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             hugepage_flags_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 #define khugepaged_defrag()                                     \
 765         (transparent_hugepage_flags &                           \
 766          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 767
 768 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 769 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 770 {
 771         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 772 }
 773
 774 #ifdef CONFIG_NUMA
 775 static int khugepaged_find_target_node(void)
 776 {
 777         static int last_khugepaged_target_node = NUMA_NO_NODE;
 778         int nid, target_node = 0, max_value = 0;
 779
 780         /* find first node with max normal pages hit */
 781         for (nid = 0; nid < MAX_NUMNODES; nid++)
 782                 if (khugepaged_node_load[nid] > max_value) {
 783                         max_value = khugepaged_node_load[nid];
 784                         target_node = nid;
 785                 }
 786
 787         /* do some balance if several nodes have the same hit record */
 788         if (target_node <= last_khugepaged_target_node)
 789                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 790                                 nid++)
 791                         if (max_value == khugepaged_node_load[nid]) {
 792                                 target_node = nid;
 793                                 break;
 794                         }
 795
 796         last_khugepaged_target_node = target_node;
 797         return target_node;
 798 }
 799
 800 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 801 {
 802         if (IS_ERR(*hpage)) {
 803                 if (!*wait)
 804                         return false;
 805
 806                 *wait = false;
 807                 *hpage = NULL;
 808                 khugepaged_alloc_sleep();
 809         } else if (*hpage) {
 810                 put_page(*hpage);
 811                 *hpage = NULL;
 812         }
 813
 814         return true;
 815 }
 816
 817 static struct page *
 818 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 819 {
 820         VM_BUG_ON_PAGE(*hpage, *hpage);
 821
 822         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 823         if (unlikely(!*hpage)) {
 824                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 825                 *hpage = ERR_PTR(-ENOMEM);
 826                 return NULL;
 827         }
 828
 829         prep_transhuge_page(*hpage);
 830         count_vm_event(THP_COLLAPSE_ALLOC);
 831         return *hpage;
 832 }
 833 #else
 834 static int khugepaged_find_target_node(void)
 835 {
 836         return 0;
 837 }
 838
 839 static inline struct page *alloc_khugepaged_hugepage(void)
 840 {
 841         struct page *page;
 842
 843         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 844                            HPAGE_PMD_ORDER);
 845         if (page)
 846                 prep_transhuge_page(page);
 847         return page;
 848 }
 849
 850 static struct page *khugepaged_alloc_hugepage(bool *wait)
 851 {
 852         struct page *hpage;
 853
 854         do {
 855                 hpage = alloc_khugepaged_hugepage();
 856                 if (!hpage) {
 857                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 858                         if (!*wait)
 859                                 return NULL;
 860
 861                         *wait = false;
 862                         khugepaged_alloc_sleep();
 863                 } else
 864                         count_vm_event(THP_COLLAPSE_ALLOC);
 865         } while (unlikely(!hpage) && likely(hugepage_flags_enabled()));
 866
 867         return hpage;
 868 }
 869
 870 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 871 {
 872         /*
 873          * If the hpage allocated earlier was briefly exposed in page cache
 874          * before collapse_file() failed, it is possible that racing lookups
 875          * have not yet completed, and would then be unpleasantly surprised by
 876          * finding the hpage reused for the same mapping at a different offset.
 877          * Just release the previous allocation if there is any danger of that.
 878          */
 879         if (*hpage && page_count(*hpage) > 1) {
 880                 put_page(*hpage);
 881                 *hpage = NULL;
 882         }
 883
 884         if (!*hpage)
 885                 *hpage = khugepaged_alloc_hugepage(wait);
 886
 887         if (unlikely(!*hpage))
 888                 return false;
 889
 890         return true;
 891 }
 892
 893 static struct page *
 894 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 895 {
 896         VM_BUG_ON(!*hpage);
 897
 898         return  *hpage;
 899 }
 900 #endif
 901
 902 /*
 903  * If mmap_lock temporarily dropped, revalidate vma
 904  * before taking mmap_lock.
 905  * Return 0 if succeeds, otherwise return none-zero
 906  * value (scan code).
 907  */
 908
 909 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 910                 struct vm_area_struct **vmap)
 911 {
 912         struct vm_area_struct *vma;
 913
 914         if (unlikely(khugepaged_test_exit(mm)))
 915                 return SCAN_ANY_PROCESS;
 916
 917         *vmap = vma = find_vma(mm, address);
 918         if (!vma)
 919                 return SCAN_VMA_NULL;
 920
 921         if (!transhuge_vma_suitable(vma, address))
 922                 return SCAN_ADDRESS_RANGE;
 923         if (!hugepage_vma_check(vma, vma->vm_flags, false, false))
 924                 return SCAN_VMA_CHECK;
 925         /*
 926          * Anon VMA expected, the address may be unmapped then
 927          * remapped to file after khugepaged reaquired the mmap_lock.
 928          *
 929          * hugepage_vma_check may return true for qualified file
 930          * vmas.
 931          */
 932         if (!vma->anon_vma || !vma_is_anonymous(vma))
 933                 return SCAN_VMA_CHECK;
 934         return 0;
 935 }
 936
 937 /*
 938  * Bring missing pages in from swap, to complete THP collapse.
 939  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 940  *
 941  * Called and returns without pte mapped or spinlocks held.
 942  * Note that if false is returned, mmap_lock will be released.
 943  */
 944
 945 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 946                                         struct vm_area_struct *vma,
 947                                         unsigned long haddr, pmd_t *pmd,
 948                                         int referenced)
 949 {
 950         int swapped_in = 0;
 951         vm_fault_t ret = 0;
 952         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 953
 954         for (address = haddr; address < end; address += PAGE_SIZE) {
 955                 struct vm_fault vmf = {
 956                         .vma = vma,
 957                         .address = address,
 958                         .pgoff = linear_page_index(vma, haddr),
 959                         .flags = FAULT_FLAG_ALLOW_RETRY,
 960                         .pmd = pmd,
 961                 };
 962
 963                 vmf.pte = pte_offset_map(pmd, address);
 964                 vmf.orig_pte = *vmf.pte;
 965                 if (!is_swap_pte(vmf.orig_pte)) {
 966                         pte_unmap(vmf.pte);
 967                         continue;
 968                 }
 969                 ret = do_swap_page(&vmf);
 970
 971                 /*
 972                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
 973                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
 974                  * we do not retry here and swap entry will remain in pagetable
 975                  * resulting in later failure.
 976                  */
 977                 if (ret & VM_FAULT_RETRY) {
 978                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 979                         return false;
 980                 }
 981                 if (ret & VM_FAULT_ERROR) {
 982                         mmap_read_unlock(mm);
 983                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 984                         return false;
 985                 }
 986                 swapped_in++;
 987         }
 988
 989         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
 990         if (swapped_in)
 991                 lru_add_drain();
 992
 993         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 994         return true;
 995 }
 996
 997 static void collapse_huge_page(struct mm_struct *mm,
 998                                    unsigned long address,
 999                                    struct page **hpage,
1000                                    int node, int referenced, int unmapped)
1001 {
1002         LIST_HEAD(compound_pagelist);
1003         pmd_t *pmd, _pmd;
1004         pte_t *pte;
1005         pgtable_t pgtable;
1006         struct page *new_page;
1007         spinlock_t *pmd_ptl, *pte_ptl;
1008         int isolated = 0, result = 0;
1009         struct vm_area_struct *vma;
1010         struct mmu_notifier_range range;
1011         gfp_t gfp;
1012
1013         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1014
1015         /* Only allocate from the target node */
1016         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1017
1018         /*
1019          * Before allocating the hugepage, release the mmap_lock read lock.
1020          * The allocation can take potentially a long time if it involves
1021          * sync compaction, and we do not need to hold the mmap_lock during
1022          * that. We will recheck the vma after taking it again in write mode.
1023          */
1024         mmap_read_unlock(mm);
1025         new_page = khugepaged_alloc_page(hpage, gfp, node);
1026         if (!new_page) {
1027                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1028                 goto out_nolock;
1029         }
1030
1031         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1032                 result = SCAN_CGROUP_CHARGE_FAIL;
1033                 goto out_nolock;
1034         }
1035         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1036
1037         mmap_read_lock(mm);
1038         result = hugepage_vma_revalidate(mm, address, &vma);
1039         if (result) {
1040                 mmap_read_unlock(mm);
1041                 goto out_nolock;
1042         }
1043
1044         pmd = mm_find_pmd(mm, address);
1045         if (!pmd) {
1046                 result = SCAN_PMD_NULL;
1047                 mmap_read_unlock(mm);
1048                 goto out_nolock;
1049         }
1050
1051         /*
1052          * __collapse_huge_page_swapin will return with mmap_lock released
1053          * when it fails. So we jump out_nolock directly in that case.
1054          * Continuing to collapse causes inconsistency.
1055          */
1056         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1057                                                      pmd, referenced)) {
1058                 goto out_nolock;
1059         }
1060
1061         mmap_read_unlock(mm);
1062         /*
1063          * Prevent all access to pagetables with the exception of
1064          * gup_fast later handled by the ptep_clear_flush and the VM
1065          * handled by the anon_vma lock + PG_lock.
1066          */
1067         mmap_write_lock(mm);
1068         result = hugepage_vma_revalidate(mm, address, &vma);
1069         if (result)
1070                 goto out_up_write;
1071         /* check if the pmd is still valid */
1072         if (mm_find_pmd(mm, address) != pmd)
1073                 goto out_up_write;
1074
1075         anon_vma_lock_write(vma->anon_vma);
1076
1077         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1078                                 address, address + HPAGE_PMD_SIZE);
1079         mmu_notifier_invalidate_range_start(&range);
1080
1081         pte = pte_offset_map(pmd, address);
1082         pte_ptl = pte_lockptr(mm, pmd);
1083
1084         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1085         /*
1086          * After this gup_fast can't run anymore. This also removes
1087          * any huge TLB entry from the CPU so we won't allow
1088          * huge and small TLB entries for the same virtual address
1089          * to avoid the risk of CPU bugs in that area.
1090          */
1091         _pmd = pmdp_collapse_flush(vma, address, pmd);
1092         spin_unlock(pmd_ptl);
1093         mmu_notifier_invalidate_range_end(&range);
1094
1095         spin_lock(pte_ptl);
1096         isolated = __collapse_huge_page_isolate(vma, address, pte,
1097                         &compound_pagelist);
1098         spin_unlock(pte_ptl);
1099
1100         if (unlikely(!isolated)) {
1101                 pte_unmap(pte);
1102                 spin_lock(pmd_ptl);
1103                 BUG_ON(!pmd_none(*pmd));
1104                 /*
1105                  * We can only use set_pmd_at when establishing
1106                  * hugepmds and never for establishing regular pmds that
1107                  * points to regular pagetables. Use pmd_populate for that
1108                  */
1109                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1110                 spin_unlock(pmd_ptl);
1111                 anon_vma_unlock_write(vma->anon_vma);
1112                 result = SCAN_FAIL;
1113                 goto out_up_write;
1114         }
1115
1116         /*
1117          * All pages are isolated and locked so anon_vma rmap
1118          * can't run anymore.
1119          */
1120         anon_vma_unlock_write(vma->anon_vma);
1121
1122         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1123                         &compound_pagelist);
1124         pte_unmap(pte);
1125         /*
1126          * spin_lock() below is not the equivalent of smp_wmb(), but
1127          * the smp_wmb() inside __SetPageUptodate() can be reused to
1128          * avoid the copy_huge_page writes to become visible after
1129          * the set_pmd_at() write.
1130          */
1131         __SetPageUptodate(new_page);
1132         pgtable = pmd_pgtable(_pmd);
1133
1134         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1135         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1136
1137         spin_lock(pmd_ptl);
1138         BUG_ON(!pmd_none(*pmd));
1139         page_add_new_anon_rmap(new_page, vma, address);
1140         lru_cache_add_inactive_or_unevictable(new_page, vma);
1141         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1142         set_pmd_at(mm, address, pmd, _pmd);
1143         update_mmu_cache_pmd(vma, address, pmd);
1144         spin_unlock(pmd_ptl);
1145
1146         *hpage = NULL;
1147
1148         khugepaged_pages_collapsed++;
1149         result = SCAN_SUCCEED;
1150 out_up_write:
1151         mmap_write_unlock(mm);
1152 out_nolock:
1153         if (!IS_ERR_OR_NULL(*hpage))
1154                 mem_cgroup_uncharge(page_folio(*hpage));
1155         trace_mm_collapse_huge_page(mm, isolated, result);
1156         return;
1157 }
1158
1159 static int khugepaged_scan_pmd(struct mm_struct *mm,
1160                                struct vm_area_struct *vma,
1161                                unsigned long address,
1162                                struct page **hpage)
1163 {
1164         pmd_t *pmd;
1165         pte_t *pte, *_pte;
1166         int ret = 0, result = 0, referenced = 0;
1167         int none_or_zero = 0, shared = 0;
1168         struct page *page = NULL;
1169         unsigned long _address;
1170         spinlock_t *ptl;
1171         int node = NUMA_NO_NODE, unmapped = 0;
1172         bool writable = false;
1173
1174         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1175
1176         pmd = mm_find_pmd(mm, address);
1177         if (!pmd) {
1178                 result = SCAN_PMD_NULL;
1179                 goto out;
1180         }
1181
1182         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1183         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1184         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1185              _pte++, _address += PAGE_SIZE) {
1186                 pte_t pteval = *_pte;
1187                 if (is_swap_pte(pteval)) {
1188                         if (++unmapped <= khugepaged_max_ptes_swap) {
1189                                 /*
1190                                  * Always be strict with uffd-wp
1191                                  * enabled swap entries.  Please see
1192                                  * comment below for pte_uffd_wp().
1193                                  */
1194                                 if (pte_swp_uffd_wp(pteval)) {
1195                                         result = SCAN_PTE_UFFD_WP;
1196                                         goto out_unmap;
1197                                 }
1198                                 continue;
1199                         } else {
1200                                 result = SCAN_EXCEED_SWAP_PTE;
1201                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1202                                 goto out_unmap;
1203                         }
1204                 }
1205                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1206                         if (!userfaultfd_armed(vma) &&
1207                             ++none_or_zero <= khugepaged_max_ptes_none) {
1208                                 continue;
1209                         } else {
1210                                 result = SCAN_EXCEED_NONE_PTE;
1211                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1212                                 goto out_unmap;
1213                         }
1214                 }
1215                 if (pte_uffd_wp(pteval)) {
1216                         /*
1217                          * Don't collapse the page if any of the small
1218                          * PTEs are armed with uffd write protection.
1219                          * Here we can also mark the new huge pmd as
1220                          * write protected if any of the small ones is
1221                          * marked but that could bring unknown
1222                          * userfault messages that falls outside of
1223                          * the registered range.  So, just be simple.
1224                          */
1225                         result = SCAN_PTE_UFFD_WP;
1226                         goto out_unmap;
1227                 }
1228                 if (pte_write(pteval))
1229                         writable = true;
1230
1231                 page = vm_normal_page(vma, _address, pteval);
1232                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1233                         result = SCAN_PAGE_NULL;
1234                         goto out_unmap;
1235                 }
1236
1237                 if (page_mapcount(page) > 1 &&
1238                                 ++shared > khugepaged_max_ptes_shared) {
1239                         result = SCAN_EXCEED_SHARED_PTE;
1240                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1241                         goto out_unmap;
1242                 }
1243
1244                 page = compound_head(page);
1245
1246                 /*
1247                  * Record which node the original page is from and save this
1248                  * information to khugepaged_node_load[].
1249                  * Khugepaged will allocate hugepage from the node has the max
1250                  * hit record.
1251                  */
1252                 node = page_to_nid(page);
1253                 if (khugepaged_scan_abort(node)) {
1254                         result = SCAN_SCAN_ABORT;
1255                         goto out_unmap;
1256                 }
1257                 khugepaged_node_load[node]++;
1258                 if (!PageLRU(page)) {
1259                         result = SCAN_PAGE_LRU;
1260                         goto out_unmap;
1261                 }
1262                 if (PageLocked(page)) {
1263                         result = SCAN_PAGE_LOCK;
1264                         goto out_unmap;
1265                 }
1266                 if (!PageAnon(page)) {
1267                         result = SCAN_PAGE_ANON;
1268                         goto out_unmap;
1269                 }
1270
1271                 /*
1272                  * Check if the page has any GUP (or other external) pins.
1273                  *
1274                  * Here the check is racy it may see total_mapcount > refcount
1275                  * in some cases.
1276                  * For example, one process with one forked child process.
1277                  * The parent has the PMD split due to MADV_DONTNEED, then
1278                  * the child is trying unmap the whole PMD, but khugepaged
1279                  * may be scanning the parent between the child has
1280                  * PageDoubleMap flag cleared and dec the mapcount.  So
1281                  * khugepaged may see total_mapcount > refcount.
1282                  *
1283                  * But such case is ephemeral we could always retry collapse
1284                  * later.  However it may report false positive if the page
1285                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1286                  * will be done again later the risk seems low.
1287                  */
1288                 if (!is_refcount_suitable(page)) {
1289                         result = SCAN_PAGE_COUNT;
1290                         goto out_unmap;
1291                 }
1292                 if (pte_young(pteval) ||
1293                     page_is_young(page) || PageReferenced(page) ||
1294                     mmu_notifier_test_young(vma->vm_mm, address))
1295                         referenced++;
1296         }
1297         if (!writable) {
1298                 result = SCAN_PAGE_RO;
1299         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1300                 result = SCAN_LACK_REFERENCED_PAGE;
1301         } else {
1302                 result = SCAN_SUCCEED;
1303                 ret = 1;
1304         }
1305 out_unmap:
1306         pte_unmap_unlock(pte, ptl);
1307         if (ret) {
1308                 node = khugepaged_find_target_node();
1309                 /* collapse_huge_page will return with the mmap_lock released */
1310                 collapse_huge_page(mm, address, hpage, node,
1311                                 referenced, unmapped);
1312         }
1313 out:
1314         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1315                                      none_or_zero, result, unmapped);
1316         return ret;
1317 }
1318
1319 static void collect_mm_slot(struct mm_slot *mm_slot)
1320 {
1321         struct mm_struct *mm = mm_slot->mm;
1322
1323         lockdep_assert_held(&khugepaged_mm_lock);
1324
1325         if (khugepaged_test_exit(mm)) {
1326                 /* free mm_slot */
1327                 hash_del(&mm_slot->hash);
1328                 list_del(&mm_slot->mm_node);
1329
1330                 /*
1331                  * Not strictly needed because the mm exited already.
1332                  *
1333                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1334                  */
1335
1336                 /* khugepaged_mm_lock actually not necessary for the below */
1337                 free_mm_slot(mm_slot);
1338                 mmdrop(mm);
1339         }
1340 }
1341
1342 #ifdef CONFIG_SHMEM
1343 /*
1344  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1345  * khugepaged should try to collapse the page table.
1346  */
1347 static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1348                                           unsigned long addr)
1349 {
1350         struct mm_slot *mm_slot;
1351
1352         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1353
1354         spin_lock(&khugepaged_mm_lock);
1355         mm_slot = get_mm_slot(mm);
1356         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1357                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1358         spin_unlock(&khugepaged_mm_lock);
1359 }
1360
1361 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1362                                   unsigned long addr, pmd_t *pmdp)
1363 {
1364         spinlock_t *ptl;
1365         pmd_t pmd;
1366
1367         mmap_assert_write_locked(mm);
1368         ptl = pmd_lock(vma->vm_mm, pmdp);
1369         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1370         spin_unlock(ptl);
1371         mm_dec_nr_ptes(mm);
1372         page_table_check_pte_clear_range(mm, addr, pmd);
1373         pte_free(mm, pmd_pgtable(pmd));
1374 }
1375
1376 /**
1377  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1378  * address haddr.
1379  *
1380  * @mm: process address space where collapse happens
1381  * @addr: THP collapse address
1382  *
1383  * This function checks whether all the PTEs in the PMD are pointing to the
1384  * right THP. If so, retract the page table so the THP can refault in with
1385  * as pmd-mapped.
1386  */
1387 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1388 {
1389         unsigned long haddr = addr & HPAGE_PMD_MASK;
1390         struct vm_area_struct *vma = find_vma(mm, haddr);
1391         struct page *hpage;
1392         pte_t *start_pte, *pte;
1393         pmd_t *pmd;
1394         spinlock_t *ptl;
1395         int count = 0;
1396         int i;
1397
1398         if (!vma || !vma->vm_file ||
1399             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1400                 return;
1401
1402         /*
1403          * This vm_flags may not have VM_HUGEPAGE if the page was not
1404          * collapsed by this mm. But we can still collapse if the page is
1405          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1406          * will not fail the vma for missing VM_HUGEPAGE
1407          */
1408         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE, false, false))
1409                 return;
1410
1411         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1412         if (userfaultfd_wp(vma))
1413                 return;
1414
1415         hpage = find_lock_page(vma->vm_file->f_mapping,
1416                                linear_page_index(vma, haddr));
1417         if (!hpage)
1418                 return;
1419
1420         if (!PageHead(hpage))
1421                 goto drop_hpage;
1422
1423         pmd = mm_find_pmd(mm, haddr);
1424         if (!pmd)
1425                 goto drop_hpage;
1426
1427         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1428
1429         /* step 1: check all mapped PTEs are to the right huge page */
1430         for (i = 0, addr = haddr, pte = start_pte;
1431              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1432                 struct page *page;
1433
1434                 /* empty pte, skip */
1435                 if (pte_none(*pte))
1436                         continue;
1437
1438                 /* page swapped out, abort */
1439                 if (!pte_present(*pte))
1440                         goto abort;
1441
1442                 page = vm_normal_page(vma, addr, *pte);
1443                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1444                         page = NULL;
1445                 /*
1446                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1447                  * page table, but the new page will not be a subpage of hpage.
1448                  */
1449                 if (hpage + i != page)
1450                         goto abort;
1451                 count++;
1452         }
1453
1454         /* step 2: adjust rmap */
1455         for (i = 0, addr = haddr, pte = start_pte;
1456              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1457                 struct page *page;
1458
1459                 if (pte_none(*pte))
1460                         continue;
1461                 page = vm_normal_page(vma, addr, *pte);
1462                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1463                         goto abort;
1464                 page_remove_rmap(page, vma, false);
1465         }
1466
1467         pte_unmap_unlock(start_pte, ptl);
1468
1469         /* step 3: set proper refcount and mm_counters. */
1470         if (count) {
1471                 page_ref_sub(hpage, count);
1472                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1473         }
1474
1475         /* step 4: collapse pmd */
1476         collapse_and_free_pmd(mm, vma, haddr, pmd);
1477 drop_hpage:
1478         unlock_page(hpage);
1479         put_page(hpage);
1480         return;
1481
1482 abort:
1483         pte_unmap_unlock(start_pte, ptl);
1484         goto drop_hpage;
1485 }
1486
1487 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1488 {
1489         struct mm_struct *mm = mm_slot->mm;
1490         int i;
1491
1492         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1493                 return;
1494
1495         if (!mmap_write_trylock(mm))
1496                 return;
1497
1498         if (unlikely(khugepaged_test_exit(mm)))
1499                 goto out;
1500
1501         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1502                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1503
1504 out:
1505         mm_slot->nr_pte_mapped_thp = 0;
1506         mmap_write_unlock(mm);
1507 }
1508
1509 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1510 {
1511         struct vm_area_struct *vma;
1512         struct mm_struct *mm;
1513         unsigned long addr;
1514         pmd_t *pmd;
1515
1516         i_mmap_lock_write(mapping);
1517         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1518                 /*
1519                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1520                  * got written to. These VMAs are likely not worth investing
1521                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1522                  * later.
1523                  *
1524                  * Note that vma->anon_vma check is racy: it can be set up after
1525                  * the check but before we took mmap_lock by the fault path.
1526                  * But page lock would prevent establishing any new ptes of the
1527                  * page, so we are safe.
1528                  *
1529                  * An alternative would be drop the check, but check that page
1530                  * table is clear before calling pmdp_collapse_flush() under
1531                  * ptl. It has higher chance to recover THP for the VMA, but
1532                  * has higher cost too.
1533                  */
1534                 if (vma->anon_vma)
1535                         continue;
1536                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1537                 if (addr & ~HPAGE_PMD_MASK)
1538                         continue;
1539                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1540                         continue;
1541                 mm = vma->vm_mm;
1542                 pmd = mm_find_pmd(mm, addr);
1543                 if (!pmd)
1544                         continue;
1545                 /*
1546                  * We need exclusive mmap_lock to retract page table.
1547                  *
1548                  * We use trylock due to lock inversion: we need to acquire
1549                  * mmap_lock while holding page lock. Fault path does it in
1550                  * reverse order. Trylock is a way to avoid deadlock.
1551                  */
1552                 if (mmap_write_trylock(mm)) {
1553                         /*
1554                          * When a vma is registered with uffd-wp, we can't
1555                          * recycle the pmd pgtable because there can be pte
1556                          * markers installed.  Skip it only, so the rest mm/vma
1557                          * can still have the same file mapped hugely, however
1558                          * it'll always mapped in small page size for uffd-wp
1559                          * registered ranges.
1560                          */
1561                         if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma))
1562                                 collapse_and_free_pmd(mm, vma, addr, pmd);
1563                         mmap_write_unlock(mm);
1564                 } else {
1565                         /* Try again later */
1566                         khugepaged_add_pte_mapped_thp(mm, addr);
1567                 }
1568         }
1569         i_mmap_unlock_write(mapping);
1570 }
1571
1572 /**
1573  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1574  *
1575  * @mm: process address space where collapse happens
1576  * @file: file that collapse on
1577  * @start: collapse start address
1578  * @hpage: new allocated huge page for collapse
1579  * @node: appointed node the new huge page allocate from
1580  *
1581  * Basic scheme is simple, details are more complex:
1582  *  - allocate and lock a new huge page;
1583  *  - scan page cache replacing old pages with the new one
1584  *    + swap/gup in pages if necessary;
1585  *    + fill in gaps;
1586  *    + keep old pages around in case rollback is required;
1587  *  - if replacing succeeds:
1588  *    + copy data over;
1589  *    + free old pages;
1590  *    + unlock huge page;
1591  *  - if replacing failed;
1592  *    + put all pages back and unfreeze them;
1593  *    + restore gaps in the page cache;
1594  *    + unlock and free huge page;
1595  */
1596 static void collapse_file(struct mm_struct *mm,
1597                 struct file *file, pgoff_t start,
1598                 struct page **hpage, int node)
1599 {
1600         struct address_space *mapping = file->f_mapping;
1601         gfp_t gfp;
1602         struct page *new_page;
1603         pgoff_t index, end = start + HPAGE_PMD_NR;
1604         LIST_HEAD(pagelist);
1605         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1606         int nr_none = 0, result = SCAN_SUCCEED;
1607         bool is_shmem = shmem_file(file);
1608         int nr;
1609
1610         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1611         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1612
1613         /* Only allocate from the target node */
1614         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1615
1616         new_page = khugepaged_alloc_page(hpage, gfp, node);
1617         if (!new_page) {
1618                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1619                 goto out;
1620         }
1621
1622         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1623                 result = SCAN_CGROUP_CHARGE_FAIL;
1624                 goto out;
1625         }
1626         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1627
1628         /*
1629          * Ensure we have slots for all the pages in the range.  This is
1630          * almost certainly a no-op because most of the pages must be present
1631          */
1632         do {
1633                 xas_lock_irq(&xas);
1634                 xas_create_range(&xas);
1635                 if (!xas_error(&xas))
1636                         break;
1637                 xas_unlock_irq(&xas);
1638                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1639                         result = SCAN_FAIL;
1640                         goto out;
1641                 }
1642         } while (1);
1643
1644         __SetPageLocked(new_page);
1645         if (is_shmem)
1646                 __SetPageSwapBacked(new_page);
1647         new_page->index = start;
1648         new_page->mapping = mapping;
1649
1650         /*
1651          * At this point the new_page is locked and not up-to-date.
1652          * It's safe to insert it into the page cache, because nobody would
1653          * be able to map it or use it in another way until we unlock it.
1654          */
1655
1656         xas_set(&xas, start);
1657         for (index = start; index < end; index++) {
1658                 struct page *page = xas_next(&xas);
1659
1660                 VM_BUG_ON(index != xas.xa_index);
1661                 if (is_shmem) {
1662                         if (!page) {
1663                                 /*
1664                                  * Stop if extent has been truncated or
1665                                  * hole-punched, and is now completely
1666                                  * empty.
1667                                  */
1668                                 if (index == start) {
1669                                         if (!xas_next_entry(&xas, end - 1)) {
1670                                                 result = SCAN_TRUNCATED;
1671                                                 goto xa_locked;
1672                                         }
1673                                         xas_set(&xas, index);
1674                                 }
1675                                 if (!shmem_charge(mapping->host, 1)) {
1676                                         result = SCAN_FAIL;
1677                                         goto xa_locked;
1678                                 }
1679                                 xas_store(&xas, new_page);
1680                                 nr_none++;
1681                                 continue;
1682                         }
1683
1684                         if (xa_is_value(page) || !PageUptodate(page)) {
1685                                 xas_unlock_irq(&xas);
1686                                 /* swap in or instantiate fallocated page */
1687                                 if (shmem_getpage(mapping->host, index, &page,
1688                                                   SGP_NOALLOC)) {
1689                                         result = SCAN_FAIL;
1690                                         goto xa_unlocked;
1691                                 }
1692                         } else if (trylock_page(page)) {
1693                                 get_page(page);
1694                                 xas_unlock_irq(&xas);
1695                         } else {
1696                                 result = SCAN_PAGE_LOCK;
1697                                 goto xa_locked;
1698                         }
1699                 } else {        /* !is_shmem */
1700                         if (!page || xa_is_value(page)) {
1701                                 xas_unlock_irq(&xas);
1702                                 page_cache_sync_readahead(mapping, &file->f_ra,
1703                                                           file, index,
1704                                                           end - index);
1705                                 /* drain pagevecs to help isolate_lru_page() */
1706                                 lru_add_drain();
1707                                 page = find_lock_page(mapping, index);
1708                                 if (unlikely(page == NULL)) {
1709                                         result = SCAN_FAIL;
1710                                         goto xa_unlocked;
1711                                 }
1712                         } else if (PageDirty(page)) {
1713                                 /*
1714                                  * khugepaged only works on read-only fd,
1715                                  * so this page is dirty because it hasn't
1716                                  * been flushed since first write. There
1717                                  * won't be new dirty pages.
1718                                  *
1719                                  * Trigger async flush here and hope the
1720                                  * writeback is done when khugepaged
1721                                  * revisits this page.
1722                                  *
1723                                  * This is a one-off situation. We are not
1724                                  * forcing writeback in loop.
1725                                  */
1726                                 xas_unlock_irq(&xas);
1727                                 filemap_flush(mapping);
1728                                 result = SCAN_FAIL;
1729                                 goto xa_unlocked;
1730                         } else if (PageWriteback(page)) {
1731                                 xas_unlock_irq(&xas);
1732                                 result = SCAN_FAIL;
1733                                 goto xa_unlocked;
1734                         } else if (trylock_page(page)) {
1735                                 get_page(page);
1736                                 xas_unlock_irq(&xas);
1737                         } else {
1738                                 result = SCAN_PAGE_LOCK;
1739                                 goto xa_locked;
1740                         }
1741                 }
1742
1743                 /*
1744                  * The page must be locked, so we can drop the i_pages lock
1745                  * without racing with truncate.
1746                  */
1747                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1748
1749                 /* make sure the page is up to date */
1750                 if (unlikely(!PageUptodate(page))) {
1751                         result = SCAN_FAIL;
1752                         goto out_unlock;
1753                 }
1754
1755                 /*
1756                  * If file was truncated then extended, or hole-punched, before
1757                  * we locked the first page, then a THP might be there already.
1758                  */
1759                 if (PageTransCompound(page)) {
1760                         result = SCAN_PAGE_COMPOUND;
1761                         goto out_unlock;
1762                 }
1763
1764                 if (page_mapping(page) != mapping) {
1765                         result = SCAN_TRUNCATED;
1766                         goto out_unlock;
1767                 }
1768
1769                 if (!is_shmem && (PageDirty(page) ||
1770                                   PageWriteback(page))) {
1771                         /*
1772                          * khugepaged only works on read-only fd, so this
1773                          * page is dirty because it hasn't been flushed
1774                          * since first write.
1775                          */
1776                         result = SCAN_FAIL;
1777                         goto out_unlock;
1778                 }
1779
1780                 if (isolate_lru_page(page)) {
1781                         result = SCAN_DEL_PAGE_LRU;
1782                         goto out_unlock;
1783                 }
1784
1785                 if (page_has_private(page) &&
1786                     !try_to_release_page(page, GFP_KERNEL)) {
1787                         result = SCAN_PAGE_HAS_PRIVATE;
1788                         putback_lru_page(page);
1789                         goto out_unlock;
1790                 }
1791
1792                 if (page_mapped(page))
1793                         try_to_unmap(page_folio(page),
1794                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1795
1796                 xas_lock_irq(&xas);
1797                 xas_set(&xas, index);
1798
1799                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1800
1801                 /*
1802                  * The page is expected to have page_count() == 3:
1803                  *  - we hold a pin on it;
1804                  *  - one reference from page cache;
1805                  *  - one from isolate_lru_page;
1806                  */
1807                 if (!page_ref_freeze(page, 3)) {
1808                         result = SCAN_PAGE_COUNT;
1809                         xas_unlock_irq(&xas);
1810                         putback_lru_page(page);
1811                         goto out_unlock;
1812                 }
1813
1814                 /*
1815                  * Add the page to the list to be able to undo the collapse if
1816                  * something go wrong.
1817                  */
1818                 list_add_tail(&page->lru, &pagelist);
1819
1820                 /* Finally, replace with the new page. */
1821                 xas_store(&xas, new_page);
1822                 continue;
1823 out_unlock:
1824                 unlock_page(page);
1825                 put_page(page);
1826                 goto xa_unlocked;
1827         }
1828         nr = thp_nr_pages(new_page);
1829
1830         if (is_shmem)
1831                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1832         else {
1833                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1834                 filemap_nr_thps_inc(mapping);
1835                 /*
1836                  * Paired with smp_mb() in do_dentry_open() to ensure
1837                  * i_writecount is up to date and the update to nr_thps is
1838                  * visible. Ensures the page cache will be truncated if the
1839                  * file is opened writable.
1840                  */
1841                 smp_mb();
1842                 if (inode_is_open_for_write(mapping->host)) {
1843                         result = SCAN_FAIL;
1844                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1845                         filemap_nr_thps_dec(mapping);
1846                         goto xa_locked;
1847                 }
1848         }
1849
1850         if (nr_none) {
1851                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1852                 /* nr_none is always 0 for non-shmem. */
1853                 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1854         }
1855
1856         /* Join all the small entries into a single multi-index entry */
1857         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1858         xas_store(&xas, new_page);
1859 xa_locked:
1860         xas_unlock_irq(&xas);
1861 xa_unlocked:
1862
1863         /*
1864          * If collapse is successful, flush must be done now before copying.
1865          * If collapse is unsuccessful, does flush actually need to be done?
1866          * Do it anyway, to clear the state.
1867          */
1868         try_to_unmap_flush();
1869
1870         if (result == SCAN_SUCCEED) {
1871                 struct page *page, *tmp;
1872
1873                 /*
1874                  * Replacing old pages with new one has succeeded, now we
1875                  * need to copy the content and free the old pages.
1876                  */
1877                 index = start;
1878                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1879                         while (index < page->index) {
1880                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1881                                 index++;
1882                         }
1883                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1884                                         page);
1885                         list_del(&page->lru);
1886                         page->mapping = NULL;
1887                         page_ref_unfreeze(page, 1);
1888                         ClearPageActive(page);
1889                         ClearPageUnevictable(page);
1890                         unlock_page(page);
1891                         put_page(page);
1892                         index++;
1893                 }
1894                 while (index < end) {
1895                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1896                         index++;
1897                 }
1898
1899                 SetPageUptodate(new_page);
1900                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1901                 if (is_shmem)
1902                         set_page_dirty(new_page);
1903                 lru_cache_add(new_page);
1904
1905                 /*
1906                  * Remove pte page tables, so we can re-fault the page as huge.
1907                  */
1908                 retract_page_tables(mapping, start);
1909                 *hpage = NULL;
1910
1911                 khugepaged_pages_collapsed++;
1912         } else {
1913                 struct page *page;
1914
1915                 /* Something went wrong: roll back page cache changes */
1916                 xas_lock_irq(&xas);
1917                 if (nr_none) {
1918                         mapping->nrpages -= nr_none;
1919                         shmem_uncharge(mapping->host, nr_none);
1920                 }
1921
1922                 xas_set(&xas, start);
1923                 xas_for_each(&xas, page, end - 1) {
1924                         page = list_first_entry_or_null(&pagelist,
1925                                         struct page, lru);
1926                         if (!page || xas.xa_index < page->index) {
1927                                 if (!nr_none)
1928                                         break;
1929                                 nr_none--;
1930                                 /* Put holes back where they were */
1931                                 xas_store(&xas, NULL);
1932                                 continue;
1933                         }
1934
1935                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1936
1937                         /* Unfreeze the page. */
1938                         list_del(&page->lru);
1939                         page_ref_unfreeze(page, 2);
1940                         xas_store(&xas, page);
1941                         xas_pause(&xas);
1942                         xas_unlock_irq(&xas);
1943                         unlock_page(page);
1944                         putback_lru_page(page);
1945                         xas_lock_irq(&xas);
1946                 }
1947                 VM_BUG_ON(nr_none);
1948                 xas_unlock_irq(&xas);
1949
1950                 new_page->mapping = NULL;
1951         }
1952
1953         unlock_page(new_page);
1954 out:
1955         VM_BUG_ON(!list_empty(&pagelist));
1956         if (!IS_ERR_OR_NULL(*hpage))
1957                 mem_cgroup_uncharge(page_folio(*hpage));
1958         /* TODO: tracepoints */
1959 }
1960
1961 static void khugepaged_scan_file(struct mm_struct *mm,
1962                 struct file *file, pgoff_t start, struct page **hpage)
1963 {
1964         struct page *page = NULL;
1965         struct address_space *mapping = file->f_mapping;
1966         XA_STATE(xas, &mapping->i_pages, start);
1967         int present, swap;
1968         int node = NUMA_NO_NODE;
1969         int result = SCAN_SUCCEED;
1970
1971         present = 0;
1972         swap = 0;
1973         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1974         rcu_read_lock();
1975         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1976                 if (xas_retry(&xas, page))
1977                         continue;
1978
1979                 if (xa_is_value(page)) {
1980                         if (++swap > khugepaged_max_ptes_swap) {
1981                                 result = SCAN_EXCEED_SWAP_PTE;
1982                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1983                                 break;
1984                         }
1985                         continue;
1986                 }
1987
1988                 /*
1989                  * XXX: khugepaged should compact smaller compound pages
1990                  * into a PMD sized page
1991                  */
1992                 if (PageTransCompound(page)) {
1993                         result = SCAN_PAGE_COMPOUND;
1994                         break;
1995                 }
1996
1997                 node = page_to_nid(page);
1998                 if (khugepaged_scan_abort(node)) {
1999                         result = SCAN_SCAN_ABORT;
2000                         break;
2001                 }
2002                 khugepaged_node_load[node]++;
2003
2004                 if (!PageLRU(page)) {
2005                         result = SCAN_PAGE_LRU;
2006                         break;
2007                 }
2008
2009                 if (page_count(page) !=
2010                     1 + page_mapcount(page) + page_has_private(page)) {
2011                         result = SCAN_PAGE_COUNT;
2012                         break;
2013                 }
2014
2015                 /*
2016                  * We probably should check if the page is referenced here, but
2017                  * nobody would transfer pte_young() to PageReferenced() for us.
2018                  * And rmap walk here is just too costly...
2019                  */
2020
2021                 present++;
2022
2023                 if (need_resched()) {
2024                         xas_pause(&xas);
2025                         cond_resched_rcu();
2026                 }
2027         }
2028         rcu_read_unlock();
2029
2030         if (result == SCAN_SUCCEED) {
2031                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2032                         result = SCAN_EXCEED_NONE_PTE;
2033                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2034                 } else {
2035                         node = khugepaged_find_target_node();
2036                         collapse_file(mm, file, start, hpage, node);
2037                 }
2038         }
2039
2040         /* TODO: tracepoints */
2041 }
2042 #else
2043 static void khugepaged_scan_file(struct mm_struct *mm,
2044                 struct file *file, pgoff_t start, struct page **hpage)
2045 {
2046         BUILD_BUG();
2047 }
2048
2049 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2050 {
2051 }
2052 #endif
2053
2054 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2055                                             struct page **hpage)
2056         __releases(&khugepaged_mm_lock)
2057         __acquires(&khugepaged_mm_lock)
2058 {
2059         struct mm_slot *mm_slot;
2060         struct mm_struct *mm;
2061         struct vm_area_struct *vma;
2062         int progress = 0;
2063
2064         VM_BUG_ON(!pages);
2065         lockdep_assert_held(&khugepaged_mm_lock);
2066
2067         if (khugepaged_scan.mm_slot)
2068                 mm_slot = khugepaged_scan.mm_slot;
2069         else {
2070                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2071                                      struct mm_slot, mm_node);
2072                 khugepaged_scan.address = 0;
2073                 khugepaged_scan.mm_slot = mm_slot;
2074         }
2075         spin_unlock(&khugepaged_mm_lock);
2076         khugepaged_collapse_pte_mapped_thps(mm_slot);
2077
2078         mm = mm_slot->mm;
2079         /*
2080          * Don't wait for semaphore (to avoid long wait times).  Just move to
2081          * the next mm on the list.
2082          */
2083         vma = NULL;
2084         if (unlikely(!mmap_read_trylock(mm)))
2085                 goto breakouterloop_mmap_lock;
2086         if (likely(!khugepaged_test_exit(mm)))
2087                 vma = find_vma(mm, khugepaged_scan.address);
2088
2089         progress++;
2090         for (; vma; vma = vma->vm_next) {
2091                 unsigned long hstart, hend;
2092
2093                 cond_resched();
2094                 if (unlikely(khugepaged_test_exit(mm))) {
2095                         progress++;
2096                         break;
2097                 }
2098                 if (!hugepage_vma_check(vma, vma->vm_flags, false, false)) {
2099 skip:
2100                         progress++;
2101                         continue;
2102                 }
2103                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2104                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2105                 if (khugepaged_scan.address > hend)
2106                         goto skip;
2107                 if (khugepaged_scan.address < hstart)
2108                         khugepaged_scan.address = hstart;
2109                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2110
2111                 while (khugepaged_scan.address < hend) {
2112                         int ret;
2113                         cond_resched();
2114                         if (unlikely(khugepaged_test_exit(mm)))
2115                                 goto breakouterloop;
2116
2117                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2118                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2119                                   hend);
2120                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2121                                 struct file *file = get_file(vma->vm_file);
2122                                 pgoff_t pgoff = linear_page_index(vma,
2123                                                 khugepaged_scan.address);
2124
2125                                 mmap_read_unlock(mm);
2126                                 ret = 1;
2127                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2128                                 fput(file);
2129                         } else {
2130                                 ret = khugepaged_scan_pmd(mm, vma,
2131                                                 khugepaged_scan.address,
2132                                                 hpage);
2133                         }
2134                         /* move to next address */
2135                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2136                         progress += HPAGE_PMD_NR;
2137                         if (ret)
2138                                 /* we released mmap_lock so break loop */
2139                                 goto breakouterloop_mmap_lock;
2140                         if (progress >= pages)
2141                                 goto breakouterloop;
2142                 }
2143         }
2144 breakouterloop:
2145         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2146 breakouterloop_mmap_lock:
2147
2148         spin_lock(&khugepaged_mm_lock);
2149         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2150         /*
2151          * Release the current mm_slot if this mm is about to die, or
2152          * if we scanned all vmas of this mm.
2153          */
2154         if (khugepaged_test_exit(mm) || !vma) {
2155                 /*
2156                  * Make sure that if mm_users is reaching zero while
2157                  * khugepaged runs here, khugepaged_exit will find
2158                  * mm_slot not pointing to the exiting mm.
2159                  */
2160                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2161                         khugepaged_scan.mm_slot = list_entry(
2162                                 mm_slot->mm_node.next,
2163                                 struct mm_slot, mm_node);
2164                         khugepaged_scan.address = 0;
2165                 } else {
2166                         khugepaged_scan.mm_slot = NULL;
2167                         khugepaged_full_scans++;
2168                 }
2169
2170                 collect_mm_slot(mm_slot);
2171         }
2172
2173         return progress;
2174 }
2175
2176 static int khugepaged_has_work(void)
2177 {
2178         return !list_empty(&khugepaged_scan.mm_head) &&
2179                 hugepage_flags_enabled();
2180 }
2181
2182 static int khugepaged_wait_event(void)
2183 {
2184         return !list_empty(&khugepaged_scan.mm_head) ||
2185                 kthread_should_stop();
2186 }
2187
2188 static void khugepaged_do_scan(void)
2189 {
2190         struct page *hpage = NULL;
2191         unsigned int progress = 0, pass_through_head = 0;
2192         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2193         bool wait = true;
2194
2195         lru_add_drain_all();
2196
2197         while (progress < pages) {
2198                 if (!khugepaged_prealloc_page(&hpage, &wait))
2199                         break;
2200
2201                 cond_resched();
2202
2203                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2204                         break;
2205
2206                 spin_lock(&khugepaged_mm_lock);
2207                 if (!khugepaged_scan.mm_slot)
2208                         pass_through_head++;
2209                 if (khugepaged_has_work() &&
2210                     pass_through_head < 2)
2211                         progress += khugepaged_scan_mm_slot(pages - progress,
2212                                                             &hpage);
2213                 else
2214                         progress = pages;
2215                 spin_unlock(&khugepaged_mm_lock);
2216         }
2217
2218         if (!IS_ERR_OR_NULL(hpage))
2219                 put_page(hpage);
2220 }
2221
2222 static bool khugepaged_should_wakeup(void)
2223 {
2224         return kthread_should_stop() ||
2225                time_after_eq(jiffies, khugepaged_sleep_expire);
2226 }
2227
2228 static void khugepaged_wait_work(void)
2229 {
2230         if (khugepaged_has_work()) {
2231                 const unsigned long scan_sleep_jiffies =
2232                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2233
2234                 if (!scan_sleep_jiffies)
2235                         return;
2236
2237                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2238                 wait_event_freezable_timeout(khugepaged_wait,
2239                                              khugepaged_should_wakeup(),
2240                                              scan_sleep_jiffies);
2241                 return;
2242         }
2243
2244         if (hugepage_flags_enabled())
2245                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2246 }
2247
2248 static int khugepaged(void *none)
2249 {
2250         struct mm_slot *mm_slot;
2251
2252         set_freezable();
2253         set_user_nice(current, MAX_NICE);
2254
2255         while (!kthread_should_stop()) {
2256                 khugepaged_do_scan();
2257                 khugepaged_wait_work();
2258         }
2259
2260         spin_lock(&khugepaged_mm_lock);
2261         mm_slot = khugepaged_scan.mm_slot;
2262         khugepaged_scan.mm_slot = NULL;
2263         if (mm_slot)
2264                 collect_mm_slot(mm_slot);
2265         spin_unlock(&khugepaged_mm_lock);
2266         return 0;
2267 }
2268
2269 static void set_recommended_min_free_kbytes(void)
2270 {
2271         struct zone *zone;
2272         int nr_zones = 0;
2273         unsigned long recommended_min;
2274
2275         if (!hugepage_flags_enabled()) {
2276                 calculate_min_free_kbytes();
2277                 goto update_wmarks;
2278         }
2279
2280         for_each_populated_zone(zone) {
2281                 /*
2282                  * We don't need to worry about fragmentation of
2283                  * ZONE_MOVABLE since it only has movable pages.
2284                  */
2285                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2286                         continue;
2287
2288                 nr_zones++;
2289         }
2290
2291         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2292         recommended_min = pageblock_nr_pages * nr_zones * 2;
2293
2294         /*
2295          * Make sure that on average at least two pageblocks are almost free
2296          * of another type, one for a migratetype to fall back to and a
2297          * second to avoid subsequent fallbacks of other types There are 3
2298          * MIGRATE_TYPES we care about.
2299          */
2300         recommended_min += pageblock_nr_pages * nr_zones *
2301                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2302
2303         /* don't ever allow to reserve more than 5% of the lowmem */
2304         recommended_min = min(recommended_min,
2305                               (unsigned long) nr_free_buffer_pages() / 20);
2306         recommended_min <<= (PAGE_SHIFT-10);
2307
2308         if (recommended_min > min_free_kbytes) {
2309                 if (user_min_free_kbytes >= 0)
2310                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2311                                 min_free_kbytes, recommended_min);
2312
2313                 min_free_kbytes = recommended_min;
2314         }
2315
2316 update_wmarks:
2317         setup_per_zone_wmarks();
2318 }
2319
2320 int start_stop_khugepaged(void)
2321 {
2322         int err = 0;
2323
2324         mutex_lock(&khugepaged_mutex);
2325         if (hugepage_flags_enabled()) {
2326                 if (!khugepaged_thread)
2327                         khugepaged_thread = kthread_run(khugepaged, NULL,
2328                                                         "khugepaged");
2329                 if (IS_ERR(khugepaged_thread)) {
2330                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2331                         err = PTR_ERR(khugepaged_thread);
2332                         khugepaged_thread = NULL;
2333                         goto fail;
2334                 }
2335
2336                 if (!list_empty(&khugepaged_scan.mm_head))
2337                         wake_up_interruptible(&khugepaged_wait);
2338         } else if (khugepaged_thread) {
2339                 kthread_stop(khugepaged_thread);
2340                 khugepaged_thread = NULL;
2341         }
2342         set_recommended_min_free_kbytes();
2343 fail:
2344         mutex_unlock(&khugepaged_mutex);
2345         return err;
2346 }
2347
2348 void khugepaged_min_free_kbytes_update(void)
2349 {
2350         mutex_lock(&khugepaged_mutex);
2351         if (hugepage_flags_enabled() && khugepaged_thread)
2352                 set_recommended_min_free_kbytes();
2353         mutex_unlock(&khugepaged_mutex);
2354 }