mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_EXCEED_SWAP_PTE,
  32         SCAN_EXCEED_SHARED_PTE,
  33         SCAN_PTE_NON_PRESENT,
  34         SCAN_PTE_UFFD_WP,
  35         SCAN_PAGE_RO,
  36         SCAN_LACK_REFERENCED_PAGE,
  37         SCAN_PAGE_NULL,
  38         SCAN_SCAN_ABORT,
  39         SCAN_PAGE_COUNT,
  40         SCAN_PAGE_LRU,
  41         SCAN_PAGE_LOCK,
  42         SCAN_PAGE_ANON,
  43         SCAN_PAGE_COMPOUND,
  44         SCAN_ANY_PROCESS,
  45         SCAN_VMA_NULL,
  46         SCAN_VMA_CHECK,
  47         SCAN_ADDRESS_RANGE,
  48         SCAN_SWAP_CACHE_PAGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  */
  94 struct mm_slot {
  95         struct hlist_node hash;
  96         struct list_head mm_node;
  97         struct mm_struct *mm;
  98
  99         /* pte-mapped THP in this mm */
 100         int nr_pte_mapped_thp;
 101         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 102 };
 103
 104 /**
 105  * struct khugepaged_scan - cursor for scanning
 106  * @mm_head: the head of the mm list to scan
 107  * @mm_slot: the current mm_slot we are scanning
 108  * @address: the next address inside that to be scanned
 109  *
 110  * There is only the one khugepaged_scan instance of this cursor structure.
 111  */
 112 struct khugepaged_scan {
 113         struct list_head mm_head;
 114         struct mm_slot *mm_slot;
 115         unsigned long address;
 116 };
 117
 118 static struct khugepaged_scan khugepaged_scan = {
 119         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 120 };
 121
 122 #ifdef CONFIG_SYSFS
 123 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 124                                          struct kobj_attribute *attr,
 125                                          char *buf)
 126 {
 127         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 128 }
 129
 130 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 131                                           struct kobj_attribute *attr,
 132                                           const char *buf, size_t count)
 133 {
 134         unsigned long msecs;
 135         int err;
 136
 137         err = kstrtoul(buf, 10, &msecs);
 138         if (err || msecs > UINT_MAX)
 139                 return -EINVAL;
 140
 141         khugepaged_scan_sleep_millisecs = msecs;
 142         khugepaged_sleep_expire = 0;
 143         wake_up_interruptible(&khugepaged_wait);
 144
 145         return count;
 146 }
 147 static struct kobj_attribute scan_sleep_millisecs_attr =
 148         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 149                scan_sleep_millisecs_store);
 150
 151 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 152                                           struct kobj_attribute *attr,
 153                                           char *buf)
 154 {
 155         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 156 }
 157
 158 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 159                                            struct kobj_attribute *attr,
 160                                            const char *buf, size_t count)
 161 {
 162         unsigned long msecs;
 163         int err;
 164
 165         err = kstrtoul(buf, 10, &msecs);
 166         if (err || msecs > UINT_MAX)
 167                 return -EINVAL;
 168
 169         khugepaged_alloc_sleep_millisecs = msecs;
 170         khugepaged_sleep_expire = 0;
 171         wake_up_interruptible(&khugepaged_wait);
 172
 173         return count;
 174 }
 175 static struct kobj_attribute alloc_sleep_millisecs_attr =
 176         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 177                alloc_sleep_millisecs_store);
 178
 179 static ssize_t pages_to_scan_show(struct kobject *kobj,
 180                                   struct kobj_attribute *attr,
 181                                   char *buf)
 182 {
 183         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 184 }
 185 static ssize_t pages_to_scan_store(struct kobject *kobj,
 186                                    struct kobj_attribute *attr,
 187                                    const char *buf, size_t count)
 188 {
 189         int err;
 190         unsigned long pages;
 191
 192         err = kstrtoul(buf, 10, &pages);
 193         if (err || !pages || pages > UINT_MAX)
 194                 return -EINVAL;
 195
 196         khugepaged_pages_to_scan = pages;
 197
 198         return count;
 199 }
 200 static struct kobj_attribute pages_to_scan_attr =
 201         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 202                pages_to_scan_store);
 203
 204 static ssize_t pages_collapsed_show(struct kobject *kobj,
 205                                     struct kobj_attribute *attr,
 206                                     char *buf)
 207 {
 208         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 209 }
 210 static struct kobj_attribute pages_collapsed_attr =
 211         __ATTR_RO(pages_collapsed);
 212
 213 static ssize_t full_scans_show(struct kobject *kobj,
 214                                struct kobj_attribute *attr,
 215                                char *buf)
 216 {
 217         return sprintf(buf, "%u\n", khugepaged_full_scans);
 218 }
 219 static struct kobj_attribute full_scans_attr =
 220         __ATTR_RO(full_scans);
 221
 222 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 223                                       struct kobj_attribute *attr, char *buf)
 224 {
 225         return single_hugepage_flag_show(kobj, attr, buf,
 226                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 227 }
 228 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 229                                        struct kobj_attribute *attr,
 230                                        const char *buf, size_t count)
 231 {
 232         return single_hugepage_flag_store(kobj, attr, buf, count,
 233                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 234 }
 235 static struct kobj_attribute khugepaged_defrag_attr =
 236         __ATTR(defrag, 0644, khugepaged_defrag_show,
 237                khugepaged_defrag_store);
 238
 239 /*
 240  * max_ptes_none controls if khugepaged should collapse hugepages over
 241  * any unmapped ptes in turn potentially increasing the memory
 242  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 243  * reduce the available free memory in the system as it
 244  * runs. Increasing max_ptes_none will instead potentially reduce the
 245  * free memory in the system during the khugepaged scan.
 246  */
 247 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 248                                              struct kobj_attribute *attr,
 249                                              char *buf)
 250 {
 251         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 252 }
 253 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 254                                               struct kobj_attribute *attr,
 255                                               const char *buf, size_t count)
 256 {
 257         int err;
 258         unsigned long max_ptes_none;
 259
 260         err = kstrtoul(buf, 10, &max_ptes_none);
 261         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 262                 return -EINVAL;
 263
 264         khugepaged_max_ptes_none = max_ptes_none;
 265
 266         return count;
 267 }
 268 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 269         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 270                khugepaged_max_ptes_none_store);
 271
 272 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 273                                              struct kobj_attribute *attr,
 274                                              char *buf)
 275 {
 276         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 277 }
 278
 279 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 280                                               struct kobj_attribute *attr,
 281                                               const char *buf, size_t count)
 282 {
 283         int err;
 284         unsigned long max_ptes_swap;
 285
 286         err  = kstrtoul(buf, 10, &max_ptes_swap);
 287         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 288                 return -EINVAL;
 289
 290         khugepaged_max_ptes_swap = max_ptes_swap;
 291
 292         return count;
 293 }
 294
 295 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 296         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 297                khugepaged_max_ptes_swap_store);
 298
 299 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 300                                              struct kobj_attribute *attr,
 301                                              char *buf)
 302 {
 303         return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
 304 }
 305
 306 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 307                                               struct kobj_attribute *attr,
 308                                               const char *buf, size_t count)
 309 {
 310         int err;
 311         unsigned long max_ptes_shared;
 312
 313         err  = kstrtoul(buf, 10, &max_ptes_shared);
 314         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 315                 return -EINVAL;
 316
 317         khugepaged_max_ptes_shared = max_ptes_shared;
 318
 319         return count;
 320 }
 321
 322 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 323         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 324                khugepaged_max_ptes_shared_store);
 325
 326 static struct attribute *khugepaged_attr[] = {
 327         &khugepaged_defrag_attr.attr,
 328         &khugepaged_max_ptes_none_attr.attr,
 329         &khugepaged_max_ptes_swap_attr.attr,
 330         &khugepaged_max_ptes_shared_attr.attr,
 331         &pages_to_scan_attr.attr,
 332         &pages_collapsed_attr.attr,
 333         &full_scans_attr.attr,
 334         &scan_sleep_millisecs_attr.attr,
 335         &alloc_sleep_millisecs_attr.attr,
 336         NULL,
 337 };
 338
 339 struct attribute_group khugepaged_attr_group = {
 340         .attrs = khugepaged_attr,
 341         .name = "khugepaged",
 342 };
 343 #endif /* CONFIG_SYSFS */
 344
 345 int hugepage_madvise(struct vm_area_struct *vma,
 346                      unsigned long *vm_flags, int advice)
 347 {
 348         switch (advice) {
 349         case MADV_HUGEPAGE:
 350 #ifdef CONFIG_S390
 351                 /*
 352                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 353                  * can't handle this properly after s390_enable_sie, so we simply
 354                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 355                  */
 356                 if (mm_has_pgste(vma->vm_mm))
 357                         return 0;
 358 #endif
 359                 *vm_flags &= ~VM_NOHUGEPAGE;
 360                 *vm_flags |= VM_HUGEPAGE;
 361                 /*
 362                  * If the vma become good for khugepaged to scan,
 363                  * register it here without waiting a page fault that
 364                  * may not happen any time soon.
 365                  */
 366                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 367                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 368                         return -ENOMEM;
 369                 break;
 370         case MADV_NOHUGEPAGE:
 371                 *vm_flags &= ~VM_HUGEPAGE;
 372                 *vm_flags |= VM_NOHUGEPAGE;
 373                 /*
 374                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 375                  * this vma even if we leave the mm registered in khugepaged if
 376                  * it got registered before VM_NOHUGEPAGE was set.
 377                  */
 378                 break;
 379         }
 380
 381         return 0;
 382 }
 383
 384 int __init khugepaged_init(void)
 385 {
 386         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 387                                           sizeof(struct mm_slot),
 388                                           __alignof__(struct mm_slot), 0, NULL);
 389         if (!mm_slot_cache)
 390                 return -ENOMEM;
 391
 392         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 393         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 394         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 395         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 396
 397         return 0;
 398 }
 399
 400 void __init khugepaged_destroy(void)
 401 {
 402         kmem_cache_destroy(mm_slot_cache);
 403 }
 404
 405 static inline struct mm_slot *alloc_mm_slot(void)
 406 {
 407         if (!mm_slot_cache)     /* initialization failed */
 408                 return NULL;
 409         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 410 }
 411
 412 static inline void free_mm_slot(struct mm_slot *mm_slot)
 413 {
 414         kmem_cache_free(mm_slot_cache, mm_slot);
 415 }
 416
 417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 418 {
 419         struct mm_slot *mm_slot;
 420
 421         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 422                 if (mm == mm_slot->mm)
 423                         return mm_slot;
 424
 425         return NULL;
 426 }
 427
 428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 429                                     struct mm_slot *mm_slot)
 430 {
 431         mm_slot->mm = mm;
 432         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 433 }
 434
 435 static inline int khugepaged_test_exit(struct mm_struct *mm)
 436 {
 437         return atomic_read(&mm->mm_users) == 0;
 438 }
 439
 440 static bool hugepage_vma_check(struct vm_area_struct *vma,
 441                                unsigned long vm_flags)
 442 {
 443         /* Explicitly disabled through madvise. */
 444         if ((vm_flags & VM_NOHUGEPAGE) ||
 445             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 446                 return false;
 447
 448         /* Enabled via shmem mount options or sysfs settings. */
 449         if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
 450                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 451                                 HPAGE_PMD_NR);
 452         }
 453
 454         /* THP settings require madvise. */
 455         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 456                 return false;
 457
 458         /* Read-only file mappings need to be aligned for THP to work. */
 459         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 460             (vm_flags & VM_DENYWRITE)) {
 461                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 462                                 HPAGE_PMD_NR);
 463         }
 464
 465         if (!vma->anon_vma || vma->vm_ops)
 466                 return false;
 467         if (vma_is_temporary_stack(vma))
 468                 return false;
 469         return !(vm_flags & VM_NO_KHUGEPAGED);
 470 }
 471
 472 int __khugepaged_enter(struct mm_struct *mm)
 473 {
 474         struct mm_slot *mm_slot;
 475         int wakeup;
 476
 477         mm_slot = alloc_mm_slot();
 478         if (!mm_slot)
 479                 return -ENOMEM;
 480
 481         /* __khugepaged_exit() must not run from under us */
 482         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 483         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 484                 free_mm_slot(mm_slot);
 485                 return 0;
 486         }
 487
 488         spin_lock(&khugepaged_mm_lock);
 489         insert_to_mm_slots_hash(mm, mm_slot);
 490         /*
 491          * Insert just behind the scanning cursor, to let the area settle
 492          * down a little.
 493          */
 494         wakeup = list_empty(&khugepaged_scan.mm_head);
 495         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 496         spin_unlock(&khugepaged_mm_lock);
 497
 498         mmgrab(mm);
 499         if (wakeup)
 500                 wake_up_interruptible(&khugepaged_wait);
 501
 502         return 0;
 503 }
 504
 505 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 506                                unsigned long vm_flags)
 507 {
 508         unsigned long hstart, hend;
 509
 510         /*
 511          * khugepaged only supports read-only files for non-shmem files.
 512          * khugepaged does not yet work on special mappings. And
 513          * file-private shmem THP is not supported.
 514          */
 515         if (!hugepage_vma_check(vma, vm_flags))
 516                 return 0;
 517
 518         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 519         hend = vma->vm_end & HPAGE_PMD_MASK;
 520         if (hstart < hend)
 521                 return khugepaged_enter(vma, vm_flags);
 522         return 0;
 523 }
 524
 525 void __khugepaged_exit(struct mm_struct *mm)
 526 {
 527         struct mm_slot *mm_slot;
 528         int free = 0;
 529
 530         spin_lock(&khugepaged_mm_lock);
 531         mm_slot = get_mm_slot(mm);
 532         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 533                 hash_del(&mm_slot->hash);
 534                 list_del(&mm_slot->mm_node);
 535                 free = 1;
 536         }
 537         spin_unlock(&khugepaged_mm_lock);
 538
 539         if (free) {
 540                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 541                 free_mm_slot(mm_slot);
 542                 mmdrop(mm);
 543         } else if (mm_slot) {
 544                 /*
 545                  * This is required to serialize against
 546                  * khugepaged_test_exit() (which is guaranteed to run
 547                  * under mmap sem read mode). Stop here (after we
 548                  * return all pagetables will be destroyed) until
 549                  * khugepaged has finished working on the pagetables
 550                  * under the mmap_lock.
 551                  */
 552                 mmap_write_lock(mm);
 553                 mmap_write_unlock(mm);
 554         }
 555 }
 556
 557 static void release_pte_page(struct page *page)
 558 {
 559         mod_node_page_state(page_pgdat(page),
 560                         NR_ISOLATED_ANON + page_is_file_lru(page),
 561                         -compound_nr(page));
 562         unlock_page(page);
 563         putback_lru_page(page);
 564 }
 565
 566 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 567                 struct list_head *compound_pagelist)
 568 {
 569         struct page *page, *tmp;
 570
 571         while (--_pte >= pte) {
 572                 pte_t pteval = *_pte;
 573
 574                 page = pte_page(pteval);
 575                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 576                                 !PageCompound(page))
 577                         release_pte_page(page);
 578         }
 579
 580         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 581                 list_del(&page->lru);
 582                 release_pte_page(page);
 583         }
 584 }
 585
 586 static bool is_refcount_suitable(struct page *page)
 587 {
 588         int expected_refcount;
 589
 590         expected_refcount = total_mapcount(page);
 591         if (PageSwapCache(page))
 592                 expected_refcount += compound_nr(page);
 593
 594         return page_count(page) == expected_refcount;
 595 }
 596
 597 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 598                                         unsigned long address,
 599                                         pte_t *pte,
 600                                         struct list_head *compound_pagelist)
 601 {
 602         struct page *page = NULL;
 603         pte_t *_pte;
 604         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 605         bool writable = false;
 606
 607         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 608              _pte++, address += PAGE_SIZE) {
 609                 pte_t pteval = *_pte;
 610                 if (pte_none(pteval) || (pte_present(pteval) &&
 611                                 is_zero_pfn(pte_pfn(pteval)))) {
 612                         if (!userfaultfd_armed(vma) &&
 613                             ++none_or_zero <= khugepaged_max_ptes_none) {
 614                                 continue;
 615                         } else {
 616                                 result = SCAN_EXCEED_NONE_PTE;
 617                                 goto out;
 618                         }
 619                 }
 620                 if (!pte_present(pteval)) {
 621                         result = SCAN_PTE_NON_PRESENT;
 622                         goto out;
 623                 }
 624                 page = vm_normal_page(vma, address, pteval);
 625                 if (unlikely(!page)) {
 626                         result = SCAN_PAGE_NULL;
 627                         goto out;
 628                 }
 629
 630                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 631
 632                 if (page_mapcount(page) > 1 &&
 633                                 ++shared > khugepaged_max_ptes_shared) {
 634                         result = SCAN_EXCEED_SHARED_PTE;
 635                         goto out;
 636                 }
 637
 638                 if (PageCompound(page)) {
 639                         struct page *p;
 640                         page = compound_head(page);
 641
 642                         /*
 643                          * Check if we have dealt with the compound page
 644                          * already
 645                          */
 646                         list_for_each_entry(p, compound_pagelist, lru) {
 647                                 if (page == p)
 648                                         goto next;
 649                         }
 650                 }
 651
 652                 /*
 653                  * We can do it before isolate_lru_page because the
 654                  * page can't be freed from under us. NOTE: PG_lock
 655                  * is needed to serialize against split_huge_page
 656                  * when invoked from the VM.
 657                  */
 658                 if (!trylock_page(page)) {
 659                         result = SCAN_PAGE_LOCK;
 660                         goto out;
 661                 }
 662
 663                 /*
 664                  * Check if the page has any GUP (or other external) pins.
 665                  *
 666                  * The page table that maps the page has been already unlinked
 667                  * from the page table tree and this process cannot get
 668                  * an additinal pin on the page.
 669                  *
 670                  * New pins can come later if the page is shared across fork,
 671                  * but not from this process. The other process cannot write to
 672                  * the page, only trigger CoW.
 673                  */
 674                 if (!is_refcount_suitable(page)) {
 675                         unlock_page(page);
 676                         result = SCAN_PAGE_COUNT;
 677                         goto out;
 678                 }
 679                 if (!pte_write(pteval) && PageSwapCache(page) &&
 680                                 !reuse_swap_page(page, NULL)) {
 681                         /*
 682                          * Page is in the swap cache and cannot be re-used.
 683                          * It cannot be collapsed into a THP.
 684                          */
 685                         unlock_page(page);
 686                         result = SCAN_SWAP_CACHE_PAGE;
 687                         goto out;
 688                 }
 689
 690                 /*
 691                  * Isolate the page to avoid collapsing an hugepage
 692                  * currently in use by the VM.
 693                  */
 694                 if (isolate_lru_page(page)) {
 695                         unlock_page(page);
 696                         result = SCAN_DEL_PAGE_LRU;
 697                         goto out;
 698                 }
 699                 mod_node_page_state(page_pgdat(page),
 700                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 701                                 compound_nr(page));
 702                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 703                 VM_BUG_ON_PAGE(PageLRU(page), page);
 704
 705                 if (PageCompound(page))
 706                         list_add_tail(&page->lru, compound_pagelist);
 707 next:
 708                 /* There should be enough young pte to collapse the page */
 709                 if (pte_young(pteval) ||
 710                     page_is_young(page) || PageReferenced(page) ||
 711                     mmu_notifier_test_young(vma->vm_mm, address))
 712                         referenced++;
 713
 714                 if (pte_write(pteval))
 715                         writable = true;
 716         }
 717         if (likely(writable)) {
 718                 if (likely(referenced)) {
 719                         result = SCAN_SUCCEED;
 720                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 721                                                             referenced, writable, result);
 722                         return 1;
 723                 }
 724         } else {
 725                 result = SCAN_PAGE_RO;
 726         }
 727
 728 out:
 729         release_pte_pages(pte, _pte, compound_pagelist);
 730         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 731                                             referenced, writable, result);
 732         return 0;
 733 }
 734
 735 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 736                                       struct vm_area_struct *vma,
 737                                       unsigned long address,
 738                                       spinlock_t *ptl,
 739                                       struct list_head *compound_pagelist)
 740 {
 741         struct page *src_page, *tmp;
 742         pte_t *_pte;
 743         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 744                                 _pte++, page++, address += PAGE_SIZE) {
 745                 pte_t pteval = *_pte;
 746
 747                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 748                         clear_user_highpage(page, address);
 749                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 750                         if (is_zero_pfn(pte_pfn(pteval))) {
 751                                 /*
 752                                  * ptl mostly unnecessary.
 753                                  */
 754                                 spin_lock(ptl);
 755                                 /*
 756                                  * paravirt calls inside pte_clear here are
 757                                  * superfluous.
 758                                  */
 759                                 pte_clear(vma->vm_mm, address, _pte);
 760                                 spin_unlock(ptl);
 761                         }
 762                 } else {
 763                         src_page = pte_page(pteval);
 764                         copy_user_highpage(page, src_page, address, vma);
 765                         if (!PageCompound(src_page))
 766                                 release_pte_page(src_page);
 767                         /*
 768                          * ptl mostly unnecessary, but preempt has to
 769                          * be disabled to update the per-cpu stats
 770                          * inside page_remove_rmap().
 771                          */
 772                         spin_lock(ptl);
 773                         /*
 774                          * paravirt calls inside pte_clear here are
 775                          * superfluous.
 776                          */
 777                         pte_clear(vma->vm_mm, address, _pte);
 778                         page_remove_rmap(src_page, false);
 779                         spin_unlock(ptl);
 780                         free_page_and_swap_cache(src_page);
 781                 }
 782         }
 783
 784         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 785                 list_del(&src_page->lru);
 786                 release_pte_page(src_page);
 787         }
 788 }
 789
 790 static void khugepaged_alloc_sleep(void)
 791 {
 792         DEFINE_WAIT(wait);
 793
 794         add_wait_queue(&khugepaged_wait, &wait);
 795         freezable_schedule_timeout_interruptible(
 796                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 797         remove_wait_queue(&khugepaged_wait, &wait);
 798 }
 799
 800 static int khugepaged_node_load[MAX_NUMNODES];
 801
 802 static bool khugepaged_scan_abort(int nid)
 803 {
 804         int i;
 805
 806         /*
 807          * If node_reclaim_mode is disabled, then no extra effort is made to
 808          * allocate memory locally.
 809          */
 810         if (!node_reclaim_mode)
 811                 return false;
 812
 813         /* If there is a count for this node already, it must be acceptable */
 814         if (khugepaged_node_load[nid])
 815                 return false;
 816
 817         for (i = 0; i < MAX_NUMNODES; i++) {
 818                 if (!khugepaged_node_load[i])
 819                         continue;
 820                 if (node_distance(nid, i) > node_reclaim_distance)
 821                         return true;
 822         }
 823         return false;
 824 }
 825
 826 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 827 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 828 {
 829         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 830 }
 831
 832 #ifdef CONFIG_NUMA
 833 static int khugepaged_find_target_node(void)
 834 {
 835         static int last_khugepaged_target_node = NUMA_NO_NODE;
 836         int nid, target_node = 0, max_value = 0;
 837
 838         /* find first node with max normal pages hit */
 839         for (nid = 0; nid < MAX_NUMNODES; nid++)
 840                 if (khugepaged_node_load[nid] > max_value) {
 841                         max_value = khugepaged_node_load[nid];
 842                         target_node = nid;
 843                 }
 844
 845         /* do some balance if several nodes have the same hit record */
 846         if (target_node <= last_khugepaged_target_node)
 847                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 848                                 nid++)
 849                         if (max_value == khugepaged_node_load[nid]) {
 850                                 target_node = nid;
 851                                 break;
 852                         }
 853
 854         last_khugepaged_target_node = target_node;
 855         return target_node;
 856 }
 857
 858 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 859 {
 860         if (IS_ERR(*hpage)) {
 861                 if (!*wait)
 862                         return false;
 863
 864                 *wait = false;
 865                 *hpage = NULL;
 866                 khugepaged_alloc_sleep();
 867         } else if (*hpage) {
 868                 put_page(*hpage);
 869                 *hpage = NULL;
 870         }
 871
 872         return true;
 873 }
 874
 875 static struct page *
 876 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 877 {
 878         VM_BUG_ON_PAGE(*hpage, *hpage);
 879
 880         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 881         if (unlikely(!*hpage)) {
 882                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 883                 *hpage = ERR_PTR(-ENOMEM);
 884                 return NULL;
 885         }
 886
 887         prep_transhuge_page(*hpage);
 888         count_vm_event(THP_COLLAPSE_ALLOC);
 889         return *hpage;
 890 }
 891 #else
 892 static int khugepaged_find_target_node(void)
 893 {
 894         return 0;
 895 }
 896
 897 static inline struct page *alloc_khugepaged_hugepage(void)
 898 {
 899         struct page *page;
 900
 901         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 902                            HPAGE_PMD_ORDER);
 903         if (page)
 904                 prep_transhuge_page(page);
 905         return page;
 906 }
 907
 908 static struct page *khugepaged_alloc_hugepage(bool *wait)
 909 {
 910         struct page *hpage;
 911
 912         do {
 913                 hpage = alloc_khugepaged_hugepage();
 914                 if (!hpage) {
 915                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 916                         if (!*wait)
 917                                 return NULL;
 918
 919                         *wait = false;
 920                         khugepaged_alloc_sleep();
 921                 } else
 922                         count_vm_event(THP_COLLAPSE_ALLOC);
 923         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 924
 925         return hpage;
 926 }
 927
 928 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 929 {
 930         /*
 931          * If the hpage allocated earlier was briefly exposed in page cache
 932          * before collapse_file() failed, it is possible that racing lookups
 933          * have not yet completed, and would then be unpleasantly surprised by
 934          * finding the hpage reused for the same mapping at a different offset.
 935          * Just release the previous allocation if there is any danger of that.
 936          */
 937         if (*hpage && page_count(*hpage) > 1) {
 938                 put_page(*hpage);
 939                 *hpage = NULL;
 940         }
 941
 942         if (!*hpage)
 943                 *hpage = khugepaged_alloc_hugepage(wait);
 944
 945         if (unlikely(!*hpage))
 946                 return false;
 947
 948         return true;
 949 }
 950
 951 static struct page *
 952 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 953 {
 954         VM_BUG_ON(!*hpage);
 955
 956         return  *hpage;
 957 }
 958 #endif
 959
 960 /*
 961  * If mmap_lock temporarily dropped, revalidate vma
 962  * before taking mmap_lock.
 963  * Return 0 if succeeds, otherwise return none-zero
 964  * value (scan code).
 965  */
 966
 967 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 968                 struct vm_area_struct **vmap)
 969 {
 970         struct vm_area_struct *vma;
 971         unsigned long hstart, hend;
 972
 973         if (unlikely(khugepaged_test_exit(mm)))
 974                 return SCAN_ANY_PROCESS;
 975
 976         *vmap = vma = find_vma(mm, address);
 977         if (!vma)
 978                 return SCAN_VMA_NULL;
 979
 980         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 981         hend = vma->vm_end & HPAGE_PMD_MASK;
 982         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 983                 return SCAN_ADDRESS_RANGE;
 984         if (!hugepage_vma_check(vma, vma->vm_flags))
 985                 return SCAN_VMA_CHECK;
 986         /* Anon VMA expected */
 987         if (!vma->anon_vma || vma->vm_ops)
 988                 return SCAN_VMA_CHECK;
 989         return 0;
 990 }
 991
 992 /*
 993  * Bring missing pages in from swap, to complete THP collapse.
 994  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 995  *
 996  * Called and returns without pte mapped or spinlocks held,
 997  * but with mmap_lock held to protect against vma changes.
 998  */
 999
1000 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1001                                         struct vm_area_struct *vma,
1002                                         unsigned long address, pmd_t *pmd,
1003                                         int referenced)
1004 {
1005         int swapped_in = 0;
1006         vm_fault_t ret = 0;
1007         struct vm_fault vmf = {
1008                 .vma = vma,
1009                 .address = address,
1010                 .flags = FAULT_FLAG_ALLOW_RETRY,
1011                 .pmd = pmd,
1012                 .pgoff = linear_page_index(vma, address),
1013         };
1014
1015         vmf.pte = pte_offset_map(pmd, address);
1016         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1017                         vmf.pte++, vmf.address += PAGE_SIZE) {
1018                 vmf.orig_pte = *vmf.pte;
1019                 if (!is_swap_pte(vmf.orig_pte))
1020                         continue;
1021                 swapped_in++;
1022                 ret = do_swap_page(&vmf);
1023
1024                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1025                 if (ret & VM_FAULT_RETRY) {
1026                         mmap_read_lock(mm);
1027                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1028                                 /* vma is no longer available, don't continue to swapin */
1029                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1030                                 return false;
1031                         }
1032                         /* check if the pmd is still valid */
1033                         if (mm_find_pmd(mm, address) != pmd) {
1034                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1035                                 return false;
1036                         }
1037                 }
1038                 if (ret & VM_FAULT_ERROR) {
1039                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1040                         return false;
1041                 }
1042                 /* pte is unmapped now, we need to map it */
1043                 vmf.pte = pte_offset_map(pmd, vmf.address);
1044         }
1045         vmf.pte--;
1046         pte_unmap(vmf.pte);
1047
1048         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1049         if (swapped_in)
1050                 lru_add_drain();
1051
1052         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1053         return true;
1054 }
1055
1056 static void collapse_huge_page(struct mm_struct *mm,
1057                                    unsigned long address,
1058                                    struct page **hpage,
1059                                    int node, int referenced, int unmapped)
1060 {
1061         LIST_HEAD(compound_pagelist);
1062         pmd_t *pmd, _pmd;
1063         pte_t *pte;
1064         pgtable_t pgtable;
1065         struct page *new_page;
1066         spinlock_t *pmd_ptl, *pte_ptl;
1067         int isolated = 0, result = 0;
1068         struct vm_area_struct *vma;
1069         struct mmu_notifier_range range;
1070         gfp_t gfp;
1071
1072         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1073
1074         /* Only allocate from the target node */
1075         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1076
1077         /*
1078          * Before allocating the hugepage, release the mmap_lock read lock.
1079          * The allocation can take potentially a long time if it involves
1080          * sync compaction, and we do not need to hold the mmap_lock during
1081          * that. We will recheck the vma after taking it again in write mode.
1082          */
1083         mmap_read_unlock(mm);
1084         new_page = khugepaged_alloc_page(hpage, gfp, node);
1085         if (!new_page) {
1086                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1087                 goto out_nolock;
1088         }
1089
1090         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1091                 result = SCAN_CGROUP_CHARGE_FAIL;
1092                 goto out_nolock;
1093         }
1094         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1095
1096         mmap_read_lock(mm);
1097         result = hugepage_vma_revalidate(mm, address, &vma);
1098         if (result) {
1099                 mmap_read_unlock(mm);
1100                 goto out_nolock;
1101         }
1102
1103         pmd = mm_find_pmd(mm, address);
1104         if (!pmd) {
1105                 result = SCAN_PMD_NULL;
1106                 mmap_read_unlock(mm);
1107                 goto out_nolock;
1108         }
1109
1110         /*
1111          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1112          * If it fails, we release mmap_lock and jump out_nolock.
1113          * Continuing to collapse causes inconsistency.
1114          */
1115         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1116                                                      pmd, referenced)) {
1117                 mmap_read_unlock(mm);
1118                 goto out_nolock;
1119         }
1120
1121         mmap_read_unlock(mm);
1122         /*
1123          * Prevent all access to pagetables with the exception of
1124          * gup_fast later handled by the ptep_clear_flush and the VM
1125          * handled by the anon_vma lock + PG_lock.
1126          */
1127         mmap_write_lock(mm);
1128         result = hugepage_vma_revalidate(mm, address, &vma);
1129         if (result)
1130                 goto out;
1131         /* check if the pmd is still valid */
1132         if (mm_find_pmd(mm, address) != pmd)
1133                 goto out;
1134
1135         anon_vma_lock_write(vma->anon_vma);
1136
1137         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1138                                 address, address + HPAGE_PMD_SIZE);
1139         mmu_notifier_invalidate_range_start(&range);
1140
1141         pte = pte_offset_map(pmd, address);
1142         pte_ptl = pte_lockptr(mm, pmd);
1143
1144         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1145         /*
1146          * After this gup_fast can't run anymore. This also removes
1147          * any huge TLB entry from the CPU so we won't allow
1148          * huge and small TLB entries for the same virtual address
1149          * to avoid the risk of CPU bugs in that area.
1150          */
1151         _pmd = pmdp_collapse_flush(vma, address, pmd);
1152         spin_unlock(pmd_ptl);
1153         mmu_notifier_invalidate_range_end(&range);
1154
1155         spin_lock(pte_ptl);
1156         isolated = __collapse_huge_page_isolate(vma, address, pte,
1157                         &compound_pagelist);
1158         spin_unlock(pte_ptl);
1159
1160         if (unlikely(!isolated)) {
1161                 pte_unmap(pte);
1162                 spin_lock(pmd_ptl);
1163                 BUG_ON(!pmd_none(*pmd));
1164                 /*
1165                  * We can only use set_pmd_at when establishing
1166                  * hugepmds and never for establishing regular pmds that
1167                  * points to regular pagetables. Use pmd_populate for that
1168                  */
1169                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1170                 spin_unlock(pmd_ptl);
1171                 anon_vma_unlock_write(vma->anon_vma);
1172                 result = SCAN_FAIL;
1173                 goto out;
1174         }
1175
1176         /*
1177          * All pages are isolated and locked so anon_vma rmap
1178          * can't run anymore.
1179          */
1180         anon_vma_unlock_write(vma->anon_vma);
1181
1182         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1183                         &compound_pagelist);
1184         pte_unmap(pte);
1185         __SetPageUptodate(new_page);
1186         pgtable = pmd_pgtable(_pmd);
1187
1188         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1189         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1190
1191         /*
1192          * spin_lock() below is not the equivalent of smp_wmb(), so
1193          * this is needed to avoid the copy_huge_page writes to become
1194          * visible after the set_pmd_at() write.
1195          */
1196         smp_wmb();
1197
1198         spin_lock(pmd_ptl);
1199         BUG_ON(!pmd_none(*pmd));
1200         page_add_new_anon_rmap(new_page, vma, address, true);
1201         lru_cache_add_inactive_or_unevictable(new_page, vma);
1202         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1203         set_pmd_at(mm, address, pmd, _pmd);
1204         update_mmu_cache_pmd(vma, address, pmd);
1205         spin_unlock(pmd_ptl);
1206
1207         *hpage = NULL;
1208
1209         khugepaged_pages_collapsed++;
1210         result = SCAN_SUCCEED;
1211 out_up_write:
1212         mmap_write_unlock(mm);
1213 out_nolock:
1214         if (!IS_ERR_OR_NULL(*hpage))
1215                 mem_cgroup_uncharge(*hpage);
1216         trace_mm_collapse_huge_page(mm, isolated, result);
1217         return;
1218 out:
1219         goto out_up_write;
1220 }
1221
1222 static int khugepaged_scan_pmd(struct mm_struct *mm,
1223                                struct vm_area_struct *vma,
1224                                unsigned long address,
1225                                struct page **hpage)
1226 {
1227         pmd_t *pmd;
1228         pte_t *pte, *_pte;
1229         int ret = 0, result = 0, referenced = 0;
1230         int none_or_zero = 0, shared = 0;
1231         struct page *page = NULL;
1232         unsigned long _address;
1233         spinlock_t *ptl;
1234         int node = NUMA_NO_NODE, unmapped = 0;
1235         bool writable = false;
1236
1237         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1238
1239         pmd = mm_find_pmd(mm, address);
1240         if (!pmd) {
1241                 result = SCAN_PMD_NULL;
1242                 goto out;
1243         }
1244
1245         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1246         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1247         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1248              _pte++, _address += PAGE_SIZE) {
1249                 pte_t pteval = *_pte;
1250                 if (is_swap_pte(pteval)) {
1251                         if (++unmapped <= khugepaged_max_ptes_swap) {
1252                                 /*
1253                                  * Always be strict with uffd-wp
1254                                  * enabled swap entries.  Please see
1255                                  * comment below for pte_uffd_wp().
1256                                  */
1257                                 if (pte_swp_uffd_wp(pteval)) {
1258                                         result = SCAN_PTE_UFFD_WP;
1259                                         goto out_unmap;
1260                                 }
1261                                 continue;
1262                         } else {
1263                                 result = SCAN_EXCEED_SWAP_PTE;
1264                                 goto out_unmap;
1265                         }
1266                 }
1267                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1268                         if (!userfaultfd_armed(vma) &&
1269                             ++none_or_zero <= khugepaged_max_ptes_none) {
1270                                 continue;
1271                         } else {
1272                                 result = SCAN_EXCEED_NONE_PTE;
1273                                 goto out_unmap;
1274                         }
1275                 }
1276                 if (!pte_present(pteval)) {
1277                         result = SCAN_PTE_NON_PRESENT;
1278                         goto out_unmap;
1279                 }
1280                 if (pte_uffd_wp(pteval)) {
1281                         /*
1282                          * Don't collapse the page if any of the small
1283                          * PTEs are armed with uffd write protection.
1284                          * Here we can also mark the new huge pmd as
1285                          * write protected if any of the small ones is
1286                          * marked but that could bring uknown
1287                          * userfault messages that falls outside of
1288                          * the registered range.  So, just be simple.
1289                          */
1290                         result = SCAN_PTE_UFFD_WP;
1291                         goto out_unmap;
1292                 }
1293                 if (pte_write(pteval))
1294                         writable = true;
1295
1296                 page = vm_normal_page(vma, _address, pteval);
1297                 if (unlikely(!page)) {
1298                         result = SCAN_PAGE_NULL;
1299                         goto out_unmap;
1300                 }
1301
1302                 if (page_mapcount(page) > 1 &&
1303                                 ++shared > khugepaged_max_ptes_shared) {
1304                         result = SCAN_EXCEED_SHARED_PTE;
1305                         goto out_unmap;
1306                 }
1307
1308                 page = compound_head(page);
1309
1310                 /*
1311                  * Record which node the original page is from and save this
1312                  * information to khugepaged_node_load[].
1313                  * Khupaged will allocate hugepage from the node has the max
1314                  * hit record.
1315                  */
1316                 node = page_to_nid(page);
1317                 if (khugepaged_scan_abort(node)) {
1318                         result = SCAN_SCAN_ABORT;
1319                         goto out_unmap;
1320                 }
1321                 khugepaged_node_load[node]++;
1322                 if (!PageLRU(page)) {
1323                         result = SCAN_PAGE_LRU;
1324                         goto out_unmap;
1325                 }
1326                 if (PageLocked(page)) {
1327                         result = SCAN_PAGE_LOCK;
1328                         goto out_unmap;
1329                 }
1330                 if (!PageAnon(page)) {
1331                         result = SCAN_PAGE_ANON;
1332                         goto out_unmap;
1333                 }
1334
1335                 /*
1336                  * Check if the page has any GUP (or other external) pins.
1337                  *
1338                  * Here the check is racy it may see totmal_mapcount > refcount
1339                  * in some cases.
1340                  * For example, one process with one forked child process.
1341                  * The parent has the PMD split due to MADV_DONTNEED, then
1342                  * the child is trying unmap the whole PMD, but khugepaged
1343                  * may be scanning the parent between the child has
1344                  * PageDoubleMap flag cleared and dec the mapcount.  So
1345                  * khugepaged may see total_mapcount > refcount.
1346                  *
1347                  * But such case is ephemeral we could always retry collapse
1348                  * later.  However it may report false positive if the page
1349                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1350                  * will be done again later the risk seems low.
1351                  */
1352                 if (!is_refcount_suitable(page)) {
1353                         result = SCAN_PAGE_COUNT;
1354                         goto out_unmap;
1355                 }
1356                 if (pte_young(pteval) ||
1357                     page_is_young(page) || PageReferenced(page) ||
1358                     mmu_notifier_test_young(vma->vm_mm, address))
1359                         referenced++;
1360         }
1361         if (!writable) {
1362                 result = SCAN_PAGE_RO;
1363         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1364                 result = SCAN_LACK_REFERENCED_PAGE;
1365         } else {
1366                 result = SCAN_SUCCEED;
1367                 ret = 1;
1368         }
1369 out_unmap:
1370         pte_unmap_unlock(pte, ptl);
1371         if (ret) {
1372                 node = khugepaged_find_target_node();
1373                 /* collapse_huge_page will return with the mmap_lock released */
1374                 collapse_huge_page(mm, address, hpage, node,
1375                                 referenced, unmapped);
1376         }
1377 out:
1378         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1379                                      none_or_zero, result, unmapped);
1380         return ret;
1381 }
1382
1383 static void collect_mm_slot(struct mm_slot *mm_slot)
1384 {
1385         struct mm_struct *mm = mm_slot->mm;
1386
1387         lockdep_assert_held(&khugepaged_mm_lock);
1388
1389         if (khugepaged_test_exit(mm)) {
1390                 /* free mm_slot */
1391                 hash_del(&mm_slot->hash);
1392                 list_del(&mm_slot->mm_node);
1393
1394                 /*
1395                  * Not strictly needed because the mm exited already.
1396                  *
1397                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1398                  */
1399
1400                 /* khugepaged_mm_lock actually not necessary for the below */
1401                 free_mm_slot(mm_slot);
1402                 mmdrop(mm);
1403         }
1404 }
1405
1406 #ifdef CONFIG_SHMEM
1407 /*
1408  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1409  * khugepaged should try to collapse the page table.
1410  */
1411 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1412                                          unsigned long addr)
1413 {
1414         struct mm_slot *mm_slot;
1415
1416         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1417
1418         spin_lock(&khugepaged_mm_lock);
1419         mm_slot = get_mm_slot(mm);
1420         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1421                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1422         spin_unlock(&khugepaged_mm_lock);
1423         return 0;
1424 }
1425
1426 /**
1427  * Try to collapse a pte-mapped THP for mm at address haddr.
1428  *
1429  * This function checks whether all the PTEs in the PMD are pointing to the
1430  * right THP. If so, retract the page table so the THP can refault in with
1431  * as pmd-mapped.
1432  */
1433 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1434 {
1435         unsigned long haddr = addr & HPAGE_PMD_MASK;
1436         struct vm_area_struct *vma = find_vma(mm, haddr);
1437         struct page *hpage;
1438         pte_t *start_pte, *pte;
1439         pmd_t *pmd, _pmd;
1440         spinlock_t *ptl;
1441         int count = 0;
1442         int i;
1443
1444         if (!vma || !vma->vm_file ||
1445             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1446                 return;
1447
1448         /*
1449          * This vm_flags may not have VM_HUGEPAGE if the page was not
1450          * collapsed by this mm. But we can still collapse if the page is
1451          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1452          * will not fail the vma for missing VM_HUGEPAGE
1453          */
1454         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1455                 return;
1456
1457         hpage = find_lock_page(vma->vm_file->f_mapping,
1458                                linear_page_index(vma, haddr));
1459         if (!hpage)
1460                 return;
1461
1462         if (!PageHead(hpage))
1463                 goto drop_hpage;
1464
1465         pmd = mm_find_pmd(mm, haddr);
1466         if (!pmd)
1467                 goto drop_hpage;
1468
1469         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1470
1471         /* step 1: check all mapped PTEs are to the right huge page */
1472         for (i = 0, addr = haddr, pte = start_pte;
1473              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1474                 struct page *page;
1475
1476                 /* empty pte, skip */
1477                 if (pte_none(*pte))
1478                         continue;
1479
1480                 /* page swapped out, abort */
1481                 if (!pte_present(*pte))
1482                         goto abort;
1483
1484                 page = vm_normal_page(vma, addr, *pte);
1485
1486                 /*
1487                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1488                  * page table, but the new page will not be a subpage of hpage.
1489                  */
1490                 if (hpage + i != page)
1491                         goto abort;
1492                 count++;
1493         }
1494
1495         /* step 2: adjust rmap */
1496         for (i = 0, addr = haddr, pte = start_pte;
1497              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1498                 struct page *page;
1499
1500                 if (pte_none(*pte))
1501                         continue;
1502                 page = vm_normal_page(vma, addr, *pte);
1503                 page_remove_rmap(page, false);
1504         }
1505
1506         pte_unmap_unlock(start_pte, ptl);
1507
1508         /* step 3: set proper refcount and mm_counters. */
1509         if (count) {
1510                 page_ref_sub(hpage, count);
1511                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1512         }
1513
1514         /* step 4: collapse pmd */
1515         ptl = pmd_lock(vma->vm_mm, pmd);
1516         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1517         spin_unlock(ptl);
1518         mm_dec_nr_ptes(mm);
1519         pte_free(mm, pmd_pgtable(_pmd));
1520
1521 drop_hpage:
1522         unlock_page(hpage);
1523         put_page(hpage);
1524         return;
1525
1526 abort:
1527         pte_unmap_unlock(start_pte, ptl);
1528         goto drop_hpage;
1529 }
1530
1531 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1532 {
1533         struct mm_struct *mm = mm_slot->mm;
1534         int i;
1535
1536         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1537                 return 0;
1538
1539         if (!mmap_write_trylock(mm))
1540                 return -EBUSY;
1541
1542         if (unlikely(khugepaged_test_exit(mm)))
1543                 goto out;
1544
1545         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1546                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1547
1548 out:
1549         mm_slot->nr_pte_mapped_thp = 0;
1550         mmap_write_unlock(mm);
1551         return 0;
1552 }
1553
1554 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1555 {
1556         struct vm_area_struct *vma;
1557         struct mm_struct *mm;
1558         unsigned long addr;
1559         pmd_t *pmd, _pmd;
1560
1561         i_mmap_lock_write(mapping);
1562         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1563                 /*
1564                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1565                  * got written to. These VMAs are likely not worth investing
1566                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1567                  * later.
1568                  *
1569                  * Not that vma->anon_vma check is racy: it can be set up after
1570                  * the check but before we took mmap_lock by the fault path.
1571                  * But page lock would prevent establishing any new ptes of the
1572                  * page, so we are safe.
1573                  *
1574                  * An alternative would be drop the check, but check that page
1575                  * table is clear before calling pmdp_collapse_flush() under
1576                  * ptl. It has higher chance to recover THP for the VMA, but
1577                  * has higher cost too.
1578                  */
1579                 if (vma->anon_vma)
1580                         continue;
1581                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1582                 if (addr & ~HPAGE_PMD_MASK)
1583                         continue;
1584                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1585                         continue;
1586                 mm = vma->vm_mm;
1587                 pmd = mm_find_pmd(mm, addr);
1588                 if (!pmd)
1589                         continue;
1590                 /*
1591                  * We need exclusive mmap_lock to retract page table.
1592                  *
1593                  * We use trylock due to lock inversion: we need to acquire
1594                  * mmap_lock while holding page lock. Fault path does it in
1595                  * reverse order. Trylock is a way to avoid deadlock.
1596                  */
1597                 if (mmap_write_trylock(mm)) {
1598                         if (!khugepaged_test_exit(mm)) {
1599                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1600                                 /* assume page table is clear */
1601                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1602                                 spin_unlock(ptl);
1603                                 mm_dec_nr_ptes(mm);
1604                                 pte_free(mm, pmd_pgtable(_pmd));
1605                         }
1606                         mmap_write_unlock(mm);
1607                 } else {
1608                         /* Try again later */
1609                         khugepaged_add_pte_mapped_thp(mm, addr);
1610                 }
1611         }
1612         i_mmap_unlock_write(mapping);
1613 }
1614
1615 /**
1616  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1617  *
1618  * Basic scheme is simple, details are more complex:
1619  *  - allocate and lock a new huge page;
1620  *  - scan page cache replacing old pages with the new one
1621  *    + swap/gup in pages if necessary;
1622  *    + fill in gaps;
1623  *    + keep old pages around in case rollback is required;
1624  *  - if replacing succeeds:
1625  *    + copy data over;
1626  *    + free old pages;
1627  *    + unlock huge page;
1628  *  - if replacing failed;
1629  *    + put all pages back and unfreeze them;
1630  *    + restore gaps in the page cache;
1631  *    + unlock and free huge page;
1632  */
1633 static void collapse_file(struct mm_struct *mm,
1634                 struct file *file, pgoff_t start,
1635                 struct page **hpage, int node)
1636 {
1637         struct address_space *mapping = file->f_mapping;
1638         gfp_t gfp;
1639         struct page *new_page;
1640         pgoff_t index, end = start + HPAGE_PMD_NR;
1641         LIST_HEAD(pagelist);
1642         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1643         int nr_none = 0, result = SCAN_SUCCEED;
1644         bool is_shmem = shmem_file(file);
1645
1646         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1647         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1648
1649         /* Only allocate from the target node */
1650         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1651
1652         new_page = khugepaged_alloc_page(hpage, gfp, node);
1653         if (!new_page) {
1654                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1655                 goto out;
1656         }
1657
1658         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1659                 result = SCAN_CGROUP_CHARGE_FAIL;
1660                 goto out;
1661         }
1662         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1663
1664         /* This will be less messy when we use multi-index entries */
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_NOHUGE)) {
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 (trylock_page(page)) {
1764                                 get_page(page);
1765                                 xas_unlock_irq(&xas);
1766                         } else {
1767                                 result = SCAN_PAGE_LOCK;
1768                                 goto xa_locked;
1769                         }
1770                 }
1771
1772                 /*
1773                  * The page must be locked, so we can drop the i_pages lock
1774                  * without racing with truncate.
1775                  */
1776                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1777
1778                 /* make sure the page is up to date */
1779                 if (unlikely(!PageUptodate(page))) {
1780                         result = SCAN_FAIL;
1781                         goto out_unlock;
1782                 }
1783
1784                 /*
1785                  * If file was truncated then extended, or hole-punched, before
1786                  * we locked the first page, then a THP might be there already.
1787                  */
1788                 if (PageTransCompound(page)) {
1789                         result = SCAN_PAGE_COMPOUND;
1790                         goto out_unlock;
1791                 }
1792
1793                 if (page_mapping(page) != mapping) {
1794                         result = SCAN_TRUNCATED;
1795                         goto out_unlock;
1796                 }
1797
1798                 if (!is_shmem && PageDirty(page)) {
1799                         /*
1800                          * khugepaged only works on read-only fd, so this
1801                          * page is dirty because it hasn't been flushed
1802                          * since first write.
1803                          */
1804                         result = SCAN_FAIL;
1805                         goto out_unlock;
1806                 }
1807
1808                 if (isolate_lru_page(page)) {
1809                         result = SCAN_DEL_PAGE_LRU;
1810                         goto out_unlock;
1811                 }
1812
1813                 if (page_has_private(page) &&
1814                     !try_to_release_page(page, GFP_KERNEL)) {
1815                         result = SCAN_PAGE_HAS_PRIVATE;
1816                         putback_lru_page(page);
1817                         goto out_unlock;
1818                 }
1819
1820                 if (page_mapped(page))
1821                         unmap_mapping_pages(mapping, index, 1, false);
1822
1823                 xas_lock_irq(&xas);
1824                 xas_set(&xas, index);
1825
1826                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1827                 VM_BUG_ON_PAGE(page_mapped(page), page);
1828
1829                 /*
1830                  * The page is expected to have page_count() == 3:
1831                  *  - we hold a pin on it;
1832                  *  - one reference from page cache;
1833                  *  - one from isolate_lru_page;
1834                  */
1835                 if (!page_ref_freeze(page, 3)) {
1836                         result = SCAN_PAGE_COUNT;
1837                         xas_unlock_irq(&xas);
1838                         putback_lru_page(page);
1839                         goto out_unlock;
1840                 }
1841
1842                 /*
1843                  * Add the page to the list to be able to undo the collapse if
1844                  * something go wrong.
1845                  */
1846                 list_add_tail(&page->lru, &pagelist);
1847
1848                 /* Finally, replace with the new page. */
1849                 xas_store(&xas, new_page);
1850                 continue;
1851 out_unlock:
1852                 unlock_page(page);
1853                 put_page(page);
1854                 goto xa_unlocked;
1855         }
1856
1857         if (is_shmem)
1858                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1859         else {
1860                 __inc_node_page_state(new_page, NR_FILE_THPS);
1861                 filemap_nr_thps_inc(mapping);
1862         }
1863
1864         if (nr_none) {
1865                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1866                 if (is_shmem)
1867                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1868         }
1869
1870 xa_locked:
1871         xas_unlock_irq(&xas);
1872 xa_unlocked:
1873
1874         if (result == SCAN_SUCCEED) {
1875                 struct page *page, *tmp;
1876
1877                 /*
1878                  * Replacing old pages with new one has succeeded, now we
1879                  * need to copy the content and free the old pages.
1880                  */
1881                 index = start;
1882                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1883                         while (index < page->index) {
1884                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1885                                 index++;
1886                         }
1887                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1888                                         page);
1889                         list_del(&page->lru);
1890                         page->mapping = NULL;
1891                         page_ref_unfreeze(page, 1);
1892                         ClearPageActive(page);
1893                         ClearPageUnevictable(page);
1894                         unlock_page(page);
1895                         put_page(page);
1896                         index++;
1897                 }
1898                 while (index < end) {
1899                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1900                         index++;
1901                 }
1902
1903                 SetPageUptodate(new_page);
1904                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1905                 if (is_shmem)
1906                         set_page_dirty(new_page);
1907                 lru_cache_add(new_page);
1908
1909                 /*
1910                  * Remove pte page tables, so we can re-fault the page as huge.
1911                  */
1912                 retract_page_tables(mapping, start);
1913                 *hpage = NULL;
1914
1915                 khugepaged_pages_collapsed++;
1916         } else {
1917                 struct page *page;
1918
1919                 /* Something went wrong: roll back page cache changes */
1920                 xas_lock_irq(&xas);
1921                 mapping->nrpages -= nr_none;
1922
1923                 if (is_shmem)
1924                         shmem_uncharge(mapping->host, nr_none);
1925
1926                 xas_set(&xas, start);
1927                 xas_for_each(&xas, page, end - 1) {
1928                         page = list_first_entry_or_null(&pagelist,
1929                                         struct page, lru);
1930                         if (!page || xas.xa_index < page->index) {
1931                                 if (!nr_none)
1932                                         break;
1933                                 nr_none--;
1934                                 /* Put holes back where they were */
1935                                 xas_store(&xas, NULL);
1936                                 continue;
1937                         }
1938
1939                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1940
1941                         /* Unfreeze the page. */
1942                         list_del(&page->lru);
1943                         page_ref_unfreeze(page, 2);
1944                         xas_store(&xas, page);
1945                         xas_pause(&xas);
1946                         xas_unlock_irq(&xas);
1947                         unlock_page(page);
1948                         putback_lru_page(page);
1949                         xas_lock_irq(&xas);
1950                 }
1951                 VM_BUG_ON(nr_none);
1952                 xas_unlock_irq(&xas);
1953
1954                 new_page->mapping = NULL;
1955         }
1956
1957         unlock_page(new_page);
1958 out:
1959         VM_BUG_ON(!list_empty(&pagelist));
1960         if (!IS_ERR_OR_NULL(*hpage))
1961                 mem_cgroup_uncharge(*hpage);
1962         /* TODO: tracepoints */
1963 }
1964
1965 static void khugepaged_scan_file(struct mm_struct *mm,
1966                 struct file *file, pgoff_t start, struct page **hpage)
1967 {
1968         struct page *page = NULL;
1969         struct address_space *mapping = file->f_mapping;
1970         XA_STATE(xas, &mapping->i_pages, start);
1971         int present, swap;
1972         int node = NUMA_NO_NODE;
1973         int result = SCAN_SUCCEED;
1974
1975         present = 0;
1976         swap = 0;
1977         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1978         rcu_read_lock();
1979         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1980                 if (xas_retry(&xas, page))
1981                         continue;
1982
1983                 if (xa_is_value(page)) {
1984                         if (++swap > khugepaged_max_ptes_swap) {
1985                                 result = SCAN_EXCEED_SWAP_PTE;
1986                                 break;
1987                         }
1988                         continue;
1989                 }
1990
1991                 if (PageTransCompound(page)) {
1992                         result = SCAN_PAGE_COMPOUND;
1993                         break;
1994                 }
1995
1996                 node = page_to_nid(page);
1997                 if (khugepaged_scan_abort(node)) {
1998                         result = SCAN_SCAN_ABORT;
1999                         break;
2000                 }
2001                 khugepaged_node_load[node]++;
2002
2003                 if (!PageLRU(page)) {
2004                         result = SCAN_PAGE_LRU;
2005                         break;
2006                 }
2007
2008                 if (page_count(page) !=
2009                     1 + page_mapcount(page) + page_has_private(page)) {
2010                         result = SCAN_PAGE_COUNT;
2011                         break;
2012                 }
2013
2014                 /*
2015                  * We probably should check if the page is referenced here, but
2016                  * nobody would transfer pte_young() to PageReferenced() for us.
2017                  * And rmap walk here is just too costly...
2018                  */
2019
2020                 present++;
2021
2022                 if (need_resched()) {
2023                         xas_pause(&xas);
2024                         cond_resched_rcu();
2025                 }
2026         }
2027         rcu_read_unlock();
2028
2029         if (result == SCAN_SUCCEED) {
2030                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2031                         result = SCAN_EXCEED_NONE_PTE;
2032                 } else {
2033                         node = khugepaged_find_target_node();
2034                         collapse_file(mm, file, start, hpage, node);
2035                 }
2036         }
2037
2038         /* TODO: tracepoints */
2039 }
2040 #else
2041 static void khugepaged_scan_file(struct mm_struct *mm,
2042                 struct file *file, pgoff_t start, struct page **hpage)
2043 {
2044         BUILD_BUG();
2045 }
2046
2047 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2048 {
2049         return 0;
2050 }
2051 #endif
2052
2053 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2054                                             struct page **hpage)
2055         __releases(&khugepaged_mm_lock)
2056         __acquires(&khugepaged_mm_lock)
2057 {
2058         struct mm_slot *mm_slot;
2059         struct mm_struct *mm;
2060         struct vm_area_struct *vma;
2061         int progress = 0;
2062
2063         VM_BUG_ON(!pages);
2064         lockdep_assert_held(&khugepaged_mm_lock);
2065
2066         if (khugepaged_scan.mm_slot)
2067                 mm_slot = khugepaged_scan.mm_slot;
2068         else {
2069                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2070                                      struct mm_slot, mm_node);
2071                 khugepaged_scan.address = 0;
2072                 khugepaged_scan.mm_slot = mm_slot;
2073         }
2074         spin_unlock(&khugepaged_mm_lock);
2075         khugepaged_collapse_pte_mapped_thps(mm_slot);
2076
2077         mm = mm_slot->mm;
2078         /*
2079          * Don't wait for semaphore (to avoid long wait times).  Just move to
2080          * the next mm on the list.
2081          */
2082         vma = NULL;
2083         if (unlikely(!mmap_read_trylock(mm)))
2084                 goto breakouterloop_mmap_lock;
2085         if (likely(!khugepaged_test_exit(mm)))
2086                 vma = find_vma(mm, khugepaged_scan.address);
2087
2088         progress++;
2089         for (; vma; vma = vma->vm_next) {
2090                 unsigned long hstart, hend;
2091
2092                 cond_resched();
2093                 if (unlikely(khugepaged_test_exit(mm))) {
2094                         progress++;
2095                         break;
2096                 }
2097                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2098 skip:
2099                         progress++;
2100                         continue;
2101                 }
2102                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2103                 hend = vma->vm_end & HPAGE_PMD_MASK;
2104                 if (hstart >= hend)
2105                         goto skip;
2106                 if (khugepaged_scan.address > hend)
2107                         goto skip;
2108                 if (khugepaged_scan.address < hstart)
2109                         khugepaged_scan.address = hstart;
2110                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2111                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2112                         goto skip;
2113
2114                 while (khugepaged_scan.address < hend) {
2115                         int ret;
2116                         cond_resched();
2117                         if (unlikely(khugepaged_test_exit(mm)))
2118                                 goto breakouterloop;
2119
2120                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2121                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2122                                   hend);
2123                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2124                                 struct file *file = get_file(vma->vm_file);
2125                                 pgoff_t pgoff = linear_page_index(vma,
2126                                                 khugepaged_scan.address);
2127
2128                                 mmap_read_unlock(mm);
2129                                 ret = 1;
2130                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2131                                 fput(file);
2132                         } else {
2133                                 ret = khugepaged_scan_pmd(mm, vma,
2134                                                 khugepaged_scan.address,
2135                                                 hpage);
2136                         }
2137                         /* move to next address */
2138                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2139                         progress += HPAGE_PMD_NR;
2140                         if (ret)
2141                                 /* we released mmap_lock so break loop */
2142                                 goto breakouterloop_mmap_lock;
2143                         if (progress >= pages)
2144                                 goto breakouterloop;
2145                 }
2146         }
2147 breakouterloop:
2148         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2149 breakouterloop_mmap_lock:
2150
2151         spin_lock(&khugepaged_mm_lock);
2152         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2153         /*
2154          * Release the current mm_slot if this mm is about to die, or
2155          * if we scanned all vmas of this mm.
2156          */
2157         if (khugepaged_test_exit(mm) || !vma) {
2158                 /*
2159                  * Make sure that if mm_users is reaching zero while
2160                  * khugepaged runs here, khugepaged_exit will find
2161                  * mm_slot not pointing to the exiting mm.
2162                  */
2163                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2164                         khugepaged_scan.mm_slot = list_entry(
2165                                 mm_slot->mm_node.next,
2166                                 struct mm_slot, mm_node);
2167                         khugepaged_scan.address = 0;
2168                 } else {
2169                         khugepaged_scan.mm_slot = NULL;
2170                         khugepaged_full_scans++;
2171                 }
2172
2173                 collect_mm_slot(mm_slot);
2174         }
2175
2176         return progress;
2177 }
2178
2179 static int khugepaged_has_work(void)
2180 {
2181         return !list_empty(&khugepaged_scan.mm_head) &&
2182                 khugepaged_enabled();
2183 }
2184
2185 static int khugepaged_wait_event(void)
2186 {
2187         return !list_empty(&khugepaged_scan.mm_head) ||
2188                 kthread_should_stop();
2189 }
2190
2191 static void khugepaged_do_scan(void)
2192 {
2193         struct page *hpage = NULL;
2194         unsigned int progress = 0, pass_through_head = 0;
2195         unsigned int pages = khugepaged_pages_to_scan;
2196         bool wait = true;
2197
2198         barrier(); /* write khugepaged_pages_to_scan to local stack */
2199
2200         lru_add_drain_all();
2201
2202         while (progress < pages) {
2203                 if (!khugepaged_prealloc_page(&hpage, &wait))
2204                         break;
2205
2206                 cond_resched();
2207
2208                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2209                         break;
2210
2211                 spin_lock(&khugepaged_mm_lock);
2212                 if (!khugepaged_scan.mm_slot)
2213                         pass_through_head++;
2214                 if (khugepaged_has_work() &&
2215                     pass_through_head < 2)
2216                         progress += khugepaged_scan_mm_slot(pages - progress,
2217                                                             &hpage);
2218                 else
2219                         progress = pages;
2220                 spin_unlock(&khugepaged_mm_lock);
2221         }
2222
2223         if (!IS_ERR_OR_NULL(hpage))
2224                 put_page(hpage);
2225 }
2226
2227 static bool khugepaged_should_wakeup(void)
2228 {
2229         return kthread_should_stop() ||
2230                time_after_eq(jiffies, khugepaged_sleep_expire);
2231 }
2232
2233 static void khugepaged_wait_work(void)
2234 {
2235         if (khugepaged_has_work()) {
2236                 const unsigned long scan_sleep_jiffies =
2237                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2238
2239                 if (!scan_sleep_jiffies)
2240                         return;
2241
2242                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2243                 wait_event_freezable_timeout(khugepaged_wait,
2244                                              khugepaged_should_wakeup(),
2245                                              scan_sleep_jiffies);
2246                 return;
2247         }
2248
2249         if (khugepaged_enabled())
2250                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2251 }
2252
2253 static int khugepaged(void *none)
2254 {
2255         struct mm_slot *mm_slot;
2256
2257         set_freezable();
2258         set_user_nice(current, MAX_NICE);
2259
2260         while (!kthread_should_stop()) {
2261                 khugepaged_do_scan();
2262                 khugepaged_wait_work();
2263         }
2264
2265         spin_lock(&khugepaged_mm_lock);
2266         mm_slot = khugepaged_scan.mm_slot;
2267         khugepaged_scan.mm_slot = NULL;
2268         if (mm_slot)
2269                 collect_mm_slot(mm_slot);
2270         spin_unlock(&khugepaged_mm_lock);
2271         return 0;
2272 }
2273
2274 static void set_recommended_min_free_kbytes(void)
2275 {
2276         struct zone *zone;
2277         int nr_zones = 0;
2278         unsigned long recommended_min;
2279
2280         for_each_populated_zone(zone) {
2281                 /*
2282                  * We don't need to worry about fragmentation of
2283                  * ZONE_MOVABLE since it only has movable pages.
2284                  */
2285                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2286                         continue;
2287
2288                 nr_zones++;
2289         }
2290
2291         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2292         recommended_min = pageblock_nr_pages * nr_zones * 2;
2293
2294         /*
2295          * Make sure that on average at least two pageblocks are almost free
2296          * of another type, one for a migratetype to fall back to and a
2297          * second to avoid subsequent fallbacks of other types There are 3
2298          * MIGRATE_TYPES we care about.
2299          */
2300         recommended_min += pageblock_nr_pages * nr_zones *
2301                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2302
2303         /* don't ever allow to reserve more than 5% of the lowmem */
2304         recommended_min = min(recommended_min,
2305                               (unsigned long) nr_free_buffer_pages() / 20);
2306         recommended_min <<= (PAGE_SHIFT-10);
2307
2308         if (recommended_min > min_free_kbytes) {
2309                 if (user_min_free_kbytes >= 0)
2310                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2311                                 min_free_kbytes, recommended_min);
2312
2313                 min_free_kbytes = recommended_min;
2314         }
2315         setup_per_zone_wmarks();
2316 }
2317
2318 int start_stop_khugepaged(void)
2319 {
2320         int err = 0;
2321
2322         mutex_lock(&khugepaged_mutex);
2323         if (khugepaged_enabled()) {
2324                 if (!khugepaged_thread)
2325                         khugepaged_thread = kthread_run(khugepaged, NULL,
2326                                                         "khugepaged");
2327                 if (IS_ERR(khugepaged_thread)) {
2328                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2329                         err = PTR_ERR(khugepaged_thread);
2330                         khugepaged_thread = NULL;
2331                         goto fail;
2332                 }
2333
2334                 if (!list_empty(&khugepaged_scan.mm_head))
2335                         wake_up_interruptible(&khugepaged_wait);
2336
2337                 set_recommended_min_free_kbytes();
2338         } else if (khugepaged_thread) {
2339                 kthread_stop(khugepaged_thread);
2340                 khugepaged_thread = NULL;
2341         }
2342 fail:
2343         mutex_unlock(&khugepaged_mutex);
2344         return err;
2345 }
2346
2347 void khugepaged_min_free_kbytes_update(void)
2348 {
2349         mutex_lock(&khugepaged_mutex);
2350         if (khugepaged_enabled() && khugepaged_thread)
2351                 set_recommended_min_free_kbytes();
2352         mutex_unlock(&khugepaged_mutex);
2353 }