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