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         if (!transhuge_vma_enabled(vma, vm_flags))
 444                 return false;
 445
 446         if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
 447                                 vma->vm_pgoff, HPAGE_PMD_NR))
 448                 return false;
 449
 450         /* Enabled via shmem mount options or sysfs settings. */
 451         if (shmem_file(vma->vm_file))
 452                 return shmem_huge_enabled(vma);
 453
 454         /* THP settings require madvise. */
 455         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 456                 return false;
 457
 458         /* Only regular file is valid */
 459         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 460             (vm_flags & VM_DENYWRITE)) {
 461                 struct inode *inode = vma->vm_file->f_inode;
 462
 463                 return S_ISREG(inode->i_mode);
 464         }
 465
 466         if (!vma->anon_vma || vma->vm_ops)
 467                 return false;
 468         if (vma_is_temporary_stack(vma))
 469                 return false;
 470         return !(vm_flags & VM_NO_KHUGEPAGED);
 471 }
 472
 473 int __khugepaged_enter(struct mm_struct *mm)
 474 {
 475         struct mm_slot *mm_slot;
 476         int wakeup;
 477
 478         mm_slot = alloc_mm_slot();
 479         if (!mm_slot)
 480                 return -ENOMEM;
 481
 482         /* __khugepaged_exit() must not run from under us */
 483         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 484         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 485                 free_mm_slot(mm_slot);
 486                 return 0;
 487         }
 488
 489         spin_lock(&khugepaged_mm_lock);
 490         insert_to_mm_slots_hash(mm, mm_slot);
 491         /*
 492          * Insert just behind the scanning cursor, to let the area settle
 493          * down a little.
 494          */
 495         wakeup = list_empty(&khugepaged_scan.mm_head);
 496         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 497         spin_unlock(&khugepaged_mm_lock);
 498
 499         mmgrab(mm);
 500         if (wakeup)
 501                 wake_up_interruptible(&khugepaged_wait);
 502
 503         return 0;
 504 }
 505
 506 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 507                                unsigned long vm_flags)
 508 {
 509         unsigned long hstart, hend;
 510
 511         /*
 512          * khugepaged only supports read-only files for non-shmem files.
 513          * khugepaged does not yet work on special mappings. And
 514          * file-private shmem THP is not supported.
 515          */
 516         if (!hugepage_vma_check(vma, vm_flags))
 517                 return 0;
 518
 519         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 520         hend = vma->vm_end & HPAGE_PMD_MASK;
 521         if (hstart < hend)
 522                 return khugepaged_enter(vma, vm_flags);
 523         return 0;
 524 }
 525
 526 void __khugepaged_exit(struct mm_struct *mm)
 527 {
 528         struct mm_slot *mm_slot;
 529         int free = 0;
 530
 531         spin_lock(&khugepaged_mm_lock);
 532         mm_slot = get_mm_slot(mm);
 533         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 534                 hash_del(&mm_slot->hash);
 535                 list_del(&mm_slot->mm_node);
 536                 free = 1;
 537         }
 538         spin_unlock(&khugepaged_mm_lock);
 539
 540         if (free) {
 541                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 542                 free_mm_slot(mm_slot);
 543                 mmdrop(mm);
 544         } else if (mm_slot) {
 545                 /*
 546                  * This is required to serialize against
 547                  * khugepaged_test_exit() (which is guaranteed to run
 548                  * under mmap sem read mode). Stop here (after we
 549                  * return all pagetables will be destroyed) until
 550                  * khugepaged has finished working on the pagetables
 551                  * under the mmap_lock.
 552                  */
 553                 mmap_write_lock(mm);
 554                 mmap_write_unlock(mm);
 555         }
 556 }
 557
 558 static void release_pte_page(struct page *page)
 559 {
 560         mod_node_page_state(page_pgdat(page),
 561                         NR_ISOLATED_ANON + page_is_file_lru(page),
 562                         -compound_nr(page));
 563         unlock_page(page);
 564         putback_lru_page(page);
 565 }
 566
 567 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 568                 struct list_head *compound_pagelist)
 569 {
 570         struct page *page, *tmp;
 571
 572         while (--_pte >= pte) {
 573                 pte_t pteval = *_pte;
 574
 575                 page = pte_page(pteval);
 576                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 577                                 !PageCompound(page))
 578                         release_pte_page(page);
 579         }
 580
 581         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 582                 list_del(&page->lru);
 583                 release_pte_page(page);
 584         }
 585 }
 586
 587 static bool is_refcount_suitable(struct page *page)
 588 {
 589         int expected_refcount;
 590
 591         expected_refcount = total_mapcount(page);
 592         if (PageSwapCache(page))
 593                 expected_refcount += compound_nr(page);
 594
 595         return page_count(page) == expected_refcount;
 596 }
 597
 598 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 599                                         unsigned long address,
 600                                         pte_t *pte,
 601                                         struct list_head *compound_pagelist)
 602 {
 603         struct page *page = NULL;
 604         pte_t *_pte;
 605         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 606         bool writable = false;
 607
 608         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 609              _pte++, address += PAGE_SIZE) {
 610                 pte_t pteval = *_pte;
 611                 if (pte_none(pteval) || (pte_present(pteval) &&
 612                                 is_zero_pfn(pte_pfn(pteval)))) {
 613                         if (!userfaultfd_armed(vma) &&
 614                             ++none_or_zero <= khugepaged_max_ptes_none) {
 615                                 continue;
 616                         } else {
 617                                 result = SCAN_EXCEED_NONE_PTE;
 618                                 goto out;
 619                         }
 620                 }
 621                 if (!pte_present(pteval)) {
 622                         result = SCAN_PTE_NON_PRESENT;
 623                         goto out;
 624                 }
 625                 if (pte_uffd_wp(pteval)) {
 626                         result = SCAN_PTE_UFFD_WP;
 627                         goto out;
 628                 }
 629                 page = vm_normal_page(vma, address, pteval);
 630                 if (unlikely(!page)) {
 631                         result = SCAN_PAGE_NULL;
 632                         goto out;
 633                 }
 634
 635                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 636
 637                 if (page_mapcount(page) > 1 &&
 638                                 ++shared > khugepaged_max_ptes_shared) {
 639                         result = SCAN_EXCEED_SHARED_PTE;
 640                         goto out;
 641                 }
 642
 643                 if (PageCompound(page)) {
 644                         struct page *p;
 645                         page = compound_head(page);
 646
 647                         /*
 648                          * Check if we have dealt with the compound page
 649                          * already
 650                          */
 651                         list_for_each_entry(p, compound_pagelist, lru) {
 652                                 if (page == p)
 653                                         goto next;
 654                         }
 655                 }
 656
 657                 /*
 658                  * We can do it before isolate_lru_page because the
 659                  * page can't be freed from under us. NOTE: PG_lock
 660                  * is needed to serialize against split_huge_page
 661                  * when invoked from the VM.
 662                  */
 663                 if (!trylock_page(page)) {
 664                         result = SCAN_PAGE_LOCK;
 665                         goto out;
 666                 }
 667
 668                 /*
 669                  * Check if the page has any GUP (or other external) pins.
 670                  *
 671                  * The page table that maps the page has been already unlinked
 672                  * from the page table tree and this process cannot get
 673                  * an additinal pin on the page.
 674                  *
 675                  * New pins can come later if the page is shared across fork,
 676                  * but not from this process. The other process cannot write to
 677                  * the page, only trigger CoW.
 678                  */
 679                 if (!is_refcount_suitable(page)) {
 680                         unlock_page(page);
 681                         result = SCAN_PAGE_COUNT;
 682                         goto out;
 683                 }
 684                 if (!pte_write(pteval) && PageSwapCache(page) &&
 685                                 !reuse_swap_page(page, NULL)) {
 686                         /*
 687                          * Page is in the swap cache and cannot be re-used.
 688                          * It cannot be collapsed into a THP.
 689                          */
 690                         unlock_page(page);
 691                         result = SCAN_SWAP_CACHE_PAGE;
 692                         goto out;
 693                 }
 694
 695                 /*
 696                  * Isolate the page to avoid collapsing an hugepage
 697                  * currently in use by the VM.
 698                  */
 699                 if (isolate_lru_page(page)) {
 700                         unlock_page(page);
 701                         result = SCAN_DEL_PAGE_LRU;
 702                         goto out;
 703                 }
 704                 mod_node_page_state(page_pgdat(page),
 705                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 706                                 compound_nr(page));
 707                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 708                 VM_BUG_ON_PAGE(PageLRU(page), page);
 709
 710                 if (PageCompound(page))
 711                         list_add_tail(&page->lru, compound_pagelist);
 712 next:
 713                 /* There should be enough young pte to collapse the page */
 714                 if (pte_young(pteval) ||
 715                     page_is_young(page) || PageReferenced(page) ||
 716                     mmu_notifier_test_young(vma->vm_mm, address))
 717                         referenced++;
 718
 719                 if (pte_write(pteval))
 720                         writable = true;
 721         }
 722
 723         if (unlikely(!writable)) {
 724                 result = SCAN_PAGE_RO;
 725         } else if (unlikely(!referenced)) {
 726                 result = SCAN_LACK_REFERENCED_PAGE;
 727         } else {
 728                 result = SCAN_SUCCEED;
 729                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 730                                                     referenced, writable, result);
 731                 return 1;
 732         }
 733 out:
 734         release_pte_pages(pte, _pte, compound_pagelist);
 735         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 736                                             referenced, writable, result);
 737         return 0;
 738 }
 739
 740 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 741                                       struct vm_area_struct *vma,
 742                                       unsigned long address,
 743                                       spinlock_t *ptl,
 744                                       struct list_head *compound_pagelist)
 745 {
 746         struct page *src_page, *tmp;
 747         pte_t *_pte;
 748         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 749                                 _pte++, page++, address += PAGE_SIZE) {
 750                 pte_t pteval = *_pte;
 751
 752                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 753                         clear_user_highpage(page, address);
 754                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 755                         if (is_zero_pfn(pte_pfn(pteval))) {
 756                                 /*
 757                                  * ptl mostly unnecessary.
 758                                  */
 759                                 spin_lock(ptl);
 760                                 /*
 761                                  * paravirt calls inside pte_clear here are
 762                                  * superfluous.
 763                                  */
 764                                 pte_clear(vma->vm_mm, address, _pte);
 765                                 spin_unlock(ptl);
 766                         }
 767                 } else {
 768                         src_page = pte_page(pteval);
 769                         copy_user_highpage(page, src_page, address, vma);
 770                         if (!PageCompound(src_page))
 771                                 release_pte_page(src_page);
 772                         /*
 773                          * ptl mostly unnecessary, but preempt has to
 774                          * be disabled to update the per-cpu stats
 775                          * inside page_remove_rmap().
 776                          */
 777                         spin_lock(ptl);
 778                         /*
 779                          * paravirt calls inside pte_clear here are
 780                          * superfluous.
 781                          */
 782                         pte_clear(vma->vm_mm, address, _pte);
 783                         page_remove_rmap(src_page, false);
 784                         spin_unlock(ptl);
 785                         free_page_and_swap_cache(src_page);
 786                 }
 787         }
 788
 789         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 790                 list_del(&src_page->lru);
 791                 release_pte_page(src_page);
 792         }
 793 }
 794
 795 static void khugepaged_alloc_sleep(void)
 796 {
 797         DEFINE_WAIT(wait);
 798
 799         add_wait_queue(&khugepaged_wait, &wait);
 800         freezable_schedule_timeout_interruptible(
 801                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 802         remove_wait_queue(&khugepaged_wait, &wait);
 803 }
 804
 805 static int khugepaged_node_load[MAX_NUMNODES];
 806
 807 static bool khugepaged_scan_abort(int nid)
 808 {
 809         int i;
 810
 811         /*
 812          * If node_reclaim_mode is disabled, then no extra effort is made to
 813          * allocate memory locally.
 814          */
 815         if (!node_reclaim_mode)
 816                 return false;
 817
 818         /* If there is a count for this node already, it must be acceptable */
 819         if (khugepaged_node_load[nid])
 820                 return false;
 821
 822         for (i = 0; i < MAX_NUMNODES; i++) {
 823                 if (!khugepaged_node_load[i])
 824                         continue;
 825                 if (node_distance(nid, i) > node_reclaim_distance)
 826                         return true;
 827         }
 828         return false;
 829 }
 830
 831 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 832 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 833 {
 834         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 835 }
 836
 837 #ifdef CONFIG_NUMA
 838 static int khugepaged_find_target_node(void)
 839 {
 840         static int last_khugepaged_target_node = NUMA_NO_NODE;
 841         int nid, target_node = 0, max_value = 0;
 842
 843         /* find first node with max normal pages hit */
 844         for (nid = 0; nid < MAX_NUMNODES; nid++)
 845                 if (khugepaged_node_load[nid] > max_value) {
 846                         max_value = khugepaged_node_load[nid];
 847                         target_node = nid;
 848                 }
 849
 850         /* do some balance if several nodes have the same hit record */
 851         if (target_node <= last_khugepaged_target_node)
 852                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 853                                 nid++)
 854                         if (max_value == khugepaged_node_load[nid]) {
 855                                 target_node = nid;
 856                                 break;
 857                         }
 858
 859         last_khugepaged_target_node = target_node;
 860         return target_node;
 861 }
 862
 863 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 864 {
 865         if (IS_ERR(*hpage)) {
 866                 if (!*wait)
 867                         return false;
 868
 869                 *wait = false;
 870                 *hpage = NULL;
 871                 khugepaged_alloc_sleep();
 872         } else if (*hpage) {
 873                 put_page(*hpage);
 874                 *hpage = NULL;
 875         }
 876
 877         return true;
 878 }
 879
 880 static struct page *
 881 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 882 {
 883         VM_BUG_ON_PAGE(*hpage, *hpage);
 884
 885         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 886         if (unlikely(!*hpage)) {
 887                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 888                 *hpage = ERR_PTR(-ENOMEM);
 889                 return NULL;
 890         }
 891
 892         prep_transhuge_page(*hpage);
 893         count_vm_event(THP_COLLAPSE_ALLOC);
 894         return *hpage;
 895 }
 896 #else
 897 static int khugepaged_find_target_node(void)
 898 {
 899         return 0;
 900 }
 901
 902 static inline struct page *alloc_khugepaged_hugepage(void)
 903 {
 904         struct page *page;
 905
 906         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 907                            HPAGE_PMD_ORDER);
 908         if (page)
 909                 prep_transhuge_page(page);
 910         return page;
 911 }
 912
 913 static struct page *khugepaged_alloc_hugepage(bool *wait)
 914 {
 915         struct page *hpage;
 916
 917         do {
 918                 hpage = alloc_khugepaged_hugepage();
 919                 if (!hpage) {
 920                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 921                         if (!*wait)
 922                                 return NULL;
 923
 924                         *wait = false;
 925                         khugepaged_alloc_sleep();
 926                 } else
 927                         count_vm_event(THP_COLLAPSE_ALLOC);
 928         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 929
 930         return hpage;
 931 }
 932
 933 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 934 {
 935         /*
 936          * If the hpage allocated earlier was briefly exposed in page cache
 937          * before collapse_file() failed, it is possible that racing lookups
 938          * have not yet completed, and would then be unpleasantly surprised by
 939          * finding the hpage reused for the same mapping at a different offset.
 940          * Just release the previous allocation if there is any danger of that.
 941          */
 942         if (*hpage && page_count(*hpage) > 1) {
 943                 put_page(*hpage);
 944                 *hpage = NULL;
 945         }
 946
 947         if (!*hpage)
 948                 *hpage = khugepaged_alloc_hugepage(wait);
 949
 950         if (unlikely(!*hpage))
 951                 return false;
 952
 953         return true;
 954 }
 955
 956 static struct page *
 957 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 958 {
 959         VM_BUG_ON(!*hpage);
 960
 961         return  *hpage;
 962 }
 963 #endif
 964
 965 /*
 966  * If mmap_lock temporarily dropped, revalidate vma
 967  * before taking mmap_lock.
 968  * Return 0 if succeeds, otherwise return none-zero
 969  * value (scan code).
 970  */
 971
 972 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 973                 struct vm_area_struct **vmap)
 974 {
 975         struct vm_area_struct *vma;
 976         unsigned long hstart, hend;
 977
 978         if (unlikely(khugepaged_test_exit(mm)))
 979                 return SCAN_ANY_PROCESS;
 980
 981         *vmap = vma = find_vma(mm, address);
 982         if (!vma)
 983                 return SCAN_VMA_NULL;
 984
 985         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 986         hend = vma->vm_end & HPAGE_PMD_MASK;
 987         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 988                 return SCAN_ADDRESS_RANGE;
 989         if (!hugepage_vma_check(vma, vma->vm_flags))
 990                 return SCAN_VMA_CHECK;
 991         /* Anon VMA expected */
 992         if (!vma->anon_vma || vma->vm_ops)
 993                 return SCAN_VMA_CHECK;
 994         return 0;
 995 }
 996
 997 /*
 998  * Bring missing pages in from swap, to complete THP collapse.
 999  * Only done if khugepaged_scan_pmd believes it is worthwhile.
1000  *
1001  * Called and returns without pte mapped or spinlocks held,
1002  * but with mmap_lock held to protect against vma changes.
1003  */
1004
1005 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1006                                         struct vm_area_struct *vma,
1007                                         unsigned long address, pmd_t *pmd,
1008                                         int referenced)
1009 {
1010         int swapped_in = 0;
1011         vm_fault_t ret = 0;
1012         struct vm_fault vmf = {
1013                 .vma = vma,
1014                 .address = address,
1015                 .flags = FAULT_FLAG_ALLOW_RETRY,
1016                 .pmd = pmd,
1017                 .pgoff = linear_page_index(vma, address),
1018         };
1019
1020         vmf.pte = pte_offset_map(pmd, address);
1021         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1022                         vmf.pte++, vmf.address += PAGE_SIZE) {
1023                 vmf.orig_pte = *vmf.pte;
1024                 if (!is_swap_pte(vmf.orig_pte))
1025                         continue;
1026                 swapped_in++;
1027                 ret = do_swap_page(&vmf);
1028
1029                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1030                 if (ret & VM_FAULT_RETRY) {
1031                         mmap_read_lock(mm);
1032                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1033                                 /* vma is no longer available, don't continue to swapin */
1034                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1035                                 return false;
1036                         }
1037                         /* check if the pmd is still valid */
1038                         if (mm_find_pmd(mm, address) != pmd) {
1039                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1040                                 return false;
1041                         }
1042                 }
1043                 if (ret & VM_FAULT_ERROR) {
1044                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1045                         return false;
1046                 }
1047                 /* pte is unmapped now, we need to map it */
1048                 vmf.pte = pte_offset_map(pmd, vmf.address);
1049         }
1050         vmf.pte--;
1051         pte_unmap(vmf.pte);
1052
1053         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1054         if (swapped_in)
1055                 lru_add_drain();
1056
1057         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1058         return true;
1059 }
1060
1061 static void collapse_huge_page(struct mm_struct *mm,
1062                                    unsigned long address,
1063                                    struct page **hpage,
1064                                    int node, int referenced, int unmapped)
1065 {
1066         LIST_HEAD(compound_pagelist);
1067         pmd_t *pmd, _pmd;
1068         pte_t *pte;
1069         pgtable_t pgtable;
1070         struct page *new_page;
1071         spinlock_t *pmd_ptl, *pte_ptl;
1072         int isolated = 0, result = 0;
1073         struct vm_area_struct *vma;
1074         struct mmu_notifier_range range;
1075         gfp_t gfp;
1076
1077         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1078
1079         /* Only allocate from the target node */
1080         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1081
1082         /*
1083          * Before allocating the hugepage, release the mmap_lock read lock.
1084          * The allocation can take potentially a long time if it involves
1085          * sync compaction, and we do not need to hold the mmap_lock during
1086          * that. We will recheck the vma after taking it again in write mode.
1087          */
1088         mmap_read_unlock(mm);
1089         new_page = khugepaged_alloc_page(hpage, gfp, node);
1090         if (!new_page) {
1091                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1092                 goto out_nolock;
1093         }
1094
1095         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1096                 result = SCAN_CGROUP_CHARGE_FAIL;
1097                 goto out_nolock;
1098         }
1099         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1100
1101         mmap_read_lock(mm);
1102         result = hugepage_vma_revalidate(mm, address, &vma);
1103         if (result) {
1104                 mmap_read_unlock(mm);
1105                 goto out_nolock;
1106         }
1107
1108         pmd = mm_find_pmd(mm, address);
1109         if (!pmd) {
1110                 result = SCAN_PMD_NULL;
1111                 mmap_read_unlock(mm);
1112                 goto out_nolock;
1113         }
1114
1115         /*
1116          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1117          * If it fails, we release mmap_lock and jump out_nolock.
1118          * Continuing to collapse causes inconsistency.
1119          */
1120         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1121                                                      pmd, referenced)) {
1122                 mmap_read_unlock(mm);
1123                 goto out_nolock;
1124         }
1125
1126         mmap_read_unlock(mm);
1127         /*
1128          * Prevent all access to pagetables with the exception of
1129          * gup_fast later handled by the ptep_clear_flush and the VM
1130          * handled by the anon_vma lock + PG_lock.
1131          */
1132         mmap_write_lock(mm);
1133         result = hugepage_vma_revalidate(mm, address, &vma);
1134         if (result)
1135                 goto out;
1136         /* check if the pmd is still valid */
1137         if (mm_find_pmd(mm, address) != pmd)
1138                 goto out;
1139
1140         anon_vma_lock_write(vma->anon_vma);
1141
1142         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1143                                 address, address + HPAGE_PMD_SIZE);
1144         mmu_notifier_invalidate_range_start(&range);
1145
1146         pte = pte_offset_map(pmd, address);
1147         pte_ptl = pte_lockptr(mm, pmd);
1148
1149         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1150         /*
1151          * This removes any huge TLB entry from the CPU so we won't allow
1152          * huge and small TLB entries for the same virtual address to
1153          * avoid the risk of CPU bugs in that area.
1154          *
1155          * Parallel fast GUP is fine since fast GUP will back off when
1156          * it detects PMD is changed.
1157          */
1158         _pmd = pmdp_collapse_flush(vma, address, pmd);
1159         spin_unlock(pmd_ptl);
1160         mmu_notifier_invalidate_range_end(&range);
1161         tlb_remove_table_sync_one();
1162
1163         spin_lock(pte_ptl);
1164         isolated = __collapse_huge_page_isolate(vma, address, pte,
1165                         &compound_pagelist);
1166         spin_unlock(pte_ptl);
1167
1168         if (unlikely(!isolated)) {
1169                 pte_unmap(pte);
1170                 spin_lock(pmd_ptl);
1171                 BUG_ON(!pmd_none(*pmd));
1172                 /*
1173                  * We can only use set_pmd_at when establishing
1174                  * hugepmds and never for establishing regular pmds that
1175                  * points to regular pagetables. Use pmd_populate for that
1176                  */
1177                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1178                 spin_unlock(pmd_ptl);
1179                 anon_vma_unlock_write(vma->anon_vma);
1180                 result = SCAN_FAIL;
1181                 goto out;
1182         }
1183
1184         /*
1185          * All pages are isolated and locked so anon_vma rmap
1186          * can't run anymore.
1187          */
1188         anon_vma_unlock_write(vma->anon_vma);
1189
1190         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1191                         &compound_pagelist);
1192         pte_unmap(pte);
1193         __SetPageUptodate(new_page);
1194         pgtable = pmd_pgtable(_pmd);
1195
1196         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1197         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1198
1199         /*
1200          * spin_lock() below is not the equivalent of smp_wmb(), so
1201          * this is needed to avoid the copy_huge_page writes to become
1202          * visible after the set_pmd_at() write.
1203          */
1204         smp_wmb();
1205
1206         spin_lock(pmd_ptl);
1207         BUG_ON(!pmd_none(*pmd));
1208         page_add_new_anon_rmap(new_page, vma, address, true);
1209         lru_cache_add_inactive_or_unevictable(new_page, vma);
1210         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1211         set_pmd_at(mm, address, pmd, _pmd);
1212         update_mmu_cache_pmd(vma, address, pmd);
1213         spin_unlock(pmd_ptl);
1214
1215         *hpage = NULL;
1216
1217         khugepaged_pages_collapsed++;
1218         result = SCAN_SUCCEED;
1219 out_up_write:
1220         mmap_write_unlock(mm);
1221 out_nolock:
1222         if (!IS_ERR_OR_NULL(*hpage))
1223                 mem_cgroup_uncharge(*hpage);
1224         trace_mm_collapse_huge_page(mm, isolated, result);
1225         return;
1226 out:
1227         goto out_up_write;
1228 }
1229
1230 static int khugepaged_scan_pmd(struct mm_struct *mm,
1231                                struct vm_area_struct *vma,
1232                                unsigned long address,
1233                                struct page **hpage)
1234 {
1235         pmd_t *pmd;
1236         pte_t *pte, *_pte;
1237         int ret = 0, result = 0, referenced = 0;
1238         int none_or_zero = 0, shared = 0;
1239         struct page *page = NULL;
1240         unsigned long _address;
1241         spinlock_t *ptl;
1242         int node = NUMA_NO_NODE, unmapped = 0;
1243         bool writable = false;
1244
1245         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1246
1247         pmd = mm_find_pmd(mm, address);
1248         if (!pmd) {
1249                 result = SCAN_PMD_NULL;
1250                 goto out;
1251         }
1252
1253         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1254         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1255         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1256              _pte++, _address += PAGE_SIZE) {
1257                 pte_t pteval = *_pte;
1258                 if (is_swap_pte(pteval)) {
1259                         if (++unmapped <= khugepaged_max_ptes_swap) {
1260                                 /*
1261                                  * Always be strict with uffd-wp
1262                                  * enabled swap entries.  Please see
1263                                  * comment below for pte_uffd_wp().
1264                                  */
1265                                 if (pte_swp_uffd_wp(pteval)) {
1266                                         result = SCAN_PTE_UFFD_WP;
1267                                         goto out_unmap;
1268                                 }
1269                                 continue;
1270                         } else {
1271                                 result = SCAN_EXCEED_SWAP_PTE;
1272                                 goto out_unmap;
1273                         }
1274                 }
1275                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1276                         if (!userfaultfd_armed(vma) &&
1277                             ++none_or_zero <= khugepaged_max_ptes_none) {
1278                                 continue;
1279                         } else {
1280                                 result = SCAN_EXCEED_NONE_PTE;
1281                                 goto out_unmap;
1282                         }
1283                 }
1284                 if (!pte_present(pteval)) {
1285                         result = SCAN_PTE_NON_PRESENT;
1286                         goto out_unmap;
1287                 }
1288                 if (pte_uffd_wp(pteval)) {
1289                         /*
1290                          * Don't collapse the page if any of the small
1291                          * PTEs are armed with uffd write protection.
1292                          * Here we can also mark the new huge pmd as
1293                          * write protected if any of the small ones is
1294                          * marked but that could bring uknown
1295                          * userfault messages that falls outside of
1296                          * the registered range.  So, just be simple.
1297                          */
1298                         result = SCAN_PTE_UFFD_WP;
1299                         goto out_unmap;
1300                 }
1301                 if (pte_write(pteval))
1302                         writable = true;
1303
1304                 page = vm_normal_page(vma, _address, pteval);
1305                 if (unlikely(!page)) {
1306                         result = SCAN_PAGE_NULL;
1307                         goto out_unmap;
1308                 }
1309
1310                 if (page_mapcount(page) > 1 &&
1311                                 ++shared > khugepaged_max_ptes_shared) {
1312                         result = SCAN_EXCEED_SHARED_PTE;
1313                         goto out_unmap;
1314                 }
1315
1316                 page = compound_head(page);
1317
1318                 /*
1319                  * Record which node the original page is from and save this
1320                  * information to khugepaged_node_load[].
1321                  * Khupaged will allocate hugepage from the node has the max
1322                  * hit record.
1323                  */
1324                 node = page_to_nid(page);
1325                 if (khugepaged_scan_abort(node)) {
1326                         result = SCAN_SCAN_ABORT;
1327                         goto out_unmap;
1328                 }
1329                 khugepaged_node_load[node]++;
1330                 if (!PageLRU(page)) {
1331                         result = SCAN_PAGE_LRU;
1332                         goto out_unmap;
1333                 }
1334                 if (PageLocked(page)) {
1335                         result = SCAN_PAGE_LOCK;
1336                         goto out_unmap;
1337                 }
1338                 if (!PageAnon(page)) {
1339                         result = SCAN_PAGE_ANON;
1340                         goto out_unmap;
1341                 }
1342
1343                 /*
1344                  * Check if the page has any GUP (or other external) pins.
1345                  *
1346                  * Here the check is racy it may see totmal_mapcount > refcount
1347                  * in some cases.
1348                  * For example, one process with one forked child process.
1349                  * The parent has the PMD split due to MADV_DONTNEED, then
1350                  * the child is trying unmap the whole PMD, but khugepaged
1351                  * may be scanning the parent between the child has
1352                  * PageDoubleMap flag cleared and dec the mapcount.  So
1353                  * khugepaged may see total_mapcount > refcount.
1354                  *
1355                  * But such case is ephemeral we could always retry collapse
1356                  * later.  However it may report false positive if the page
1357                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1358                  * will be done again later the risk seems low.
1359                  */
1360                 if (!is_refcount_suitable(page)) {
1361                         result = SCAN_PAGE_COUNT;
1362                         goto out_unmap;
1363                 }
1364                 if (pte_young(pteval) ||
1365                     page_is_young(page) || PageReferenced(page) ||
1366                     mmu_notifier_test_young(vma->vm_mm, address))
1367                         referenced++;
1368         }
1369         if (!writable) {
1370                 result = SCAN_PAGE_RO;
1371         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1372                 result = SCAN_LACK_REFERENCED_PAGE;
1373         } else {
1374                 result = SCAN_SUCCEED;
1375                 ret = 1;
1376         }
1377 out_unmap:
1378         pte_unmap_unlock(pte, ptl);
1379         if (ret) {
1380                 node = khugepaged_find_target_node();
1381                 /* collapse_huge_page will return with the mmap_lock released */
1382                 collapse_huge_page(mm, address, hpage, node,
1383                                 referenced, unmapped);
1384         }
1385 out:
1386         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1387                                      none_or_zero, result, unmapped);
1388         return ret;
1389 }
1390
1391 static void collect_mm_slot(struct mm_slot *mm_slot)
1392 {
1393         struct mm_struct *mm = mm_slot->mm;
1394
1395         lockdep_assert_held(&khugepaged_mm_lock);
1396
1397         if (khugepaged_test_exit(mm)) {
1398                 /* free mm_slot */
1399                 hash_del(&mm_slot->hash);
1400                 list_del(&mm_slot->mm_node);
1401
1402                 /*
1403                  * Not strictly needed because the mm exited already.
1404                  *
1405                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1406                  */
1407
1408                 /* khugepaged_mm_lock actually not necessary for the below */
1409                 free_mm_slot(mm_slot);
1410                 mmdrop(mm);
1411         }
1412 }
1413
1414 #ifdef CONFIG_SHMEM
1415 /*
1416  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1417  * khugepaged should try to collapse the page table.
1418  */
1419 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1420                                          unsigned long addr)
1421 {
1422         struct mm_slot *mm_slot;
1423
1424         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1425
1426         spin_lock(&khugepaged_mm_lock);
1427         mm_slot = get_mm_slot(mm);
1428         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1429                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1430         spin_unlock(&khugepaged_mm_lock);
1431         return 0;
1432 }
1433
1434 /**
1435  * Try to collapse a pte-mapped THP for mm at address haddr.
1436  *
1437  * This function checks whether all the PTEs in the PMD are pointing to the
1438  * right THP. If so, retract the page table so the THP can refault in with
1439  * as pmd-mapped.
1440  */
1441 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1442 {
1443         unsigned long haddr = addr & HPAGE_PMD_MASK;
1444         struct vm_area_struct *vma = find_vma(mm, haddr);
1445         struct page *hpage;
1446         pte_t *start_pte, *pte;
1447         pmd_t *pmd, _pmd;
1448         spinlock_t *ptl;
1449         int count = 0;
1450         int i;
1451         struct mmu_notifier_range range;
1452
1453         if (!vma || !vma->vm_file ||
1454             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1455                 return;
1456
1457         /*
1458          * This vm_flags may not have VM_HUGEPAGE if the page was not
1459          * collapsed by this mm. But we can still collapse if the page is
1460          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1461          * will not fail the vma for missing VM_HUGEPAGE
1462          */
1463         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1464                 return;
1465
1466         hpage = find_lock_page(vma->vm_file->f_mapping,
1467                                linear_page_index(vma, haddr));
1468         if (!hpage)
1469                 return;
1470
1471         if (!PageHead(hpage))
1472                 goto drop_hpage;
1473
1474         pmd = mm_find_pmd(mm, haddr);
1475         if (!pmd)
1476                 goto drop_hpage;
1477
1478         /*
1479          * We need to lock the mapping so that from here on, only GUP-fast and
1480          * hardware page walks can access the parts of the page tables that
1481          * we're operating on.
1482          */
1483         i_mmap_lock_write(vma->vm_file->f_mapping);
1484
1485         /*
1486          * This spinlock should be unnecessary: Nobody else should be accessing
1487          * the page tables under spinlock protection here, only
1488          * lockless_pages_from_mm() and the hardware page walker can access page
1489          * tables while all the high-level locks are held in write mode.
1490          */
1491         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1492
1493         /* step 1: check all mapped PTEs are to the right huge page */
1494         for (i = 0, addr = haddr, pte = start_pte;
1495              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1496                 struct page *page;
1497
1498                 /* empty pte, skip */
1499                 if (pte_none(*pte))
1500                         continue;
1501
1502                 /* page swapped out, abort */
1503                 if (!pte_present(*pte))
1504                         goto abort;
1505
1506                 page = vm_normal_page(vma, addr, *pte);
1507
1508                 /*
1509                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1510                  * page table, but the new page will not be a subpage of hpage.
1511                  */
1512                 if (hpage + i != page)
1513                         goto abort;
1514                 count++;
1515         }
1516
1517         /* step 2: adjust rmap */
1518         for (i = 0, addr = haddr, pte = start_pte;
1519              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1520                 struct page *page;
1521
1522                 if (pte_none(*pte))
1523                         continue;
1524                 page = vm_normal_page(vma, addr, *pte);
1525                 page_remove_rmap(page, false);
1526         }
1527
1528         pte_unmap_unlock(start_pte, ptl);
1529
1530         /* step 3: set proper refcount and mm_counters. */
1531         if (count) {
1532                 page_ref_sub(hpage, count);
1533                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1534         }
1535
1536         /* step 4: collapse pmd */
1537         /* we make no change to anon, but protect concurrent anon page lookup */
1538         if (vma->anon_vma)
1539                 anon_vma_lock_write(vma->anon_vma);
1540
1541         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1542                                 haddr + HPAGE_PMD_SIZE);
1543         mmu_notifier_invalidate_range_start(&range);
1544         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1545         mm_dec_nr_ptes(mm);
1546         tlb_remove_table_sync_one();
1547         mmu_notifier_invalidate_range_end(&range);
1548         pte_free(mm, pmd_pgtable(_pmd));
1549
1550         if (vma->anon_vma)
1551                 anon_vma_unlock_write(vma->anon_vma);
1552         i_mmap_unlock_write(vma->vm_file->f_mapping);
1553
1554 drop_hpage:
1555         unlock_page(hpage);
1556         put_page(hpage);
1557         return;
1558
1559 abort:
1560         pte_unmap_unlock(start_pte, ptl);
1561         i_mmap_unlock_write(vma->vm_file->f_mapping);
1562         goto drop_hpage;
1563 }
1564
1565 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1566 {
1567         struct mm_struct *mm = mm_slot->mm;
1568         int i;
1569
1570         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1571                 return 0;
1572
1573         if (!mmap_write_trylock(mm))
1574                 return -EBUSY;
1575
1576         if (unlikely(khugepaged_test_exit(mm)))
1577                 goto out;
1578
1579         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1580                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1581
1582 out:
1583         mm_slot->nr_pte_mapped_thp = 0;
1584         mmap_write_unlock(mm);
1585         return 0;
1586 }
1587
1588 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1589 {
1590         struct vm_area_struct *vma;
1591         struct mm_struct *mm;
1592         unsigned long addr;
1593         pmd_t *pmd, _pmd;
1594
1595         i_mmap_lock_write(mapping);
1596         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1597                 /*
1598                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1599                  * got written to. These VMAs are likely not worth investing
1600                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1601                  * later.
1602                  *
1603                  * Not that vma->anon_vma check is racy: it can be set up after
1604                  * the check but before we took mmap_lock by the fault path.
1605                  * But page lock would prevent establishing any new ptes of the
1606                  * page, so we are safe.
1607                  *
1608                  * An alternative would be drop the check, but check that page
1609                  * table is clear before calling pmdp_collapse_flush() under
1610                  * ptl. It has higher chance to recover THP for the VMA, but
1611                  * has higher cost too. It would also probably require locking
1612                  * the anon_vma.
1613                  */
1614                 if (vma->anon_vma)
1615                         continue;
1616                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1617                 if (addr & ~HPAGE_PMD_MASK)
1618                         continue;
1619                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1620                         continue;
1621                 mm = vma->vm_mm;
1622                 pmd = mm_find_pmd(mm, addr);
1623                 if (!pmd)
1624                         continue;
1625                 /*
1626                  * We need exclusive mmap_lock to retract page table.
1627                  *
1628                  * We use trylock due to lock inversion: we need to acquire
1629                  * mmap_lock while holding page lock. Fault path does it in
1630                  * reverse order. Trylock is a way to avoid deadlock.
1631                  */
1632                 if (mmap_write_trylock(mm)) {
1633                         if (!khugepaged_test_exit(mm)) {
1634                                 struct mmu_notifier_range range;
1635
1636                                 mmu_notifier_range_init(&range,
1637                                                         MMU_NOTIFY_CLEAR, 0,
1638                                                         NULL, mm, addr,
1639                                                         addr + HPAGE_PMD_SIZE);
1640                                 mmu_notifier_invalidate_range_start(&range);
1641                                 /* assume page table is clear */
1642                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1643                                 mm_dec_nr_ptes(mm);
1644                                 tlb_remove_table_sync_one();
1645                                 pte_free(mm, pmd_pgtable(_pmd));
1646                                 mmu_notifier_invalidate_range_end(&range);
1647                         }
1648                         mmap_write_unlock(mm);
1649                 } else {
1650                         /* Try again later */
1651                         khugepaged_add_pte_mapped_thp(mm, addr);
1652                 }
1653         }
1654         i_mmap_unlock_write(mapping);
1655 }
1656
1657 /**
1658  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1659  *
1660  * Basic scheme is simple, details are more complex:
1661  *  - allocate and lock a new huge page;
1662  *  - scan page cache replacing old pages with the new one
1663  *    + swap/gup in pages if necessary;
1664  *    + fill in gaps;
1665  *    + keep old pages around in case rollback is required;
1666  *  - if replacing succeeds:
1667  *    + copy data over;
1668  *    + free old pages;
1669  *    + unlock huge page;
1670  *  - if replacing failed;
1671  *    + put all pages back and unfreeze them;
1672  *    + restore gaps in the page cache;
1673  *    + unlock and free huge page;
1674  */
1675 static void collapse_file(struct mm_struct *mm,
1676                 struct file *file, pgoff_t start,
1677                 struct page **hpage, int node)
1678 {
1679         struct address_space *mapping = file->f_mapping;
1680         gfp_t gfp;
1681         struct page *new_page;
1682         pgoff_t index, end = start + HPAGE_PMD_NR;
1683         LIST_HEAD(pagelist);
1684         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1685         int nr_none = 0, result = SCAN_SUCCEED;
1686         bool is_shmem = shmem_file(file);
1687
1688         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1689         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1690
1691         /* Only allocate from the target node */
1692         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1693
1694         new_page = khugepaged_alloc_page(hpage, gfp, node);
1695         if (!new_page) {
1696                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1697                 goto out;
1698         }
1699
1700         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1701                 result = SCAN_CGROUP_CHARGE_FAIL;
1702                 goto out;
1703         }
1704         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1705
1706         /* This will be less messy when we use multi-index entries */
1707         do {
1708                 xas_lock_irq(&xas);
1709                 xas_create_range(&xas);
1710                 if (!xas_error(&xas))
1711                         break;
1712                 xas_unlock_irq(&xas);
1713                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1714                         result = SCAN_FAIL;
1715                         goto out;
1716                 }
1717         } while (1);
1718
1719         __SetPageLocked(new_page);
1720         if (is_shmem)
1721                 __SetPageSwapBacked(new_page);
1722         new_page->index = start;
1723         new_page->mapping = mapping;
1724
1725         /*
1726          * At this point the new_page is locked and not up-to-date.
1727          * It's safe to insert it into the page cache, because nobody would
1728          * be able to map it or use it in another way until we unlock it.
1729          */
1730
1731         xas_set(&xas, start);
1732         for (index = start; index < end; index++) {
1733                 struct page *page = xas_next(&xas);
1734
1735                 VM_BUG_ON(index != xas.xa_index);
1736                 if (is_shmem) {
1737                         if (!page) {
1738                                 /*
1739                                  * Stop if extent has been truncated or
1740                                  * hole-punched, and is now completely
1741                                  * empty.
1742                                  */
1743                                 if (index == start) {
1744                                         if (!xas_next_entry(&xas, end - 1)) {
1745                                                 result = SCAN_TRUNCATED;
1746                                                 goto xa_locked;
1747                                         }
1748                                         xas_set(&xas, index);
1749                                 }
1750                                 if (!shmem_charge(mapping->host, 1)) {
1751                                         result = SCAN_FAIL;
1752                                         goto xa_locked;
1753                                 }
1754                                 xas_store(&xas, new_page);
1755                                 nr_none++;
1756                                 continue;
1757                         }
1758
1759                         if (xa_is_value(page) || !PageUptodate(page)) {
1760                                 xas_unlock_irq(&xas);
1761                                 /* swap in or instantiate fallocated page */
1762                                 if (shmem_getpage(mapping->host, index, &page,
1763                                                   SGP_NOHUGE)) {
1764                                         result = SCAN_FAIL;
1765                                         goto xa_unlocked;
1766                                 }
1767                         } else if (trylock_page(page)) {
1768                                 get_page(page);
1769                                 xas_unlock_irq(&xas);
1770                         } else {
1771                                 result = SCAN_PAGE_LOCK;
1772                                 goto xa_locked;
1773                         }
1774                 } else {        /* !is_shmem */
1775                         if (!page || xa_is_value(page)) {
1776                                 xas_unlock_irq(&xas);
1777                                 page_cache_sync_readahead(mapping, &file->f_ra,
1778                                                           file, index,
1779                                                           end - index);
1780                                 /* drain pagevecs to help isolate_lru_page() */
1781                                 lru_add_drain();
1782                                 page = find_lock_page(mapping, index);
1783                                 if (unlikely(page == NULL)) {
1784                                         result = SCAN_FAIL;
1785                                         goto xa_unlocked;
1786                                 }
1787                         } else if (PageDirty(page)) {
1788                                 /*
1789                                  * khugepaged only works on read-only fd,
1790                                  * so this page is dirty because it hasn't
1791                                  * been flushed since first write. There
1792                                  * won't be new dirty pages.
1793                                  *
1794                                  * Trigger async flush here and hope the
1795                                  * writeback is done when khugepaged
1796                                  * revisits this page.
1797                                  *
1798                                  * This is a one-off situation. We are not
1799                                  * forcing writeback in loop.
1800                                  */
1801                                 xas_unlock_irq(&xas);
1802                                 filemap_flush(mapping);
1803                                 result = SCAN_FAIL;
1804                                 goto xa_unlocked;
1805                         } else if (PageWriteback(page)) {
1806                                 xas_unlock_irq(&xas);
1807                                 result = SCAN_FAIL;
1808                                 goto xa_unlocked;
1809                         } else if (trylock_page(page)) {
1810                                 get_page(page);
1811                                 xas_unlock_irq(&xas);
1812                         } else {
1813                                 result = SCAN_PAGE_LOCK;
1814                                 goto xa_locked;
1815                         }
1816                 }
1817
1818                 /*
1819                  * The page must be locked, so we can drop the i_pages lock
1820                  * without racing with truncate.
1821                  */
1822                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1823
1824                 /* make sure the page is up to date */
1825                 if (unlikely(!PageUptodate(page))) {
1826                         result = SCAN_FAIL;
1827                         goto out_unlock;
1828                 }
1829
1830                 /*
1831                  * If file was truncated then extended, or hole-punched, before
1832                  * we locked the first page, then a THP might be there already.
1833                  */
1834                 if (PageTransCompound(page)) {
1835                         result = SCAN_PAGE_COMPOUND;
1836                         goto out_unlock;
1837                 }
1838
1839                 if (page_mapping(page) != mapping) {
1840                         result = SCAN_TRUNCATED;
1841                         goto out_unlock;
1842                 }
1843
1844                 if (!is_shmem && (PageDirty(page) ||
1845                                   PageWriteback(page))) {
1846                         /*
1847                          * khugepaged only works on read-only fd, so this
1848                          * page is dirty because it hasn't been flushed
1849                          * since first write.
1850                          */
1851                         result = SCAN_FAIL;
1852                         goto out_unlock;
1853                 }
1854
1855                 if (isolate_lru_page(page)) {
1856                         result = SCAN_DEL_PAGE_LRU;
1857                         goto out_unlock;
1858                 }
1859
1860                 if (page_has_private(page) &&
1861                     !try_to_release_page(page, GFP_KERNEL)) {
1862                         result = SCAN_PAGE_HAS_PRIVATE;
1863                         putback_lru_page(page);
1864                         goto out_unlock;
1865                 }
1866
1867                 if (page_mapped(page))
1868                         unmap_mapping_pages(mapping, index, 1, false);
1869
1870                 xas_lock_irq(&xas);
1871                 xas_set(&xas, index);
1872
1873                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1874                 VM_BUG_ON_PAGE(page_mapped(page), page);
1875
1876                 /*
1877                  * The page is expected to have page_count() == 3:
1878                  *  - we hold a pin on it;
1879                  *  - one reference from page cache;
1880                  *  - one from isolate_lru_page;
1881                  */
1882                 if (!page_ref_freeze(page, 3)) {
1883                         result = SCAN_PAGE_COUNT;
1884                         xas_unlock_irq(&xas);
1885                         putback_lru_page(page);
1886                         goto out_unlock;
1887                 }
1888
1889                 /*
1890                  * Add the page to the list to be able to undo the collapse if
1891                  * something go wrong.
1892                  */
1893                 list_add_tail(&page->lru, &pagelist);
1894
1895                 /* Finally, replace with the new page. */
1896                 xas_store(&xas, new_page);
1897                 continue;
1898 out_unlock:
1899                 unlock_page(page);
1900                 put_page(page);
1901                 goto xa_unlocked;
1902         }
1903
1904         if (is_shmem)
1905                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1906         else {
1907                 __inc_node_page_state(new_page, NR_FILE_THPS);
1908                 filemap_nr_thps_inc(mapping);
1909         }
1910
1911         if (nr_none) {
1912                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1913                 if (is_shmem)
1914                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1915         }
1916
1917 xa_locked:
1918         xas_unlock_irq(&xas);
1919 xa_unlocked:
1920
1921         if (result == SCAN_SUCCEED) {
1922                 struct page *page, *tmp;
1923
1924                 /*
1925                  * Replacing old pages with new one has succeeded, now we
1926                  * need to copy the content and free the old pages.
1927                  */
1928                 index = start;
1929                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1930                         while (index < page->index) {
1931                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1932                                 index++;
1933                         }
1934                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1935                                         page);
1936                         list_del(&page->lru);
1937                         page->mapping = NULL;
1938                         page_ref_unfreeze(page, 1);
1939                         ClearPageActive(page);
1940                         ClearPageUnevictable(page);
1941                         unlock_page(page);
1942                         put_page(page);
1943                         index++;
1944                 }
1945                 while (index < end) {
1946                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1947                         index++;
1948                 }
1949
1950                 SetPageUptodate(new_page);
1951                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1952                 if (is_shmem)
1953                         set_page_dirty(new_page);
1954                 lru_cache_add(new_page);
1955
1956                 /*
1957                  * Remove pte page tables, so we can re-fault the page as huge.
1958                  */
1959                 retract_page_tables(mapping, start);
1960                 *hpage = NULL;
1961
1962                 khugepaged_pages_collapsed++;
1963         } else {
1964                 struct page *page;
1965
1966                 /* Something went wrong: roll back page cache changes */
1967                 xas_lock_irq(&xas);
1968                 mapping->nrpages -= nr_none;
1969
1970                 if (is_shmem)
1971                         shmem_uncharge(mapping->host, nr_none);
1972
1973                 xas_set(&xas, start);
1974                 xas_for_each(&xas, page, end - 1) {
1975                         page = list_first_entry_or_null(&pagelist,
1976                                         struct page, lru);
1977                         if (!page || xas.xa_index < page->index) {
1978                                 if (!nr_none)
1979                                         break;
1980                                 nr_none--;
1981                                 /* Put holes back where they were */
1982                                 xas_store(&xas, NULL);
1983                                 continue;
1984                         }
1985
1986                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1987
1988                         /* Unfreeze the page. */
1989                         list_del(&page->lru);
1990                         page_ref_unfreeze(page, 2);
1991                         xas_store(&xas, page);
1992                         xas_pause(&xas);
1993                         xas_unlock_irq(&xas);
1994                         unlock_page(page);
1995                         putback_lru_page(page);
1996                         xas_lock_irq(&xas);
1997                 }
1998                 VM_BUG_ON(nr_none);
1999                 xas_unlock_irq(&xas);
2000
2001                 new_page->mapping = NULL;
2002         }
2003
2004         unlock_page(new_page);
2005 out:
2006         VM_BUG_ON(!list_empty(&pagelist));
2007         if (!IS_ERR_OR_NULL(*hpage))
2008                 mem_cgroup_uncharge(*hpage);
2009         /* TODO: tracepoints */
2010 }
2011
2012 static void khugepaged_scan_file(struct mm_struct *mm,
2013                 struct file *file, pgoff_t start, struct page **hpage)
2014 {
2015         struct page *page = NULL;
2016         struct address_space *mapping = file->f_mapping;
2017         XA_STATE(xas, &mapping->i_pages, start);
2018         int present, swap;
2019         int node = NUMA_NO_NODE;
2020         int result = SCAN_SUCCEED;
2021
2022         present = 0;
2023         swap = 0;
2024         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2025         rcu_read_lock();
2026         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2027                 if (xas_retry(&xas, page))
2028                         continue;
2029
2030                 if (xa_is_value(page)) {
2031                         if (++swap > khugepaged_max_ptes_swap) {
2032                                 result = SCAN_EXCEED_SWAP_PTE;
2033                                 break;
2034                         }
2035                         continue;
2036                 }
2037
2038                 if (PageTransCompound(page)) {
2039                         result = SCAN_PAGE_COMPOUND;
2040                         break;
2041                 }
2042
2043                 node = page_to_nid(page);
2044                 if (khugepaged_scan_abort(node)) {
2045                         result = SCAN_SCAN_ABORT;
2046                         break;
2047                 }
2048                 khugepaged_node_load[node]++;
2049
2050                 if (!PageLRU(page)) {
2051                         result = SCAN_PAGE_LRU;
2052                         break;
2053                 }
2054
2055                 if (page_count(page) !=
2056                     1 + page_mapcount(page) + page_has_private(page)) {
2057                         result = SCAN_PAGE_COUNT;
2058                         break;
2059                 }
2060
2061                 /*
2062                  * We probably should check if the page is referenced here, but
2063                  * nobody would transfer pte_young() to PageReferenced() for us.
2064                  * And rmap walk here is just too costly...
2065                  */
2066
2067                 present++;
2068
2069                 if (need_resched()) {
2070                         xas_pause(&xas);
2071                         cond_resched_rcu();
2072                 }
2073         }
2074         rcu_read_unlock();
2075
2076         if (result == SCAN_SUCCEED) {
2077                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2078                         result = SCAN_EXCEED_NONE_PTE;
2079                 } else {
2080                         node = khugepaged_find_target_node();
2081                         collapse_file(mm, file, start, hpage, node);
2082                 }
2083         }
2084
2085         /* TODO: tracepoints */
2086 }
2087 #else
2088 static void khugepaged_scan_file(struct mm_struct *mm,
2089                 struct file *file, pgoff_t start, struct page **hpage)
2090 {
2091         BUILD_BUG();
2092 }
2093
2094 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2095 {
2096         return 0;
2097 }
2098 #endif
2099
2100 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2101                                             struct page **hpage)
2102         __releases(&khugepaged_mm_lock)
2103         __acquires(&khugepaged_mm_lock)
2104 {
2105         struct mm_slot *mm_slot;
2106         struct mm_struct *mm;
2107         struct vm_area_struct *vma;
2108         int progress = 0;
2109
2110         VM_BUG_ON(!pages);
2111         lockdep_assert_held(&khugepaged_mm_lock);
2112
2113         if (khugepaged_scan.mm_slot)
2114                 mm_slot = khugepaged_scan.mm_slot;
2115         else {
2116                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2117                                      struct mm_slot, mm_node);
2118                 khugepaged_scan.address = 0;
2119                 khugepaged_scan.mm_slot = mm_slot;
2120         }
2121         spin_unlock(&khugepaged_mm_lock);
2122         khugepaged_collapse_pte_mapped_thps(mm_slot);
2123
2124         mm = mm_slot->mm;
2125         /*
2126          * Don't wait for semaphore (to avoid long wait times).  Just move to
2127          * the next mm on the list.
2128          */
2129         vma = NULL;
2130         if (unlikely(!mmap_read_trylock(mm)))
2131                 goto breakouterloop_mmap_lock;
2132         if (likely(!khugepaged_test_exit(mm)))
2133                 vma = find_vma(mm, khugepaged_scan.address);
2134
2135         progress++;
2136         for (; vma; vma = vma->vm_next) {
2137                 unsigned long hstart, hend;
2138
2139                 cond_resched();
2140                 if (unlikely(khugepaged_test_exit(mm))) {
2141                         progress++;
2142                         break;
2143                 }
2144                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2145 skip:
2146                         progress++;
2147                         continue;
2148                 }
2149                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2150                 hend = vma->vm_end & HPAGE_PMD_MASK;
2151                 if (hstart >= hend)
2152                         goto skip;
2153                 if (khugepaged_scan.address > hend)
2154                         goto skip;
2155                 if (khugepaged_scan.address < hstart)
2156                         khugepaged_scan.address = hstart;
2157                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2158                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2159                         goto skip;
2160
2161                 while (khugepaged_scan.address < hend) {
2162                         int ret;
2163                         cond_resched();
2164                         if (unlikely(khugepaged_test_exit(mm)))
2165                                 goto breakouterloop;
2166
2167                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2168                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2169                                   hend);
2170                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2171                                 struct file *file = get_file(vma->vm_file);
2172                                 pgoff_t pgoff = linear_page_index(vma,
2173                                                 khugepaged_scan.address);
2174
2175                                 mmap_read_unlock(mm);
2176                                 ret = 1;
2177                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2178                                 fput(file);
2179                         } else {
2180                                 ret = khugepaged_scan_pmd(mm, vma,
2181                                                 khugepaged_scan.address,
2182                                                 hpage);
2183                         }
2184                         /* move to next address */
2185                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2186                         progress += HPAGE_PMD_NR;
2187                         if (ret)
2188                                 /* we released mmap_lock so break loop */
2189                                 goto breakouterloop_mmap_lock;
2190                         if (progress >= pages)
2191                                 goto breakouterloop;
2192                 }
2193         }
2194 breakouterloop:
2195         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2196 breakouterloop_mmap_lock:
2197
2198         spin_lock(&khugepaged_mm_lock);
2199         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2200         /*
2201          * Release the current mm_slot if this mm is about to die, or
2202          * if we scanned all vmas of this mm.
2203          */
2204         if (khugepaged_test_exit(mm) || !vma) {
2205                 /*
2206                  * Make sure that if mm_users is reaching zero while
2207                  * khugepaged runs here, khugepaged_exit will find
2208                  * mm_slot not pointing to the exiting mm.
2209                  */
2210                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2211                         khugepaged_scan.mm_slot = list_entry(
2212                                 mm_slot->mm_node.next,
2213                                 struct mm_slot, mm_node);
2214                         khugepaged_scan.address = 0;
2215                 } else {
2216                         khugepaged_scan.mm_slot = NULL;
2217                         khugepaged_full_scans++;
2218                 }
2219
2220                 collect_mm_slot(mm_slot);
2221         }
2222
2223         return progress;
2224 }
2225
2226 static int khugepaged_has_work(void)
2227 {
2228         return !list_empty(&khugepaged_scan.mm_head) &&
2229                 khugepaged_enabled();
2230 }
2231
2232 static int khugepaged_wait_event(void)
2233 {
2234         return !list_empty(&khugepaged_scan.mm_head) ||
2235                 kthread_should_stop();
2236 }
2237
2238 static void khugepaged_do_scan(void)
2239 {
2240         struct page *hpage = NULL;
2241         unsigned int progress = 0, pass_through_head = 0;
2242         unsigned int pages = khugepaged_pages_to_scan;
2243         bool wait = true;
2244
2245         barrier(); /* write khugepaged_pages_to_scan to local stack */
2246
2247         lru_add_drain_all();
2248
2249         while (progress < pages) {
2250                 if (!khugepaged_prealloc_page(&hpage, &wait))
2251                         break;
2252
2253                 cond_resched();
2254
2255                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2256                         break;
2257
2258                 spin_lock(&khugepaged_mm_lock);
2259                 if (!khugepaged_scan.mm_slot)
2260                         pass_through_head++;
2261                 if (khugepaged_has_work() &&
2262                     pass_through_head < 2)
2263                         progress += khugepaged_scan_mm_slot(pages - progress,
2264                                                             &hpage);
2265                 else
2266                         progress = pages;
2267                 spin_unlock(&khugepaged_mm_lock);
2268         }
2269
2270         if (!IS_ERR_OR_NULL(hpage))
2271                 put_page(hpage);
2272 }
2273
2274 static bool khugepaged_should_wakeup(void)
2275 {
2276         return kthread_should_stop() ||
2277                time_after_eq(jiffies, khugepaged_sleep_expire);
2278 }
2279
2280 static void khugepaged_wait_work(void)
2281 {
2282         if (khugepaged_has_work()) {
2283                 const unsigned long scan_sleep_jiffies =
2284                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2285
2286                 if (!scan_sleep_jiffies)
2287                         return;
2288
2289                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2290                 wait_event_freezable_timeout(khugepaged_wait,
2291                                              khugepaged_should_wakeup(),
2292                                              scan_sleep_jiffies);
2293                 return;
2294         }
2295
2296         if (khugepaged_enabled())
2297                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2298 }
2299
2300 static int khugepaged(void *none)
2301 {
2302         struct mm_slot *mm_slot;
2303
2304         set_freezable();
2305         set_user_nice(current, MAX_NICE);
2306
2307         while (!kthread_should_stop()) {
2308                 khugepaged_do_scan();
2309                 khugepaged_wait_work();
2310         }
2311
2312         spin_lock(&khugepaged_mm_lock);
2313         mm_slot = khugepaged_scan.mm_slot;
2314         khugepaged_scan.mm_slot = NULL;
2315         if (mm_slot)
2316                 collect_mm_slot(mm_slot);
2317         spin_unlock(&khugepaged_mm_lock);
2318         return 0;
2319 }
2320
2321 static void set_recommended_min_free_kbytes(void)
2322 {
2323         struct zone *zone;
2324         int nr_zones = 0;
2325         unsigned long recommended_min;
2326
2327         for_each_populated_zone(zone) {
2328                 /*
2329                  * We don't need to worry about fragmentation of
2330                  * ZONE_MOVABLE since it only has movable pages.
2331                  */
2332                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2333                         continue;
2334
2335                 nr_zones++;
2336         }
2337
2338         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2339         recommended_min = pageblock_nr_pages * nr_zones * 2;
2340
2341         /*
2342          * Make sure that on average at least two pageblocks are almost free
2343          * of another type, one for a migratetype to fall back to and a
2344          * second to avoid subsequent fallbacks of other types There are 3
2345          * MIGRATE_TYPES we care about.
2346          */
2347         recommended_min += pageblock_nr_pages * nr_zones *
2348                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2349
2350         /* don't ever allow to reserve more than 5% of the lowmem */
2351         recommended_min = min(recommended_min,
2352                               (unsigned long) nr_free_buffer_pages() / 20);
2353         recommended_min <<= (PAGE_SHIFT-10);
2354
2355         if (recommended_min > min_free_kbytes) {
2356                 if (user_min_free_kbytes >= 0)
2357                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2358                                 min_free_kbytes, recommended_min);
2359
2360                 min_free_kbytes = recommended_min;
2361         }
2362         setup_per_zone_wmarks();
2363 }
2364
2365 int start_stop_khugepaged(void)
2366 {
2367         int err = 0;
2368
2369         mutex_lock(&khugepaged_mutex);
2370         if (khugepaged_enabled()) {
2371                 if (!khugepaged_thread)
2372                         khugepaged_thread = kthread_run(khugepaged, NULL,
2373                                                         "khugepaged");
2374                 if (IS_ERR(khugepaged_thread)) {
2375                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2376                         err = PTR_ERR(khugepaged_thread);
2377                         khugepaged_thread = NULL;
2378                         goto fail;
2379                 }
2380
2381                 if (!list_empty(&khugepaged_scan.mm_head))
2382                         wake_up_interruptible(&khugepaged_wait);
2383
2384                 set_recommended_min_free_kbytes();
2385         } else if (khugepaged_thread) {
2386                 kthread_stop(khugepaged_thread);
2387                 khugepaged_thread = NULL;
2388         }
2389 fail:
2390         mutex_unlock(&khugepaged_mutex);
2391         return err;
2392 }
2393
2394 void khugepaged_min_free_kbytes_update(void)
2395 {
2396         mutex_lock(&khugepaged_mutex);
2397         if (khugepaged_enabled() && khugepaged_thread)
2398                 set_recommended_min_free_kbytes();
2399         mutex_unlock(&khugepaged_mutex);
2400 }