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_SWAP_CACHE_PAGE,
  50         SCAN_DEL_PAGE_LRU,
  51         SCAN_ALLOC_HUGE_PAGE_FAIL,
  52         SCAN_CGROUP_CHARGE_FAIL,
  53         SCAN_TRUNCATED,
  54         SCAN_PAGE_HAS_PRIVATE,
  55 };
  56
  57 #define CREATE_TRACE_POINTS
  58 #include <trace/events/huge_memory.h>
  59
  60 static struct task_struct *khugepaged_thread __read_mostly;
  61 static DEFINE_MUTEX(khugepaged_mutex);
  62
  63 /* default scan 8*512 pte (or vmas) every 30 second */
  64 static unsigned int khugepaged_pages_to_scan __read_mostly;
  65 static unsigned int khugepaged_pages_collapsed;
  66 static unsigned int khugepaged_full_scans;
  67 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  68 /* during fragmentation poll the hugepage allocator once every minute */
  69 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  70 static unsigned long khugepaged_sleep_expire;
  71 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  72 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  73 /*
  74  * default collapse hugepages if there is at least one pte mapped like
  75  * it would have happened if the vma was large enough during page
  76  * fault.
  77  */
  78 static unsigned int khugepaged_max_ptes_none __read_mostly;
  79 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  80 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  81
  82 #define MM_SLOTS_HASH_BITS 10
  83 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  84
  85 static struct kmem_cache *mm_slot_cache __read_mostly;
  86
  87 #define MAX_PTE_MAPPED_THP 8
  88
  89 /**
  90  * struct mm_slot - hash lookup from mm to mm_slot
  91  * @hash: hash collision list
  92  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  93  * @mm: the mm that this information is valid for
  94  * @nr_pte_mapped_thp: number of pte mapped THP
  95  * @pte_mapped_thp: address array corresponding pte mapped THP
  96  */
  97 struct mm_slot {
  98         struct hlist_node hash;
  99         struct list_head mm_node;
 100         struct mm_struct *mm;
 101
 102         /* pte-mapped THP in this mm */
 103         int nr_pte_mapped_thp;
 104         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 105 };
 106
 107 /**
 108  * struct khugepaged_scan - cursor for scanning
 109  * @mm_head: the head of the mm list to scan
 110  * @mm_slot: the current mm_slot we are scanning
 111  * @address: the next address inside that to be scanned
 112  *
 113  * There is only the one khugepaged_scan instance of this cursor structure.
 114  */
 115 struct khugepaged_scan {
 116         struct list_head mm_head;
 117         struct mm_slot *mm_slot;
 118         unsigned long address;
 119 };
 120
 121 static struct khugepaged_scan khugepaged_scan = {
 122         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 123 };
 124
 125 #ifdef CONFIG_SYSFS
 126 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 127                                          struct kobj_attribute *attr,
 128                                          char *buf)
 129 {
 130         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 131 }
 132
 133 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 134                                           struct kobj_attribute *attr,
 135                                           const char *buf, size_t count)
 136 {
 137         unsigned int msecs;
 138         int err;
 139
 140         err = kstrtouint(buf, 10, &msecs);
 141         if (err)
 142                 return -EINVAL;
 143
 144         khugepaged_scan_sleep_millisecs = msecs;
 145         khugepaged_sleep_expire = 0;
 146         wake_up_interruptible(&khugepaged_wait);
 147
 148         return count;
 149 }
 150 static struct kobj_attribute scan_sleep_millisecs_attr =
 151         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 152                scan_sleep_millisecs_store);
 153
 154 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 155                                           struct kobj_attribute *attr,
 156                                           char *buf)
 157 {
 158         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 159 }
 160
 161 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 162                                            struct kobj_attribute *attr,
 163                                            const char *buf, size_t count)
 164 {
 165         unsigned int msecs;
 166         int err;
 167
 168         err = kstrtouint(buf, 10, &msecs);
 169         if (err)
 170                 return -EINVAL;
 171
 172         khugepaged_alloc_sleep_millisecs = msecs;
 173         khugepaged_sleep_expire = 0;
 174         wake_up_interruptible(&khugepaged_wait);
 175
 176         return count;
 177 }
 178 static struct kobj_attribute alloc_sleep_millisecs_attr =
 179         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 180                alloc_sleep_millisecs_store);
 181
 182 static ssize_t pages_to_scan_show(struct kobject *kobj,
 183                                   struct kobj_attribute *attr,
 184                                   char *buf)
 185 {
 186         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 187 }
 188 static ssize_t pages_to_scan_store(struct kobject *kobj,
 189                                    struct kobj_attribute *attr,
 190                                    const char *buf, size_t count)
 191 {
 192         unsigned int pages;
 193         int err;
 194
 195         err = kstrtouint(buf, 10, &pages);
 196         if (err || !pages)
 197                 return -EINVAL;
 198
 199         khugepaged_pages_to_scan = pages;
 200
 201         return count;
 202 }
 203 static struct kobj_attribute pages_to_scan_attr =
 204         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 205                pages_to_scan_store);
 206
 207 static ssize_t pages_collapsed_show(struct kobject *kobj,
 208                                     struct kobj_attribute *attr,
 209                                     char *buf)
 210 {
 211         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 212 }
 213 static struct kobj_attribute pages_collapsed_attr =
 214         __ATTR_RO(pages_collapsed);
 215
 216 static ssize_t full_scans_show(struct kobject *kobj,
 217                                struct kobj_attribute *attr,
 218                                char *buf)
 219 {
 220         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 221 }
 222 static struct kobj_attribute full_scans_attr =
 223         __ATTR_RO(full_scans);
 224
 225 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 226                                       struct kobj_attribute *attr, char *buf)
 227 {
 228         return single_hugepage_flag_show(kobj, attr, buf,
 229                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 230 }
 231 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 232                                        struct kobj_attribute *attr,
 233                                        const char *buf, size_t count)
 234 {
 235         return single_hugepage_flag_store(kobj, attr, buf, count,
 236                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 237 }
 238 static struct kobj_attribute khugepaged_defrag_attr =
 239         __ATTR(defrag, 0644, khugepaged_defrag_show,
 240                khugepaged_defrag_store);
 241
 242 /*
 243  * max_ptes_none controls if khugepaged should collapse hugepages over
 244  * any unmapped ptes in turn potentially increasing the memory
 245  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 246  * reduce the available free memory in the system as it
 247  * runs. Increasing max_ptes_none will instead potentially reduce the
 248  * free memory in the system during the khugepaged scan.
 249  */
 250 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 251                                              struct kobj_attribute *attr,
 252                                              char *buf)
 253 {
 254         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 255 }
 256 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 257                                               struct kobj_attribute *attr,
 258                                               const char *buf, size_t count)
 259 {
 260         int err;
 261         unsigned long max_ptes_none;
 262
 263         err = kstrtoul(buf, 10, &max_ptes_none);
 264         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 265                 return -EINVAL;
 266
 267         khugepaged_max_ptes_none = max_ptes_none;
 268
 269         return count;
 270 }
 271 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 272         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 273                khugepaged_max_ptes_none_store);
 274
 275 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 276                                              struct kobj_attribute *attr,
 277                                              char *buf)
 278 {
 279         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 280 }
 281
 282 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 283                                               struct kobj_attribute *attr,
 284                                               const char *buf, size_t count)
 285 {
 286         int err;
 287         unsigned long max_ptes_swap;
 288
 289         err  = kstrtoul(buf, 10, &max_ptes_swap);
 290         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 291                 return -EINVAL;
 292
 293         khugepaged_max_ptes_swap = max_ptes_swap;
 294
 295         return count;
 296 }
 297
 298 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 299         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 300                khugepaged_max_ptes_swap_store);
 301
 302 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 303                                                struct kobj_attribute *attr,
 304                                                char *buf)
 305 {
 306         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 307 }
 308
 309 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 310                                               struct kobj_attribute *attr,
 311                                               const char *buf, size_t count)
 312 {
 313         int err;
 314         unsigned long max_ptes_shared;
 315
 316         err  = kstrtoul(buf, 10, &max_ptes_shared);
 317         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 318                 return -EINVAL;
 319
 320         khugepaged_max_ptes_shared = max_ptes_shared;
 321
 322         return count;
 323 }
 324
 325 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 326         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 327                khugepaged_max_ptes_shared_store);
 328
 329 static struct attribute *khugepaged_attr[] = {
 330         &khugepaged_defrag_attr.attr,
 331         &khugepaged_max_ptes_none_attr.attr,
 332         &khugepaged_max_ptes_swap_attr.attr,
 333         &khugepaged_max_ptes_shared_attr.attr,
 334         &pages_to_scan_attr.attr,
 335         &pages_collapsed_attr.attr,
 336         &full_scans_attr.attr,
 337         &scan_sleep_millisecs_attr.attr,
 338         &alloc_sleep_millisecs_attr.attr,
 339         NULL,
 340 };
 341
 342 struct attribute_group khugepaged_attr_group = {
 343         .attrs = khugepaged_attr,
 344         .name = "khugepaged",
 345 };
 346 #endif /* CONFIG_SYSFS */
 347
 348 int hugepage_madvise(struct vm_area_struct *vma,
 349                      unsigned long *vm_flags, int advice)
 350 {
 351         switch (advice) {
 352         case MADV_HUGEPAGE:
 353 #ifdef CONFIG_S390
 354                 /*
 355                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 356                  * can't handle this properly after s390_enable_sie, so we simply
 357                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 358                  */
 359                 if (mm_has_pgste(vma->vm_mm))
 360                         return 0;
 361 #endif
 362                 *vm_flags &= ~VM_NOHUGEPAGE;
 363                 *vm_flags |= VM_HUGEPAGE;
 364                 /*
 365                  * If the vma become good for khugepaged to scan,
 366                  * register it here without waiting a page fault that
 367                  * may not happen any time soon.
 368                  */
 369                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 370                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 371                         return -ENOMEM;
 372                 break;
 373         case MADV_NOHUGEPAGE:
 374                 *vm_flags &= ~VM_HUGEPAGE;
 375                 *vm_flags |= VM_NOHUGEPAGE;
 376                 /*
 377                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 378                  * this vma even if we leave the mm registered in khugepaged if
 379                  * it got registered before VM_NOHUGEPAGE was set.
 380                  */
 381                 break;
 382         }
 383
 384         return 0;
 385 }
 386
 387 int __init khugepaged_init(void)
 388 {
 389         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 390                                           sizeof(struct mm_slot),
 391                                           __alignof__(struct mm_slot), 0, NULL);
 392         if (!mm_slot_cache)
 393                 return -ENOMEM;
 394
 395         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 396         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 397         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 398         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 399
 400         return 0;
 401 }
 402
 403 void __init khugepaged_destroy(void)
 404 {
 405         kmem_cache_destroy(mm_slot_cache);
 406 }
 407
 408 static inline struct mm_slot *alloc_mm_slot(void)
 409 {
 410         if (!mm_slot_cache)     /* initialization failed */
 411                 return NULL;
 412         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 413 }
 414
 415 static inline void free_mm_slot(struct mm_slot *mm_slot)
 416 {
 417         kmem_cache_free(mm_slot_cache, mm_slot);
 418 }
 419
 420 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 421 {
 422         struct mm_slot *mm_slot;
 423
 424         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 425                 if (mm == mm_slot->mm)
 426                         return mm_slot;
 427
 428         return NULL;
 429 }
 430
 431 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 432                                     struct mm_slot *mm_slot)
 433 {
 434         mm_slot->mm = mm;
 435         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 436 }
 437
 438 static inline int khugepaged_test_exit(struct mm_struct *mm)
 439 {
 440         return atomic_read(&mm->mm_users) == 0;
 441 }
 442
 443 static bool hugepage_vma_check(struct vm_area_struct *vma,
 444                                unsigned long vm_flags)
 445 {
 446         if (!transhuge_vma_enabled(vma, vm_flags))
 447                 return false;
 448
 449         if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
 450                                 vma->vm_pgoff, HPAGE_PMD_NR))
 451                 return false;
 452
 453         /* Enabled via shmem mount options or sysfs settings. */
 454         if (shmem_file(vma->vm_file))
 455                 return shmem_huge_enabled(vma);
 456
 457         /* THP settings require madvise. */
 458         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 459                 return false;
 460
 461         /* Only regular file is valid */
 462         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 463             (vm_flags & VM_EXEC)) {
 464                 struct inode *inode = vma->vm_file->f_inode;
 465
 466                 return !inode_is_open_for_write(inode) &&
 467                         S_ISREG(inode->i_mode);
 468         }
 469
 470         if (!vma->anon_vma || vma->vm_ops)
 471                 return false;
 472         if (vma_is_temporary_stack(vma))
 473                 return false;
 474         return !(vm_flags & VM_NO_KHUGEPAGED);
 475 }
 476
 477 int __khugepaged_enter(struct mm_struct *mm)
 478 {
 479         struct mm_slot *mm_slot;
 480         int wakeup;
 481
 482         mm_slot = alloc_mm_slot();
 483         if (!mm_slot)
 484                 return -ENOMEM;
 485
 486         /* __khugepaged_exit() must not run from under us */
 487         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 488         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 489                 free_mm_slot(mm_slot);
 490                 return 0;
 491         }
 492
 493         spin_lock(&khugepaged_mm_lock);
 494         insert_to_mm_slots_hash(mm, mm_slot);
 495         /*
 496          * Insert just behind the scanning cursor, to let the area settle
 497          * down a little.
 498          */
 499         wakeup = list_empty(&khugepaged_scan.mm_head);
 500         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 501         spin_unlock(&khugepaged_mm_lock);
 502
 503         mmgrab(mm);
 504         if (wakeup)
 505                 wake_up_interruptible(&khugepaged_wait);
 506
 507         return 0;
 508 }
 509
 510 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 511                                unsigned long vm_flags)
 512 {
 513         unsigned long hstart, hend;
 514
 515         /*
 516          * khugepaged only supports read-only files for non-shmem files.
 517          * khugepaged does not yet work on special mappings. And
 518          * file-private shmem THP is not supported.
 519          */
 520         if (!hugepage_vma_check(vma, vm_flags))
 521                 return 0;
 522
 523         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 524         hend = vma->vm_end & HPAGE_PMD_MASK;
 525         if (hstart < hend)
 526                 return khugepaged_enter(vma, vm_flags);
 527         return 0;
 528 }
 529
 530 void __khugepaged_exit(struct mm_struct *mm)
 531 {
 532         struct mm_slot *mm_slot;
 533         int free = 0;
 534
 535         spin_lock(&khugepaged_mm_lock);
 536         mm_slot = get_mm_slot(mm);
 537         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 538                 hash_del(&mm_slot->hash);
 539                 list_del(&mm_slot->mm_node);
 540                 free = 1;
 541         }
 542         spin_unlock(&khugepaged_mm_lock);
 543
 544         if (free) {
 545                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 546                 free_mm_slot(mm_slot);
 547                 mmdrop(mm);
 548         } else if (mm_slot) {
 549                 /*
 550                  * This is required to serialize against
 551                  * khugepaged_test_exit() (which is guaranteed to run
 552                  * under mmap sem read mode). Stop here (after we
 553                  * return all pagetables will be destroyed) until
 554                  * khugepaged has finished working on the pagetables
 555                  * under the mmap_lock.
 556                  */
 557                 mmap_write_lock(mm);
 558                 mmap_write_unlock(mm);
 559         }
 560 }
 561
 562 static void release_pte_page(struct page *page)
 563 {
 564         mod_node_page_state(page_pgdat(page),
 565                         NR_ISOLATED_ANON + page_is_file_lru(page),
 566                         -compound_nr(page));
 567         unlock_page(page);
 568         putback_lru_page(page);
 569 }
 570
 571 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 572                 struct list_head *compound_pagelist)
 573 {
 574         struct page *page, *tmp;
 575
 576         while (--_pte >= pte) {
 577                 pte_t pteval = *_pte;
 578
 579                 page = pte_page(pteval);
 580                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 581                                 !PageCompound(page))
 582                         release_pte_page(page);
 583         }
 584
 585         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 586                 list_del(&page->lru);
 587                 release_pte_page(page);
 588         }
 589 }
 590
 591 static bool is_refcount_suitable(struct page *page)
 592 {
 593         int expected_refcount;
 594
 595         expected_refcount = total_mapcount(page);
 596         if (PageSwapCache(page))
 597                 expected_refcount += compound_nr(page);
 598
 599         return page_count(page) == expected_refcount;
 600 }
 601
 602 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 603                                         unsigned long address,
 604                                         pte_t *pte,
 605                                         struct list_head *compound_pagelist)
 606 {
 607         struct page *page = NULL;
 608         pte_t *_pte;
 609         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 610         bool writable = false;
 611
 612         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 613              _pte++, address += PAGE_SIZE) {
 614                 pte_t pteval = *_pte;
 615                 if (pte_none(pteval) || (pte_present(pteval) &&
 616                                 is_zero_pfn(pte_pfn(pteval)))) {
 617                         if (!userfaultfd_armed(vma) &&
 618                             ++none_or_zero <= khugepaged_max_ptes_none) {
 619                                 continue;
 620                         } else {
 621                                 result = SCAN_EXCEED_NONE_PTE;
 622                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 623                                 goto out;
 624                         }
 625                 }
 626                 if (!pte_present(pteval)) {
 627                         result = SCAN_PTE_NON_PRESENT;
 628                         goto out;
 629                 }
 630                 page = vm_normal_page(vma, address, pteval);
 631                 if (unlikely(!page)) {
 632                         result = SCAN_PAGE_NULL;
 633                         goto out;
 634                 }
 635
 636                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 637
 638                 if (page_mapcount(page) > 1 &&
 639                                 ++shared > khugepaged_max_ptes_shared) {
 640                         result = SCAN_EXCEED_SHARED_PTE;
 641                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 642                         goto out;
 643                 }
 644
 645                 if (PageCompound(page)) {
 646                         struct page *p;
 647                         page = compound_head(page);
 648
 649                         /*
 650                          * Check if we have dealt with the compound page
 651                          * already
 652                          */
 653                         list_for_each_entry(p, compound_pagelist, lru) {
 654                                 if (page == p)
 655                                         goto next;
 656                         }
 657                 }
 658
 659                 /*
 660                  * We can do it before isolate_lru_page because the
 661                  * page can't be freed from under us. NOTE: PG_lock
 662                  * is needed to serialize against split_huge_page
 663                  * when invoked from the VM.
 664                  */
 665                 if (!trylock_page(page)) {
 666                         result = SCAN_PAGE_LOCK;
 667                         goto out;
 668                 }
 669
 670                 /*
 671                  * Check if the page has any GUP (or other external) pins.
 672                  *
 673                  * The page table that maps the page has been already unlinked
 674                  * from the page table tree and this process cannot get
 675                  * an additional pin on the page.
 676                  *
 677                  * New pins can come later if the page is shared across fork,
 678                  * but not from this process. The other process cannot write to
 679                  * the page, only trigger CoW.
 680                  */
 681                 if (!is_refcount_suitable(page)) {
 682                         unlock_page(page);
 683                         result = SCAN_PAGE_COUNT;
 684                         goto out;
 685                 }
 686                 if (!pte_write(pteval) && PageSwapCache(page) &&
 687                                 !reuse_swap_page(page)) {
 688                         /*
 689                          * Page is in the swap cache and cannot be re-used.
 690                          * It cannot be collapsed into a THP.
 691                          */
 692                         unlock_page(page);
 693                         result = SCAN_SWAP_CACHE_PAGE;
 694                         goto out;
 695                 }
 696
 697                 /*
 698                  * Isolate the page to avoid collapsing an hugepage
 699                  * currently in use by the VM.
 700                  */
 701                 if (isolate_lru_page(page)) {
 702                         unlock_page(page);
 703                         result = SCAN_DEL_PAGE_LRU;
 704                         goto out;
 705                 }
 706                 mod_node_page_state(page_pgdat(page),
 707                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 708                                 compound_nr(page));
 709                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 710                 VM_BUG_ON_PAGE(PageLRU(page), page);
 711
 712                 if (PageCompound(page))
 713                         list_add_tail(&page->lru, compound_pagelist);
 714 next:
 715                 /* There should be enough young pte to collapse the page */
 716                 if (pte_young(pteval) ||
 717                     page_is_young(page) || PageReferenced(page) ||
 718                     mmu_notifier_test_young(vma->vm_mm, address))
 719                         referenced++;
 720
 721                 if (pte_write(pteval))
 722                         writable = true;
 723         }
 724
 725         if (unlikely(!writable)) {
 726                 result = SCAN_PAGE_RO;
 727         } else if (unlikely(!referenced)) {
 728                 result = SCAN_LACK_REFERENCED_PAGE;
 729         } else {
 730                 result = SCAN_SUCCEED;
 731                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 732                                                     referenced, writable, result);
 733                 return 1;
 734         }
 735 out:
 736         release_pte_pages(pte, _pte, compound_pagelist);
 737         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 738                                             referenced, writable, result);
 739         return 0;
 740 }
 741
 742 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 743                                       struct vm_area_struct *vma,
 744                                       unsigned long address,
 745                                       spinlock_t *ptl,
 746                                       struct list_head *compound_pagelist)
 747 {
 748         struct page *src_page, *tmp;
 749         pte_t *_pte;
 750         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 751                                 _pte++, page++, address += PAGE_SIZE) {
 752                 pte_t pteval = *_pte;
 753
 754                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 755                         clear_user_highpage(page, address);
 756                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 757                         if (is_zero_pfn(pte_pfn(pteval))) {
 758                                 /*
 759                                  * ptl mostly unnecessary.
 760                                  */
 761                                 spin_lock(ptl);
 762                                 ptep_clear(vma->vm_mm, address, _pte);
 763                                 spin_unlock(ptl);
 764                         }
 765                 } else {
 766                         src_page = pte_page(pteval);
 767                         copy_user_highpage(page, src_page, address, vma);
 768                         if (!PageCompound(src_page))
 769                                 release_pte_page(src_page);
 770                         /*
 771                          * ptl mostly unnecessary, but preempt has to
 772                          * be disabled to update the per-cpu stats
 773                          * inside page_remove_rmap().
 774                          */
 775                         spin_lock(ptl);
 776                         ptep_clear(vma->vm_mm, address, _pte);
 777                         page_remove_rmap(src_page, vma, false);
 778                         spin_unlock(ptl);
 779                         free_page_and_swap_cache(src_page);
 780                 }
 781         }
 782
 783         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 784                 list_del(&src_page->lru);
 785                 release_pte_page(src_page);
 786         }
 787 }
 788
 789 static void khugepaged_alloc_sleep(void)
 790 {
 791         DEFINE_WAIT(wait);
 792
 793         add_wait_queue(&khugepaged_wait, &wait);
 794         freezable_schedule_timeout_interruptible(
 795                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 796         remove_wait_queue(&khugepaged_wait, &wait);
 797 }
 798
 799 static int khugepaged_node_load[MAX_NUMNODES];
 800
 801 static bool khugepaged_scan_abort(int nid)
 802 {
 803         int i;
 804
 805         /*
 806          * If node_reclaim_mode is disabled, then no extra effort is made to
 807          * allocate memory locally.
 808          */
 809         if (!node_reclaim_enabled())
 810                 return false;
 811
 812         /* If there is a count for this node already, it must be acceptable */
 813         if (khugepaged_node_load[nid])
 814                 return false;
 815
 816         for (i = 0; i < MAX_NUMNODES; i++) {
 817                 if (!khugepaged_node_load[i])
 818                         continue;
 819                 if (node_distance(nid, i) > node_reclaim_distance)
 820                         return true;
 821         }
 822         return false;
 823 }
 824
 825 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 826 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 827 {
 828         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 829 }
 830
 831 #ifdef CONFIG_NUMA
 832 static int khugepaged_find_target_node(void)
 833 {
 834         static int last_khugepaged_target_node = NUMA_NO_NODE;
 835         int nid, target_node = 0, max_value = 0;
 836
 837         /* find first node with max normal pages hit */
 838         for (nid = 0; nid < MAX_NUMNODES; nid++)
 839                 if (khugepaged_node_load[nid] > max_value) {
 840                         max_value = khugepaged_node_load[nid];
 841                         target_node = nid;
 842                 }
 843
 844         /* do some balance if several nodes have the same hit record */
 845         if (target_node <= last_khugepaged_target_node)
 846                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 847                                 nid++)
 848                         if (max_value == khugepaged_node_load[nid]) {
 849                                 target_node = nid;
 850                                 break;
 851                         }
 852
 853         last_khugepaged_target_node = target_node;
 854         return target_node;
 855 }
 856
 857 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 858 {
 859         if (IS_ERR(*hpage)) {
 860                 if (!*wait)
 861                         return false;
 862
 863                 *wait = false;
 864                 *hpage = NULL;
 865                 khugepaged_alloc_sleep();
 866         } else if (*hpage) {
 867                 put_page(*hpage);
 868                 *hpage = NULL;
 869         }
 870
 871         return true;
 872 }
 873
 874 static struct page *
 875 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 876 {
 877         VM_BUG_ON_PAGE(*hpage, *hpage);
 878
 879         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 880         if (unlikely(!*hpage)) {
 881                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 882                 *hpage = ERR_PTR(-ENOMEM);
 883                 return NULL;
 884         }
 885
 886         prep_transhuge_page(*hpage);
 887         count_vm_event(THP_COLLAPSE_ALLOC);
 888         return *hpage;
 889 }
 890 #else
 891 static int khugepaged_find_target_node(void)
 892 {
 893         return 0;
 894 }
 895
 896 static inline struct page *alloc_khugepaged_hugepage(void)
 897 {
 898         struct page *page;
 899
 900         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 901                            HPAGE_PMD_ORDER);
 902         if (page)
 903                 prep_transhuge_page(page);
 904         return page;
 905 }
 906
 907 static struct page *khugepaged_alloc_hugepage(bool *wait)
 908 {
 909         struct page *hpage;
 910
 911         do {
 912                 hpage = alloc_khugepaged_hugepage();
 913                 if (!hpage) {
 914                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 915                         if (!*wait)
 916                                 return NULL;
 917
 918                         *wait = false;
 919                         khugepaged_alloc_sleep();
 920                 } else
 921                         count_vm_event(THP_COLLAPSE_ALLOC);
 922         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 923
 924         return hpage;
 925 }
 926
 927 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 928 {
 929         /*
 930          * If the hpage allocated earlier was briefly exposed in page cache
 931          * before collapse_file() failed, it is possible that racing lookups
 932          * have not yet completed, and would then be unpleasantly surprised by
 933          * finding the hpage reused for the same mapping at a different offset.
 934          * Just release the previous allocation if there is any danger of that.
 935          */
 936         if (*hpage && page_count(*hpage) > 1) {
 937                 put_page(*hpage);
 938                 *hpage = NULL;
 939         }
 940
 941         if (!*hpage)
 942                 *hpage = khugepaged_alloc_hugepage(wait);
 943
 944         if (unlikely(!*hpage))
 945                 return false;
 946
 947         return true;
 948 }
 949
 950 static struct page *
 951 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 952 {
 953         VM_BUG_ON(!*hpage);
 954
 955         return  *hpage;
 956 }
 957 #endif
 958
 959 /*
 960  * If mmap_lock temporarily dropped, revalidate vma
 961  * before taking mmap_lock.
 962  * Return 0 if succeeds, otherwise return none-zero
 963  * value (scan code).
 964  */
 965
 966 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 967                 struct vm_area_struct **vmap)
 968 {
 969         struct vm_area_struct *vma;
 970         unsigned long hstart, hend;
 971
 972         if (unlikely(khugepaged_test_exit(mm)))
 973                 return SCAN_ANY_PROCESS;
 974
 975         *vmap = vma = find_vma(mm, address);
 976         if (!vma)
 977                 return SCAN_VMA_NULL;
 978
 979         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 980         hend = vma->vm_end & HPAGE_PMD_MASK;
 981         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 982                 return SCAN_ADDRESS_RANGE;
 983         if (!hugepage_vma_check(vma, vma->vm_flags))
 984                 return SCAN_VMA_CHECK;
 985         /* Anon VMA expected */
 986         if (!vma->anon_vma || vma->vm_ops)
 987                 return SCAN_VMA_CHECK;
 988         return 0;
 989 }
 990
 991 /*
 992  * Bring missing pages in from swap, to complete THP collapse.
 993  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 994  *
 995  * Called and returns without pte mapped or spinlocks held,
 996  * but with mmap_lock held to protect against vma changes.
 997  */
 998
 999 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1000                                         struct vm_area_struct *vma,
1001                                         unsigned long haddr, pmd_t *pmd,
1002                                         int referenced)
1003 {
1004         int swapped_in = 0;
1005         vm_fault_t ret = 0;
1006         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1007
1008         for (address = haddr; address < end; address += PAGE_SIZE) {
1009                 struct vm_fault vmf = {
1010                         .vma = vma,
1011                         .address = address,
1012                         .pgoff = linear_page_index(vma, haddr),
1013                         .flags = FAULT_FLAG_ALLOW_RETRY,
1014                         .pmd = pmd,
1015                 };
1016
1017                 vmf.pte = pte_offset_map(pmd, address);
1018                 vmf.orig_pte = *vmf.pte;
1019                 if (!is_swap_pte(vmf.orig_pte)) {
1020                         pte_unmap(vmf.pte);
1021                         continue;
1022                 }
1023                 swapped_in++;
1024                 ret = do_swap_page(&vmf);
1025
1026                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1027                 if (ret & VM_FAULT_RETRY) {
1028                         mmap_read_lock(mm);
1029                         if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1030                                 /* vma is no longer available, don't continue to swapin */
1031                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1032                                 return false;
1033                         }
1034                         /* check if the pmd is still valid */
1035                         if (mm_find_pmd(mm, haddr) != pmd) {
1036                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1037                                 return false;
1038                         }
1039                 }
1040                 if (ret & VM_FAULT_ERROR) {
1041                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1042                         return false;
1043                 }
1044         }
1045
1046         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1047         if (swapped_in)
1048                 lru_add_drain();
1049
1050         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1051         return true;
1052 }
1053
1054 static void collapse_huge_page(struct mm_struct *mm,
1055                                    unsigned long address,
1056                                    struct page **hpage,
1057                                    int node, int referenced, int unmapped)
1058 {
1059         LIST_HEAD(compound_pagelist);
1060         pmd_t *pmd, _pmd;
1061         pte_t *pte;
1062         pgtable_t pgtable;
1063         struct page *new_page;
1064         spinlock_t *pmd_ptl, *pte_ptl;
1065         int isolated = 0, result = 0;
1066         struct vm_area_struct *vma;
1067         struct mmu_notifier_range range;
1068         gfp_t gfp;
1069
1070         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1071
1072         /* Only allocate from the target node */
1073         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1074
1075         /*
1076          * Before allocating the hugepage, release the mmap_lock read lock.
1077          * The allocation can take potentially a long time if it involves
1078          * sync compaction, and we do not need to hold the mmap_lock during
1079          * that. We will recheck the vma after taking it again in write mode.
1080          */
1081         mmap_read_unlock(mm);
1082         new_page = khugepaged_alloc_page(hpage, gfp, node);
1083         if (!new_page) {
1084                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1085                 goto out_nolock;
1086         }
1087
1088         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1089                 result = SCAN_CGROUP_CHARGE_FAIL;
1090                 goto out_nolock;
1091         }
1092         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1093
1094         mmap_read_lock(mm);
1095         result = hugepage_vma_revalidate(mm, address, &vma);
1096         if (result) {
1097                 mmap_read_unlock(mm);
1098                 goto out_nolock;
1099         }
1100
1101         pmd = mm_find_pmd(mm, address);
1102         if (!pmd) {
1103                 result = SCAN_PMD_NULL;
1104                 mmap_read_unlock(mm);
1105                 goto out_nolock;
1106         }
1107
1108         /*
1109          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1110          * If it fails, we release mmap_lock and jump out_nolock.
1111          * Continuing to collapse causes inconsistency.
1112          */
1113         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1114                                                      pmd, referenced)) {
1115                 mmap_read_unlock(mm);
1116                 goto out_nolock;
1117         }
1118
1119         mmap_read_unlock(mm);
1120         /*
1121          * Prevent all access to pagetables with the exception of
1122          * gup_fast later handled by the ptep_clear_flush and the VM
1123          * handled by the anon_vma lock + PG_lock.
1124          */
1125         mmap_write_lock(mm);
1126         result = hugepage_vma_revalidate(mm, address, &vma);
1127         if (result)
1128                 goto out_up_write;
1129         /* check if the pmd is still valid */
1130         if (mm_find_pmd(mm, address) != pmd)
1131                 goto out_up_write;
1132
1133         anon_vma_lock_write(vma->anon_vma);
1134
1135         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1136                                 address, address + HPAGE_PMD_SIZE);
1137         mmu_notifier_invalidate_range_start(&range);
1138
1139         pte = pte_offset_map(pmd, address);
1140         pte_ptl = pte_lockptr(mm, pmd);
1141
1142         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1143         /*
1144          * After this gup_fast can't run anymore. This also removes
1145          * any huge TLB entry from the CPU so we won't allow
1146          * huge and small TLB entries for the same virtual address
1147          * to avoid the risk of CPU bugs in that area.
1148          */
1149         _pmd = pmdp_collapse_flush(vma, address, pmd);
1150         spin_unlock(pmd_ptl);
1151         mmu_notifier_invalidate_range_end(&range);
1152
1153         spin_lock(pte_ptl);
1154         isolated = __collapse_huge_page_isolate(vma, address, pte,
1155                         &compound_pagelist);
1156         spin_unlock(pte_ptl);
1157
1158         if (unlikely(!isolated)) {
1159                 pte_unmap(pte);
1160                 spin_lock(pmd_ptl);
1161                 BUG_ON(!pmd_none(*pmd));
1162                 /*
1163                  * We can only use set_pmd_at when establishing
1164                  * hugepmds and never for establishing regular pmds that
1165                  * points to regular pagetables. Use pmd_populate for that
1166                  */
1167                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1168                 spin_unlock(pmd_ptl);
1169                 anon_vma_unlock_write(vma->anon_vma);
1170                 result = SCAN_FAIL;
1171                 goto out_up_write;
1172         }
1173
1174         /*
1175          * All pages are isolated and locked so anon_vma rmap
1176          * can't run anymore.
1177          */
1178         anon_vma_unlock_write(vma->anon_vma);
1179
1180         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1181                         &compound_pagelist);
1182         pte_unmap(pte);
1183         /*
1184          * spin_lock() below is not the equivalent of smp_wmb(), but
1185          * the smp_wmb() inside __SetPageUptodate() can be reused to
1186          * avoid the copy_huge_page writes to become visible after
1187          * the set_pmd_at() write.
1188          */
1189         __SetPageUptodate(new_page);
1190         pgtable = pmd_pgtable(_pmd);
1191
1192         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1193         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1194
1195         spin_lock(pmd_ptl);
1196         BUG_ON(!pmd_none(*pmd));
1197         page_add_new_anon_rmap(new_page, vma, address, true);
1198         lru_cache_add_inactive_or_unevictable(new_page, vma);
1199         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1200         set_pmd_at(mm, address, pmd, _pmd);
1201         update_mmu_cache_pmd(vma, address, pmd);
1202         spin_unlock(pmd_ptl);
1203
1204         *hpage = NULL;
1205
1206         khugepaged_pages_collapsed++;
1207         result = SCAN_SUCCEED;
1208 out_up_write:
1209         mmap_write_unlock(mm);
1210 out_nolock:
1211         if (!IS_ERR_OR_NULL(*hpage))
1212                 mem_cgroup_uncharge(page_folio(*hpage));
1213         trace_mm_collapse_huge_page(mm, isolated, result);
1214         return;
1215 }
1216
1217 static int khugepaged_scan_pmd(struct mm_struct *mm,
1218                                struct vm_area_struct *vma,
1219                                unsigned long address,
1220                                struct page **hpage)
1221 {
1222         pmd_t *pmd;
1223         pte_t *pte, *_pte;
1224         int ret = 0, result = 0, referenced = 0;
1225         int none_or_zero = 0, shared = 0;
1226         struct page *page = NULL;
1227         unsigned long _address;
1228         spinlock_t *ptl;
1229         int node = NUMA_NO_NODE, unmapped = 0;
1230         bool writable = false;
1231
1232         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1233
1234         pmd = mm_find_pmd(mm, address);
1235         if (!pmd) {
1236                 result = SCAN_PMD_NULL;
1237                 goto out;
1238         }
1239
1240         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1241         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1242         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1243              _pte++, _address += PAGE_SIZE) {
1244                 pte_t pteval = *_pte;
1245                 if (is_swap_pte(pteval)) {
1246                         if (++unmapped <= khugepaged_max_ptes_swap) {
1247                                 /*
1248                                  * Always be strict with uffd-wp
1249                                  * enabled swap entries.  Please see
1250                                  * comment below for pte_uffd_wp().
1251                                  */
1252                                 if (pte_swp_uffd_wp(pteval)) {
1253                                         result = SCAN_PTE_UFFD_WP;
1254                                         goto out_unmap;
1255                                 }
1256                                 continue;
1257                         } else {
1258                                 result = SCAN_EXCEED_SWAP_PTE;
1259                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1260                                 goto out_unmap;
1261                         }
1262                 }
1263                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1264                         if (!userfaultfd_armed(vma) &&
1265                             ++none_or_zero <= khugepaged_max_ptes_none) {
1266                                 continue;
1267                         } else {
1268                                 result = SCAN_EXCEED_NONE_PTE;
1269                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1270                                 goto out_unmap;
1271                         }
1272                 }
1273                 if (pte_uffd_wp(pteval)) {
1274                         /*
1275                          * Don't collapse the page if any of the small
1276                          * PTEs are armed with uffd write protection.
1277                          * Here we can also mark the new huge pmd as
1278                          * write protected if any of the small ones is
1279                          * marked but that could bring unknown
1280                          * userfault messages that falls outside of
1281                          * the registered range.  So, just be simple.
1282                          */
1283                         result = SCAN_PTE_UFFD_WP;
1284                         goto out_unmap;
1285                 }
1286                 if (pte_write(pteval))
1287                         writable = true;
1288
1289                 page = vm_normal_page(vma, _address, pteval);
1290                 if (unlikely(!page)) {
1291                         result = SCAN_PAGE_NULL;
1292                         goto out_unmap;
1293                 }
1294
1295                 if (page_mapcount(page) > 1 &&
1296                                 ++shared > khugepaged_max_ptes_shared) {
1297                         result = SCAN_EXCEED_SHARED_PTE;
1298                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1299                         goto out_unmap;
1300                 }
1301
1302                 page = compound_head(page);
1303
1304                 /*
1305                  * Record which node the original page is from and save this
1306                  * information to khugepaged_node_load[].
1307                  * Khugepaged will allocate hugepage from the node has the max
1308                  * hit record.
1309                  */
1310                 node = page_to_nid(page);
1311                 if (khugepaged_scan_abort(node)) {
1312                         result = SCAN_SCAN_ABORT;
1313                         goto out_unmap;
1314                 }
1315                 khugepaged_node_load[node]++;
1316                 if (!PageLRU(page)) {
1317                         result = SCAN_PAGE_LRU;
1318                         goto out_unmap;
1319                 }
1320                 if (PageLocked(page)) {
1321                         result = SCAN_PAGE_LOCK;
1322                         goto out_unmap;
1323                 }
1324                 if (!PageAnon(page)) {
1325                         result = SCAN_PAGE_ANON;
1326                         goto out_unmap;
1327                 }
1328
1329                 /*
1330                  * Check if the page has any GUP (or other external) pins.
1331                  *
1332                  * Here the check is racy it may see totmal_mapcount > refcount
1333                  * in some cases.
1334                  * For example, one process with one forked child process.
1335                  * The parent has the PMD split due to MADV_DONTNEED, then
1336                  * the child is trying unmap the whole PMD, but khugepaged
1337                  * may be scanning the parent between the child has
1338                  * PageDoubleMap flag cleared and dec the mapcount.  So
1339                  * khugepaged may see total_mapcount > refcount.
1340                  *
1341                  * But such case is ephemeral we could always retry collapse
1342                  * later.  However it may report false positive if the page
1343                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1344                  * will be done again later the risk seems low.
1345                  */
1346                 if (!is_refcount_suitable(page)) {
1347                         result = SCAN_PAGE_COUNT;
1348                         goto out_unmap;
1349                 }
1350                 if (pte_young(pteval) ||
1351                     page_is_young(page) || PageReferenced(page) ||
1352                     mmu_notifier_test_young(vma->vm_mm, address))
1353                         referenced++;
1354         }
1355         if (!writable) {
1356                 result = SCAN_PAGE_RO;
1357         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1358                 result = SCAN_LACK_REFERENCED_PAGE;
1359         } else {
1360                 result = SCAN_SUCCEED;
1361                 ret = 1;
1362         }
1363 out_unmap:
1364         pte_unmap_unlock(pte, ptl);
1365         if (ret) {
1366                 node = khugepaged_find_target_node();
1367                 /* collapse_huge_page will return with the mmap_lock released */
1368                 collapse_huge_page(mm, address, hpage, node,
1369                                 referenced, unmapped);
1370         }
1371 out:
1372         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1373                                      none_or_zero, result, unmapped);
1374         return ret;
1375 }
1376
1377 static void collect_mm_slot(struct mm_slot *mm_slot)
1378 {
1379         struct mm_struct *mm = mm_slot->mm;
1380
1381         lockdep_assert_held(&khugepaged_mm_lock);
1382
1383         if (khugepaged_test_exit(mm)) {
1384                 /* free mm_slot */
1385                 hash_del(&mm_slot->hash);
1386                 list_del(&mm_slot->mm_node);
1387
1388                 /*
1389                  * Not strictly needed because the mm exited already.
1390                  *
1391                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1392                  */
1393
1394                 /* khugepaged_mm_lock actually not necessary for the below */
1395                 free_mm_slot(mm_slot);
1396                 mmdrop(mm);
1397         }
1398 }
1399
1400 #ifdef CONFIG_SHMEM
1401 /*
1402  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1403  * khugepaged should try to collapse the page table.
1404  */
1405 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1406                                          unsigned long addr)
1407 {
1408         struct mm_slot *mm_slot;
1409
1410         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1411
1412         spin_lock(&khugepaged_mm_lock);
1413         mm_slot = get_mm_slot(mm);
1414         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1415                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1416         spin_unlock(&khugepaged_mm_lock);
1417         return 0;
1418 }
1419
1420 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1421                                   unsigned long addr, pmd_t *pmdp)
1422 {
1423         spinlock_t *ptl;
1424         pmd_t pmd;
1425
1426         mmap_assert_write_locked(mm);
1427         ptl = pmd_lock(vma->vm_mm, pmdp);
1428         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1429         spin_unlock(ptl);
1430         mm_dec_nr_ptes(mm);
1431         page_table_check_pte_clear_range(mm, addr, pmd);
1432         pte_free(mm, pmd_pgtable(pmd));
1433 }
1434
1435 /**
1436  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1437  * address haddr.
1438  *
1439  * @mm: process address space where collapse happens
1440  * @addr: THP collapse address
1441  *
1442  * This function checks whether all the PTEs in the PMD are pointing to the
1443  * right THP. If so, retract the page table so the THP can refault in with
1444  * as pmd-mapped.
1445  */
1446 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1447 {
1448         unsigned long haddr = addr & HPAGE_PMD_MASK;
1449         struct vm_area_struct *vma = find_vma(mm, haddr);
1450         struct page *hpage;
1451         pte_t *start_pte, *pte;
1452         pmd_t *pmd;
1453         spinlock_t *ptl;
1454         int count = 0;
1455         int i;
1456
1457         if (!vma || !vma->vm_file ||
1458             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1459                 return;
1460
1461         /*
1462          * This vm_flags may not have VM_HUGEPAGE if the page was not
1463          * collapsed by this mm. But we can still collapse if the page is
1464          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1465          * will not fail the vma for missing VM_HUGEPAGE
1466          */
1467         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1468                 return;
1469
1470         hpage = find_lock_page(vma->vm_file->f_mapping,
1471                                linear_page_index(vma, haddr));
1472         if (!hpage)
1473                 return;
1474
1475         if (!PageHead(hpage))
1476                 goto drop_hpage;
1477
1478         pmd = mm_find_pmd(mm, haddr);
1479         if (!pmd)
1480                 goto drop_hpage;
1481
1482         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1483
1484         /* step 1: check all mapped PTEs are to the right huge page */
1485         for (i = 0, addr = haddr, pte = start_pte;
1486              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1487                 struct page *page;
1488
1489                 /* empty pte, skip */
1490                 if (pte_none(*pte))
1491                         continue;
1492
1493                 /* page swapped out, abort */
1494                 if (!pte_present(*pte))
1495                         goto abort;
1496
1497                 page = vm_normal_page(vma, addr, *pte);
1498
1499                 /*
1500                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1501                  * page table, but the new page will not be a subpage of hpage.
1502                  */
1503                 if (hpage + i != page)
1504                         goto abort;
1505                 count++;
1506         }
1507
1508         /* step 2: adjust rmap */
1509         for (i = 0, addr = haddr, pte = start_pte;
1510              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1511                 struct page *page;
1512
1513                 if (pte_none(*pte))
1514                         continue;
1515                 page = vm_normal_page(vma, addr, *pte);
1516                 page_remove_rmap(page, vma, false);
1517         }
1518
1519         pte_unmap_unlock(start_pte, ptl);
1520
1521         /* step 3: set proper refcount and mm_counters. */
1522         if (count) {
1523                 page_ref_sub(hpage, count);
1524                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1525         }
1526
1527         /* step 4: collapse pmd */
1528         collapse_and_free_pmd(mm, vma, haddr, pmd);
1529 drop_hpage:
1530         unlock_page(hpage);
1531         put_page(hpage);
1532         return;
1533
1534 abort:
1535         pte_unmap_unlock(start_pte, ptl);
1536         goto drop_hpage;
1537 }
1538
1539 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1540 {
1541         struct mm_struct *mm = mm_slot->mm;
1542         int i;
1543
1544         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1545                 return;
1546
1547         if (!mmap_write_trylock(mm))
1548                 return;
1549
1550         if (unlikely(khugepaged_test_exit(mm)))
1551                 goto out;
1552
1553         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1554                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1555
1556 out:
1557         mm_slot->nr_pte_mapped_thp = 0;
1558         mmap_write_unlock(mm);
1559 }
1560
1561 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1562 {
1563         struct vm_area_struct *vma;
1564         struct mm_struct *mm;
1565         unsigned long addr;
1566         pmd_t *pmd;
1567
1568         i_mmap_lock_write(mapping);
1569         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1570                 /*
1571                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1572                  * got written to. These VMAs are likely not worth investing
1573                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1574                  * later.
1575                  *
1576                  * Not that vma->anon_vma check is racy: it can be set up after
1577                  * the check but before we took mmap_lock by the fault path.
1578                  * But page lock would prevent establishing any new ptes of the
1579                  * page, so we are safe.
1580                  *
1581                  * An alternative would be drop the check, but check that page
1582                  * table is clear before calling pmdp_collapse_flush() under
1583                  * ptl. It has higher chance to recover THP for the VMA, but
1584                  * has higher cost too.
1585                  */
1586                 if (vma->anon_vma)
1587                         continue;
1588                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1589                 if (addr & ~HPAGE_PMD_MASK)
1590                         continue;
1591                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1592                         continue;
1593                 mm = vma->vm_mm;
1594                 pmd = mm_find_pmd(mm, addr);
1595                 if (!pmd)
1596                         continue;
1597                 /*
1598                  * We need exclusive mmap_lock to retract page table.
1599                  *
1600                  * We use trylock due to lock inversion: we need to acquire
1601                  * mmap_lock while holding page lock. Fault path does it in
1602                  * reverse order. Trylock is a way to avoid deadlock.
1603                  */
1604                 if (mmap_write_trylock(mm)) {
1605                         if (!khugepaged_test_exit(mm))
1606                                 collapse_and_free_pmd(mm, vma, addr, pmd);
1607                         mmap_write_unlock(mm);
1608                 } else {
1609                         /* Try again later */
1610                         khugepaged_add_pte_mapped_thp(mm, addr);
1611                 }
1612         }
1613         i_mmap_unlock_write(mapping);
1614 }
1615
1616 /**
1617  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1618  *
1619  * @mm: process address space where collapse happens
1620  * @file: file that collapse on
1621  * @start: collapse start address
1622  * @hpage: new allocated huge page for collapse
1623  * @node: appointed node the new huge page allocate from
1624  *
1625  * Basic scheme is simple, details are more complex:
1626  *  - allocate and lock a new huge page;
1627  *  - scan page cache replacing old pages with the new one
1628  *    + swap/gup in pages if necessary;
1629  *    + fill in gaps;
1630  *    + keep old pages around in case rollback is required;
1631  *  - if replacing succeeds:
1632  *    + copy data over;
1633  *    + free old pages;
1634  *    + unlock huge page;
1635  *  - if replacing failed;
1636  *    + put all pages back and unfreeze them;
1637  *    + restore gaps in the page cache;
1638  *    + unlock and free huge page;
1639  */
1640 static void collapse_file(struct mm_struct *mm,
1641                 struct file *file, pgoff_t start,
1642                 struct page **hpage, int node)
1643 {
1644         struct address_space *mapping = file->f_mapping;
1645         gfp_t gfp;
1646         struct page *new_page;
1647         pgoff_t index, end = start + HPAGE_PMD_NR;
1648         LIST_HEAD(pagelist);
1649         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1650         int nr_none = 0, result = SCAN_SUCCEED;
1651         bool is_shmem = shmem_file(file);
1652         int nr;
1653
1654         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1655         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1656
1657         /* Only allocate from the target node */
1658         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1659
1660         new_page = khugepaged_alloc_page(hpage, gfp, node);
1661         if (!new_page) {
1662                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1663                 goto out;
1664         }
1665
1666         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1667                 result = SCAN_CGROUP_CHARGE_FAIL;
1668                 goto out;
1669         }
1670         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1671
1672         /*
1673          * Ensure we have slots for all the pages in the range.  This is
1674          * almost certainly a no-op because most of the pages must be present
1675          */
1676         do {
1677                 xas_lock_irq(&xas);
1678                 xas_create_range(&xas);
1679                 if (!xas_error(&xas))
1680                         break;
1681                 xas_unlock_irq(&xas);
1682                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1683                         result = SCAN_FAIL;
1684                         goto out;
1685                 }
1686         } while (1);
1687
1688         __SetPageLocked(new_page);
1689         if (is_shmem)
1690                 __SetPageSwapBacked(new_page);
1691         new_page->index = start;
1692         new_page->mapping = mapping;
1693
1694         /*
1695          * At this point the new_page is locked and not up-to-date.
1696          * It's safe to insert it into the page cache, because nobody would
1697          * be able to map it or use it in another way until we unlock it.
1698          */
1699
1700         xas_set(&xas, start);
1701         for (index = start; index < end; index++) {
1702                 struct page *page = xas_next(&xas);
1703
1704                 VM_BUG_ON(index != xas.xa_index);
1705                 if (is_shmem) {
1706                         if (!page) {
1707                                 /*
1708                                  * Stop if extent has been truncated or
1709                                  * hole-punched, and is now completely
1710                                  * empty.
1711                                  */
1712                                 if (index == start) {
1713                                         if (!xas_next_entry(&xas, end - 1)) {
1714                                                 result = SCAN_TRUNCATED;
1715                                                 goto xa_locked;
1716                                         }
1717                                         xas_set(&xas, index);
1718                                 }
1719                                 if (!shmem_charge(mapping->host, 1)) {
1720                                         result = SCAN_FAIL;
1721                                         goto xa_locked;
1722                                 }
1723                                 xas_store(&xas, new_page);
1724                                 nr_none++;
1725                                 continue;
1726                         }
1727
1728                         if (xa_is_value(page) || !PageUptodate(page)) {
1729                                 xas_unlock_irq(&xas);
1730                                 /* swap in or instantiate fallocated page */
1731                                 if (shmem_getpage(mapping->host, index, &page,
1732                                                   SGP_NOALLOC)) {
1733                                         result = SCAN_FAIL;
1734                                         goto xa_unlocked;
1735                                 }
1736                         } else if (trylock_page(page)) {
1737                                 get_page(page);
1738                                 xas_unlock_irq(&xas);
1739                         } else {
1740                                 result = SCAN_PAGE_LOCK;
1741                                 goto xa_locked;
1742                         }
1743                 } else {        /* !is_shmem */
1744                         if (!page || xa_is_value(page)) {
1745                                 xas_unlock_irq(&xas);
1746                                 page_cache_sync_readahead(mapping, &file->f_ra,
1747                                                           file, index,
1748                                                           end - index);
1749                                 /* drain pagevecs to help isolate_lru_page() */
1750                                 lru_add_drain();
1751                                 page = find_lock_page(mapping, index);
1752                                 if (unlikely(page == NULL)) {
1753                                         result = SCAN_FAIL;
1754                                         goto xa_unlocked;
1755                                 }
1756                         } else if (PageDirty(page)) {
1757                                 /*
1758                                  * khugepaged only works on read-only fd,
1759                                  * so this page is dirty because it hasn't
1760                                  * been flushed since first write. There
1761                                  * won't be new dirty pages.
1762                                  *
1763                                  * Trigger async flush here and hope the
1764                                  * writeback is done when khugepaged
1765                                  * revisits this page.
1766                                  *
1767                                  * This is a one-off situation. We are not
1768                                  * forcing writeback in loop.
1769                                  */
1770                                 xas_unlock_irq(&xas);
1771                                 filemap_flush(mapping);
1772                                 result = SCAN_FAIL;
1773                                 goto xa_unlocked;
1774                         } else if (PageWriteback(page)) {
1775                                 xas_unlock_irq(&xas);
1776                                 result = SCAN_FAIL;
1777                                 goto xa_unlocked;
1778                         } else if (trylock_page(page)) {
1779                                 get_page(page);
1780                                 xas_unlock_irq(&xas);
1781                         } else {
1782                                 result = SCAN_PAGE_LOCK;
1783                                 goto xa_locked;
1784                         }
1785                 }
1786
1787                 /*
1788                  * The page must be locked, so we can drop the i_pages lock
1789                  * without racing with truncate.
1790                  */
1791                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1792
1793                 /* make sure the page is up to date */
1794                 if (unlikely(!PageUptodate(page))) {
1795                         result = SCAN_FAIL;
1796                         goto out_unlock;
1797                 }
1798
1799                 /*
1800                  * If file was truncated then extended, or hole-punched, before
1801                  * we locked the first page, then a THP might be there already.
1802                  */
1803                 if (PageTransCompound(page)) {
1804                         result = SCAN_PAGE_COMPOUND;
1805                         goto out_unlock;
1806                 }
1807
1808                 if (page_mapping(page) != mapping) {
1809                         result = SCAN_TRUNCATED;
1810                         goto out_unlock;
1811                 }
1812
1813                 if (!is_shmem && (PageDirty(page) ||
1814                                   PageWriteback(page))) {
1815                         /*
1816                          * khugepaged only works on read-only fd, so this
1817                          * page is dirty because it hasn't been flushed
1818                          * since first write.
1819                          */
1820                         result = SCAN_FAIL;
1821                         goto out_unlock;
1822                 }
1823
1824                 if (isolate_lru_page(page)) {
1825                         result = SCAN_DEL_PAGE_LRU;
1826                         goto out_unlock;
1827                 }
1828
1829                 if (page_has_private(page) &&
1830                     !try_to_release_page(page, GFP_KERNEL)) {
1831                         result = SCAN_PAGE_HAS_PRIVATE;
1832                         putback_lru_page(page);
1833                         goto out_unlock;
1834                 }
1835
1836                 if (page_mapped(page))
1837                         unmap_mapping_pages(mapping, index, 1, false);
1838
1839                 xas_lock_irq(&xas);
1840                 xas_set(&xas, index);
1841
1842                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1843                 VM_BUG_ON_PAGE(page_mapped(page), page);
1844
1845                 /*
1846                  * The page is expected to have page_count() == 3:
1847                  *  - we hold a pin on it;
1848                  *  - one reference from page cache;
1849                  *  - one from isolate_lru_page;
1850                  */
1851                 if (!page_ref_freeze(page, 3)) {
1852                         result = SCAN_PAGE_COUNT;
1853                         xas_unlock_irq(&xas);
1854                         putback_lru_page(page);
1855                         goto out_unlock;
1856                 }
1857
1858                 /*
1859                  * Add the page to the list to be able to undo the collapse if
1860                  * something go wrong.
1861                  */
1862                 list_add_tail(&page->lru, &pagelist);
1863
1864                 /* Finally, replace with the new page. */
1865                 xas_store(&xas, new_page);
1866                 continue;
1867 out_unlock:
1868                 unlock_page(page);
1869                 put_page(page);
1870                 goto xa_unlocked;
1871         }
1872         nr = thp_nr_pages(new_page);
1873
1874         if (is_shmem)
1875                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1876         else {
1877                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1878                 filemap_nr_thps_inc(mapping);
1879                 /*
1880                  * Paired with smp_mb() in do_dentry_open() to ensure
1881                  * i_writecount is up to date and the update to nr_thps is
1882                  * visible. Ensures the page cache will be truncated if the
1883                  * file is opened writable.
1884                  */
1885                 smp_mb();
1886                 if (inode_is_open_for_write(mapping->host)) {
1887                         result = SCAN_FAIL;
1888                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1889                         filemap_nr_thps_dec(mapping);
1890                         goto xa_locked;
1891                 }
1892         }
1893
1894         if (nr_none) {
1895                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1896                 if (is_shmem)
1897                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1898         }
1899
1900         /* Join all the small entries into a single multi-index entry */
1901         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1902         xas_store(&xas, new_page);
1903 xa_locked:
1904         xas_unlock_irq(&xas);
1905 xa_unlocked:
1906
1907         if (result == SCAN_SUCCEED) {
1908                 struct page *page, *tmp;
1909
1910                 /*
1911                  * Replacing old pages with new one has succeeded, now we
1912                  * need to copy the content and free the old pages.
1913                  */
1914                 index = start;
1915                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1916                         while (index < page->index) {
1917                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1918                                 index++;
1919                         }
1920                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1921                                         page);
1922                         list_del(&page->lru);
1923                         page->mapping = NULL;
1924                         page_ref_unfreeze(page, 1);
1925                         ClearPageActive(page);
1926                         ClearPageUnevictable(page);
1927                         unlock_page(page);
1928                         put_page(page);
1929                         index++;
1930                 }
1931                 while (index < end) {
1932                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1933                         index++;
1934                 }
1935
1936                 SetPageUptodate(new_page);
1937                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1938                 if (is_shmem)
1939                         set_page_dirty(new_page);
1940                 lru_cache_add(new_page);
1941
1942                 /*
1943                  * Remove pte page tables, so we can re-fault the page as huge.
1944                  */
1945                 retract_page_tables(mapping, start);
1946                 *hpage = NULL;
1947
1948                 khugepaged_pages_collapsed++;
1949         } else {
1950                 struct page *page;
1951
1952                 /* Something went wrong: roll back page cache changes */
1953                 xas_lock_irq(&xas);
1954                 mapping->nrpages -= nr_none;
1955
1956                 if (is_shmem)
1957                         shmem_uncharge(mapping->host, nr_none);
1958
1959                 xas_set(&xas, start);
1960                 xas_for_each(&xas, page, end - 1) {
1961                         page = list_first_entry_or_null(&pagelist,
1962                                         struct page, lru);
1963                         if (!page || xas.xa_index < page->index) {
1964                                 if (!nr_none)
1965                                         break;
1966                                 nr_none--;
1967                                 /* Put holes back where they were */
1968                                 xas_store(&xas, NULL);
1969                                 continue;
1970                         }
1971
1972                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1973
1974                         /* Unfreeze the page. */
1975                         list_del(&page->lru);
1976                         page_ref_unfreeze(page, 2);
1977                         xas_store(&xas, page);
1978                         xas_pause(&xas);
1979                         xas_unlock_irq(&xas);
1980                         unlock_page(page);
1981                         putback_lru_page(page);
1982                         xas_lock_irq(&xas);
1983                 }
1984                 VM_BUG_ON(nr_none);
1985                 xas_unlock_irq(&xas);
1986
1987                 new_page->mapping = NULL;
1988         }
1989
1990         unlock_page(new_page);
1991 out:
1992         VM_BUG_ON(!list_empty(&pagelist));
1993         if (!IS_ERR_OR_NULL(*hpage))
1994                 mem_cgroup_uncharge(page_folio(*hpage));
1995         /* TODO: tracepoints */
1996 }
1997
1998 static void khugepaged_scan_file(struct mm_struct *mm,
1999                 struct file *file, pgoff_t start, struct page **hpage)
2000 {
2001         struct page *page = NULL;
2002         struct address_space *mapping = file->f_mapping;
2003         XA_STATE(xas, &mapping->i_pages, start);
2004         int present, swap;
2005         int node = NUMA_NO_NODE;
2006         int result = SCAN_SUCCEED;
2007
2008         present = 0;
2009         swap = 0;
2010         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2011         rcu_read_lock();
2012         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2013                 if (xas_retry(&xas, page))
2014                         continue;
2015
2016                 if (xa_is_value(page)) {
2017                         if (++swap > khugepaged_max_ptes_swap) {
2018                                 result = SCAN_EXCEED_SWAP_PTE;
2019                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2020                                 break;
2021                         }
2022                         continue;
2023                 }
2024
2025                 /*
2026                  * XXX: khugepaged should compact smaller compound pages
2027                  * into a PMD sized page
2028                  */
2029                 if (PageTransCompound(page)) {
2030                         result = SCAN_PAGE_COMPOUND;
2031                         break;
2032                 }
2033
2034                 node = page_to_nid(page);
2035                 if (khugepaged_scan_abort(node)) {
2036                         result = SCAN_SCAN_ABORT;
2037                         break;
2038                 }
2039                 khugepaged_node_load[node]++;
2040
2041                 if (!PageLRU(page)) {
2042                         result = SCAN_PAGE_LRU;
2043                         break;
2044                 }
2045
2046                 if (page_count(page) !=
2047                     1 + page_mapcount(page) + page_has_private(page)) {
2048                         result = SCAN_PAGE_COUNT;
2049                         break;
2050                 }
2051
2052                 /*
2053                  * We probably should check if the page is referenced here, but
2054                  * nobody would transfer pte_young() to PageReferenced() for us.
2055                  * And rmap walk here is just too costly...
2056                  */
2057
2058                 present++;
2059
2060                 if (need_resched()) {
2061                         xas_pause(&xas);
2062                         cond_resched_rcu();
2063                 }
2064         }
2065         rcu_read_unlock();
2066
2067         if (result == SCAN_SUCCEED) {
2068                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2069                         result = SCAN_EXCEED_NONE_PTE;
2070                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2071                 } else {
2072                         node = khugepaged_find_target_node();
2073                         collapse_file(mm, file, start, hpage, node);
2074                 }
2075         }
2076
2077         /* TODO: tracepoints */
2078 }
2079 #else
2080 static void khugepaged_scan_file(struct mm_struct *mm,
2081                 struct file *file, pgoff_t start, struct page **hpage)
2082 {
2083         BUILD_BUG();
2084 }
2085
2086 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2087 {
2088 }
2089 #endif
2090
2091 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2092                                             struct page **hpage)
2093         __releases(&khugepaged_mm_lock)
2094         __acquires(&khugepaged_mm_lock)
2095 {
2096         struct mm_slot *mm_slot;
2097         struct mm_struct *mm;
2098         struct vm_area_struct *vma;
2099         int progress = 0;
2100
2101         VM_BUG_ON(!pages);
2102         lockdep_assert_held(&khugepaged_mm_lock);
2103
2104         if (khugepaged_scan.mm_slot)
2105                 mm_slot = khugepaged_scan.mm_slot;
2106         else {
2107                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2108                                      struct mm_slot, mm_node);
2109                 khugepaged_scan.address = 0;
2110                 khugepaged_scan.mm_slot = mm_slot;
2111         }
2112         spin_unlock(&khugepaged_mm_lock);
2113         khugepaged_collapse_pte_mapped_thps(mm_slot);
2114
2115         mm = mm_slot->mm;
2116         /*
2117          * Don't wait for semaphore (to avoid long wait times).  Just move to
2118          * the next mm on the list.
2119          */
2120         vma = NULL;
2121         if (unlikely(!mmap_read_trylock(mm)))
2122                 goto breakouterloop_mmap_lock;
2123         if (likely(!khugepaged_test_exit(mm)))
2124                 vma = find_vma(mm, khugepaged_scan.address);
2125
2126         progress++;
2127         for (; vma; vma = vma->vm_next) {
2128                 unsigned long hstart, hend;
2129
2130                 cond_resched();
2131                 if (unlikely(khugepaged_test_exit(mm))) {
2132                         progress++;
2133                         break;
2134                 }
2135                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2136 skip:
2137                         progress++;
2138                         continue;
2139                 }
2140                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2141                 hend = vma->vm_end & HPAGE_PMD_MASK;
2142                 if (hstart >= hend)
2143                         goto skip;
2144                 if (khugepaged_scan.address > hend)
2145                         goto skip;
2146                 if (khugepaged_scan.address < hstart)
2147                         khugepaged_scan.address = hstart;
2148                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2149                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2150                         goto skip;
2151
2152                 while (khugepaged_scan.address < hend) {
2153                         int ret;
2154                         cond_resched();
2155                         if (unlikely(khugepaged_test_exit(mm)))
2156                                 goto breakouterloop;
2157
2158                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2159                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2160                                   hend);
2161                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2162                                 struct file *file = get_file(vma->vm_file);
2163                                 pgoff_t pgoff = linear_page_index(vma,
2164                                                 khugepaged_scan.address);
2165
2166                                 mmap_read_unlock(mm);
2167                                 ret = 1;
2168                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2169                                 fput(file);
2170                         } else {
2171                                 ret = khugepaged_scan_pmd(mm, vma,
2172                                                 khugepaged_scan.address,
2173                                                 hpage);
2174                         }
2175                         /* move to next address */
2176                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2177                         progress += HPAGE_PMD_NR;
2178                         if (ret)
2179                                 /* we released mmap_lock so break loop */
2180                                 goto breakouterloop_mmap_lock;
2181                         if (progress >= pages)
2182                                 goto breakouterloop;
2183                 }
2184         }
2185 breakouterloop:
2186         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2187 breakouterloop_mmap_lock:
2188
2189         spin_lock(&khugepaged_mm_lock);
2190         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2191         /*
2192          * Release the current mm_slot if this mm is about to die, or
2193          * if we scanned all vmas of this mm.
2194          */
2195         if (khugepaged_test_exit(mm) || !vma) {
2196                 /*
2197                  * Make sure that if mm_users is reaching zero while
2198                  * khugepaged runs here, khugepaged_exit will find
2199                  * mm_slot not pointing to the exiting mm.
2200                  */
2201                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2202                         khugepaged_scan.mm_slot = list_entry(
2203                                 mm_slot->mm_node.next,
2204                                 struct mm_slot, mm_node);
2205                         khugepaged_scan.address = 0;
2206                 } else {
2207                         khugepaged_scan.mm_slot = NULL;
2208                         khugepaged_full_scans++;
2209                 }
2210
2211                 collect_mm_slot(mm_slot);
2212         }
2213
2214         return progress;
2215 }
2216
2217 static int khugepaged_has_work(void)
2218 {
2219         return !list_empty(&khugepaged_scan.mm_head) &&
2220                 khugepaged_enabled();
2221 }
2222
2223 static int khugepaged_wait_event(void)
2224 {
2225         return !list_empty(&khugepaged_scan.mm_head) ||
2226                 kthread_should_stop();
2227 }
2228
2229 static void khugepaged_do_scan(void)
2230 {
2231         struct page *hpage = NULL;
2232         unsigned int progress = 0, pass_through_head = 0;
2233         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2234         bool wait = true;
2235
2236         lru_add_drain_all();
2237
2238         while (progress < pages) {
2239                 if (!khugepaged_prealloc_page(&hpage, &wait))
2240                         break;
2241
2242                 cond_resched();
2243
2244                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2245                         break;
2246
2247                 spin_lock(&khugepaged_mm_lock);
2248                 if (!khugepaged_scan.mm_slot)
2249                         pass_through_head++;
2250                 if (khugepaged_has_work() &&
2251                     pass_through_head < 2)
2252                         progress += khugepaged_scan_mm_slot(pages - progress,
2253                                                             &hpage);
2254                 else
2255                         progress = pages;
2256                 spin_unlock(&khugepaged_mm_lock);
2257         }
2258
2259         if (!IS_ERR_OR_NULL(hpage))
2260                 put_page(hpage);
2261 }
2262
2263 static bool khugepaged_should_wakeup(void)
2264 {
2265         return kthread_should_stop() ||
2266                time_after_eq(jiffies, khugepaged_sleep_expire);
2267 }
2268
2269 static void khugepaged_wait_work(void)
2270 {
2271         if (khugepaged_has_work()) {
2272                 const unsigned long scan_sleep_jiffies =
2273                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2274
2275                 if (!scan_sleep_jiffies)
2276                         return;
2277
2278                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2279                 wait_event_freezable_timeout(khugepaged_wait,
2280                                              khugepaged_should_wakeup(),
2281                                              scan_sleep_jiffies);
2282                 return;
2283         }
2284
2285         if (khugepaged_enabled())
2286                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2287 }
2288
2289 static int khugepaged(void *none)
2290 {
2291         struct mm_slot *mm_slot;
2292
2293         set_freezable();
2294         set_user_nice(current, MAX_NICE);
2295
2296         while (!kthread_should_stop()) {
2297                 khugepaged_do_scan();
2298                 khugepaged_wait_work();
2299         }
2300
2301         spin_lock(&khugepaged_mm_lock);
2302         mm_slot = khugepaged_scan.mm_slot;
2303         khugepaged_scan.mm_slot = NULL;
2304         if (mm_slot)
2305                 collect_mm_slot(mm_slot);
2306         spin_unlock(&khugepaged_mm_lock);
2307         return 0;
2308 }
2309
2310 static void set_recommended_min_free_kbytes(void)
2311 {
2312         struct zone *zone;
2313         int nr_zones = 0;
2314         unsigned long recommended_min;
2315
2316         if (!khugepaged_enabled()) {
2317                 calculate_min_free_kbytes();
2318                 goto update_wmarks;
2319         }
2320
2321         for_each_populated_zone(zone) {
2322                 /*
2323                  * We don't need to worry about fragmentation of
2324                  * ZONE_MOVABLE since it only has movable pages.
2325                  */
2326                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2327                         continue;
2328
2329                 nr_zones++;
2330         }
2331
2332         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2333         recommended_min = pageblock_nr_pages * nr_zones * 2;
2334
2335         /*
2336          * Make sure that on average at least two pageblocks are almost free
2337          * of another type, one for a migratetype to fall back to and a
2338          * second to avoid subsequent fallbacks of other types There are 3
2339          * MIGRATE_TYPES we care about.
2340          */
2341         recommended_min += pageblock_nr_pages * nr_zones *
2342                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2343
2344         /* don't ever allow to reserve more than 5% of the lowmem */
2345         recommended_min = min(recommended_min,
2346                               (unsigned long) nr_free_buffer_pages() / 20);
2347         recommended_min <<= (PAGE_SHIFT-10);
2348
2349         if (recommended_min > min_free_kbytes) {
2350                 if (user_min_free_kbytes >= 0)
2351                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2352                                 min_free_kbytes, recommended_min);
2353
2354                 min_free_kbytes = recommended_min;
2355         }
2356
2357 update_wmarks:
2358         setup_per_zone_wmarks();
2359 }
2360
2361 int start_stop_khugepaged(void)
2362 {
2363         int err = 0;
2364
2365         mutex_lock(&khugepaged_mutex);
2366         if (khugepaged_enabled()) {
2367                 if (!khugepaged_thread)
2368                         khugepaged_thread = kthread_run(khugepaged, NULL,
2369                                                         "khugepaged");
2370                 if (IS_ERR(khugepaged_thread)) {
2371                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2372                         err = PTR_ERR(khugepaged_thread);
2373                         khugepaged_thread = NULL;
2374                         goto fail;
2375                 }
2376
2377                 if (!list_empty(&khugepaged_scan.mm_head))
2378                         wake_up_interruptible(&khugepaged_wait);
2379         } else if (khugepaged_thread) {
2380                 kthread_stop(khugepaged_thread);
2381                 khugepaged_thread = NULL;
2382         }
2383         set_recommended_min_free_kbytes();
2384 fail:
2385         mutex_unlock(&khugepaged_mutex);
2386         return err;
2387 }
2388
2389 void khugepaged_min_free_kbytes_update(void)
2390 {
2391         mutex_lock(&khugepaged_mutex);
2392         if (khugepaged_enabled() && khugepaged_thread)
2393                 set_recommended_min_free_kbytes();
2394         mutex_unlock(&khugepaged_mutex);
2395 }