mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_EXCEED_SWAP_PTE,
  32         SCAN_EXCEED_SHARED_PTE,
  33         SCAN_PTE_NON_PRESENT,
  34         SCAN_PTE_UFFD_WP,
  35         SCAN_PAGE_RO,
  36         SCAN_LACK_REFERENCED_PAGE,
  37         SCAN_PAGE_NULL,
  38         SCAN_SCAN_ABORT,
  39         SCAN_PAGE_COUNT,
  40         SCAN_PAGE_LRU,
  41         SCAN_PAGE_LOCK,
  42         SCAN_PAGE_ANON,
  43         SCAN_PAGE_COMPOUND,
  44         SCAN_ANY_PROCESS,
  45         SCAN_VMA_NULL,
  46         SCAN_VMA_CHECK,
  47         SCAN_ADDRESS_RANGE,
  48         SCAN_SWAP_CACHE_PAGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 151                scan_sleep_millisecs_store);
 152
 153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 154                                           struct kobj_attribute *attr,
 155                                           char *buf)
 156 {
 157         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 158 }
 159
 160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 161                                            struct kobj_attribute *attr,
 162                                            const char *buf, size_t count)
 163 {
 164         unsigned int msecs;
 165         int err;
 166
 167         err = kstrtouint(buf, 10, &msecs);
 168         if (err)
 169                 return -EINVAL;
 170
 171         khugepaged_alloc_sleep_millisecs = msecs;
 172         khugepaged_sleep_expire = 0;
 173         wake_up_interruptible(&khugepaged_wait);
 174
 175         return count;
 176 }
 177 static struct kobj_attribute alloc_sleep_millisecs_attr =
 178         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 179                alloc_sleep_millisecs_store);
 180
 181 static ssize_t pages_to_scan_show(struct kobject *kobj,
 182                                   struct kobj_attribute *attr,
 183                                   char *buf)
 184 {
 185         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 186 }
 187 static ssize_t pages_to_scan_store(struct kobject *kobj,
 188                                    struct kobj_attribute *attr,
 189                                    const char *buf, size_t count)
 190 {
 191         unsigned int pages;
 192         int err;
 193
 194         err = kstrtouint(buf, 10, &pages);
 195         if (err || !pages)
 196                 return -EINVAL;
 197
 198         khugepaged_pages_to_scan = pages;
 199
 200         return count;
 201 }
 202 static struct kobj_attribute pages_to_scan_attr =
 203         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 204                pages_to_scan_store);
 205
 206 static ssize_t pages_collapsed_show(struct kobject *kobj,
 207                                     struct kobj_attribute *attr,
 208                                     char *buf)
 209 {
 210         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 211 }
 212 static struct kobj_attribute pages_collapsed_attr =
 213         __ATTR_RO(pages_collapsed);
 214
 215 static ssize_t full_scans_show(struct kobject *kobj,
 216                                struct kobj_attribute *attr,
 217                                char *buf)
 218 {
 219         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 220 }
 221 static struct kobj_attribute full_scans_attr =
 222         __ATTR_RO(full_scans);
 223
 224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 225                                       struct kobj_attribute *attr, char *buf)
 226 {
 227         return single_hugepage_flag_show(kobj, attr, buf,
 228                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 229 }
 230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 231                                        struct kobj_attribute *attr,
 232                                        const char *buf, size_t count)
 233 {
 234         return single_hugepage_flag_store(kobj, attr, buf, count,
 235                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 236 }
 237 static struct kobj_attribute khugepaged_defrag_attr =
 238         __ATTR(defrag, 0644, khugepaged_defrag_show,
 239                khugepaged_defrag_store);
 240
 241 /*
 242  * max_ptes_none controls if khugepaged should collapse hugepages over
 243  * any unmapped ptes in turn potentially increasing the memory
 244  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 245  * reduce the available free memory in the system as it
 246  * runs. Increasing max_ptes_none will instead potentially reduce the
 247  * free memory in the system during the khugepaged scan.
 248  */
 249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 250                                              struct kobj_attribute *attr,
 251                                              char *buf)
 252 {
 253         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 254 }
 255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 256                                               struct kobj_attribute *attr,
 257                                               const char *buf, size_t count)
 258 {
 259         int err;
 260         unsigned long max_ptes_none;
 261
 262         err = kstrtoul(buf, 10, &max_ptes_none);
 263         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 264                 return -EINVAL;
 265
 266         khugepaged_max_ptes_none = max_ptes_none;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 271         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 272                khugepaged_max_ptes_none_store);
 273
 274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 275                                              struct kobj_attribute *attr,
 276                                              char *buf)
 277 {
 278         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 279 }
 280
 281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 282                                               struct kobj_attribute *attr,
 283                                               const char *buf, size_t count)
 284 {
 285         int err;
 286         unsigned long max_ptes_swap;
 287
 288         err  = kstrtoul(buf, 10, &max_ptes_swap);
 289         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 290                 return -EINVAL;
 291
 292         khugepaged_max_ptes_swap = max_ptes_swap;
 293
 294         return count;
 295 }
 296
 297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 298         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 299                khugepaged_max_ptes_swap_store);
 300
 301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 302                                                struct kobj_attribute *attr,
 303                                                char *buf)
 304 {
 305         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 306 }
 307
 308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 309                                               struct kobj_attribute *attr,
 310                                               const char *buf, size_t count)
 311 {
 312         int err;
 313         unsigned long max_ptes_shared;
 314
 315         err  = kstrtoul(buf, 10, &max_ptes_shared);
 316         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 317                 return -EINVAL;
 318
 319         khugepaged_max_ptes_shared = max_ptes_shared;
 320
 321         return count;
 322 }
 323
 324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 325         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 326                khugepaged_max_ptes_shared_store);
 327
 328 static struct attribute *khugepaged_attr[] = {
 329         &khugepaged_defrag_attr.attr,
 330         &khugepaged_max_ptes_none_attr.attr,
 331         &khugepaged_max_ptes_swap_attr.attr,
 332         &khugepaged_max_ptes_shared_attr.attr,
 333         &pages_to_scan_attr.attr,
 334         &pages_collapsed_attr.attr,
 335         &full_scans_attr.attr,
 336         &scan_sleep_millisecs_attr.attr,
 337         &alloc_sleep_millisecs_attr.attr,
 338         NULL,
 339 };
 340
 341 struct attribute_group khugepaged_attr_group = {
 342         .attrs = khugepaged_attr,
 343         .name = "khugepaged",
 344 };
 345 #endif /* CONFIG_SYSFS */
 346
 347 int hugepage_madvise(struct vm_area_struct *vma,
 348                      unsigned long *vm_flags, int advice)
 349 {
 350         switch (advice) {
 351         case MADV_HUGEPAGE:
 352 #ifdef CONFIG_S390
 353                 /*
 354                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 355                  * can't handle this properly after s390_enable_sie, so we simply
 356                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 357                  */
 358                 if (mm_has_pgste(vma->vm_mm))
 359                         return 0;
 360 #endif
 361                 *vm_flags &= ~VM_NOHUGEPAGE;
 362                 *vm_flags |= VM_HUGEPAGE;
 363                 /*
 364                  * If the vma become good for khugepaged to scan,
 365                  * register it here without waiting a page fault that
 366                  * may not happen any time soon.
 367                  */
 368                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 369                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 370                         return -ENOMEM;
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct mm_slot),
 390                                           __alignof__(struct mm_slot), 0, NULL);
 391         if (!mm_slot_cache)
 392                 return -ENOMEM;
 393
 394         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 395         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 396         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 397         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 398
 399         return 0;
 400 }
 401
 402 void __init khugepaged_destroy(void)
 403 {
 404         kmem_cache_destroy(mm_slot_cache);
 405 }
 406
 407 static inline struct mm_slot *alloc_mm_slot(void)
 408 {
 409         if (!mm_slot_cache)     /* initialization failed */
 410                 return NULL;
 411         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 412 }
 413
 414 static inline void free_mm_slot(struct mm_slot *mm_slot)
 415 {
 416         kmem_cache_free(mm_slot_cache, mm_slot);
 417 }
 418
 419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 420 {
 421         struct mm_slot *mm_slot;
 422
 423         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 424                 if (mm == mm_slot->mm)
 425                         return mm_slot;
 426
 427         return NULL;
 428 }
 429
 430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 431                                     struct mm_slot *mm_slot)
 432 {
 433         mm_slot->mm = mm;
 434         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 435 }
 436
 437 static inline int khugepaged_test_exit(struct mm_struct *mm)
 438 {
 439         return atomic_read(&mm->mm_users) == 0;
 440 }
 441
 442 static bool hugepage_vma_check(struct vm_area_struct *vma,
 443                                unsigned long vm_flags)
 444 {
 445         /* Explicitly disabled through madvise. */
 446         if ((vm_flags & VM_NOHUGEPAGE) ||
 447             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 448                 return false;
 449
 450         /* Enabled via shmem mount options or sysfs settings. */
 451         if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
 452                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 453                                 HPAGE_PMD_NR);
 454         }
 455
 456         /* THP settings require madvise. */
 457         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 458                 return false;
 459
 460         /* Read-only file mappings need to be aligned for THP to work. */
 461         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 462             (vm_flags & VM_DENYWRITE)) {
 463                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 464                                 HPAGE_PMD_NR);
 465         }
 466
 467         if (!vma->anon_vma || vma->vm_ops)
 468                 return false;
 469         if (vma_is_temporary_stack(vma))
 470                 return false;
 471         return !(vm_flags & VM_NO_KHUGEPAGED);
 472 }
 473
 474 int __khugepaged_enter(struct mm_struct *mm)
 475 {
 476         struct mm_slot *mm_slot;
 477         int wakeup;
 478
 479         mm_slot = alloc_mm_slot();
 480         if (!mm_slot)
 481                 return -ENOMEM;
 482
 483         /* __khugepaged_exit() must not run from under us */
 484         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 485         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 486                 free_mm_slot(mm_slot);
 487                 return 0;
 488         }
 489
 490         spin_lock(&khugepaged_mm_lock);
 491         insert_to_mm_slots_hash(mm, mm_slot);
 492         /*
 493          * Insert just behind the scanning cursor, to let the area settle
 494          * down a little.
 495          */
 496         wakeup = list_empty(&khugepaged_scan.mm_head);
 497         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 498         spin_unlock(&khugepaged_mm_lock);
 499
 500         mmgrab(mm);
 501         if (wakeup)
 502                 wake_up_interruptible(&khugepaged_wait);
 503
 504         return 0;
 505 }
 506
 507 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 508                                unsigned long vm_flags)
 509 {
 510         unsigned long hstart, hend;
 511
 512         /*
 513          * khugepaged only supports read-only files for non-shmem files.
 514          * khugepaged does not yet work on special mappings. And
 515          * file-private shmem THP is not supported.
 516          */
 517         if (!hugepage_vma_check(vma, vm_flags))
 518                 return 0;
 519
 520         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 521         hend = vma->vm_end & HPAGE_PMD_MASK;
 522         if (hstart < hend)
 523                 return khugepaged_enter(vma, vm_flags);
 524         return 0;
 525 }
 526
 527 void __khugepaged_exit(struct mm_struct *mm)
 528 {
 529         struct mm_slot *mm_slot;
 530         int free = 0;
 531
 532         spin_lock(&khugepaged_mm_lock);
 533         mm_slot = get_mm_slot(mm);
 534         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 535                 hash_del(&mm_slot->hash);
 536                 list_del(&mm_slot->mm_node);
 537                 free = 1;
 538         }
 539         spin_unlock(&khugepaged_mm_lock);
 540
 541         if (free) {
 542                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 543                 free_mm_slot(mm_slot);
 544                 mmdrop(mm);
 545         } else if (mm_slot) {
 546                 /*
 547                  * This is required to serialize against
 548                  * khugepaged_test_exit() (which is guaranteed to run
 549                  * under mmap sem read mode). Stop here (after we
 550                  * return all pagetables will be destroyed) until
 551                  * khugepaged has finished working on the pagetables
 552                  * under the mmap_lock.
 553                  */
 554                 mmap_write_lock(mm);
 555                 mmap_write_unlock(mm);
 556         }
 557 }
 558
 559 static void release_pte_page(struct page *page)
 560 {
 561         mod_node_page_state(page_pgdat(page),
 562                         NR_ISOLATED_ANON + page_is_file_lru(page),
 563                         -compound_nr(page));
 564         unlock_page(page);
 565         putback_lru_page(page);
 566 }
 567
 568 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 569                 struct list_head *compound_pagelist)
 570 {
 571         struct page *page, *tmp;
 572
 573         while (--_pte >= pte) {
 574                 pte_t pteval = *_pte;
 575
 576                 page = pte_page(pteval);
 577                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 578                                 !PageCompound(page))
 579                         release_pte_page(page);
 580         }
 581
 582         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 583                 list_del(&page->lru);
 584                 release_pte_page(page);
 585         }
 586 }
 587
 588 static bool is_refcount_suitable(struct page *page)
 589 {
 590         int expected_refcount;
 591
 592         expected_refcount = total_mapcount(page);
 593         if (PageSwapCache(page))
 594                 expected_refcount += compound_nr(page);
 595
 596         return page_count(page) == expected_refcount;
 597 }
 598
 599 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 600                                         unsigned long address,
 601                                         pte_t *pte,
 602                                         struct list_head *compound_pagelist)
 603 {
 604         struct page *page = NULL;
 605         pte_t *_pte;
 606         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 607         bool writable = false;
 608
 609         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 610              _pte++, address += PAGE_SIZE) {
 611                 pte_t pteval = *_pte;
 612                 if (pte_none(pteval) || (pte_present(pteval) &&
 613                                 is_zero_pfn(pte_pfn(pteval)))) {
 614                         if (!userfaultfd_armed(vma) &&
 615                             ++none_or_zero <= khugepaged_max_ptes_none) {
 616                                 continue;
 617                         } else {
 618                                 result = SCAN_EXCEED_NONE_PTE;
 619                                 goto out;
 620                         }
 621                 }
 622                 if (!pte_present(pteval)) {
 623                         result = SCAN_PTE_NON_PRESENT;
 624                         goto out;
 625                 }
 626                 page = vm_normal_page(vma, address, pteval);
 627                 if (unlikely(!page)) {
 628                         result = SCAN_PAGE_NULL;
 629                         goto out;
 630                 }
 631
 632                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 633
 634                 if (page_mapcount(page) > 1 &&
 635                                 ++shared > khugepaged_max_ptes_shared) {
 636                         result = SCAN_EXCEED_SHARED_PTE;
 637                         goto out;
 638                 }
 639
 640                 if (PageCompound(page)) {
 641                         struct page *p;
 642                         page = compound_head(page);
 643
 644                         /*
 645                          * Check if we have dealt with the compound page
 646                          * already
 647                          */
 648                         list_for_each_entry(p, compound_pagelist, lru) {
 649                                 if (page == p)
 650                                         goto next;
 651                         }
 652                 }
 653
 654                 /*
 655                  * We can do it before isolate_lru_page because the
 656                  * page can't be freed from under us. NOTE: PG_lock
 657                  * is needed to serialize against split_huge_page
 658                  * when invoked from the VM.
 659                  */
 660                 if (!trylock_page(page)) {
 661                         result = SCAN_PAGE_LOCK;
 662                         goto out;
 663                 }
 664
 665                 /*
 666                  * Check if the page has any GUP (or other external) pins.
 667                  *
 668                  * The page table that maps the page has been already unlinked
 669                  * from the page table tree and this process cannot get
 670                  * an additinal pin on the page.
 671                  *
 672                  * New pins can come later if the page is shared across fork,
 673                  * but not from this process. The other process cannot write to
 674                  * the page, only trigger CoW.
 675                  */
 676                 if (!is_refcount_suitable(page)) {
 677                         unlock_page(page);
 678                         result = SCAN_PAGE_COUNT;
 679                         goto out;
 680                 }
 681                 if (!pte_write(pteval) && PageSwapCache(page) &&
 682                                 !reuse_swap_page(page, NULL)) {
 683                         /*
 684                          * Page is in the swap cache and cannot be re-used.
 685                          * It cannot be collapsed into a THP.
 686                          */
 687                         unlock_page(page);
 688                         result = SCAN_SWAP_CACHE_PAGE;
 689                         goto out;
 690                 }
 691
 692                 /*
 693                  * Isolate the page to avoid collapsing an hugepage
 694                  * currently in use by the VM.
 695                  */
 696                 if (isolate_lru_page(page)) {
 697                         unlock_page(page);
 698                         result = SCAN_DEL_PAGE_LRU;
 699                         goto out;
 700                 }
 701                 mod_node_page_state(page_pgdat(page),
 702                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 703                                 compound_nr(page));
 704                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 705                 VM_BUG_ON_PAGE(PageLRU(page), page);
 706
 707                 if (PageCompound(page))
 708                         list_add_tail(&page->lru, compound_pagelist);
 709 next:
 710                 /* There should be enough young pte to collapse the page */
 711                 if (pte_young(pteval) ||
 712                     page_is_young(page) || PageReferenced(page) ||
 713                     mmu_notifier_test_young(vma->vm_mm, address))
 714                         referenced++;
 715
 716                 if (pte_write(pteval))
 717                         writable = true;
 718         }
 719         if (likely(writable)) {
 720                 if (likely(referenced)) {
 721                         result = SCAN_SUCCEED;
 722                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 723                                                             referenced, writable, result);
 724                         return 1;
 725                 }
 726         } else {
 727                 result = SCAN_PAGE_RO;
 728         }
 729
 730 out:
 731         release_pte_pages(pte, _pte, compound_pagelist);
 732         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 733                                             referenced, writable, result);
 734         return 0;
 735 }
 736
 737 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 738                                       struct vm_area_struct *vma,
 739                                       unsigned long address,
 740                                       spinlock_t *ptl,
 741                                       struct list_head *compound_pagelist)
 742 {
 743         struct page *src_page, *tmp;
 744         pte_t *_pte;
 745         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 746                                 _pte++, page++, address += PAGE_SIZE) {
 747                 pte_t pteval = *_pte;
 748
 749                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 750                         clear_user_highpage(page, address);
 751                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 752                         if (is_zero_pfn(pte_pfn(pteval))) {
 753                                 /*
 754                                  * ptl mostly unnecessary.
 755                                  */
 756                                 spin_lock(ptl);
 757                                 /*
 758                                  * paravirt calls inside pte_clear here are
 759                                  * superfluous.
 760                                  */
 761                                 pte_clear(vma->vm_mm, address, _pte);
 762                                 spin_unlock(ptl);
 763                         }
 764                 } else {
 765                         src_page = pte_page(pteval);
 766                         copy_user_highpage(page, src_page, address, vma);
 767                         if (!PageCompound(src_page))
 768                                 release_pte_page(src_page);
 769                         /*
 770                          * ptl mostly unnecessary, but preempt has to
 771                          * be disabled to update the per-cpu stats
 772                          * inside page_remove_rmap().
 773                          */
 774                         spin_lock(ptl);
 775                         /*
 776                          * paravirt calls inside pte_clear here are
 777                          * superfluous.
 778                          */
 779                         pte_clear(vma->vm_mm, address, _pte);
 780                         page_remove_rmap(src_page, false);
 781                         spin_unlock(ptl);
 782                         free_page_and_swap_cache(src_page);
 783                 }
 784         }
 785
 786         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 787                 list_del(&src_page->lru);
 788                 release_pte_page(src_page);
 789         }
 790 }
 791
 792 static void khugepaged_alloc_sleep(void)
 793 {
 794         DEFINE_WAIT(wait);
 795
 796         add_wait_queue(&khugepaged_wait, &wait);
 797         freezable_schedule_timeout_interruptible(
 798                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 799         remove_wait_queue(&khugepaged_wait, &wait);
 800 }
 801
 802 static int khugepaged_node_load[MAX_NUMNODES];
 803
 804 static bool khugepaged_scan_abort(int nid)
 805 {
 806         int i;
 807
 808         /*
 809          * If node_reclaim_mode is disabled, then no extra effort is made to
 810          * allocate memory locally.
 811          */
 812         if (!node_reclaim_mode)
 813                 return false;
 814
 815         /* If there is a count for this node already, it must be acceptable */
 816         if (khugepaged_node_load[nid])
 817                 return false;
 818
 819         for (i = 0; i < MAX_NUMNODES; i++) {
 820                 if (!khugepaged_node_load[i])
 821                         continue;
 822                 if (node_distance(nid, i) > node_reclaim_distance)
 823                         return true;
 824         }
 825         return false;
 826 }
 827
 828 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 829 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 830 {
 831         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 832 }
 833
 834 #ifdef CONFIG_NUMA
 835 static int khugepaged_find_target_node(void)
 836 {
 837         static int last_khugepaged_target_node = NUMA_NO_NODE;
 838         int nid, target_node = 0, max_value = 0;
 839
 840         /* find first node with max normal pages hit */
 841         for (nid = 0; nid < MAX_NUMNODES; nid++)
 842                 if (khugepaged_node_load[nid] > max_value) {
 843                         max_value = khugepaged_node_load[nid];
 844                         target_node = nid;
 845                 }
 846
 847         /* do some balance if several nodes have the same hit record */
 848         if (target_node <= last_khugepaged_target_node)
 849                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 850                                 nid++)
 851                         if (max_value == khugepaged_node_load[nid]) {
 852                                 target_node = nid;
 853                                 break;
 854                         }
 855
 856         last_khugepaged_target_node = target_node;
 857         return target_node;
 858 }
 859
 860 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 861 {
 862         if (IS_ERR(*hpage)) {
 863                 if (!*wait)
 864                         return false;
 865
 866                 *wait = false;
 867                 *hpage = NULL;
 868                 khugepaged_alloc_sleep();
 869         } else if (*hpage) {
 870                 put_page(*hpage);
 871                 *hpage = NULL;
 872         }
 873
 874         return true;
 875 }
 876
 877 static struct page *
 878 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 879 {
 880         VM_BUG_ON_PAGE(*hpage, *hpage);
 881
 882         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 883         if (unlikely(!*hpage)) {
 884                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 885                 *hpage = ERR_PTR(-ENOMEM);
 886                 return NULL;
 887         }
 888
 889         prep_transhuge_page(*hpage);
 890         count_vm_event(THP_COLLAPSE_ALLOC);
 891         return *hpage;
 892 }
 893 #else
 894 static int khugepaged_find_target_node(void)
 895 {
 896         return 0;
 897 }
 898
 899 static inline struct page *alloc_khugepaged_hugepage(void)
 900 {
 901         struct page *page;
 902
 903         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 904                            HPAGE_PMD_ORDER);
 905         if (page)
 906                 prep_transhuge_page(page);
 907         return page;
 908 }
 909
 910 static struct page *khugepaged_alloc_hugepage(bool *wait)
 911 {
 912         struct page *hpage;
 913
 914         do {
 915                 hpage = alloc_khugepaged_hugepage();
 916                 if (!hpage) {
 917                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 918                         if (!*wait)
 919                                 return NULL;
 920
 921                         *wait = false;
 922                         khugepaged_alloc_sleep();
 923                 } else
 924                         count_vm_event(THP_COLLAPSE_ALLOC);
 925         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 926
 927         return hpage;
 928 }
 929
 930 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 931 {
 932         /*
 933          * If the hpage allocated earlier was briefly exposed in page cache
 934          * before collapse_file() failed, it is possible that racing lookups
 935          * have not yet completed, and would then be unpleasantly surprised by
 936          * finding the hpage reused for the same mapping at a different offset.
 937          * Just release the previous allocation if there is any danger of that.
 938          */
 939         if (*hpage && page_count(*hpage) > 1) {
 940                 put_page(*hpage);
 941                 *hpage = NULL;
 942         }
 943
 944         if (!*hpage)
 945                 *hpage = khugepaged_alloc_hugepage(wait);
 946
 947         if (unlikely(!*hpage))
 948                 return false;
 949
 950         return true;
 951 }
 952
 953 static struct page *
 954 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 955 {
 956         VM_BUG_ON(!*hpage);
 957
 958         return  *hpage;
 959 }
 960 #endif
 961
 962 /*
 963  * If mmap_lock temporarily dropped, revalidate vma
 964  * before taking mmap_lock.
 965  * Return 0 if succeeds, otherwise return none-zero
 966  * value (scan code).
 967  */
 968
 969 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 970                 struct vm_area_struct **vmap)
 971 {
 972         struct vm_area_struct *vma;
 973         unsigned long hstart, hend;
 974
 975         if (unlikely(khugepaged_test_exit(mm)))
 976                 return SCAN_ANY_PROCESS;
 977
 978         *vmap = vma = find_vma(mm, address);
 979         if (!vma)
 980                 return SCAN_VMA_NULL;
 981
 982         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 983         hend = vma->vm_end & HPAGE_PMD_MASK;
 984         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 985                 return SCAN_ADDRESS_RANGE;
 986         if (!hugepage_vma_check(vma, vma->vm_flags))
 987                 return SCAN_VMA_CHECK;
 988         /* Anon VMA expected */
 989         if (!vma->anon_vma || vma->vm_ops)
 990                 return SCAN_VMA_CHECK;
 991         return 0;
 992 }
 993
 994 /*
 995  * Bring missing pages in from swap, to complete THP collapse.
 996  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 997  *
 998  * Called and returns without pte mapped or spinlocks held,
 999  * but with mmap_lock held to protect against vma changes.
1000  */
1001
1002 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1003                                         struct vm_area_struct *vma,
1004                                         unsigned long haddr, pmd_t *pmd,
1005                                         int referenced)
1006 {
1007         int swapped_in = 0;
1008         vm_fault_t ret = 0;
1009         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1010
1011         for (address = haddr; address < end; address += PAGE_SIZE) {
1012                 struct vm_fault vmf = {
1013                         .vma = vma,
1014                         .address = address,
1015                         .pgoff = linear_page_index(vma, haddr),
1016                         .flags = FAULT_FLAG_ALLOW_RETRY,
1017                         .pmd = pmd,
1018                 };
1019
1020                 vmf.pte = pte_offset_map(pmd, address);
1021                 vmf.orig_pte = *vmf.pte;
1022                 if (!is_swap_pte(vmf.orig_pte)) {
1023                         pte_unmap(vmf.pte);
1024                         continue;
1025                 }
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, haddr, &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, haddr) != 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         }
1048
1049         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1050         if (swapped_in)
1051                 lru_add_drain();
1052
1053         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1054         return true;
1055 }
1056
1057 static void collapse_huge_page(struct mm_struct *mm,
1058                                    unsigned long address,
1059                                    struct page **hpage,
1060                                    int node, int referenced, int unmapped)
1061 {
1062         LIST_HEAD(compound_pagelist);
1063         pmd_t *pmd, _pmd;
1064         pte_t *pte;
1065         pgtable_t pgtable;
1066         struct page *new_page;
1067         spinlock_t *pmd_ptl, *pte_ptl;
1068         int isolated = 0, result = 0;
1069         struct vm_area_struct *vma;
1070         struct mmu_notifier_range range;
1071         gfp_t gfp;
1072
1073         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1074
1075         /* Only allocate from the target node */
1076         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1077
1078         /*
1079          * Before allocating the hugepage, release the mmap_lock read lock.
1080          * The allocation can take potentially a long time if it involves
1081          * sync compaction, and we do not need to hold the mmap_lock during
1082          * that. We will recheck the vma after taking it again in write mode.
1083          */
1084         mmap_read_unlock(mm);
1085         new_page = khugepaged_alloc_page(hpage, gfp, node);
1086         if (!new_page) {
1087                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1088                 goto out_nolock;
1089         }
1090
1091         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1092                 result = SCAN_CGROUP_CHARGE_FAIL;
1093                 goto out_nolock;
1094         }
1095         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1096
1097         mmap_read_lock(mm);
1098         result = hugepage_vma_revalidate(mm, address, &vma);
1099         if (result) {
1100                 mmap_read_unlock(mm);
1101                 goto out_nolock;
1102         }
1103
1104         pmd = mm_find_pmd(mm, address);
1105         if (!pmd) {
1106                 result = SCAN_PMD_NULL;
1107                 mmap_read_unlock(mm);
1108                 goto out_nolock;
1109         }
1110
1111         /*
1112          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1113          * If it fails, we release mmap_lock and jump out_nolock.
1114          * Continuing to collapse causes inconsistency.
1115          */
1116         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1117                                                      pmd, referenced)) {
1118                 mmap_read_unlock(mm);
1119                 goto out_nolock;
1120         }
1121
1122         mmap_read_unlock(mm);
1123         /*
1124          * Prevent all access to pagetables with the exception of
1125          * gup_fast later handled by the ptep_clear_flush and the VM
1126          * handled by the anon_vma lock + PG_lock.
1127          */
1128         mmap_write_lock(mm);
1129         result = hugepage_vma_revalidate(mm, address, &vma);
1130         if (result)
1131                 goto out;
1132         /* check if the pmd is still valid */
1133         if (mm_find_pmd(mm, address) != pmd)
1134                 goto out;
1135
1136         anon_vma_lock_write(vma->anon_vma);
1137
1138         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1139                                 address, address + HPAGE_PMD_SIZE);
1140         mmu_notifier_invalidate_range_start(&range);
1141
1142         pte = pte_offset_map(pmd, address);
1143         pte_ptl = pte_lockptr(mm, pmd);
1144
1145         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1146         /*
1147          * After this gup_fast can't run anymore. This also removes
1148          * any huge TLB entry from the CPU so we won't allow
1149          * huge and small TLB entries for the same virtual address
1150          * to avoid the risk of CPU bugs in that area.
1151          */
1152         _pmd = pmdp_collapse_flush(vma, address, pmd);
1153         spin_unlock(pmd_ptl);
1154         mmu_notifier_invalidate_range_end(&range);
1155
1156         spin_lock(pte_ptl);
1157         isolated = __collapse_huge_page_isolate(vma, address, pte,
1158                         &compound_pagelist);
1159         spin_unlock(pte_ptl);
1160
1161         if (unlikely(!isolated)) {
1162                 pte_unmap(pte);
1163                 spin_lock(pmd_ptl);
1164                 BUG_ON(!pmd_none(*pmd));
1165                 /*
1166                  * We can only use set_pmd_at when establishing
1167                  * hugepmds and never for establishing regular pmds that
1168                  * points to regular pagetables. Use pmd_populate for that
1169                  */
1170                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1171                 spin_unlock(pmd_ptl);
1172                 anon_vma_unlock_write(vma->anon_vma);
1173                 result = SCAN_FAIL;
1174                 goto out;
1175         }
1176
1177         /*
1178          * All pages are isolated and locked so anon_vma rmap
1179          * can't run anymore.
1180          */
1181         anon_vma_unlock_write(vma->anon_vma);
1182
1183         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1184                         &compound_pagelist);
1185         pte_unmap(pte);
1186         __SetPageUptodate(new_page);
1187         pgtable = pmd_pgtable(_pmd);
1188
1189         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1190         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1191
1192         /*
1193          * spin_lock() below is not the equivalent of smp_wmb(), so
1194          * this is needed to avoid the copy_huge_page writes to become
1195          * visible after the set_pmd_at() write.
1196          */
1197         smp_wmb();
1198
1199         spin_lock(pmd_ptl);
1200         BUG_ON(!pmd_none(*pmd));
1201         page_add_new_anon_rmap(new_page, vma, address, true);
1202         lru_cache_add_inactive_or_unevictable(new_page, vma);
1203         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1204         set_pmd_at(mm, address, pmd, _pmd);
1205         update_mmu_cache_pmd(vma, address, pmd);
1206         spin_unlock(pmd_ptl);
1207
1208         *hpage = NULL;
1209
1210         khugepaged_pages_collapsed++;
1211         result = SCAN_SUCCEED;
1212 out_up_write:
1213         mmap_write_unlock(mm);
1214 out_nolock:
1215         if (!IS_ERR_OR_NULL(*hpage))
1216                 mem_cgroup_uncharge(*hpage);
1217         trace_mm_collapse_huge_page(mm, isolated, result);
1218         return;
1219 out:
1220         goto out_up_write;
1221 }
1222
1223 static int khugepaged_scan_pmd(struct mm_struct *mm,
1224                                struct vm_area_struct *vma,
1225                                unsigned long address,
1226                                struct page **hpage)
1227 {
1228         pmd_t *pmd;
1229         pte_t *pte, *_pte;
1230         int ret = 0, result = 0, referenced = 0;
1231         int none_or_zero = 0, shared = 0;
1232         struct page *page = NULL;
1233         unsigned long _address;
1234         spinlock_t *ptl;
1235         int node = NUMA_NO_NODE, unmapped = 0;
1236         bool writable = false;
1237
1238         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1239
1240         pmd = mm_find_pmd(mm, address);
1241         if (!pmd) {
1242                 result = SCAN_PMD_NULL;
1243                 goto out;
1244         }
1245
1246         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1247         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1248         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1249              _pte++, _address += PAGE_SIZE) {
1250                 pte_t pteval = *_pte;
1251                 if (is_swap_pte(pteval)) {
1252                         if (++unmapped <= khugepaged_max_ptes_swap) {
1253                                 /*
1254                                  * Always be strict with uffd-wp
1255                                  * enabled swap entries.  Please see
1256                                  * comment below for pte_uffd_wp().
1257                                  */
1258                                 if (pte_swp_uffd_wp(pteval)) {
1259                                         result = SCAN_PTE_UFFD_WP;
1260                                         goto out_unmap;
1261                                 }
1262                                 continue;
1263                         } else {
1264                                 result = SCAN_EXCEED_SWAP_PTE;
1265                                 goto out_unmap;
1266                         }
1267                 }
1268                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1269                         if (!userfaultfd_armed(vma) &&
1270                             ++none_or_zero <= khugepaged_max_ptes_none) {
1271                                 continue;
1272                         } else {
1273                                 result = SCAN_EXCEED_NONE_PTE;
1274                                 goto out_unmap;
1275                         }
1276                 }
1277                 if (!pte_present(pteval)) {
1278                         result = SCAN_PTE_NON_PRESENT;
1279                         goto out_unmap;
1280                 }
1281                 if (pte_uffd_wp(pteval)) {
1282                         /*
1283                          * Don't collapse the page if any of the small
1284                          * PTEs are armed with uffd write protection.
1285                          * Here we can also mark the new huge pmd as
1286                          * write protected if any of the small ones is
1287                          * marked but that could bring unknown
1288                          * userfault messages that falls outside of
1289                          * the registered range.  So, just be simple.
1290                          */
1291                         result = SCAN_PTE_UFFD_WP;
1292                         goto out_unmap;
1293                 }
1294                 if (pte_write(pteval))
1295                         writable = true;
1296
1297                 page = vm_normal_page(vma, _address, pteval);
1298                 if (unlikely(!page)) {
1299                         result = SCAN_PAGE_NULL;
1300                         goto out_unmap;
1301                 }
1302
1303                 if (page_mapcount(page) > 1 &&
1304                                 ++shared > khugepaged_max_ptes_shared) {
1305                         result = SCAN_EXCEED_SHARED_PTE;
1306                         goto out_unmap;
1307                 }
1308
1309                 page = compound_head(page);
1310
1311                 /*
1312                  * Record which node the original page is from and save this
1313                  * information to khugepaged_node_load[].
1314                  * Khupaged will allocate hugepage from the node has the max
1315                  * hit record.
1316                  */
1317                 node = page_to_nid(page);
1318                 if (khugepaged_scan_abort(node)) {
1319                         result = SCAN_SCAN_ABORT;
1320                         goto out_unmap;
1321                 }
1322                 khugepaged_node_load[node]++;
1323                 if (!PageLRU(page)) {
1324                         result = SCAN_PAGE_LRU;
1325                         goto out_unmap;
1326                 }
1327                 if (PageLocked(page)) {
1328                         result = SCAN_PAGE_LOCK;
1329                         goto out_unmap;
1330                 }
1331                 if (!PageAnon(page)) {
1332                         result = SCAN_PAGE_ANON;
1333                         goto out_unmap;
1334                 }
1335
1336                 /*
1337                  * Check if the page has any GUP (or other external) pins.
1338                  *
1339                  * Here the check is racy it may see totmal_mapcount > refcount
1340                  * in some cases.
1341                  * For example, one process with one forked child process.
1342                  * The parent has the PMD split due to MADV_DONTNEED, then
1343                  * the child is trying unmap the whole PMD, but khugepaged
1344                  * may be scanning the parent between the child has
1345                  * PageDoubleMap flag cleared and dec the mapcount.  So
1346                  * khugepaged may see total_mapcount > refcount.
1347                  *
1348                  * But such case is ephemeral we could always retry collapse
1349                  * later.  However it may report false positive if the page
1350                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1351                  * will be done again later the risk seems low.
1352                  */
1353                 if (!is_refcount_suitable(page)) {
1354                         result = SCAN_PAGE_COUNT;
1355                         goto out_unmap;
1356                 }
1357                 if (pte_young(pteval) ||
1358                     page_is_young(page) || PageReferenced(page) ||
1359                     mmu_notifier_test_young(vma->vm_mm, address))
1360                         referenced++;
1361         }
1362         if (!writable) {
1363                 result = SCAN_PAGE_RO;
1364         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1365                 result = SCAN_LACK_REFERENCED_PAGE;
1366         } else {
1367                 result = SCAN_SUCCEED;
1368                 ret = 1;
1369         }
1370 out_unmap:
1371         pte_unmap_unlock(pte, ptl);
1372         if (ret) {
1373                 node = khugepaged_find_target_node();
1374                 /* collapse_huge_page will return with the mmap_lock released */
1375                 collapse_huge_page(mm, address, hpage, node,
1376                                 referenced, unmapped);
1377         }
1378 out:
1379         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1380                                      none_or_zero, result, unmapped);
1381         return ret;
1382 }
1383
1384 static void collect_mm_slot(struct mm_slot *mm_slot)
1385 {
1386         struct mm_struct *mm = mm_slot->mm;
1387
1388         lockdep_assert_held(&khugepaged_mm_lock);
1389
1390         if (khugepaged_test_exit(mm)) {
1391                 /* free mm_slot */
1392                 hash_del(&mm_slot->hash);
1393                 list_del(&mm_slot->mm_node);
1394
1395                 /*
1396                  * Not strictly needed because the mm exited already.
1397                  *
1398                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1399                  */
1400
1401                 /* khugepaged_mm_lock actually not necessary for the below */
1402                 free_mm_slot(mm_slot);
1403                 mmdrop(mm);
1404         }
1405 }
1406
1407 #ifdef CONFIG_SHMEM
1408 /*
1409  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1410  * khugepaged should try to collapse the page table.
1411  */
1412 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1413                                          unsigned long addr)
1414 {
1415         struct mm_slot *mm_slot;
1416
1417         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1418
1419         spin_lock(&khugepaged_mm_lock);
1420         mm_slot = get_mm_slot(mm);
1421         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1422                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1423         spin_unlock(&khugepaged_mm_lock);
1424         return 0;
1425 }
1426
1427 /**
1428  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1429  * address haddr.
1430  *
1431  * @mm: process address space where collapse happens
1432  * @addr: THP collapse address
1433  *
1434  * This function checks whether all the PTEs in the PMD are pointing to the
1435  * right THP. If so, retract the page table so the THP can refault in with
1436  * as pmd-mapped.
1437  */
1438 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1439 {
1440         unsigned long haddr = addr & HPAGE_PMD_MASK;
1441         struct vm_area_struct *vma = find_vma(mm, haddr);
1442         struct page *hpage;
1443         pte_t *start_pte, *pte;
1444         pmd_t *pmd, _pmd;
1445         spinlock_t *ptl;
1446         int count = 0;
1447         int i;
1448
1449         if (!vma || !vma->vm_file ||
1450             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1451                 return;
1452
1453         /*
1454          * This vm_flags may not have VM_HUGEPAGE if the page was not
1455          * collapsed by this mm. But we can still collapse if the page is
1456          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1457          * will not fail the vma for missing VM_HUGEPAGE
1458          */
1459         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1460                 return;
1461
1462         hpage = find_lock_page(vma->vm_file->f_mapping,
1463                                linear_page_index(vma, haddr));
1464         if (!hpage)
1465                 return;
1466
1467         if (!PageHead(hpage))
1468                 goto drop_hpage;
1469
1470         pmd = mm_find_pmd(mm, haddr);
1471         if (!pmd)
1472                 goto drop_hpage;
1473
1474         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1475
1476         /* step 1: check all mapped PTEs are to the right huge page */
1477         for (i = 0, addr = haddr, pte = start_pte;
1478              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1479                 struct page *page;
1480
1481                 /* empty pte, skip */
1482                 if (pte_none(*pte))
1483                         continue;
1484
1485                 /* page swapped out, abort */
1486                 if (!pte_present(*pte))
1487                         goto abort;
1488
1489                 page = vm_normal_page(vma, addr, *pte);
1490
1491                 /*
1492                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1493                  * page table, but the new page will not be a subpage of hpage.
1494                  */
1495                 if (hpage + i != page)
1496                         goto abort;
1497                 count++;
1498         }
1499
1500         /* step 2: adjust rmap */
1501         for (i = 0, addr = haddr, pte = start_pte;
1502              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1503                 struct page *page;
1504
1505                 if (pte_none(*pte))
1506                         continue;
1507                 page = vm_normal_page(vma, addr, *pte);
1508                 page_remove_rmap(page, false);
1509         }
1510
1511         pte_unmap_unlock(start_pte, ptl);
1512
1513         /* step 3: set proper refcount and mm_counters. */
1514         if (count) {
1515                 page_ref_sub(hpage, count);
1516                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1517         }
1518
1519         /* step 4: collapse pmd */
1520         ptl = pmd_lock(vma->vm_mm, pmd);
1521         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1522         spin_unlock(ptl);
1523         mm_dec_nr_ptes(mm);
1524         pte_free(mm, pmd_pgtable(_pmd));
1525
1526 drop_hpage:
1527         unlock_page(hpage);
1528         put_page(hpage);
1529         return;
1530
1531 abort:
1532         pte_unmap_unlock(start_pte, ptl);
1533         goto drop_hpage;
1534 }
1535
1536 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1537 {
1538         struct mm_struct *mm = mm_slot->mm;
1539         int i;
1540
1541         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1542                 return 0;
1543
1544         if (!mmap_write_trylock(mm))
1545                 return -EBUSY;
1546
1547         if (unlikely(khugepaged_test_exit(mm)))
1548                 goto out;
1549
1550         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1551                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1552
1553 out:
1554         mm_slot->nr_pte_mapped_thp = 0;
1555         mmap_write_unlock(mm);
1556         return 0;
1557 }
1558
1559 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1560 {
1561         struct vm_area_struct *vma;
1562         struct mm_struct *mm;
1563         unsigned long addr;
1564         pmd_t *pmd, _pmd;
1565
1566         i_mmap_lock_write(mapping);
1567         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1568                 /*
1569                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1570                  * got written to. These VMAs are likely not worth investing
1571                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1572                  * later.
1573                  *
1574                  * Not that vma->anon_vma check is racy: it can be set up after
1575                  * the check but before we took mmap_lock by the fault path.
1576                  * But page lock would prevent establishing any new ptes of the
1577                  * page, so we are safe.
1578                  *
1579                  * An alternative would be drop the check, but check that page
1580                  * table is clear before calling pmdp_collapse_flush() under
1581                  * ptl. It has higher chance to recover THP for the VMA, but
1582                  * has higher cost too.
1583                  */
1584                 if (vma->anon_vma)
1585                         continue;
1586                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1587                 if (addr & ~HPAGE_PMD_MASK)
1588                         continue;
1589                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1590                         continue;
1591                 mm = vma->vm_mm;
1592                 pmd = mm_find_pmd(mm, addr);
1593                 if (!pmd)
1594                         continue;
1595                 /*
1596                  * We need exclusive mmap_lock to retract page table.
1597                  *
1598                  * We use trylock due to lock inversion: we need to acquire
1599                  * mmap_lock while holding page lock. Fault path does it in
1600                  * reverse order. Trylock is a way to avoid deadlock.
1601                  */
1602                 if (mmap_write_trylock(mm)) {
1603                         if (!khugepaged_test_exit(mm)) {
1604                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1605                                 /* assume page table is clear */
1606                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1607                                 spin_unlock(ptl);
1608                                 mm_dec_nr_ptes(mm);
1609                                 pte_free(mm, pmd_pgtable(_pmd));
1610                         }
1611                         mmap_write_unlock(mm);
1612                 } else {
1613                         /* Try again later */
1614                         khugepaged_add_pte_mapped_thp(mm, addr);
1615                 }
1616         }
1617         i_mmap_unlock_write(mapping);
1618 }
1619
1620 /**
1621  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1622  *
1623  * @mm: process address space where collapse happens
1624  * @file: file that collapse on
1625  * @start: collapse start address
1626  * @hpage: new allocated huge page for collapse
1627  * @node: appointed node the new huge page allocate from
1628  *
1629  * Basic scheme is simple, details are more complex:
1630  *  - allocate and lock a new huge page;
1631  *  - scan page cache replacing old pages with the new one
1632  *    + swap/gup in pages if necessary;
1633  *    + fill in gaps;
1634  *    + keep old pages around in case rollback is required;
1635  *  - if replacing succeeds:
1636  *    + copy data over;
1637  *    + free old pages;
1638  *    + unlock huge page;
1639  *  - if replacing failed;
1640  *    + put all pages back and unfreeze them;
1641  *    + restore gaps in the page cache;
1642  *    + unlock and free huge page;
1643  */
1644 static void collapse_file(struct mm_struct *mm,
1645                 struct file *file, pgoff_t start,
1646                 struct page **hpage, int node)
1647 {
1648         struct address_space *mapping = file->f_mapping;
1649         gfp_t gfp;
1650         struct page *new_page;
1651         pgoff_t index, end = start + HPAGE_PMD_NR;
1652         LIST_HEAD(pagelist);
1653         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1654         int nr_none = 0, result = SCAN_SUCCEED;
1655         bool is_shmem = shmem_file(file);
1656         int nr;
1657
1658         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1659         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1660
1661         /* Only allocate from the target node */
1662         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1663
1664         new_page = khugepaged_alloc_page(hpage, gfp, node);
1665         if (!new_page) {
1666                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1667                 goto out;
1668         }
1669
1670         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1671                 result = SCAN_CGROUP_CHARGE_FAIL;
1672                 goto out;
1673         }
1674         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1675
1676         /* This will be less messy when we use multi-index entries */
1677         do {
1678                 xas_lock_irq(&xas);
1679                 xas_create_range(&xas);
1680                 if (!xas_error(&xas))
1681                         break;
1682                 xas_unlock_irq(&xas);
1683                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1684                         result = SCAN_FAIL;
1685                         goto out;
1686                 }
1687         } while (1);
1688
1689         __SetPageLocked(new_page);
1690         if (is_shmem)
1691                 __SetPageSwapBacked(new_page);
1692         new_page->index = start;
1693         new_page->mapping = mapping;
1694
1695         /*
1696          * At this point the new_page is locked and not up-to-date.
1697          * It's safe to insert it into the page cache, because nobody would
1698          * be able to map it or use it in another way until we unlock it.
1699          */
1700
1701         xas_set(&xas, start);
1702         for (index = start; index < end; index++) {
1703                 struct page *page = xas_next(&xas);
1704
1705                 VM_BUG_ON(index != xas.xa_index);
1706                 if (is_shmem) {
1707                         if (!page) {
1708                                 /*
1709                                  * Stop if extent has been truncated or
1710                                  * hole-punched, and is now completely
1711                                  * empty.
1712                                  */
1713                                 if (index == start) {
1714                                         if (!xas_next_entry(&xas, end - 1)) {
1715                                                 result = SCAN_TRUNCATED;
1716                                                 goto xa_locked;
1717                                         }
1718                                         xas_set(&xas, index);
1719                                 }
1720                                 if (!shmem_charge(mapping->host, 1)) {
1721                                         result = SCAN_FAIL;
1722                                         goto xa_locked;
1723                                 }
1724                                 xas_store(&xas, new_page);
1725                                 nr_none++;
1726                                 continue;
1727                         }
1728
1729                         if (xa_is_value(page) || !PageUptodate(page)) {
1730                                 xas_unlock_irq(&xas);
1731                                 /* swap in or instantiate fallocated page */
1732                                 if (shmem_getpage(mapping->host, index, &page,
1733                                                   SGP_NOHUGE)) {
1734                                         result = SCAN_FAIL;
1735                                         goto xa_unlocked;
1736                                 }
1737                         } else if (trylock_page(page)) {
1738                                 get_page(page);
1739                                 xas_unlock_irq(&xas);
1740                         } else {
1741                                 result = SCAN_PAGE_LOCK;
1742                                 goto xa_locked;
1743                         }
1744                 } else {        /* !is_shmem */
1745                         if (!page || xa_is_value(page)) {
1746                                 xas_unlock_irq(&xas);
1747                                 page_cache_sync_readahead(mapping, &file->f_ra,
1748                                                           file, index,
1749                                                           end - index);
1750                                 /* drain pagevecs to help isolate_lru_page() */
1751                                 lru_add_drain();
1752                                 page = find_lock_page(mapping, index);
1753                                 if (unlikely(page == NULL)) {
1754                                         result = SCAN_FAIL;
1755                                         goto xa_unlocked;
1756                                 }
1757                         } else if (PageDirty(page)) {
1758                                 /*
1759                                  * khugepaged only works on read-only fd,
1760                                  * so this page is dirty because it hasn't
1761                                  * been flushed since first write. There
1762                                  * won't be new dirty pages.
1763                                  *
1764                                  * Trigger async flush here and hope the
1765                                  * writeback is done when khugepaged
1766                                  * revisits this page.
1767                                  *
1768                                  * This is a one-off situation. We are not
1769                                  * forcing writeback in loop.
1770                                  */
1771                                 xas_unlock_irq(&xas);
1772                                 filemap_flush(mapping);
1773                                 result = SCAN_FAIL;
1774                                 goto xa_unlocked;
1775                         } else if (trylock_page(page)) {
1776                                 get_page(page);
1777                                 xas_unlock_irq(&xas);
1778                         } else {
1779                                 result = SCAN_PAGE_LOCK;
1780                                 goto xa_locked;
1781                         }
1782                 }
1783
1784                 /*
1785                  * The page must be locked, so we can drop the i_pages lock
1786                  * without racing with truncate.
1787                  */
1788                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1789
1790                 /* make sure the page is up to date */
1791                 if (unlikely(!PageUptodate(page))) {
1792                         result = SCAN_FAIL;
1793                         goto out_unlock;
1794                 }
1795
1796                 /*
1797                  * If file was truncated then extended, or hole-punched, before
1798                  * we locked the first page, then a THP might be there already.
1799                  */
1800                 if (PageTransCompound(page)) {
1801                         result = SCAN_PAGE_COMPOUND;
1802                         goto out_unlock;
1803                 }
1804
1805                 if (page_mapping(page) != mapping) {
1806                         result = SCAN_TRUNCATED;
1807                         goto out_unlock;
1808                 }
1809
1810                 if (!is_shmem && PageDirty(page)) {
1811                         /*
1812                          * khugepaged only works on read-only fd, so this
1813                          * page is dirty because it hasn't been flushed
1814                          * since first write.
1815                          */
1816                         result = SCAN_FAIL;
1817                         goto out_unlock;
1818                 }
1819
1820                 if (isolate_lru_page(page)) {
1821                         result = SCAN_DEL_PAGE_LRU;
1822                         goto out_unlock;
1823                 }
1824
1825                 if (page_has_private(page) &&
1826                     !try_to_release_page(page, GFP_KERNEL)) {
1827                         result = SCAN_PAGE_HAS_PRIVATE;
1828                         putback_lru_page(page);
1829                         goto out_unlock;
1830                 }
1831
1832                 if (page_mapped(page))
1833                         unmap_mapping_pages(mapping, index, 1, false);
1834
1835                 xas_lock_irq(&xas);
1836                 xas_set(&xas, index);
1837
1838                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1839                 VM_BUG_ON_PAGE(page_mapped(page), page);
1840
1841                 /*
1842                  * The page is expected to have page_count() == 3:
1843                  *  - we hold a pin on it;
1844                  *  - one reference from page cache;
1845                  *  - one from isolate_lru_page;
1846                  */
1847                 if (!page_ref_freeze(page, 3)) {
1848                         result = SCAN_PAGE_COUNT;
1849                         xas_unlock_irq(&xas);
1850                         putback_lru_page(page);
1851                         goto out_unlock;
1852                 }
1853
1854                 /*
1855                  * Add the page to the list to be able to undo the collapse if
1856                  * something go wrong.
1857                  */
1858                 list_add_tail(&page->lru, &pagelist);
1859
1860                 /* Finally, replace with the new page. */
1861                 xas_store(&xas, new_page);
1862                 continue;
1863 out_unlock:
1864                 unlock_page(page);
1865                 put_page(page);
1866                 goto xa_unlocked;
1867         }
1868         nr = thp_nr_pages(new_page);
1869
1870         if (is_shmem)
1871                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1872         else {
1873                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1874                 filemap_nr_thps_inc(mapping);
1875         }
1876
1877         if (nr_none) {
1878                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1879                 if (is_shmem)
1880                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1881         }
1882
1883 xa_locked:
1884         xas_unlock_irq(&xas);
1885 xa_unlocked:
1886
1887         if (result == SCAN_SUCCEED) {
1888                 struct page *page, *tmp;
1889
1890                 /*
1891                  * Replacing old pages with new one has succeeded, now we
1892                  * need to copy the content and free the old pages.
1893                  */
1894                 index = start;
1895                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1896                         while (index < page->index) {
1897                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1898                                 index++;
1899                         }
1900                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1901                                         page);
1902                         list_del(&page->lru);
1903                         page->mapping = NULL;
1904                         page_ref_unfreeze(page, 1);
1905                         ClearPageActive(page);
1906                         ClearPageUnevictable(page);
1907                         unlock_page(page);
1908                         put_page(page);
1909                         index++;
1910                 }
1911                 while (index < end) {
1912                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1913                         index++;
1914                 }
1915
1916                 SetPageUptodate(new_page);
1917                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1918                 if (is_shmem)
1919                         set_page_dirty(new_page);
1920                 lru_cache_add(new_page);
1921
1922                 /*
1923                  * Remove pte page tables, so we can re-fault the page as huge.
1924                  */
1925                 retract_page_tables(mapping, start);
1926                 *hpage = NULL;
1927
1928                 khugepaged_pages_collapsed++;
1929         } else {
1930                 struct page *page;
1931
1932                 /* Something went wrong: roll back page cache changes */
1933                 xas_lock_irq(&xas);
1934                 mapping->nrpages -= nr_none;
1935
1936                 if (is_shmem)
1937                         shmem_uncharge(mapping->host, nr_none);
1938
1939                 xas_set(&xas, start);
1940                 xas_for_each(&xas, page, end - 1) {
1941                         page = list_first_entry_or_null(&pagelist,
1942                                         struct page, lru);
1943                         if (!page || xas.xa_index < page->index) {
1944                                 if (!nr_none)
1945                                         break;
1946                                 nr_none--;
1947                                 /* Put holes back where they were */
1948                                 xas_store(&xas, NULL);
1949                                 continue;
1950                         }
1951
1952                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1953
1954                         /* Unfreeze the page. */
1955                         list_del(&page->lru);
1956                         page_ref_unfreeze(page, 2);
1957                         xas_store(&xas, page);
1958                         xas_pause(&xas);
1959                         xas_unlock_irq(&xas);
1960                         unlock_page(page);
1961                         putback_lru_page(page);
1962                         xas_lock_irq(&xas);
1963                 }
1964                 VM_BUG_ON(nr_none);
1965                 xas_unlock_irq(&xas);
1966
1967                 new_page->mapping = NULL;
1968         }
1969
1970         unlock_page(new_page);
1971 out:
1972         VM_BUG_ON(!list_empty(&pagelist));
1973         if (!IS_ERR_OR_NULL(*hpage))
1974                 mem_cgroup_uncharge(*hpage);
1975         /* TODO: tracepoints */
1976 }
1977
1978 static void khugepaged_scan_file(struct mm_struct *mm,
1979                 struct file *file, pgoff_t start, struct page **hpage)
1980 {
1981         struct page *page = NULL;
1982         struct address_space *mapping = file->f_mapping;
1983         XA_STATE(xas, &mapping->i_pages, start);
1984         int present, swap;
1985         int node = NUMA_NO_NODE;
1986         int result = SCAN_SUCCEED;
1987
1988         present = 0;
1989         swap = 0;
1990         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1991         rcu_read_lock();
1992         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1993                 if (xas_retry(&xas, page))
1994                         continue;
1995
1996                 if (xa_is_value(page)) {
1997                         if (++swap > khugepaged_max_ptes_swap) {
1998                                 result = SCAN_EXCEED_SWAP_PTE;
1999                                 break;
2000                         }
2001                         continue;
2002                 }
2003
2004                 if (PageTransCompound(page)) {
2005                         result = SCAN_PAGE_COMPOUND;
2006                         break;
2007                 }
2008
2009                 node = page_to_nid(page);
2010                 if (khugepaged_scan_abort(node)) {
2011                         result = SCAN_SCAN_ABORT;
2012                         break;
2013                 }
2014                 khugepaged_node_load[node]++;
2015
2016                 if (!PageLRU(page)) {
2017                         result = SCAN_PAGE_LRU;
2018                         break;
2019                 }
2020
2021                 if (page_count(page) !=
2022                     1 + page_mapcount(page) + page_has_private(page)) {
2023                         result = SCAN_PAGE_COUNT;
2024                         break;
2025                 }
2026
2027                 /*
2028                  * We probably should check if the page is referenced here, but
2029                  * nobody would transfer pte_young() to PageReferenced() for us.
2030                  * And rmap walk here is just too costly...
2031                  */
2032
2033                 present++;
2034
2035                 if (need_resched()) {
2036                         xas_pause(&xas);
2037                         cond_resched_rcu();
2038                 }
2039         }
2040         rcu_read_unlock();
2041
2042         if (result == SCAN_SUCCEED) {
2043                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2044                         result = SCAN_EXCEED_NONE_PTE;
2045                 } else {
2046                         node = khugepaged_find_target_node();
2047                         collapse_file(mm, file, start, hpage, node);
2048                 }
2049         }
2050
2051         /* TODO: tracepoints */
2052 }
2053 #else
2054 static void khugepaged_scan_file(struct mm_struct *mm,
2055                 struct file *file, pgoff_t start, struct page **hpage)
2056 {
2057         BUILD_BUG();
2058 }
2059
2060 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2061 {
2062         return 0;
2063 }
2064 #endif
2065
2066 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2067                                             struct page **hpage)
2068         __releases(&khugepaged_mm_lock)
2069         __acquires(&khugepaged_mm_lock)
2070 {
2071         struct mm_slot *mm_slot;
2072         struct mm_struct *mm;
2073         struct vm_area_struct *vma;
2074         int progress = 0;
2075
2076         VM_BUG_ON(!pages);
2077         lockdep_assert_held(&khugepaged_mm_lock);
2078
2079         if (khugepaged_scan.mm_slot)
2080                 mm_slot = khugepaged_scan.mm_slot;
2081         else {
2082                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2083                                      struct mm_slot, mm_node);
2084                 khugepaged_scan.address = 0;
2085                 khugepaged_scan.mm_slot = mm_slot;
2086         }
2087         spin_unlock(&khugepaged_mm_lock);
2088         khugepaged_collapse_pte_mapped_thps(mm_slot);
2089
2090         mm = mm_slot->mm;
2091         /*
2092          * Don't wait for semaphore (to avoid long wait times).  Just move to
2093          * the next mm on the list.
2094          */
2095         vma = NULL;
2096         if (unlikely(!mmap_read_trylock(mm)))
2097                 goto breakouterloop_mmap_lock;
2098         if (likely(!khugepaged_test_exit(mm)))
2099                 vma = find_vma(mm, khugepaged_scan.address);
2100
2101         progress++;
2102         for (; vma; vma = vma->vm_next) {
2103                 unsigned long hstart, hend;
2104
2105                 cond_resched();
2106                 if (unlikely(khugepaged_test_exit(mm))) {
2107                         progress++;
2108                         break;
2109                 }
2110                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2111 skip:
2112                         progress++;
2113                         continue;
2114                 }
2115                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2116                 hend = vma->vm_end & HPAGE_PMD_MASK;
2117                 if (hstart >= hend)
2118                         goto skip;
2119                 if (khugepaged_scan.address > hend)
2120                         goto skip;
2121                 if (khugepaged_scan.address < hstart)
2122                         khugepaged_scan.address = hstart;
2123                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2124                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2125                         goto skip;
2126
2127                 while (khugepaged_scan.address < hend) {
2128                         int ret;
2129                         cond_resched();
2130                         if (unlikely(khugepaged_test_exit(mm)))
2131                                 goto breakouterloop;
2132
2133                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2134                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2135                                   hend);
2136                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2137                                 struct file *file = get_file(vma->vm_file);
2138                                 pgoff_t pgoff = linear_page_index(vma,
2139                                                 khugepaged_scan.address);
2140
2141                                 mmap_read_unlock(mm);
2142                                 ret = 1;
2143                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2144                                 fput(file);
2145                         } else {
2146                                 ret = khugepaged_scan_pmd(mm, vma,
2147                                                 khugepaged_scan.address,
2148                                                 hpage);
2149                         }
2150                         /* move to next address */
2151                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2152                         progress += HPAGE_PMD_NR;
2153                         if (ret)
2154                                 /* we released mmap_lock so break loop */
2155                                 goto breakouterloop_mmap_lock;
2156                         if (progress >= pages)
2157                                 goto breakouterloop;
2158                 }
2159         }
2160 breakouterloop:
2161         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2162 breakouterloop_mmap_lock:
2163
2164         spin_lock(&khugepaged_mm_lock);
2165         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2166         /*
2167          * Release the current mm_slot if this mm is about to die, or
2168          * if we scanned all vmas of this mm.
2169          */
2170         if (khugepaged_test_exit(mm) || !vma) {
2171                 /*
2172                  * Make sure that if mm_users is reaching zero while
2173                  * khugepaged runs here, khugepaged_exit will find
2174                  * mm_slot not pointing to the exiting mm.
2175                  */
2176                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2177                         khugepaged_scan.mm_slot = list_entry(
2178                                 mm_slot->mm_node.next,
2179                                 struct mm_slot, mm_node);
2180                         khugepaged_scan.address = 0;
2181                 } else {
2182                         khugepaged_scan.mm_slot = NULL;
2183                         khugepaged_full_scans++;
2184                 }
2185
2186                 collect_mm_slot(mm_slot);
2187         }
2188
2189         return progress;
2190 }
2191
2192 static int khugepaged_has_work(void)
2193 {
2194         return !list_empty(&khugepaged_scan.mm_head) &&
2195                 khugepaged_enabled();
2196 }
2197
2198 static int khugepaged_wait_event(void)
2199 {
2200         return !list_empty(&khugepaged_scan.mm_head) ||
2201                 kthread_should_stop();
2202 }
2203
2204 static void khugepaged_do_scan(void)
2205 {
2206         struct page *hpage = NULL;
2207         unsigned int progress = 0, pass_through_head = 0;
2208         unsigned int pages = khugepaged_pages_to_scan;
2209         bool wait = true;
2210
2211         barrier(); /* write khugepaged_pages_to_scan to local stack */
2212
2213         lru_add_drain_all();
2214
2215         while (progress < pages) {
2216                 if (!khugepaged_prealloc_page(&hpage, &wait))
2217                         break;
2218
2219                 cond_resched();
2220
2221                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2222                         break;
2223
2224                 spin_lock(&khugepaged_mm_lock);
2225                 if (!khugepaged_scan.mm_slot)
2226                         pass_through_head++;
2227                 if (khugepaged_has_work() &&
2228                     pass_through_head < 2)
2229                         progress += khugepaged_scan_mm_slot(pages - progress,
2230                                                             &hpage);
2231                 else
2232                         progress = pages;
2233                 spin_unlock(&khugepaged_mm_lock);
2234         }
2235
2236         if (!IS_ERR_OR_NULL(hpage))
2237                 put_page(hpage);
2238 }
2239
2240 static bool khugepaged_should_wakeup(void)
2241 {
2242         return kthread_should_stop() ||
2243                time_after_eq(jiffies, khugepaged_sleep_expire);
2244 }
2245
2246 static void khugepaged_wait_work(void)
2247 {
2248         if (khugepaged_has_work()) {
2249                 const unsigned long scan_sleep_jiffies =
2250                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2251
2252                 if (!scan_sleep_jiffies)
2253                         return;
2254
2255                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2256                 wait_event_freezable_timeout(khugepaged_wait,
2257                                              khugepaged_should_wakeup(),
2258                                              scan_sleep_jiffies);
2259                 return;
2260         }
2261
2262         if (khugepaged_enabled())
2263                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2264 }
2265
2266 static int khugepaged(void *none)
2267 {
2268         struct mm_slot *mm_slot;
2269
2270         set_freezable();
2271         set_user_nice(current, MAX_NICE);
2272
2273         while (!kthread_should_stop()) {
2274                 khugepaged_do_scan();
2275                 khugepaged_wait_work();
2276         }
2277
2278         spin_lock(&khugepaged_mm_lock);
2279         mm_slot = khugepaged_scan.mm_slot;
2280         khugepaged_scan.mm_slot = NULL;
2281         if (mm_slot)
2282                 collect_mm_slot(mm_slot);
2283         spin_unlock(&khugepaged_mm_lock);
2284         return 0;
2285 }
2286
2287 static void set_recommended_min_free_kbytes(void)
2288 {
2289         struct zone *zone;
2290         int nr_zones = 0;
2291         unsigned long recommended_min;
2292
2293         for_each_populated_zone(zone) {
2294                 /*
2295                  * We don't need to worry about fragmentation of
2296                  * ZONE_MOVABLE since it only has movable pages.
2297                  */
2298                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2299                         continue;
2300
2301                 nr_zones++;
2302         }
2303
2304         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2305         recommended_min = pageblock_nr_pages * nr_zones * 2;
2306
2307         /*
2308          * Make sure that on average at least two pageblocks are almost free
2309          * of another type, one for a migratetype to fall back to and a
2310          * second to avoid subsequent fallbacks of other types There are 3
2311          * MIGRATE_TYPES we care about.
2312          */
2313         recommended_min += pageblock_nr_pages * nr_zones *
2314                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2315
2316         /* don't ever allow to reserve more than 5% of the lowmem */
2317         recommended_min = min(recommended_min,
2318                               (unsigned long) nr_free_buffer_pages() / 20);
2319         recommended_min <<= (PAGE_SHIFT-10);
2320
2321         if (recommended_min > min_free_kbytes) {
2322                 if (user_min_free_kbytes >= 0)
2323                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2324                                 min_free_kbytes, recommended_min);
2325
2326                 min_free_kbytes = recommended_min;
2327         }
2328         setup_per_zone_wmarks();
2329 }
2330
2331 int start_stop_khugepaged(void)
2332 {
2333         int err = 0;
2334
2335         mutex_lock(&khugepaged_mutex);
2336         if (khugepaged_enabled()) {
2337                 if (!khugepaged_thread)
2338                         khugepaged_thread = kthread_run(khugepaged, NULL,
2339                                                         "khugepaged");
2340                 if (IS_ERR(khugepaged_thread)) {
2341                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2342                         err = PTR_ERR(khugepaged_thread);
2343                         khugepaged_thread = NULL;
2344                         goto fail;
2345                 }
2346
2347                 if (!list_empty(&khugepaged_scan.mm_head))
2348                         wake_up_interruptible(&khugepaged_wait);
2349
2350                 set_recommended_min_free_kbytes();
2351         } else if (khugepaged_thread) {
2352                 kthread_stop(khugepaged_thread);
2353                 khugepaged_thread = NULL;
2354         }
2355 fail:
2356         mutex_unlock(&khugepaged_mutex);
2357         return err;
2358 }
2359
2360 void khugepaged_min_free_kbytes_update(void)
2361 {
2362         mutex_lock(&khugepaged_mutex);
2363         if (khugepaged_enabled() && khugepaged_thread)
2364                 set_recommended_min_free_kbytes();
2365         mutex_unlock(&khugepaged_mutex);
2366 }