mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/page_table_check.h>
  20 #include <linux/swapops.h>
  21 #include <linux/shmem_fs.h>
  22
  23 #include <asm/tlb.h>
  24 #include <asm/pgalloc.h>
  25 #include "internal.h"
  26
  27 enum scan_result {
  28         SCAN_FAIL,
  29         SCAN_SUCCEED,
  30         SCAN_PMD_NULL,
  31         SCAN_EXCEED_NONE_PTE,
  32         SCAN_EXCEED_SWAP_PTE,
  33         SCAN_EXCEED_SHARED_PTE,
  34         SCAN_PTE_NON_PRESENT,
  35         SCAN_PTE_UFFD_WP,
  36         SCAN_PAGE_RO,
  37         SCAN_LACK_REFERENCED_PAGE,
  38         SCAN_PAGE_NULL,
  39         SCAN_SCAN_ABORT,
  40         SCAN_PAGE_COUNT,
  41         SCAN_PAGE_LRU,
  42         SCAN_PAGE_LOCK,
  43         SCAN_PAGE_ANON,
  44         SCAN_PAGE_COMPOUND,
  45         SCAN_ANY_PROCESS,
  46         SCAN_VMA_NULL,
  47         SCAN_VMA_CHECK,
  48         SCAN_ADDRESS_RANGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR(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                 khugepaged_enter_vma(vma, *vm_flags);
 369                 break;
 370         case MADV_NOHUGEPAGE:
 371                 *vm_flags &= ~VM_HUGEPAGE;
 372                 *vm_flags |= VM_NOHUGEPAGE;
 373                 /*
 374                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 375                  * this vma even if we leave the mm registered in khugepaged if
 376                  * it got registered before VM_NOHUGEPAGE was set.
 377                  */
 378                 break;
 379         }
 380
 381         return 0;
 382 }
 383
 384 int __init khugepaged_init(void)
 385 {
 386         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 387                                           sizeof(struct mm_slot),
 388                                           __alignof__(struct mm_slot), 0, NULL);
 389         if (!mm_slot_cache)
 390                 return -ENOMEM;
 391
 392         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 393         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 394         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 395         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 396
 397         return 0;
 398 }
 399
 400 void __init khugepaged_destroy(void)
 401 {
 402         kmem_cache_destroy(mm_slot_cache);
 403 }
 404
 405 static inline struct mm_slot *alloc_mm_slot(void)
 406 {
 407         if (!mm_slot_cache)     /* initialization failed */
 408                 return NULL;
 409         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 410 }
 411
 412 static inline void free_mm_slot(struct mm_slot *mm_slot)
 413 {
 414         kmem_cache_free(mm_slot_cache, mm_slot);
 415 }
 416
 417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 418 {
 419         struct mm_slot *mm_slot;
 420
 421         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 422                 if (mm == mm_slot->mm)
 423                         return mm_slot;
 424
 425         return NULL;
 426 }
 427
 428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 429                                     struct mm_slot *mm_slot)
 430 {
 431         mm_slot->mm = mm;
 432         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 433 }
 434
 435 static inline int khugepaged_test_exit(struct mm_struct *mm)
 436 {
 437         return atomic_read(&mm->mm_users) == 0;
 438 }
 439
 440 bool hugepage_vma_check(struct vm_area_struct *vma,
 441                         unsigned long vm_flags)
 442 {
 443         if (!transhuge_vma_enabled(vma, vm_flags))
 444                 return false;
 445
 446         if (vm_flags & VM_NO_KHUGEPAGED)
 447                 return false;
 448
 449         /* Don't run khugepaged against DAX vma */
 450         if (vma_is_dax(vma))
 451                 return false;
 452
 453         if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
 454                                 vma->vm_pgoff, HPAGE_PMD_NR))
 455                 return false;
 456
 457         /* Enabled via shmem mount options or sysfs settings. */
 458         if (shmem_file(vma->vm_file))
 459                 return shmem_huge_enabled(vma);
 460
 461         /* THP settings require madvise. */
 462         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 463                 return false;
 464
 465         /* Only regular file is valid */
 466         if (file_thp_enabled(vma))
 467                 return true;
 468
 469         if (!vma->anon_vma || !vma_is_anonymous(vma))
 470                 return false;
 471         if (vma_is_temporary_stack(vma))
 472                 return false;
 473
 474         return true;
 475 }
 476
 477 void __khugepaged_enter(struct mm_struct *mm)
 478 {
 479         struct mm_slot *mm_slot;
 480         int wakeup;
 481
 482         mm_slot = alloc_mm_slot();
 483         if (!mm_slot)
 484                 return;
 485
 486         /* __khugepaged_exit() must not run from under us */
 487         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 488         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 489                 free_mm_slot(mm_slot);
 490                 return;
 491         }
 492
 493         spin_lock(&khugepaged_mm_lock);
 494         insert_to_mm_slots_hash(mm, mm_slot);
 495         /*
 496          * Insert just behind the scanning cursor, to let the area settle
 497          * down a little.
 498          */
 499         wakeup = list_empty(&khugepaged_scan.mm_head);
 500         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 501         spin_unlock(&khugepaged_mm_lock);
 502
 503         mmgrab(mm);
 504         if (wakeup)
 505                 wake_up_interruptible(&khugepaged_wait);
 506 }
 507
 508 void khugepaged_enter_vma(struct vm_area_struct *vma,
 509                           unsigned long vm_flags)
 510 {
 511         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 512             khugepaged_enabled() &&
 513             (((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
 514              (vma->vm_end & HPAGE_PMD_MASK))) {
 515                 if (hugepage_vma_check(vma, vm_flags))
 516                         __khugepaged_enter(vma->vm_mm);
 517         }
 518 }
 519
 520 void __khugepaged_exit(struct mm_struct *mm)
 521 {
 522         struct mm_slot *mm_slot;
 523         int free = 0;
 524
 525         spin_lock(&khugepaged_mm_lock);
 526         mm_slot = get_mm_slot(mm);
 527         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 528                 hash_del(&mm_slot->hash);
 529                 list_del(&mm_slot->mm_node);
 530                 free = 1;
 531         }
 532         spin_unlock(&khugepaged_mm_lock);
 533
 534         if (free) {
 535                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 536                 free_mm_slot(mm_slot);
 537                 mmdrop(mm);
 538         } else if (mm_slot) {
 539                 /*
 540                  * This is required to serialize against
 541                  * khugepaged_test_exit() (which is guaranteed to run
 542                  * under mmap sem read mode). Stop here (after we
 543                  * return all pagetables will be destroyed) until
 544                  * khugepaged has finished working on the pagetables
 545                  * under the mmap_lock.
 546                  */
 547                 mmap_write_lock(mm);
 548                 mmap_write_unlock(mm);
 549         }
 550 }
 551
 552 static void release_pte_page(struct page *page)
 553 {
 554         mod_node_page_state(page_pgdat(page),
 555                         NR_ISOLATED_ANON + page_is_file_lru(page),
 556                         -compound_nr(page));
 557         unlock_page(page);
 558         putback_lru_page(page);
 559 }
 560
 561 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 562                 struct list_head *compound_pagelist)
 563 {
 564         struct page *page, *tmp;
 565
 566         while (--_pte >= pte) {
 567                 pte_t pteval = *_pte;
 568
 569                 page = pte_page(pteval);
 570                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 571                                 !PageCompound(page))
 572                         release_pte_page(page);
 573         }
 574
 575         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 576                 list_del(&page->lru);
 577                 release_pte_page(page);
 578         }
 579 }
 580
 581 static bool is_refcount_suitable(struct page *page)
 582 {
 583         int expected_refcount;
 584
 585         expected_refcount = total_mapcount(page);
 586         if (PageSwapCache(page))
 587                 expected_refcount += compound_nr(page);
 588
 589         return page_count(page) == expected_refcount;
 590 }
 591
 592 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 593                                         unsigned long address,
 594                                         pte_t *pte,
 595                                         struct list_head *compound_pagelist)
 596 {
 597         struct page *page = NULL;
 598         pte_t *_pte;
 599         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 600         bool writable = false;
 601
 602         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 603              _pte++, address += PAGE_SIZE) {
 604                 pte_t pteval = *_pte;
 605                 if (pte_none(pteval) || (pte_present(pteval) &&
 606                                 is_zero_pfn(pte_pfn(pteval)))) {
 607                         if (!userfaultfd_armed(vma) &&
 608                             ++none_or_zero <= khugepaged_max_ptes_none) {
 609                                 continue;
 610                         } else {
 611                                 result = SCAN_EXCEED_NONE_PTE;
 612                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 613                                 goto out;
 614                         }
 615                 }
 616                 if (!pte_present(pteval)) {
 617                         result = SCAN_PTE_NON_PRESENT;
 618                         goto out;
 619                 }
 620                 page = vm_normal_page(vma, address, pteval);
 621                 if (unlikely(!page)) {
 622                         result = SCAN_PAGE_NULL;
 623                         goto out;
 624                 }
 625
 626                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 627
 628                 if (page_mapcount(page) > 1 &&
 629                                 ++shared > khugepaged_max_ptes_shared) {
 630                         result = SCAN_EXCEED_SHARED_PTE;
 631                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 632                         goto out;
 633                 }
 634
 635                 if (PageCompound(page)) {
 636                         struct page *p;
 637                         page = compound_head(page);
 638
 639                         /*
 640                          * Check if we have dealt with the compound page
 641                          * already
 642                          */
 643                         list_for_each_entry(p, compound_pagelist, lru) {
 644                                 if (page == p)
 645                                         goto next;
 646                         }
 647                 }
 648
 649                 /*
 650                  * We can do it before isolate_lru_page because the
 651                  * page can't be freed from under us. NOTE: PG_lock
 652                  * is needed to serialize against split_huge_page
 653                  * when invoked from the VM.
 654                  */
 655                 if (!trylock_page(page)) {
 656                         result = SCAN_PAGE_LOCK;
 657                         goto out;
 658                 }
 659
 660                 /*
 661                  * Check if the page has any GUP (or other external) pins.
 662                  *
 663                  * The page table that maps the page has been already unlinked
 664                  * from the page table tree and this process cannot get
 665                  * an additional pin on the page.
 666                  *
 667                  * New pins can come later if the page is shared across fork,
 668                  * but not from this process. The other process cannot write to
 669                  * the page, only trigger CoW.
 670                  */
 671                 if (!is_refcount_suitable(page)) {
 672                         unlock_page(page);
 673                         result = SCAN_PAGE_COUNT;
 674                         goto out;
 675                 }
 676
 677                 /*
 678                  * Isolate the page to avoid collapsing an hugepage
 679                  * currently in use by the VM.
 680                  */
 681                 if (isolate_lru_page(page)) {
 682                         unlock_page(page);
 683                         result = SCAN_DEL_PAGE_LRU;
 684                         goto out;
 685                 }
 686                 mod_node_page_state(page_pgdat(page),
 687                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 688                                 compound_nr(page));
 689                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 690                 VM_BUG_ON_PAGE(PageLRU(page), page);
 691
 692                 if (PageCompound(page))
 693                         list_add_tail(&page->lru, compound_pagelist);
 694 next:
 695                 /* There should be enough young pte to collapse the page */
 696                 if (pte_young(pteval) ||
 697                     page_is_young(page) || PageReferenced(page) ||
 698                     mmu_notifier_test_young(vma->vm_mm, address))
 699                         referenced++;
 700
 701                 if (pte_write(pteval))
 702                         writable = true;
 703         }
 704
 705         if (unlikely(!writable)) {
 706                 result = SCAN_PAGE_RO;
 707         } else if (unlikely(!referenced)) {
 708                 result = SCAN_LACK_REFERENCED_PAGE;
 709         } else {
 710                 result = SCAN_SUCCEED;
 711                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 712                                                     referenced, writable, result);
 713                 return 1;
 714         }
 715 out:
 716         release_pte_pages(pte, _pte, compound_pagelist);
 717         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 718                                             referenced, writable, result);
 719         return 0;
 720 }
 721
 722 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 723                                       struct vm_area_struct *vma,
 724                                       unsigned long address,
 725                                       spinlock_t *ptl,
 726                                       struct list_head *compound_pagelist)
 727 {
 728         struct page *src_page, *tmp;
 729         pte_t *_pte;
 730         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 731                                 _pte++, page++, address += PAGE_SIZE) {
 732                 pte_t pteval = *_pte;
 733
 734                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 735                         clear_user_highpage(page, address);
 736                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 737                         if (is_zero_pfn(pte_pfn(pteval))) {
 738                                 /*
 739                                  * ptl mostly unnecessary.
 740                                  */
 741                                 spin_lock(ptl);
 742                                 ptep_clear(vma->vm_mm, address, _pte);
 743                                 spin_unlock(ptl);
 744                         }
 745                 } else {
 746                         src_page = pte_page(pteval);
 747                         copy_user_highpage(page, src_page, address, vma);
 748                         if (!PageCompound(src_page))
 749                                 release_pte_page(src_page);
 750                         /*
 751                          * ptl mostly unnecessary, but preempt has to
 752                          * be disabled to update the per-cpu stats
 753                          * inside page_remove_rmap().
 754                          */
 755                         spin_lock(ptl);
 756                         ptep_clear(vma->vm_mm, address, _pte);
 757                         page_remove_rmap(src_page, vma, false);
 758                         spin_unlock(ptl);
 759                         free_page_and_swap_cache(src_page);
 760                 }
 761         }
 762
 763         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 764                 list_del(&src_page->lru);
 765                 release_pte_page(src_page);
 766         }
 767 }
 768
 769 static void khugepaged_alloc_sleep(void)
 770 {
 771         DEFINE_WAIT(wait);
 772
 773         add_wait_queue(&khugepaged_wait, &wait);
 774         freezable_schedule_timeout_interruptible(
 775                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 776         remove_wait_queue(&khugepaged_wait, &wait);
 777 }
 778
 779 static int khugepaged_node_load[MAX_NUMNODES];
 780
 781 static bool khugepaged_scan_abort(int nid)
 782 {
 783         int i;
 784
 785         /*
 786          * If node_reclaim_mode is disabled, then no extra effort is made to
 787          * allocate memory locally.
 788          */
 789         if (!node_reclaim_enabled())
 790                 return false;
 791
 792         /* If there is a count for this node already, it must be acceptable */
 793         if (khugepaged_node_load[nid])
 794                 return false;
 795
 796         for (i = 0; i < MAX_NUMNODES; i++) {
 797                 if (!khugepaged_node_load[i])
 798                         continue;
 799                 if (node_distance(nid, i) > node_reclaim_distance)
 800                         return true;
 801         }
 802         return false;
 803 }
 804
 805 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 806 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 807 {
 808         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 809 }
 810
 811 #ifdef CONFIG_NUMA
 812 static int khugepaged_find_target_node(void)
 813 {
 814         static int last_khugepaged_target_node = NUMA_NO_NODE;
 815         int nid, target_node = 0, max_value = 0;
 816
 817         /* find first node with max normal pages hit */
 818         for (nid = 0; nid < MAX_NUMNODES; nid++)
 819                 if (khugepaged_node_load[nid] > max_value) {
 820                         max_value = khugepaged_node_load[nid];
 821                         target_node = nid;
 822                 }
 823
 824         /* do some balance if several nodes have the same hit record */
 825         if (target_node <= last_khugepaged_target_node)
 826                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 827                                 nid++)
 828                         if (max_value == khugepaged_node_load[nid]) {
 829                                 target_node = nid;
 830                                 break;
 831                         }
 832
 833         last_khugepaged_target_node = target_node;
 834         return target_node;
 835 }
 836
 837 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 838 {
 839         if (IS_ERR(*hpage)) {
 840                 if (!*wait)
 841                         return false;
 842
 843                 *wait = false;
 844                 *hpage = NULL;
 845                 khugepaged_alloc_sleep();
 846         } else if (*hpage) {
 847                 put_page(*hpage);
 848                 *hpage = NULL;
 849         }
 850
 851         return true;
 852 }
 853
 854 static struct page *
 855 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 856 {
 857         VM_BUG_ON_PAGE(*hpage, *hpage);
 858
 859         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 860         if (unlikely(!*hpage)) {
 861                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 862                 *hpage = ERR_PTR(-ENOMEM);
 863                 return NULL;
 864         }
 865
 866         prep_transhuge_page(*hpage);
 867         count_vm_event(THP_COLLAPSE_ALLOC);
 868         return *hpage;
 869 }
 870 #else
 871 static int khugepaged_find_target_node(void)
 872 {
 873         return 0;
 874 }
 875
 876 static inline struct page *alloc_khugepaged_hugepage(void)
 877 {
 878         struct page *page;
 879
 880         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 881                            HPAGE_PMD_ORDER);
 882         if (page)
 883                 prep_transhuge_page(page);
 884         return page;
 885 }
 886
 887 static struct page *khugepaged_alloc_hugepage(bool *wait)
 888 {
 889         struct page *hpage;
 890
 891         do {
 892                 hpage = alloc_khugepaged_hugepage();
 893                 if (!hpage) {
 894                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 895                         if (!*wait)
 896                                 return NULL;
 897
 898                         *wait = false;
 899                         khugepaged_alloc_sleep();
 900                 } else
 901                         count_vm_event(THP_COLLAPSE_ALLOC);
 902         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 903
 904         return hpage;
 905 }
 906
 907 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 908 {
 909         /*
 910          * If the hpage allocated earlier was briefly exposed in page cache
 911          * before collapse_file() failed, it is possible that racing lookups
 912          * have not yet completed, and would then be unpleasantly surprised by
 913          * finding the hpage reused for the same mapping at a different offset.
 914          * Just release the previous allocation if there is any danger of that.
 915          */
 916         if (*hpage && page_count(*hpage) > 1) {
 917                 put_page(*hpage);
 918                 *hpage = NULL;
 919         }
 920
 921         if (!*hpage)
 922                 *hpage = khugepaged_alloc_hugepage(wait);
 923
 924         if (unlikely(!*hpage))
 925                 return false;
 926
 927         return true;
 928 }
 929
 930 static struct page *
 931 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 932 {
 933         VM_BUG_ON(!*hpage);
 934
 935         return  *hpage;
 936 }
 937 #endif
 938
 939 /*
 940  * If mmap_lock temporarily dropped, revalidate vma
 941  * before taking mmap_lock.
 942  * Return 0 if succeeds, otherwise return none-zero
 943  * value (scan code).
 944  */
 945
 946 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 947                 struct vm_area_struct **vmap)
 948 {
 949         struct vm_area_struct *vma;
 950         unsigned long hstart, hend;
 951
 952         if (unlikely(khugepaged_test_exit(mm)))
 953                 return SCAN_ANY_PROCESS;
 954
 955         *vmap = vma = find_vma(mm, address);
 956         if (!vma)
 957                 return SCAN_VMA_NULL;
 958
 959         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 960         hend = vma->vm_end & HPAGE_PMD_MASK;
 961         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 962                 return SCAN_ADDRESS_RANGE;
 963         if (!hugepage_vma_check(vma, vma->vm_flags))
 964                 return SCAN_VMA_CHECK;
 965         /* Anon VMA expected */
 966         if (!vma->anon_vma || !vma_is_anonymous(vma))
 967                 return SCAN_VMA_CHECK;
 968         return 0;
 969 }
 970
 971 /*
 972  * Bring missing pages in from swap, to complete THP collapse.
 973  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 974  *
 975  * Called and returns without pte mapped or spinlocks held,
 976  * but with mmap_lock held to protect against vma changes.
 977  */
 978
 979 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 980                                         struct vm_area_struct *vma,
 981                                         unsigned long haddr, pmd_t *pmd,
 982                                         int referenced)
 983 {
 984         int swapped_in = 0;
 985         vm_fault_t ret = 0;
 986         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 987
 988         for (address = haddr; address < end; address += PAGE_SIZE) {
 989                 struct vm_fault vmf = {
 990                         .vma = vma,
 991                         .address = address,
 992                         .pgoff = linear_page_index(vma, haddr),
 993                         .flags = FAULT_FLAG_ALLOW_RETRY,
 994                         .pmd = pmd,
 995                 };
 996
 997                 vmf.pte = pte_offset_map(pmd, address);
 998                 vmf.orig_pte = *vmf.pte;
 999                 if (!is_swap_pte(vmf.orig_pte)) {
1000                         pte_unmap(vmf.pte);
1001                         continue;
1002                 }
1003                 swapped_in++;
1004                 ret = do_swap_page(&vmf);
1005
1006                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1007                 if (ret & VM_FAULT_RETRY) {
1008                         mmap_read_lock(mm);
1009                         if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1010                                 /* vma is no longer available, don't continue to swapin */
1011                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1012                                 return false;
1013                         }
1014                         /* check if the pmd is still valid */
1015                         if (mm_find_pmd(mm, haddr) != pmd) {
1016                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1017                                 return false;
1018                         }
1019                 }
1020                 if (ret & VM_FAULT_ERROR) {
1021                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1022                         return false;
1023                 }
1024         }
1025
1026         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1027         if (swapped_in)
1028                 lru_add_drain();
1029
1030         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1031         return true;
1032 }
1033
1034 static void collapse_huge_page(struct mm_struct *mm,
1035                                    unsigned long address,
1036                                    struct page **hpage,
1037                                    int node, int referenced, int unmapped)
1038 {
1039         LIST_HEAD(compound_pagelist);
1040         pmd_t *pmd, _pmd;
1041         pte_t *pte;
1042         pgtable_t pgtable;
1043         struct page *new_page;
1044         spinlock_t *pmd_ptl, *pte_ptl;
1045         int isolated = 0, result = 0;
1046         struct vm_area_struct *vma;
1047         struct mmu_notifier_range range;
1048         gfp_t gfp;
1049
1050         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1051
1052         /* Only allocate from the target node */
1053         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1054
1055         /*
1056          * Before allocating the hugepage, release the mmap_lock read lock.
1057          * The allocation can take potentially a long time if it involves
1058          * sync compaction, and we do not need to hold the mmap_lock during
1059          * that. We will recheck the vma after taking it again in write mode.
1060          */
1061         mmap_read_unlock(mm);
1062         new_page = khugepaged_alloc_page(hpage, gfp, node);
1063         if (!new_page) {
1064                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1065                 goto out_nolock;
1066         }
1067
1068         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1069                 result = SCAN_CGROUP_CHARGE_FAIL;
1070                 goto out_nolock;
1071         }
1072         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1073
1074         mmap_read_lock(mm);
1075         result = hugepage_vma_revalidate(mm, address, &vma);
1076         if (result) {
1077                 mmap_read_unlock(mm);
1078                 goto out_nolock;
1079         }
1080
1081         pmd = mm_find_pmd(mm, address);
1082         if (!pmd) {
1083                 result = SCAN_PMD_NULL;
1084                 mmap_read_unlock(mm);
1085                 goto out_nolock;
1086         }
1087
1088         /*
1089          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1090          * If it fails, we release mmap_lock and jump out_nolock.
1091          * Continuing to collapse causes inconsistency.
1092          */
1093         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1094                                                      pmd, referenced)) {
1095                 mmap_read_unlock(mm);
1096                 goto out_nolock;
1097         }
1098
1099         mmap_read_unlock(mm);
1100         /*
1101          * Prevent all access to pagetables with the exception of
1102          * gup_fast later handled by the ptep_clear_flush and the VM
1103          * handled by the anon_vma lock + PG_lock.
1104          */
1105         mmap_write_lock(mm);
1106         result = hugepage_vma_revalidate(mm, address, &vma);
1107         if (result)
1108                 goto out_up_write;
1109         /* check if the pmd is still valid */
1110         if (mm_find_pmd(mm, address) != pmd)
1111                 goto out_up_write;
1112
1113         anon_vma_lock_write(vma->anon_vma);
1114
1115         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1116                                 address, address + HPAGE_PMD_SIZE);
1117         mmu_notifier_invalidate_range_start(&range);
1118
1119         pte = pte_offset_map(pmd, address);
1120         pte_ptl = pte_lockptr(mm, pmd);
1121
1122         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1123         /*
1124          * After this gup_fast can't run anymore. This also removes
1125          * any huge TLB entry from the CPU so we won't allow
1126          * huge and small TLB entries for the same virtual address
1127          * to avoid the risk of CPU bugs in that area.
1128          */
1129         _pmd = pmdp_collapse_flush(vma, address, pmd);
1130         spin_unlock(pmd_ptl);
1131         mmu_notifier_invalidate_range_end(&range);
1132
1133         spin_lock(pte_ptl);
1134         isolated = __collapse_huge_page_isolate(vma, address, pte,
1135                         &compound_pagelist);
1136         spin_unlock(pte_ptl);
1137
1138         if (unlikely(!isolated)) {
1139                 pte_unmap(pte);
1140                 spin_lock(pmd_ptl);
1141                 BUG_ON(!pmd_none(*pmd));
1142                 /*
1143                  * We can only use set_pmd_at when establishing
1144                  * hugepmds and never for establishing regular pmds that
1145                  * points to regular pagetables. Use pmd_populate for that
1146                  */
1147                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1148                 spin_unlock(pmd_ptl);
1149                 anon_vma_unlock_write(vma->anon_vma);
1150                 result = SCAN_FAIL;
1151                 goto out_up_write;
1152         }
1153
1154         /*
1155          * All pages are isolated and locked so anon_vma rmap
1156          * can't run anymore.
1157          */
1158         anon_vma_unlock_write(vma->anon_vma);
1159
1160         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1161                         &compound_pagelist);
1162         pte_unmap(pte);
1163         /*
1164          * spin_lock() below is not the equivalent of smp_wmb(), but
1165          * the smp_wmb() inside __SetPageUptodate() can be reused to
1166          * avoid the copy_huge_page writes to become visible after
1167          * the set_pmd_at() write.
1168          */
1169         __SetPageUptodate(new_page);
1170         pgtable = pmd_pgtable(_pmd);
1171
1172         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1173         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1174
1175         spin_lock(pmd_ptl);
1176         BUG_ON(!pmd_none(*pmd));
1177         page_add_new_anon_rmap(new_page, vma, address);
1178         lru_cache_add_inactive_or_unevictable(new_page, vma);
1179         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1180         set_pmd_at(mm, address, pmd, _pmd);
1181         update_mmu_cache_pmd(vma, address, pmd);
1182         spin_unlock(pmd_ptl);
1183
1184         *hpage = NULL;
1185
1186         khugepaged_pages_collapsed++;
1187         result = SCAN_SUCCEED;
1188 out_up_write:
1189         mmap_write_unlock(mm);
1190 out_nolock:
1191         if (!IS_ERR_OR_NULL(*hpage))
1192                 mem_cgroup_uncharge(page_folio(*hpage));
1193         trace_mm_collapse_huge_page(mm, isolated, result);
1194         return;
1195 }
1196
1197 static int khugepaged_scan_pmd(struct mm_struct *mm,
1198                                struct vm_area_struct *vma,
1199                                unsigned long address,
1200                                struct page **hpage)
1201 {
1202         pmd_t *pmd;
1203         pte_t *pte, *_pte;
1204         int ret = 0, result = 0, referenced = 0;
1205         int none_or_zero = 0, shared = 0;
1206         struct page *page = NULL;
1207         unsigned long _address;
1208         spinlock_t *ptl;
1209         int node = NUMA_NO_NODE, unmapped = 0;
1210         bool writable = false;
1211
1212         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1213
1214         pmd = mm_find_pmd(mm, address);
1215         if (!pmd) {
1216                 result = SCAN_PMD_NULL;
1217                 goto out;
1218         }
1219
1220         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1221         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1222         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1223              _pte++, _address += PAGE_SIZE) {
1224                 pte_t pteval = *_pte;
1225                 if (is_swap_pte(pteval)) {
1226                         if (++unmapped <= khugepaged_max_ptes_swap) {
1227                                 /*
1228                                  * Always be strict with uffd-wp
1229                                  * enabled swap entries.  Please see
1230                                  * comment below for pte_uffd_wp().
1231                                  */
1232                                 if (pte_swp_uffd_wp(pteval)) {
1233                                         result = SCAN_PTE_UFFD_WP;
1234                                         goto out_unmap;
1235                                 }
1236                                 continue;
1237                         } else {
1238                                 result = SCAN_EXCEED_SWAP_PTE;
1239                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1240                                 goto out_unmap;
1241                         }
1242                 }
1243                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1244                         if (!userfaultfd_armed(vma) &&
1245                             ++none_or_zero <= khugepaged_max_ptes_none) {
1246                                 continue;
1247                         } else {
1248                                 result = SCAN_EXCEED_NONE_PTE;
1249                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1250                                 goto out_unmap;
1251                         }
1252                 }
1253                 if (pte_uffd_wp(pteval)) {
1254                         /*
1255                          * Don't collapse the page if any of the small
1256                          * PTEs are armed with uffd write protection.
1257                          * Here we can also mark the new huge pmd as
1258                          * write protected if any of the small ones is
1259                          * marked but that could bring unknown
1260                          * userfault messages that falls outside of
1261                          * the registered range.  So, just be simple.
1262                          */
1263                         result = SCAN_PTE_UFFD_WP;
1264                         goto out_unmap;
1265                 }
1266                 if (pte_write(pteval))
1267                         writable = true;
1268
1269                 page = vm_normal_page(vma, _address, pteval);
1270                 if (unlikely(!page)) {
1271                         result = SCAN_PAGE_NULL;
1272                         goto out_unmap;
1273                 }
1274
1275                 if (page_mapcount(page) > 1 &&
1276                                 ++shared > khugepaged_max_ptes_shared) {
1277                         result = SCAN_EXCEED_SHARED_PTE;
1278                         count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1279                         goto out_unmap;
1280                 }
1281
1282                 page = compound_head(page);
1283
1284                 /*
1285                  * Record which node the original page is from and save this
1286                  * information to khugepaged_node_load[].
1287                  * Khugepaged will allocate hugepage from the node has the max
1288                  * hit record.
1289                  */
1290                 node = page_to_nid(page);
1291                 if (khugepaged_scan_abort(node)) {
1292                         result = SCAN_SCAN_ABORT;
1293                         goto out_unmap;
1294                 }
1295                 khugepaged_node_load[node]++;
1296                 if (!PageLRU(page)) {
1297                         result = SCAN_PAGE_LRU;
1298                         goto out_unmap;
1299                 }
1300                 if (PageLocked(page)) {
1301                         result = SCAN_PAGE_LOCK;
1302                         goto out_unmap;
1303                 }
1304                 if (!PageAnon(page)) {
1305                         result = SCAN_PAGE_ANON;
1306                         goto out_unmap;
1307                 }
1308
1309                 /*
1310                  * Check if the page has any GUP (or other external) pins.
1311                  *
1312                  * Here the check is racy it may see totmal_mapcount > refcount
1313                  * in some cases.
1314                  * For example, one process with one forked child process.
1315                  * The parent has the PMD split due to MADV_DONTNEED, then
1316                  * the child is trying unmap the whole PMD, but khugepaged
1317                  * may be scanning the parent between the child has
1318                  * PageDoubleMap flag cleared and dec the mapcount.  So
1319                  * khugepaged may see total_mapcount > refcount.
1320                  *
1321                  * But such case is ephemeral we could always retry collapse
1322                  * later.  However it may report false positive if the page
1323                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1324                  * will be done again later the risk seems low.
1325                  */
1326                 if (!is_refcount_suitable(page)) {
1327                         result = SCAN_PAGE_COUNT;
1328                         goto out_unmap;
1329                 }
1330                 if (pte_young(pteval) ||
1331                     page_is_young(page) || PageReferenced(page) ||
1332                     mmu_notifier_test_young(vma->vm_mm, address))
1333                         referenced++;
1334         }
1335         if (!writable) {
1336                 result = SCAN_PAGE_RO;
1337         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1338                 result = SCAN_LACK_REFERENCED_PAGE;
1339         } else {
1340                 result = SCAN_SUCCEED;
1341                 ret = 1;
1342         }
1343 out_unmap:
1344         pte_unmap_unlock(pte, ptl);
1345         if (ret) {
1346                 node = khugepaged_find_target_node();
1347                 /* collapse_huge_page will return with the mmap_lock released */
1348                 collapse_huge_page(mm, address, hpage, node,
1349                                 referenced, unmapped);
1350         }
1351 out:
1352         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1353                                      none_or_zero, result, unmapped);
1354         return ret;
1355 }
1356
1357 static void collect_mm_slot(struct mm_slot *mm_slot)
1358 {
1359         struct mm_struct *mm = mm_slot->mm;
1360
1361         lockdep_assert_held(&khugepaged_mm_lock);
1362
1363         if (khugepaged_test_exit(mm)) {
1364                 /* free mm_slot */
1365                 hash_del(&mm_slot->hash);
1366                 list_del(&mm_slot->mm_node);
1367
1368                 /*
1369                  * Not strictly needed because the mm exited already.
1370                  *
1371                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1372                  */
1373
1374                 /* khugepaged_mm_lock actually not necessary for the below */
1375                 free_mm_slot(mm_slot);
1376                 mmdrop(mm);
1377         }
1378 }
1379
1380 #ifdef CONFIG_SHMEM
1381 /*
1382  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1383  * khugepaged should try to collapse the page table.
1384  */
1385 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1386                                          unsigned long addr)
1387 {
1388         struct mm_slot *mm_slot;
1389
1390         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1391
1392         spin_lock(&khugepaged_mm_lock);
1393         mm_slot = get_mm_slot(mm);
1394         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1395                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1396         spin_unlock(&khugepaged_mm_lock);
1397         return 0;
1398 }
1399
1400 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1401                                   unsigned long addr, pmd_t *pmdp)
1402 {
1403         spinlock_t *ptl;
1404         pmd_t pmd;
1405
1406         mmap_assert_write_locked(mm);
1407         ptl = pmd_lock(vma->vm_mm, pmdp);
1408         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1409         spin_unlock(ptl);
1410         mm_dec_nr_ptes(mm);
1411         page_table_check_pte_clear_range(mm, addr, pmd);
1412         pte_free(mm, pmd_pgtable(pmd));
1413 }
1414
1415 /**
1416  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1417  * address haddr.
1418  *
1419  * @mm: process address space where collapse happens
1420  * @addr: THP collapse address
1421  *
1422  * This function checks whether all the PTEs in the PMD are pointing to the
1423  * right THP. If so, retract the page table so the THP can refault in with
1424  * as pmd-mapped.
1425  */
1426 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1427 {
1428         unsigned long haddr = addr & HPAGE_PMD_MASK;
1429         struct vm_area_struct *vma = find_vma(mm, haddr);
1430         struct page *hpage;
1431         pte_t *start_pte, *pte;
1432         pmd_t *pmd;
1433         spinlock_t *ptl;
1434         int count = 0;
1435         int i;
1436
1437         if (!vma || !vma->vm_file ||
1438             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1439                 return;
1440
1441         /*
1442          * This vm_flags may not have VM_HUGEPAGE if the page was not
1443          * collapsed by this mm. But we can still collapse if the page is
1444          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1445          * will not fail the vma for missing VM_HUGEPAGE
1446          */
1447         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1448                 return;
1449
1450         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1451         if (userfaultfd_wp(vma))
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, vma, 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         collapse_and_free_pmd(mm, vma, haddr, pmd);
1513 drop_hpage:
1514         unlock_page(hpage);
1515         put_page(hpage);
1516         return;
1517
1518 abort:
1519         pte_unmap_unlock(start_pte, ptl);
1520         goto drop_hpage;
1521 }
1522
1523 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1524 {
1525         struct mm_struct *mm = mm_slot->mm;
1526         int i;
1527
1528         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1529                 return;
1530
1531         if (!mmap_write_trylock(mm))
1532                 return;
1533
1534         if (unlikely(khugepaged_test_exit(mm)))
1535                 goto out;
1536
1537         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1538                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1539
1540 out:
1541         mm_slot->nr_pte_mapped_thp = 0;
1542         mmap_write_unlock(mm);
1543 }
1544
1545 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1546 {
1547         struct vm_area_struct *vma;
1548         struct mm_struct *mm;
1549         unsigned long addr;
1550         pmd_t *pmd;
1551
1552         i_mmap_lock_write(mapping);
1553         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1554                 /*
1555                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1556                  * got written to. These VMAs are likely not worth investing
1557                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1558                  * later.
1559                  *
1560                  * Not that vma->anon_vma check is racy: it can be set up after
1561                  * the check but before we took mmap_lock by the fault path.
1562                  * But page lock would prevent establishing any new ptes of the
1563                  * page, so we are safe.
1564                  *
1565                  * An alternative would be drop the check, but check that page
1566                  * table is clear before calling pmdp_collapse_flush() under
1567                  * ptl. It has higher chance to recover THP for the VMA, but
1568                  * has higher cost too.
1569                  */
1570                 if (vma->anon_vma)
1571                         continue;
1572                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1573                 if (addr & ~HPAGE_PMD_MASK)
1574                         continue;
1575                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1576                         continue;
1577                 mm = vma->vm_mm;
1578                 pmd = mm_find_pmd(mm, addr);
1579                 if (!pmd)
1580                         continue;
1581                 /*
1582                  * We need exclusive mmap_lock to retract page table.
1583                  *
1584                  * We use trylock due to lock inversion: we need to acquire
1585                  * mmap_lock while holding page lock. Fault path does it in
1586                  * reverse order. Trylock is a way to avoid deadlock.
1587                  */
1588                 if (mmap_write_trylock(mm)) {
1589                         /*
1590                          * When a vma is registered with uffd-wp, we can't
1591                          * recycle the pmd pgtable because there can be pte
1592                          * markers installed.  Skip it only, so the rest mm/vma
1593                          * can still have the same file mapped hugely, however
1594                          * it'll always mapped in small page size for uffd-wp
1595                          * registered ranges.
1596                          */
1597                         if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma))
1598                                 collapse_and_free_pmd(mm, vma, addr, pmd);
1599                         mmap_write_unlock(mm);
1600                 } else {
1601                         /* Try again later */
1602                         khugepaged_add_pte_mapped_thp(mm, addr);
1603                 }
1604         }
1605         i_mmap_unlock_write(mapping);
1606 }
1607
1608 /**
1609  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1610  *
1611  * @mm: process address space where collapse happens
1612  * @file: file that collapse on
1613  * @start: collapse start address
1614  * @hpage: new allocated huge page for collapse
1615  * @node: appointed node the new huge page allocate from
1616  *
1617  * Basic scheme is simple, details are more complex:
1618  *  - allocate and lock a new huge page;
1619  *  - scan page cache replacing old pages with the new one
1620  *    + swap/gup in pages if necessary;
1621  *    + fill in gaps;
1622  *    + keep old pages around in case rollback is required;
1623  *  - if replacing succeeds:
1624  *    + copy data over;
1625  *    + free old pages;
1626  *    + unlock huge page;
1627  *  - if replacing failed;
1628  *    + put all pages back and unfreeze them;
1629  *    + restore gaps in the page cache;
1630  *    + unlock and free huge page;
1631  */
1632 static void collapse_file(struct mm_struct *mm,
1633                 struct file *file, pgoff_t start,
1634                 struct page **hpage, int node)
1635 {
1636         struct address_space *mapping = file->f_mapping;
1637         gfp_t gfp;
1638         struct page *new_page;
1639         pgoff_t index, end = start + HPAGE_PMD_NR;
1640         LIST_HEAD(pagelist);
1641         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1642         int nr_none = 0, result = SCAN_SUCCEED;
1643         bool is_shmem = shmem_file(file);
1644         int nr;
1645
1646         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1647         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1648
1649         /* Only allocate from the target node */
1650         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1651
1652         new_page = khugepaged_alloc_page(hpage, gfp, node);
1653         if (!new_page) {
1654                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1655                 goto out;
1656         }
1657
1658         if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1659                 result = SCAN_CGROUP_CHARGE_FAIL;
1660                 goto out;
1661         }
1662         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1663
1664         /*
1665          * Ensure we have slots for all the pages in the range.  This is
1666          * almost certainly a no-op because most of the pages must be present
1667          */
1668         do {
1669                 xas_lock_irq(&xas);
1670                 xas_create_range(&xas);
1671                 if (!xas_error(&xas))
1672                         break;
1673                 xas_unlock_irq(&xas);
1674                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1675                         result = SCAN_FAIL;
1676                         goto out;
1677                 }
1678         } while (1);
1679
1680         __SetPageLocked(new_page);
1681         if (is_shmem)
1682                 __SetPageSwapBacked(new_page);
1683         new_page->index = start;
1684         new_page->mapping = mapping;
1685
1686         /*
1687          * At this point the new_page is locked and not up-to-date.
1688          * It's safe to insert it into the page cache, because nobody would
1689          * be able to map it or use it in another way until we unlock it.
1690          */
1691
1692         xas_set(&xas, start);
1693         for (index = start; index < end; index++) {
1694                 struct page *page = xas_next(&xas);
1695
1696                 VM_BUG_ON(index != xas.xa_index);
1697                 if (is_shmem) {
1698                         if (!page) {
1699                                 /*
1700                                  * Stop if extent has been truncated or
1701                                  * hole-punched, and is now completely
1702                                  * empty.
1703                                  */
1704                                 if (index == start) {
1705                                         if (!xas_next_entry(&xas, end - 1)) {
1706                                                 result = SCAN_TRUNCATED;
1707                                                 goto xa_locked;
1708                                         }
1709                                         xas_set(&xas, index);
1710                                 }
1711                                 if (!shmem_charge(mapping->host, 1)) {
1712                                         result = SCAN_FAIL;
1713                                         goto xa_locked;
1714                                 }
1715                                 xas_store(&xas, new_page);
1716                                 nr_none++;
1717                                 continue;
1718                         }
1719
1720                         if (xa_is_value(page) || !PageUptodate(page)) {
1721                                 xas_unlock_irq(&xas);
1722                                 /* swap in or instantiate fallocated page */
1723                                 if (shmem_getpage(mapping->host, index, &page,
1724                                                   SGP_NOALLOC)) {
1725                                         result = SCAN_FAIL;
1726                                         goto xa_unlocked;
1727                                 }
1728                         } else if (trylock_page(page)) {
1729                                 get_page(page);
1730                                 xas_unlock_irq(&xas);
1731                         } else {
1732                                 result = SCAN_PAGE_LOCK;
1733                                 goto xa_locked;
1734                         }
1735                 } else {        /* !is_shmem */
1736                         if (!page || xa_is_value(page)) {
1737                                 xas_unlock_irq(&xas);
1738                                 page_cache_sync_readahead(mapping, &file->f_ra,
1739                                                           file, index,
1740                                                           end - index);
1741                                 /* drain pagevecs to help isolate_lru_page() */
1742                                 lru_add_drain();
1743                                 page = find_lock_page(mapping, index);
1744                                 if (unlikely(page == NULL)) {
1745                                         result = SCAN_FAIL;
1746                                         goto xa_unlocked;
1747                                 }
1748                         } else if (PageDirty(page)) {
1749                                 /*
1750                                  * khugepaged only works on read-only fd,
1751                                  * so this page is dirty because it hasn't
1752                                  * been flushed since first write. There
1753                                  * won't be new dirty pages.
1754                                  *
1755                                  * Trigger async flush here and hope the
1756                                  * writeback is done when khugepaged
1757                                  * revisits this page.
1758                                  *
1759                                  * This is a one-off situation. We are not
1760                                  * forcing writeback in loop.
1761                                  */
1762                                 xas_unlock_irq(&xas);
1763                                 filemap_flush(mapping);
1764                                 result = SCAN_FAIL;
1765                                 goto xa_unlocked;
1766                         } else if (PageWriteback(page)) {
1767                                 xas_unlock_irq(&xas);
1768                                 result = SCAN_FAIL;
1769                                 goto xa_unlocked;
1770                         } else if (trylock_page(page)) {
1771                                 get_page(page);
1772                                 xas_unlock_irq(&xas);
1773                         } else {
1774                                 result = SCAN_PAGE_LOCK;
1775                                 goto xa_locked;
1776                         }
1777                 }
1778
1779                 /*
1780                  * The page must be locked, so we can drop the i_pages lock
1781                  * without racing with truncate.
1782                  */
1783                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1784
1785                 /* make sure the page is up to date */
1786                 if (unlikely(!PageUptodate(page))) {
1787                         result = SCAN_FAIL;
1788                         goto out_unlock;
1789                 }
1790
1791                 /*
1792                  * If file was truncated then extended, or hole-punched, before
1793                  * we locked the first page, then a THP might be there already.
1794                  */
1795                 if (PageTransCompound(page)) {
1796                         result = SCAN_PAGE_COMPOUND;
1797                         goto out_unlock;
1798                 }
1799
1800                 if (page_mapping(page) != mapping) {
1801                         result = SCAN_TRUNCATED;
1802                         goto out_unlock;
1803                 }
1804
1805                 if (!is_shmem && (PageDirty(page) ||
1806                                   PageWriteback(page))) {
1807                         /*
1808                          * khugepaged only works on read-only fd, so this
1809                          * page is dirty because it hasn't been flushed
1810                          * since first write.
1811                          */
1812                         result = SCAN_FAIL;
1813                         goto out_unlock;
1814                 }
1815
1816                 if (isolate_lru_page(page)) {
1817                         result = SCAN_DEL_PAGE_LRU;
1818                         goto out_unlock;
1819                 }
1820
1821                 if (page_has_private(page) &&
1822                     !try_to_release_page(page, GFP_KERNEL)) {
1823                         result = SCAN_PAGE_HAS_PRIVATE;
1824                         putback_lru_page(page);
1825                         goto out_unlock;
1826                 }
1827
1828                 if (page_mapped(page))
1829                         try_to_unmap(page_folio(page),
1830                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1831
1832                 xas_lock_irq(&xas);
1833                 xas_set(&xas, index);
1834
1835                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1836
1837                 /*
1838                  * The page is expected to have page_count() == 3:
1839                  *  - we hold a pin on it;
1840                  *  - one reference from page cache;
1841                  *  - one from isolate_lru_page;
1842                  */
1843                 if (!page_ref_freeze(page, 3)) {
1844                         result = SCAN_PAGE_COUNT;
1845                         xas_unlock_irq(&xas);
1846                         putback_lru_page(page);
1847                         goto out_unlock;
1848                 }
1849
1850                 /*
1851                  * Add the page to the list to be able to undo the collapse if
1852                  * something go wrong.
1853                  */
1854                 list_add_tail(&page->lru, &pagelist);
1855
1856                 /* Finally, replace with the new page. */
1857                 xas_store(&xas, new_page);
1858                 continue;
1859 out_unlock:
1860                 unlock_page(page);
1861                 put_page(page);
1862                 goto xa_unlocked;
1863         }
1864         nr = thp_nr_pages(new_page);
1865
1866         if (is_shmem)
1867                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1868         else {
1869                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1870                 filemap_nr_thps_inc(mapping);
1871                 /*
1872                  * Paired with smp_mb() in do_dentry_open() to ensure
1873                  * i_writecount is up to date and the update to nr_thps is
1874                  * visible. Ensures the page cache will be truncated if the
1875                  * file is opened writable.
1876                  */
1877                 smp_mb();
1878                 if (inode_is_open_for_write(mapping->host)) {
1879                         result = SCAN_FAIL;
1880                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1881                         filemap_nr_thps_dec(mapping);
1882                         goto xa_locked;
1883                 }
1884         }
1885
1886         if (nr_none) {
1887                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1888                 if (is_shmem)
1889                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1890         }
1891
1892         /* Join all the small entries into a single multi-index entry */
1893         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1894         xas_store(&xas, new_page);
1895 xa_locked:
1896         xas_unlock_irq(&xas);
1897 xa_unlocked:
1898
1899         /*
1900          * If collapse is successful, flush must be done now before copying.
1901          * If collapse is unsuccessful, does flush actually need to be done?
1902          * Do it anyway, to clear the state.
1903          */
1904         try_to_unmap_flush();
1905
1906         if (result == SCAN_SUCCEED) {
1907                 struct page *page, *tmp;
1908
1909                 /*
1910                  * Replacing old pages with new one has succeeded, now we
1911                  * need to copy the content and free the old pages.
1912                  */
1913                 index = start;
1914                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1915                         while (index < page->index) {
1916                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1917                                 index++;
1918                         }
1919                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1920                                         page);
1921                         list_del(&page->lru);
1922                         page->mapping = NULL;
1923                         page_ref_unfreeze(page, 1);
1924                         ClearPageActive(page);
1925                         ClearPageUnevictable(page);
1926                         unlock_page(page);
1927                         put_page(page);
1928                         index++;
1929                 }
1930                 while (index < end) {
1931                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1932                         index++;
1933                 }
1934
1935                 SetPageUptodate(new_page);
1936                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1937                 if (is_shmem)
1938                         set_page_dirty(new_page);
1939                 lru_cache_add(new_page);
1940
1941                 /*
1942                  * Remove pte page tables, so we can re-fault the page as huge.
1943                  */
1944                 retract_page_tables(mapping, start);
1945                 *hpage = NULL;
1946
1947                 khugepaged_pages_collapsed++;
1948         } else {
1949                 struct page *page;
1950
1951                 /* Something went wrong: roll back page cache changes */
1952                 xas_lock_irq(&xas);
1953                 mapping->nrpages -= nr_none;
1954
1955                 if (is_shmem)
1956                         shmem_uncharge(mapping->host, nr_none);
1957
1958                 xas_set(&xas, start);
1959                 xas_for_each(&xas, page, end - 1) {
1960                         page = list_first_entry_or_null(&pagelist,
1961                                         struct page, lru);
1962                         if (!page || xas.xa_index < page->index) {
1963                                 if (!nr_none)
1964                                         break;
1965                                 nr_none--;
1966                                 /* Put holes back where they were */
1967                                 xas_store(&xas, NULL);
1968                                 continue;
1969                         }
1970
1971                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1972
1973                         /* Unfreeze the page. */
1974                         list_del(&page->lru);
1975                         page_ref_unfreeze(page, 2);
1976                         xas_store(&xas, page);
1977                         xas_pause(&xas);
1978                         xas_unlock_irq(&xas);
1979                         unlock_page(page);
1980                         putback_lru_page(page);
1981                         xas_lock_irq(&xas);
1982                 }
1983                 VM_BUG_ON(nr_none);
1984                 xas_unlock_irq(&xas);
1985
1986                 new_page->mapping = NULL;
1987         }
1988
1989         unlock_page(new_page);
1990 out:
1991         VM_BUG_ON(!list_empty(&pagelist));
1992         if (!IS_ERR_OR_NULL(*hpage))
1993                 mem_cgroup_uncharge(page_folio(*hpage));
1994         /* TODO: tracepoints */
1995 }
1996
1997 static void khugepaged_scan_file(struct mm_struct *mm,
1998                 struct file *file, pgoff_t start, struct page **hpage)
1999 {
2000         struct page *page = NULL;
2001         struct address_space *mapping = file->f_mapping;
2002         XA_STATE(xas, &mapping->i_pages, start);
2003         int present, swap;
2004         int node = NUMA_NO_NODE;
2005         int result = SCAN_SUCCEED;
2006
2007         present = 0;
2008         swap = 0;
2009         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2010         rcu_read_lock();
2011         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2012                 if (xas_retry(&xas, page))
2013                         continue;
2014
2015                 if (xa_is_value(page)) {
2016                         if (++swap > khugepaged_max_ptes_swap) {
2017                                 result = SCAN_EXCEED_SWAP_PTE;
2018                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2019                                 break;
2020                         }
2021                         continue;
2022                 }
2023
2024                 /*
2025                  * XXX: khugepaged should compact smaller compound pages
2026                  * into a PMD sized page
2027                  */
2028                 if (PageTransCompound(page)) {
2029                         result = SCAN_PAGE_COMPOUND;
2030                         break;
2031                 }
2032
2033                 node = page_to_nid(page);
2034                 if (khugepaged_scan_abort(node)) {
2035                         result = SCAN_SCAN_ABORT;
2036                         break;
2037                 }
2038                 khugepaged_node_load[node]++;
2039
2040                 if (!PageLRU(page)) {
2041                         result = SCAN_PAGE_LRU;
2042                         break;
2043                 }
2044
2045                 if (page_count(page) !=
2046                     1 + page_mapcount(page) + page_has_private(page)) {
2047                         result = SCAN_PAGE_COUNT;
2048                         break;
2049                 }
2050
2051                 /*
2052                  * We probably should check if the page is referenced here, but
2053                  * nobody would transfer pte_young() to PageReferenced() for us.
2054                  * And rmap walk here is just too costly...
2055                  */
2056
2057                 present++;
2058
2059                 if (need_resched()) {
2060                         xas_pause(&xas);
2061                         cond_resched_rcu();
2062                 }
2063         }
2064         rcu_read_unlock();
2065
2066         if (result == SCAN_SUCCEED) {
2067                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2068                         result = SCAN_EXCEED_NONE_PTE;
2069                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2070                 } else {
2071                         node = khugepaged_find_target_node();
2072                         collapse_file(mm, file, start, hpage, node);
2073                 }
2074         }
2075
2076         /* TODO: tracepoints */
2077 }
2078 #else
2079 static void khugepaged_scan_file(struct mm_struct *mm,
2080                 struct file *file, pgoff_t start, struct page **hpage)
2081 {
2082         BUILD_BUG();
2083 }
2084
2085 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2086 {
2087 }
2088 #endif
2089
2090 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2091                                             struct page **hpage)
2092         __releases(&khugepaged_mm_lock)
2093         __acquires(&khugepaged_mm_lock)
2094 {
2095         struct mm_slot *mm_slot;
2096         struct mm_struct *mm;
2097         struct vm_area_struct *vma;
2098         int progress = 0;
2099
2100         VM_BUG_ON(!pages);
2101         lockdep_assert_held(&khugepaged_mm_lock);
2102
2103         if (khugepaged_scan.mm_slot)
2104                 mm_slot = khugepaged_scan.mm_slot;
2105         else {
2106                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2107                                      struct mm_slot, mm_node);
2108                 khugepaged_scan.address = 0;
2109                 khugepaged_scan.mm_slot = mm_slot;
2110         }
2111         spin_unlock(&khugepaged_mm_lock);
2112         khugepaged_collapse_pte_mapped_thps(mm_slot);
2113
2114         mm = mm_slot->mm;
2115         /*
2116          * Don't wait for semaphore (to avoid long wait times).  Just move to
2117          * the next mm on the list.
2118          */
2119         vma = NULL;
2120         if (unlikely(!mmap_read_trylock(mm)))
2121                 goto breakouterloop_mmap_lock;
2122         if (likely(!khugepaged_test_exit(mm)))
2123                 vma = find_vma(mm, khugepaged_scan.address);
2124
2125         progress++;
2126         for (; vma; vma = vma->vm_next) {
2127                 unsigned long hstart, hend;
2128
2129                 cond_resched();
2130                 if (unlikely(khugepaged_test_exit(mm))) {
2131                         progress++;
2132                         break;
2133                 }
2134                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2135 skip:
2136                         progress++;
2137                         continue;
2138                 }
2139                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2140                 hend = vma->vm_end & HPAGE_PMD_MASK;
2141                 if (hstart >= hend)
2142                         goto skip;
2143                 if (khugepaged_scan.address > hend)
2144                         goto skip;
2145                 if (khugepaged_scan.address < hstart)
2146                         khugepaged_scan.address = hstart;
2147                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2148                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2149                         goto skip;
2150
2151                 while (khugepaged_scan.address < hend) {
2152                         int ret;
2153                         cond_resched();
2154                         if (unlikely(khugepaged_test_exit(mm)))
2155                                 goto breakouterloop;
2156
2157                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2158                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2159                                   hend);
2160                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2161                                 struct file *file = get_file(vma->vm_file);
2162                                 pgoff_t pgoff = linear_page_index(vma,
2163                                                 khugepaged_scan.address);
2164
2165                                 mmap_read_unlock(mm);
2166                                 ret = 1;
2167                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2168                                 fput(file);
2169                         } else {
2170                                 ret = khugepaged_scan_pmd(mm, vma,
2171                                                 khugepaged_scan.address,
2172                                                 hpage);
2173                         }
2174                         /* move to next address */
2175                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2176                         progress += HPAGE_PMD_NR;
2177                         if (ret)
2178                                 /* we released mmap_lock so break loop */
2179                                 goto breakouterloop_mmap_lock;
2180                         if (progress >= pages)
2181                                 goto breakouterloop;
2182                 }
2183         }
2184 breakouterloop:
2185         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2186 breakouterloop_mmap_lock:
2187
2188         spin_lock(&khugepaged_mm_lock);
2189         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2190         /*
2191          * Release the current mm_slot if this mm is about to die, or
2192          * if we scanned all vmas of this mm.
2193          */
2194         if (khugepaged_test_exit(mm) || !vma) {
2195                 /*
2196                  * Make sure that if mm_users is reaching zero while
2197                  * khugepaged runs here, khugepaged_exit will find
2198                  * mm_slot not pointing to the exiting mm.
2199                  */
2200                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2201                         khugepaged_scan.mm_slot = list_entry(
2202                                 mm_slot->mm_node.next,
2203                                 struct mm_slot, mm_node);
2204                         khugepaged_scan.address = 0;
2205                 } else {
2206                         khugepaged_scan.mm_slot = NULL;
2207                         khugepaged_full_scans++;
2208                 }
2209
2210                 collect_mm_slot(mm_slot);
2211         }
2212
2213         return progress;
2214 }
2215
2216 static int khugepaged_has_work(void)
2217 {
2218         return !list_empty(&khugepaged_scan.mm_head) &&
2219                 khugepaged_enabled();
2220 }
2221
2222 static int khugepaged_wait_event(void)
2223 {
2224         return !list_empty(&khugepaged_scan.mm_head) ||
2225                 kthread_should_stop();
2226 }
2227
2228 static void khugepaged_do_scan(void)
2229 {
2230         struct page *hpage = NULL;
2231         unsigned int progress = 0, pass_through_head = 0;
2232         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2233         bool wait = true;
2234
2235         lru_add_drain_all();
2236
2237         while (progress < pages) {
2238                 if (!khugepaged_prealloc_page(&hpage, &wait))
2239                         break;
2240
2241                 cond_resched();
2242
2243                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2244                         break;
2245
2246                 spin_lock(&khugepaged_mm_lock);
2247                 if (!khugepaged_scan.mm_slot)
2248                         pass_through_head++;
2249                 if (khugepaged_has_work() &&
2250                     pass_through_head < 2)
2251                         progress += khugepaged_scan_mm_slot(pages - progress,
2252                                                             &hpage);
2253                 else
2254                         progress = pages;
2255                 spin_unlock(&khugepaged_mm_lock);
2256         }
2257
2258         if (!IS_ERR_OR_NULL(hpage))
2259                 put_page(hpage);
2260 }
2261
2262 static bool khugepaged_should_wakeup(void)
2263 {
2264         return kthread_should_stop() ||
2265                time_after_eq(jiffies, khugepaged_sleep_expire);
2266 }
2267
2268 static void khugepaged_wait_work(void)
2269 {
2270         if (khugepaged_has_work()) {
2271                 const unsigned long scan_sleep_jiffies =
2272                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2273
2274                 if (!scan_sleep_jiffies)
2275                         return;
2276
2277                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2278                 wait_event_freezable_timeout(khugepaged_wait,
2279                                              khugepaged_should_wakeup(),
2280                                              scan_sleep_jiffies);
2281                 return;
2282         }
2283
2284         if (khugepaged_enabled())
2285                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2286 }
2287
2288 static int khugepaged(void *none)
2289 {
2290         struct mm_slot *mm_slot;
2291
2292         set_freezable();
2293         set_user_nice(current, MAX_NICE);
2294
2295         while (!kthread_should_stop()) {
2296                 khugepaged_do_scan();
2297                 khugepaged_wait_work();
2298         }
2299
2300         spin_lock(&khugepaged_mm_lock);
2301         mm_slot = khugepaged_scan.mm_slot;
2302         khugepaged_scan.mm_slot = NULL;
2303         if (mm_slot)
2304                 collect_mm_slot(mm_slot);
2305         spin_unlock(&khugepaged_mm_lock);
2306         return 0;
2307 }
2308
2309 static void set_recommended_min_free_kbytes(void)
2310 {
2311         struct zone *zone;
2312         int nr_zones = 0;
2313         unsigned long recommended_min;
2314
2315         if (!khugepaged_enabled()) {
2316                 calculate_min_free_kbytes();
2317                 goto update_wmarks;
2318         }
2319
2320         for_each_populated_zone(zone) {
2321                 /*
2322                  * We don't need to worry about fragmentation of
2323                  * ZONE_MOVABLE since it only has movable pages.
2324                  */
2325                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2326                         continue;
2327
2328                 nr_zones++;
2329         }
2330
2331         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2332         recommended_min = pageblock_nr_pages * nr_zones * 2;
2333
2334         /*
2335          * Make sure that on average at least two pageblocks are almost free
2336          * of another type, one for a migratetype to fall back to and a
2337          * second to avoid subsequent fallbacks of other types There are 3
2338          * MIGRATE_TYPES we care about.
2339          */
2340         recommended_min += pageblock_nr_pages * nr_zones *
2341                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2342
2343         /* don't ever allow to reserve more than 5% of the lowmem */
2344         recommended_min = min(recommended_min,
2345                               (unsigned long) nr_free_buffer_pages() / 20);
2346         recommended_min <<= (PAGE_SHIFT-10);
2347
2348         if (recommended_min > min_free_kbytes) {
2349                 if (user_min_free_kbytes >= 0)
2350                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2351                                 min_free_kbytes, recommended_min);
2352
2353                 min_free_kbytes = recommended_min;
2354         }
2355
2356 update_wmarks:
2357         setup_per_zone_wmarks();
2358 }
2359
2360 int start_stop_khugepaged(void)
2361 {
2362         int err = 0;
2363
2364         mutex_lock(&khugepaged_mutex);
2365         if (khugepaged_enabled()) {
2366                 if (!khugepaged_thread)
2367                         khugepaged_thread = kthread_run(khugepaged, NULL,
2368                                                         "khugepaged");
2369                 if (IS_ERR(khugepaged_thread)) {
2370                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2371                         err = PTR_ERR(khugepaged_thread);
2372                         khugepaged_thread = NULL;
2373                         goto fail;
2374                 }
2375
2376                 if (!list_empty(&khugepaged_scan.mm_head))
2377                         wake_up_interruptible(&khugepaged_wait);
2378         } else if (khugepaged_thread) {
2379                 kthread_stop(khugepaged_thread);
2380                 khugepaged_thread = NULL;
2381         }
2382         set_recommended_min_free_kbytes();
2383 fail:
2384         mutex_unlock(&khugepaged_mutex);
2385         return err;
2386 }
2387
2388 void khugepaged_min_free_kbytes_update(void)
2389 {
2390         mutex_lock(&khugepaged_mutex);
2391         if (khugepaged_enabled() && khugepaged_thread)
2392                 set_recommended_min_free_kbytes();
2393         mutex_unlock(&khugepaged_mutex);
2394 }