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