_unlock(ptl); 2010 mmu_notifier_invalidate_range_end(&range); 2011 } 2012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2013 2014 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2015 unsigned long haddr, pmd_t *pmd) 2016 { 2017 struct mm_struct *mm = vma->vm_mm; 2018 pgtable_t pgtable; 2019 pmd_t _pmd, old_pmd; 2020 unsigned long addr; 2021 pte_t *pte; 2022 int i; 2023 2024 /* 2025 * Leave pmd empty until pte is filled note that it is fine to delay 2026 * notification until mmu_notifier_invalidate_range_end() as we are 2027 * replacing a zero pmd write protected page with a zero pte write 2028 * protected page. 2029 * 2030 * See Documentation/mm/mmu_notifier.rst 2031 */ 2032 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2033 2034 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2035 pmd_populate(mm, &_pmd, pgtable); 2036 2037 pte = pte_offset_map(&_pmd, haddr); 2038 VM_BUG_ON(!pte); 2039 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2040 pte_t entry; 2041 2042 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot); 2043 entry = pte_mkspecial(entry); 2044 if (pmd_uffd_wp(old_pmd)) 2045 entry = pte_mkuffd_wp(entry); 2046 VM_BUG_ON(!pte_none(ptep_get(pte))); 2047 set_pte_at(mm, addr, pte, entry); 2048 pte++; 2049 } 2050 pte_unmap(pte - 1); 2051 smp_wmb(); /* make pte visible before pmd */ 2052 pmd_populate(mm, pmd, pgtable); 2053 } 2054 2055 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2056 unsigned long haddr, bool freeze) 2057 { 2058 struct mm_struct *mm = vma->vm_mm; 2059 struct page *page; 2060 pgtable_t pgtable; 2061 pmd_t old_pmd, _pmd; 2062 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 2063 bool anon_exclusive = false, dirty = false; 2064 unsigned long addr; 2065 pte_t *pte; 2066 int i; 2067 2068 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2069 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2070 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2071 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2072 && !pmd_devmap(*pmd)); 2073 2074 count_vm_event(THP_SPLIT_PMD); 2075 2076 if (!vma_is_anonymous(vma)) { 2077 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2078 /* 2079 * We are going to unmap this huge page. So 2080 * just go ahead and zap it 2081 */ 2082 if (arch_needs_pgtable_deposit()) 2083 zap_deposited_table(mm, pmd); 2084 if (vma_is_special_huge(vma)) 2085 return; 2086 if (unlikely(is_pmd_migration_entry(old_pmd))) { 2087 swp_entry_t entry; 2088 2089 entry = pmd_to_swp_entry(old_pmd); 2090 page = pfn_swap_entry_to_page(entry); 2091 } else { 2092 page = pmd_page(old_pmd); 2093 if (!PageDirty(page) && pmd_dirty(old_pmd)) 2094 set_page_dirty(page); 2095 if (!PageReferenced(page) && pmd_young(old_pmd)) 2096 SetPageReferenced(page); 2097 page_remove_rmap(page, vma, true); 2098 put_page(page); 2099 } 2100 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2101 return; 2102 } 2103 2104 if (is_huge_zero_pmd(*pmd)) { 2105 /* 2106 * FIXME: Do we want to invalidate secondary mmu by calling 2107 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below 2108 * inside __split_huge_pmd() ? 2109 * 2110 * We are going from a zero huge page write protected to zero 2111 * small page also write protected so it does not seems useful 2112 * to invalidate secondary mmu at this time. 2113 */ 2114 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2115 } 2116 2117 pmd_migration = is_pmd_migration_entry(*pmd); 2118 if (unlikely(pmd_migration)) { 2119 swp_entry_t entry; 2120 2121 old_pmd = *pmd; 2122 entry = pmd_to_swp_entry(old_pmd); 2123 page = pfn_swap_entry_to_page(entry); 2124 write = is_writable_migration_entry(entry); 2125 if (PageAnon(page)) 2126 anon_exclusive = is_readable_exclusive_migration_entry(entry); 2127 young = is_migration_entry_young(entry); 2128 dirty = is_migration_entry_dirty(entry); 2129 soft_dirty = pmd_swp_soft_dirty(old_pmd); 2130 uffd_wp = pmd_swp_uffd_wp(old_pmd); 2131 } else { 2132 /* 2133 * Up to this point the pmd is present and huge and userland has 2134 * the whole access to the hugepage during the split (which 2135 * happens in place). If we overwrite the pmd with the not-huge 2136 * version pointing to the pte here (which of course we could if 2137 * all CPUs were bug free), userland could trigger a small page 2138 * size TLB miss on the small sized TLB while the hugepage TLB 2139 * entry is still established in the huge TLB. Some CPU doesn't 2140 * like that. See 2141 * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 2142 * 383 on page 105. Intel should be safe but is also warns that 2143 * it's only safe if the permission and cache attributes of the 2144 * two entries loaded in the two TLB is identical (which should 2145 * be the case here). But it is generally safer to never allow 2146 * small and huge TLB entries for the same virtual address to be 2147 * loaded simultaneously. So instead of doing "pmd_populate(); 2148 * flush_pmd_tlb_range();" we first mark the current pmd 2149 * notpresent (atomically because here the pmd_trans_huge must 2150 * remain set at all times on the pmd until the split is 2151 * complete for this pmd), then we flush the SMP TLB and finally 2152 * we write the non-huge version of the pmd entry with 2153 * pmd_populate. 2154 */ 2155 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2156 page = pmd_page(old_pmd); 2157 if (pmd_dirty(old_pmd)) { 2158 dirty = true; 2159 SetPageDirty(page); 2160 } 2161 write = pmd_write(old_pmd); 2162 young = pmd_young(old_pmd); 2163 soft_dirty = pmd_soft_dirty(old_pmd); 2164 uffd_wp = pmd_uffd_wp(old_pmd); 2165 2166 VM_BUG_ON_PAGE(!page_count(page), page); 2167 2168 /* 2169 * Without "freeze", we'll simply split the PMD, propagating the 2170 * PageAnonExclusive() flag for each PTE by setting it for 2171 * each subpage -- no need to (temporarily) clear. 2172 * 2173 * With "freeze" we want to replace mapped pages by 2174 * migration entries right away. This is only possible if we 2175 * managed to clear PageAnonExclusive() -- see 2176 * set_pmd_migration_entry(). 2177 * 2178 * In case we cannot clear PageAnonExclusive(), split the PMD 2179 * only and let try_to_migrate_one() fail later. 2180 * 2181 * See page_try_share_anon_rmap(): invalidate PMD first. 2182 */ 2183 anon_exclusive = PageAnon(page) && PageAnonExclusive(page); 2184 if (freeze && anon_exclusive && page_try_share_anon_rmap(page)) 2185 freeze = false; 2186 if (!freeze) 2187 page_ref_add(page, HPAGE_PMD_NR - 1); 2188 } 2189 2190 /* 2191 * Withdraw the table only after we mark the pmd entry invalid. 2192 * This's critical for some architectures (Power). 2193 */ 2194 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2195 pmd_populate(mm, &_pmd, pgtable); 2196 2197 pte = pte_offset_map(&_pmd, haddr); 2198 VM_BUG_ON(!pte); 2199 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2200 pte_t entry; 2201 /* 2202 * Note that NUMA hinting access restrictions are not 2203 * transferred to avoid any possibility of altering 2204 * permissions across VMAs. 2205 */ 2206 if (freeze || pmd_migration) { 2207 swp_entry_t swp_entry; 2208 if (write) 2209 swp_entry = make_writable_migration_entry( 2210 page_to_pfn(page + i)); 2211 else if (anon_exclusive) 2212 swp_entry = make_readable_exclusive_migration_entry( 2213 page_to_pfn(page + i)); 2214 else 2215 swp_entry = make_readable_migration_entry( 2216 page_to_pfn(page + i)); 2217 if (young) 2218 swp_entry = make_migration_entry_young(swp_entry); 2219 if (dirty) 2220 swp_entry = make_migration_entry_dirty(swp_entry); 2221 entry = swp_entry_to_pte(swp_entry); 2222 if (soft_dirty) 2223 entry = pte_swp_mksoft_dirty(entry); 2224 if (uffd_wp) 2225 entry = pte_swp_mkuffd_wp(entry); 2226 } else { 2227 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2228 if (write) 2229 entry = pte_mkwrite(entry, vma); 2230 if (anon_exclusive) 2231 SetPageAnonExclusive(page + i); 2232 if (!young) 2233 entry = pte_mkold(entry); 2234 /* NOTE: this may set soft-dirty too on some archs */ 2235 if (dirty) 2236 entry = pte_mkdirty(entry); 2237 if (soft_dirty) 2238 entry = pte_mksoft_dirty(entry); 2239 if (uffd_wp) 2240 entry = pte_mkuffd_wp(entry); 2241 page_add_anon_rmap(page + i, vma, addr, RMAP_NONE); 2242 } 2243 VM_BUG_ON(!pte_none(ptep_get(pte))); 2244 set_pte_at(mm, addr, pte, entry); 2245 pte++; 2246 } 2247 pte_unmap(pte - 1); 2248 2249 if (!pmd_migration) 2250 page_remove_rmap(page, vma, true); 2251 if (freeze) 2252 put_page(page); 2253 2254 smp_wmb(); /* make pte visible before pmd */ 2255 pmd_populate(mm, pmd, pgtable); 2256 } 2257 2258 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2259 unsigned long address, bool freeze, struct folio *folio) 2260 { 2261 spinlock_t *ptl; 2262 struct mmu_notifier_range range; 2263 2264 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 2265 address & HPAGE_PMD_MASK, 2266 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2267 mmu_notifier_invalidate_range_start(&range); 2268 ptl = pmd_lock(vma->vm_mm, pmd); 2269 2270 /* 2271 * If caller asks to setup a migration entry, we need a folio to check 2272 * pmd against. Otherwise we can end up replacing wrong folio. 2273 */ 2274 VM_BUG_ON(freeze && !folio); 2275 VM_WARN_ON_ONCE(folio && !folio_test_locked(folio)); 2276 2277 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || 2278 is_pmd_migration_entry(*pmd)) { 2279 /* 2280 * It's safe to call pmd_page when folio is set because it's 2281 * guaranteed that pmd is present. 2282 */ 2283 if (folio && folio != page_folio(pmd_page(*pmd))) 2284 goto out; 2285 __split_huge_pmd_locked(vma, pmd, range.start, freeze); 2286 } 2287 2288 out: 2289 spin_unlock(ptl); 2290 mmu_notifier_invalidate_range_end(&range); 2291 } 2292 2293 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2294 bool freeze, struct folio *folio) 2295 { 2296 pmd_t *pmd = mm_find_pmd(vma->vm_mm, address); 2297 2298 if (!pmd) 2299 return; 2300 2301 __split_huge_pmd(vma, pmd, address, freeze, folio); 2302 } 2303 2304 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 2305 { 2306 /* 2307 * If the new address isn't hpage aligned and it could previously 2308 * contain an hugepage: check if we need to split an huge pmd. 2309 */ 2310 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 2311 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 2312 ALIGN(address, HPAGE_PMD_SIZE))) 2313 split_huge_pmd_address(vma, address, false, NULL); 2314 } 2315 2316 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2317 unsigned long start, 2318 unsigned long end, 2319 long adjust_next) 2320 { 2321 /* Check if we need to split start first. */ 2322 split_huge_pmd_if_needed(vma, start); 2323 2324 /* Check if we need to split end next. */ 2325 split_huge_pmd_if_needed(vma, end); 2326 2327 /* 2328 * If we're also updating the next vma vm_start, 2329 * check if we need to split it. 2330 */ 2331 if (adjust_next > 0) { 2332 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end); 2333 unsigned long nstart = next->vm_start; 2334 nstart += adjust_next; 2335 split_huge_pmd_if_needed(next, nstart); 2336 } 2337 } 2338 2339 static void unmap_folio(struct folio *folio) 2340 { 2341 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | 2342 TTU_SYNC; 2343 2344 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 2345 2346 /* 2347 * Anon pages need migration entries to preserve them, but file 2348 * pages can simply be left unmapped, then faulted back on demand. 2349 * If that is ever changed (perhaps for mlock), update remap_page(). 2350 */ 2351 if (folio_test_anon(folio)) 2352 try_to_migrate(folio, ttu_flags); 2353 else 2354 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK); 2355 } 2356 2357 static void remap_page(struct folio *folio, unsigned long nr) 2358 { 2359 int i = 0; 2360 2361 /* If unmap_folio() uses try_to_migrate() on file, remove this check */ 2362 if (!folio_test_anon(folio)) 2363 return; 2364 for (;;) { 2365 remove_migration_ptes(folio, folio, true); 2366 i += folio_nr_pages(folio); 2367 if (i >= nr) 2368 break; 2369 folio = folio_next(folio); 2370 } 2371 } 2372 2373 static void lru_add_page_tail(struct page *head, struct page *tail, 2374 struct lruvec *lruvec, struct list_head *list) 2375 { 2376 VM_BUG_ON_PAGE(!PageHead(head), head); 2377 VM_BUG_ON_PAGE(PageCompound(tail), head); 2378 VM_BUG_ON_PAGE(PageLRU(tail), head); 2379 lockdep_assert_held(&lruvec->lru_lock); 2380 2381 if (list) { 2382 /* page reclaim is reclaiming a huge page */ 2383 VM_WARN_ON(PageLRU(head)); 2384 get_page(tail); 2385 list_add_tail(&tail->lru, list); 2386 } else { 2387 /* head is still on lru (and we have it frozen) */ 2388 VM_WARN_ON(!PageLRU(head)); 2389 if (PageUnevictable(tail)) 2390 tail->mlock_count = 0; 2391 else 2392 list_add_tail(&tail->lru, &head->lru); 2393 SetPageLRU(tail); 2394 } 2395 } 2396 2397 static void __split_huge_page_tail(struct folio *folio, int tail, 2398 struct lruvec *lruvec, struct list_head *list) 2399 { 2400 struct page *head = &folio->page; 2401 struct page *page_tail = head + tail; 2402 /* 2403 * Careful: new_folio is not a "real" folio before we cleared PageTail. 2404 * Don't pass it around before clear_compound_head(). 2405 */ 2406 struct folio *new_folio = (struct folio *)page_tail; 2407 2408 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2409 2410 /* 2411 * Clone page flags before unfreezing refcount. 2412 * 2413 * After successful get_page_unless_zero() might follow flags change, 2414 * for example lock_page() which set PG_waiters. 2415 * 2416 * Note that for mapped sub-pages of an anonymous THP, 2417 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in 2418 * the migration entry instead from where remap_page() will restore it. 2419 * We can still have PG_anon_exclusive set on effectively unmapped and 2420 * unreferenced sub-pages of an anonymous THP: we can simply drop 2421 * PG_anon_exclusive (-> PG_mappedtodisk) for these here. 2422 */ 2423 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2424 page_tail->flags |= (head->flags & 2425 ((1L << PG_referenced) | 2426 (1L << PG_swapbacked) | 2427 (1L << PG_swapcache) | 2428 (1L << PG_mlocked) | 2429 (1L << PG_uptodate) | 2430 (1L << PG_active) | 2431 (1L << PG_workingset) | 2432 (1L << PG_locked) | 2433 (1L << PG_unevictable) | 2434 #ifdef CONFIG_ARCH_USES_PG_ARCH_X 2435 (1L << PG_arch_2) | 2436 (1L << PG_arch_3) | 2437 #endif 2438 (1L << PG_dirty) | 2439 LRU_GEN_MASK | LRU_REFS_MASK)); 2440 2441 /* ->mapping in first and second tail page is replaced by other uses */ 2442 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2443 page_tail); 2444 page_tail->mapping = head->mapping; 2445 page_tail->index = head->index + tail; 2446 2447 /* 2448 * page->private should not be set in tail pages. Fix up and warn once 2449 * if private is unexpectedly set. 2450 */ 2451 if (unlikely(page_tail->private)) { 2452 VM_WARN_ON_ONCE_PAGE(true, page_tail); 2453 page_tail->private = 0; 2454 } 2455 if (folio_test_swapcache(folio)) 2456 new_folio->swap.val = folio->swap.val + tail; 2457 2458 /* Page flags must be visible before we make the page non-compound. */ 2459 smp_wmb(); 2460 2461 /* 2462 * Clear PageTail before unfreezing page refcount. 2463 * 2464 * After successful get_page_unless_zero() might follow put_page() 2465 * which needs correct compound_head(). 2466 */ 2467 clear_compound_head(page_tail); 2468 2469 /* Finally unfreeze refcount. Additional reference from page cache. */ 2470 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2471 PageSwapCache(head))); 2472 2473 if (page_is_young(head)) 2474 set_page_young(page_tail); 2475 if (page_is_idle(head)) 2476 set_page_idle(page_tail); 2477 2478 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2479 2480 /* 2481 * always add to the tail because some iterators expect new 2482 * pages to show after the currently processed elements - e.g. 2483 * migrate_pages 2484 */ 2485 lru_add_page_tail(head, page_tail, lruvec, list); 2486 } 2487 2488 static void __split_huge_page(struct page *page, struct list_head *list, 2489 pgoff_t end) 2490 { 2491 struct folio *folio = page_folio(page); 2492 struct page *head = &folio->page; 2493 struct lruvec *lruvec; 2494 struct address_space *swap_cache = NULL; 2495 unsigned long offset = 0; 2496 unsigned int nr = thp_nr_pages(head); 2497 int i, nr_dropped = 0; 2498 2499 /* complete memcg works before add pages to LRU */ 2500 split_page_memcg(head, nr); 2501 2502 if (folio_test_anon(folio) && folio_test_swapcache(folio)) { 2503 offset = swp_offset(folio->swap); 2504 swap_cache = swap_address_space(folio->swap); 2505 xa_lock(&swap_cache->i_pages); 2506 } 2507 2508 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 2509 lruvec = folio_lruvec_lock(folio); 2510 2511 ClearPageHasHWPoisoned(head); 2512 2513 for (i = nr - 1; i >= 1; i--) { 2514 __split_huge_page_tail(folio, i, lruvec, list); 2515 /* Some pages can be beyond EOF: drop them from page cache */ 2516 if (head[i].index >= end) { 2517 struct folio *tail = page_folio(head + i); 2518 2519 if (shmem_mapping(head->mapping)) 2520 nr_dropped++; 2521 else if (folio_test_clear_dirty(tail)) 2522 folio_account_cleaned(tail, 2523 inode_to_wb(folio->mapping->host)); 2524 __filemap_remove_folio(tail, NULL); 2525 folio_put(tail); 2526 } else if (!PageAnon(page)) { 2527 __xa_store(&head->mapping->i_pages, head[i].index, 2528 head + i, 0); 2529 } else if (swap_cache) { 2530 __xa_store(&swap_cache->i_pages, offset + i, 2531 head + i, 0); 2532 } 2533 } 2534 2535 ClearPageCompound(head); 2536 unlock_page_lruvec(lruvec); 2537 /* Caller disabled irqs, so they are still disabled here */ 2538 2539 split_page_owner(head, nr); 2540 2541 /* See comment in __split_huge_page_tail() */ 2542 if (PageAnon(head)) { 2543 /* Additional pin to swap cache */ 2544 if (PageSwapCache(head)) { 2545 page_ref_add(head, 2); 2546 xa_unlock(&swap_cache->i_pages); 2547 } else { 2548 page_ref_inc(head); 2549 } 2550 } else { 2551 /* Additional pin to page cache */ 2552 page_ref_add(head, 2); 2553 xa_unlock(&head->mapping->i_pages); 2554 } 2555 local_irq_enable(); 2556 2557 if (nr_dropped) 2558 shmem_uncharge(head->mapping->host, nr_dropped); 2559 remap_page(folio, nr); 2560 2561 if (folio_test_swapcache(folio)) 2562 split_swap_cluster(folio->swap); 2563 2564 for (i = 0; i < nr; i++) { 2565 struct page *subpage = head + i; 2566 if (subpage == page) 2567 continue; 2568 unlock_page(subpage); 2569 2570 /* 2571 * Subpages may be freed if there wasn't any mapping 2572 * like if add_to_swap() is running on a lru page that 2573 * had its mapping zapped. And freeing these pages 2574 * requires taking the lru_lock so we do the put_page 2575 * of the tail pages after the split is complete. 2576 */ 2577 free_page_and_swap_cache(subpage); 2578 } 2579 } 2580 2581 /* Racy check whether the huge page can be split */ 2582 bool can_split_folio(struct folio *folio, int *pextra_pins) 2583 { 2584 int extra_pins; 2585 2586 /* Additional pins from page cache */ 2587 if (folio_test_anon(folio)) 2588 extra_pins = folio_test_swapcache(folio) ? 2589 folio_nr_pages(folio) : 0; 2590 else 2591 extra_pins = folio_nr_pages(folio); 2592 if (pextra_pins) 2593 *pextra_pins = extra_pins; 2594 return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1; 2595 } 2596 2597 /* 2598 * This function splits huge page into normal pages. @page can point to any 2599 * subpage of huge page to split. Split doesn't change the position of @page. 2600 * 2601 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2602 * The huge page must be locked. 2603 * 2604 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2605 * 2606 * Both head page and tail pages will inherit mapping, flags, and so on from 2607 * the hugepage. 2608 * 2609 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2610 * they are not mapped. 2611 * 2612 * Returns 0 if the hugepage is split successfully. 2613 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2614 * us. 2615 */ 2616 int split_huge_page_to_list(struct page *page, struct list_head *list) 2617 { 2618 struct folio *folio = page_folio(page); 2619 struct deferred_split *ds_queue = get_deferred_split_queue(folio); 2620 XA_STATE(xas, &folio->mapping->i_pages, folio->index); 2621 struct anon_vma *anon_vma = NULL; 2622 struct address_space *mapping = NULL; 2623 int extra_pins, ret; 2624 pgoff_t end; 2625 bool is_hzp; 2626 2627 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 2628 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 2629 2630 is_hzp = is_huge_zero_page(&folio->page); 2631 if (is_hzp) { 2632 pr_warn_ratelimited("Called split_huge_page for huge zero page\n"); 2633 return -EBUSY; 2634 } 2635 2636 if (folio_test_writeback(folio)) 2637 return -EBUSY; 2638 2639 if (folio_test_anon(folio)) { 2640 /* 2641 * The caller does not necessarily hold an mmap_lock that would 2642 * prevent the anon_vma disappearing so we first we take a 2643 * reference to it and then lock the anon_vma for write. This 2644 * is similar to folio_lock_anon_vma_read except the write lock 2645 * is taken to serialise against parallel split or collapse 2646 * operations. 2647 */ 2648 anon_vma = folio_get_anon_vma(folio); 2649 if (!anon_vma) { 2650 ret = -EBUSY; 2651 goto out; 2652 } 2653 end = -1; 2654 mapping = NULL; 2655 anon_vma_lock_write(anon_vma); 2656 } else { 2657 gfp_t gfp; 2658 2659 mapping = folio->mapping; 2660 2661 /* Truncated ? */ 2662 if (!mapping) { 2663 ret = -EBUSY; 2664 goto out; 2665 } 2666 2667 gfp = current_gfp_context(mapping_gfp_mask(mapping) & 2668 GFP_RECLAIM_MASK); 2669 2670 if (!filemap_release_folio(folio, gfp)) { 2671 ret = -EBUSY; 2672 goto out; 2673 } 2674 2675 xas_split_alloc(&xas, folio, folio_order(folio), gfp); 2676 if (xas_error(&xas)) { 2677 ret = xas_error(&xas); 2678 goto out; 2679 } 2680 2681 anon_vma = NULL; 2682 i_mmap_lock_read(mapping); 2683 2684 /* 2685 *__split_huge_page() may need to trim off pages beyond EOF: 2686 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2687 * which cannot be nested inside the page tree lock. So note 2688 * end now: i_size itself may be changed at any moment, but 2689 * folio lock is good enough to serialize the trimming. 2690 */ 2691 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2692 if (shmem_mapping(mapping)) 2693 end = shmem_fallocend(mapping->host, end); 2694 } 2695 2696 /* 2697 * Racy check if we can split the page, before unmap_folio() will 2698 * split PMDs 2699 */ 2700 if (!can_split_folio(folio, &extra_pins)) { 2701 ret = -EAGAIN; 2702 goto out_unlock; 2703 } 2704 2705 unmap_folio(folio); 2706 2707 /* block interrupt reentry in xa_lock and spinlock */ 2708 local_irq_disable(); 2709 if (mapping) { 2710 /* 2711 * Check if the folio is present in page cache. 2712 * We assume all tail are present too, if folio is there. 2713 */ 2714 xas_lock(&xas); 2715 xas_reset(&xas); 2716 if (xas_load(&xas) != folio) 2717 goto fail; 2718 } 2719 2720 /* Prevent deferred_split_scan() touching ->_refcount */ 2721 spin_lock(&ds_queue->split_queue_lock); 2722 if (folio_ref_freeze(folio, 1 + extra_pins)) { 2723 if (folio_order(folio) > 1 && 2724 !list_empty(&folio->_deferred_list)) { 2725 ds_queue->split_queue_len--; 2726 list_del_init(&folio->_deferred_list); 2727 } 2728 spin_unlock(&ds_queue->split_queue_lock); 2729 if (mapping) { 2730 int nr = folio_nr_pages(folio); 2731 2732 xas_split(&xas, folio, folio_order(folio)); 2733 if (folio_test_pmd_mappable(folio)) { 2734 if (folio_test_swapbacked(folio)) { 2735 __lruvec_stat_mod_folio(folio, 2736 NR_SHMEM_THPS, -nr); 2737 } else { 2738 __lruvec_stat_mod_folio(folio, 2739 NR_FILE_THPS, -nr); 2740 filemap_nr_thps_dec(mapping); 2741 } 2742 } 2743 } 2744 2745 __split_huge_page(page, list, end); 2746 ret = 0; 2747 } else { 2748 spin_unlock(&ds_queue->split_queue_lock); 2749 fail: 2750 if (mapping) 2751 xas_unlock(&xas); 2752 local_irq_enable(); 2753 remap_page(folio, folio_nr_pages(folio)); 2754 ret = -EAGAIN; 2755 } 2756 2757 out_unlock: 2758 if (anon_vma) { 2759 anon_vma_unlock_write(anon_vma); 2760 put_anon_vma(anon_vma); 2761 } 2762 if (mapping) 2763 i_mmap_unlock_read(mapping); 2764 out: 2765 xas_destroy(&xas); 2766 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2767 return ret; 2768 } 2769 2770 /* 2771 * __folio_unqueue_deferred_split() is not to be called directly: 2772 * the folio_unqueue_deferred_split() inline wrapper in mm/internal.h 2773 * limits its calls to those folios which may have a _deferred_list for 2774 * queueing THP splits, and that list is (racily observed to be) non-empty. 2775 * 2776 * It is unsafe to call folio_unqueue_deferred_split() until folio refcount is 2777 * zero: because even when split_queue_lock is held, a non-empty _deferred_list 2778 * might be in use on deferred_split_scan()'s unlocked on-stack list. 2779 * 2780 * If memory cgroups are enabled, split_queue_lock is in the mem_cgroup: it is 2781 * therefore important to unqueue deferred split before changing folio memcg. 2782 */ 2783 bool __folio_unqueue_deferred_split(struct folio *folio) 2784 { 2785 struct deferred_split *ds_queue; 2786 unsigned long flags; 2787 bool unqueued = false; 2788 2789 WARN_ON_ONCE(folio_ref_count(folio)); 2790 WARN_ON_ONCE(!mem_cgroup_disabled() && !folio_memcg(folio)); 2791 2792 ds_queue = get_deferred_split_queue(folio); 2793 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2794 if (!list_empty(&folio->_deferred_list)) { 2795 ds_queue->split_queue_len--; 2796 list_del_init(&folio->_deferred_list); 2797 unqueued = true; 2798 } 2799 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2800 2801 return unqueued; /* useful for debug warnings */ 2802 } 2803 2804 void deferred_split_folio(struct folio *folio) 2805 { 2806 struct deferred_split *ds_queue = get_deferred_split_queue(folio); 2807 #ifdef CONFIG_MEMCG 2808 struct mem_cgroup *memcg = folio_memcg(folio); 2809 #endif 2810 unsigned long flags; 2811 2812 /* 2813 * Order 1 folios have no space for a deferred list, but we also 2814 * won't waste much memory by not adding them to the deferred list. 2815 */ 2816 if (folio_order(folio) <= 1) 2817 return; 2818 2819 /* 2820 * Exclude swapcache: originally to avoid a corrupt deferred split 2821 * queue. Nowadays that is fully prevented by mem_cgroup_swapout(); 2822 * but if page reclaim is already handling the same folio, it is 2823 * unnecessary to handle it again in the shrinker, so excluding 2824 * swapcache here may still be a useful optimization. 2825 */ 2826 if (folio_test_swapcache(folio)) 2827 return; 2828 2829 if (!list_empty(&folio->_deferred_list)) 2830 return; 2831 2832 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2833 if (list_empty(&folio->_deferred_list)) { 2834 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2835 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue); 2836 ds_queue->split_queue_len++; 2837 #ifdef CONFIG_MEMCG 2838 if (memcg) 2839 set_shrinker_bit(memcg, folio_nid(folio), 2840 deferred_split_shrinker.id); 2841 #endif 2842 } 2843 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2844 } 2845 2846 static unsigned long deferred_split_count(struct shrinker *shrink, 2847 struct shrink_control *sc) 2848 { 2849 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2850 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2851 2852 #ifdef CONFIG_MEMCG 2853 if (sc->memcg) 2854 ds_queue = &sc->memcg->deferred_split_queue; 2855 #endif 2856 return READ_ONCE(ds_queue->split_queue_len); 2857 } 2858 2859 static unsigned long deferred_split_scan(struct shrinker *shrink, 2860 struct shrink_control *sc) 2861 { 2862 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2863 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2864 unsigned long flags; 2865 LIST_HEAD(list); 2866 struct folio *folio, *next; 2867 int split = 0; 2868 2869 #ifdef CONFIG_MEMCG 2870 if (sc->memcg) 2871 ds_queue = &sc->memcg->deferred_split_queue; 2872 #endif 2873 2874 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2875 /* Take pin on all head pages to avoid freeing them under us */ 2876 list_for_each_entry_safe(folio, next, &ds_queue->split_queue, 2877 _deferred_list) { 2878 if (folio_try_get(folio)) { 2879 list_move(&folio->_deferred_list, &list); 2880 } else { 2881 /* We lost race with folio_put() */ 2882 list_del_init(&folio->_deferred_list); 2883 ds_queue->split_queue_len--; 2884 } 2885 if (!--sc->nr_to_scan) 2886 break; 2887 } 2888 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2889 2890 list_for_each_entry_safe(folio, next, &list, _deferred_list) { 2891 if (!folio_trylock(folio)) 2892 goto next; 2893 /* split_huge_page() removes page from list on success */ 2894 if (!split_folio(folio)) 2895 split++; 2896 folio_unlock(folio); 2897 next: 2898 folio_put(folio); 2899 } 2900 2901 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2902 list_splice_tail(&list, &ds_queue->split_queue); 2903 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2904 2905 /* 2906 * Stop shrinker if we didn't split any page, but the queue is empty. 2907 * This can happen if pages were freed under us. 2908 */ 2909 if (!split && list_empty(&ds_queue->split_queue)) 2910 return SHRINK_STOP; 2911 return split; 2912 } 2913 2914 static struct shrinker deferred_split_shrinker = { 2915 .count_objects = deferred_split_count, 2916 .scan_objects = deferred_split_scan, 2917 .seeks = DEFAULT_SEEKS, 2918 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | 2919 SHRINKER_NONSLAB, 2920 }; 2921 2922 #ifdef CONFIG_DEBUG_FS 2923 static void split_huge_pages_all(void) 2924 { 2925 struct zone *zone; 2926 struct page *page; 2927 struct folio *folio; 2928 unsigned long pfn, max_zone_pfn; 2929 unsigned long total = 0, split = 0; 2930 2931 pr_debug("Split all THPs\n"); 2932 for_each_zone(zone) { 2933 if (!managed_zone(zone)) 2934 continue; 2935 max_zone_pfn = zone_end_pfn(zone); 2936 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2937 int nr_pages; 2938 2939 page = pfn_to_online_page(pfn); 2940 if (!page || PageTail(page)) 2941 continue; 2942 folio = page_folio(page); 2943 if (!folio_try_get(folio)) 2944 continue; 2945 2946 if (unlikely(page_folio(page) != folio)) 2947 goto next; 2948 2949 if (zone != folio_zone(folio)) 2950 goto next; 2951 2952 if (!folio_test_large(folio) 2953 || folio_test_hugetlb(folio) 2954 || !folio_test_lru(folio)) 2955 goto next; 2956 2957 total++; 2958 folio_lock(folio); 2959 nr_pages = folio_nr_pages(folio); 2960 if (!split_folio(folio)) 2961 split++; 2962 pfn += nr_pages - 1; 2963 folio_unlock(folio); 2964 next: 2965 folio_put(folio); 2966 cond_resched(); 2967 } 2968 } 2969 2970 pr_debug("%lu of %lu THP split\n", split, total); 2971 } 2972 2973 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 2974 { 2975 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 2976 is_vm_hugetlb_page(vma); 2977 } 2978 2979 static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 2980 unsigned long vaddr_end) 2981 { 2982 int ret = 0; 2983 struct task_struct *task; 2984 struct mm_struct *mm; 2985 unsigned long total = 0, split = 0; 2986 unsigned long addr; 2987 2988 vaddr_start &= PAGE_MASK; 2989 vaddr_end &= PAGE_MASK; 2990 2991 /* Find the task_struct from pid */ 2992 rcu_read_lock(); 2993 task = find_task_by_vpid(pid); 2994 if (!task) { 2995 rcu_read_unlock(); 2996 ret = -ESRCH; 2997 goto out; 2998 } 2999 get_task_struct(task); 3000 rcu_read_unlock(); 3001 3002 /* Find the mm_struct */ 3003 mm = get_task_mm(task); 3004 put_task_struct(task); 3005 3006 if (!mm) { 3007 ret = -EINVAL; 3008 goto out; 3009 } 3010 3011 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 3012 pid, vaddr_start, vaddr_end); 3013 3014 mmap_read_lock(mm); 3015 /* 3016 * always increase addr by PAGE_SIZE, since we could have a PTE page 3017 * table filled with PTE-mapped THPs, each of which is distinct. 3018 */ 3019 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 3020 struct vm_area_struct *vma = vma_lookup(mm, addr); 3021 struct page *page; 3022 struct folio *folio; 3023 3024 if (!vma) 3025 break; 3026 3027 /* skip special VMA and hugetlb VMA */ 3028 if (vma_not_suitable_for_thp_split(vma)) { 3029 addr = vma->vm_end; 3030 continue; 3031 } 3032 3033 /* FOLL_DUMP to ignore special (like zero) pages */ 3034 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 3035 3036 if (IS_ERR_OR_NULL(page)) 3037 continue; 3038 3039 folio = page_folio(page); 3040 if (!is_transparent_hugepage(folio)) 3041 goto next; 3042 3043 total++; 3044 if (!can_split_folio(folio, NULL)) 3045 goto next; 3046 3047 if (!folio_trylock(folio)) 3048 goto next; 3049 3050 if (!split_folio(folio)) 3051 split++; 3052 3053 folio_unlock(folio); 3054 next: 3055 folio_put(folio); 3056 cond_resched(); 3057 } 3058 mmap_read_unlock(mm); 3059 mmput(mm); 3060 3061 pr_debug("%lu of %lu THP split\n", split, total); 3062 3063 out: 3064 return ret; 3065 } 3066 3067 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 3068 pgoff_t off_end) 3069 { 3070 struct filename *file; 3071 struct file *candidate; 3072 struct address_space *mapping; 3073 int ret = -EINVAL; 3074 pgoff_t index; 3075 int nr_pages = 1; 3076 unsigned long total = 0, split = 0; 3077 3078 file = getname_kernel(file_path); 3079 if (IS_ERR(file)) 3080 return ret; 3081 3082 candidate = file_open_name(file, O_RDONLY, 0); 3083 if (IS_ERR(candidate)) 3084 goto out; 3085 3086 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 3087 file_path, off_start, off_end); 3088 3089 mapping = candidate->f_mapping; 3090 3091 for (index = off_start; index < off_end; index += nr_pages) { 3092 struct folio *folio = filemap_get_folio(mapping, index); 3093 3094 nr_pages = 1; 3095 if (IS_ERR(folio)) 3096 continue; 3097 3098 if (!folio_test_large(folio)) 3099 goto next; 3100 3101 total++; 3102 nr_pages = folio_nr_pages(folio); 3103 3104 if (!folio_trylock(folio)) 3105 goto next; 3106 3107 if (!split_folio(folio)) 3108 split++; 3109 3110 folio_unlock(folio); 3111 next: 3112 folio_put(folio); 3113 cond_resched(); 3114 } 3115 3116 filp_close(candidate, NULL); 3117 ret = 0; 3118 3119 pr_debug("%lu of %lu file-backed THP split\n", split, total); 3120 out: 3121 putname(file); 3122 return ret; 3123 } 3124 3125 #define MAX_INPUT_BUF_SZ 255 3126 3127 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 3128 size_t count, loff_t *ppops) 3129 { 3130 static DEFINE_MUTEX(split_debug_mutex); 3131 ssize_t ret; 3132 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */ 3133 char input_buf[MAX_INPUT_BUF_SZ]; 3134 int pid; 3135 unsigned long vaddr_start, vaddr_end; 3136 3137 ret = mutex_lock_interruptible(&split_debug_mutex); 3138 if (ret) 3139 return ret; 3140 3141 ret = -EFAULT; 3142 3143 memset(input_buf, 0, MAX_INPUT_BUF_SZ); 3144 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 3145 goto out; 3146 3147 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 3148 3149 if (input_buf[0] == '/') { 3150 char *tok; 3151 char *buf = input_buf; 3152 char file_path[MAX_INPUT_BUF_SZ]; 3153 pgoff_t off_start = 0, off_end = 0; 3154 size_t input_len = strlen(input_buf); 3155 3156 tok = strsep(&buf, ","); 3157 if (tok) { 3158 strcpy(file_path, tok); 3159 } else { 3160 ret = -EINVAL; 3161 goto out; 3162 } 3163 3164 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end); 3165 if (ret != 2) { 3166 ret = -EINVAL; 3167 goto out; 3168 } 3169 ret = split_huge_pages_in_file(file_path, off_start, off_end); 3170 if (!ret) 3171 ret = input_len; 3172 3173 goto out; 3174 } 3175 3176 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end); 3177 if (ret == 1 && pid == 1) { 3178 split_huge_pages_all(); 3179 ret = strlen(input_buf); 3180 goto out; 3181 } else if (ret != 3) { 3182 ret = -EINVAL; 3183 goto out; 3184 } 3185 3186 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end); 3187 if (!ret) 3188 ret = strlen(input_buf); 3189 out: 3190 mutex_unlock(&split_debug_mutex); 3191 return ret; 3192 3193 } 3194 3195 static const struct file_operations split_huge_pages_fops = { 3196 .owner = THIS_MODULE, 3197 .write = split_huge_pages_write, 3198 .llseek = no_llseek, 3199 }; 3200 3201 static int __init split_huge_pages_debugfs(void) 3202 { 3203 debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 3204 &split_huge_pages_fops); 3205 return 0; 3206 } 3207 late_initcall(split_huge_pages_debugfs); 3208 #endif 3209 3210 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 3211 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 3212 struct page *page) 3213 { 3214 struct vm_area_struct *vma = pvmw->vma; 3215 struct mm_struct *mm = vma->vm_mm; 3216 unsigned long address = pvmw->address; 3217 bool anon_exclusive; 3218 pmd_t pmdval; 3219 swp_entry_t entry; 3220 pmd_t pmdswp; 3221 3222 if (!(pvmw->pmd && !pvmw->pte)) 3223 return 0; 3224 3225 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 3226 pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 3227 3228 /* See page_try_share_anon_rmap(): invalidate PMD first. */ 3229 anon_exclusive = PageAnon(page) && PageAnonExclusive(page); 3230 if (anon_exclusive && page_try_share_anon_rmap(page)) { 3231 set_pmd_at(mm, address, pvmw->pmd, pmdval); 3232 return -EBUSY; 3233 } 3234 3235 if (pmd_dirty(pmdval)) 3236 set_page_dirty(page); 3237 if (pmd_write(pmdval)) 3238 entry = make_writable_migration_entry(page_to_pfn(page)); 3239 else if (anon_exclusive) 3240 entry = make_readable_exclusive_migration_entry(page_to_pfn(page)); 3241 else 3242 entry = make_readable_migration_entry(page_to_pfn(page)); 3243 if (pmd_young(pmdval)) 3244 entry = make_migration_entry_young(entry); 3245 if (pmd_dirty(pmdval)) 3246 entry = make_migration_entry_dirty(entry); 3247 pmdswp = swp_entry_to_pmd(entry); 3248 if (pmd_soft_dirty(pmdval)) 3249 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 3250 if (pmd_uffd_wp(pmdval)) 3251 pmdswp = pmd_swp_mkuffd_wp(pmdswp); 3252 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 3253 page_remove_rmap(page, vma, true); 3254 put_page(page); 3255 trace_set_migration_pmd(address, pmd_val(pmdswp)); 3256 3257 return 0; 3258 } 3259 3260 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 3261 { 3262 struct vm_area_struct *vma = pvmw->vma; 3263 struct mm_struct *mm = vma->vm_mm; 3264 unsigned long address = pvmw->address; 3265 unsigned long haddr = address & HPAGE_PMD_MASK; 3266 pmd_t pmde; 3267 swp_entry_t entry; 3268 3269 if (!(pvmw->pmd && !pvmw->pte)) 3270 return; 3271 3272 entry = pmd_to_swp_entry(*pvmw->pmd); 3273 get_page(new); 3274 pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)); 3275 if (pmd_swp_soft_dirty(*pvmw->pmd)) 3276 pmde = pmd_mksoft_dirty(pmde); 3277 if (is_writable_migration_entry(entry)) 3278 pmde = pmd_mkwrite(pmde, vma); 3279 if (pmd_swp_uffd_wp(*pvmw->pmd)) 3280 pmde = pmd_mkuffd_wp(pmde); 3281 if (!is_migration_entry_young(entry)) 3282 pmde = pmd_mkold(pmde); 3283 /* NOTE: this may contain setting soft-dirty on some archs */ 3284 if (PageDirty(new) && is_migration_entry_dirty(entry)) 3285 pmde = pmd_mkdirty(pmde); 3286 3287 if (PageAnon(new)) { 3288 rmap_t rmap_flags = RMAP_COMPOUND; 3289 3290 if (!is_readable_migration_entry(entry)) 3291 rmap_flags |= RMAP_EXCLUSIVE; 3292 3293 page_add_anon_rmap(new, vma, haddr, rmap_flags); 3294 } else { 3295 page_add_file_rmap(new, vma, true); 3296 } 3297 VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new)); 3298 set_pmd_at(mm, haddr, pvmw->pmd, pmde); 3299 3300 /* No need to invalidate - it was non-present before */ 3301 update_mmu_cache_pmd(vma, address, pvmw->pmd); 3302 trace_remove_migration_pmd(address, pmd_val(pmde)); 3303 } 3304 #endif 3305