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/numa_balancing.h> 12 #include <linux/highmem.h> 13 #include <linux/hugetlb.h> 14 #include <linux/mmu_notifier.h> 15 #include <linux/rmap.h> 16 #include <linux/swap.h> 17 #include <linux/shrinker.h> 18 #include <linux/mm_inline.h> 19 #include <linux/swapops.h> 20 #include <linux/backing-dev.h> 21 #include <linux/dax.h> 22 #include <linux/mm_types.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 #include <linux/pgalloc_tag.h> 42 #include <linux/pagewalk.h> 43 44 #include <asm/tlb.h> 45 #include <asm/pgalloc.h> 46 #include "internal.h" 47 #include "swap.h" 48 49 #define CREATE_TRACE_POINTS 50 #include <trace/events/thp.h> 51 52 /* 53 * By default, transparent hugepage support is disabled in order to avoid 54 * risking an increased memory footprint for applications that are not 55 * guaranteed to benefit from it. When transparent hugepage support is 56 * enabled, it is for all mappings, and khugepaged scans all mappings. 57 * Defrag is invoked by khugepaged hugepage allocations and by page faults 58 * for all hugepage allocations. 59 */ 60 unsigned long transparent_hugepage_flags __read_mostly = 61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 62 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 63 #endif 64 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 65 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 66 #endif 67 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 68 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 69 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 70 71 static struct shrinker *deferred_split_shrinker; 72 static unsigned long deferred_split_count(struct shrinker *shrink, 73 struct shrink_control *sc); 74 static unsigned long deferred_split_scan(struct shrinker *shrink, 75 struct shrink_control *sc); 76 static bool split_underused_thp = true; 77 78 static atomic_t huge_zero_refcount; 79 struct folio *huge_zero_folio __read_mostly; 80 unsigned long huge_zero_pfn __read_mostly = ~0UL; 81 unsigned long huge_anon_orders_always __read_mostly; 82 unsigned long huge_anon_orders_madvise __read_mostly; 83 unsigned long huge_anon_orders_inherit __read_mostly; 84 static bool anon_orders_configured __initdata; 85 86 static inline bool file_thp_enabled(struct vm_area_struct *vma) 87 { 88 struct inode *inode; 89 90 if (!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) 91 return false; 92 93 if (!vma->vm_file) 94 return false; 95 96 inode = file_inode(vma->vm_file); 97 98 return !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode); 99 } 100 101 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma, 102 vm_flags_t vm_flags, 103 unsigned long tva_flags, 104 unsigned long orders) 105 { 106 bool smaps = tva_flags & TVA_SMAPS; 107 bool in_pf = tva_flags & TVA_IN_PF; 108 bool enforce_sysfs = tva_flags & TVA_ENFORCE_SYSFS; 109 unsigned long supported_orders; 110 111 /* Check the intersection of requested and supported orders. */ 112 if (vma_is_anonymous(vma)) 113 supported_orders = THP_ORDERS_ALL_ANON; 114 else if (vma_is_special_huge(vma)) 115 supported_orders = THP_ORDERS_ALL_SPECIAL; 116 else 117 supported_orders = THP_ORDERS_ALL_FILE_DEFAULT; 118 119 orders &= supported_orders; 120 if (!orders) 121 return 0; 122 123 if (!vma->vm_mm) /* vdso */ 124 return 0; 125 126 if (thp_disabled_by_hw() || vma_thp_disabled(vma, vm_flags)) 127 return 0; 128 129 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */ 130 if (vma_is_dax(vma)) 131 return in_pf ? orders : 0; 132 133 /* 134 * khugepaged special VMA and hugetlb VMA. 135 * Must be checked after dax since some dax mappings may have 136 * VM_MIXEDMAP set. 137 */ 138 if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED)) 139 return 0; 140 141 /* 142 * Check alignment for file vma and size for both file and anon vma by 143 * filtering out the unsuitable orders. 144 * 145 * Skip the check for page fault. Huge fault does the check in fault 146 * handlers. 147 */ 148 if (!in_pf) { 149 int order = highest_order(orders); 150 unsigned long addr; 151 152 while (orders) { 153 addr = vma->vm_end - (PAGE_SIZE << order); 154 if (thp_vma_suitable_order(vma, addr, order)) 155 break; 156 order = next_order(&orders, order); 157 } 158 159 if (!orders) 160 return 0; 161 } 162 163 /* 164 * Enabled via shmem mount options or sysfs settings. 165 * Must be done before hugepage flags check since shmem has its 166 * own flags. 167 */ 168 if (!in_pf && shmem_file(vma->vm_file)) 169 return orders & shmem_allowable_huge_orders(file_inode(vma->vm_file), 170 vma, vma->vm_pgoff, 0, 171 !enforce_sysfs); 172 173 if (!vma_is_anonymous(vma)) { 174 /* 175 * Enforce sysfs THP requirements as necessary. Anonymous vmas 176 * were already handled in thp_vma_allowable_orders(). 177 */ 178 if (enforce_sysfs && 179 (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) && 180 !hugepage_global_always()))) 181 return 0; 182 183 /* 184 * Trust that ->huge_fault() handlers know what they are doing 185 * in fault path. 186 */ 187 if (((in_pf || smaps)) && vma->vm_ops->huge_fault) 188 return orders; 189 /* Only regular file is valid in collapse path */ 190 if (((!in_pf || smaps)) && file_thp_enabled(vma)) 191 return orders; 192 return 0; 193 } 194 195 if (vma_is_temporary_stack(vma)) 196 return 0; 197 198 /* 199 * THPeligible bit of smaps should show 1 for proper VMAs even 200 * though anon_vma is not initialized yet. 201 * 202 * Allow page fault since anon_vma may be not initialized until 203 * the first page fault. 204 */ 205 if (!vma->anon_vma) 206 return (smaps || in_pf) ? orders : 0; 207 208 return orders; 209 } 210 211 static bool get_huge_zero_page(void) 212 { 213 struct folio *zero_folio; 214 retry: 215 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 216 return true; 217 218 zero_folio = folio_alloc((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 219 HPAGE_PMD_ORDER); 220 if (!zero_folio) { 221 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 222 return false; 223 } 224 /* Ensure zero folio won't have large_rmappable flag set. */ 225 folio_clear_large_rmappable(zero_folio); 226 preempt_disable(); 227 if (cmpxchg(&huge_zero_folio, NULL, zero_folio)) { 228 preempt_enable(); 229 folio_put(zero_folio); 230 goto retry; 231 } 232 WRITE_ONCE(huge_zero_pfn, folio_pfn(zero_folio)); 233 234 /* We take additional reference here. It will be put back by shrinker */ 235 atomic_set(&huge_zero_refcount, 2); 236 preempt_enable(); 237 count_vm_event(THP_ZERO_PAGE_ALLOC); 238 return true; 239 } 240 241 static void put_huge_zero_page(void) 242 { 243 /* 244 * Counter should never go to zero here. Only shrinker can put 245 * last reference. 246 */ 247 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 248 } 249 250 struct folio *mm_get_huge_zero_folio(struct mm_struct *mm) 251 { 252 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 253 return READ_ONCE(huge_zero_folio); 254 255 if (!get_huge_zero_page()) 256 return NULL; 257 258 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 259 put_huge_zero_page(); 260 261 return READ_ONCE(huge_zero_folio); 262 } 263 264 void mm_put_huge_zero_folio(struct mm_struct *mm) 265 { 266 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 267 put_huge_zero_page(); 268 } 269 270 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 271 struct shrink_control *sc) 272 { 273 /* we can free zero page only if last reference remains */ 274 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 275 } 276 277 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 278 struct shrink_control *sc) 279 { 280 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 281 struct folio *zero_folio = xchg(&huge_zero_folio, NULL); 282 BUG_ON(zero_folio == NULL); 283 WRITE_ONCE(huge_zero_pfn, ~0UL); 284 folio_put(zero_folio); 285 return HPAGE_PMD_NR; 286 } 287 288 return 0; 289 } 290 291 static struct shrinker *huge_zero_page_shrinker; 292 293 #ifdef CONFIG_SYSFS 294 static ssize_t enabled_show(struct kobject *kobj, 295 struct kobj_attribute *attr, char *buf) 296 { 297 const char *output; 298 299 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 300 output = "[always] madvise never"; 301 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 302 &transparent_hugepage_flags)) 303 output = "always [madvise] never"; 304 else 305 output = "always madvise [never]"; 306 307 return sysfs_emit(buf, "%s\n", output); 308 } 309 310 static ssize_t enabled_store(struct kobject *kobj, 311 struct kobj_attribute *attr, 312 const char *buf, size_t count) 313 { 314 ssize_t ret = count; 315 316 if (sysfs_streq(buf, "always")) { 317 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 318 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 319 } else if (sysfs_streq(buf, "madvise")) { 320 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 321 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 322 } else if (sysfs_streq(buf, "never")) { 323 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 324 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 325 } else 326 ret = -EINVAL; 327 328 if (ret > 0) { 329 int err = start_stop_khugepaged(); 330 if (err) 331 ret = err; 332 } 333 return ret; 334 } 335 336 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled); 337 338 ssize_t single_hugepage_flag_show(struct kobject *kobj, 339 struct kobj_attribute *attr, char *buf, 340 enum transparent_hugepage_flag flag) 341 { 342 return sysfs_emit(buf, "%d\n", 343 !!test_bit(flag, &transparent_hugepage_flags)); 344 } 345 346 ssize_t single_hugepage_flag_store(struct kobject *kobj, 347 struct kobj_attribute *attr, 348 const char *buf, size_t count, 349 enum transparent_hugepage_flag flag) 350 { 351 unsigned long value; 352 int ret; 353 354 ret = kstrtoul(buf, 10, &value); 355 if (ret < 0) 356 return ret; 357 if (value > 1) 358 return -EINVAL; 359 360 if (value) 361 set_bit(flag, &transparent_hugepage_flags); 362 else 363 clear_bit(flag, &transparent_hugepage_flags); 364 365 return count; 366 } 367 368 static ssize_t defrag_show(struct kobject *kobj, 369 struct kobj_attribute *attr, char *buf) 370 { 371 const char *output; 372 373 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 374 &transparent_hugepage_flags)) 375 output = "[always] defer defer+madvise madvise never"; 376 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 377 &transparent_hugepage_flags)) 378 output = "always [defer] defer+madvise madvise never"; 379 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, 380 &transparent_hugepage_flags)) 381 output = "always defer [defer+madvise] madvise never"; 382 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, 383 &transparent_hugepage_flags)) 384 output = "always defer defer+madvise [madvise] never"; 385 else 386 output = "always defer defer+madvise madvise [never]"; 387 388 return sysfs_emit(buf, "%s\n", output); 389 } 390 391 static ssize_t defrag_store(struct kobject *kobj, 392 struct kobj_attribute *attr, 393 const char *buf, size_t count) 394 { 395 if (sysfs_streq(buf, "always")) { 396 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 397 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 398 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 399 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 400 } else if (sysfs_streq(buf, "defer+madvise")) { 401 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 402 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 403 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 404 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 405 } else if (sysfs_streq(buf, "defer")) { 406 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 407 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 408 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 409 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 410 } else if (sysfs_streq(buf, "madvise")) { 411 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 412 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 413 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 414 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 415 } else if (sysfs_streq(buf, "never")) { 416 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 417 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 418 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 419 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 420 } else 421 return -EINVAL; 422 423 return count; 424 } 425 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag); 426 427 static ssize_t use_zero_page_show(struct kobject *kobj, 428 struct kobj_attribute *attr, char *buf) 429 { 430 return single_hugepage_flag_show(kobj, attr, buf, 431 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 432 } 433 static ssize_t use_zero_page_store(struct kobject *kobj, 434 struct kobj_attribute *attr, const char *buf, size_t count) 435 { 436 return single_hugepage_flag_store(kobj, attr, buf, count, 437 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 438 } 439 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page); 440 441 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 442 struct kobj_attribute *attr, char *buf) 443 { 444 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); 445 } 446 static struct kobj_attribute hpage_pmd_size_attr = 447 __ATTR_RO(hpage_pmd_size); 448 449 static ssize_t split_underused_thp_show(struct kobject *kobj, 450 struct kobj_attribute *attr, char *buf) 451 { 452 return sysfs_emit(buf, "%d\n", split_underused_thp); 453 } 454 455 static ssize_t split_underused_thp_store(struct kobject *kobj, 456 struct kobj_attribute *attr, 457 const char *buf, size_t count) 458 { 459 int err = kstrtobool(buf, &split_underused_thp); 460 461 if (err < 0) 462 return err; 463 464 return count; 465 } 466 467 static struct kobj_attribute split_underused_thp_attr = __ATTR( 468 shrink_underused, 0644, split_underused_thp_show, split_underused_thp_store); 469 470 static struct attribute *hugepage_attr[] = { 471 &enabled_attr.attr, 472 &defrag_attr.attr, 473 &use_zero_page_attr.attr, 474 &hpage_pmd_size_attr.attr, 475 #ifdef CONFIG_SHMEM 476 &shmem_enabled_attr.attr, 477 #endif 478 &split_underused_thp_attr.attr, 479 NULL, 480 }; 481 482 static const struct attribute_group hugepage_attr_group = { 483 .attrs = hugepage_attr, 484 }; 485 486 static void hugepage_exit_sysfs(struct kobject *hugepage_kobj); 487 static void thpsize_release(struct kobject *kobj); 488 static DEFINE_SPINLOCK(huge_anon_orders_lock); 489 static LIST_HEAD(thpsize_list); 490 491 static ssize_t anon_enabled_show(struct kobject *kobj, 492 struct kobj_attribute *attr, char *buf) 493 { 494 int order = to_thpsize(kobj)->order; 495 const char *output; 496 497 if (test_bit(order, &huge_anon_orders_always)) 498 output = "[always] inherit madvise never"; 499 else if (test_bit(order, &huge_anon_orders_inherit)) 500 output = "always [inherit] madvise never"; 501 else if (test_bit(order, &huge_anon_orders_madvise)) 502 output = "always inherit [madvise] never"; 503 else 504 output = "always inherit madvise [never]"; 505 506 return sysfs_emit(buf, "%s\n", output); 507 } 508 509 static ssize_t anon_enabled_store(struct kobject *kobj, 510 struct kobj_attribute *attr, 511 const char *buf, size_t count) 512 { 513 int order = to_thpsize(kobj)->order; 514 ssize_t ret = count; 515 516 if (sysfs_streq(buf, "always")) { 517 spin_lock(&huge_anon_orders_lock); 518 clear_bit(order, &huge_anon_orders_inherit); 519 clear_bit(order, &huge_anon_orders_madvise); 520 set_bit(order, &huge_anon_orders_always); 521 spin_unlock(&huge_anon_orders_lock); 522 } else if (sysfs_streq(buf, "inherit")) { 523 spin_lock(&huge_anon_orders_lock); 524 clear_bit(order, &huge_anon_orders_always); 525 clear_bit(order, &huge_anon_orders_madvise); 526 set_bit(order, &huge_anon_orders_inherit); 527 spin_unlock(&huge_anon_orders_lock); 528 } else if (sysfs_streq(buf, "madvise")) { 529 spin_lock(&huge_anon_orders_lock); 530 clear_bit(order, &huge_anon_orders_always); 531 clear_bit(order, &huge_anon_orders_inherit); 532 set_bit(order, &huge_anon_orders_madvise); 533 spin_unlock(&huge_anon_orders_lock); 534 } else if (sysfs_streq(buf, "never")) { 535 spin_lock(&huge_anon_orders_lock); 536 clear_bit(order, &huge_anon_orders_always); 537 clear_bit(order, &huge_anon_orders_inherit); 538 clear_bit(order, &huge_anon_orders_madvise); 539 spin_unlock(&huge_anon_orders_lock); 540 } else 541 ret = -EINVAL; 542 543 if (ret > 0) { 544 int err; 545 546 err = start_stop_khugepaged(); 547 if (err) 548 ret = err; 549 } 550 return ret; 551 } 552 553 static struct kobj_attribute anon_enabled_attr = 554 __ATTR(enabled, 0644, anon_enabled_show, anon_enabled_store); 555 556 static struct attribute *anon_ctrl_attrs[] = { 557 &anon_enabled_attr.attr, 558 NULL, 559 }; 560 561 static const struct attribute_group anon_ctrl_attr_grp = { 562 .attrs = anon_ctrl_attrs, 563 }; 564 565 static struct attribute *file_ctrl_attrs[] = { 566 #ifdef CONFIG_SHMEM 567 &thpsize_shmem_enabled_attr.attr, 568 #endif 569 NULL, 570 }; 571 572 static const struct attribute_group file_ctrl_attr_grp = { 573 .attrs = file_ctrl_attrs, 574 }; 575 576 static struct attribute *any_ctrl_attrs[] = { 577 NULL, 578 }; 579 580 static const struct attribute_group any_ctrl_attr_grp = { 581 .attrs = any_ctrl_attrs, 582 }; 583 584 static const struct kobj_type thpsize_ktype = { 585 .release = &thpsize_release, 586 .sysfs_ops = &kobj_sysfs_ops, 587 }; 588 589 DEFINE_PER_CPU(struct mthp_stat, mthp_stats) = {{{0}}}; 590 591 static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item) 592 { 593 unsigned long sum = 0; 594 int cpu; 595 596 for_each_possible_cpu(cpu) { 597 struct mthp_stat *this = &per_cpu(mthp_stats, cpu); 598 599 sum += this->stats[order][item]; 600 } 601 602 return sum; 603 } 604 605 #define DEFINE_MTHP_STAT_ATTR(_name, _index) \ 606 static ssize_t _name##_show(struct kobject *kobj, \ 607 struct kobj_attribute *attr, char *buf) \ 608 { \ 609 int order = to_thpsize(kobj)->order; \ 610 \ 611 return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index)); \ 612 } \ 613 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 614 615 DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC); 616 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK); 617 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE); 618 DEFINE_MTHP_STAT_ATTR(zswpout, MTHP_STAT_ZSWPOUT); 619 DEFINE_MTHP_STAT_ATTR(swpin, MTHP_STAT_SWPIN); 620 DEFINE_MTHP_STAT_ATTR(swpin_fallback, MTHP_STAT_SWPIN_FALLBACK); 621 DEFINE_MTHP_STAT_ATTR(swpin_fallback_charge, MTHP_STAT_SWPIN_FALLBACK_CHARGE); 622 DEFINE_MTHP_STAT_ATTR(swpout, MTHP_STAT_SWPOUT); 623 DEFINE_MTHP_STAT_ATTR(swpout_fallback, MTHP_STAT_SWPOUT_FALLBACK); 624 #ifdef CONFIG_SHMEM 625 DEFINE_MTHP_STAT_ATTR(shmem_alloc, MTHP_STAT_SHMEM_ALLOC); 626 DEFINE_MTHP_STAT_ATTR(shmem_fallback, MTHP_STAT_SHMEM_FALLBACK); 627 DEFINE_MTHP_STAT_ATTR(shmem_fallback_charge, MTHP_STAT_SHMEM_FALLBACK_CHARGE); 628 #endif 629 DEFINE_MTHP_STAT_ATTR(split, MTHP_STAT_SPLIT); 630 DEFINE_MTHP_STAT_ATTR(split_failed, MTHP_STAT_SPLIT_FAILED); 631 DEFINE_MTHP_STAT_ATTR(split_deferred, MTHP_STAT_SPLIT_DEFERRED); 632 DEFINE_MTHP_STAT_ATTR(nr_anon, MTHP_STAT_NR_ANON); 633 DEFINE_MTHP_STAT_ATTR(nr_anon_partially_mapped, MTHP_STAT_NR_ANON_PARTIALLY_MAPPED); 634 635 static struct attribute *anon_stats_attrs[] = { 636 &anon_fault_alloc_attr.attr, 637 &anon_fault_fallback_attr.attr, 638 &anon_fault_fallback_charge_attr.attr, 639 #ifndef CONFIG_SHMEM 640 &zswpout_attr.attr, 641 &swpin_attr.attr, 642 &swpin_fallback_attr.attr, 643 &swpin_fallback_charge_attr.attr, 644 &swpout_attr.attr, 645 &swpout_fallback_attr.attr, 646 #endif 647 &split_deferred_attr.attr, 648 &nr_anon_attr.attr, 649 &nr_anon_partially_mapped_attr.attr, 650 NULL, 651 }; 652 653 static struct attribute_group anon_stats_attr_grp = { 654 .name = "stats", 655 .attrs = anon_stats_attrs, 656 }; 657 658 static struct attribute *file_stats_attrs[] = { 659 #ifdef CONFIG_SHMEM 660 &shmem_alloc_attr.attr, 661 &shmem_fallback_attr.attr, 662 &shmem_fallback_charge_attr.attr, 663 #endif 664 NULL, 665 }; 666 667 static struct attribute_group file_stats_attr_grp = { 668 .name = "stats", 669 .attrs = file_stats_attrs, 670 }; 671 672 static struct attribute *any_stats_attrs[] = { 673 #ifdef CONFIG_SHMEM 674 &zswpout_attr.attr, 675 &swpin_attr.attr, 676 &swpin_fallback_attr.attr, 677 &swpin_fallback_charge_attr.attr, 678 &swpout_attr.attr, 679 &swpout_fallback_attr.attr, 680 #endif 681 &split_attr.attr, 682 &split_failed_attr.attr, 683 NULL, 684 }; 685 686 static struct attribute_group any_stats_attr_grp = { 687 .name = "stats", 688 .attrs = any_stats_attrs, 689 }; 690 691 static int sysfs_add_group(struct kobject *kobj, 692 const struct attribute_group *grp) 693 { 694 int ret = -ENOENT; 695 696 /* 697 * If the group is named, try to merge first, assuming the subdirectory 698 * was already created. This avoids the warning emitted by 699 * sysfs_create_group() if the directory already exists. 700 */ 701 if (grp->name) 702 ret = sysfs_merge_group(kobj, grp); 703 if (ret) 704 ret = sysfs_create_group(kobj, grp); 705 706 return ret; 707 } 708 709 static struct thpsize *thpsize_create(int order, struct kobject *parent) 710 { 711 unsigned long size = (PAGE_SIZE << order) / SZ_1K; 712 struct thpsize *thpsize; 713 int ret = -ENOMEM; 714 715 thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL); 716 if (!thpsize) 717 goto err; 718 719 thpsize->order = order; 720 721 ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent, 722 "hugepages-%lukB", size); 723 if (ret) { 724 kfree(thpsize); 725 goto err; 726 } 727 728 729 ret = sysfs_add_group(&thpsize->kobj, &any_ctrl_attr_grp); 730 if (ret) 731 goto err_put; 732 733 ret = sysfs_add_group(&thpsize->kobj, &any_stats_attr_grp); 734 if (ret) 735 goto err_put; 736 737 if (BIT(order) & THP_ORDERS_ALL_ANON) { 738 ret = sysfs_add_group(&thpsize->kobj, &anon_ctrl_attr_grp); 739 if (ret) 740 goto err_put; 741 742 ret = sysfs_add_group(&thpsize->kobj, &anon_stats_attr_grp); 743 if (ret) 744 goto err_put; 745 } 746 747 if (BIT(order) & THP_ORDERS_ALL_FILE_DEFAULT) { 748 ret = sysfs_add_group(&thpsize->kobj, &file_ctrl_attr_grp); 749 if (ret) 750 goto err_put; 751 752 ret = sysfs_add_group(&thpsize->kobj, &file_stats_attr_grp); 753 if (ret) 754 goto err_put; 755 } 756 757 return thpsize; 758 err_put: 759 kobject_put(&thpsize->kobj); 760 err: 761 return ERR_PTR(ret); 762 } 763 764 static void thpsize_release(struct kobject *kobj) 765 { 766 kfree(to_thpsize(kobj)); 767 } 768 769 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 770 { 771 int err; 772 struct thpsize *thpsize; 773 unsigned long orders; 774 int order; 775 776 /* 777 * Default to setting PMD-sized THP to inherit the global setting and 778 * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time 779 * constant so we have to do this here. 780 */ 781 if (!anon_orders_configured) 782 huge_anon_orders_inherit = BIT(PMD_ORDER); 783 784 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 785 if (unlikely(!*hugepage_kobj)) { 786 pr_err("failed to create transparent hugepage kobject\n"); 787 return -ENOMEM; 788 } 789 790 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 791 if (err) { 792 pr_err("failed to register transparent hugepage group\n"); 793 goto delete_obj; 794 } 795 796 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 797 if (err) { 798 pr_err("failed to register transparent hugepage group\n"); 799 goto remove_hp_group; 800 } 801 802 orders = THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE_DEFAULT; 803 order = highest_order(orders); 804 while (orders) { 805 thpsize = thpsize_create(order, *hugepage_kobj); 806 if (IS_ERR(thpsize)) { 807 pr_err("failed to create thpsize for order %d\n", order); 808 err = PTR_ERR(thpsize); 809 goto remove_all; 810 } 811 list_add(&thpsize->node, &thpsize_list); 812 order = next_order(&orders, order); 813 } 814 815 return 0; 816 817 remove_all: 818 hugepage_exit_sysfs(*hugepage_kobj); 819 return err; 820 remove_hp_group: 821 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 822 delete_obj: 823 kobject_put(*hugepage_kobj); 824 return err; 825 } 826 827 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 828 { 829 struct thpsize *thpsize, *tmp; 830 831 list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) { 832 list_del(&thpsize->node); 833 kobject_put(&thpsize->kobj); 834 } 835 836 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 837 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 838 kobject_put(hugepage_kobj); 839 } 840 #else 841 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 842 { 843 return 0; 844 } 845 846 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 847 { 848 } 849 #endif /* CONFIG_SYSFS */ 850 851 static int __init thp_shrinker_init(void) 852 { 853 huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero"); 854 if (!huge_zero_page_shrinker) 855 return -ENOMEM; 856 857 deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | 858 SHRINKER_MEMCG_AWARE | 859 SHRINKER_NONSLAB, 860 "thp-deferred_split"); 861 if (!deferred_split_shrinker) { 862 shrinker_free(huge_zero_page_shrinker); 863 return -ENOMEM; 864 } 865 866 huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count; 867 huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan; 868 shrinker_register(huge_zero_page_shrinker); 869 870 deferred_split_shrinker->count_objects = deferred_split_count; 871 deferred_split_shrinker->scan_objects = deferred_split_scan; 872 shrinker_register(deferred_split_shrinker); 873 874 return 0; 875 } 876 877 static void __init thp_shrinker_exit(void) 878 { 879 shrinker_free(huge_zero_page_shrinker); 880 shrinker_free(deferred_split_shrinker); 881 } 882 883 static int __init hugepage_init(void) 884 { 885 int err; 886 struct kobject *hugepage_kobj; 887 888 if (!has_transparent_hugepage()) { 889 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED; 890 return -EINVAL; 891 } 892 893 /* 894 * hugepages can't be allocated by the buddy allocator 895 */ 896 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER); 897 898 err = hugepage_init_sysfs(&hugepage_kobj); 899 if (err) 900 goto err_sysfs; 901 902 err = khugepaged_init(); 903 if (err) 904 goto err_slab; 905 906 err = thp_shrinker_init(); 907 if (err) 908 goto err_shrinker; 909 910 /* 911 * By default disable transparent hugepages on smaller systems, 912 * where the extra memory used could hurt more than TLB overhead 913 * is likely to save. The admin can still enable it through /sys. 914 */ 915 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 916 transparent_hugepage_flags = 0; 917 return 0; 918 } 919 920 err = start_stop_khugepaged(); 921 if (err) 922 goto err_khugepaged; 923 924 return 0; 925 err_khugepaged: 926 thp_shrinker_exit(); 927 err_shrinker: 928 khugepaged_destroy(); 929 err_slab: 930 hugepage_exit_sysfs(hugepage_kobj); 931 err_sysfs: 932 return err; 933 } 934 subsys_initcall(hugepage_init); 935 936 static int __init setup_transparent_hugepage(char *str) 937 { 938 int ret = 0; 939 if (!str) 940 goto out; 941 if (!strcmp(str, "always")) { 942 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 943 &transparent_hugepage_flags); 944 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 945 &transparent_hugepage_flags); 946 ret = 1; 947 } else if (!strcmp(str, "madvise")) { 948 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 949 &transparent_hugepage_flags); 950 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 951 &transparent_hugepage_flags); 952 ret = 1; 953 } else if (!strcmp(str, "never")) { 954 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 955 &transparent_hugepage_flags); 956 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 957 &transparent_hugepage_flags); 958 ret = 1; 959 } 960 out: 961 if (!ret) 962 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 963 return ret; 964 } 965 __setup("transparent_hugepage=", setup_transparent_hugepage); 966 967 static char str_dup[PAGE_SIZE] __initdata; 968 static int __init setup_thp_anon(char *str) 969 { 970 char *token, *range, *policy, *subtoken; 971 unsigned long always, inherit, madvise; 972 char *start_size, *end_size; 973 int start, end, nr; 974 char *p; 975 976 if (!str || strlen(str) + 1 > PAGE_SIZE) 977 goto err; 978 strscpy(str_dup, str); 979 980 always = huge_anon_orders_always; 981 madvise = huge_anon_orders_madvise; 982 inherit = huge_anon_orders_inherit; 983 p = str_dup; 984 while ((token = strsep(&p, ";")) != NULL) { 985 range = strsep(&token, ":"); 986 policy = token; 987 988 if (!policy) 989 goto err; 990 991 while ((subtoken = strsep(&range, ",")) != NULL) { 992 if (strchr(subtoken, '-')) { 993 start_size = strsep(&subtoken, "-"); 994 end_size = subtoken; 995 996 start = get_order_from_str(start_size, THP_ORDERS_ALL_ANON); 997 end = get_order_from_str(end_size, THP_ORDERS_ALL_ANON); 998 } else { 999 start_size = end_size = subtoken; 1000 start = end = get_order_from_str(subtoken, 1001 THP_ORDERS_ALL_ANON); 1002 } 1003 1004 if (start == -EINVAL) { 1005 pr_err("invalid size %s in thp_anon boot parameter\n", start_size); 1006 goto err; 1007 } 1008 1009 if (end == -EINVAL) { 1010 pr_err("invalid size %s in thp_anon boot parameter\n", end_size); 1011 goto err; 1012 } 1013 1014 if (start < 0 || end < 0 || start > end) 1015 goto err; 1016 1017 nr = end - start + 1; 1018 if (!strcmp(policy, "always")) { 1019 bitmap_set(&always, start, nr); 1020 bitmap_clear(&inherit, start, nr); 1021 bitmap_clear(&madvise, start, nr); 1022 } else if (!strcmp(policy, "madvise")) { 1023 bitmap_set(&madvise, start, nr); 1024 bitmap_clear(&inherit, start, nr); 1025 bitmap_clear(&always, start, nr); 1026 } else if (!strcmp(policy, "inherit")) { 1027 bitmap_set(&inherit, start, nr); 1028 bitmap_clear(&madvise, start, nr); 1029 bitmap_clear(&always, start, nr); 1030 } else if (!strcmp(policy, "never")) { 1031 bitmap_clear(&inherit, start, nr); 1032 bitmap_clear(&madvise, start, nr); 1033 bitmap_clear(&always, start, nr); 1034 } else { 1035 pr_err("invalid policy %s in thp_anon boot parameter\n", policy); 1036 goto err; 1037 } 1038 } 1039 } 1040 1041 huge_anon_orders_always = always; 1042 huge_anon_orders_madvise = madvise; 1043 huge_anon_orders_inherit = inherit; 1044 anon_orders_configured = true; 1045 return 1; 1046 1047 err: 1048 pr_warn("thp_anon=%s: error parsing string, ignoring setting\n", str); 1049 return 0; 1050 } 1051 __setup("thp_anon=", setup_thp_anon); 1052 1053 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 1054 { 1055 if (likely(vma->vm_flags & VM_WRITE)) 1056 pmd = pmd_mkwrite(pmd, vma); 1057 return pmd; 1058 } 1059 1060 #ifdef CONFIG_MEMCG 1061 static inline 1062 struct deferred_split *get_deferred_split_queue(struct folio *folio) 1063 { 1064 struct mem_cgroup *memcg = folio_memcg(folio); 1065 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 1066 1067 if (memcg) 1068 return &memcg->deferred_split_queue; 1069 else 1070 return &pgdat->deferred_split_queue; 1071 } 1072 #else 1073 static inline 1074 struct deferred_split *get_deferred_split_queue(struct folio *folio) 1075 { 1076 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 1077 1078 return &pgdat->deferred_split_queue; 1079 } 1080 #endif 1081 1082 static inline bool is_transparent_hugepage(const struct folio *folio) 1083 { 1084 if (!folio_test_large(folio)) 1085 return false; 1086 1087 return is_huge_zero_folio(folio) || 1088 folio_test_large_rmappable(folio); 1089 } 1090 1091 static unsigned long __thp_get_unmapped_area(struct file *filp, 1092 unsigned long addr, unsigned long len, 1093 loff_t off, unsigned long flags, unsigned long size, 1094 vm_flags_t vm_flags) 1095 { 1096 loff_t off_end = off + len; 1097 loff_t off_align = round_up(off, size); 1098 unsigned long len_pad, ret, off_sub; 1099 1100 if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall()) 1101 return 0; 1102 1103 if (off_end <= off_align || (off_end - off_align) < size) 1104 return 0; 1105 1106 len_pad = len + size; 1107 if (len_pad < len || (off + len_pad) < off) 1108 return 0; 1109 1110 ret = mm_get_unmapped_area_vmflags(current->mm, filp, addr, len_pad, 1111 off >> PAGE_SHIFT, flags, vm_flags); 1112 1113 /* 1114 * The failure might be due to length padding. The caller will retry 1115 * without the padding. 1116 */ 1117 if (IS_ERR_VALUE(ret)) 1118 return 0; 1119 1120 /* 1121 * Do not try to align to THP boundary if allocation at the address 1122 * hint succeeds. 1123 */ 1124 if (ret == addr) 1125 return addr; 1126 1127 off_sub = (off - ret) & (size - 1); 1128 1129 if (test_bit(MMF_TOPDOWN, ¤t->mm->flags) && !off_sub) 1130 return ret + size; 1131 1132 ret += off_sub; 1133 return ret; 1134 } 1135 1136 unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr, 1137 unsigned long len, unsigned long pgoff, unsigned long flags, 1138 vm_flags_t vm_flags) 1139 { 1140 unsigned long ret; 1141 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 1142 1143 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE, vm_flags); 1144 if (ret) 1145 return ret; 1146 1147 return mm_get_unmapped_area_vmflags(current->mm, filp, addr, len, pgoff, flags, 1148 vm_flags); 1149 } 1150 1151 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 1152 unsigned long len, unsigned long pgoff, unsigned long flags) 1153 { 1154 return thp_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags, 0); 1155 } 1156 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 1157 1158 static struct folio *vma_alloc_anon_folio_pmd(struct vm_area_struct *vma, 1159 unsigned long addr) 1160 { 1161 gfp_t gfp = vma_thp_gfp_mask(vma); 1162 const int order = HPAGE_PMD_ORDER; 1163 struct folio *folio; 1164 1165 folio = vma_alloc_folio(gfp, order, vma, addr & HPAGE_PMD_MASK); 1166 1167 if (unlikely(!folio)) { 1168 count_vm_event(THP_FAULT_FALLBACK); 1169 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK); 1170 return NULL; 1171 } 1172 1173 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 1174 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { 1175 folio_put(folio); 1176 count_vm_event(THP_FAULT_FALLBACK); 1177 count_vm_event(THP_FAULT_FALLBACK_CHARGE); 1178 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK); 1179 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE); 1180 return NULL; 1181 } 1182 folio_throttle_swaprate(folio, gfp); 1183 1184 /* 1185 * When a folio is not zeroed during allocation (__GFP_ZERO not used) 1186 * or user folios require special handling, folio_zero_user() is used to 1187 * make sure that the page corresponding to the faulting address will be 1188 * hot in the cache after zeroing. 1189 */ 1190 if (user_alloc_needs_zeroing()) 1191 folio_zero_user(folio, addr); 1192 /* 1193 * The memory barrier inside __folio_mark_uptodate makes sure that 1194 * folio_zero_user writes become visible before the set_pmd_at() 1195 * write. 1196 */ 1197 __folio_mark_uptodate(folio); 1198 return folio; 1199 } 1200 1201 static void map_anon_folio_pmd(struct folio *folio, pmd_t *pmd, 1202 struct vm_area_struct *vma, unsigned long haddr) 1203 { 1204 pmd_t entry; 1205 1206 entry = folio_mk_pmd(folio, vma->vm_page_prot); 1207 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1208 folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE); 1209 folio_add_lru_vma(folio, vma); 1210 set_pmd_at(vma->vm_mm, haddr, pmd, entry); 1211 update_mmu_cache_pmd(vma, haddr, pmd); 1212 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1213 count_vm_event(THP_FAULT_ALLOC); 1214 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC); 1215 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 1216 } 1217 1218 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf) 1219 { 1220 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1221 struct vm_area_struct *vma = vmf->vma; 1222 struct folio *folio; 1223 pgtable_t pgtable; 1224 vm_fault_t ret = 0; 1225 1226 folio = vma_alloc_anon_folio_pmd(vma, vmf->address); 1227 if (unlikely(!folio)) 1228 return VM_FAULT_FALLBACK; 1229 1230 pgtable = pte_alloc_one(vma->vm_mm); 1231 if (unlikely(!pgtable)) { 1232 ret = VM_FAULT_OOM; 1233 goto release; 1234 } 1235 1236 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1237 if (unlikely(!pmd_none(*vmf->pmd))) { 1238 goto unlock_release; 1239 } else { 1240 ret = check_stable_address_space(vma->vm_mm); 1241 if (ret) 1242 goto unlock_release; 1243 1244 /* Deliver the page fault to userland */ 1245 if (userfaultfd_missing(vma)) { 1246 spin_unlock(vmf->ptl); 1247 folio_put(folio); 1248 pte_free(vma->vm_mm, pgtable); 1249 ret = handle_userfault(vmf, VM_UFFD_MISSING); 1250 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 1251 return ret; 1252 } 1253 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 1254 map_anon_folio_pmd(folio, vmf->pmd, vma, haddr); 1255 mm_inc_nr_ptes(vma->vm_mm); 1256 deferred_split_folio(folio, false); 1257 spin_unlock(vmf->ptl); 1258 } 1259 1260 return 0; 1261 unlock_release: 1262 spin_unlock(vmf->ptl); 1263 release: 1264 if (pgtable) 1265 pte_free(vma->vm_mm, pgtable); 1266 folio_put(folio); 1267 return ret; 1268 1269 } 1270 1271 /* 1272 * always: directly stall for all thp allocations 1273 * defer: wake kswapd and fail if not immediately available 1274 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 1275 * fail if not immediately available 1276 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 1277 * available 1278 * never: never stall for any thp allocation 1279 */ 1280 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 1281 { 1282 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 1283 1284 /* Always do synchronous compaction */ 1285 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 1286 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 1287 1288 /* Kick kcompactd and fail quickly */ 1289 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 1290 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 1291 1292 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 1293 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 1294 return GFP_TRANSHUGE_LIGHT | 1295 (vma_madvised ? __GFP_DIRECT_RECLAIM : 1296 __GFP_KSWAPD_RECLAIM); 1297 1298 /* Only do synchronous compaction if madvised */ 1299 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 1300 return GFP_TRANSHUGE_LIGHT | 1301 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 1302 1303 return GFP_TRANSHUGE_LIGHT; 1304 } 1305 1306 /* Caller must hold page table lock. */ 1307 static void set_huge_zero_folio(pgtable_t pgtable, struct mm_struct *mm, 1308 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 1309 struct folio *zero_folio) 1310 { 1311 pmd_t entry; 1312 entry = folio_mk_pmd(zero_folio, vma->vm_page_prot); 1313 pgtable_trans_huge_deposit(mm, pmd, pgtable); 1314 set_pmd_at(mm, haddr, pmd, entry); 1315 mm_inc_nr_ptes(mm); 1316 } 1317 1318 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 1319 { 1320 struct vm_area_struct *vma = vmf->vma; 1321 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1322 vm_fault_t ret; 1323 1324 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER)) 1325 return VM_FAULT_FALLBACK; 1326 ret = vmf_anon_prepare(vmf); 1327 if (ret) 1328 return ret; 1329 khugepaged_enter_vma(vma, vma->vm_flags); 1330 1331 if (!(vmf->flags & FAULT_FLAG_WRITE) && 1332 !mm_forbids_zeropage(vma->vm_mm) && 1333 transparent_hugepage_use_zero_page()) { 1334 pgtable_t pgtable; 1335 struct folio *zero_folio; 1336 vm_fault_t ret; 1337 1338 pgtable = pte_alloc_one(vma->vm_mm); 1339 if (unlikely(!pgtable)) 1340 return VM_FAULT_OOM; 1341 zero_folio = mm_get_huge_zero_folio(vma->vm_mm); 1342 if (unlikely(!zero_folio)) { 1343 pte_free(vma->vm_mm, pgtable); 1344 count_vm_event(THP_FAULT_FALLBACK); 1345 return VM_FAULT_FALLBACK; 1346 } 1347 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1348 ret = 0; 1349 if (pmd_none(*vmf->pmd)) { 1350 ret = check_stable_address_space(vma->vm_mm); 1351 if (ret) { 1352 spin_unlock(vmf->ptl); 1353 pte_free(vma->vm_mm, pgtable); 1354 } else if (userfaultfd_missing(vma)) { 1355 spin_unlock(vmf->ptl); 1356 pte_free(vma->vm_mm, pgtable); 1357 ret = handle_userfault(vmf, VM_UFFD_MISSING); 1358 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 1359 } else { 1360 set_huge_zero_folio(pgtable, vma->vm_mm, vma, 1361 haddr, vmf->pmd, zero_folio); 1362 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1363 spin_unlock(vmf->ptl); 1364 } 1365 } else { 1366 spin_unlock(vmf->ptl); 1367 pte_free(vma->vm_mm, pgtable); 1368 } 1369 return ret; 1370 } 1371 1372 return __do_huge_pmd_anonymous_page(vmf); 1373 } 1374 1375 struct folio_or_pfn { 1376 union { 1377 struct folio *folio; 1378 pfn_t pfn; 1379 }; 1380 bool is_folio; 1381 }; 1382 1383 static int insert_pmd(struct vm_area_struct *vma, unsigned long addr, 1384 pmd_t *pmd, struct folio_or_pfn fop, pgprot_t prot, 1385 bool write, pgtable_t pgtable) 1386 { 1387 struct mm_struct *mm = vma->vm_mm; 1388 pmd_t entry; 1389 1390 lockdep_assert_held(pmd_lockptr(mm, pmd)); 1391 1392 if (!pmd_none(*pmd)) { 1393 const unsigned long pfn = fop.is_folio ? folio_pfn(fop.folio) : 1394 pfn_t_to_pfn(fop.pfn); 1395 1396 if (write) { 1397 if (pmd_pfn(*pmd) != pfn) { 1398 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 1399 return -EEXIST; 1400 } 1401 entry = pmd_mkyoung(*pmd); 1402 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1403 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 1404 update_mmu_cache_pmd(vma, addr, pmd); 1405 } 1406 1407 return -EEXIST; 1408 } 1409 1410 if (fop.is_folio) { 1411 entry = folio_mk_pmd(fop.folio, vma->vm_page_prot); 1412 1413 folio_get(fop.folio); 1414 folio_add_file_rmap_pmd(fop.folio, &fop.folio->page, vma); 1415 add_mm_counter(mm, mm_counter_file(fop.folio), HPAGE_PMD_NR); 1416 } else { 1417 entry = pmd_mkhuge(pfn_t_pmd(fop.pfn, prot)); 1418 entry = pmd_mkspecial(entry); 1419 } 1420 if (write) { 1421 entry = pmd_mkyoung(pmd_mkdirty(entry)); 1422 entry = maybe_pmd_mkwrite(entry, vma); 1423 } 1424 1425 if (pgtable) { 1426 pgtable_trans_huge_deposit(mm, pmd, pgtable); 1427 mm_inc_nr_ptes(mm); 1428 } 1429 1430 set_pmd_at(mm, addr, pmd, entry); 1431 update_mmu_cache_pmd(vma, addr, pmd); 1432 return 0; 1433 } 1434 1435 /** 1436 * vmf_insert_pfn_pmd - insert a pmd size pfn 1437 * @vmf: Structure describing the fault 1438 * @pfn: pfn to insert 1439 * @write: whether it's a write fault 1440 * 1441 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info. 1442 * 1443 * Return: vm_fault_t value. 1444 */ 1445 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write) 1446 { 1447 unsigned long addr = vmf->address & PMD_MASK; 1448 struct vm_area_struct *vma = vmf->vma; 1449 pgprot_t pgprot = vma->vm_page_prot; 1450 struct folio_or_pfn fop = { 1451 .pfn = pfn, 1452 }; 1453 pgtable_t pgtable = NULL; 1454 spinlock_t *ptl; 1455 int error; 1456 1457 /* 1458 * If we had pmd_special, we could avoid all these restrictions, 1459 * but we need to be consistent with PTEs and architectures that 1460 * can't support a 'special' bit. 1461 */ 1462 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 1463 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1464 (VM_PFNMAP|VM_MIXEDMAP)); 1465 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1466 1467 if (addr < vma->vm_start || addr >= vma->vm_end) 1468 return VM_FAULT_SIGBUS; 1469 1470 if (arch_needs_pgtable_deposit()) { 1471 pgtable = pte_alloc_one(vma->vm_mm); 1472 if (!pgtable) 1473 return VM_FAULT_OOM; 1474 } 1475 1476 pfnmap_setup_cachemode_pfn(pfn_t_to_pfn(pfn), &pgprot); 1477 1478 ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1479 error = insert_pmd(vma, addr, vmf->pmd, fop, pgprot, write, 1480 pgtable); 1481 spin_unlock(ptl); 1482 if (error && pgtable) 1483 pte_free(vma->vm_mm, pgtable); 1484 1485 return VM_FAULT_NOPAGE; 1486 } 1487 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 1488 1489 vm_fault_t vmf_insert_folio_pmd(struct vm_fault *vmf, struct folio *folio, 1490 bool write) 1491 { 1492 struct vm_area_struct *vma = vmf->vma; 1493 unsigned long addr = vmf->address & PMD_MASK; 1494 struct mm_struct *mm = vma->vm_mm; 1495 struct folio_or_pfn fop = { 1496 .folio = folio, 1497 .is_folio = true, 1498 }; 1499 spinlock_t *ptl; 1500 pgtable_t pgtable = NULL; 1501 int error; 1502 1503 if (addr < vma->vm_start || addr >= vma->vm_end) 1504 return VM_FAULT_SIGBUS; 1505 1506 if (WARN_ON_ONCE(folio_order(folio) != PMD_ORDER)) 1507 return VM_FAULT_SIGBUS; 1508 1509 if (arch_needs_pgtable_deposit()) { 1510 pgtable = pte_alloc_one(vma->vm_mm); 1511 if (!pgtable) 1512 return VM_FAULT_OOM; 1513 } 1514 1515 ptl = pmd_lock(mm, vmf->pmd); 1516 error = insert_pmd(vma, addr, vmf->pmd, fop, vma->vm_page_prot, 1517 write, pgtable); 1518 spin_unlock(ptl); 1519 if (error && pgtable) 1520 pte_free(mm, pgtable); 1521 1522 return VM_FAULT_NOPAGE; 1523 } 1524 EXPORT_SYMBOL_GPL(vmf_insert_folio_pmd); 1525 1526 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1527 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 1528 { 1529 if (likely(vma->vm_flags & VM_WRITE)) 1530 pud = pud_mkwrite(pud); 1531 return pud; 1532 } 1533 1534 static void insert_pud(struct vm_area_struct *vma, unsigned long addr, 1535 pud_t *pud, struct folio_or_pfn fop, pgprot_t prot, bool write) 1536 { 1537 struct mm_struct *mm = vma->vm_mm; 1538 pud_t entry; 1539 1540 if (!pud_none(*pud)) { 1541 const unsigned long pfn = fop.is_folio ? folio_pfn(fop.folio) : 1542 pfn_t_to_pfn(fop.pfn); 1543 1544 if (write) { 1545 if (WARN_ON_ONCE(pud_pfn(*pud) != pfn)) 1546 return; 1547 entry = pud_mkyoung(*pud); 1548 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 1549 if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 1550 update_mmu_cache_pud(vma, addr, pud); 1551 } 1552 return; 1553 } 1554 1555 if (fop.is_folio) { 1556 entry = folio_mk_pud(fop.folio, vma->vm_page_prot); 1557 1558 folio_get(fop.folio); 1559 folio_add_file_rmap_pud(fop.folio, &fop.folio->page, vma); 1560 add_mm_counter(mm, mm_counter_file(fop.folio), HPAGE_PUD_NR); 1561 } else { 1562 entry = pud_mkhuge(pfn_t_pud(fop.pfn, prot)); 1563 entry = pud_mkspecial(entry); 1564 } 1565 if (write) { 1566 entry = pud_mkyoung(pud_mkdirty(entry)); 1567 entry = maybe_pud_mkwrite(entry, vma); 1568 } 1569 set_pud_at(mm, addr, pud, entry); 1570 update_mmu_cache_pud(vma, addr, pud); 1571 } 1572 1573 /** 1574 * vmf_insert_pfn_pud - insert a pud size pfn 1575 * @vmf: Structure describing the fault 1576 * @pfn: pfn to insert 1577 * @write: whether it's a write fault 1578 * 1579 * Insert a pud size pfn. See vmf_insert_pfn() for additional info. 1580 * 1581 * Return: vm_fault_t value. 1582 */ 1583 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write) 1584 { 1585 unsigned long addr = vmf->address & PUD_MASK; 1586 struct vm_area_struct *vma = vmf->vma; 1587 pgprot_t pgprot = vma->vm_page_prot; 1588 struct folio_or_pfn fop = { 1589 .pfn = pfn, 1590 }; 1591 spinlock_t *ptl; 1592 1593 /* 1594 * If we had pud_special, we could avoid all these restrictions, 1595 * but we need to be consistent with PTEs and architectures that 1596 * can't support a 'special' bit. 1597 */ 1598 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 1599 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1600 (VM_PFNMAP|VM_MIXEDMAP)); 1601 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1602 1603 if (addr < vma->vm_start || addr >= vma->vm_end) 1604 return VM_FAULT_SIGBUS; 1605 1606 pfnmap_setup_cachemode_pfn(pfn_t_to_pfn(pfn), &pgprot); 1607 1608 ptl = pud_lock(vma->vm_mm, vmf->pud); 1609 insert_pud(vma, addr, vmf->pud, fop, pgprot, write); 1610 spin_unlock(ptl); 1611 1612 return VM_FAULT_NOPAGE; 1613 } 1614 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 1615 1616 /** 1617 * vmf_insert_folio_pud - insert a pud size folio mapped by a pud entry 1618 * @vmf: Structure describing the fault 1619 * @folio: folio to insert 1620 * @write: whether it's a write fault 1621 * 1622 * Return: vm_fault_t value. 1623 */ 1624 vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio, 1625 bool write) 1626 { 1627 struct vm_area_struct *vma = vmf->vma; 1628 unsigned long addr = vmf->address & PUD_MASK; 1629 pud_t *pud = vmf->pud; 1630 struct mm_struct *mm = vma->vm_mm; 1631 struct folio_or_pfn fop = { 1632 .folio = folio, 1633 .is_folio = true, 1634 }; 1635 spinlock_t *ptl; 1636 1637 if (addr < vma->vm_start || addr >= vma->vm_end) 1638 return VM_FAULT_SIGBUS; 1639 1640 if (WARN_ON_ONCE(folio_order(folio) != PUD_ORDER)) 1641 return VM_FAULT_SIGBUS; 1642 1643 ptl = pud_lock(mm, pud); 1644 insert_pud(vma, addr, vmf->pud, fop, vma->vm_page_prot, write); 1645 spin_unlock(ptl); 1646 1647 return VM_FAULT_NOPAGE; 1648 } 1649 EXPORT_SYMBOL_GPL(vmf_insert_folio_pud); 1650 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1651 1652 void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 1653 pmd_t *pmd, bool write) 1654 { 1655 pmd_t _pmd; 1656 1657 _pmd = pmd_mkyoung(*pmd); 1658 if (write) 1659 _pmd = pmd_mkdirty(_pmd); 1660 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1661 pmd, _pmd, write)) 1662 update_mmu_cache_pmd(vma, addr, pmd); 1663 } 1664 1665 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1666 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1667 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1668 { 1669 spinlock_t *dst_ptl, *src_ptl; 1670 struct page *src_page; 1671 struct folio *src_folio; 1672 pmd_t pmd; 1673 pgtable_t pgtable = NULL; 1674 int ret = -ENOMEM; 1675 1676 pmd = pmdp_get_lockless(src_pmd); 1677 if (unlikely(pmd_present(pmd) && pmd_special(pmd))) { 1678 dst_ptl = pmd_lock(dst_mm, dst_pmd); 1679 src_ptl = pmd_lockptr(src_mm, src_pmd); 1680 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1681 /* 1682 * No need to recheck the pmd, it can't change with write 1683 * mmap lock held here. 1684 * 1685 * Meanwhile, making sure it's not a CoW VMA with writable 1686 * mapping, otherwise it means either the anon page wrongly 1687 * applied special bit, or we made the PRIVATE mapping be 1688 * able to wrongly write to the backend MMIO. 1689 */ 1690 VM_WARN_ON_ONCE(is_cow_mapping(src_vma->vm_flags) && pmd_write(pmd)); 1691 goto set_pmd; 1692 } 1693 1694 /* Skip if can be re-fill on fault */ 1695 if (!vma_is_anonymous(dst_vma)) 1696 return 0; 1697 1698 pgtable = pte_alloc_one(dst_mm); 1699 if (unlikely(!pgtable)) 1700 goto out; 1701 1702 dst_ptl = pmd_lock(dst_mm, dst_pmd); 1703 src_ptl = pmd_lockptr(src_mm, src_pmd); 1704 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1705 1706 ret = -EAGAIN; 1707 pmd = *src_pmd; 1708 1709 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1710 if (unlikely(is_swap_pmd(pmd))) { 1711 swp_entry_t entry = pmd_to_swp_entry(pmd); 1712 1713 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 1714 if (!is_readable_migration_entry(entry)) { 1715 entry = make_readable_migration_entry( 1716 swp_offset(entry)); 1717 pmd = swp_entry_to_pmd(entry); 1718 if (pmd_swp_soft_dirty(*src_pmd)) 1719 pmd = pmd_swp_mksoft_dirty(pmd); 1720 if (pmd_swp_uffd_wp(*src_pmd)) 1721 pmd = pmd_swp_mkuffd_wp(pmd); 1722 set_pmd_at(src_mm, addr, src_pmd, pmd); 1723 } 1724 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1725 mm_inc_nr_ptes(dst_mm); 1726 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1727 if (!userfaultfd_wp(dst_vma)) 1728 pmd = pmd_swp_clear_uffd_wp(pmd); 1729 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1730 ret = 0; 1731 goto out_unlock; 1732 } 1733 #endif 1734 1735 if (unlikely(!pmd_trans_huge(pmd))) { 1736 pte_free(dst_mm, pgtable); 1737 goto out_unlock; 1738 } 1739 /* 1740 * When page table lock is held, the huge zero pmd should not be 1741 * under splitting since we don't split the page itself, only pmd to 1742 * a page table. 1743 */ 1744 if (is_huge_zero_pmd(pmd)) { 1745 /* 1746 * mm_get_huge_zero_folio() will never allocate a new 1747 * folio here, since we already have a zero page to 1748 * copy. It just takes a reference. 1749 */ 1750 mm_get_huge_zero_folio(dst_mm); 1751 goto out_zero_page; 1752 } 1753 1754 src_page = pmd_page(pmd); 1755 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 1756 src_folio = page_folio(src_page); 1757 1758 folio_get(src_folio); 1759 if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, dst_vma, src_vma))) { 1760 /* Page maybe pinned: split and retry the fault on PTEs. */ 1761 folio_put(src_folio); 1762 pte_free(dst_mm, pgtable); 1763 spin_unlock(src_ptl); 1764 spin_unlock(dst_ptl); 1765 __split_huge_pmd(src_vma, src_pmd, addr, false); 1766 return -EAGAIN; 1767 } 1768 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1769 out_zero_page: 1770 mm_inc_nr_ptes(dst_mm); 1771 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1772 pmdp_set_wrprotect(src_mm, addr, src_pmd); 1773 if (!userfaultfd_wp(dst_vma)) 1774 pmd = pmd_clear_uffd_wp(pmd); 1775 pmd = pmd_wrprotect(pmd); 1776 set_pmd: 1777 pmd = pmd_mkold(pmd); 1778 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1779 1780 ret = 0; 1781 out_unlock: 1782 spin_unlock(src_ptl); 1783 spin_unlock(dst_ptl); 1784 out: 1785 return ret; 1786 } 1787 1788 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1789 void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1790 pud_t *pud, bool write) 1791 { 1792 pud_t _pud; 1793 1794 _pud = pud_mkyoung(*pud); 1795 if (write) 1796 _pud = pud_mkdirty(_pud); 1797 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1798 pud, _pud, write)) 1799 update_mmu_cache_pud(vma, addr, pud); 1800 } 1801 1802 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1803 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1804 struct vm_area_struct *vma) 1805 { 1806 spinlock_t *dst_ptl, *src_ptl; 1807 pud_t pud; 1808 int ret; 1809 1810 dst_ptl = pud_lock(dst_mm, dst_pud); 1811 src_ptl = pud_lockptr(src_mm, src_pud); 1812 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1813 1814 ret = -EAGAIN; 1815 pud = *src_pud; 1816 if (unlikely(!pud_trans_huge(pud))) 1817 goto out_unlock; 1818 1819 /* 1820 * TODO: once we support anonymous pages, use 1821 * folio_try_dup_anon_rmap_*() and split if duplicating fails. 1822 */ 1823 if (is_cow_mapping(vma->vm_flags) && pud_write(pud)) { 1824 pudp_set_wrprotect(src_mm, addr, src_pud); 1825 pud = pud_wrprotect(pud); 1826 } 1827 pud = pud_mkold(pud); 1828 set_pud_at(dst_mm, addr, dst_pud, pud); 1829 1830 ret = 0; 1831 out_unlock: 1832 spin_unlock(src_ptl); 1833 spin_unlock(dst_ptl); 1834 return ret; 1835 } 1836 1837 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1838 { 1839 bool write = vmf->flags & FAULT_FLAG_WRITE; 1840 1841 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1842 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1843 goto unlock; 1844 1845 touch_pud(vmf->vma, vmf->address, vmf->pud, write); 1846 unlock: 1847 spin_unlock(vmf->ptl); 1848 } 1849 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1850 1851 void huge_pmd_set_accessed(struct vm_fault *vmf) 1852 { 1853 bool write = vmf->flags & FAULT_FLAG_WRITE; 1854 1855 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1856 if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd))) 1857 goto unlock; 1858 1859 touch_pmd(vmf->vma, vmf->address, vmf->pmd, write); 1860 1861 unlock: 1862 spin_unlock(vmf->ptl); 1863 } 1864 1865 static vm_fault_t do_huge_zero_wp_pmd(struct vm_fault *vmf) 1866 { 1867 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1868 struct vm_area_struct *vma = vmf->vma; 1869 struct mmu_notifier_range range; 1870 struct folio *folio; 1871 vm_fault_t ret = 0; 1872 1873 folio = vma_alloc_anon_folio_pmd(vma, vmf->address); 1874 if (unlikely(!folio)) 1875 return VM_FAULT_FALLBACK; 1876 1877 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, haddr, 1878 haddr + HPAGE_PMD_SIZE); 1879 mmu_notifier_invalidate_range_start(&range); 1880 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1881 if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd))) 1882 goto release; 1883 ret = check_stable_address_space(vma->vm_mm); 1884 if (ret) 1885 goto release; 1886 (void)pmdp_huge_clear_flush(vma, haddr, vmf->pmd); 1887 map_anon_folio_pmd(folio, vmf->pmd, vma, haddr); 1888 goto unlock; 1889 release: 1890 folio_put(folio); 1891 unlock: 1892 spin_unlock(vmf->ptl); 1893 mmu_notifier_invalidate_range_end(&range); 1894 return ret; 1895 } 1896 1897 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf) 1898 { 1899 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 1900 struct vm_area_struct *vma = vmf->vma; 1901 struct folio *folio; 1902 struct page *page; 1903 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1904 pmd_t orig_pmd = vmf->orig_pmd; 1905 1906 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1907 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1908 1909 if (is_huge_zero_pmd(orig_pmd)) { 1910 vm_fault_t ret = do_huge_zero_wp_pmd(vmf); 1911 1912 if (!(ret & VM_FAULT_FALLBACK)) 1913 return ret; 1914 1915 /* Fallback to splitting PMD if THP cannot be allocated */ 1916 goto fallback; 1917 } 1918 1919 spin_lock(vmf->ptl); 1920 1921 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1922 spin_unlock(vmf->ptl); 1923 return 0; 1924 } 1925 1926 page = pmd_page(orig_pmd); 1927 folio = page_folio(page); 1928 VM_BUG_ON_PAGE(!PageHead(page), page); 1929 1930 /* Early check when only holding the PT lock. */ 1931 if (PageAnonExclusive(page)) 1932 goto reuse; 1933 1934 if (!folio_trylock(folio)) { 1935 folio_get(folio); 1936 spin_unlock(vmf->ptl); 1937 folio_lock(folio); 1938 spin_lock(vmf->ptl); 1939 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1940 spin_unlock(vmf->ptl); 1941 folio_unlock(folio); 1942 folio_put(folio); 1943 return 0; 1944 } 1945 folio_put(folio); 1946 } 1947 1948 /* Recheck after temporarily dropping the PT lock. */ 1949 if (PageAnonExclusive(page)) { 1950 folio_unlock(folio); 1951 goto reuse; 1952 } 1953 1954 /* 1955 * See do_wp_page(): we can only reuse the folio exclusively if 1956 * there are no additional references. Note that we always drain 1957 * the LRU cache immediately after adding a THP. 1958 */ 1959 if (folio_ref_count(folio) > 1960 1 + folio_test_swapcache(folio) * folio_nr_pages(folio)) 1961 goto unlock_fallback; 1962 if (folio_test_swapcache(folio)) 1963 folio_free_swap(folio); 1964 if (folio_ref_count(folio) == 1) { 1965 pmd_t entry; 1966 1967 folio_move_anon_rmap(folio, vma); 1968 SetPageAnonExclusive(page); 1969 folio_unlock(folio); 1970 reuse: 1971 if (unlikely(unshare)) { 1972 spin_unlock(vmf->ptl); 1973 return 0; 1974 } 1975 entry = pmd_mkyoung(orig_pmd); 1976 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1977 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1978 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1979 spin_unlock(vmf->ptl); 1980 return 0; 1981 } 1982 1983 unlock_fallback: 1984 folio_unlock(folio); 1985 spin_unlock(vmf->ptl); 1986 fallback: 1987 __split_huge_pmd(vma, vmf->pmd, vmf->address, false); 1988 return VM_FAULT_FALLBACK; 1989 } 1990 1991 static inline bool can_change_pmd_writable(struct vm_area_struct *vma, 1992 unsigned long addr, pmd_t pmd) 1993 { 1994 struct page *page; 1995 1996 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE))) 1997 return false; 1998 1999 /* Don't touch entries that are not even readable (NUMA hinting). */ 2000 if (pmd_protnone(pmd)) 2001 return false; 2002 2003 /* Do we need write faults for softdirty tracking? */ 2004 if (pmd_needs_soft_dirty_wp(vma, pmd)) 2005 return false; 2006 2007 /* Do we need write faults for uffd-wp tracking? */ 2008 if (userfaultfd_huge_pmd_wp(vma, pmd)) 2009 return false; 2010 2011 if (!(vma->vm_flags & VM_SHARED)) { 2012 /* See can_change_pte_writable(). */ 2013 page = vm_normal_page_pmd(vma, addr, pmd); 2014 return page && PageAnon(page) && PageAnonExclusive(page); 2015 } 2016 2017 /* See can_change_pte_writable(). */ 2018 return pmd_dirty(pmd); 2019 } 2020 2021 /* NUMA hinting page fault entry point for trans huge pmds */ 2022 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf) 2023 { 2024 struct vm_area_struct *vma = vmf->vma; 2025 struct folio *folio; 2026 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 2027 int nid = NUMA_NO_NODE; 2028 int target_nid, last_cpupid; 2029 pmd_t pmd, old_pmd; 2030 bool writable = false; 2031 int flags = 0; 2032 2033 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 2034 old_pmd = pmdp_get(vmf->pmd); 2035 2036 if (unlikely(!pmd_same(old_pmd, vmf->orig_pmd))) { 2037 spin_unlock(vmf->ptl); 2038 return 0; 2039 } 2040 2041 pmd = pmd_modify(old_pmd, vma->vm_page_prot); 2042 2043 /* 2044 * Detect now whether the PMD could be writable; this information 2045 * is only valid while holding the PT lock. 2046 */ 2047 writable = pmd_write(pmd); 2048 if (!writable && vma_wants_manual_pte_write_upgrade(vma) && 2049 can_change_pmd_writable(vma, vmf->address, pmd)) 2050 writable = true; 2051 2052 folio = vm_normal_folio_pmd(vma, haddr, pmd); 2053 if (!folio) 2054 goto out_map; 2055 2056 nid = folio_nid(folio); 2057 2058 target_nid = numa_migrate_check(folio, vmf, haddr, &flags, writable, 2059 &last_cpupid); 2060 if (target_nid == NUMA_NO_NODE) 2061 goto out_map; 2062 if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) { 2063 flags |= TNF_MIGRATE_FAIL; 2064 goto out_map; 2065 } 2066 /* The folio is isolated and isolation code holds a folio reference. */ 2067 spin_unlock(vmf->ptl); 2068 writable = false; 2069 2070 if (!migrate_misplaced_folio(folio, target_nid)) { 2071 flags |= TNF_MIGRATED; 2072 nid = target_nid; 2073 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags); 2074 return 0; 2075 } 2076 2077 flags |= TNF_MIGRATE_FAIL; 2078 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 2079 if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd))) { 2080 spin_unlock(vmf->ptl); 2081 return 0; 2082 } 2083 out_map: 2084 /* Restore the PMD */ 2085 pmd = pmd_modify(pmdp_get(vmf->pmd), vma->vm_page_prot); 2086 pmd = pmd_mkyoung(pmd); 2087 if (writable) 2088 pmd = pmd_mkwrite(pmd, vma); 2089 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 2090 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 2091 spin_unlock(vmf->ptl); 2092 2093 if (nid != NUMA_NO_NODE) 2094 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags); 2095 return 0; 2096 } 2097 2098 /* 2099 * Return true if we do MADV_FREE successfully on entire pmd page. 2100 * Otherwise, return false. 2101 */ 2102 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 2103 pmd_t *pmd, unsigned long addr, unsigned long next) 2104 { 2105 spinlock_t *ptl; 2106 pmd_t orig_pmd; 2107 struct folio *folio; 2108 struct mm_struct *mm = tlb->mm; 2109 bool ret = false; 2110 2111 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 2112 2113 ptl = pmd_trans_huge_lock(pmd, vma); 2114 if (!ptl) 2115 goto out_unlocked; 2116 2117 orig_pmd = *pmd; 2118 if (is_huge_zero_pmd(orig_pmd)) 2119 goto out; 2120 2121 if (unlikely(!pmd_present(orig_pmd))) { 2122 VM_BUG_ON(thp_migration_supported() && 2123 !is_pmd_migration_entry(orig_pmd)); 2124 goto out; 2125 } 2126 2127 folio = pmd_folio(orig_pmd); 2128 /* 2129 * If other processes are mapping this folio, we couldn't discard 2130 * the folio unless they all do MADV_FREE so let's skip the folio. 2131 */ 2132 if (folio_maybe_mapped_shared(folio)) 2133 goto out; 2134 2135 if (!folio_trylock(folio)) 2136 goto out; 2137 2138 /* 2139 * If user want to discard part-pages of THP, split it so MADV_FREE 2140 * will deactivate only them. 2141 */ 2142 if (next - addr != HPAGE_PMD_SIZE) { 2143 folio_get(folio); 2144 spin_unlock(ptl); 2145 split_folio(folio); 2146 folio_unlock(folio); 2147 folio_put(folio); 2148 goto out_unlocked; 2149 } 2150 2151 if (folio_test_dirty(folio)) 2152 folio_clear_dirty(folio); 2153 folio_unlock(folio); 2154 2155 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 2156 pmdp_invalidate(vma, addr, pmd); 2157 orig_pmd = pmd_mkold(orig_pmd); 2158 orig_pmd = pmd_mkclean(orig_pmd); 2159 2160 set_pmd_at(mm, addr, pmd, orig_pmd); 2161 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 2162 } 2163 2164 folio_mark_lazyfree(folio); 2165 ret = true; 2166 out: 2167 spin_unlock(ptl); 2168 out_unlocked: 2169 return ret; 2170 } 2171 2172 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 2173 { 2174 pgtable_t pgtable; 2175 2176 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2177 pte_free(mm, pgtable); 2178 mm_dec_nr_ptes(mm); 2179 } 2180 2181 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 2182 pmd_t *pmd, unsigned long addr) 2183 { 2184 pmd_t orig_pmd; 2185 spinlock_t *ptl; 2186 2187 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 2188 2189 ptl = __pmd_trans_huge_lock(pmd, vma); 2190 if (!ptl) 2191 return 0; 2192 /* 2193 * For architectures like ppc64 we look at deposited pgtable 2194 * when calling pmdp_huge_get_and_clear. So do the 2195 * pgtable_trans_huge_withdraw after finishing pmdp related 2196 * operations. 2197 */ 2198 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 2199 tlb->fullmm); 2200 arch_check_zapped_pmd(vma, orig_pmd); 2201 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 2202 if (!vma_is_dax(vma) && vma_is_special_huge(vma)) { 2203 if (arch_needs_pgtable_deposit()) 2204 zap_deposited_table(tlb->mm, pmd); 2205 spin_unlock(ptl); 2206 } else if (is_huge_zero_pmd(orig_pmd)) { 2207 if (!vma_is_dax(vma) || arch_needs_pgtable_deposit()) 2208 zap_deposited_table(tlb->mm, pmd); 2209 spin_unlock(ptl); 2210 } else { 2211 struct folio *folio = NULL; 2212 int flush_needed = 1; 2213 2214 if (pmd_present(orig_pmd)) { 2215 struct page *page = pmd_page(orig_pmd); 2216 2217 folio = page_folio(page); 2218 folio_remove_rmap_pmd(folio, page, vma); 2219 WARN_ON_ONCE(folio_mapcount(folio) < 0); 2220 VM_BUG_ON_PAGE(!PageHead(page), page); 2221 } else if (thp_migration_supported()) { 2222 swp_entry_t entry; 2223 2224 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 2225 entry = pmd_to_swp_entry(orig_pmd); 2226 folio = pfn_swap_entry_folio(entry); 2227 flush_needed = 0; 2228 } else 2229 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 2230 2231 if (folio_test_anon(folio)) { 2232 zap_deposited_table(tlb->mm, pmd); 2233 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 2234 } else { 2235 if (arch_needs_pgtable_deposit()) 2236 zap_deposited_table(tlb->mm, pmd); 2237 add_mm_counter(tlb->mm, mm_counter_file(folio), 2238 -HPAGE_PMD_NR); 2239 2240 /* 2241 * Use flush_needed to indicate whether the PMD entry 2242 * is present, instead of checking pmd_present() again. 2243 */ 2244 if (flush_needed && pmd_young(orig_pmd) && 2245 likely(vma_has_recency(vma))) 2246 folio_mark_accessed(folio); 2247 } 2248 2249 spin_unlock(ptl); 2250 if (flush_needed) 2251 tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE); 2252 } 2253 return 1; 2254 } 2255 2256 #ifndef pmd_move_must_withdraw 2257 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 2258 spinlock_t *old_pmd_ptl, 2259 struct vm_area_struct *vma) 2260 { 2261 /* 2262 * With split pmd lock we also need to move preallocated 2263 * PTE page table if new_pmd is on different PMD page table. 2264 * 2265 * We also don't deposit and withdraw tables for file pages. 2266 */ 2267 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 2268 } 2269 #endif 2270 2271 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 2272 { 2273 #ifdef CONFIG_MEM_SOFT_DIRTY 2274 if (unlikely(is_pmd_migration_entry(pmd))) 2275 pmd = pmd_swp_mksoft_dirty(pmd); 2276 else if (pmd_present(pmd)) 2277 pmd = pmd_mksoft_dirty(pmd); 2278 #endif 2279 return pmd; 2280 } 2281 2282 static pmd_t clear_uffd_wp_pmd(pmd_t pmd) 2283 { 2284 if (pmd_present(pmd)) 2285 pmd = pmd_clear_uffd_wp(pmd); 2286 else if (is_swap_pmd(pmd)) 2287 pmd = pmd_swp_clear_uffd_wp(pmd); 2288 2289 return pmd; 2290 } 2291 2292 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 2293 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 2294 { 2295 spinlock_t *old_ptl, *new_ptl; 2296 pmd_t pmd; 2297 struct mm_struct *mm = vma->vm_mm; 2298 bool force_flush = false; 2299 2300 /* 2301 * The destination pmd shouldn't be established, free_pgtables() 2302 * should have released it; but move_page_tables() might have already 2303 * inserted a page table, if racing against shmem/file collapse. 2304 */ 2305 if (!pmd_none(*new_pmd)) { 2306 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 2307 return false; 2308 } 2309 2310 /* 2311 * We don't have to worry about the ordering of src and dst 2312 * ptlocks because exclusive mmap_lock prevents deadlock. 2313 */ 2314 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 2315 if (old_ptl) { 2316 new_ptl = pmd_lockptr(mm, new_pmd); 2317 if (new_ptl != old_ptl) 2318 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 2319 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 2320 if (pmd_present(pmd)) 2321 force_flush = true; 2322 VM_BUG_ON(!pmd_none(*new_pmd)); 2323 2324 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 2325 pgtable_t pgtable; 2326 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 2327 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 2328 } 2329 pmd = move_soft_dirty_pmd(pmd); 2330 if (vma_has_uffd_without_event_remap(vma)) 2331 pmd = clear_uffd_wp_pmd(pmd); 2332 set_pmd_at(mm, new_addr, new_pmd, pmd); 2333 if (force_flush) 2334 flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 2335 if (new_ptl != old_ptl) 2336 spin_unlock(new_ptl); 2337 spin_unlock(old_ptl); 2338 return true; 2339 } 2340 return false; 2341 } 2342 2343 /* 2344 * Returns 2345 * - 0 if PMD could not be locked 2346 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 2347 * or if prot_numa but THP migration is not supported 2348 * - HPAGE_PMD_NR if protections changed and TLB flush necessary 2349 */ 2350 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 2351 pmd_t *pmd, unsigned long addr, pgprot_t newprot, 2352 unsigned long cp_flags) 2353 { 2354 struct mm_struct *mm = vma->vm_mm; 2355 spinlock_t *ptl; 2356 pmd_t oldpmd, entry; 2357 bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 2358 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 2359 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 2360 int ret = 1; 2361 2362 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 2363 2364 if (prot_numa && !thp_migration_supported()) 2365 return 1; 2366 2367 ptl = __pmd_trans_huge_lock(pmd, vma); 2368 if (!ptl) 2369 return 0; 2370 2371 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2372 if (is_swap_pmd(*pmd)) { 2373 swp_entry_t entry = pmd_to_swp_entry(*pmd); 2374 struct folio *folio = pfn_swap_entry_folio(entry); 2375 pmd_t newpmd; 2376 2377 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 2378 if (is_writable_migration_entry(entry)) { 2379 /* 2380 * A protection check is difficult so 2381 * just be safe and disable write 2382 */ 2383 if (folio_test_anon(folio)) 2384 entry = make_readable_exclusive_migration_entry(swp_offset(entry)); 2385 else 2386 entry = make_readable_migration_entry(swp_offset(entry)); 2387 newpmd = swp_entry_to_pmd(entry); 2388 if (pmd_swp_soft_dirty(*pmd)) 2389 newpmd = pmd_swp_mksoft_dirty(newpmd); 2390 } else { 2391 newpmd = *pmd; 2392 } 2393 2394 if (uffd_wp) 2395 newpmd = pmd_swp_mkuffd_wp(newpmd); 2396 else if (uffd_wp_resolve) 2397 newpmd = pmd_swp_clear_uffd_wp(newpmd); 2398 if (!pmd_same(*pmd, newpmd)) 2399 set_pmd_at(mm, addr, pmd, newpmd); 2400 goto unlock; 2401 } 2402 #endif 2403 2404 if (prot_numa) { 2405 struct folio *folio; 2406 bool toptier; 2407 /* 2408 * Avoid trapping faults against the zero page. The read-only 2409 * data is likely to be read-cached on the local CPU and 2410 * local/remote hits to the zero page are not interesting. 2411 */ 2412 if (is_huge_zero_pmd(*pmd)) 2413 goto unlock; 2414 2415 if (pmd_protnone(*pmd)) 2416 goto unlock; 2417 2418 folio = pmd_folio(*pmd); 2419 toptier = node_is_toptier(folio_nid(folio)); 2420 /* 2421 * Skip scanning top tier node if normal numa 2422 * balancing is disabled 2423 */ 2424 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 2425 toptier) 2426 goto unlock; 2427 2428 if (folio_use_access_time(folio)) 2429 folio_xchg_access_time(folio, 2430 jiffies_to_msecs(jiffies)); 2431 } 2432 /* 2433 * In case prot_numa, we are under mmap_read_lock(mm). It's critical 2434 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 2435 * which is also under mmap_read_lock(mm): 2436 * 2437 * CPU0: CPU1: 2438 * change_huge_pmd(prot_numa=1) 2439 * pmdp_huge_get_and_clear_notify() 2440 * madvise_dontneed() 2441 * zap_pmd_range() 2442 * pmd_trans_huge(*pmd) == 0 (without ptl) 2443 * // skip the pmd 2444 * set_pmd_at(); 2445 * // pmd is re-established 2446 * 2447 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 2448 * which may break userspace. 2449 * 2450 * pmdp_invalidate_ad() is required to make sure we don't miss 2451 * dirty/young flags set by hardware. 2452 */ 2453 oldpmd = pmdp_invalidate_ad(vma, addr, pmd); 2454 2455 entry = pmd_modify(oldpmd, newprot); 2456 if (uffd_wp) 2457 entry = pmd_mkuffd_wp(entry); 2458 else if (uffd_wp_resolve) 2459 /* 2460 * Leave the write bit to be handled by PF interrupt 2461 * handler, then things like COW could be properly 2462 * handled. 2463 */ 2464 entry = pmd_clear_uffd_wp(entry); 2465 2466 /* See change_pte_range(). */ 2467 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) && 2468 can_change_pmd_writable(vma, addr, entry)) 2469 entry = pmd_mkwrite(entry, vma); 2470 2471 ret = HPAGE_PMD_NR; 2472 set_pmd_at(mm, addr, pmd, entry); 2473 2474 if (huge_pmd_needs_flush(oldpmd, entry)) 2475 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE); 2476 unlock: 2477 spin_unlock(ptl); 2478 return ret; 2479 } 2480 2481 /* 2482 * Returns: 2483 * 2484 * - 0: if pud leaf changed from under us 2485 * - 1: if pud can be skipped 2486 * - HPAGE_PUD_NR: if pud was successfully processed 2487 */ 2488 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 2489 int change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 2490 pud_t *pudp, unsigned long addr, pgprot_t newprot, 2491 unsigned long cp_flags) 2492 { 2493 struct mm_struct *mm = vma->vm_mm; 2494 pud_t oldpud, entry; 2495 spinlock_t *ptl; 2496 2497 tlb_change_page_size(tlb, HPAGE_PUD_SIZE); 2498 2499 /* NUMA balancing doesn't apply to dax */ 2500 if (cp_flags & MM_CP_PROT_NUMA) 2501 return 1; 2502 2503 /* 2504 * Huge entries on userfault-wp only works with anonymous, while we 2505 * don't have anonymous PUDs yet. 2506 */ 2507 if (WARN_ON_ONCE(cp_flags & MM_CP_UFFD_WP_ALL)) 2508 return 1; 2509 2510 ptl = __pud_trans_huge_lock(pudp, vma); 2511 if (!ptl) 2512 return 0; 2513 2514 /* 2515 * Can't clear PUD or it can race with concurrent zapping. See 2516 * change_huge_pmd(). 2517 */ 2518 oldpud = pudp_invalidate(vma, addr, pudp); 2519 entry = pud_modify(oldpud, newprot); 2520 set_pud_at(mm, addr, pudp, entry); 2521 tlb_flush_pud_range(tlb, addr, HPAGE_PUD_SIZE); 2522 2523 spin_unlock(ptl); 2524 return HPAGE_PUD_NR; 2525 } 2526 #endif 2527 2528 #ifdef CONFIG_USERFAULTFD 2529 /* 2530 * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by 2531 * the caller, but it must return after releasing the page_table_lock. 2532 * Just move the page from src_pmd to dst_pmd if possible. 2533 * Return zero if succeeded in moving the page, -EAGAIN if it needs to be 2534 * repeated by the caller, or other errors in case of failure. 2535 */ 2536 int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval, 2537 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 2538 unsigned long dst_addr, unsigned long src_addr) 2539 { 2540 pmd_t _dst_pmd, src_pmdval; 2541 struct page *src_page; 2542 struct folio *src_folio; 2543 struct anon_vma *src_anon_vma; 2544 spinlock_t *src_ptl, *dst_ptl; 2545 pgtable_t src_pgtable; 2546 struct mmu_notifier_range range; 2547 int err = 0; 2548 2549 src_pmdval = *src_pmd; 2550 src_ptl = pmd_lockptr(mm, src_pmd); 2551 2552 lockdep_assert_held(src_ptl); 2553 vma_assert_locked(src_vma); 2554 vma_assert_locked(dst_vma); 2555 2556 /* Sanity checks before the operation */ 2557 if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) || 2558 WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) { 2559 spin_unlock(src_ptl); 2560 return -EINVAL; 2561 } 2562 2563 if (!pmd_trans_huge(src_pmdval)) { 2564 spin_unlock(src_ptl); 2565 if (is_pmd_migration_entry(src_pmdval)) { 2566 pmd_migration_entry_wait(mm, &src_pmdval); 2567 return -EAGAIN; 2568 } 2569 return -ENOENT; 2570 } 2571 2572 src_page = pmd_page(src_pmdval); 2573 2574 if (!is_huge_zero_pmd(src_pmdval)) { 2575 if (unlikely(!PageAnonExclusive(src_page))) { 2576 spin_unlock(src_ptl); 2577 return -EBUSY; 2578 } 2579 2580 src_folio = page_folio(src_page); 2581 folio_get(src_folio); 2582 } else 2583 src_folio = NULL; 2584 2585 spin_unlock(src_ptl); 2586 2587 flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE); 2588 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr, 2589 src_addr + HPAGE_PMD_SIZE); 2590 mmu_notifier_invalidate_range_start(&range); 2591 2592 if (src_folio) { 2593 folio_lock(src_folio); 2594 2595 /* 2596 * split_huge_page walks the anon_vma chain without the page 2597 * lock. Serialize against it with the anon_vma lock, the page 2598 * lock is not enough. 2599 */ 2600 src_anon_vma = folio_get_anon_vma(src_folio); 2601 if (!src_anon_vma) { 2602 err = -EAGAIN; 2603 goto unlock_folio; 2604 } 2605 anon_vma_lock_write(src_anon_vma); 2606 } else 2607 src_anon_vma = NULL; 2608 2609 dst_ptl = pmd_lockptr(mm, dst_pmd); 2610 double_pt_lock(src_ptl, dst_ptl); 2611 if (unlikely(!pmd_same(*src_pmd, src_pmdval) || 2612 !pmd_same(*dst_pmd, dst_pmdval))) { 2613 err = -EAGAIN; 2614 goto unlock_ptls; 2615 } 2616 if (src_folio) { 2617 if (folio_maybe_dma_pinned(src_folio) || 2618 !PageAnonExclusive(&src_folio->page)) { 2619 err = -EBUSY; 2620 goto unlock_ptls; 2621 } 2622 2623 if (WARN_ON_ONCE(!folio_test_head(src_folio)) || 2624 WARN_ON_ONCE(!folio_test_anon(src_folio))) { 2625 err = -EBUSY; 2626 goto unlock_ptls; 2627 } 2628 2629 src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd); 2630 /* Folio got pinned from under us. Put it back and fail the move. */ 2631 if (folio_maybe_dma_pinned(src_folio)) { 2632 set_pmd_at(mm, src_addr, src_pmd, src_pmdval); 2633 err = -EBUSY; 2634 goto unlock_ptls; 2635 } 2636 2637 folio_move_anon_rmap(src_folio, dst_vma); 2638 src_folio->index = linear_page_index(dst_vma, dst_addr); 2639 2640 _dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot); 2641 /* Follow mremap() behavior and treat the entry dirty after the move */ 2642 _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma); 2643 } else { 2644 src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd); 2645 _dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot); 2646 } 2647 set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd); 2648 2649 src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd); 2650 pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable); 2651 unlock_ptls: 2652 double_pt_unlock(src_ptl, dst_ptl); 2653 if (src_anon_vma) { 2654 anon_vma_unlock_write(src_anon_vma); 2655 put_anon_vma(src_anon_vma); 2656 } 2657 unlock_folio: 2658 /* unblock rmap walks */ 2659 if (src_folio) 2660 folio_unlock(src_folio); 2661 mmu_notifier_invalidate_range_end(&range); 2662 if (src_folio) 2663 folio_put(src_folio); 2664 return err; 2665 } 2666 #endif /* CONFIG_USERFAULTFD */ 2667 2668 /* 2669 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 2670 * 2671 * Note that if it returns page table lock pointer, this routine returns without 2672 * unlocking page table lock. So callers must unlock it. 2673 */ 2674 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 2675 { 2676 spinlock_t *ptl; 2677 ptl = pmd_lock(vma->vm_mm, pmd); 2678 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd))) 2679 return ptl; 2680 spin_unlock(ptl); 2681 return NULL; 2682 } 2683 2684 /* 2685 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise. 2686 * 2687 * Note that if it returns page table lock pointer, this routine returns without 2688 * unlocking page table lock. So callers must unlock it. 2689 */ 2690 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 2691 { 2692 spinlock_t *ptl; 2693 2694 ptl = pud_lock(vma->vm_mm, pud); 2695 if (likely(pud_trans_huge(*pud))) 2696 return ptl; 2697 spin_unlock(ptl); 2698 return NULL; 2699 } 2700 2701 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 2702 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 2703 pud_t *pud, unsigned long addr) 2704 { 2705 spinlock_t *ptl; 2706 pud_t orig_pud; 2707 2708 ptl = __pud_trans_huge_lock(pud, vma); 2709 if (!ptl) 2710 return 0; 2711 2712 orig_pud = pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm); 2713 arch_check_zapped_pud(vma, orig_pud); 2714 tlb_remove_pud_tlb_entry(tlb, pud, addr); 2715 if (!vma_is_dax(vma) && vma_is_special_huge(vma)) { 2716 spin_unlock(ptl); 2717 /* No zero page support yet */ 2718 } else { 2719 struct page *page = NULL; 2720 struct folio *folio; 2721 2722 /* No support for anonymous PUD pages or migration yet */ 2723 VM_WARN_ON_ONCE(vma_is_anonymous(vma) || 2724 !pud_present(orig_pud)); 2725 2726 page = pud_page(orig_pud); 2727 folio = page_folio(page); 2728 folio_remove_rmap_pud(folio, page, vma); 2729 add_mm_counter(tlb->mm, mm_counter_file(folio), -HPAGE_PUD_NR); 2730 2731 spin_unlock(ptl); 2732 tlb_remove_page_size(tlb, page, HPAGE_PUD_SIZE); 2733 } 2734 return 1; 2735 } 2736 2737 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2738 unsigned long haddr) 2739 { 2740 struct folio *folio; 2741 struct page *page; 2742 pud_t old_pud; 2743 2744 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2745 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2746 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2747 VM_BUG_ON(!pud_trans_huge(*pud)); 2748 2749 count_vm_event(THP_SPLIT_PUD); 2750 2751 old_pud = pudp_huge_clear_flush(vma, haddr, pud); 2752 2753 if (!vma_is_dax(vma)) 2754 return; 2755 2756 page = pud_page(old_pud); 2757 folio = page_folio(page); 2758 2759 if (!folio_test_dirty(folio) && pud_dirty(old_pud)) 2760 folio_mark_dirty(folio); 2761 if (!folio_test_referenced(folio) && pud_young(old_pud)) 2762 folio_set_referenced(folio); 2763 folio_remove_rmap_pud(folio, page, vma); 2764 folio_put(folio); 2765 add_mm_counter(vma->vm_mm, mm_counter_file(folio), 2766 -HPAGE_PUD_NR); 2767 } 2768 2769 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2770 unsigned long address) 2771 { 2772 spinlock_t *ptl; 2773 struct mmu_notifier_range range; 2774 2775 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 2776 address & HPAGE_PUD_MASK, 2777 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 2778 mmu_notifier_invalidate_range_start(&range); 2779 ptl = pud_lock(vma->vm_mm, pud); 2780 if (unlikely(!pud_trans_huge(*pud))) 2781 goto out; 2782 __split_huge_pud_locked(vma, pud, range.start); 2783 2784 out: 2785 spin_unlock(ptl); 2786 mmu_notifier_invalidate_range_end(&range); 2787 } 2788 #else 2789 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2790 unsigned long address) 2791 { 2792 } 2793 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2794 2795 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2796 unsigned long haddr, pmd_t *pmd) 2797 { 2798 struct mm_struct *mm = vma->vm_mm; 2799 pgtable_t pgtable; 2800 pmd_t _pmd, old_pmd; 2801 unsigned long addr; 2802 pte_t *pte; 2803 int i; 2804 2805 /* 2806 * Leave pmd empty until pte is filled note that it is fine to delay 2807 * notification until mmu_notifier_invalidate_range_end() as we are 2808 * replacing a zero pmd write protected page with a zero pte write 2809 * protected page. 2810 * 2811 * See Documentation/mm/mmu_notifier.rst 2812 */ 2813 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2814 2815 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2816 pmd_populate(mm, &_pmd, pgtable); 2817 2818 pte = pte_offset_map(&_pmd, haddr); 2819 VM_BUG_ON(!pte); 2820 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2821 pte_t entry; 2822 2823 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot); 2824 entry = pte_mkspecial(entry); 2825 if (pmd_uffd_wp(old_pmd)) 2826 entry = pte_mkuffd_wp(entry); 2827 VM_BUG_ON(!pte_none(ptep_get(pte))); 2828 set_pte_at(mm, addr, pte, entry); 2829 pte++; 2830 } 2831 pte_unmap(pte - 1); 2832 smp_wmb(); /* make pte visible before pmd */ 2833 pmd_populate(mm, pmd, pgtable); 2834 } 2835 2836 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2837 unsigned long haddr, bool freeze) 2838 { 2839 struct mm_struct *mm = vma->vm_mm; 2840 struct folio *folio; 2841 struct page *page; 2842 pgtable_t pgtable; 2843 pmd_t old_pmd, _pmd; 2844 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 2845 bool anon_exclusive = false, dirty = false; 2846 unsigned long addr; 2847 pte_t *pte; 2848 int i; 2849 2850 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2851 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2852 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2853 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)); 2854 2855 count_vm_event(THP_SPLIT_PMD); 2856 2857 if (!vma_is_anonymous(vma)) { 2858 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2859 /* 2860 * We are going to unmap this huge page. So 2861 * just go ahead and zap it 2862 */ 2863 if (arch_needs_pgtable_deposit()) 2864 zap_deposited_table(mm, pmd); 2865 if (!vma_is_dax(vma) && vma_is_special_huge(vma)) 2866 return; 2867 if (unlikely(is_pmd_migration_entry(old_pmd))) { 2868 swp_entry_t entry; 2869 2870 entry = pmd_to_swp_entry(old_pmd); 2871 folio = pfn_swap_entry_folio(entry); 2872 } else if (is_huge_zero_pmd(old_pmd)) { 2873 return; 2874 } else { 2875 page = pmd_page(old_pmd); 2876 folio = page_folio(page); 2877 if (!folio_test_dirty(folio) && pmd_dirty(old_pmd)) 2878 folio_mark_dirty(folio); 2879 if (!folio_test_referenced(folio) && pmd_young(old_pmd)) 2880 folio_set_referenced(folio); 2881 folio_remove_rmap_pmd(folio, page, vma); 2882 folio_put(folio); 2883 } 2884 add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR); 2885 return; 2886 } 2887 2888 if (is_huge_zero_pmd(*pmd)) { 2889 /* 2890 * FIXME: Do we want to invalidate secondary mmu by calling 2891 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below 2892 * inside __split_huge_pmd() ? 2893 * 2894 * We are going from a zero huge page write protected to zero 2895 * small page also write protected so it does not seems useful 2896 * to invalidate secondary mmu at this time. 2897 */ 2898 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2899 } 2900 2901 pmd_migration = is_pmd_migration_entry(*pmd); 2902 if (unlikely(pmd_migration)) { 2903 swp_entry_t entry; 2904 2905 old_pmd = *pmd; 2906 entry = pmd_to_swp_entry(old_pmd); 2907 page = pfn_swap_entry_to_page(entry); 2908 write = is_writable_migration_entry(entry); 2909 if (PageAnon(page)) 2910 anon_exclusive = is_readable_exclusive_migration_entry(entry); 2911 young = is_migration_entry_young(entry); 2912 dirty = is_migration_entry_dirty(entry); 2913 soft_dirty = pmd_swp_soft_dirty(old_pmd); 2914 uffd_wp = pmd_swp_uffd_wp(old_pmd); 2915 } else { 2916 /* 2917 * Up to this point the pmd is present and huge and userland has 2918 * the whole access to the hugepage during the split (which 2919 * happens in place). If we overwrite the pmd with the not-huge 2920 * version pointing to the pte here (which of course we could if 2921 * all CPUs were bug free), userland could trigger a small page 2922 * size TLB miss on the small sized TLB while the hugepage TLB 2923 * entry is still established in the huge TLB. Some CPU doesn't 2924 * like that. See 2925 * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 2926 * 383 on page 105. Intel should be safe but is also warns that 2927 * it's only safe if the permission and cache attributes of the 2928 * two entries loaded in the two TLB is identical (which should 2929 * be the case here). But it is generally safer to never allow 2930 * small and huge TLB entries for the same virtual address to be 2931 * loaded simultaneously. So instead of doing "pmd_populate(); 2932 * flush_pmd_tlb_range();" we first mark the current pmd 2933 * notpresent (atomically because here the pmd_trans_huge must 2934 * remain set at all times on the pmd until the split is 2935 * complete for this pmd), then we flush the SMP TLB and finally 2936 * we write the non-huge version of the pmd entry with 2937 * pmd_populate. 2938 */ 2939 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2940 page = pmd_page(old_pmd); 2941 folio = page_folio(page); 2942 if (pmd_dirty(old_pmd)) { 2943 dirty = true; 2944 folio_set_dirty(folio); 2945 } 2946 write = pmd_write(old_pmd); 2947 young = pmd_young(old_pmd); 2948 soft_dirty = pmd_soft_dirty(old_pmd); 2949 uffd_wp = pmd_uffd_wp(old_pmd); 2950 2951 VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio); 2952 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); 2953 2954 /* 2955 * Without "freeze", we'll simply split the PMD, propagating the 2956 * PageAnonExclusive() flag for each PTE by setting it for 2957 * each subpage -- no need to (temporarily) clear. 2958 * 2959 * With "freeze" we want to replace mapped pages by 2960 * migration entries right away. This is only possible if we 2961 * managed to clear PageAnonExclusive() -- see 2962 * set_pmd_migration_entry(). 2963 * 2964 * In case we cannot clear PageAnonExclusive(), split the PMD 2965 * only and let try_to_migrate_one() fail later. 2966 * 2967 * See folio_try_share_anon_rmap_pmd(): invalidate PMD first. 2968 */ 2969 anon_exclusive = PageAnonExclusive(page); 2970 if (freeze && anon_exclusive && 2971 folio_try_share_anon_rmap_pmd(folio, page)) 2972 freeze = false; 2973 if (!freeze) { 2974 rmap_t rmap_flags = RMAP_NONE; 2975 2976 folio_ref_add(folio, HPAGE_PMD_NR - 1); 2977 if (anon_exclusive) 2978 rmap_flags |= RMAP_EXCLUSIVE; 2979 folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR, 2980 vma, haddr, rmap_flags); 2981 } 2982 } 2983 2984 /* 2985 * Withdraw the table only after we mark the pmd entry invalid. 2986 * This's critical for some architectures (Power). 2987 */ 2988 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2989 pmd_populate(mm, &_pmd, pgtable); 2990 2991 pte = pte_offset_map(&_pmd, haddr); 2992 VM_BUG_ON(!pte); 2993 2994 /* 2995 * Note that NUMA hinting access restrictions are not transferred to 2996 * avoid any possibility of altering permissions across VMAs. 2997 */ 2998 if (freeze || pmd_migration) { 2999 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 3000 pte_t entry; 3001 swp_entry_t swp_entry; 3002 3003 if (write) 3004 swp_entry = make_writable_migration_entry( 3005 page_to_pfn(page + i)); 3006 else if (anon_exclusive) 3007 swp_entry = make_readable_exclusive_migration_entry( 3008 page_to_pfn(page + i)); 3009 else 3010 swp_entry = make_readable_migration_entry( 3011 page_to_pfn(page + i)); 3012 if (young) 3013 swp_entry = make_migration_entry_young(swp_entry); 3014 if (dirty) 3015 swp_entry = make_migration_entry_dirty(swp_entry); 3016 entry = swp_entry_to_pte(swp_entry); 3017 if (soft_dirty) 3018 entry = pte_swp_mksoft_dirty(entry); 3019 if (uffd_wp) 3020 entry = pte_swp_mkuffd_wp(entry); 3021 3022 VM_WARN_ON(!pte_none(ptep_get(pte + i))); 3023 set_pte_at(mm, addr, pte + i, entry); 3024 } 3025 } else { 3026 pte_t entry; 3027 3028 entry = mk_pte(page, READ_ONCE(vma->vm_page_prot)); 3029 if (write) 3030 entry = pte_mkwrite(entry, vma); 3031 if (!young) 3032 entry = pte_mkold(entry); 3033 /* NOTE: this may set soft-dirty too on some archs */ 3034 if (dirty) 3035 entry = pte_mkdirty(entry); 3036 if (soft_dirty) 3037 entry = pte_mksoft_dirty(entry); 3038 if (uffd_wp) 3039 entry = pte_mkuffd_wp(entry); 3040 3041 for (i = 0; i < HPAGE_PMD_NR; i++) 3042 VM_WARN_ON(!pte_none(ptep_get(pte + i))); 3043 3044 set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR); 3045 } 3046 pte_unmap(pte); 3047 3048 if (!pmd_migration) 3049 folio_remove_rmap_pmd(folio, page, vma); 3050 if (freeze) 3051 put_page(page); 3052 3053 smp_wmb(); /* make pte visible before pmd */ 3054 pmd_populate(mm, pmd, pgtable); 3055 } 3056 3057 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address, 3058 pmd_t *pmd, bool freeze) 3059 { 3060 VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE)); 3061 if (pmd_trans_huge(*pmd) || is_pmd_migration_entry(*pmd)) 3062 __split_huge_pmd_locked(vma, pmd, address, freeze); 3063 } 3064 3065 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 3066 unsigned long address, bool freeze) 3067 { 3068 spinlock_t *ptl; 3069 struct mmu_notifier_range range; 3070 3071 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 3072 address & HPAGE_PMD_MASK, 3073 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 3074 mmu_notifier_invalidate_range_start(&range); 3075 ptl = pmd_lock(vma->vm_mm, pmd); 3076 split_huge_pmd_locked(vma, range.start, pmd, freeze); 3077 spin_unlock(ptl); 3078 mmu_notifier_invalidate_range_end(&range); 3079 } 3080 3081 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 3082 bool freeze) 3083 { 3084 pmd_t *pmd = mm_find_pmd(vma->vm_mm, address); 3085 3086 if (!pmd) 3087 return; 3088 3089 __split_huge_pmd(vma, pmd, address, freeze); 3090 } 3091 3092 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 3093 { 3094 /* 3095 * If the new address isn't hpage aligned and it could previously 3096 * contain an hugepage: check if we need to split an huge pmd. 3097 */ 3098 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 3099 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 3100 ALIGN(address, HPAGE_PMD_SIZE))) 3101 split_huge_pmd_address(vma, address, false); 3102 } 3103 3104 void vma_adjust_trans_huge(struct vm_area_struct *vma, 3105 unsigned long start, 3106 unsigned long end, 3107 struct vm_area_struct *next) 3108 { 3109 /* Check if we need to split start first. */ 3110 split_huge_pmd_if_needed(vma, start); 3111 3112 /* Check if we need to split end next. */ 3113 split_huge_pmd_if_needed(vma, end); 3114 3115 /* If we're incrementing next->vm_start, we might need to split it. */ 3116 if (next) 3117 split_huge_pmd_if_needed(next, end); 3118 } 3119 3120 static void unmap_folio(struct folio *folio) 3121 { 3122 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC | 3123 TTU_BATCH_FLUSH; 3124 3125 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 3126 3127 if (folio_test_pmd_mappable(folio)) 3128 ttu_flags |= TTU_SPLIT_HUGE_PMD; 3129 3130 /* 3131 * Anon pages need migration entries to preserve them, but file 3132 * pages can simply be left unmapped, then faulted back on demand. 3133 * If that is ever changed (perhaps for mlock), update remap_page(). 3134 */ 3135 if (folio_test_anon(folio)) 3136 try_to_migrate(folio, ttu_flags); 3137 else 3138 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK); 3139 3140 try_to_unmap_flush(); 3141 } 3142 3143 static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma, 3144 unsigned long addr, pmd_t *pmdp, 3145 struct folio *folio) 3146 { 3147 struct mm_struct *mm = vma->vm_mm; 3148 int ref_count, map_count; 3149 pmd_t orig_pmd = *pmdp; 3150 3151 if (pmd_dirty(orig_pmd)) 3152 folio_set_dirty(folio); 3153 if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) { 3154 folio_set_swapbacked(folio); 3155 return false; 3156 } 3157 3158 orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp); 3159 3160 /* 3161 * Syncing against concurrent GUP-fast: 3162 * - clear PMD; barrier; read refcount 3163 * - inc refcount; barrier; read PMD 3164 */ 3165 smp_mb(); 3166 3167 ref_count = folio_ref_count(folio); 3168 map_count = folio_mapcount(folio); 3169 3170 /* 3171 * Order reads for folio refcount and dirty flag 3172 * (see comments in __remove_mapping()). 3173 */ 3174 smp_rmb(); 3175 3176 /* 3177 * If the folio or its PMD is redirtied at this point, or if there 3178 * are unexpected references, we will give up to discard this folio 3179 * and remap it. 3180 * 3181 * The only folio refs must be one from isolation plus the rmap(s). 3182 */ 3183 if (pmd_dirty(orig_pmd)) 3184 folio_set_dirty(folio); 3185 if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) { 3186 folio_set_swapbacked(folio); 3187 set_pmd_at(mm, addr, pmdp, orig_pmd); 3188 return false; 3189 } 3190 3191 if (ref_count != map_count + 1) { 3192 set_pmd_at(mm, addr, pmdp, orig_pmd); 3193 return false; 3194 } 3195 3196 folio_remove_rmap_pmd(folio, pmd_page(orig_pmd), vma); 3197 zap_deposited_table(mm, pmdp); 3198 add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR); 3199 if (vma->vm_flags & VM_LOCKED) 3200 mlock_drain_local(); 3201 folio_put(folio); 3202 3203 return true; 3204 } 3205 3206 bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr, 3207 pmd_t *pmdp, struct folio *folio) 3208 { 3209 VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio); 3210 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio); 3211 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); 3212 VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio); 3213 VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE)); 3214 3215 return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio); 3216 } 3217 3218 static void remap_page(struct folio *folio, unsigned long nr, int flags) 3219 { 3220 int i = 0; 3221 3222 /* If unmap_folio() uses try_to_migrate() on file, remove this check */ 3223 if (!folio_test_anon(folio)) 3224 return; 3225 for (;;) { 3226 remove_migration_ptes(folio, folio, RMP_LOCKED | flags); 3227 i += folio_nr_pages(folio); 3228 if (i >= nr) 3229 break; 3230 folio = folio_next(folio); 3231 } 3232 } 3233 3234 static void lru_add_split_folio(struct folio *folio, struct folio *new_folio, 3235 struct lruvec *lruvec, struct list_head *list) 3236 { 3237 VM_BUG_ON_FOLIO(folio_test_lru(new_folio), folio); 3238 lockdep_assert_held(&lruvec->lru_lock); 3239 3240 if (list) { 3241 /* page reclaim is reclaiming a huge page */ 3242 VM_WARN_ON(folio_test_lru(folio)); 3243 folio_get(new_folio); 3244 list_add_tail(&new_folio->lru, list); 3245 } else { 3246 /* head is still on lru (and we have it frozen) */ 3247 VM_WARN_ON(!folio_test_lru(folio)); 3248 if (folio_test_unevictable(folio)) 3249 new_folio->mlock_count = 0; 3250 else 3251 list_add_tail(&new_folio->lru, &folio->lru); 3252 folio_set_lru(new_folio); 3253 } 3254 } 3255 3256 /* Racy check whether the huge page can be split */ 3257 bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins) 3258 { 3259 int extra_pins; 3260 3261 /* Additional pins from page cache */ 3262 if (folio_test_anon(folio)) 3263 extra_pins = folio_test_swapcache(folio) ? 3264 folio_nr_pages(folio) : 0; 3265 else 3266 extra_pins = folio_nr_pages(folio); 3267 if (pextra_pins) 3268 *pextra_pins = extra_pins; 3269 return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 3270 caller_pins; 3271 } 3272 3273 /* 3274 * It splits @folio into @new_order folios and copies the @folio metadata to 3275 * all the resulting folios. 3276 */ 3277 static void __split_folio_to_order(struct folio *folio, int old_order, 3278 int new_order) 3279 { 3280 long new_nr_pages = 1 << new_order; 3281 long nr_pages = 1 << old_order; 3282 long i; 3283 3284 /* 3285 * Skip the first new_nr_pages, since the new folio from them have all 3286 * the flags from the original folio. 3287 */ 3288 for (i = new_nr_pages; i < nr_pages; i += new_nr_pages) { 3289 struct page *new_head = &folio->page + i; 3290 3291 /* 3292 * Careful: new_folio is not a "real" folio before we cleared PageTail. 3293 * Don't pass it around before clear_compound_head(). 3294 */ 3295 struct folio *new_folio = (struct folio *)new_head; 3296 3297 VM_BUG_ON_PAGE(atomic_read(&new_folio->_mapcount) != -1, new_head); 3298 3299 /* 3300 * Clone page flags before unfreezing refcount. 3301 * 3302 * After successful get_page_unless_zero() might follow flags change, 3303 * for example lock_page() which set PG_waiters. 3304 * 3305 * Note that for mapped sub-pages of an anonymous THP, 3306 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in 3307 * the migration entry instead from where remap_page() will restore it. 3308 * We can still have PG_anon_exclusive set on effectively unmapped and 3309 * unreferenced sub-pages of an anonymous THP: we can simply drop 3310 * PG_anon_exclusive (-> PG_mappedtodisk) for these here. 3311 */ 3312 new_folio->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 3313 new_folio->flags |= (folio->flags & 3314 ((1L << PG_referenced) | 3315 (1L << PG_swapbacked) | 3316 (1L << PG_swapcache) | 3317 (1L << PG_mlocked) | 3318 (1L << PG_uptodate) | 3319 (1L << PG_active) | 3320 (1L << PG_workingset) | 3321 (1L << PG_locked) | 3322 (1L << PG_unevictable) | 3323 #ifdef CONFIG_ARCH_USES_PG_ARCH_2 3324 (1L << PG_arch_2) | 3325 #endif 3326 #ifdef CONFIG_ARCH_USES_PG_ARCH_3 3327 (1L << PG_arch_3) | 3328 #endif 3329 (1L << PG_dirty) | 3330 LRU_GEN_MASK | LRU_REFS_MASK)); 3331 3332 new_folio->mapping = folio->mapping; 3333 new_folio->index = folio->index + i; 3334 3335 /* 3336 * page->private should not be set in tail pages. Fix up and warn once 3337 * if private is unexpectedly set. 3338 */ 3339 if (unlikely(new_folio->private)) { 3340 VM_WARN_ON_ONCE_PAGE(true, new_head); 3341 new_folio->private = NULL; 3342 } 3343 3344 if (folio_test_swapcache(folio)) 3345 new_folio->swap.val = folio->swap.val + i; 3346 3347 /* Page flags must be visible before we make the page non-compound. */ 3348 smp_wmb(); 3349 3350 /* 3351 * Clear PageTail before unfreezing page refcount. 3352 * 3353 * After successful get_page_unless_zero() might follow put_page() 3354 * which needs correct compound_head(). 3355 */ 3356 clear_compound_head(new_head); 3357 if (new_order) { 3358 prep_compound_page(new_head, new_order); 3359 folio_set_large_rmappable(new_folio); 3360 } 3361 3362 if (folio_test_young(folio)) 3363 folio_set_young(new_folio); 3364 if (folio_test_idle(folio)) 3365 folio_set_idle(new_folio); 3366 #ifdef CONFIG_MEMCG 3367 new_folio->memcg_data = folio->memcg_data; 3368 #endif 3369 3370 folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio)); 3371 } 3372 3373 if (new_order) 3374 folio_set_order(folio, new_order); 3375 else 3376 ClearPageCompound(&folio->page); 3377 } 3378 3379 /* 3380 * It splits an unmapped @folio to lower order smaller folios in two ways. 3381 * @folio: the to-be-split folio 3382 * @new_order: the smallest order of the after split folios (since buddy 3383 * allocator like split generates folios with orders from @folio's 3384 * order - 1 to new_order). 3385 * @split_at: in buddy allocator like split, the folio containing @split_at 3386 * will be split until its order becomes @new_order. 3387 * @lock_at: the folio containing @lock_at is left locked for caller. 3388 * @list: the after split folios will be added to @list if it is not NULL, 3389 * otherwise to LRU lists. 3390 * @end: the end of the file @folio maps to. -1 if @folio is anonymous memory. 3391 * @xas: xa_state pointing to folio->mapping->i_pages and locked by caller 3392 * @mapping: @folio->mapping 3393 * @uniform_split: if the split is uniform or not (buddy allocator like split) 3394 * 3395 * 3396 * 1. uniform split: the given @folio into multiple @new_order small folios, 3397 * where all small folios have the same order. This is done when 3398 * uniform_split is true. 3399 * 2. buddy allocator like (non-uniform) split: the given @folio is split into 3400 * half and one of the half (containing the given page) is split into half 3401 * until the given @page's order becomes @new_order. This is done when 3402 * uniform_split is false. 3403 * 3404 * The high level flow for these two methods are: 3405 * 1. uniform split: a single __split_folio_to_order() is called to split the 3406 * @folio into @new_order, then we traverse all the resulting folios one by 3407 * one in PFN ascending order and perform stats, unfreeze, adding to list, 3408 * and file mapping index operations. 3409 * 2. non-uniform split: in general, folio_order - @new_order calls to 3410 * __split_folio_to_order() are made in a for loop to split the @folio 3411 * to one lower order at a time. The resulting small folios are processed 3412 * like what is done during the traversal in 1, except the one containing 3413 * @page, which is split in next for loop. 3414 * 3415 * After splitting, the caller's folio reference will be transferred to the 3416 * folio containing @page. The other folios may be freed if they are not mapped. 3417 * 3418 * In terms of locking, after splitting, 3419 * 1. uniform split leaves @page (or the folio contains it) locked; 3420 * 2. buddy allocator like (non-uniform) split leaves @folio locked. 3421 * 3422 * 3423 * For !uniform_split, when -ENOMEM is returned, the original folio might be 3424 * split. The caller needs to check the input folio. 3425 */ 3426 static int __split_unmapped_folio(struct folio *folio, int new_order, 3427 struct page *split_at, struct page *lock_at, 3428 struct list_head *list, pgoff_t end, 3429 struct xa_state *xas, struct address_space *mapping, 3430 bool uniform_split) 3431 { 3432 struct lruvec *lruvec; 3433 struct address_space *swap_cache = NULL; 3434 struct folio *origin_folio = folio; 3435 struct folio *next_folio = folio_next(folio); 3436 struct folio *new_folio; 3437 struct folio *next; 3438 int order = folio_order(folio); 3439 int split_order; 3440 int start_order = uniform_split ? new_order : order - 1; 3441 int nr_dropped = 0; 3442 int ret = 0; 3443 bool stop_split = false; 3444 3445 if (folio_test_swapcache(folio)) { 3446 VM_BUG_ON(mapping); 3447 3448 /* a swapcache folio can only be uniformly split to order-0 */ 3449 if (!uniform_split || new_order != 0) 3450 return -EINVAL; 3451 3452 swap_cache = swap_address_space(folio->swap); 3453 xa_lock(&swap_cache->i_pages); 3454 } 3455 3456 if (folio_test_anon(folio)) 3457 mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1); 3458 3459 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 3460 lruvec = folio_lruvec_lock(folio); 3461 3462 folio_clear_has_hwpoisoned(folio); 3463 3464 /* 3465 * split to new_order one order at a time. For uniform split, 3466 * folio is split to new_order directly. 3467 */ 3468 for (split_order = start_order; 3469 split_order >= new_order && !stop_split; 3470 split_order--) { 3471 int old_order = folio_order(folio); 3472 struct folio *release; 3473 struct folio *end_folio = folio_next(folio); 3474 3475 /* order-1 anonymous folio is not supported */ 3476 if (folio_test_anon(folio) && split_order == 1) 3477 continue; 3478 if (uniform_split && split_order != new_order) 3479 continue; 3480 3481 if (mapping) { 3482 /* 3483 * uniform split has xas_split_alloc() called before 3484 * irq is disabled to allocate enough memory, whereas 3485 * non-uniform split can handle ENOMEM. 3486 */ 3487 if (uniform_split) 3488 xas_split(xas, folio, old_order); 3489 else { 3490 xas_set_order(xas, folio->index, split_order); 3491 xas_try_split(xas, folio, old_order); 3492 if (xas_error(xas)) { 3493 ret = xas_error(xas); 3494 stop_split = true; 3495 goto after_split; 3496 } 3497 } 3498 } 3499 3500 folio_split_memcg_refs(folio, old_order, split_order); 3501 split_page_owner(&folio->page, old_order, split_order); 3502 pgalloc_tag_split(folio, old_order, split_order); 3503 3504 __split_folio_to_order(folio, old_order, split_order); 3505 3506 after_split: 3507 /* 3508 * Iterate through after-split folios and perform related 3509 * operations. But in buddy allocator like split, the folio 3510 * containing the specified page is skipped until its order 3511 * is new_order, since the folio will be worked on in next 3512 * iteration. 3513 */ 3514 for (release = folio; release != end_folio; release = next) { 3515 next = folio_next(release); 3516 /* 3517 * for buddy allocator like split, the folio containing 3518 * page will be split next and should not be released, 3519 * until the folio's order is new_order or stop_split 3520 * is set to true by the above xas_split() failure. 3521 */ 3522 if (release == page_folio(split_at)) { 3523 folio = release; 3524 if (split_order != new_order && !stop_split) 3525 continue; 3526 } 3527 if (folio_test_anon(release)) { 3528 mod_mthp_stat(folio_order(release), 3529 MTHP_STAT_NR_ANON, 1); 3530 } 3531 3532 /* 3533 * origin_folio should be kept frozon until page cache 3534 * entries are updated with all the other after-split 3535 * folios to prevent others seeing stale page cache 3536 * entries. 3537 */ 3538 if (release == origin_folio) 3539 continue; 3540 3541 folio_ref_unfreeze(release, 1 + 3542 ((mapping || swap_cache) ? 3543 folio_nr_pages(release) : 0)); 3544 3545 lru_add_split_folio(origin_folio, release, lruvec, 3546 list); 3547 3548 /* Some pages can be beyond EOF: drop them from cache */ 3549 if (release->index >= end) { 3550 if (shmem_mapping(mapping)) 3551 nr_dropped += folio_nr_pages(release); 3552 else if (folio_test_clear_dirty(release)) 3553 folio_account_cleaned(release, 3554 inode_to_wb(mapping->host)); 3555 __filemap_remove_folio(release, NULL); 3556 folio_put_refs(release, folio_nr_pages(release)); 3557 } else if (mapping) { 3558 __xa_store(&mapping->i_pages, 3559 release->index, release, 0); 3560 } else if (swap_cache) { 3561 __xa_store(&swap_cache->i_pages, 3562 swap_cache_index(release->swap), 3563 release, 0); 3564 } 3565 } 3566 } 3567 3568 /* 3569 * Unfreeze origin_folio only after all page cache entries, which used 3570 * to point to it, have been updated with new folios. Otherwise, 3571 * a parallel folio_try_get() can grab origin_folio and its caller can 3572 * see stale page cache entries. 3573 */ 3574 folio_ref_unfreeze(origin_folio, 1 + 3575 ((mapping || swap_cache) ? folio_nr_pages(origin_folio) : 0)); 3576 3577 unlock_page_lruvec(lruvec); 3578 3579 if (swap_cache) 3580 xa_unlock(&swap_cache->i_pages); 3581 if (mapping) 3582 xa_unlock(&mapping->i_pages); 3583 3584 /* Caller disabled irqs, so they are still disabled here */ 3585 local_irq_enable(); 3586 3587 if (nr_dropped) 3588 shmem_uncharge(mapping->host, nr_dropped); 3589 3590 remap_page(origin_folio, 1 << order, 3591 folio_test_anon(origin_folio) ? 3592 RMP_USE_SHARED_ZEROPAGE : 0); 3593 3594 /* 3595 * At this point, folio should contain the specified page. 3596 * For uniform split, it is left for caller to unlock. 3597 * For buddy allocator like split, the first after-split folio is left 3598 * for caller to unlock. 3599 */ 3600 for (new_folio = origin_folio; new_folio != next_folio; new_folio = next) { 3601 next = folio_next(new_folio); 3602 if (new_folio == page_folio(lock_at)) 3603 continue; 3604 3605 folio_unlock(new_folio); 3606 /* 3607 * Subpages may be freed if there wasn't any mapping 3608 * like if add_to_swap() is running on a lru page that 3609 * had its mapping zapped. And freeing these pages 3610 * requires taking the lru_lock so we do the put_page 3611 * of the tail pages after the split is complete. 3612 */ 3613 free_folio_and_swap_cache(new_folio); 3614 } 3615 return ret; 3616 } 3617 3618 bool non_uniform_split_supported(struct folio *folio, unsigned int new_order, 3619 bool warns) 3620 { 3621 if (folio_test_anon(folio)) { 3622 /* order-1 is not supported for anonymous THP. */ 3623 VM_WARN_ONCE(warns && new_order == 1, 3624 "Cannot split to order-1 folio"); 3625 return new_order != 1; 3626 } else if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && 3627 !mapping_large_folio_support(folio->mapping)) { 3628 /* 3629 * No split if the file system does not support large folio. 3630 * Note that we might still have THPs in such mappings due to 3631 * CONFIG_READ_ONLY_THP_FOR_FS. But in that case, the mapping 3632 * does not actually support large folios properly. 3633 */ 3634 VM_WARN_ONCE(warns, 3635 "Cannot split file folio to non-0 order"); 3636 return false; 3637 } 3638 3639 /* Only swapping a whole PMD-mapped folio is supported */ 3640 if (folio_test_swapcache(folio)) { 3641 VM_WARN_ONCE(warns, 3642 "Cannot split swapcache folio to non-0 order"); 3643 return false; 3644 } 3645 3646 return true; 3647 } 3648 3649 /* See comments in non_uniform_split_supported() */ 3650 bool uniform_split_supported(struct folio *folio, unsigned int new_order, 3651 bool warns) 3652 { 3653 if (folio_test_anon(folio)) { 3654 VM_WARN_ONCE(warns && new_order == 1, 3655 "Cannot split to order-1 folio"); 3656 return new_order != 1; 3657 } else if (new_order) { 3658 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && 3659 !mapping_large_folio_support(folio->mapping)) { 3660 VM_WARN_ONCE(warns, 3661 "Cannot split file folio to non-0 order"); 3662 return false; 3663 } 3664 } 3665 3666 if (new_order && folio_test_swapcache(folio)) { 3667 VM_WARN_ONCE(warns, 3668 "Cannot split swapcache folio to non-0 order"); 3669 return false; 3670 } 3671 3672 return true; 3673 } 3674 3675 /* 3676 * __folio_split: split a folio at @split_at to a @new_order folio 3677 * @folio: folio to split 3678 * @new_order: the order of the new folio 3679 * @split_at: a page within the new folio 3680 * @lock_at: a page within @folio to be left locked to caller 3681 * @list: after-split folios will be put on it if non NULL 3682 * @uniform_split: perform uniform split or not (non-uniform split) 3683 * 3684 * It calls __split_unmapped_folio() to perform uniform and non-uniform split. 3685 * It is in charge of checking whether the split is supported or not and 3686 * preparing @folio for __split_unmapped_folio(). 3687 * 3688 * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be 3689 * split but not to @new_order, the caller needs to check) 3690 */ 3691 static int __folio_split(struct folio *folio, unsigned int new_order, 3692 struct page *split_at, struct page *lock_at, 3693 struct list_head *list, bool uniform_split) 3694 { 3695 struct deferred_split *ds_queue = get_deferred_split_queue(folio); 3696 XA_STATE(xas, &folio->mapping->i_pages, folio->index); 3697 bool is_anon = folio_test_anon(folio); 3698 struct address_space *mapping = NULL; 3699 struct anon_vma *anon_vma = NULL; 3700 int order = folio_order(folio); 3701 int extra_pins, ret; 3702 pgoff_t end; 3703 bool is_hzp; 3704 3705 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 3706 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 3707 3708 if (folio != page_folio(split_at) || folio != page_folio(lock_at)) 3709 return -EINVAL; 3710 3711 if (new_order >= folio_order(folio)) 3712 return -EINVAL; 3713 3714 if (uniform_split && !uniform_split_supported(folio, new_order, true)) 3715 return -EINVAL; 3716 3717 if (!uniform_split && 3718 !non_uniform_split_supported(folio, new_order, true)) 3719 return -EINVAL; 3720 3721 is_hzp = is_huge_zero_folio(folio); 3722 if (is_hzp) { 3723 pr_warn_ratelimited("Called split_huge_page for huge zero page\n"); 3724 return -EBUSY; 3725 } 3726 3727 if (folio_test_writeback(folio)) 3728 return -EBUSY; 3729 3730 if (is_anon) { 3731 /* 3732 * The caller does not necessarily hold an mmap_lock that would 3733 * prevent the anon_vma disappearing so we first we take a 3734 * reference to it and then lock the anon_vma for write. This 3735 * is similar to folio_lock_anon_vma_read except the write lock 3736 * is taken to serialise against parallel split or collapse 3737 * operations. 3738 */ 3739 anon_vma = folio_get_anon_vma(folio); 3740 if (!anon_vma) { 3741 ret = -EBUSY; 3742 goto out; 3743 } 3744 end = -1; 3745 mapping = NULL; 3746 anon_vma_lock_write(anon_vma); 3747 } else { 3748 unsigned int min_order; 3749 gfp_t gfp; 3750 3751 mapping = folio->mapping; 3752 3753 /* Truncated ? */ 3754 /* 3755 * TODO: add support for large shmem folio in swap cache. 3756 * When shmem is in swap cache, mapping is NULL and 3757 * folio_test_swapcache() is true. 3758 */ 3759 if (!mapping) { 3760 ret = -EBUSY; 3761 goto out; 3762 } 3763 3764 min_order = mapping_min_folio_order(folio->mapping); 3765 if (new_order < min_order) { 3766 VM_WARN_ONCE(1, "Cannot split mapped folio below min-order: %u", 3767 min_order); 3768 ret = -EINVAL; 3769 goto out; 3770 } 3771 3772 gfp = current_gfp_context(mapping_gfp_mask(mapping) & 3773 GFP_RECLAIM_MASK); 3774 3775 if (!filemap_release_folio(folio, gfp)) { 3776 ret = -EBUSY; 3777 goto out; 3778 } 3779 3780 if (uniform_split) { 3781 xas_set_order(&xas, folio->index, new_order); 3782 xas_split_alloc(&xas, folio, folio_order(folio), gfp); 3783 if (xas_error(&xas)) { 3784 ret = xas_error(&xas); 3785 goto out; 3786 } 3787 } 3788 3789 anon_vma = NULL; 3790 i_mmap_lock_read(mapping); 3791 3792 /* 3793 *__split_unmapped_folio() may need to trim off pages beyond 3794 * EOF: but on 32-bit, i_size_read() takes an irq-unsafe 3795 * seqlock, which cannot be nested inside the page tree lock. 3796 * So note end now: i_size itself may be changed at any moment, 3797 * but folio lock is good enough to serialize the trimming. 3798 */ 3799 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 3800 if (shmem_mapping(mapping)) 3801 end = shmem_fallocend(mapping->host, end); 3802 } 3803 3804 /* 3805 * Racy check if we can split the page, before unmap_folio() will 3806 * split PMDs 3807 */ 3808 if (!can_split_folio(folio, 1, &extra_pins)) { 3809 ret = -EAGAIN; 3810 goto out_unlock; 3811 } 3812 3813 unmap_folio(folio); 3814 3815 /* block interrupt reentry in xa_lock and spinlock */ 3816 local_irq_disable(); 3817 if (mapping) { 3818 /* 3819 * Check if the folio is present in page cache. 3820 * We assume all tail are present too, if folio is there. 3821 */ 3822 xas_lock(&xas); 3823 xas_reset(&xas); 3824 if (xas_load(&xas) != folio) 3825 goto fail; 3826 } 3827 3828 /* Prevent deferred_split_scan() touching ->_refcount */ 3829 spin_lock(&ds_queue->split_queue_lock); 3830 if (folio_ref_freeze(folio, 1 + extra_pins)) { 3831 if (folio_order(folio) > 1 && 3832 !list_empty(&folio->_deferred_list)) { 3833 ds_queue->split_queue_len--; 3834 if (folio_test_partially_mapped(folio)) { 3835 folio_clear_partially_mapped(folio); 3836 mod_mthp_stat(folio_order(folio), 3837 MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1); 3838 } 3839 /* 3840 * Reinitialize page_deferred_list after removing the 3841 * page from the split_queue, otherwise a subsequent 3842 * split will see list corruption when checking the 3843 * page_deferred_list. 3844 */ 3845 list_del_init(&folio->_deferred_list); 3846 } 3847 spin_unlock(&ds_queue->split_queue_lock); 3848 if (mapping) { 3849 int nr = folio_nr_pages(folio); 3850 3851 if (folio_test_pmd_mappable(folio) && 3852 new_order < HPAGE_PMD_ORDER) { 3853 if (folio_test_swapbacked(folio)) { 3854 __lruvec_stat_mod_folio(folio, 3855 NR_SHMEM_THPS, -nr); 3856 } else { 3857 __lruvec_stat_mod_folio(folio, 3858 NR_FILE_THPS, -nr); 3859 filemap_nr_thps_dec(mapping); 3860 } 3861 } 3862 } 3863 3864 ret = __split_unmapped_folio(folio, new_order, 3865 split_at, lock_at, list, end, &xas, mapping, 3866 uniform_split); 3867 } else { 3868 spin_unlock(&ds_queue->split_queue_lock); 3869 fail: 3870 if (mapping) 3871 xas_unlock(&xas); 3872 local_irq_enable(); 3873 remap_page(folio, folio_nr_pages(folio), 0); 3874 ret = -EAGAIN; 3875 } 3876 3877 out_unlock: 3878 if (anon_vma) { 3879 anon_vma_unlock_write(anon_vma); 3880 put_anon_vma(anon_vma); 3881 } 3882 if (mapping) 3883 i_mmap_unlock_read(mapping); 3884 out: 3885 xas_destroy(&xas); 3886 if (order == HPAGE_PMD_ORDER) 3887 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 3888 count_mthp_stat(order, !ret ? MTHP_STAT_SPLIT : MTHP_STAT_SPLIT_FAILED); 3889 return ret; 3890 } 3891 3892 /* 3893 * This function splits a large folio into smaller folios of order @new_order. 3894 * @page can point to any page of the large folio to split. The split operation 3895 * does not change the position of @page. 3896 * 3897 * Prerequisites: 3898 * 3899 * 1) The caller must hold a reference on the @page's owning folio, also known 3900 * as the large folio. 3901 * 3902 * 2) The large folio must be locked. 3903 * 3904 * 3) The folio must not be pinned. Any unexpected folio references, including 3905 * GUP pins, will result in the folio not getting split; instead, the caller 3906 * will receive an -EAGAIN. 3907 * 3908 * 4) @new_order > 1, usually. Splitting to order-1 anonymous folios is not 3909 * supported for non-file-backed folios, because folio->_deferred_list, which 3910 * is used by partially mapped folios, is stored in subpage 2, but an order-1 3911 * folio only has subpages 0 and 1. File-backed order-1 folios are supported, 3912 * since they do not use _deferred_list. 3913 * 3914 * After splitting, the caller's folio reference will be transferred to @page, 3915 * resulting in a raised refcount of @page after this call. The other pages may 3916 * be freed if they are not mapped. 3917 * 3918 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 3919 * 3920 * Pages in @new_order will inherit the mapping, flags, and so on from the 3921 * huge page. 3922 * 3923 * Returns 0 if the huge page was split successfully. 3924 * 3925 * Returns -EAGAIN if the folio has unexpected reference (e.g., GUP) or if 3926 * the folio was concurrently removed from the page cache. 3927 * 3928 * Returns -EBUSY when trying to split the huge zeropage, if the folio is 3929 * under writeback, if fs-specific folio metadata cannot currently be 3930 * released, or if some unexpected race happened (e.g., anon VMA disappeared, 3931 * truncation). 3932 * 3933 * Callers should ensure that the order respects the address space mapping 3934 * min-order if one is set for non-anonymous folios. 3935 * 3936 * Returns -EINVAL when trying to split to an order that is incompatible 3937 * with the folio. Splitting to order 0 is compatible with all folios. 3938 */ 3939 int split_huge_page_to_list_to_order(struct page *page, struct list_head *list, 3940 unsigned int new_order) 3941 { 3942 struct folio *folio = page_folio(page); 3943 3944 return __folio_split(folio, new_order, &folio->page, page, list, true); 3945 } 3946 3947 /* 3948 * folio_split: split a folio at @split_at to a @new_order folio 3949 * @folio: folio to split 3950 * @new_order: the order of the new folio 3951 * @split_at: a page within the new folio 3952 * 3953 * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be 3954 * split but not to @new_order, the caller needs to check) 3955 * 3956 * It has the same prerequisites and returns as 3957 * split_huge_page_to_list_to_order(). 3958 * 3959 * Split a folio at @split_at to a new_order folio, leave the 3960 * remaining subpages of the original folio as large as possible. For example, 3961 * in the case of splitting an order-9 folio at its third order-3 subpages to 3962 * an order-3 folio, there are 2^(9-3)=64 order-3 subpages in the order-9 folio. 3963 * After the split, there will be a group of folios with different orders and 3964 * the new folio containing @split_at is marked in bracket: 3965 * [order-4, {order-3}, order-3, order-5, order-6, order-7, order-8]. 3966 * 3967 * After split, folio is left locked for caller. 3968 */ 3969 int folio_split(struct folio *folio, unsigned int new_order, 3970 struct page *split_at, struct list_head *list) 3971 { 3972 return __folio_split(folio, new_order, split_at, &folio->page, list, 3973 false); 3974 } 3975 3976 int min_order_for_split(struct folio *folio) 3977 { 3978 if (folio_test_anon(folio)) 3979 return 0; 3980 3981 if (!folio->mapping) { 3982 if (folio_test_pmd_mappable(folio)) 3983 count_vm_event(THP_SPLIT_PAGE_FAILED); 3984 return -EBUSY; 3985 } 3986 3987 return mapping_min_folio_order(folio->mapping); 3988 } 3989 3990 int split_folio_to_list(struct folio *folio, struct list_head *list) 3991 { 3992 int ret = min_order_for_split(folio); 3993 3994 if (ret < 0) 3995 return ret; 3996 3997 return split_huge_page_to_list_to_order(&folio->page, list, ret); 3998 } 3999 4000 /* 4001 * __folio_unqueue_deferred_split() is not to be called directly: 4002 * the folio_unqueue_deferred_split() inline wrapper in mm/internal.h 4003 * limits its calls to those folios which may have a _deferred_list for 4004 * queueing THP splits, and that list is (racily observed to be) non-empty. 4005 * 4006 * It is unsafe to call folio_unqueue_deferred_split() until folio refcount is 4007 * zero: because even when split_queue_lock is held, a non-empty _deferred_list 4008 * might be in use on deferred_split_scan()'s unlocked on-stack list. 4009 * 4010 * If memory cgroups are enabled, split_queue_lock is in the mem_cgroup: it is 4011 * therefore important to unqueue deferred split before changing folio memcg. 4012 */ 4013 bool __folio_unqueue_deferred_split(struct folio *folio) 4014 { 4015 struct deferred_split *ds_queue; 4016 unsigned long flags; 4017 bool unqueued = false; 4018 4019 WARN_ON_ONCE(folio_ref_count(folio)); 4020 WARN_ON_ONCE(!mem_cgroup_disabled() && !folio_memcg(folio)); 4021 4022 ds_queue = get_deferred_split_queue(folio); 4023 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 4024 if (!list_empty(&folio->_deferred_list)) { 4025 ds_queue->split_queue_len--; 4026 if (folio_test_partially_mapped(folio)) { 4027 folio_clear_partially_mapped(folio); 4028 mod_mthp_stat(folio_order(folio), 4029 MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1); 4030 } 4031 list_del_init(&folio->_deferred_list); 4032 unqueued = true; 4033 } 4034 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 4035 4036 return unqueued; /* useful for debug warnings */ 4037 } 4038 4039 /* partially_mapped=false won't clear PG_partially_mapped folio flag */ 4040 void deferred_split_folio(struct folio *folio, bool partially_mapped) 4041 { 4042 struct deferred_split *ds_queue = get_deferred_split_queue(folio); 4043 #ifdef CONFIG_MEMCG 4044 struct mem_cgroup *memcg = folio_memcg(folio); 4045 #endif 4046 unsigned long flags; 4047 4048 /* 4049 * Order 1 folios have no space for a deferred list, but we also 4050 * won't waste much memory by not adding them to the deferred list. 4051 */ 4052 if (folio_order(folio) <= 1) 4053 return; 4054 4055 if (!partially_mapped && !split_underused_thp) 4056 return; 4057 4058 /* 4059 * Exclude swapcache: originally to avoid a corrupt deferred split 4060 * queue. Nowadays that is fully prevented by memcg1_swapout(); 4061 * but if page reclaim is already handling the same folio, it is 4062 * unnecessary to handle it again in the shrinker, so excluding 4063 * swapcache here may still be a useful optimization. 4064 */ 4065 if (folio_test_swapcache(folio)) 4066 return; 4067 4068 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 4069 if (partially_mapped) { 4070 if (!folio_test_partially_mapped(folio)) { 4071 folio_set_partially_mapped(folio); 4072 if (folio_test_pmd_mappable(folio)) 4073 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 4074 count_mthp_stat(folio_order(folio), MTHP_STAT_SPLIT_DEFERRED); 4075 mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, 1); 4076 4077 } 4078 } else { 4079 /* partially mapped folios cannot become non-partially mapped */ 4080 VM_WARN_ON_FOLIO(folio_test_partially_mapped(folio), folio); 4081 } 4082 if (list_empty(&folio->_deferred_list)) { 4083 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue); 4084 ds_queue->split_queue_len++; 4085 #ifdef CONFIG_MEMCG 4086 if (memcg) 4087 set_shrinker_bit(memcg, folio_nid(folio), 4088 deferred_split_shrinker->id); 4089 #endif 4090 } 4091 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 4092 } 4093 4094 static unsigned long deferred_split_count(struct shrinker *shrink, 4095 struct shrink_control *sc) 4096 { 4097 struct pglist_data *pgdata = NODE_DATA(sc->nid); 4098 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 4099 4100 #ifdef CONFIG_MEMCG 4101 if (sc->memcg) 4102 ds_queue = &sc->memcg->deferred_split_queue; 4103 #endif 4104 return READ_ONCE(ds_queue->split_queue_len); 4105 } 4106 4107 static bool thp_underused(struct folio *folio) 4108 { 4109 int num_zero_pages = 0, num_filled_pages = 0; 4110 void *kaddr; 4111 int i; 4112 4113 if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1) 4114 return false; 4115 4116 for (i = 0; i < folio_nr_pages(folio); i++) { 4117 kaddr = kmap_local_folio(folio, i * PAGE_SIZE); 4118 if (!memchr_inv(kaddr, 0, PAGE_SIZE)) { 4119 num_zero_pages++; 4120 if (num_zero_pages > khugepaged_max_ptes_none) { 4121 kunmap_local(kaddr); 4122 return true; 4123 } 4124 } else { 4125 /* 4126 * Another path for early exit once the number 4127 * of non-zero filled pages exceeds threshold. 4128 */ 4129 num_filled_pages++; 4130 if (num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none) { 4131 kunmap_local(kaddr); 4132 return false; 4133 } 4134 } 4135 kunmap_local(kaddr); 4136 } 4137 return false; 4138 } 4139 4140 static unsigned long deferred_split_scan(struct shrinker *shrink, 4141 struct shrink_control *sc) 4142 { 4143 struct pglist_data *pgdata = NODE_DATA(sc->nid); 4144 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 4145 unsigned long flags; 4146 LIST_HEAD(list); 4147 struct folio *folio, *next, *prev = NULL; 4148 int split = 0, removed = 0; 4149 4150 #ifdef CONFIG_MEMCG 4151 if (sc->memcg) 4152 ds_queue = &sc->memcg->deferred_split_queue; 4153 #endif 4154 4155 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 4156 /* Take pin on all head pages to avoid freeing them under us */ 4157 list_for_each_entry_safe(folio, next, &ds_queue->split_queue, 4158 _deferred_list) { 4159 if (folio_try_get(folio)) { 4160 list_move(&folio->_deferred_list, &list); 4161 } else { 4162 /* We lost race with folio_put() */ 4163 if (folio_test_partially_mapped(folio)) { 4164 folio_clear_partially_mapped(folio); 4165 mod_mthp_stat(folio_order(folio), 4166 MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1); 4167 } 4168 list_del_init(&folio->_deferred_list); 4169 ds_queue->split_queue_len--; 4170 } 4171 if (!--sc->nr_to_scan) 4172 break; 4173 } 4174 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 4175 4176 list_for_each_entry_safe(folio, next, &list, _deferred_list) { 4177 bool did_split = false; 4178 bool underused = false; 4179 4180 if (!folio_test_partially_mapped(folio)) { 4181 underused = thp_underused(folio); 4182 if (!underused) 4183 goto next; 4184 } 4185 if (!folio_trylock(folio)) 4186 goto next; 4187 if (!split_folio(folio)) { 4188 did_split = true; 4189 if (underused) 4190 count_vm_event(THP_UNDERUSED_SPLIT_PAGE); 4191 split++; 4192 } 4193 folio_unlock(folio); 4194 next: 4195 /* 4196 * split_folio() removes folio from list on success. 4197 * Only add back to the queue if folio is partially mapped. 4198 * If thp_underused returns false, or if split_folio fails 4199 * in the case it was underused, then consider it used and 4200 * don't add it back to split_queue. 4201 */ 4202 if (did_split) { 4203 ; /* folio already removed from list */ 4204 } else if (!folio_test_partially_mapped(folio)) { 4205 list_del_init(&folio->_deferred_list); 4206 removed++; 4207 } else { 4208 /* 4209 * That unlocked list_del_init() above would be unsafe, 4210 * unless its folio is separated from any earlier folios 4211 * left on the list (which may be concurrently unqueued) 4212 * by one safe folio with refcount still raised. 4213 */ 4214 swap(folio, prev); 4215 } 4216 if (folio) 4217 folio_put(folio); 4218 } 4219 4220 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 4221 list_splice_tail(&list, &ds_queue->split_queue); 4222 ds_queue->split_queue_len -= removed; 4223 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 4224 4225 if (prev) 4226 folio_put(prev); 4227 4228 /* 4229 * Stop shrinker if we didn't split any page, but the queue is empty. 4230 * This can happen if pages were freed under us. 4231 */ 4232 if (!split && list_empty(&ds_queue->split_queue)) 4233 return SHRINK_STOP; 4234 return split; 4235 } 4236 4237 #ifdef CONFIG_DEBUG_FS 4238 static void split_huge_pages_all(void) 4239 { 4240 struct zone *zone; 4241 struct page *page; 4242 struct folio *folio; 4243 unsigned long pfn, max_zone_pfn; 4244 unsigned long total = 0, split = 0; 4245 4246 pr_debug("Split all THPs\n"); 4247 for_each_zone(zone) { 4248 if (!managed_zone(zone)) 4249 continue; 4250 max_zone_pfn = zone_end_pfn(zone); 4251 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 4252 int nr_pages; 4253 4254 page = pfn_to_online_page(pfn); 4255 if (!page || PageTail(page)) 4256 continue; 4257 folio = page_folio(page); 4258 if (!folio_try_get(folio)) 4259 continue; 4260 4261 if (unlikely(page_folio(page) != folio)) 4262 goto next; 4263 4264 if (zone != folio_zone(folio)) 4265 goto next; 4266 4267 if (!folio_test_large(folio) 4268 || folio_test_hugetlb(folio) 4269 || !folio_test_lru(folio)) 4270 goto next; 4271 4272 total++; 4273 folio_lock(folio); 4274 nr_pages = folio_nr_pages(folio); 4275 if (!split_folio(folio)) 4276 split++; 4277 pfn += nr_pages - 1; 4278 folio_unlock(folio); 4279 next: 4280 folio_put(folio); 4281 cond_resched(); 4282 } 4283 } 4284 4285 pr_debug("%lu of %lu THP split\n", split, total); 4286 } 4287 4288 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 4289 { 4290 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 4291 is_vm_hugetlb_page(vma); 4292 } 4293 4294 static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 4295 unsigned long vaddr_end, unsigned int new_order, 4296 long in_folio_offset) 4297 { 4298 int ret = 0; 4299 struct task_struct *task; 4300 struct mm_struct *mm; 4301 unsigned long total = 0, split = 0; 4302 unsigned long addr; 4303 4304 vaddr_start &= PAGE_MASK; 4305 vaddr_end &= PAGE_MASK; 4306 4307 task = find_get_task_by_vpid(pid); 4308 if (!task) { 4309 ret = -ESRCH; 4310 goto out; 4311 } 4312 4313 /* Find the mm_struct */ 4314 mm = get_task_mm(task); 4315 put_task_struct(task); 4316 4317 if (!mm) { 4318 ret = -EINVAL; 4319 goto out; 4320 } 4321 4322 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 4323 pid, vaddr_start, vaddr_end); 4324 4325 mmap_read_lock(mm); 4326 /* 4327 * always increase addr by PAGE_SIZE, since we could have a PTE page 4328 * table filled with PTE-mapped THPs, each of which is distinct. 4329 */ 4330 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 4331 struct vm_area_struct *vma = vma_lookup(mm, addr); 4332 struct folio_walk fw; 4333 struct folio *folio; 4334 struct address_space *mapping; 4335 unsigned int target_order = new_order; 4336 4337 if (!vma) 4338 break; 4339 4340 /* skip special VMA and hugetlb VMA */ 4341 if (vma_not_suitable_for_thp_split(vma)) { 4342 addr = vma->vm_end; 4343 continue; 4344 } 4345 4346 folio = folio_walk_start(&fw, vma, addr, 0); 4347 if (!folio) 4348 continue; 4349 4350 if (!is_transparent_hugepage(folio)) 4351 goto next; 4352 4353 if (!folio_test_anon(folio)) { 4354 mapping = folio->mapping; 4355 target_order = max(new_order, 4356 mapping_min_folio_order(mapping)); 4357 } 4358 4359 if (target_order >= folio_order(folio)) 4360 goto next; 4361 4362 total++; 4363 /* 4364 * For folios with private, split_huge_page_to_list_to_order() 4365 * will try to drop it before split and then check if the folio 4366 * can be split or not. So skip the check here. 4367 */ 4368 if (!folio_test_private(folio) && 4369 !can_split_folio(folio, 0, NULL)) 4370 goto next; 4371 4372 if (!folio_trylock(folio)) 4373 goto next; 4374 folio_get(folio); 4375 folio_walk_end(&fw, vma); 4376 4377 if (!folio_test_anon(folio) && folio->mapping != mapping) 4378 goto unlock; 4379 4380 if (in_folio_offset < 0 || 4381 in_folio_offset >= folio_nr_pages(folio)) { 4382 if (!split_folio_to_order(folio, target_order)) 4383 split++; 4384 } else { 4385 struct page *split_at = folio_page(folio, 4386 in_folio_offset); 4387 if (!folio_split(folio, target_order, split_at, NULL)) 4388 split++; 4389 } 4390 4391 unlock: 4392 4393 folio_unlock(folio); 4394 folio_put(folio); 4395 4396 cond_resched(); 4397 continue; 4398 next: 4399 folio_walk_end(&fw, vma); 4400 cond_resched(); 4401 } 4402 mmap_read_unlock(mm); 4403 mmput(mm); 4404 4405 pr_debug("%lu of %lu THP split\n", split, total); 4406 4407 out: 4408 return ret; 4409 } 4410 4411 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 4412 pgoff_t off_end, unsigned int new_order, 4413 long in_folio_offset) 4414 { 4415 struct filename *file; 4416 struct file *candidate; 4417 struct address_space *mapping; 4418 int ret = -EINVAL; 4419 pgoff_t index; 4420 int nr_pages = 1; 4421 unsigned long total = 0, split = 0; 4422 unsigned int min_order; 4423 unsigned int target_order; 4424 4425 file = getname_kernel(file_path); 4426 if (IS_ERR(file)) 4427 return ret; 4428 4429 candidate = file_open_name(file, O_RDONLY, 0); 4430 if (IS_ERR(candidate)) 4431 goto out; 4432 4433 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 4434 file_path, off_start, off_end); 4435 4436 mapping = candidate->f_mapping; 4437 min_order = mapping_min_folio_order(mapping); 4438 target_order = max(new_order, min_order); 4439 4440 for (index = off_start; index < off_end; index += nr_pages) { 4441 struct folio *folio = filemap_get_folio(mapping, index); 4442 4443 nr_pages = 1; 4444 if (IS_ERR(folio)) 4445 continue; 4446 4447 if (!folio_test_large(folio)) 4448 goto next; 4449 4450 total++; 4451 nr_pages = folio_nr_pages(folio); 4452 4453 if (target_order >= folio_order(folio)) 4454 goto next; 4455 4456 if (!folio_trylock(folio)) 4457 goto next; 4458 4459 if (folio->mapping != mapping) 4460 goto unlock; 4461 4462 if (in_folio_offset < 0 || in_folio_offset >= nr_pages) { 4463 if (!split_folio_to_order(folio, target_order)) 4464 split++; 4465 } else { 4466 struct page *split_at = folio_page(folio, 4467 in_folio_offset); 4468 if (!folio_split(folio, target_order, split_at, NULL)) 4469 split++; 4470 } 4471 4472 unlock: 4473 folio_unlock(folio); 4474 next: 4475 folio_put(folio); 4476 cond_resched(); 4477 } 4478 4479 filp_close(candidate, NULL); 4480 ret = 0; 4481 4482 pr_debug("%lu of %lu file-backed THP split\n", split, total); 4483 out: 4484 putname(file); 4485 return ret; 4486 } 4487 4488 #define MAX_INPUT_BUF_SZ 255 4489 4490 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 4491 size_t count, loff_t *ppops) 4492 { 4493 static DEFINE_MUTEX(split_debug_mutex); 4494 ssize_t ret; 4495 /* 4496 * hold pid, start_vaddr, end_vaddr, new_order or 4497 * file_path, off_start, off_end, new_order 4498 */ 4499 char input_buf[MAX_INPUT_BUF_SZ]; 4500 int pid; 4501 unsigned long vaddr_start, vaddr_end; 4502 unsigned int new_order = 0; 4503 long in_folio_offset = -1; 4504 4505 ret = mutex_lock_interruptible(&split_debug_mutex); 4506 if (ret) 4507 return ret; 4508 4509 ret = -EFAULT; 4510 4511 memset(input_buf, 0, MAX_INPUT_BUF_SZ); 4512 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 4513 goto out; 4514 4515 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 4516 4517 if (input_buf[0] == '/') { 4518 char *tok; 4519 char *tok_buf = input_buf; 4520 char file_path[MAX_INPUT_BUF_SZ]; 4521 pgoff_t off_start = 0, off_end = 0; 4522 size_t input_len = strlen(input_buf); 4523 4524 tok = strsep(&tok_buf, ","); 4525 if (tok && tok_buf) { 4526 strscpy(file_path, tok); 4527 } else { 4528 ret = -EINVAL; 4529 goto out; 4530 } 4531 4532 ret = sscanf(tok_buf, "0x%lx,0x%lx,%d,%ld", &off_start, &off_end, 4533 &new_order, &in_folio_offset); 4534 if (ret != 2 && ret != 3 && ret != 4) { 4535 ret = -EINVAL; 4536 goto out; 4537 } 4538 ret = split_huge_pages_in_file(file_path, off_start, off_end, 4539 new_order, in_folio_offset); 4540 if (!ret) 4541 ret = input_len; 4542 4543 goto out; 4544 } 4545 4546 ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d,%ld", &pid, &vaddr_start, 4547 &vaddr_end, &new_order, &in_folio_offset); 4548 if (ret == 1 && pid == 1) { 4549 split_huge_pages_all(); 4550 ret = strlen(input_buf); 4551 goto out; 4552 } else if (ret != 3 && ret != 4 && ret != 5) { 4553 ret = -EINVAL; 4554 goto out; 4555 } 4556 4557 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order, 4558 in_folio_offset); 4559 if (!ret) 4560 ret = strlen(input_buf); 4561 out: 4562 mutex_unlock(&split_debug_mutex); 4563 return ret; 4564 4565 } 4566 4567 static const struct file_operations split_huge_pages_fops = { 4568 .owner = THIS_MODULE, 4569 .write = split_huge_pages_write, 4570 }; 4571 4572 static int __init split_huge_pages_debugfs(void) 4573 { 4574 debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 4575 &split_huge_pages_fops); 4576 return 0; 4577 } 4578 late_initcall(split_huge_pages_debugfs); 4579 #endif 4580 4581 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 4582 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 4583 struct page *page) 4584 { 4585 struct folio *folio = page_folio(page); 4586 struct vm_area_struct *vma = pvmw->vma; 4587 struct mm_struct *mm = vma->vm_mm; 4588 unsigned long address = pvmw->address; 4589 bool anon_exclusive; 4590 pmd_t pmdval; 4591 swp_entry_t entry; 4592 pmd_t pmdswp; 4593 4594 if (!(pvmw->pmd && !pvmw->pte)) 4595 return 0; 4596 4597 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 4598 pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 4599 4600 /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */ 4601 anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page); 4602 if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) { 4603 set_pmd_at(mm, address, pvmw->pmd, pmdval); 4604 return -EBUSY; 4605 } 4606 4607 if (pmd_dirty(pmdval)) 4608 folio_mark_dirty(folio); 4609 if (pmd_write(pmdval)) 4610 entry = make_writable_migration_entry(page_to_pfn(page)); 4611 else if (anon_exclusive) 4612 entry = make_readable_exclusive_migration_entry(page_to_pfn(page)); 4613 else 4614 entry = make_readable_migration_entry(page_to_pfn(page)); 4615 if (pmd_young(pmdval)) 4616 entry = make_migration_entry_young(entry); 4617 if (pmd_dirty(pmdval)) 4618 entry = make_migration_entry_dirty(entry); 4619 pmdswp = swp_entry_to_pmd(entry); 4620 if (pmd_soft_dirty(pmdval)) 4621 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 4622 if (pmd_uffd_wp(pmdval)) 4623 pmdswp = pmd_swp_mkuffd_wp(pmdswp); 4624 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 4625 folio_remove_rmap_pmd(folio, page, vma); 4626 folio_put(folio); 4627 trace_set_migration_pmd(address, pmd_val(pmdswp)); 4628 4629 return 0; 4630 } 4631 4632 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 4633 { 4634 struct folio *folio = page_folio(new); 4635 struct vm_area_struct *vma = pvmw->vma; 4636 struct mm_struct *mm = vma->vm_mm; 4637 unsigned long address = pvmw->address; 4638 unsigned long haddr = address & HPAGE_PMD_MASK; 4639 pmd_t pmde; 4640 swp_entry_t entry; 4641 4642 if (!(pvmw->pmd && !pvmw->pte)) 4643 return; 4644 4645 entry = pmd_to_swp_entry(*pvmw->pmd); 4646 folio_get(folio); 4647 pmde = folio_mk_pmd(folio, READ_ONCE(vma->vm_page_prot)); 4648 if (pmd_swp_soft_dirty(*pvmw->pmd)) 4649 pmde = pmd_mksoft_dirty(pmde); 4650 if (is_writable_migration_entry(entry)) 4651 pmde = pmd_mkwrite(pmde, vma); 4652 if (pmd_swp_uffd_wp(*pvmw->pmd)) 4653 pmde = pmd_mkuffd_wp(pmde); 4654 if (!is_migration_entry_young(entry)) 4655 pmde = pmd_mkold(pmde); 4656 /* NOTE: this may contain setting soft-dirty on some archs */ 4657 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 4658 pmde = pmd_mkdirty(pmde); 4659 4660 if (folio_test_anon(folio)) { 4661 rmap_t rmap_flags = RMAP_NONE; 4662 4663 if (!is_readable_migration_entry(entry)) 4664 rmap_flags |= RMAP_EXCLUSIVE; 4665 4666 folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags); 4667 } else { 4668 folio_add_file_rmap_pmd(folio, new, vma); 4669 } 4670 VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new)); 4671 set_pmd_at(mm, haddr, pvmw->pmd, pmde); 4672 4673 /* No need to invalidate - it was non-present before */ 4674 update_mmu_cache_pmd(vma, address, pvmw->pmd); 4675 trace_remove_migration_pmd(address, pmd_val(pmde)); 4676 } 4677 #endif 4678