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/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_EXCEED_SWAP_PTE,
  32         SCAN_EXCEED_SHARED_PTE,
  33         SCAN_PTE_NON_PRESENT,
  34         SCAN_PTE_UFFD_WP,
  35         SCAN_PAGE_RO,
  36         SCAN_LACK_REFERENCED_PAGE,
  37         SCAN_PAGE_NULL,
  38         SCAN_SCAN_ABORT,
  39         SCAN_PAGE_COUNT,
  40         SCAN_PAGE_LRU,
  41         SCAN_PAGE_LOCK,
  42         SCAN_PAGE_ANON,
  43         SCAN_PAGE_COMPOUND,
  44         SCAN_ANY_PROCESS,
  45         SCAN_VMA_NULL,
  46         SCAN_VMA_CHECK,
  47         SCAN_ADDRESS_RANGE,
  48         SCAN_SWAP_CACHE_PAGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 151                scan_sleep_millisecs_store);
 152
 153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 154                                           struct kobj_attribute *attr,
 155                                           char *buf)
 156 {
 157         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 158 }
 159
 160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 161                                            struct kobj_attribute *attr,
 162                                            const char *buf, size_t count)
 163 {
 164         unsigned int msecs;
 165         int err;
 166
 167         err = kstrtouint(buf, 10, &msecs);
 168         if (err)
 169                 return -EINVAL;
 170
 171         khugepaged_alloc_sleep_millisecs = msecs;
 172         khugepaged_sleep_expire = 0;
 173         wake_up_interruptible(&khugepaged_wait);
 174
 175         return count;
 176 }
 177 static struct kobj_attribute alloc_sleep_millisecs_attr =
 178         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 179                alloc_sleep_millisecs_store);
 180
 181 static ssize_t pages_to_scan_show(struct kobject *kobj,
 182                                   struct kobj_attribute *attr,
 183                                   char *buf)
 184 {
 185         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 186 }
 187 static ssize_t pages_to_scan_store(struct kobject *kobj,
 188                                    struct kobj_attribute *attr,
 189                                    const char *buf, size_t count)
 190 {
 191         unsigned int pages;
 192         int err;
 193
 194         err = kstrtouint(buf, 10, &pages);
 195         if (err || !pages)
 196                 return -EINVAL;
 197
 198         khugepaged_pages_to_scan = pages;
 199
 200         return count;
 201 }
 202 static struct kobj_attribute pages_to_scan_attr =
 203         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 204                pages_to_scan_store);
 205
 206 static ssize_t pages_collapsed_show(struct kobject *kobj,
 207                                     struct kobj_attribute *attr,
 208                                     char *buf)
 209 {
 210         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 211 }
 212 static struct kobj_attribute pages_collapsed_attr =
 213         __ATTR_RO(pages_collapsed);
 214
 215 static ssize_t full_scans_show(struct kobject *kobj,
 216                                struct kobj_attribute *attr,
 217                                char *buf)
 218 {
 219         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 220 }
 221 static struct kobj_attribute full_scans_attr =
 222         __ATTR_RO(full_scans);
 223
 224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 225                                       struct kobj_attribute *attr, char *buf)
 226 {
 227         return single_hugepage_flag_show(kobj, attr, buf,
 228                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 229 }
 230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 231                                        struct kobj_attribute *attr,
 232                                        const char *buf, size_t count)
 233 {
 234         return single_hugepage_flag_store(kobj, attr, buf, count,
 235                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 236 }
 237 static struct kobj_attribute khugepaged_defrag_attr =
 238         __ATTR(defrag, 0644, khugepaged_defrag_show,
 239                khugepaged_defrag_store);
 240
 241 /*
 242  * max_ptes_none controls if khugepaged should collapse hugepages over
 243  * any unmapped ptes in turn potentially increasing the memory
 244  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 245  * reduce the available free memory in the system as it
 246  * runs. Increasing max_ptes_none will instead potentially reduce the
 247  * free memory in the system during the khugepaged scan.
 248  */
 249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 250                                              struct kobj_attribute *attr,
 251                                              char *buf)
 252 {
 253         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 254 }
 255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 256                                               struct kobj_attribute *attr,
 257                                               const char *buf, size_t count)
 258 {
 259         int err;
 260         unsigned long max_ptes_none;
 261
 262         err = kstrtoul(buf, 10, &max_ptes_none);
 263         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 264                 return -EINVAL;
 265
 266         khugepaged_max_ptes_none = max_ptes_none;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 271         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 272                khugepaged_max_ptes_none_store);
 273
 274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 275                                              struct kobj_attribute *attr,
 276                                              char *buf)
 277 {
 278         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 279 }
 280
 281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 282                                               struct kobj_attribute *attr,
 283                                               const char *buf, size_t count)
 284 {
 285         int err;
 286         unsigned long max_ptes_swap;
 287
 288         err  = kstrtoul(buf, 10, &max_ptes_swap);
 289         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 290                 return -EINVAL;
 291
 292         khugepaged_max_ptes_swap = max_ptes_swap;
 293
 294         return count;
 295 }
 296
 297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 298         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 299                khugepaged_max_ptes_swap_store);
 300
 301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 302                                                struct kobj_attribute *attr,
 303                                                char *buf)
 304 {
 305         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 306 }
 307
 308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 309                                               struct kobj_attribute *attr,
 310                                               const char *buf, size_t count)
 311 {
 312         int err;
 313         unsigned long max_ptes_shared;
 314
 315         err  = kstrtoul(buf, 10, &max_ptes_shared);
 316         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 317                 return -EINVAL;
 318
 319         khugepaged_max_ptes_shared = max_ptes_shared;
 320
 321         return count;
 322 }
 323
 324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 325         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 326                khugepaged_max_ptes_shared_store);
 327
 328 static struct attribute *khugepaged_attr[] = {
 329         &khugepaged_defrag_attr.attr,
 330         &khugepaged_max_ptes_none_attr.attr,
 331         &khugepaged_max_ptes_swap_attr.attr,
 332         &khugepaged_max_ptes_shared_attr.attr,
 333         &pages_to_scan_attr.attr,
 334         &pages_collapsed_attr.attr,
 335         &full_scans_attr.attr,
 336         &scan_sleep_millisecs_attr.attr,
 337         &alloc_sleep_millisecs_attr.attr,
 338         NULL,
 339 };
 340
 341 struct attribute_group khugepaged_attr_group = {
 342         .attrs = khugepaged_attr,
 343         .name = "khugepaged",
 344 };
 345 #endif /* CONFIG_SYSFS */
 346
 347 int hugepage_madvise(struct vm_area_struct *vma,
 348                      unsigned long *vm_flags, int advice)
 349 {
 350         switch (advice) {
 351         case MADV_HUGEPAGE:
 352 #ifdef CONFIG_S390
 353                 /*
 354                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 355                  * can't handle this properly after s390_enable_sie, so we simply
 356                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 357                  */
 358                 if (mm_has_pgste(vma->vm_mm))
 359                         return 0;
 360 #endif
 361                 *vm_flags &= ~VM_NOHUGEPAGE;
 362                 *vm_flags |= VM_HUGEPAGE;
 363                 /*
 364                  * If the vma become good for khugepaged to scan,
 365                  * register it here without waiting a page fault that
 366                  * may not happen any time soon.
 367                  */
 368                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 369                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 370                         return -ENOMEM;
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct mm_slot),
 390                                           __alignof__(struct mm_slot), 0, NULL);
 391         if (!mm_slot_cache)
 392                 return -ENOMEM;
 393
 394         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 395         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 396         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 397         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 398
 399         return 0;
 400 }
 401
 402 void __init khugepaged_destroy(void)
 403 {
 404         kmem_cache_destroy(mm_slot_cache);
 405 }
 406
 407 static inline struct mm_slot *alloc_mm_slot(void)
 408 {
 409         if (!mm_slot_cache)     /* initialization failed */
 410                 return NULL;
 411         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 412 }
 413
 414 static inline void free_mm_slot(struct mm_slot *mm_slot)
 415 {
 416         kmem_cache_free(mm_slot_cache, mm_slot);
 417 }
 418
 419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 420 {
 421         struct mm_slot *mm_slot;
 422
 423         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 424                 if (mm == mm_slot->mm)
 425                         return mm_slot;
 426
 427         return NULL;
 428 }
 429
 430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 431                                     struct mm_slot *mm_slot)
 432 {
 433         mm_slot->mm = mm;
 434         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 435 }
 436
 437 static inline int khugepaged_test_exit(struct mm_struct *mm)
 438 {
 439         return atomic_read(&mm->mm_users) == 0;
 440 }
 441
 442 static bool hugepage_vma_check(struct vm_area_struct *vma,
 443                                unsigned long vm_flags)
 444 {
 445         if (!transhuge_vma_enabled(vma, vm_flags))
 446                 return false;
 447
 448         if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
 449                                 vma->vm_pgoff, HPAGE_PMD_NR))
 450                 return false;
 451
 452         /* Enabled via shmem mount options or sysfs settings. */
 453         if (shmem_file(vma->vm_file))
 454                 return shmem_huge_enabled(vma);
 455
 456         /* THP settings require madvise. */
 457         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 458                 return false;
 459
 460         /* Only regular file is valid */
 461         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 462             (vm_flags & VM_EXEC)) {
 463                 struct inode *inode = vma->vm_file->f_inode;
 464
 465                 return !inode_is_open_for_write(inode) &&
 466                         S_ISREG(inode->i_mode);
 467         }
 468
 469         if (!vma->anon_vma || vma->vm_ops)
 470                 return false;
 471         if (vma_is_temporary_stack(vma))
 472                 return false;
 473         return !(vm_flags & VM_NO_KHUGEPAGED);
 474 }
 475
 476 int __khugepaged_enter(struct mm_struct *mm)
 477 {
 478         struct mm_slot *mm_slot;
 479         int wakeup;
 480
 481         mm_slot = alloc_mm_slot();
 482         if (!mm_slot)
 483                 return -ENOMEM;
 484
 485         /* __khugepaged_exit() must not run from under us */
 486         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 487         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 488                 free_mm_slot(mm_slot);
 489                 return 0;
 490         }
 491
 492         spin_lock(&khugepaged_mm_lock);
 493         insert_to_mm_slots_hash(mm, mm_slot);
 494         /*
 495          * Insert just behind the scanning cursor, to let the area settle
 496          * down a little.
 497          */
 498         wakeup = list_empty(&khugepaged_scan.mm_head);
 499         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 500         spin_unlock(&khugepaged_mm_lock);
 501
 502         mmgrab(mm);
 503         if (wakeup)
 504                 wake_up_interruptible(&khugepaged_wait);
 505
 506         return 0;
 507 }
 508
 509 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 510                                unsigned long vm_flags)
 511 {
 512         unsigned long hstart, hend;
 513
 514         /*
 515          * khugepaged only supports read-only files for non-shmem files.
 516          * khugepaged does not yet work on special mappings. And
 517          * file-private shmem THP is not supported.
 518          */
 519         if (!hugepage_vma_check(vma, vm_flags))
 520                 return 0;
 521
 522         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 523         hend = vma->vm_end & HPAGE_PMD_MASK;
 524         if (hstart < hend)
 525                 return khugepaged_enter(vma, vm_flags);
 526         return 0;
 527 }
 528
 529 void __khugepaged_exit(struct mm_struct *mm)
 530 {
 531         struct mm_slot *mm_slot;
 532         int free = 0;
 533
 534         spin_lock(&khugepaged_mm_lock);
 535         mm_slot = get_mm_slot(mm);
 536         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 537                 hash_del(&mm_slot->hash);
 538                 list_del(&mm_slot->mm_node);
 539                 free = 1;
 540         }
 541         spin_unlock(&khugepaged_mm_lock);
 542
 543         if (free) {
 544                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 545                 free_mm_slot(mm_slot);
 546                 mmdrop(mm);
 547         } else if (mm_slot) {
 548                 /*
 549                  * This is required to serialize against
 550                  * khugepaged_test_exit() (which is guaranteed to run
 551                  * under mmap sem read mode). Stop here (after we
 552                  * return all pagetables will be destroyed) until
 553                  * khugepaged has finished working on the pagetables
 554                  * under the mmap_lock.
 555                  */
 556                 mmap_write_lock(mm);
 557                 mmap_write_unlock(mm);
 558         }
 559 }
 560
 561 static void release_pte_page(struct page *page)
 562 {
 563         mod_node_page_state(page_pgdat(page),
 564                         NR_ISOLATED_ANON + page_is_file_lru(page),
 565                         -compound_nr(page));
 566         unlock_page(page);
 567         putback_lru_page(page);
 568 }
 569
 570 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 571                 struct list_head *compound_pagelist)
 572 {
 573         struct page *page, *tmp;
 574
 575         while (--_pte >= pte) {
 576                 pte_t pteval = *_pte;
 577
 578                 page = pte_page(pteval);
 579                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 580                                 !PageCompound(page))
 581                         release_pte_page(page);
 582         }
 583
 584         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 585                 list_del(&page->lru);
 586                 release_pte_page(page);
 587         }
 588 }
 589
 590 static bool is_refcount_suitable(struct page *page)
 591 {
 592         int expected_refcount;
 593
 594         expected_refcount = total_mapcount(page);
 595         if (PageSwapCache(page))
 596                 expected_refcount += compound_nr(page);
 597
 598         return page_count(page) == expected_refcount;
 599 }
 600
 601 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 602                                         unsigned long address,
 603                                         pte_t *pte,
 604                                         struct list_head *compound_pagelist)
 605 {
 606         struct page *page = NULL;
 607         pte_t *_pte;
 608         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 609         bool writable = false;
 610
 611         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 612              _pte++, address += PAGE_SIZE) {
 613                 pte_t pteval = *_pte;
 614                 if (pte_none(pteval) || (pte_present(pteval) &&
 615                                 is_zero_pfn(pte_pfn(pteval)))) {
 616                         if (!userfaultfd_armed(vma) &&
 617                             ++none_or_zero <= khugepaged_max_ptes_none) {
 618                                 continue;
 619                         } else {
 620                                 result = SCAN_EXCEED_NONE_PTE;
 621                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 622                                 goto out;
 623                         }
 624                 }
 625                 if (!pte_present(pteval)) {
 626                         result = SCAN_PTE_NON_PRESENT;
 627                         goto out;
 628                 }
 629                 page = vm_normal_page(vma, address, pteval);
 630                 if (unlikely(!page)) {
 631                         result = SCAN_PAGE_NULL;
 632                         goto out;
 633                 }
 634
 635                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 636
 637                 if (page_mapcount(page) > 1 &&
 638                                 ++shared > khugepaged_max_ptes_shared) {
 639                         result = SCAN_EXCEED_SHARED_PTE;
 640                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 641                         goto out;
 642                 }
 643
 644                 if (PageCompound(page)) {
 645                         struct page *p;
 646                         page = compound_head(page);
 647
 648                         /*
 649                          * Check if we have dealt with the compound page
 650                          * already
 651                          */
 652                         list_for_each_entry(p, compound_pagelist, lru) {
 653                                 if (page == p)
 654                                         goto next;
 655                         }
 656                 }
 657
 658                 /*
 659                  * We can do it before isolate_lru_page because the
 660                  * page can't be freed from under us. NOTE: PG_lock
 661                  * is needed to serialize against split_huge_page
 662                  * when invoked from the VM.
 663                  */
 664                 if (!trylock_page(page)) {
 665                         result = SCAN_PAGE_LOCK;
 666                         goto out;
 667                 }
 668
 669                 /*
 670                  * Check if the page has any GUP (or other external) pins.
 671                  *
 672                  * The page table that maps the page has been already unlinked
 673                  * from the page table tree and this process cannot get
 674                  * an additional pin on the page.
 675                  *
 676                  * New pins can come later if the page is shared across fork,
 677                  * but not from this process. The other process cannot write to
 678                  * the page, only trigger CoW.
 679                  */
 680                 if (!is_refcount_suitable(page)) {
 681                         unlock_page(page);
 682                         result = SCAN_PAGE_COUNT;
 683                         goto out;
 684                 }
 685                 if (!pte_write(pteval) && PageSwapCache(page) &&
 686                                 !reuse_swap_page(page)) {
 687                         /*
 688                          * Page is in the swap cache and cannot be re-used.
 689                          * It cannot be collapsed into a THP.
 690                          */
 691                         unlock_page(page);
 692                         result = SCAN_SWAP_CACHE_PAGE;
 693                         goto out;
 694                 }
 695
 696                 /*
 697                  * Isolate the page to avoid collapsing an hugepage
 698                  * currently in use by the VM.
 699                  */
 700                 if (isolate_lru_page(page)) {
 701                         unlock_page(page);
 702                         result = SCAN_DEL_PAGE_LRU;
 703                         goto out;
 704                 }
 705                 mod_node_page_state(page_pgdat(page),
 706                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 707                                 compound_nr(page));
 708                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 709                 VM_BUG_ON_PAGE(PageLRU(page), page);
 710
 711                 if (PageCompound(page))
 712                         list_add_tail(&page->lru, compound_pagelist);
 713 next:
 714                 /* There should be enough young pte to collapse the page */
 715                 if (pte_young(pteval) ||
 716                     page_is_young(page) || PageReferenced(page) ||
 717                     mmu_notifier_test_young(vma->vm_mm, address))
 718                         referenced++;
 719
 720                 if (pte_write(pteval))
 721                         writable = true;
 722         }
 723
 724         if (unlikely(!writable)) {
 725                 result = SCAN_PAGE_RO;
 726         } else if (unlikely(!referenced)) {
 727                 result = SCAN_LACK_REFERENCED_PAGE;
 728         } else {
 729                 result = SCAN_SUCCEED;
 730                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 731                                                     referenced, writable, result);
 732                 return 1;
 733         }
 734 out:
 735         release_pte_pages(pte, _pte, compound_pagelist);
 736         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 737                                             referenced, writable, result);
 738         return 0;
 739 }
 740
 741 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 742                                       struct vm_area_struct *vma,
 743                                       unsigned long address,
 744                                       spinlock_t *ptl,
 745                                       struct list_head *compound_pagelist)
 746 {
 747         struct page *src_page, *tmp;
 748         pte_t *_pte;
 749         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 750                                 _pte++, page++, address += PAGE_SIZE) {
 751                 pte_t pteval = *_pte;
 752
 753                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 754                         clear_user_highpage(page, address);
 755                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 756                         if (is_zero_pfn(pte_pfn(pteval))) {
 757                                 /*
 758                                  * ptl mostly unnecessary.
 759                                  */
 760                                 spin_lock(ptl);
 761                                 ptep_clear(vma->vm_mm, address, _pte);
 762                                 spin_unlock(ptl);
 763                         }
 764                 } else {
 765                         src_page = pte_page(pteval);
 766                         copy_user_highpage(page, src_page, address, vma);
 767                         if (!PageCompound(src_page))
 768                                 release_pte_page(src_page);
 769                         /*
 770                          * ptl mostly unnecessary, but preempt has to
 771                          * be disabled to update the per-cpu stats
 772                          * inside page_remove_rmap().
 773                          */
 774                         spin_lock(ptl);
 775                         ptep_clear(vma->vm_mm, address, _pte);
 776                         page_remove_rmap(src_page, false);
 777                         spin_unlock(ptl);
 778                         free_page_and_swap_cache(src_page);
 779                 }
 780         }
 781
 782         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 783                 list_del(&src_page->lru);
 784                 release_pte_page(src_page);
 785         }
 786 }
 787
 788 static void khugepaged_alloc_sleep(void)
 789 {
 790         DEFINE_WAIT(wait);
 791
 792         add_wait_queue(&khugepaged_wait, &wait);
 793         freezable_schedule_timeout_interruptible(
 794                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 795         remove_wait_queue(&khugepaged_wait, &wait);
 796 }
 797
 798 static int khugepaged_node_load[MAX_NUMNODES];
 799
 800 static bool khugepaged_scan_abort(int nid)
 801 {
 802         int i;
 803
 804         /*
 805          * If node_reclaim_mode is disabled, then no extra effort is made to
 806          * allocate memory locally.
 807          */
 808         if (!node_reclaim_enabled())
 809                 return false;
 810
 811         /* If there is a count for this node already, it must be acceptable */
 812         if (khugepaged_node_load[nid])
 813                 return false;
 814
 815         for (i = 0; i < MAX_NUMNODES; i++) {
 816                 if (!khugepaged_node_load[i])
 817                         continue;
 818                 if (node_distance(nid, i) > node_reclaim_distance)
 819                         return true;
 820         }
 821         return false;
 822 }
 823
 824 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 825 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 826 {
 827         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 828 }
 829
 830 #ifdef CONFIG_NUMA
 831 static int khugepaged_find_target_node(void)
 832 {
 833         static int last_khugepaged_target_node = NUMA_NO_NODE;
 834         int nid, target_node = 0, max_value = 0;
 835
 836         /* find first node with max normal pages hit */
 837         for (nid = 0; nid < MAX_NUMNODES; nid++)
 838                 if (khugepaged_node_load[nid] > max_value) {
 839                         max_value = khugepaged_node_load[nid];
 840                         target_node = nid;
 841                 }
 842
 843         /* do some balance if several nodes have the same hit record */
 844         if (target_node <= last_khugepaged_target_node)
 845                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 846                                 nid++)
 847                         if (max_value == khugepaged_node_load[nid]) {
 848                                 target_node = nid;
 849                                 break;
 850                         }
 851
 852         last_khugepaged_target_node = target_node;
 853         return target_node;
 854 }
 855
 856 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 857 {
 858         if (IS_ERR(*hpage)) {
 859                 if (!*wait)
 860                         return false;
 861
 862                 *wait = false;
 863                 *hpage = NULL;
 864                 khugepaged_alloc_sleep();
 865         } else if (*hpage) {
 866                 put_page(*hpage);
 867                 *hpage = NULL;
 868         }
 869
 870         return true;
 871 }
 872
 873 static struct page *
 874 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 875 {
 876         VM_BUG_ON_PAGE(*hpage, *hpage);
 877
 878         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 879         if (unlikely(!*hpage)) {
 880                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 881                 *hpage = ERR_PTR(-ENOMEM);
 882                 return NULL;
 883         }
 884
 885         prep_transhuge_page(*hpage);
 886         count_vm_event(THP_COLLAPSE_ALLOC);
 887         return *hpage;
 888 }
 889 #else
 890 static int khugepaged_find_target_node(void)
 891 {
 892         return 0;
 893 }
 894
 895 static inline struct page *alloc_khugepaged_hugepage(void)
 896 {
 897         struct page *page;
 898
 899         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 900                            HPAGE_PMD_ORDER);
 901         if (page)
 902                 prep_transhuge_page(page);
 903         return page;
 904 }
 905
 906 static struct page *khugepaged_alloc_hugepage(bool *wait)
 907 {
 908         struct page *hpage;
 909
 910         do {
 911                 hpage = alloc_khugepaged_hugepage();
 912                 if (!hpage) {
 913                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 914                         if (!*wait)
 915                                 return NULL;
 916
 917                         *wait = false;
 918                         khugepaged_alloc_sleep();
 919                 } else
 920                         count_vm_event(THP_COLLAPSE_ALLOC);
 921         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 922
 923         return hpage;
 924 }
 925
 926 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 927 {
 928         /*
 929          * If the hpage allocated earlier was briefly exposed in page cache
 930          * before collapse_file() failed, it is possible that racing lookups
 931          * have not yet completed, and would then be unpleasantly surprised by
 932          * finding the hpage reused for the same mapping at a different offset.
 933          * Just release the previous allocation if there is any danger of that.
 934          */
 935         if (*hpage && page_count(*hpage) > 1) {
 936                 put_page(*hpage);
 937                 *hpage = NULL;
 938         }
 939
 940         if (!*hpage)
 941                 *hpage = khugepaged_alloc_hugepage(wait);
 942
 943         if (unlikely(!*hpage))
 944                 return false;
 945
 946         return true;
 947 }
 948
 949 static struct page *
 950 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 951 {
 952         VM_BUG_ON(!*hpage);
 953
 954         return  *hpage;
 955 }
 956 #endif
 957
 958 /*
 959  * If mmap_lock temporarily dropped, revalidate vma
 960  * before taking mmap_lock.
 961  * Return 0 if succeeds, otherwise return none-zero
 962  * value (scan code).
 963  */
 964
 965 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 966                 struct vm_area_struct **vmap)
 967 {
 968         struct vm_area_struct *vma;
 969         unsigned long hstart, hend;
 970
 971         if (unlikely(khugepaged_test_exit(mm)))
 972                 return SCAN_ANY_PROCESS;
 973
 974         *vmap = vma = find_vma(mm, address);
 975         if (!vma)
 976                 return SCAN_VMA_NULL;
 977
 978         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 979         hend = vma->vm_end & HPAGE_PMD_MASK;
 980         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 981                 return SCAN_ADDRESS_RANGE;
 982         if (!hugepage_vma_check(vma, vma->vm_flags))
 983                 return SCAN_VMA_CHECK;
 984         /* Anon VMA expected */
 985         if (!vma->anon_vma || vma->vm_ops)
 986                 return SCAN_VMA_CHECK;
 987         return 0;
 988 }
 989
 990 /*
 991  * Bring missing pages in from swap, to complete THP collapse.
 992  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 993  *
 994  * Called and returns without pte mapped or spinlocks held,
 995  * but with mmap_lock held to protect against vma changes.
 996  */
 997
 998 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 999                                         struct vm_area_struct *vma,
1000                                         unsigned long haddr, pmd_t *pmd,
1001                                         int referenced)
1002 {
1003         int swapped_in = 0;
1004         vm_fault_t ret = 0;
1005         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1006
1007         for (address = haddr; address < end; address += PAGE_SIZE) {
1008                 struct vm_fault vmf = {
1009                         .vma = vma,
1010                         .address = address,
1011                         .pgoff = linear_page_index(vma, haddr),
1012                         .flags = FAULT_FLAG_ALLOW_RETRY,
1013                         .pmd = pmd,
1014                 };
1015
1016                 vmf.pte = pte_offset_map(pmd, address);
1017                 vmf.orig_pte = *vmf.pte;
1018                 if (!is_swap_pte(vmf.orig_pte)) {
1019                         pte_unmap(vmf.pte);
1020                         continue;
1021                 }
1022                 swapped_in++;
1023                 ret = do_swap_page(&vmf);
1024
1025                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1026                 if (ret & VM_FAULT_RETRY) {
1027                         mmap_read_lock(mm);
1028                         if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1029                                 /* vma is no longer available, don't continue to swapin */
1030                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1031                                 return false;
1032                         }
1033                         /* check if the pmd is still valid */
1034                         if (mm_find_pmd(mm, haddr) != pmd) {
1035                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1036                                 return false;
1037                         }
1038                 }
1039                 if (ret & VM_FAULT_ERROR) {
1040                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1041                         return false;
1042                 }
1043         }
1044
1045         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1046         if (swapped_in)
1047                 lru_add_drain();
1048
1049         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1050         return true;
1051 }
1052
1053 static void collapse_huge_page(struct mm_struct *mm,
1054                                    unsigned long address,
1055                                    struct page **hpage,
1056                                    int node, int referenced, int unmapped)
1057 {
1058         LIST_HEAD(compound_pagelist);
1059         pmd_t *pmd, _pmd;
1060         pte_t *pte;
1061         pgtable_t pgtable;
1062         struct page *new_page;
1063         spinlock_t *pmd_ptl, *pte_ptl;
1064         int isolated = 0, result = 0;
1065         struct vm_area_struct *vma;
1066         struct mmu_notifier_range range;
1067         gfp_t gfp;
1068
1069         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1070
1071         /* Only allocate from the target node */
1072         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1073
1074         /*
1075          * Before allocating the hugepage, release the mmap_lock read lock.
1076          * The allocation can take potentially a long time if it involves
1077          * sync compaction, and we do not need to hold the mmap_lock during
1078          * that. We will recheck the vma after taking it again in write mode.
1079          */
1080         mmap_read_unlock(mm);
1081         new_page = khugepaged_alloc_page(hpage, gfp, node);
1082         if (!new_page) {
1083                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1084                 goto out_nolock;
1085         }
1086
1087         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1088                 result = SCAN_CGROUP_CHARGE_FAIL;
1089                 goto out_nolock;
1090         }
1091         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1092
1093         mmap_read_lock(mm);
1094         result = hugepage_vma_revalidate(mm, address, &vma);
1095         if (result) {
1096                 mmap_read_unlock(mm);
1097                 goto out_nolock;
1098         }
1099
1100         pmd = mm_find_pmd(mm, address);
1101         if (!pmd) {
1102                 result = SCAN_PMD_NULL;
1103                 mmap_read_unlock(mm);
1104                 goto out_nolock;
1105         }
1106
1107         /*
1108          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1109          * If it fails, we release mmap_lock and jump out_nolock.
1110          * Continuing to collapse causes inconsistency.
1111          */
1112         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1113                                                      pmd, referenced)) {
1114                 mmap_read_unlock(mm);
1115                 goto out_nolock;
1116         }
1117
1118         mmap_read_unlock(mm);
1119         /*
1120          * Prevent all access to pagetables with the exception of
1121          * gup_fast later handled by the ptep_clear_flush and the VM
1122          * handled by the anon_vma lock + PG_lock.
1123          */
1124         mmap_write_lock(mm);
1125         result = hugepage_vma_revalidate(mm, address, &vma);
1126         if (result)
1127                 goto out_up_write;
1128         /* check if the pmd is still valid */
1129         if (mm_find_pmd(mm, address) != pmd)
1130                 goto out_up_write;
1131
1132         anon_vma_lock_write(vma->anon_vma);
1133
1134         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1135                                 address, address + HPAGE_PMD_SIZE);
1136         mmu_notifier_invalidate_range_start(&range);
1137
1138         pte = pte_offset_map(pmd, address);
1139         pte_ptl = pte_lockptr(mm, pmd);
1140
1141         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1142         /*
1143          * After this gup_fast can't run anymore. This also removes
1144          * any huge TLB entry from the CPU so we won't allow
1145          * huge and small TLB entries for the same virtual address
1146          * to avoid the risk of CPU bugs in that area.
1147          */
1148         _pmd = pmdp_collapse_flush(vma, address, pmd);
1149         spin_unlock(pmd_ptl);
1150         mmu_notifier_invalidate_range_end(&range);
1151
1152         spin_lock(pte_ptl);
1153         isolated = __collapse_huge_page_isolate(vma, address, pte,
1154                         &compound_pagelist);
1155         spin_unlock(pte_ptl);
1156
1157         if (unlikely(!isolated)) {
1158                 pte_unmap(pte);
1159                 spin_lock(pmd_ptl);
1160                 BUG_ON(!pmd_none(*pmd));
1161                 /*
1162                  * We can only use set_pmd_at when establishing
1163                  * hugepmds and never for establishing regular pmds that
1164                  * points to regular pagetables. Use pmd_populate for that
1165                  */
1166                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1167                 spin_unlock(pmd_ptl);
1168                 anon_vma_unlock_write(vma->anon_vma);
1169                 result = SCAN_FAIL;
1170                 goto out_up_write;
1171         }
1172
1173         /*
1174          * All pages are isolated and locked so anon_vma rmap
1175          * can't run anymore.
1176          */
1177         anon_vma_unlock_write(vma->anon_vma);
1178
1179         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1180                         &compound_pagelist);
1181         pte_unmap(pte);
1182         /*
1183          * spin_lock() below is not the equivalent of smp_wmb(), but
1184          * the smp_wmb() inside __SetPageUptodate() can be reused to
1185          * avoid the copy_huge_page writes to become visible after
1186          * the set_pmd_at() write.
1187          */
1188         __SetPageUptodate(new_page);
1189         pgtable = pmd_pgtable(_pmd);
1190
1191         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1192         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1193
1194         spin_lock(pmd_ptl);
1195         BUG_ON(!pmd_none(*pmd));
1196         page_add_new_anon_rmap(new_page, vma, address, true);
1197         lru_cache_add_inactive_or_unevictable(new_page, vma);
1198         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1199         set_pmd_at(mm, address, pmd, _pmd);
1200         update_mmu_cache_pmd(vma, address, pmd);
1201         spin_unlock(pmd_ptl);
1202
1203         *hpage = NULL;
1204
1205         khugepaged_pages_collapsed++;
1206         result = SCAN_SUCCEED;
1207 out_up_write:
1208         mmap_write_unlock(mm);
1209 out_nolock:
1210         if (!IS_ERR_OR_NULL(*hpage))
1211                 mem_cgroup_uncharge(page_folio(*hpage));
1212         trace_mm_collapse_huge_page(mm, isolated, result);
1213         return;
1214 }
1215
1216 static int khugepaged_scan_pmd(struct mm_struct *mm,
1217                                struct vm_area_struct *vma,
1218                                unsigned long address,
1219                                struct page **hpage)
1220 {
1221         pmd_t *pmd;
1222         pte_t *pte, *_pte;
1223         int ret = 0, result = 0, referenced = 0;
1224         int none_or_zero = 0, shared = 0;
1225         struct page *page = NULL;
1226         unsigned long _address;
1227         spinlock_t *ptl;
1228         int node = NUMA_NO_NODE, unmapped = 0;
1229         bool writable = false;
1230
1231         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1232
1233         pmd = mm_find_pmd(mm, address);
1234         if (!pmd) {
1235                 result = SCAN_PMD_NULL;
1236                 goto out;
1237         }
1238
1239         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1240         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1241         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1242              _pte++, _address += PAGE_SIZE) {
1243                 pte_t pteval = *_pte;
1244                 if (is_swap_pte(pteval)) {
1245                         if (++unmapped <= khugepaged_max_ptes_swap) {
1246                                 /*
1247                                  * Always be strict with uffd-wp
1248                                  * enabled swap entries.  Please see
1249                                  * comment below for pte_uffd_wp().
1250                                  */
1251                                 if (pte_swp_uffd_wp(pteval)) {
1252                                         result = SCAN_PTE_UFFD_WP;
1253                                         goto out_unmap;
1254                                 }
1255                                 continue;
1256                         } else {
1257                                 result = SCAN_EXCEED_SWAP_PTE;
1258                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1259                                 goto out_unmap;
1260                         }
1261                 }
1262                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1263                         if (!userfaultfd_armed(vma) &&
1264                             ++none_or_zero <= khugepaged_max_ptes_none) {
1265                                 continue;
1266                         } else {
1267                                 result = SCAN_EXCEED_NONE_PTE;
1268                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1269                                 goto out_unmap;
1270                         }
1271                 }
1272                 if (pte_uffd_wp(pteval)) {
1273                         /*
1274                          * Don't collapse the page if any of the small
1275                          * PTEs are armed with uffd write protection.
1276                          * Here we can also mark the new huge pmd as
1277                          * write protected if any of the small ones is
1278                          * marked but that could bring unknown
1279                          * userfault messages that falls outside of
1280                          * the registered range.  So, just be simple.
1281                          */
1282                         result = SCAN_PTE_UFFD_WP;
1283                         goto out_unmap;
1284                 }
1285                 if (pte_write(pteval))
1286                         writable = true;
1287
1288                 page = vm_normal_page(vma, _address, pteval);
1289                 if (unlikely(!page)) {
1290                         result = SCAN_PAGE_NULL;
1291                         goto out_unmap;
1292                 }
1293
1294                 if (page_mapcount(page) > 1 &&
1295                                 ++shared > khugepaged_max_ptes_shared) {
1296                         result = SCAN_EXCEED_SHARED_PTE;
1297                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1298                         goto out_unmap;
1299                 }
1300
1301                 page = compound_head(page);
1302
1303                 /*
1304                  * Record which node the original page is from and save this
1305                  * information to khugepaged_node_load[].
1306                  * Khugepaged will allocate hugepage from the node has the max
1307                  * hit record.
1308                  */
1309                 node = page_to_nid(page);
1310                 if (khugepaged_scan_abort(node)) {
1311                         result = SCAN_SCAN_ABORT;
1312                         goto out_unmap;
1313                 }
1314                 khugepaged_node_load[node]++;
1315                 if (!PageLRU(page)) {
1316                         result = SCAN_PAGE_LRU;
1317                         goto out_unmap;
1318                 }
1319                 if (PageLocked(page)) {
1320                         result = SCAN_PAGE_LOCK;
1321                         goto out_unmap;
1322                 }
1323                 if (!PageAnon(page)) {
1324                         result = SCAN_PAGE_ANON;
1325                         goto out_unmap;
1326                 }
1327
1328                 /*
1329                  * Check if the page has any GUP (or other external) pins.
1330                  *
1331                  * Here the check is racy it may see totmal_mapcount > refcount
1332                  * in some cases.
1333                  * For example, one process with one forked child process.
1334                  * The parent has the PMD split due to MADV_DONTNEED, then
1335                  * the child is trying unmap the whole PMD, but khugepaged
1336                  * may be scanning the parent between the child has
1337                  * PageDoubleMap flag cleared and dec the mapcount.  So
1338                  * khugepaged may see total_mapcount > refcount.
1339                  *
1340                  * But such case is ephemeral we could always retry collapse
1341                  * later.  However it may report false positive if the page
1342                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1343                  * will be done again later the risk seems low.
1344                  */
1345                 if (!is_refcount_suitable(page)) {
1346                         result = SCAN_PAGE_COUNT;
1347                         goto out_unmap;
1348                 }
1349                 if (pte_young(pteval) ||
1350                     page_is_young(page) || PageReferenced(page) ||
1351                     mmu_notifier_test_young(vma->vm_mm, address))
1352                         referenced++;
1353         }
1354         if (!writable) {
1355                 result = SCAN_PAGE_RO;
1356         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1357                 result = SCAN_LACK_REFERENCED_PAGE;
1358         } else {
1359                 result = SCAN_SUCCEED;
1360                 ret = 1;
1361         }
1362 out_unmap:
1363         pte_unmap_unlock(pte, ptl);
1364         if (ret) {
1365                 node = khugepaged_find_target_node();
1366                 /* collapse_huge_page will return with the mmap_lock released */
1367                 collapse_huge_page(mm, address, hpage, node,
1368                                 referenced, unmapped);
1369         }
1370 out:
1371         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1372                                      none_or_zero, result, unmapped);
1373         return ret;
1374 }
1375
1376 static void collect_mm_slot(struct mm_slot *mm_slot)
1377 {
1378         struct mm_struct *mm = mm_slot->mm;
1379
1380         lockdep_assert_held(&khugepaged_mm_lock);
1381
1382         if (khugepaged_test_exit(mm)) {
1383                 /* free mm_slot */
1384                 hash_del(&mm_slot->hash);
1385                 list_del(&mm_slot->mm_node);
1386
1387                 /*
1388                  * Not strictly needed because the mm exited already.
1389                  *
1390                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1391                  */
1392
1393                 /* khugepaged_mm_lock actually not necessary for the below */
1394                 free_mm_slot(mm_slot);
1395                 mmdrop(mm);
1396         }
1397 }
1398
1399 #ifdef CONFIG_SHMEM
1400 /*
1401  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1402  * khugepaged should try to collapse the page table.
1403  */
1404 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1405                                          unsigned long addr)
1406 {
1407         struct mm_slot *mm_slot;
1408
1409         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1410
1411         spin_lock(&khugepaged_mm_lock);
1412         mm_slot = get_mm_slot(mm);
1413         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1414                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1415         spin_unlock(&khugepaged_mm_lock);
1416         return 0;
1417 }
1418
1419 /**
1420  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1421  * address haddr.
1422  *
1423  * @mm: process address space where collapse happens
1424  * @addr: THP collapse address
1425  *
1426  * This function checks whether all the PTEs in the PMD are pointing to the
1427  * right THP. If so, retract the page table so the THP can refault in with
1428  * as pmd-mapped.
1429  */
1430 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1431 {
1432         unsigned long haddr = addr & HPAGE_PMD_MASK;
1433         struct vm_area_struct *vma = find_vma(mm, haddr);
1434         struct page *hpage;
1435         pte_t *start_pte, *pte;
1436         pmd_t *pmd, _pmd;
1437         spinlock_t *ptl;
1438         int count = 0;
1439         int i;
1440
1441         if (!vma || !vma->vm_file ||
1442             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1443                 return;
1444
1445         /*
1446          * This vm_flags may not have VM_HUGEPAGE if the page was not
1447          * collapsed by this mm. But we can still collapse if the page is
1448          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1449          * will not fail the vma for missing VM_HUGEPAGE
1450          */
1451         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1452                 return;
1453
1454         hpage = find_lock_page(vma->vm_file->f_mapping,
1455                                linear_page_index(vma, haddr));
1456         if (!hpage)
1457                 return;
1458
1459         if (!PageHead(hpage))
1460                 goto drop_hpage;
1461
1462         pmd = mm_find_pmd(mm, haddr);
1463         if (!pmd)
1464                 goto drop_hpage;
1465
1466         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1467
1468         /* step 1: check all mapped PTEs are to the right huge page */
1469         for (i = 0, addr = haddr, pte = start_pte;
1470              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1471                 struct page *page;
1472
1473                 /* empty pte, skip */
1474                 if (pte_none(*pte))
1475                         continue;
1476
1477                 /* page swapped out, abort */
1478                 if (!pte_present(*pte))
1479                         goto abort;
1480
1481                 page = vm_normal_page(vma, addr, *pte);
1482
1483                 /*
1484                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1485                  * page table, but the new page will not be a subpage of hpage.
1486                  */
1487                 if (hpage + i != page)
1488                         goto abort;
1489                 count++;
1490         }
1491
1492         /* step 2: adjust rmap */
1493         for (i = 0, addr = haddr, pte = start_pte;
1494              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1495                 struct page *page;
1496
1497                 if (pte_none(*pte))
1498                         continue;
1499                 page = vm_normal_page(vma, addr, *pte);
1500                 page_remove_rmap(page, false);
1501         }
1502
1503         pte_unmap_unlock(start_pte, ptl);
1504
1505         /* step 3: set proper refcount and mm_counters. */
1506         if (count) {
1507                 page_ref_sub(hpage, count);
1508                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1509         }
1510
1511         /* step 4: collapse pmd */
1512         ptl = pmd_lock(vma->vm_mm, pmd);
1513         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1514         spin_unlock(ptl);
1515         mm_dec_nr_ptes(mm);
1516         pte_free(mm, pmd_pgtable(_pmd));
1517
1518 drop_hpage:
1519         unlock_page(hpage);
1520         put_page(hpage);
1521         return;
1522
1523 abort:
1524         pte_unmap_unlock(start_pte, ptl);
1525         goto drop_hpage;
1526 }
1527
1528 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1529 {
1530         struct mm_struct *mm = mm_slot->mm;
1531         int i;
1532
1533         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1534                 return;
1535
1536         if (!mmap_write_trylock(mm))
1537                 return;
1538
1539         if (unlikely(khugepaged_test_exit(mm)))
1540                 goto out;
1541
1542         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1543                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1544
1545 out:
1546         mm_slot->nr_pte_mapped_thp = 0;
1547         mmap_write_unlock(mm);
1548 }
1549
1550 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1551 {
1552         struct vm_area_struct *vma;
1553         struct mm_struct *mm;
1554         unsigned long addr;
1555         pmd_t *pmd, _pmd;
1556
1557         i_mmap_lock_write(mapping);
1558         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1559                 /*
1560                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1561                  * got written to. These VMAs are likely not worth investing
1562                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1563                  * later.
1564                  *
1565                  * Not that vma->anon_vma check is racy: it can be set up after
1566                  * the check but before we took mmap_lock by the fault path.
1567                  * But page lock would prevent establishing any new ptes of the
1568                  * page, so we are safe.
1569                  *
1570                  * An alternative would be drop the check, but check that page
1571                  * table is clear before calling pmdp_collapse_flush() under
1572                  * ptl. It has higher chance to recover THP for the VMA, but
1573                  * has higher cost too.
1574                  */
1575                 if (vma->anon_vma)
1576                         continue;
1577                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1578                 if (addr & ~HPAGE_PMD_MASK)
1579                         continue;
1580                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1581                         continue;
1582                 mm = vma->vm_mm;
1583                 pmd = mm_find_pmd(mm, addr);
1584                 if (!pmd)
1585                         continue;
1586                 /*
1587                  * We need exclusive mmap_lock to retract page table.
1588                  *
1589                  * We use trylock due to lock inversion: we need to acquire
1590                  * mmap_lock while holding page lock. Fault path does it in
1591                  * reverse order. Trylock is a way to avoid deadlock.
1592                  */
1593                 if (mmap_write_trylock(mm)) {
1594                         if (!khugepaged_test_exit(mm)) {
1595                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1596                                 /* assume page table is clear */
1597                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1598                                 spin_unlock(ptl);
1599                                 mm_dec_nr_ptes(mm);
1600                                 pte_free(mm, pmd_pgtable(_pmd));
1601                         }
1602                         mmap_write_unlock(mm);
1603                 } else {
1604                         /* Try again later */
1605                         khugepaged_add_pte_mapped_thp(mm, addr);
1606                 }
1607         }
1608         i_mmap_unlock_write(mapping);
1609 }
1610
1611 /**
1612  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1613  *
1614  * @mm: process address space where collapse happens
1615  * @file: file that collapse on
1616  * @start: collapse start address
1617  * @hpage: new allocated huge page for collapse
1618  * @node: appointed node the new huge page allocate from
1619  *
1620  * Basic scheme is simple, details are more complex:
1621  *  - allocate and lock a new huge page;
1622  *  - scan page cache replacing old pages with the new one
1623  *    + swap/gup in pages if necessary;
1624  *    + fill in gaps;
1625  *    + keep old pages around in case rollback is required;
1626  *  - if replacing succeeds:
1627  *    + copy data over;
1628  *    + free old pages;
1629  *    + unlock huge page;
1630  *  - if replacing failed;
1631  *    + put all pages back and unfreeze them;
1632  *    + restore gaps in the page cache;
1633  *    + unlock and free huge page;
1634  */
1635 static void collapse_file(struct mm_struct *mm,
1636                 struct file *file, pgoff_t start,
1637                 struct page **hpage, int node)
1638 {
1639         struct address_space *mapping = file->f_mapping;
1640         gfp_t gfp;
1641         struct page *new_page;
1642         pgoff_t index, end = start + HPAGE_PMD_NR;
1643         LIST_HEAD(pagelist);
1644         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1645         int nr_none = 0, result = SCAN_SUCCEED;
1646         bool is_shmem = shmem_file(file);
1647         int nr;
1648
1649         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1650         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1651
1652         /* Only allocate from the target node */
1653         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1654
1655         new_page = khugepaged_alloc_page(hpage, gfp, node);
1656         if (!new_page) {
1657                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1658                 goto out;
1659         }
1660
1661         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1662                 result = SCAN_CGROUP_CHARGE_FAIL;
1663                 goto out;
1664         }
1665         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1666
1667         /*
1668          * Ensure we have slots for all the pages in the range.  This is
1669          * almost certainly a no-op because most of the pages must be present
1670          */
1671         do {
1672                 xas_lock_irq(&xas);
1673                 xas_create_range(&xas);
1674                 if (!xas_error(&xas))
1675                         break;
1676                 xas_unlock_irq(&xas);
1677                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1678                         result = SCAN_FAIL;
1679                         goto out;
1680                 }
1681         } while (1);
1682
1683         __SetPageLocked(new_page);
1684         if (is_shmem)
1685                 __SetPageSwapBacked(new_page);
1686         new_page->index = start;
1687         new_page->mapping = mapping;
1688
1689         /*
1690          * At this point the new_page is locked and not up-to-date.
1691          * It's safe to insert it into the page cache, because nobody would
1692          * be able to map it or use it in another way until we unlock it.
1693          */
1694
1695         xas_set(&xas, start);
1696         for (index = start; index < end; index++) {
1697                 struct page *page = xas_next(&xas);
1698
1699                 VM_BUG_ON(index != xas.xa_index);
1700                 if (is_shmem) {
1701                         if (!page) {
1702                                 /*
1703                                  * Stop if extent has been truncated or
1704                                  * hole-punched, and is now completely
1705                                  * empty.
1706                                  */
1707                                 if (index == start) {
1708                                         if (!xas_next_entry(&xas, end - 1)) {
1709                                                 result = SCAN_TRUNCATED;
1710                                                 goto xa_locked;
1711                                         }
1712                                         xas_set(&xas, index);
1713                                 }
1714                                 if (!shmem_charge(mapping->host, 1)) {
1715                                         result = SCAN_FAIL;
1716                                         goto xa_locked;
1717                                 }
1718                                 xas_store(&xas, new_page);
1719                                 nr_none++;
1720                                 continue;
1721                         }
1722
1723                         if (xa_is_value(page) || !PageUptodate(page)) {
1724                                 xas_unlock_irq(&xas);
1725                                 /* swap in or instantiate fallocated page */
1726                                 if (shmem_getpage(mapping->host, index, &page,
1727                                                   SGP_NOALLOC)) {
1728                                         result = SCAN_FAIL;
1729                                         goto xa_unlocked;
1730                                 }
1731                         } else if (trylock_page(page)) {
1732                                 get_page(page);
1733                                 xas_unlock_irq(&xas);
1734                         } else {
1735                                 result = SCAN_PAGE_LOCK;
1736                                 goto xa_locked;
1737                         }
1738                 } else {        /* !is_shmem */
1739                         if (!page || xa_is_value(page)) {
1740                                 xas_unlock_irq(&xas);
1741                                 page_cache_sync_readahead(mapping, &file->f_ra,
1742                                                           file, index,
1743                                                           end - index);
1744                                 /* drain pagevecs to help isolate_lru_page() */
1745                                 lru_add_drain();
1746                                 page = find_lock_page(mapping, index);
1747                                 if (unlikely(page == NULL)) {
1748                                         result = SCAN_FAIL;
1749                                         goto xa_unlocked;
1750                                 }
1751                         } else if (PageDirty(page)) {
1752                                 /*
1753                                  * khugepaged only works on read-only fd,
1754                                  * so this page is dirty because it hasn't
1755                                  * been flushed since first write. There
1756                                  * won't be new dirty pages.
1757                                  *
1758                                  * Trigger async flush here and hope the
1759                                  * writeback is done when khugepaged
1760                                  * revisits this page.
1761                                  *
1762                                  * This is a one-off situation. We are not
1763                                  * forcing writeback in loop.
1764                                  */
1765                                 xas_unlock_irq(&xas);
1766                                 filemap_flush(mapping);
1767                                 result = SCAN_FAIL;
1768                                 goto xa_unlocked;
1769                         } else if (PageWriteback(page)) {
1770                                 xas_unlock_irq(&xas);
1771                                 result = SCAN_FAIL;
1772                                 goto xa_unlocked;
1773                         } else if (trylock_page(page)) {
1774                                 get_page(page);
1775                                 xas_unlock_irq(&xas);
1776                         } else {
1777                                 result = SCAN_PAGE_LOCK;
1778                                 goto xa_locked;
1779                         }
1780                 }
1781
1782                 /*
1783                  * The page must be locked, so we can drop the i_pages lock
1784                  * without racing with truncate.
1785                  */
1786                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1787
1788                 /* make sure the page is up to date */
1789                 if (unlikely(!PageUptodate(page))) {
1790                         result = SCAN_FAIL;
1791                         goto out_unlock;
1792                 }
1793
1794                 /*
1795                  * If file was truncated then extended, or hole-punched, before
1796                  * we locked the first page, then a THP might be there already.
1797                  */
1798                 if (PageTransCompound(page)) {
1799                         result = SCAN_PAGE_COMPOUND;
1800                         goto out_unlock;
1801                 }
1802
1803                 if (page_mapping(page) != mapping) {
1804                         result = SCAN_TRUNCATED;
1805                         goto out_unlock;
1806                 }
1807
1808                 if (!is_shmem && (PageDirty(page) ||
1809                                   PageWriteback(page))) {
1810                         /*
1811                          * khugepaged only works on read-only fd, so this
1812                          * page is dirty because it hasn't been flushed
1813                          * since first write.
1814                          */
1815                         result = SCAN_FAIL;
1816                         goto out_unlock;
1817                 }
1818
1819                 if (isolate_lru_page(page)) {
1820                         result = SCAN_DEL_PAGE_LRU;
1821                         goto out_unlock;
1822                 }
1823
1824                 if (page_has_private(page) &&
1825                     !try_to_release_page(page, GFP_KERNEL)) {
1826                         result = SCAN_PAGE_HAS_PRIVATE;
1827                         putback_lru_page(page);
1828                         goto out_unlock;
1829                 }
1830
1831                 if (page_mapped(page))
1832                         unmap_mapping_pages(mapping, index, 1, false);
1833
1834                 xas_lock_irq(&xas);
1835                 xas_set(&xas, index);
1836
1837                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1838                 VM_BUG_ON_PAGE(page_mapped(page), page);
1839
1840                 /*
1841                  * The page is expected to have page_count() == 3:
1842                  *  - we hold a pin on it;
1843                  *  - one reference from page cache;
1844                  *  - one from isolate_lru_page;
1845                  */
1846                 if (!page_ref_freeze(page, 3)) {
1847                         result = SCAN_PAGE_COUNT;
1848                         xas_unlock_irq(&xas);
1849                         putback_lru_page(page);
1850                         goto out_unlock;
1851                 }
1852
1853                 /*
1854                  * Add the page to the list to be able to undo the collapse if
1855                  * something go wrong.
1856                  */
1857                 list_add_tail(&page->lru, &pagelist);
1858
1859                 /* Finally, replace with the new page. */
1860                 xas_store(&xas, new_page);
1861                 continue;
1862 out_unlock:
1863                 unlock_page(page);
1864                 put_page(page);
1865                 goto xa_unlocked;
1866         }
1867         nr = thp_nr_pages(new_page);
1868
1869         if (is_shmem)
1870                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1871         else {
1872                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1873                 filemap_nr_thps_inc(mapping);
1874                 /*
1875                  * Paired with smp_mb() in do_dentry_open() to ensure
1876                  * i_writecount is up to date and the update to nr_thps is
1877                  * visible. Ensures the page cache will be truncated if the
1878                  * file is opened writable.
1879                  */
1880                 smp_mb();
1881                 if (inode_is_open_for_write(mapping->host)) {
1882                         result = SCAN_FAIL;
1883                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1884                         filemap_nr_thps_dec(mapping);
1885                         goto xa_locked;
1886                 }
1887         }
1888
1889         if (nr_none) {
1890                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1891                 if (is_shmem)
1892                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1893         }
1894
1895         /* Join all the small entries into a single multi-index entry */
1896         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1897         xas_store(&xas, new_page);
1898 xa_locked:
1899         xas_unlock_irq(&xas);
1900 xa_unlocked:
1901
1902         if (result == SCAN_SUCCEED) {
1903                 struct page *page, *tmp;
1904
1905                 /*
1906                  * Replacing old pages with new one has succeeded, now we
1907                  * need to copy the content and free the old pages.
1908                  */
1909                 index = start;
1910                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1911                         while (index < page->index) {
1912                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1913                                 index++;
1914                         }
1915                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1916                                         page);
1917                         list_del(&page->lru);
1918                         page->mapping = NULL;
1919                         page_ref_unfreeze(page, 1);
1920                         ClearPageActive(page);
1921                         ClearPageUnevictable(page);
1922                         unlock_page(page);
1923                         put_page(page);
1924                         index++;
1925                 }
1926                 while (index < end) {
1927                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1928                         index++;
1929                 }
1930
1931                 SetPageUptodate(new_page);
1932                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1933                 if (is_shmem)
1934                         set_page_dirty(new_page);
1935                 lru_cache_add(new_page);
1936
1937                 /*
1938                  * Remove pte page tables, so we can re-fault the page as huge.
1939                  */
1940                 retract_page_tables(mapping, start);
1941                 *hpage = NULL;
1942
1943                 khugepaged_pages_collapsed++;
1944         } else {
1945                 struct page *page;
1946
1947                 /* Something went wrong: roll back page cache changes */
1948                 xas_lock_irq(&xas);
1949                 mapping->nrpages -= nr_none;
1950
1951                 if (is_shmem)
1952                         shmem_uncharge(mapping->host, nr_none);
1953
1954                 xas_set(&xas, start);
1955                 xas_for_each(&xas, page, end - 1) {
1956                         page = list_first_entry_or_null(&pagelist,
1957                                         struct page, lru);
1958                         if (!page || xas.xa_index < page->index) {
1959                                 if (!nr_none)
1960                                         break;
1961                                 nr_none--;
1962                                 /* Put holes back where they were */
1963                                 xas_store(&xas, NULL);
1964                                 continue;
1965                         }
1966
1967                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1968
1969                         /* Unfreeze the page. */
1970                         list_del(&page->lru);
1971                         page_ref_unfreeze(page, 2);
1972                         xas_store(&xas, page);
1973                         xas_pause(&xas);
1974                         xas_unlock_irq(&xas);
1975                         unlock_page(page);
1976                         putback_lru_page(page);
1977                         xas_lock_irq(&xas);
1978                 }
1979                 VM_BUG_ON(nr_none);
1980                 xas_unlock_irq(&xas);
1981
1982                 new_page->mapping = NULL;
1983         }
1984
1985         unlock_page(new_page);
1986 out:
1987         VM_BUG_ON(!list_empty(&pagelist));
1988         if (!IS_ERR_OR_NULL(*hpage))
1989                 mem_cgroup_uncharge(page_folio(*hpage));
1990         /* TODO: tracepoints */
1991 }
1992
1993 static void khugepaged_scan_file(struct mm_struct *mm,
1994                 struct file *file, pgoff_t start, struct page **hpage)
1995 {
1996         struct page *page = NULL;
1997         struct address_space *mapping = file->f_mapping;
1998         XA_STATE(xas, &mapping->i_pages, start);
1999         int present, swap;
2000         int node = NUMA_NO_NODE;
2001         int result = SCAN_SUCCEED;
2002
2003         present = 0;
2004         swap = 0;
2005         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2006         rcu_read_lock();
2007         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2008                 if (xas_retry(&xas, page))
2009                         continue;
2010
2011                 if (xa_is_value(page)) {
2012                         if (++swap > khugepaged_max_ptes_swap) {
2013                                 result = SCAN_EXCEED_SWAP_PTE;
2014                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2015                                 break;
2016                         }
2017                         continue;
2018                 }
2019
2020                 /*
2021                  * XXX: khugepaged should compact smaller compound pages
2022                  * into a PMD sized page
2023                  */
2024                 if (PageTransCompound(page)) {
2025                         result = SCAN_PAGE_COMPOUND;
2026                         break;
2027                 }
2028
2029                 node = page_to_nid(page);
2030                 if (khugepaged_scan_abort(node)) {
2031                         result = SCAN_SCAN_ABORT;
2032                         break;
2033                 }
2034                 khugepaged_node_load[node]++;
2035
2036                 if (!PageLRU(page)) {
2037                         result = SCAN_PAGE_LRU;
2038                         break;
2039                 }
2040
2041                 if (page_count(page) !=
2042                     1 + page_mapcount(page) + page_has_private(page)) {
2043                         result = SCAN_PAGE_COUNT;
2044                         break;
2045                 }
2046
2047                 /*
2048                  * We probably should check if the page is referenced here, but
2049                  * nobody would transfer pte_young() to PageReferenced() for us.
2050                  * And rmap walk here is just too costly...
2051                  */
2052
2053                 present++;
2054
2055                 if (need_resched()) {
2056                         xas_pause(&xas);
2057                         cond_resched_rcu();
2058                 }
2059         }
2060         rcu_read_unlock();
2061
2062         if (result == SCAN_SUCCEED) {
2063                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2064                         result = SCAN_EXCEED_NONE_PTE;
2065                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2066                 } else {
2067                         node = khugepaged_find_target_node();
2068                         collapse_file(mm, file, start, hpage, node);
2069                 }
2070         }
2071
2072         /* TODO: tracepoints */
2073 }
2074 #else
2075 static void khugepaged_scan_file(struct mm_struct *mm,
2076                 struct file *file, pgoff_t start, struct page **hpage)
2077 {
2078         BUILD_BUG();
2079 }
2080
2081 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2082 {
2083 }
2084 #endif
2085
2086 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2087                                             struct page **hpage)
2088         __releases(&khugepaged_mm_lock)
2089         __acquires(&khugepaged_mm_lock)
2090 {
2091         struct mm_slot *mm_slot;
2092         struct mm_struct *mm;
2093         struct vm_area_struct *vma;
2094         int progress = 0;
2095
2096         VM_BUG_ON(!pages);
2097         lockdep_assert_held(&khugepaged_mm_lock);
2098
2099         if (khugepaged_scan.mm_slot)
2100                 mm_slot = khugepaged_scan.mm_slot;
2101         else {
2102                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2103                                      struct mm_slot, mm_node);
2104                 khugepaged_scan.address = 0;
2105                 khugepaged_scan.mm_slot = mm_slot;
2106         }
2107         spin_unlock(&khugepaged_mm_lock);
2108         khugepaged_collapse_pte_mapped_thps(mm_slot);
2109
2110         mm = mm_slot->mm;
2111         /*
2112          * Don't wait for semaphore (to avoid long wait times).  Just move to
2113          * the next mm on the list.
2114          */
2115         vma = NULL;
2116         if (unlikely(!mmap_read_trylock(mm)))
2117                 goto breakouterloop_mmap_lock;
2118         if (likely(!khugepaged_test_exit(mm)))
2119                 vma = find_vma(mm, khugepaged_scan.address);
2120
2121         progress++;
2122         for (; vma; vma = vma->vm_next) {
2123                 unsigned long hstart, hend;
2124
2125                 cond_resched();
2126                 if (unlikely(khugepaged_test_exit(mm))) {
2127                         progress++;
2128                         break;
2129                 }
2130                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2131 skip:
2132                         progress++;
2133                         continue;
2134                 }
2135                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2136                 hend = vma->vm_end & HPAGE_PMD_MASK;
2137                 if (hstart >= hend)
2138                         goto skip;
2139                 if (khugepaged_scan.address > hend)
2140                         goto skip;
2141                 if (khugepaged_scan.address < hstart)
2142                         khugepaged_scan.address = hstart;
2143                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2144                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2145                         goto skip;
2146
2147                 while (khugepaged_scan.address < hend) {
2148                         int ret;
2149                         cond_resched();
2150                         if (unlikely(khugepaged_test_exit(mm)))
2151                                 goto breakouterloop;
2152
2153                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2154                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2155                                   hend);
2156                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2157                                 struct file *file = get_file(vma->vm_file);
2158                                 pgoff_t pgoff = linear_page_index(vma,
2159                                                 khugepaged_scan.address);
2160
2161                                 mmap_read_unlock(mm);
2162                                 ret = 1;
2163                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2164                                 fput(file);
2165                         } else {
2166                                 ret = khugepaged_scan_pmd(mm, vma,
2167                                                 khugepaged_scan.address,
2168                                                 hpage);
2169                         }
2170                         /* move to next address */
2171                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2172                         progress += HPAGE_PMD_NR;
2173                         if (ret)
2174                                 /* we released mmap_lock so break loop */
2175                                 goto breakouterloop_mmap_lock;
2176                         if (progress >= pages)
2177                                 goto breakouterloop;
2178                 }
2179         }
2180 breakouterloop:
2181         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2182 breakouterloop_mmap_lock:
2183
2184         spin_lock(&khugepaged_mm_lock);
2185         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2186         /*
2187          * Release the current mm_slot if this mm is about to die, or
2188          * if we scanned all vmas of this mm.
2189          */
2190         if (khugepaged_test_exit(mm) || !vma) {
2191                 /*
2192                  * Make sure that if mm_users is reaching zero while
2193                  * khugepaged runs here, khugepaged_exit will find
2194                  * mm_slot not pointing to the exiting mm.
2195                  */
2196                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2197                         khugepaged_scan.mm_slot = list_entry(
2198                                 mm_slot->mm_node.next,
2199                                 struct mm_slot, mm_node);
2200                         khugepaged_scan.address = 0;
2201                 } else {
2202                         khugepaged_scan.mm_slot = NULL;
2203                         khugepaged_full_scans++;
2204                 }
2205
2206                 collect_mm_slot(mm_slot);
2207         }
2208
2209         return progress;
2210 }
2211
2212 static int khugepaged_has_work(void)
2213 {
2214         return !list_empty(&khugepaged_scan.mm_head) &&
2215                 khugepaged_enabled();
2216 }
2217
2218 static int khugepaged_wait_event(void)
2219 {
2220         return !list_empty(&khugepaged_scan.mm_head) ||
2221                 kthread_should_stop();
2222 }
2223
2224 static void khugepaged_do_scan(void)
2225 {
2226         struct page *hpage = NULL;
2227         unsigned int progress = 0, pass_through_head = 0;
2228         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2229         bool wait = true;
2230
2231         lru_add_drain_all();
2232
2233         while (progress < pages) {
2234                 if (!khugepaged_prealloc_page(&hpage, &wait))
2235                         break;
2236
2237                 cond_resched();
2238
2239                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2240                         break;
2241
2242                 spin_lock(&khugepaged_mm_lock);
2243                 if (!khugepaged_scan.mm_slot)
2244                         pass_through_head++;
2245                 if (khugepaged_has_work() &&
2246                     pass_through_head < 2)
2247                         progress += khugepaged_scan_mm_slot(pages - progress,
2248                                                             &hpage);
2249                 else
2250                         progress = pages;
2251                 spin_unlock(&khugepaged_mm_lock);
2252         }
2253
2254         if (!IS_ERR_OR_NULL(hpage))
2255                 put_page(hpage);
2256 }
2257
2258 static bool khugepaged_should_wakeup(void)
2259 {
2260         return kthread_should_stop() ||
2261                time_after_eq(jiffies, khugepaged_sleep_expire);
2262 }
2263
2264 static void khugepaged_wait_work(void)
2265 {
2266         if (khugepaged_has_work()) {
2267                 const unsigned long scan_sleep_jiffies =
2268                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2269
2270                 if (!scan_sleep_jiffies)
2271                         return;
2272
2273                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2274                 wait_event_freezable_timeout(khugepaged_wait,
2275                                              khugepaged_should_wakeup(),
2276                                              scan_sleep_jiffies);
2277                 return;
2278         }
2279
2280         if (khugepaged_enabled())
2281                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2282 }
2283
2284 static int khugepaged(void *none)
2285 {
2286         struct mm_slot *mm_slot;
2287
2288         set_freezable();
2289         set_user_nice(current, MAX_NICE);
2290
2291         while (!kthread_should_stop()) {
2292                 khugepaged_do_scan();
2293                 khugepaged_wait_work();
2294         }
2295
2296         spin_lock(&khugepaged_mm_lock);
2297         mm_slot = khugepaged_scan.mm_slot;
2298         khugepaged_scan.mm_slot = NULL;
2299         if (mm_slot)
2300                 collect_mm_slot(mm_slot);
2301         spin_unlock(&khugepaged_mm_lock);
2302         return 0;
2303 }
2304
2305 static void set_recommended_min_free_kbytes(void)
2306 {
2307         struct zone *zone;
2308         int nr_zones = 0;
2309         unsigned long recommended_min;
2310
2311         if (!khugepaged_enabled()) {
2312                 calculate_min_free_kbytes();
2313                 goto update_wmarks;
2314         }
2315
2316         for_each_populated_zone(zone) {
2317                 /*
2318                  * We don't need to worry about fragmentation of
2319                  * ZONE_MOVABLE since it only has movable pages.
2320                  */
2321                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2322                         continue;
2323
2324                 nr_zones++;
2325         }
2326
2327         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2328         recommended_min = pageblock_nr_pages * nr_zones * 2;
2329
2330         /*
2331          * Make sure that on average at least two pageblocks are almost free
2332          * of another type, one for a migratetype to fall back to and a
2333          * second to avoid subsequent fallbacks of other types There are 3
2334          * MIGRATE_TYPES we care about.
2335          */
2336         recommended_min += pageblock_nr_pages * nr_zones *
2337                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2338
2339         /* don't ever allow to reserve more than 5% of the lowmem */
2340         recommended_min = min(recommended_min,
2341                               (unsigned long) nr_free_buffer_pages() / 20);
2342         recommended_min <<= (PAGE_SHIFT-10);
2343
2344         if (recommended_min > min_free_kbytes) {
2345                 if (user_min_free_kbytes >= 0)
2346                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2347                                 min_free_kbytes, recommended_min);
2348
2349                 min_free_kbytes = recommended_min;
2350         }
2351
2352 update_wmarks:
2353         setup_per_zone_wmarks();
2354 }
2355
2356 int start_stop_khugepaged(void)
2357 {
2358         int err = 0;
2359
2360         mutex_lock(&khugepaged_mutex);
2361         if (khugepaged_enabled()) {
2362                 if (!khugepaged_thread)
2363                         khugepaged_thread = kthread_run(khugepaged, NULL,
2364                                                         "khugepaged");
2365                 if (IS_ERR(khugepaged_thread)) {
2366                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2367                         err = PTR_ERR(khugepaged_thread);
2368                         khugepaged_thread = NULL;
2369                         goto fail;
2370                 }
2371
2372                 if (!list_empty(&khugepaged_scan.mm_head))
2373                         wake_up_interruptible(&khugepaged_wait);
2374         } else if (khugepaged_thread) {
2375                 kthread_stop(khugepaged_thread);
2376                 khugepaged_thread = NULL;
2377         }
2378         set_recommended_min_free_kbytes();
2379 fail:
2380         mutex_unlock(&khugepaged_mutex);
2381         return err;
2382 }
2383
2384 void khugepaged_min_free_kbytes_update(void)
2385 {
2386         mutex_lock(&khugepaged_mutex);
2387         if (khugepaged_enabled() && khugepaged_thread)
2388                 set_recommended_min_free_kbytes();
2389         mutex_unlock(&khugepaged_mutex);
2390 }