mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/page_table_check.h>
  20 #include <linux/swapops.h>
  21 #include <linux/shmem_fs.h>
  22
  23 #include <asm/tlb.h>
  24 #include <asm/pgalloc.h>
  25 #include "internal.h"
  26
  27 enum scan_result {
  28         SCAN_FAIL,
  29         SCAN_SUCCEED,
  30         SCAN_PMD_NULL,
  31         SCAN_EXCEED_NONE_PTE,
  32         SCAN_EXCEED_SWAP_PTE,
  33         SCAN_EXCEED_SHARED_PTE,
  34         SCAN_PTE_NON_PRESENT,
  35         SCAN_PTE_UFFD_WP,
  36         SCAN_PAGE_RO,
  37         SCAN_LACK_REFERENCED_PAGE,
  38         SCAN_PAGE_NULL,
  39         SCAN_SCAN_ABORT,
  40         SCAN_PAGE_COUNT,
  41         SCAN_PAGE_LRU,
  42         SCAN_PAGE_LOCK,
  43         SCAN_PAGE_ANON,
  44         SCAN_PAGE_COMPOUND,
  45         SCAN_ANY_PROCESS,
  46         SCAN_VMA_NULL,
  47         SCAN_VMA_CHECK,
  48         SCAN_ADDRESS_RANGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR_RW(scan_sleep_millisecs);
 151
 152 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 153                                           struct kobj_attribute *attr,
 154                                           char *buf)
 155 {
 156         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 157 }
 158
 159 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 160                                            struct kobj_attribute *attr,
 161                                            const char *buf, size_t count)
 162 {
 163         unsigned int msecs;
 164         int err;
 165
 166         err = kstrtouint(buf, 10, &msecs);
 167         if (err)
 168                 return -EINVAL;
 169
 170         khugepaged_alloc_sleep_millisecs = msecs;
 171         khugepaged_sleep_expire = 0;
 172         wake_up_interruptible(&khugepaged_wait);
 173
 174         return count;
 175 }
 176 static struct kobj_attribute alloc_sleep_millisecs_attr =
 177         __ATTR_RW(alloc_sleep_millisecs);
 178
 179 static ssize_t pages_to_scan_show(struct kobject *kobj,
 180                                   struct kobj_attribute *attr,
 181                                   char *buf)
 182 {
 183         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 184 }
 185 static ssize_t pages_to_scan_store(struct kobject *kobj,
 186                                    struct kobj_attribute *attr,
 187                                    const char *buf, size_t count)
 188 {
 189         unsigned int pages;
 190         int err;
 191
 192         err = kstrtouint(buf, 10, &pages);
 193         if (err || !pages)
 194                 return -EINVAL;
 195
 196         khugepaged_pages_to_scan = pages;
 197
 198         return count;
 199 }
 200 static struct kobj_attribute pages_to_scan_attr =
 201         __ATTR_RW(pages_to_scan);
 202
 203 static ssize_t pages_collapsed_show(struct kobject *kobj,
 204                                     struct kobj_attribute *attr,
 205                                     char *buf)
 206 {
 207         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 208 }
 209 static struct kobj_attribute pages_collapsed_attr =
 210         __ATTR_RO(pages_collapsed);
 211
 212 static ssize_t full_scans_show(struct kobject *kobj,
 213                                struct kobj_attribute *attr,
 214                                char *buf)
 215 {
 216         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 217 }
 218 static struct kobj_attribute full_scans_attr =
 219         __ATTR_RO(full_scans);
 220
 221 static ssize_t defrag_show(struct kobject *kobj,
 222                            struct kobj_attribute *attr, char *buf)
 223 {
 224         return single_hugepage_flag_show(kobj, attr, buf,
 225                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 226 }
 227 static ssize_t defrag_store(struct kobject *kobj,
 228                             struct kobj_attribute *attr,
 229                             const char *buf, size_t count)
 230 {
 231         return single_hugepage_flag_store(kobj, attr, buf, count,
 232                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 233 }
 234 static struct kobj_attribute khugepaged_defrag_attr =
 235         __ATTR_RW(defrag);
 236
 237 /*
 238  * max_ptes_none controls if khugepaged should collapse hugepages over
 239  * any unmapped ptes in turn potentially increasing the memory
 240  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 241  * reduce the available free memory in the system as it
 242  * runs. Increasing max_ptes_none will instead potentially reduce the
 243  * free memory in the system during the khugepaged scan.
 244  */
 245 static ssize_t max_ptes_none_show(struct kobject *kobj,
 246                                   struct kobj_attribute *attr,
 247                                   char *buf)
 248 {
 249         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 250 }
 251 static ssize_t max_ptes_none_store(struct kobject *kobj,
 252                                    struct kobj_attribute *attr,
 253                                    const char *buf, size_t count)
 254 {
 255         int err;
 256         unsigned long max_ptes_none;
 257
 258         err = kstrtoul(buf, 10, &max_ptes_none);
 259         if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 260                 return -EINVAL;
 261
 262         khugepaged_max_ptes_none = max_ptes_none;
 263
 264         return count;
 265 }
 266 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 267         __ATTR_RW(max_ptes_none);
 268
 269 static ssize_t max_ptes_swap_show(struct kobject *kobj,
 270                                   struct kobj_attribute *attr,
 271                                   char *buf)
 272 {
 273         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 274 }
 275
 276 static ssize_t max_ptes_swap_store(struct kobject *kobj,
 277                                    struct kobj_attribute *attr,
 278                                    const char *buf, size_t count)
 279 {
 280         int err;
 281         unsigned long max_ptes_swap;
 282
 283         err  = kstrtoul(buf, 10, &max_ptes_swap);
 284         if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 285                 return -EINVAL;
 286
 287         khugepaged_max_ptes_swap = max_ptes_swap;
 288
 289         return count;
 290 }
 291
 292 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 293         __ATTR_RW(max_ptes_swap);
 294
 295 static ssize_t max_ptes_shared_show(struct kobject *kobj,
 296                                     struct kobj_attribute *attr,
 297                                     char *buf)
 298 {
 299         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 300 }
 301
 302 static ssize_t max_ptes_shared_store(struct kobject *kobj,
 303                                      struct kobj_attribute *attr,
 304                                      const char *buf, size_t count)
 305 {
 306         int err;
 307         unsigned long max_ptes_shared;
 308
 309         err  = kstrtoul(buf, 10, &max_ptes_shared);
 310         if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 311                 return -EINVAL;
 312
 313         khugepaged_max_ptes_shared = max_ptes_shared;
 314
 315         return count;
 316 }
 317
 318 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 319         __ATTR_RW(max_ptes_shared);
 320
 321 static struct attribute *khugepaged_attr[] = {
 322         &khugepaged_defrag_attr.attr,
 323         &khugepaged_max_ptes_none_attr.attr,
 324         &khugepaged_max_ptes_swap_attr.attr,
 325         &khugepaged_max_ptes_shared_attr.attr,
 326         &pages_to_scan_attr.attr,
 327         &pages_collapsed_attr.attr,
 328         &full_scans_attr.attr,
 329         &scan_sleep_millisecs_attr.attr,
 330         &alloc_sleep_millisecs_attr.attr,
 331         NULL,
 332 };
 333
 334 struct attribute_group khugepaged_attr_group = {
 335         .attrs = khugepaged_attr,
 336         .name = "khugepaged",
 337 };
 338 #endif /* CONFIG_SYSFS */
 339
 340 int hugepage_madvise(struct vm_area_struct *vma,
 341                      unsigned long *vm_flags, int advice)
 342 {
 343         switch (advice) {
 344         case MADV_HUGEPAGE:
 345 #ifdef CONFIG_S390
 346                 /*
 347                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 348                  * can't handle this properly after s390_enable_sie, so we simply
 349                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 350                  */
 351                 if (mm_has_pgste(vma->vm_mm))
 352                         return 0;
 353 #endif
 354                 *vm_flags &= ~VM_NOHUGEPAGE;
 355                 *vm_flags |= VM_HUGEPAGE;
 356                 /*
 357                  * If the vma become good for khugepaged to scan,
 358                  * register it here without waiting a page fault that
 359                  * may not happen any time soon.
 360                  */
 361                 khugepaged_enter_vma(vma, *vm_flags);
 362                 break;
 363         case MADV_NOHUGEPAGE:
 364                 *vm_flags &= ~VM_HUGEPAGE;
 365                 *vm_flags |= VM_NOHUGEPAGE;
 366                 /*
 367                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 368                  * this vma even if we leave the mm registered in khugepaged if
 369                  * it got registered before VM_NOHUGEPAGE was set.
 370                  */
 371                 break;
 372         }
 373
 374         return 0;
 375 }
 376
 377 int __init khugepaged_init(void)
 378 {
 379         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 380                                           sizeof(struct mm_slot),
 381                                           __alignof__(struct mm_slot), 0, NULL);
 382         if (!mm_slot_cache)
 383                 return -ENOMEM;
 384
 385         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 386         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 387         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 388         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 389
 390         return 0;
 391 }
 392
 393 void __init khugepaged_destroy(void)
 394 {
 395         kmem_cache_destroy(mm_slot_cache);
 396 }
 397
 398 static inline struct mm_slot *alloc_mm_slot(void)
 399 {
 400         if (!mm_slot_cache)     /* initialization failed */
 401                 return NULL;
 402         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 403 }
 404
 405 static inline void free_mm_slot(struct mm_slot *mm_slot)
 406 {
 407         kmem_cache_free(mm_slot_cache, mm_slot);
 408 }
 409
 410 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 411 {
 412         struct mm_slot *mm_slot;
 413
 414         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 415                 if (mm == mm_slot->mm)
 416                         return mm_slot;
 417
 418         return NULL;
 419 }
 420
 421 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 422                                     struct mm_slot *mm_slot)
 423 {
 424         mm_slot->mm = mm;
 425         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 426 }
 427
 428 static inline int khugepaged_test_exit(struct mm_struct *mm)
 429 {
 430         return atomic_read(&mm->mm_users) == 0;
 431 }
 432
 433 void __khugepaged_enter(struct mm_struct *mm)
 434 {
 435         struct mm_slot *mm_slot;
 436         int wakeup;
 437
 438         mm_slot = alloc_mm_slot();
 439         if (!mm_slot)
 440                 return;
 441
 442         /* __khugepaged_exit() must not run from under us */
 443         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 444         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 445                 free_mm_slot(mm_slot);
 446                 return;
 447         }
 448
 449         spin_lock(&khugepaged_mm_lock);
 450         insert_to_mm_slots_hash(mm, mm_slot);
 451         /*
 452          * Insert just behind the scanning cursor, to let the area settle
 453          * down a little.
 454          */
 455         wakeup = list_empty(&khugepaged_scan.mm_head);
 456         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 457         spin_unlock(&khugepaged_mm_lock);
 458
 459         mmgrab(mm);
 460         if (wakeup)
 461                 wake_up_interruptible(&khugepaged_wait);
 462 }
 463
 464 void khugepaged_enter_vma(struct vm_area_struct *vma,
 465                           unsigned long vm_flags)
 466 {
 467         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 468             khugepaged_enabled()) {
 469                 if (hugepage_vma_check(vma, vm_flags, false, false))
 470                         __khugepaged_enter(vma->vm_mm);
 471         }
 472 }
 473
 474 void __khugepaged_exit(struct mm_struct *mm)
 475 {
 476         struct mm_slot *mm_slot;
 477         int free = 0;
 478
 479         spin_lock(&khugepaged_mm_lock);
 480         mm_slot = get_mm_slot(mm);
 481         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 482                 hash_del(&mm_slot->hash);
 483                 list_del(&mm_slot->mm_node);
 484                 free = 1;
 485         }
 486         spin_unlock(&khugepaged_mm_lock);
 487
 488         if (free) {
 489                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 490                 free_mm_slot(mm_slot);
 491                 mmdrop(mm);
 492         } else if (mm_slot) {
 493                 /*
 494                  * This is required to serialize against
 495                  * khugepaged_test_exit() (which is guaranteed to run
 496                  * under mmap sem read mode). Stop here (after we
 497                  * return all pagetables will be destroyed) until
 498                  * khugepaged has finished working on the pagetables
 499                  * under the mmap_lock.
 500                  */
 501                 mmap_write_lock(mm);
 502                 mmap_write_unlock(mm);
 503         }
 504 }
 505
 506 static void release_pte_page(struct page *page)
 507 {
 508         mod_node_page_state(page_pgdat(page),
 509                         NR_ISOLATED_ANON + page_is_file_lru(page),
 510                         -compound_nr(page));
 511         unlock_page(page);
 512         putback_lru_page(page);
 513 }
 514
 515 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 516                 struct list_head *compound_pagelist)
 517 {
 518         struct page *page, *tmp;
 519
 520         while (--_pte >= pte) {
 521                 pte_t pteval = *_pte;
 522
 523                 page = pte_page(pteval);
 524                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 525                                 !PageCompound(page))
 526                         release_pte_page(page);
 527         }
 528
 529         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 530                 list_del(&page->lru);
 531                 release_pte_page(page);
 532         }
 533 }
 534
 535 static bool is_refcount_suitable(struct page *page)
 536 {
 537         int expected_refcount;
 538
 539         expected_refcount = total_mapcount(page);
 540         if (PageSwapCache(page))
 541                 expected_refcount += compound_nr(page);
 542
 543         return page_count(page) == expected_refcount;
 544 }
 545
 546 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 547                                         unsigned long address,
 548                                         pte_t *pte,
 549                                         struct list_head *compound_pagelist)
 550 {
 551         struct page *page = NULL;
 552         pte_t *_pte;
 553         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 554         bool writable = false;
 555
 556         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 557              _pte++, address += PAGE_SIZE) {
 558                 pte_t pteval = *_pte;
 559                 if (pte_none(pteval) || (pte_present(pteval) &&
 560                                 is_zero_pfn(pte_pfn(pteval)))) {
 561                         if (!userfaultfd_armed(vma) &&
 562                             ++none_or_zero <= khugepaged_max_ptes_none) {
 563                                 continue;
 564                         } else {
 565                                 result = SCAN_EXCEED_NONE_PTE;
 566                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 567                                 goto out;
 568                         }
 569                 }
 570                 if (!pte_present(pteval)) {
 571                         result = SCAN_PTE_NON_PRESENT;
 572                         goto out;
 573                 }
 574                 page = vm_normal_page(vma, address, pteval);
 575                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 576                         result = SCAN_PAGE_NULL;
 577                         goto out;
 578                 }
 579
 580                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 581
 582                 if (page_mapcount(page) > 1 &&
 583                                 ++shared > khugepaged_max_ptes_shared) {
 584                         result = SCAN_EXCEED_SHARED_PTE;
 585                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 586                         goto out;
 587                 }
 588
 589                 if (PageCompound(page)) {
 590                         struct page *p;
 591                         page = compound_head(page);
 592
 593                         /*
 594                          * Check if we have dealt with the compound page
 595                          * already
 596                          */
 597                         list_for_each_entry(p, compound_pagelist, lru) {
 598                                 if (page == p)
 599                                         goto next;
 600                         }
 601                 }
 602
 603                 /*
 604                  * We can do it before isolate_lru_page because the
 605                  * page can't be freed from under us. NOTE: PG_lock
 606                  * is needed to serialize against split_huge_page
 607                  * when invoked from the VM.
 608                  */
 609                 if (!trylock_page(page)) {
 610                         result = SCAN_PAGE_LOCK;
 611                         goto out;
 612                 }
 613
 614                 /*
 615                  * Check if the page has any GUP (or other external) pins.
 616                  *
 617                  * The page table that maps the page has been already unlinked
 618                  * from the page table tree and this process cannot get
 619                  * an additional pin on the page.
 620                  *
 621                  * New pins can come later if the page is shared across fork,
 622                  * but not from this process. The other process cannot write to
 623                  * the page, only trigger CoW.
 624                  */
 625                 if (!is_refcount_suitable(page)) {
 626                         unlock_page(page);
 627                         result = SCAN_PAGE_COUNT;
 628                         goto out;
 629                 }
 630
 631                 /*
 632                  * Isolate the page to avoid collapsing an hugepage
 633                  * currently in use by the VM.
 634                  */
 635                 if (isolate_lru_page(page)) {
 636                         unlock_page(page);
 637                         result = SCAN_DEL_PAGE_LRU;
 638                         goto out;
 639                 }
 640                 mod_node_page_state(page_pgdat(page),
 641                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 642                                 compound_nr(page));
 643                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 644                 VM_BUG_ON_PAGE(PageLRU(page), page);
 645
 646                 if (PageCompound(page))
 647                         list_add_tail(&page->lru, compound_pagelist);
 648 next:
 649                 /* There should be enough young pte to collapse the page */
 650                 if (pte_young(pteval) ||
 651                     page_is_young(page) || PageReferenced(page) ||
 652                     mmu_notifier_test_young(vma->vm_mm, address))
 653                         referenced++;
 654
 655                 if (pte_write(pteval))
 656                         writable = true;
 657         }
 658
 659         if (unlikely(!writable)) {
 660                 result = SCAN_PAGE_RO;
 661         } else if (unlikely(!referenced)) {
 662                 result = SCAN_LACK_REFERENCED_PAGE;
 663         } else {
 664                 result = SCAN_SUCCEED;
 665                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 666                                                     referenced, writable, result);
 667                 return 1;
 668         }
 669 out:
 670         release_pte_pages(pte, _pte, compound_pagelist);
 671         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 672                                             referenced, writable, result);
 673         return 0;
 674 }
 675
 676 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 677                                       struct vm_area_struct *vma,
 678                                       unsigned long address,
 679                                       spinlock_t *ptl,
 680                                       struct list_head *compound_pagelist)
 681 {
 682         struct page *src_page, *tmp;
 683         pte_t *_pte;
 684         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 685                                 _pte++, page++, address += PAGE_SIZE) {
 686                 pte_t pteval = *_pte;
 687
 688                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 689                         clear_user_highpage(page, address);
 690                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 691                         if (is_zero_pfn(pte_pfn(pteval))) {
 692                                 /*
 693                                  * ptl mostly unnecessary.
 694                                  */
 695                                 spin_lock(ptl);
 696                                 ptep_clear(vma->vm_mm, address, _pte);
 697                                 spin_unlock(ptl);
 698                         }
 699                 } else {
 700                         src_page = pte_page(pteval);
 701                         copy_user_highpage(page, src_page, address, vma);
 702                         if (!PageCompound(src_page))
 703                                 release_pte_page(src_page);
 704                         /*
 705                          * ptl mostly unnecessary, but preempt has to
 706                          * be disabled to update the per-cpu stats
 707                          * inside page_remove_rmap().
 708                          */
 709                         spin_lock(ptl);
 710                         ptep_clear(vma->vm_mm, address, _pte);
 711                         page_remove_rmap(src_page, vma, false);
 712                         spin_unlock(ptl);
 713                         free_page_and_swap_cache(src_page);
 714                 }
 715         }
 716
 717         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 718                 list_del(&src_page->lru);
 719                 mod_node_page_state(page_pgdat(src_page),
 720                                     NR_ISOLATED_ANON + page_is_file_lru(src_page),
 721                                     -compound_nr(src_page));
 722                 unlock_page(src_page);
 723                 free_swap_cache(src_page);
 724                 putback_lru_page(src_page);
 725         }
 726 }
 727
 728 static void khugepaged_alloc_sleep(void)
 729 {
 730         DEFINE_WAIT(wait);
 731
 732         add_wait_queue(&khugepaged_wait, &wait);
 733         freezable_schedule_timeout_interruptible(
 734                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 735         remove_wait_queue(&khugepaged_wait, &wait);
 736 }
 737
 738 static int khugepaged_node_load[MAX_NUMNODES];
 739
 740 static bool khugepaged_scan_abort(int nid)
 741 {
 742         int i;
 743
 744         /*
 745          * If node_reclaim_mode is disabled, then no extra effort is made to
 746          * allocate memory locally.
 747          */
 748         if (!node_reclaim_enabled())
 749                 return false;
 750
 751         /* If there is a count for this node already, it must be acceptable */
 752         if (khugepaged_node_load[nid])
 753                 return false;
 754
 755         for (i = 0; i < MAX_NUMNODES; i++) {
 756                 if (!khugepaged_node_load[i])
 757                         continue;
 758                 if (node_distance(nid, i) > node_reclaim_distance)
 759                         return true;
 760         }
 761         return false;
 762 }
 763
 764 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 765 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 766 {
 767         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 768 }
 769
 770 #ifdef CONFIG_NUMA
 771 static int khugepaged_find_target_node(void)
 772 {
 773         static int last_khugepaged_target_node = NUMA_NO_NODE;
 774         int nid, target_node = 0, max_value = 0;
 775
 776         /* find first node with max normal pages hit */
 777         for (nid = 0; nid < MAX_NUMNODES; nid++)
 778                 if (khugepaged_node_load[nid] > max_value) {
 779                         max_value = khugepaged_node_load[nid];
 780                         target_node = nid;
 781                 }
 782
 783         /* do some balance if several nodes have the same hit record */
 784         if (target_node <= last_khugepaged_target_node)
 785                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 786                                 nid++)
 787                         if (max_value == khugepaged_node_load[nid]) {
 788                                 target_node = nid;
 789                                 break;
 790                         }
 791
 792         last_khugepaged_target_node = target_node;
 793         return target_node;
 794 }
 795
 796 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 797 {
 798         if (IS_ERR(*hpage)) {
 799                 if (!*wait)
 800                         return false;
 801
 802                 *wait = false;
 803                 *hpage = NULL;
 804                 khugepaged_alloc_sleep();
 805         } else if (*hpage) {
 806                 put_page(*hpage);
 807                 *hpage = NULL;
 808         }
 809
 810         return true;
 811 }
 812
 813 static struct page *
 814 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 815 {
 816         VM_BUG_ON_PAGE(*hpage, *hpage);
 817
 818         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 819         if (unlikely(!*hpage)) {
 820                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 821                 *hpage = ERR_PTR(-ENOMEM);
 822                 return NULL;
 823         }
 824
 825         prep_transhuge_page(*hpage);
 826         count_vm_event(THP_COLLAPSE_ALLOC);
 827         return *hpage;
 828 }
 829 #else
 830 static int khugepaged_find_target_node(void)
 831 {
 832         return 0;
 833 }
 834
 835 static inline struct page *alloc_khugepaged_hugepage(void)
 836 {
 837         struct page *page;
 838
 839         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 840                            HPAGE_PMD_ORDER);
 841         if (page)
 842                 prep_transhuge_page(page);
 843         return page;
 844 }
 845
 846 static struct page *khugepaged_alloc_hugepage(bool *wait)
 847 {
 848         struct page *hpage;
 849
 850         do {
 851                 hpage = alloc_khugepaged_hugepage();
 852                 if (!hpage) {
 853                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 854                         if (!*wait)
 855                                 return NULL;
 856
 857                         *wait = false;
 858                         khugepaged_alloc_sleep();
 859                 } else
 860                         count_vm_event(THP_COLLAPSE_ALLOC);
 861         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 862
 863         return hpage;
 864 }
 865
 866 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 867 {
 868         /*
 869          * If the hpage allocated earlier was briefly exposed in page cache
 870          * before collapse_file() failed, it is possible that racing lookups
 871          * have not yet completed, and would then be unpleasantly surprised by
 872          * finding the hpage reused for the same mapping at a different offset.
 873          * Just release the previous allocation if there is any danger of that.
 874          */
 875         if (*hpage && page_count(*hpage) > 1) {
 876                 put_page(*hpage);
 877                 *hpage = NULL;
 878         }
 879
 880         if (!*hpage)
 881                 *hpage = khugepaged_alloc_hugepage(wait);
 882
 883         if (unlikely(!*hpage))
 884                 return false;
 885
 886         return true;
 887 }
 888
 889 static struct page *
 890 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 891 {
 892         VM_BUG_ON(!*hpage);
 893
 894         return  *hpage;
 895 }
 896 #endif
 897
 898 /*
 899  * If mmap_lock temporarily dropped, revalidate vma
 900  * before taking mmap_lock.
 901  * Return 0 if succeeds, otherwise return none-zero
 902  * value (scan code).
 903  */
 904
 905 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 906                 struct vm_area_struct **vmap)
 907 {
 908         struct vm_area_struct *vma;
 909
 910         if (unlikely(khugepaged_test_exit(mm)))
 911                 return SCAN_ANY_PROCESS;
 912
 913         *vmap = vma = find_vma(mm, address);
 914         if (!vma)
 915                 return SCAN_VMA_NULL;
 916
 917         if (!transhuge_vma_suitable(vma, address))
 918                 return SCAN_ADDRESS_RANGE;
 919         if (!hugepage_vma_check(vma, vma->vm_flags, false, false))
 920                 return SCAN_VMA_CHECK;
 921         /*
 922          * Anon VMA expected, the address may be unmapped then
 923          * remapped to file after khugepaged reaquired the mmap_lock.
 924          *
 925          * hugepage_vma_check may return true for qualified file
 926          * vmas.
 927          */
 928         if (!vma->anon_vma || !vma_is_anonymous(vma))
 929                 return SCAN_VMA_CHECK;
 930         return 0;
 931 }
 932
 933 /*
 934  * Bring missing pages in from swap, to complete THP collapse.
 935  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 936  *
 937  * Called and returns without pte mapped or spinlocks held.
 938  * Note that if false is returned, mmap_lock will be released.
 939  */
 940
 941 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 942                                         struct vm_area_struct *vma,
 943                                         unsigned long haddr, pmd_t *pmd,
 944                                         int referenced)
 945 {
 946         int swapped_in = 0;
 947         vm_fault_t ret = 0;
 948         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 949
 950         for (address = haddr; address < end; address += PAGE_SIZE) {
 951                 struct vm_fault vmf = {
 952                         .vma = vma,
 953                         .address = address,
 954                         .pgoff = linear_page_index(vma, haddr),
 955                         .flags = FAULT_FLAG_ALLOW_RETRY,
 956                         .pmd = pmd,
 957                 };
 958
 959                 vmf.pte = pte_offset_map(pmd, address);
 960                 vmf.orig_pte = *vmf.pte;
 961                 if (!is_swap_pte(vmf.orig_pte)) {
 962                         pte_unmap(vmf.pte);
 963                         continue;
 964                 }
 965                 ret = do_swap_page(&vmf);
 966
 967                 /*
 968                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
 969                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
 970                  * we do not retry here and swap entry will remain in pagetable
 971                  * resulting in later failure.
 972                  */
 973                 if (ret & VM_FAULT_RETRY) {
 974                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 975                         return false;
 976                 }
 977                 if (ret & VM_FAULT_ERROR) {
 978                         mmap_read_unlock(mm);
 979                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 980                         return false;
 981                 }
 982                 swapped_in++;
 983         }
 984
 985         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
 986         if (swapped_in)
 987                 lru_add_drain();
 988
 989         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 990         return true;
 991 }
 992
 993 static void collapse_huge_page(struct mm_struct *mm,
 994                                    unsigned long address,
 995                                    struct page **hpage,
 996                                    int node, int referenced, int unmapped)
 997 {
 998         LIST_HEAD(compound_pagelist);
 999         pmd_t *pmd, _pmd;
1000         pte_t *pte;
1001         pgtable_t pgtable;
1002         struct page *new_page;
1003         spinlock_t *pmd_ptl, *pte_ptl;
1004         int isolated = 0, result = 0;
1005         struct vm_area_struct *vma;
1006         struct mmu_notifier_range range;
1007         gfp_t gfp;
1008
1009         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1010
1011         /* Only allocate from the target node */
1012         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1013
1014         /*
1015          * Before allocating the hugepage, release the mmap_lock read lock.
1016          * The allocation can take potentially a long time if it involves
1017          * sync compaction, and we do not need to hold the mmap_lock during
1018          * that. We will recheck the vma after taking it again in write mode.
1019          */
1020         mmap_read_unlock(mm);
1021         new_page = khugepaged_alloc_page(hpage, gfp, node);
1022         if (!new_page) {
1023                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1024                 goto out_nolock;
1025         }
1026
1027         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1028                 result = SCAN_CGROUP_CHARGE_FAIL;
1029                 goto out_nolock;
1030         }
1031         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1032
1033         mmap_read_lock(mm);
1034         result = hugepage_vma_revalidate(mm, address, &vma);
1035         if (result) {
1036                 mmap_read_unlock(mm);
1037                 goto out_nolock;
1038         }
1039
1040         pmd = mm_find_pmd(mm, address);
1041         if (!pmd) {
1042                 result = SCAN_PMD_NULL;
1043                 mmap_read_unlock(mm);
1044                 goto out_nolock;
1045         }
1046
1047         /*
1048          * __collapse_huge_page_swapin will return with mmap_lock released
1049          * when it fails. So we jump out_nolock directly in that case.
1050          * Continuing to collapse causes inconsistency.
1051          */
1052         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1053                                                      pmd, referenced)) {
1054                 goto out_nolock;
1055         }
1056
1057         mmap_read_unlock(mm);
1058         /*
1059          * Prevent all access to pagetables with the exception of
1060          * gup_fast later handled by the ptep_clear_flush and the VM
1061          * handled by the anon_vma lock + PG_lock.
1062          */
1063         mmap_write_lock(mm);
1064         result = hugepage_vma_revalidate(mm, address, &vma);
1065         if (result)
1066                 goto out_up_write;
1067         /* check if the pmd is still valid */
1068         if (mm_find_pmd(mm, address) != pmd)
1069                 goto out_up_write;
1070
1071         anon_vma_lock_write(vma->anon_vma);
1072
1073         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1074                                 address, address + HPAGE_PMD_SIZE);
1075         mmu_notifier_invalidate_range_start(&range);
1076
1077         pte = pte_offset_map(pmd, address);
1078         pte_ptl = pte_lockptr(mm, pmd);
1079
1080         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1081         /*
1082          * After this gup_fast can't run anymore. This also removes
1083          * any huge TLB entry from the CPU so we won't allow
1084          * huge and small TLB entries for the same virtual address
1085          * to avoid the risk of CPU bugs in that area.
1086          */
1087         _pmd = pmdp_collapse_flush(vma, address, pmd);
1088         spin_unlock(pmd_ptl);
1089         mmu_notifier_invalidate_range_end(&range);
1090
1091         spin_lock(pte_ptl);
1092         isolated = __collapse_huge_page_isolate(vma, address, pte,
1093                         &compound_pagelist);
1094         spin_unlock(pte_ptl);
1095
1096         if (unlikely(!isolated)) {
1097                 pte_unmap(pte);
1098                 spin_lock(pmd_ptl);
1099                 BUG_ON(!pmd_none(*pmd));
1100                 /*
1101                  * We can only use set_pmd_at when establishing
1102                  * hugepmds and never for establishing regular pmds that
1103                  * points to regular pagetables. Use pmd_populate for that
1104                  */
1105                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1106                 spin_unlock(pmd_ptl);
1107                 anon_vma_unlock_write(vma->anon_vma);
1108                 result = SCAN_FAIL;
1109                 goto out_up_write;
1110         }
1111
1112         /*
1113          * All pages are isolated and locked so anon_vma rmap
1114          * can't run anymore.
1115          */
1116         anon_vma_unlock_write(vma->anon_vma);
1117
1118         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1119                         &compound_pagelist);
1120         pte_unmap(pte);
1121         /*
1122          * spin_lock() below is not the equivalent of smp_wmb(), but
1123          * the smp_wmb() inside __SetPageUptodate() can be reused to
1124          * avoid the copy_huge_page writes to become visible after
1125          * the set_pmd_at() write.
1126          */
1127         __SetPageUptodate(new_page);
1128         pgtable = pmd_pgtable(_pmd);
1129
1130         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1131         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1132
1133         spin_lock(pmd_ptl);
1134         BUG_ON(!pmd_none(*pmd));
1135         page_add_new_anon_rmap(new_page, vma, address);
1136         lru_cache_add_inactive_or_unevictable(new_page, vma);
1137         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1138         set_pmd_at(mm, address, pmd, _pmd);
1139         update_mmu_cache_pmd(vma, address, pmd);
1140         spin_unlock(pmd_ptl);
1141
1142         *hpage = NULL;
1143
1144         khugepaged_pages_collapsed++;
1145         result = SCAN_SUCCEED;
1146 out_up_write:
1147         mmap_write_unlock(mm);
1148 out_nolock:
1149         if (!IS_ERR_OR_NULL(*hpage))
1150                 mem_cgroup_uncharge(page_folio(*hpage));
1151         trace_mm_collapse_huge_page(mm, isolated, result);
1152         return;
1153 }
1154
1155 static int khugepaged_scan_pmd(struct mm_struct *mm,
1156                                struct vm_area_struct *vma,
1157                                unsigned long address,
1158                                struct page **hpage)
1159 {
1160         pmd_t *pmd;
1161         pte_t *pte, *_pte;
1162         int ret = 0, result = 0, referenced = 0;
1163         int none_or_zero = 0, shared = 0;
1164         struct page *page = NULL;
1165         unsigned long _address;
1166         spinlock_t *ptl;
1167         int node = NUMA_NO_NODE, unmapped = 0;
1168         bool writable = false;
1169
1170         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1171
1172         pmd = mm_find_pmd(mm, address);
1173         if (!pmd) {
1174                 result = SCAN_PMD_NULL;
1175                 goto out;
1176         }
1177
1178         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1179         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1180         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1181              _pte++, _address += PAGE_SIZE) {
1182                 pte_t pteval = *_pte;
1183                 if (is_swap_pte(pteval)) {
1184                         if (++unmapped <= khugepaged_max_ptes_swap) {
1185                                 /*
1186                                  * Always be strict with uffd-wp
1187                                  * enabled swap entries.  Please see
1188                                  * comment below for pte_uffd_wp().
1189                                  */
1190                                 if (pte_swp_uffd_wp(pteval)) {
1191                                         result = SCAN_PTE_UFFD_WP;
1192                                         goto out_unmap;
1193                                 }
1194                                 continue;
1195                         } else {
1196                                 result = SCAN_EXCEED_SWAP_PTE;
1197                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1198                                 goto out_unmap;
1199                         }
1200                 }
1201                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1202                         if (!userfaultfd_armed(vma) &&
1203                             ++none_or_zero <= khugepaged_max_ptes_none) {
1204                                 continue;
1205                         } else {
1206                                 result = SCAN_EXCEED_NONE_PTE;
1207                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1208                                 goto out_unmap;
1209                         }
1210                 }
1211                 if (pte_uffd_wp(pteval)) {
1212                         /*
1213                          * Don't collapse the page if any of the small
1214                          * PTEs are armed with uffd write protection.
1215                          * Here we can also mark the new huge pmd as
1216                          * write protected if any of the small ones is
1217                          * marked but that could bring unknown
1218                          * userfault messages that falls outside of
1219                          * the registered range.  So, just be simple.
1220                          */
1221                         result = SCAN_PTE_UFFD_WP;
1222                         goto out_unmap;
1223                 }
1224                 if (pte_write(pteval))
1225                         writable = true;
1226
1227                 page = vm_normal_page(vma, _address, pteval);
1228                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1229                         result = SCAN_PAGE_NULL;
1230                         goto out_unmap;
1231                 }
1232
1233                 if (page_mapcount(page) > 1 &&
1234                                 ++shared > khugepaged_max_ptes_shared) {
1235                         result = SCAN_EXCEED_SHARED_PTE;
1236                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1237                         goto out_unmap;
1238                 }
1239
1240                 page = compound_head(page);
1241
1242                 /*
1243                  * Record which node the original page is from and save this
1244                  * information to khugepaged_node_load[].
1245                  * Khugepaged will allocate hugepage from the node has the max
1246                  * hit record.
1247                  */
1248                 node = page_to_nid(page);
1249                 if (khugepaged_scan_abort(node)) {
1250                         result = SCAN_SCAN_ABORT;
1251                         goto out_unmap;
1252                 }
1253                 khugepaged_node_load[node]++;
1254                 if (!PageLRU(page)) {
1255                         result = SCAN_PAGE_LRU;
1256                         goto out_unmap;
1257                 }
1258                 if (PageLocked(page)) {
1259                         result = SCAN_PAGE_LOCK;
1260                         goto out_unmap;
1261                 }
1262                 if (!PageAnon(page)) {
1263                         result = SCAN_PAGE_ANON;
1264                         goto out_unmap;
1265                 }
1266
1267                 /*
1268                  * Check if the page has any GUP (or other external) pins.
1269                  *
1270                  * Here the check is racy it may see total_mapcount > refcount
1271                  * in some cases.
1272                  * For example, one process with one forked child process.
1273                  * The parent has the PMD split due to MADV_DONTNEED, then
1274                  * the child is trying unmap the whole PMD, but khugepaged
1275                  * may be scanning the parent between the child has
1276                  * PageDoubleMap flag cleared and dec the mapcount.  So
1277                  * khugepaged may see total_mapcount > refcount.
1278                  *
1279                  * But such case is ephemeral we could always retry collapse
1280                  * later.  However it may report false positive if the page
1281                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1282                  * will be done again later the risk seems low.
1283                  */
1284                 if (!is_refcount_suitable(page)) {
1285                         result = SCAN_PAGE_COUNT;
1286                         goto out_unmap;
1287                 }
1288                 if (pte_young(pteval) ||
1289                     page_is_young(page) || PageReferenced(page) ||
1290                     mmu_notifier_test_young(vma->vm_mm, address))
1291                         referenced++;
1292         }
1293         if (!writable) {
1294                 result = SCAN_PAGE_RO;
1295         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1296                 result = SCAN_LACK_REFERENCED_PAGE;
1297         } else {
1298                 result = SCAN_SUCCEED;
1299                 ret = 1;
1300         }
1301 out_unmap:
1302         pte_unmap_unlock(pte, ptl);
1303         if (ret) {
1304                 node = khugepaged_find_target_node();
1305                 /* collapse_huge_page will return with the mmap_lock released */
1306                 collapse_huge_page(mm, address, hpage, node,
1307                                 referenced, unmapped);
1308         }
1309 out:
1310         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1311                                      none_or_zero, result, unmapped);
1312         return ret;
1313 }
1314
1315 static void collect_mm_slot(struct mm_slot *mm_slot)
1316 {
1317         struct mm_struct *mm = mm_slot->mm;
1318
1319         lockdep_assert_held(&khugepaged_mm_lock);
1320
1321         if (khugepaged_test_exit(mm)) {
1322                 /* free mm_slot */
1323                 hash_del(&mm_slot->hash);
1324                 list_del(&mm_slot->mm_node);
1325
1326                 /*
1327                  * Not strictly needed because the mm exited already.
1328                  *
1329                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1330                  */
1331
1332                 /* khugepaged_mm_lock actually not necessary for the below */
1333                 free_mm_slot(mm_slot);
1334                 mmdrop(mm);
1335         }
1336 }
1337
1338 #ifdef CONFIG_SHMEM
1339 /*
1340  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1341  * khugepaged should try to collapse the page table.
1342  */
1343 static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1344                                           unsigned long addr)
1345 {
1346         struct mm_slot *mm_slot;
1347
1348         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1349
1350         spin_lock(&khugepaged_mm_lock);
1351         mm_slot = get_mm_slot(mm);
1352         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1353                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1354         spin_unlock(&khugepaged_mm_lock);
1355 }
1356
1357 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1358                                   unsigned long addr, pmd_t *pmdp)
1359 {
1360         spinlock_t *ptl;
1361         pmd_t pmd;
1362
1363         mmap_assert_write_locked(mm);
1364         ptl = pmd_lock(vma->vm_mm, pmdp);
1365         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1366         spin_unlock(ptl);
1367         mm_dec_nr_ptes(mm);
1368         page_table_check_pte_clear_range(mm, addr, pmd);
1369         pte_free(mm, pmd_pgtable(pmd));
1370 }
1371
1372 /**
1373  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1374  * address haddr.
1375  *
1376  * @mm: process address space where collapse happens
1377  * @addr: THP collapse address
1378  *
1379  * This function checks whether all the PTEs in the PMD are pointing to the
1380  * right THP. If so, retract the page table so the THP can refault in with
1381  * as pmd-mapped.
1382  */
1383 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1384 {
1385         unsigned long haddr = addr & HPAGE_PMD_MASK;
1386         struct vm_area_struct *vma = find_vma(mm, haddr);
1387         struct page *hpage;
1388         pte_t *start_pte, *pte;
1389         pmd_t *pmd;
1390         spinlock_t *ptl;
1391         int count = 0;
1392         int i;
1393
1394         if (!vma || !vma->vm_file ||
1395             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1396                 return;
1397
1398         /*
1399          * This vm_flags may not have VM_HUGEPAGE if the page was not
1400          * collapsed by this mm. But we can still collapse if the page is
1401          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1402          * will not fail the vma for missing VM_HUGEPAGE
1403          */
1404         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE, false, false))
1405                 return;
1406
1407         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1408         if (userfaultfd_wp(vma))
1409                 return;
1410
1411         hpage = find_lock_page(vma->vm_file->f_mapping,
1412                                linear_page_index(vma, haddr));
1413         if (!hpage)
1414                 return;
1415
1416         if (!PageHead(hpage))
1417                 goto drop_hpage;
1418
1419         pmd = mm_find_pmd(mm, haddr);
1420         if (!pmd)
1421                 goto drop_hpage;
1422
1423         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1424
1425         /* step 1: check all mapped PTEs are to the right huge page */
1426         for (i = 0, addr = haddr, pte = start_pte;
1427              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1428                 struct page *page;
1429
1430                 /* empty pte, skip */
1431                 if (pte_none(*pte))
1432                         continue;
1433
1434                 /* page swapped out, abort */
1435                 if (!pte_present(*pte))
1436                         goto abort;
1437
1438                 page = vm_normal_page(vma, addr, *pte);
1439                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1440                         page = NULL;
1441                 /*
1442                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1443                  * page table, but the new page will not be a subpage of hpage.
1444                  */
1445                 if (hpage + i != page)
1446                         goto abort;
1447                 count++;
1448         }
1449
1450         /* step 2: adjust rmap */
1451         for (i = 0, addr = haddr, pte = start_pte;
1452              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1453                 struct page *page;
1454
1455                 if (pte_none(*pte))
1456                         continue;
1457                 page = vm_normal_page(vma, addr, *pte);
1458                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1459                         goto abort;
1460                 page_remove_rmap(page, vma, false);
1461         }
1462
1463         pte_unmap_unlock(start_pte, ptl);
1464
1465         /* step 3: set proper refcount and mm_counters. */
1466         if (count) {
1467                 page_ref_sub(hpage, count);
1468                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1469         }
1470
1471         /* step 4: collapse pmd */
1472         collapse_and_free_pmd(mm, vma, haddr, pmd);
1473 drop_hpage:
1474         unlock_page(hpage);
1475         put_page(hpage);
1476         return;
1477
1478 abort:
1479         pte_unmap_unlock(start_pte, ptl);
1480         goto drop_hpage;
1481 }
1482
1483 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1484 {
1485         struct mm_struct *mm = mm_slot->mm;
1486         int i;
1487
1488         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1489                 return;
1490
1491         if (!mmap_write_trylock(mm))
1492                 return;
1493
1494         if (unlikely(khugepaged_test_exit(mm)))
1495                 goto out;
1496
1497         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1498                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1499
1500 out:
1501         mm_slot->nr_pte_mapped_thp = 0;
1502         mmap_write_unlock(mm);
1503 }
1504
1505 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1506 {
1507         struct vm_area_struct *vma;
1508         struct mm_struct *mm;
1509         unsigned long addr;
1510         pmd_t *pmd;
1511
1512         i_mmap_lock_write(mapping);
1513         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1514                 /*
1515                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1516                  * got written to. These VMAs are likely not worth investing
1517                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1518                  * later.
1519                  *
1520                  * Note that vma->anon_vma check is racy: it can be set up after
1521                  * the check but before we took mmap_lock by the fault path.
1522                  * But page lock would prevent establishing any new ptes of the
1523                  * page, so we are safe.
1524                  *
1525                  * An alternative would be drop the check, but check that page
1526                  * table is clear before calling pmdp_collapse_flush() under
1527                  * ptl. It has higher chance to recover THP for the VMA, but
1528                  * has higher cost too.
1529                  */
1530                 if (vma->anon_vma)
1531                         continue;
1532                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1533                 if (addr & ~HPAGE_PMD_MASK)
1534                         continue;
1535                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1536                         continue;
1537                 mm = vma->vm_mm;
1538                 pmd = mm_find_pmd(mm, addr);
1539                 if (!pmd)
1540                         continue;
1541                 /*
1542                  * We need exclusive mmap_lock to retract page table.
1543                  *
1544                  * We use trylock due to lock inversion: we need to acquire
1545                  * mmap_lock while holding page lock. Fault path does it in
1546                  * reverse order. Trylock is a way to avoid deadlock.
1547                  */
1548                 if (mmap_write_trylock(mm)) {
1549                         /*
1550                          * When a vma is registered with uffd-wp, we can't
1551                          * recycle the pmd pgtable because there can be pte
1552                          * markers installed.  Skip it only, so the rest mm/vma
1553                          * can still have the same file mapped hugely, however
1554                          * it'll always mapped in small page size for uffd-wp
1555                          * registered ranges.
1556                          */
1557                         if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma))
1558                                 collapse_and_free_pmd(mm, vma, addr, pmd);
1559                         mmap_write_unlock(mm);
1560                 } else {
1561                         /* Try again later */
1562                         khugepaged_add_pte_mapped_thp(mm, addr);
1563                 }
1564         }
1565         i_mmap_unlock_write(mapping);
1566 }
1567
1568 /**
1569  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1570  *
1571  * @mm: process address space where collapse happens
1572  * @file: file that collapse on
1573  * @start: collapse start address
1574  * @hpage: new allocated huge page for collapse
1575  * @node: appointed node the new huge page allocate from
1576  *
1577  * Basic scheme is simple, details are more complex:
1578  *  - allocate and lock a new huge page;
1579  *  - scan page cache replacing old pages with the new one
1580  *    + swap/gup in pages if necessary;
1581  *    + fill in gaps;
1582  *    + keep old pages around in case rollback is required;
1583  *  - if replacing succeeds:
1584  *    + copy data over;
1585  *    + free old pages;
1586  *    + unlock huge page;
1587  *  - if replacing failed;
1588  *    + put all pages back and unfreeze them;
1589  *    + restore gaps in the page cache;
1590  *    + unlock and free huge page;
1591  */
1592 static void collapse_file(struct mm_struct *mm,
1593                 struct file *file, pgoff_t start,
1594                 struct page **hpage, int node)
1595 {
1596         struct address_space *mapping = file->f_mapping;
1597         gfp_t gfp;
1598         struct page *new_page;
1599         pgoff_t index, end = start + HPAGE_PMD_NR;
1600         LIST_HEAD(pagelist);
1601         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1602         int nr_none = 0, result = SCAN_SUCCEED;
1603         bool is_shmem = shmem_file(file);
1604         int nr;
1605
1606         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1607         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1608
1609         /* Only allocate from the target node */
1610         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1611
1612         new_page = khugepaged_alloc_page(hpage, gfp, node);
1613         if (!new_page) {
1614                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1615                 goto out;
1616         }
1617
1618         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1619                 result = SCAN_CGROUP_CHARGE_FAIL;
1620                 goto out;
1621         }
1622         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1623
1624         /*
1625          * Ensure we have slots for all the pages in the range.  This is
1626          * almost certainly a no-op because most of the pages must be present
1627          */
1628         do {
1629                 xas_lock_irq(&xas);
1630                 xas_create_range(&xas);
1631                 if (!xas_error(&xas))
1632                         break;
1633                 xas_unlock_irq(&xas);
1634                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1635                         result = SCAN_FAIL;
1636                         goto out;
1637                 }
1638         } while (1);
1639
1640         __SetPageLocked(new_page);
1641         if (is_shmem)
1642                 __SetPageSwapBacked(new_page);
1643         new_page->index = start;
1644         new_page->mapping = mapping;
1645
1646         /*
1647          * At this point the new_page is locked and not up-to-date.
1648          * It's safe to insert it into the page cache, because nobody would
1649          * be able to map it or use it in another way until we unlock it.
1650          */
1651
1652         xas_set(&xas, start);
1653         for (index = start; index < end; index++) {
1654                 struct page *page = xas_next(&xas);
1655
1656                 VM_BUG_ON(index != xas.xa_index);
1657                 if (is_shmem) {
1658                         if (!page) {
1659                                 /*
1660                                  * Stop if extent has been truncated or
1661                                  * hole-punched, and is now completely
1662                                  * empty.
1663                                  */
1664                                 if (index == start) {
1665                                         if (!xas_next_entry(&xas, end - 1)) {
1666                                                 result = SCAN_TRUNCATED;
1667                                                 goto xa_locked;
1668                                         }
1669                                         xas_set(&xas, index);
1670                                 }
1671                                 if (!shmem_charge(mapping->host, 1)) {
1672                                         result = SCAN_FAIL;
1673                                         goto xa_locked;
1674                                 }
1675                                 xas_store(&xas, new_page);
1676                                 nr_none++;
1677                                 continue;
1678                         }
1679
1680                         if (xa_is_value(page) || !PageUptodate(page)) {
1681                                 xas_unlock_irq(&xas);
1682                                 /* swap in or instantiate fallocated page */
1683                                 if (shmem_getpage(mapping->host, index, &page,
1684                                                   SGP_NOALLOC)) {
1685                                         result = SCAN_FAIL;
1686                                         goto xa_unlocked;
1687                                 }
1688                         } else if (trylock_page(page)) {
1689                                 get_page(page);
1690                                 xas_unlock_irq(&xas);
1691                         } else {
1692                                 result = SCAN_PAGE_LOCK;
1693                                 goto xa_locked;
1694                         }
1695                 } else {        /* !is_shmem */
1696                         if (!page || xa_is_value(page)) {
1697                                 xas_unlock_irq(&xas);
1698                                 page_cache_sync_readahead(mapping, &file->f_ra,
1699                                                           file, index,
1700                                                           end - index);
1701                                 /* drain pagevecs to help isolate_lru_page() */
1702                                 lru_add_drain();
1703                                 page = find_lock_page(mapping, index);
1704                                 if (unlikely(page == NULL)) {
1705                                         result = SCAN_FAIL;
1706                                         goto xa_unlocked;
1707                                 }
1708                         } else if (PageDirty(page)) {
1709                                 /*
1710                                  * khugepaged only works on read-only fd,
1711                                  * so this page is dirty because it hasn't
1712                                  * been flushed since first write. There
1713                                  * won't be new dirty pages.
1714                                  *
1715                                  * Trigger async flush here and hope the
1716                                  * writeback is done when khugepaged
1717                                  * revisits this page.
1718                                  *
1719                                  * This is a one-off situation. We are not
1720                                  * forcing writeback in loop.
1721                                  */
1722                                 xas_unlock_irq(&xas);
1723                                 filemap_flush(mapping);
1724                                 result = SCAN_FAIL;
1725                                 goto xa_unlocked;
1726                         } else if (PageWriteback(page)) {
1727                                 xas_unlock_irq(&xas);
1728                                 result = SCAN_FAIL;
1729                                 goto xa_unlocked;
1730                         } else if (trylock_page(page)) {
1731                                 get_page(page);
1732                                 xas_unlock_irq(&xas);
1733                         } else {
1734                                 result = SCAN_PAGE_LOCK;
1735                                 goto xa_locked;
1736                         }
1737                 }
1738
1739                 /*
1740                  * The page must be locked, so we can drop the i_pages lock
1741                  * without racing with truncate.
1742                  */
1743                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1744
1745                 /* make sure the page is up to date */
1746                 if (unlikely(!PageUptodate(page))) {
1747                         result = SCAN_FAIL;
1748                         goto out_unlock;
1749                 }
1750
1751                 /*
1752                  * If file was truncated then extended, or hole-punched, before
1753                  * we locked the first page, then a THP might be there already.
1754                  */
1755                 if (PageTransCompound(page)) {
1756                         result = SCAN_PAGE_COMPOUND;
1757                         goto out_unlock;
1758                 }
1759
1760                 if (page_mapping(page) != mapping) {
1761                         result = SCAN_TRUNCATED;
1762                         goto out_unlock;
1763                 }
1764
1765                 if (!is_shmem && (PageDirty(page) ||
1766                                   PageWriteback(page))) {
1767                         /*
1768                          * khugepaged only works on read-only fd, so this
1769                          * page is dirty because it hasn't been flushed
1770                          * since first write.
1771                          */
1772                         result = SCAN_FAIL;
1773                         goto out_unlock;
1774                 }
1775
1776                 if (isolate_lru_page(page)) {
1777                         result = SCAN_DEL_PAGE_LRU;
1778                         goto out_unlock;
1779                 }
1780
1781                 if (page_has_private(page) &&
1782                     !try_to_release_page(page, GFP_KERNEL)) {
1783                         result = SCAN_PAGE_HAS_PRIVATE;
1784                         putback_lru_page(page);
1785                         goto out_unlock;
1786                 }
1787
1788                 if (page_mapped(page))
1789                         try_to_unmap(page_folio(page),
1790                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1791
1792                 xas_lock_irq(&xas);
1793                 xas_set(&xas, index);
1794
1795                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1796
1797                 /*
1798                  * The page is expected to have page_count() == 3:
1799                  *  - we hold a pin on it;
1800                  *  - one reference from page cache;
1801                  *  - one from isolate_lru_page;
1802                  */
1803                 if (!page_ref_freeze(page, 3)) {
1804                         result = SCAN_PAGE_COUNT;
1805                         xas_unlock_irq(&xas);
1806                         putback_lru_page(page);
1807                         goto out_unlock;
1808                 }
1809
1810                 /*
1811                  * Add the page to the list to be able to undo the collapse if
1812                  * something go wrong.
1813                  */
1814                 list_add_tail(&page->lru, &pagelist);
1815
1816                 /* Finally, replace with the new page. */
1817                 xas_store(&xas, new_page);
1818                 continue;
1819 out_unlock:
1820                 unlock_page(page);
1821                 put_page(page);
1822                 goto xa_unlocked;
1823         }
1824         nr = thp_nr_pages(new_page);
1825
1826         if (is_shmem)
1827                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1828         else {
1829                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1830                 filemap_nr_thps_inc(mapping);
1831                 /*
1832                  * Paired with smp_mb() in do_dentry_open() to ensure
1833                  * i_writecount is up to date and the update to nr_thps is
1834                  * visible. Ensures the page cache will be truncated if the
1835                  * file is opened writable.
1836                  */
1837                 smp_mb();
1838                 if (inode_is_open_for_write(mapping->host)) {
1839                         result = SCAN_FAIL;
1840                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1841                         filemap_nr_thps_dec(mapping);
1842                         goto xa_locked;
1843                 }
1844         }
1845
1846         if (nr_none) {
1847                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1848                 /* nr_none is always 0 for non-shmem. */
1849                 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1850         }
1851
1852         /* Join all the small entries into a single multi-index entry */
1853         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1854         xas_store(&xas, new_page);
1855 xa_locked:
1856         xas_unlock_irq(&xas);
1857 xa_unlocked:
1858
1859         /*
1860          * If collapse is successful, flush must be done now before copying.
1861          * If collapse is unsuccessful, does flush actually need to be done?
1862          * Do it anyway, to clear the state.
1863          */
1864         try_to_unmap_flush();
1865
1866         if (result == SCAN_SUCCEED) {
1867                 struct page *page, *tmp;
1868
1869                 /*
1870                  * Replacing old pages with new one has succeeded, now we
1871                  * need to copy the content and free the old pages.
1872                  */
1873                 index = start;
1874                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1875                         while (index < page->index) {
1876                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1877                                 index++;
1878                         }
1879                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1880                                         page);
1881                         list_del(&page->lru);
1882                         page->mapping = NULL;
1883                         page_ref_unfreeze(page, 1);
1884                         ClearPageActive(page);
1885                         ClearPageUnevictable(page);
1886                         unlock_page(page);
1887                         put_page(page);
1888                         index++;
1889                 }
1890                 while (index < end) {
1891                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1892                         index++;
1893                 }
1894
1895                 SetPageUptodate(new_page);
1896                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1897                 if (is_shmem)
1898                         set_page_dirty(new_page);
1899                 lru_cache_add(new_page);
1900
1901                 /*
1902                  * Remove pte page tables, so we can re-fault the page as huge.
1903                  */
1904                 retract_page_tables(mapping, start);
1905                 *hpage = NULL;
1906
1907                 khugepaged_pages_collapsed++;
1908         } else {
1909                 struct page *page;
1910
1911                 /* Something went wrong: roll back page cache changes */
1912                 xas_lock_irq(&xas);
1913                 if (nr_none) {
1914                         mapping->nrpages -= nr_none;
1915                         shmem_uncharge(mapping->host, nr_none);
1916                 }
1917
1918                 xas_set(&xas, start);
1919                 xas_for_each(&xas, page, end - 1) {
1920                         page = list_first_entry_or_null(&pagelist,
1921                                         struct page, lru);
1922                         if (!page || xas.xa_index < page->index) {
1923                                 if (!nr_none)
1924                                         break;
1925                                 nr_none--;
1926                                 /* Put holes back where they were */
1927                                 xas_store(&xas, NULL);
1928                                 continue;
1929                         }
1930
1931                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1932
1933                         /* Unfreeze the page. */
1934                         list_del(&page->lru);
1935                         page_ref_unfreeze(page, 2);
1936                         xas_store(&xas, page);
1937                         xas_pause(&xas);
1938                         xas_unlock_irq(&xas);
1939                         unlock_page(page);
1940                         putback_lru_page(page);
1941                         xas_lock_irq(&xas);
1942                 }
1943                 VM_BUG_ON(nr_none);
1944                 xas_unlock_irq(&xas);
1945
1946                 new_page->mapping = NULL;
1947         }
1948
1949         unlock_page(new_page);
1950 out:
1951         VM_BUG_ON(!list_empty(&pagelist));
1952         if (!IS_ERR_OR_NULL(*hpage))
1953                 mem_cgroup_uncharge(page_folio(*hpage));
1954         /* TODO: tracepoints */
1955 }
1956
1957 static void khugepaged_scan_file(struct mm_struct *mm,
1958                 struct file *file, pgoff_t start, struct page **hpage)
1959 {
1960         struct page *page = NULL;
1961         struct address_space *mapping = file->f_mapping;
1962         XA_STATE(xas, &mapping->i_pages, start);
1963         int present, swap;
1964         int node = NUMA_NO_NODE;
1965         int result = SCAN_SUCCEED;
1966
1967         present = 0;
1968         swap = 0;
1969         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1970         rcu_read_lock();
1971         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1972                 if (xas_retry(&xas, page))
1973                         continue;
1974
1975                 if (xa_is_value(page)) {
1976                         if (++swap > khugepaged_max_ptes_swap) {
1977                                 result = SCAN_EXCEED_SWAP_PTE;
1978                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1979                                 break;
1980                         }
1981                         continue;
1982                 }
1983
1984                 /*
1985                  * XXX: khugepaged should compact smaller compound pages
1986                  * into a PMD sized page
1987                  */
1988                 if (PageTransCompound(page)) {
1989                         result = SCAN_PAGE_COMPOUND;
1990                         break;
1991                 }
1992
1993                 node = page_to_nid(page);
1994                 if (khugepaged_scan_abort(node)) {
1995                         result = SCAN_SCAN_ABORT;
1996                         break;
1997                 }
1998                 khugepaged_node_load[node]++;
1999
2000                 if (!PageLRU(page)) {
2001                         result = SCAN_PAGE_LRU;
2002                         break;
2003                 }
2004
2005                 if (page_count(page) !=
2006                     1 + page_mapcount(page) + page_has_private(page)) {
2007                         result = SCAN_PAGE_COUNT;
2008                         break;
2009                 }
2010
2011                 /*
2012                  * We probably should check if the page is referenced here, but
2013                  * nobody would transfer pte_young() to PageReferenced() for us.
2014                  * And rmap walk here is just too costly...
2015                  */
2016
2017                 present++;
2018
2019                 if (need_resched()) {
2020                         xas_pause(&xas);
2021                         cond_resched_rcu();
2022                 }
2023         }
2024         rcu_read_unlock();
2025
2026         if (result == SCAN_SUCCEED) {
2027                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2028                         result = SCAN_EXCEED_NONE_PTE;
2029                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2030                 } else {
2031                         node = khugepaged_find_target_node();
2032                         collapse_file(mm, file, start, hpage, node);
2033                 }
2034         }
2035
2036         /* TODO: tracepoints */
2037 }
2038 #else
2039 static void khugepaged_scan_file(struct mm_struct *mm,
2040                 struct file *file, pgoff_t start, struct page **hpage)
2041 {
2042         BUILD_BUG();
2043 }
2044
2045 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2046 {
2047 }
2048 #endif
2049
2050 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2051                                             struct page **hpage)
2052         __releases(&khugepaged_mm_lock)
2053         __acquires(&khugepaged_mm_lock)
2054 {
2055         struct mm_slot *mm_slot;
2056         struct mm_struct *mm;
2057         struct vm_area_struct *vma;
2058         int progress = 0;
2059
2060         VM_BUG_ON(!pages);
2061         lockdep_assert_held(&khugepaged_mm_lock);
2062
2063         if (khugepaged_scan.mm_slot)
2064                 mm_slot = khugepaged_scan.mm_slot;
2065         else {
2066                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2067                                      struct mm_slot, mm_node);
2068                 khugepaged_scan.address = 0;
2069                 khugepaged_scan.mm_slot = mm_slot;
2070         }
2071         spin_unlock(&khugepaged_mm_lock);
2072         khugepaged_collapse_pte_mapped_thps(mm_slot);
2073
2074         mm = mm_slot->mm;
2075         /*
2076          * Don't wait for semaphore (to avoid long wait times).  Just move to
2077          * the next mm on the list.
2078          */
2079         vma = NULL;
2080         if (unlikely(!mmap_read_trylock(mm)))
2081                 goto breakouterloop_mmap_lock;
2082         if (likely(!khugepaged_test_exit(mm)))
2083                 vma = find_vma(mm, khugepaged_scan.address);
2084
2085         progress++;
2086         for (; vma; vma = vma->vm_next) {
2087                 unsigned long hstart, hend;
2088
2089                 cond_resched();
2090                 if (unlikely(khugepaged_test_exit(mm))) {
2091                         progress++;
2092                         break;
2093                 }
2094                 if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) {
2095 skip:
2096                         progress++;
2097                         continue;
2098                 }
2099                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2100                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2101                 if (khugepaged_scan.address > hend)
2102                         goto skip;
2103                 if (khugepaged_scan.address < hstart)
2104                         khugepaged_scan.address = hstart;
2105                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2106
2107                 while (khugepaged_scan.address < hend) {
2108                         int ret;
2109                         cond_resched();
2110                         if (unlikely(khugepaged_test_exit(mm)))
2111                                 goto breakouterloop;
2112
2113                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2114                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2115                                   hend);
2116                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2117                                 struct file *file = get_file(vma->vm_file);
2118                                 pgoff_t pgoff = linear_page_index(vma,
2119                                                 khugepaged_scan.address);
2120
2121                                 mmap_read_unlock(mm);
2122                                 ret = 1;
2123                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2124                                 fput(file);
2125                         } else {
2126                                 ret = khugepaged_scan_pmd(mm, vma,
2127                                                 khugepaged_scan.address,
2128                                                 hpage);
2129                         }
2130                         /* move to next address */
2131                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2132                         progress += HPAGE_PMD_NR;
2133                         if (ret)
2134                                 /* we released mmap_lock so break loop */
2135                                 goto breakouterloop_mmap_lock;
2136                         if (progress >= pages)
2137                                 goto breakouterloop;
2138                 }
2139         }
2140 breakouterloop:
2141         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2142 breakouterloop_mmap_lock:
2143
2144         spin_lock(&khugepaged_mm_lock);
2145         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2146         /*
2147          * Release the current mm_slot if this mm is about to die, or
2148          * if we scanned all vmas of this mm.
2149          */
2150         if (khugepaged_test_exit(mm) || !vma) {
2151                 /*
2152                  * Make sure that if mm_users is reaching zero while
2153                  * khugepaged runs here, khugepaged_exit will find
2154                  * mm_slot not pointing to the exiting mm.
2155                  */
2156                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2157                         khugepaged_scan.mm_slot = list_entry(
2158                                 mm_slot->mm_node.next,
2159                                 struct mm_slot, mm_node);
2160                         khugepaged_scan.address = 0;
2161                 } else {
2162                         khugepaged_scan.mm_slot = NULL;
2163                         khugepaged_full_scans++;
2164                 }
2165
2166                 collect_mm_slot(mm_slot);
2167         }
2168
2169         return progress;
2170 }
2171
2172 static int khugepaged_has_work(void)
2173 {
2174         return !list_empty(&khugepaged_scan.mm_head) &&
2175                 khugepaged_enabled();
2176 }
2177
2178 static int khugepaged_wait_event(void)
2179 {
2180         return !list_empty(&khugepaged_scan.mm_head) ||
2181                 kthread_should_stop();
2182 }
2183
2184 static void khugepaged_do_scan(void)
2185 {
2186         struct page *hpage = NULL;
2187         unsigned int progress = 0, pass_through_head = 0;
2188         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2189         bool wait = true;
2190
2191         lru_add_drain_all();
2192
2193         while (progress < pages) {
2194                 if (!khugepaged_prealloc_page(&hpage, &wait))
2195                         break;
2196
2197                 cond_resched();
2198
2199                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2200                         break;
2201
2202                 spin_lock(&khugepaged_mm_lock);
2203                 if (!khugepaged_scan.mm_slot)
2204                         pass_through_head++;
2205                 if (khugepaged_has_work() &&
2206                     pass_through_head < 2)
2207                         progress += khugepaged_scan_mm_slot(pages - progress,
2208                                                             &hpage);
2209                 else
2210                         progress = pages;
2211                 spin_unlock(&khugepaged_mm_lock);
2212         }
2213
2214         if (!IS_ERR_OR_NULL(hpage))
2215                 put_page(hpage);
2216 }
2217
2218 static bool khugepaged_should_wakeup(void)
2219 {
2220         return kthread_should_stop() ||
2221                time_after_eq(jiffies, khugepaged_sleep_expire);
2222 }
2223
2224 static void khugepaged_wait_work(void)
2225 {
2226         if (khugepaged_has_work()) {
2227                 const unsigned long scan_sleep_jiffies =
2228                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2229
2230                 if (!scan_sleep_jiffies)
2231                         return;
2232
2233                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2234                 wait_event_freezable_timeout(khugepaged_wait,
2235                                              khugepaged_should_wakeup(),
2236                                              scan_sleep_jiffies);
2237                 return;
2238         }
2239
2240         if (khugepaged_enabled())
2241                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2242 }
2243
2244 static int khugepaged(void *none)
2245 {
2246         struct mm_slot *mm_slot;
2247
2248         set_freezable();
2249         set_user_nice(current, MAX_NICE);
2250
2251         while (!kthread_should_stop()) {
2252                 khugepaged_do_scan();
2253                 khugepaged_wait_work();
2254         }
2255
2256         spin_lock(&khugepaged_mm_lock);
2257         mm_slot = khugepaged_scan.mm_slot;
2258         khugepaged_scan.mm_slot = NULL;
2259         if (mm_slot)
2260                 collect_mm_slot(mm_slot);
2261         spin_unlock(&khugepaged_mm_lock);
2262         return 0;
2263 }
2264
2265 static void set_recommended_min_free_kbytes(void)
2266 {
2267         struct zone *zone;
2268         int nr_zones = 0;
2269         unsigned long recommended_min;
2270
2271         if (!khugepaged_enabled()) {
2272                 calculate_min_free_kbytes();
2273                 goto update_wmarks;
2274         }
2275
2276         for_each_populated_zone(zone) {
2277                 /*
2278                  * We don't need to worry about fragmentation of
2279                  * ZONE_MOVABLE since it only has movable pages.
2280                  */
2281                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2282                         continue;
2283
2284                 nr_zones++;
2285         }
2286
2287         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2288         recommended_min = pageblock_nr_pages * nr_zones * 2;
2289
2290         /*
2291          * Make sure that on average at least two pageblocks are almost free
2292          * of another type, one for a migratetype to fall back to and a
2293          * second to avoid subsequent fallbacks of other types There are 3
2294          * MIGRATE_TYPES we care about.
2295          */
2296         recommended_min += pageblock_nr_pages * nr_zones *
2297                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2298
2299         /* don't ever allow to reserve more than 5% of the lowmem */
2300         recommended_min = min(recommended_min,
2301                               (unsigned long) nr_free_buffer_pages() / 20);
2302         recommended_min <<= (PAGE_SHIFT-10);
2303
2304         if (recommended_min > min_free_kbytes) {
2305                 if (user_min_free_kbytes >= 0)
2306                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2307                                 min_free_kbytes, recommended_min);
2308
2309                 min_free_kbytes = recommended_min;
2310         }
2311
2312 update_wmarks:
2313         setup_per_zone_wmarks();
2314 }
2315
2316 int start_stop_khugepaged(void)
2317 {
2318         int err = 0;
2319
2320         mutex_lock(&khugepaged_mutex);
2321         if (khugepaged_enabled()) {
2322                 if (!khugepaged_thread)
2323                         khugepaged_thread = kthread_run(khugepaged, NULL,
2324                                                         "khugepaged");
2325                 if (IS_ERR(khugepaged_thread)) {
2326                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2327                         err = PTR_ERR(khugepaged_thread);
2328                         khugepaged_thread = NULL;
2329                         goto fail;
2330                 }
2331
2332                 if (!list_empty(&khugepaged_scan.mm_head))
2333                         wake_up_interruptible(&khugepaged_wait);
2334         } else if (khugepaged_thread) {
2335                 kthread_stop(khugepaged_thread);
2336                 khugepaged_thread = NULL;
2337         }
2338         set_recommended_min_free_kbytes();
2339 fail:
2340         mutex_unlock(&khugepaged_mutex);
2341         return err;
2342 }
2343
2344 void khugepaged_min_free_kbytes_update(void)
2345 {
2346         mutex_lock(&khugepaged_mutex);
2347         if (khugepaged_enabled() && khugepaged_thread)
2348                 set_recommended_min_free_kbytes();
2349         mutex_unlock(&khugepaged_mutex);
2350 }