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 /* default scan 8*512 pte (or vmas) every 30 second */
  60 static unsigned int khugepaged_pages_to_scan __read_mostly;
  61 static unsigned int khugepaged_pages_collapsed;
  62 static unsigned int khugepaged_full_scans;
  63 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  64 /* during fragmentation poll the hugepage allocator once every minute */
  65 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  66 static unsigned long khugepaged_sleep_expire;
  67 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  68 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  69 /*
  70  * default collapse hugepages if there is at least one pte mapped like
  71  * it would have happened if the vma was large enough during page
  72  * fault.
  73  */
  74 static unsigned int khugepaged_max_ptes_none __read_mostly;
  75 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  76 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  77
  78 #define MM_SLOTS_HASH_BITS 10
  79 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  80
  81 static struct kmem_cache *mm_slot_cache __read_mostly;
  82
  83 #define MAX_PTE_MAPPED_THP 8
  84
  85 /**
  86  * struct mm_slot - hash lookup from mm to mm_slot
  87  * @hash: hash collision list
  88  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  89  * @mm: the mm that this information is valid for
  90  */
  91 struct mm_slot {
  92         struct hlist_node hash;
  93         struct list_head mm_node;
  94         struct mm_struct *mm;
  95
  96         /* pte-mapped THP in this mm */
  97         int nr_pte_mapped_thp;
  98         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
  99 };
 100
 101 /**
 102  * struct khugepaged_scan - cursor for scanning
 103  * @mm_head: the head of the mm list to scan
 104  * @mm_slot: the current mm_slot we are scanning
 105  * @address: the next address inside that to be scanned
 106  *
 107  * There is only the one khugepaged_scan instance of this cursor structure.
 108  */
 109 struct khugepaged_scan {
 110         struct list_head mm_head;
 111         struct mm_slot *mm_slot;
 112         unsigned long address;
 113 };
 114
 115 static struct khugepaged_scan khugepaged_scan = {
 116         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 117 };
 118
 119 #ifdef CONFIG_SYSFS
 120 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 121                                          struct kobj_attribute *attr,
 122                                          char *buf)
 123 {
 124         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 125 }
 126
 127 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 128                                           struct kobj_attribute *attr,
 129                                           const char *buf, size_t count)
 130 {
 131         unsigned long msecs;
 132         int err;
 133
 134         err = kstrtoul(buf, 10, &msecs);
 135         if (err || msecs > UINT_MAX)
 136                 return -EINVAL;
 137
 138         khugepaged_scan_sleep_millisecs = msecs;
 139         khugepaged_sleep_expire = 0;
 140         wake_up_interruptible(&khugepaged_wait);
 141
 142         return count;
 143 }
 144 static struct kobj_attribute scan_sleep_millisecs_attr =
 145         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 146                scan_sleep_millisecs_store);
 147
 148 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 149                                           struct kobj_attribute *attr,
 150                                           char *buf)
 151 {
 152         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 153 }
 154
 155 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 156                                            struct kobj_attribute *attr,
 157                                            const char *buf, size_t count)
 158 {
 159         unsigned long msecs;
 160         int err;
 161
 162         err = kstrtoul(buf, 10, &msecs);
 163         if (err || msecs > UINT_MAX)
 164                 return -EINVAL;
 165
 166         khugepaged_alloc_sleep_millisecs = msecs;
 167         khugepaged_sleep_expire = 0;
 168         wake_up_interruptible(&khugepaged_wait);
 169
 170         return count;
 171 }
 172 static struct kobj_attribute alloc_sleep_millisecs_attr =
 173         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 174                alloc_sleep_millisecs_store);
 175
 176 static ssize_t pages_to_scan_show(struct kobject *kobj,
 177                                   struct kobj_attribute *attr,
 178                                   char *buf)
 179 {
 180         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 181 }
 182 static ssize_t pages_to_scan_store(struct kobject *kobj,
 183                                    struct kobj_attribute *attr,
 184                                    const char *buf, size_t count)
 185 {
 186         int err;
 187         unsigned long pages;
 188
 189         err = kstrtoul(buf, 10, &pages);
 190         if (err || !pages || pages > UINT_MAX)
 191                 return -EINVAL;
 192
 193         khugepaged_pages_to_scan = pages;
 194
 195         return count;
 196 }
 197 static struct kobj_attribute pages_to_scan_attr =
 198         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 199                pages_to_scan_store);
 200
 201 static ssize_t pages_collapsed_show(struct kobject *kobj,
 202                                     struct kobj_attribute *attr,
 203                                     char *buf)
 204 {
 205         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 206 }
 207 static struct kobj_attribute pages_collapsed_attr =
 208         __ATTR_RO(pages_collapsed);
 209
 210 static ssize_t full_scans_show(struct kobject *kobj,
 211                                struct kobj_attribute *attr,
 212                                char *buf)
 213 {
 214         return sprintf(buf, "%u\n", khugepaged_full_scans);
 215 }
 216 static struct kobj_attribute full_scans_attr =
 217         __ATTR_RO(full_scans);
 218
 219 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 220                                       struct kobj_attribute *attr, char *buf)
 221 {
 222         return single_hugepage_flag_show(kobj, attr, buf,
 223                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 224 }
 225 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 226                                        struct kobj_attribute *attr,
 227                                        const char *buf, size_t count)
 228 {
 229         return single_hugepage_flag_store(kobj, attr, buf, count,
 230                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 231 }
 232 static struct kobj_attribute khugepaged_defrag_attr =
 233         __ATTR(defrag, 0644, khugepaged_defrag_show,
 234                khugepaged_defrag_store);
 235
 236 /*
 237  * max_ptes_none controls if khugepaged should collapse hugepages over
 238  * any unmapped ptes in turn potentially increasing the memory
 239  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 240  * reduce the available free memory in the system as it
 241  * runs. Increasing max_ptes_none will instead potentially reduce the
 242  * free memory in the system during the khugepaged scan.
 243  */
 244 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 245                                              struct kobj_attribute *attr,
 246                                              char *buf)
 247 {
 248         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 249 }
 250 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 251                                               struct kobj_attribute *attr,
 252                                               const char *buf, size_t count)
 253 {
 254         int err;
 255         unsigned long max_ptes_none;
 256
 257         err = kstrtoul(buf, 10, &max_ptes_none);
 258         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 259                 return -EINVAL;
 260
 261         khugepaged_max_ptes_none = max_ptes_none;
 262
 263         return count;
 264 }
 265 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 266         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 267                khugepaged_max_ptes_none_store);
 268
 269 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 270                                              struct kobj_attribute *attr,
 271                                              char *buf)
 272 {
 273         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 274 }
 275
 276 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 277                                               struct kobj_attribute *attr,
 278                                               const char *buf, size_t count)
 279 {
 280         int err;
 281         unsigned long max_ptes_swap;
 282
 283         err  = kstrtoul(buf, 10, &max_ptes_swap);
 284         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 285                 return -EINVAL;
 286
 287         khugepaged_max_ptes_swap = max_ptes_swap;
 288
 289         return count;
 290 }
 291
 292 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 293         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 294                khugepaged_max_ptes_swap_store);
 295
 296 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 297                                              struct kobj_attribute *attr,
 298                                              char *buf)
 299 {
 300         return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
 301 }
 302
 303 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 304                                               struct kobj_attribute *attr,
 305                                               const char *buf, size_t count)
 306 {
 307         int err;
 308         unsigned long max_ptes_shared;
 309
 310         err  = kstrtoul(buf, 10, &max_ptes_shared);
 311         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 312                 return -EINVAL;
 313
 314         khugepaged_max_ptes_shared = max_ptes_shared;
 315
 316         return count;
 317 }
 318
 319 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 320         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 321                khugepaged_max_ptes_shared_store);
 322
 323 static struct attribute *khugepaged_attr[] = {
 324         &khugepaged_defrag_attr.attr,
 325         &khugepaged_max_ptes_none_attr.attr,
 326         &khugepaged_max_ptes_swap_attr.attr,
 327         &khugepaged_max_ptes_shared_attr.attr,
 328         &pages_to_scan_attr.attr,
 329         &pages_collapsed_attr.attr,
 330         &full_scans_attr.attr,
 331         &scan_sleep_millisecs_attr.attr,
 332         &alloc_sleep_millisecs_attr.attr,
 333         NULL,
 334 };
 335
 336 struct attribute_group khugepaged_attr_group = {
 337         .attrs = khugepaged_attr,
 338         .name = "khugepaged",
 339 };
 340 #endif /* CONFIG_SYSFS */
 341
 342 int hugepage_madvise(struct vm_area_struct *vma,
 343                      unsigned long *vm_flags, int advice)
 344 {
 345         switch (advice) {
 346         case MADV_HUGEPAGE:
 347 #ifdef CONFIG_S390
 348                 /*
 349                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 350                  * can't handle this properly after s390_enable_sie, so we simply
 351                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 352                  */
 353                 if (mm_has_pgste(vma->vm_mm))
 354                         return 0;
 355 #endif
 356                 *vm_flags &= ~VM_NOHUGEPAGE;
 357                 *vm_flags |= VM_HUGEPAGE;
 358                 /*
 359                  * If the vma become good for khugepaged to scan,
 360                  * register it here without waiting a page fault that
 361                  * may not happen any time soon.
 362                  */
 363                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 364                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 365                         return -ENOMEM;
 366                 break;
 367         case MADV_NOHUGEPAGE:
 368                 *vm_flags &= ~VM_HUGEPAGE;
 369                 *vm_flags |= VM_NOHUGEPAGE;
 370                 /*
 371                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 372                  * this vma even if we leave the mm registered in khugepaged if
 373                  * it got registered before VM_NOHUGEPAGE was set.
 374                  */
 375                 break;
 376         }
 377
 378         return 0;
 379 }
 380
 381 int __init khugepaged_init(void)
 382 {
 383         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 384                                           sizeof(struct mm_slot),
 385                                           __alignof__(struct mm_slot), 0, NULL);
 386         if (!mm_slot_cache)
 387                 return -ENOMEM;
 388
 389         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 390         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 391         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 392         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 393
 394         return 0;
 395 }
 396
 397 void __init khugepaged_destroy(void)
 398 {
 399         kmem_cache_destroy(mm_slot_cache);
 400 }
 401
 402 static inline struct mm_slot *alloc_mm_slot(void)
 403 {
 404         if (!mm_slot_cache)     /* initialization failed */
 405                 return NULL;
 406         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 407 }
 408
 409 static inline void free_mm_slot(struct mm_slot *mm_slot)
 410 {
 411         kmem_cache_free(mm_slot_cache, mm_slot);
 412 }
 413
 414 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 415 {
 416         struct mm_slot *mm_slot;
 417
 418         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 419                 if (mm == mm_slot->mm)
 420                         return mm_slot;
 421
 422         return NULL;
 423 }
 424
 425 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 426                                     struct mm_slot *mm_slot)
 427 {
 428         mm_slot->mm = mm;
 429         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 430 }
 431
 432 static inline int khugepaged_test_exit(struct mm_struct *mm)
 433 {
 434         return atomic_read(&mm->mm_users) == 0;
 435 }
 436
 437 static bool hugepage_vma_check(struct vm_area_struct *vma,
 438                                unsigned long vm_flags)
 439 {
 440         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 441             (vm_flags & VM_NOHUGEPAGE) ||
 442             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 443                 return false;
 444
 445         if (shmem_file(vma->vm_file) ||
 446             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 447              vma->vm_file &&
 448              (vm_flags & VM_DENYWRITE))) {
 449                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 450                                 HPAGE_PMD_NR);
 451         }
 452         if (!vma->anon_vma || vma->vm_ops)
 453                 return false;
 454         if (vma_is_temporary_stack(vma))
 455                 return false;
 456         return !(vm_flags & VM_NO_KHUGEPAGED);
 457 }
 458
 459 int __khugepaged_enter(struct mm_struct *mm)
 460 {
 461         struct mm_slot *mm_slot;
 462         int wakeup;
 463
 464         mm_slot = alloc_mm_slot();
 465         if (!mm_slot)
 466                 return -ENOMEM;
 467
 468         /* __khugepaged_exit() must not run from under us */
 469         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 470         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 471                 free_mm_slot(mm_slot);
 472                 return 0;
 473         }
 474
 475         spin_lock(&khugepaged_mm_lock);
 476         insert_to_mm_slots_hash(mm, mm_slot);
 477         /*
 478          * Insert just behind the scanning cursor, to let the area settle
 479          * down a little.
 480          */
 481         wakeup = list_empty(&khugepaged_scan.mm_head);
 482         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 483         spin_unlock(&khugepaged_mm_lock);
 484
 485         mmgrab(mm);
 486         if (wakeup)
 487                 wake_up_interruptible(&khugepaged_wait);
 488
 489         return 0;
 490 }
 491
 492 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 493                                unsigned long vm_flags)
 494 {
 495         unsigned long hstart, hend;
 496
 497         /*
 498          * khugepaged only supports read-only files for non-shmem files.
 499          * khugepaged does not yet work on special mappings. And
 500          * file-private shmem THP is not supported.
 501          */
 502         if (!hugepage_vma_check(vma, vm_flags))
 503                 return 0;
 504
 505         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 506         hend = vma->vm_end & HPAGE_PMD_MASK;
 507         if (hstart < hend)
 508                 return khugepaged_enter(vma, vm_flags);
 509         return 0;
 510 }
 511
 512 void __khugepaged_exit(struct mm_struct *mm)
 513 {
 514         struct mm_slot *mm_slot;
 515         int free = 0;
 516
 517         spin_lock(&khugepaged_mm_lock);
 518         mm_slot = get_mm_slot(mm);
 519         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 520                 hash_del(&mm_slot->hash);
 521                 list_del(&mm_slot->mm_node);
 522                 free = 1;
 523         }
 524         spin_unlock(&khugepaged_mm_lock);
 525
 526         if (free) {
 527                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 528                 free_mm_slot(mm_slot);
 529                 mmdrop(mm);
 530         } else if (mm_slot) {
 531                 /*
 532                  * This is required to serialize against
 533                  * khugepaged_test_exit() (which is guaranteed to run
 534                  * under mmap sem read mode). Stop here (after we
 535                  * return all pagetables will be destroyed) until
 536                  * khugepaged has finished working on the pagetables
 537                  * under the mmap_lock.
 538                  */
 539                 mmap_write_lock(mm);
 540                 mmap_write_unlock(mm);
 541         }
 542 }
 543
 544 static void release_pte_page(struct page *page)
 545 {
 546         mod_node_page_state(page_pgdat(page),
 547                         NR_ISOLATED_ANON + page_is_file_lru(page),
 548                         -compound_nr(page));
 549         unlock_page(page);
 550         putback_lru_page(page);
 551 }
 552
 553 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 554                 struct list_head *compound_pagelist)
 555 {
 556         struct page *page, *tmp;
 557
 558         while (--_pte >= pte) {
 559                 pte_t pteval = *_pte;
 560
 561                 page = pte_page(pteval);
 562                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 563                                 !PageCompound(page))
 564                         release_pte_page(page);
 565         }
 566
 567         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 568                 list_del(&page->lru);
 569                 release_pte_page(page);
 570         }
 571 }
 572
 573 static bool is_refcount_suitable(struct page *page)
 574 {
 575         int expected_refcount;
 576
 577         expected_refcount = total_mapcount(page);
 578         if (PageSwapCache(page))
 579                 expected_refcount += compound_nr(page);
 580
 581         return page_count(page) == expected_refcount;
 582 }
 583
 584 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 585                                         unsigned long address,
 586                                         pte_t *pte,
 587                                         struct list_head *compound_pagelist)
 588 {
 589         struct page *page = NULL;
 590         pte_t *_pte;
 591         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 592         bool writable = false;
 593
 594         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 595              _pte++, address += PAGE_SIZE) {
 596                 pte_t pteval = *_pte;
 597                 if (pte_none(pteval) || (pte_present(pteval) &&
 598                                 is_zero_pfn(pte_pfn(pteval)))) {
 599                         if (!userfaultfd_armed(vma) &&
 600                             ++none_or_zero <= khugepaged_max_ptes_none) {
 601                                 continue;
 602                         } else {
 603                                 result = SCAN_EXCEED_NONE_PTE;
 604                                 goto out;
 605                         }
 606                 }
 607                 if (!pte_present(pteval)) {
 608                         result = SCAN_PTE_NON_PRESENT;
 609                         goto out;
 610                 }
 611                 page = vm_normal_page(vma, address, pteval);
 612                 if (unlikely(!page)) {
 613                         result = SCAN_PAGE_NULL;
 614                         goto out;
 615                 }
 616
 617                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 618
 619                 if (page_mapcount(page) > 1 &&
 620                                 ++shared > khugepaged_max_ptes_shared) {
 621                         result = SCAN_EXCEED_SHARED_PTE;
 622                         goto out;
 623                 }
 624
 625                 if (PageCompound(page)) {
 626                         struct page *p;
 627                         page = compound_head(page);
 628
 629                         /*
 630                          * Check if we have dealt with the compound page
 631                          * already
 632                          */
 633                         list_for_each_entry(p, compound_pagelist, lru) {
 634                                 if (page == p)
 635                                         goto next;
 636                         }
 637                 }
 638
 639                 /*
 640                  * We can do it before isolate_lru_page because the
 641                  * page can't be freed from under us. NOTE: PG_lock
 642                  * is needed to serialize against split_huge_page
 643                  * when invoked from the VM.
 644                  */
 645                 if (!trylock_page(page)) {
 646                         result = SCAN_PAGE_LOCK;
 647                         goto out;
 648                 }
 649
 650                 /*
 651                  * Check if the page has any GUP (or other external) pins.
 652                  *
 653                  * The page table that maps the page has been already unlinked
 654                  * from the page table tree and this process cannot get
 655                  * an additinal pin on the page.
 656                  *
 657                  * New pins can come later if the page is shared across fork,
 658                  * but not from this process. The other process cannot write to
 659                  * the page, only trigger CoW.
 660                  */
 661                 if (!is_refcount_suitable(page)) {
 662                         unlock_page(page);
 663                         result = SCAN_PAGE_COUNT;
 664                         goto out;
 665                 }
 666                 if (!pte_write(pteval) && PageSwapCache(page) &&
 667                                 !reuse_swap_page(page, NULL)) {
 668                         /*
 669                          * Page is in the swap cache and cannot be re-used.
 670                          * It cannot be collapsed into a THP.
 671                          */
 672                         unlock_page(page);
 673                         result = SCAN_SWAP_CACHE_PAGE;
 674                         goto out;
 675                 }
 676
 677                 /*
 678                  * Isolate the page to avoid collapsing an hugepage
 679                  * currently in use by the VM.
 680                  */
 681                 if (isolate_lru_page(page)) {
 682                         unlock_page(page);
 683                         result = SCAN_DEL_PAGE_LRU;
 684                         goto out;
 685                 }
 686                 mod_node_page_state(page_pgdat(page),
 687                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 688                                 compound_nr(page));
 689                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 690                 VM_BUG_ON_PAGE(PageLRU(page), page);
 691
 692                 if (PageCompound(page))
 693                         list_add_tail(&page->lru, compound_pagelist);
 694 next:
 695                 /* There should be enough young pte to collapse the page */
 696                 if (pte_young(pteval) ||
 697                     page_is_young(page) || PageReferenced(page) ||
 698                     mmu_notifier_test_young(vma->vm_mm, address))
 699                         referenced++;
 700
 701                 if (pte_write(pteval))
 702                         writable = true;
 703         }
 704         if (likely(writable)) {
 705                 if (likely(referenced)) {
 706                         result = SCAN_SUCCEED;
 707                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 708                                                             referenced, writable, result);
 709                         return 1;
 710                 }
 711         } else {
 712                 result = SCAN_PAGE_RO;
 713         }
 714
 715 out:
 716         release_pte_pages(pte, _pte, compound_pagelist);
 717         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 718                                             referenced, writable, result);
 719         return 0;
 720 }
 721
 722 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 723                                       struct vm_area_struct *vma,
 724                                       unsigned long address,
 725                                       spinlock_t *ptl,
 726                                       struct list_head *compound_pagelist)
 727 {
 728         struct page *src_page, *tmp;
 729         pte_t *_pte;
 730         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 731                                 _pte++, page++, address += PAGE_SIZE) {
 732                 pte_t pteval = *_pte;
 733
 734                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 735                         clear_user_highpage(page, address);
 736                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 737                         if (is_zero_pfn(pte_pfn(pteval))) {
 738                                 /*
 739                                  * ptl mostly unnecessary.
 740                                  */
 741                                 spin_lock(ptl);
 742                                 /*
 743                                  * paravirt calls inside pte_clear here are
 744                                  * superfluous.
 745                                  */
 746                                 pte_clear(vma->vm_mm, address, _pte);
 747                                 spin_unlock(ptl);
 748                         }
 749                 } else {
 750                         src_page = pte_page(pteval);
 751                         copy_user_highpage(page, src_page, address, vma);
 752                         if (!PageCompound(src_page))
 753                                 release_pte_page(src_page);
 754                         /*
 755                          * ptl mostly unnecessary, but preempt has to
 756                          * be disabled to update the per-cpu stats
 757                          * inside page_remove_rmap().
 758                          */
 759                         spin_lock(ptl);
 760                         /*
 761                          * paravirt calls inside pte_clear here are
 762                          * superfluous.
 763                          */
 764                         pte_clear(vma->vm_mm, address, _pte);
 765                         page_remove_rmap(src_page, false);
 766                         spin_unlock(ptl);
 767                         free_page_and_swap_cache(src_page);
 768                 }
 769         }
 770
 771         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 772                 list_del(&src_page->lru);
 773                 release_pte_page(src_page);
 774         }
 775 }
 776
 777 static void khugepaged_alloc_sleep(void)
 778 {
 779         DEFINE_WAIT(wait);
 780
 781         add_wait_queue(&khugepaged_wait, &wait);
 782         freezable_schedule_timeout_interruptible(
 783                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 784         remove_wait_queue(&khugepaged_wait, &wait);
 785 }
 786
 787 static int khugepaged_node_load[MAX_NUMNODES];
 788
 789 static bool khugepaged_scan_abort(int nid)
 790 {
 791         int i;
 792
 793         /*
 794          * If node_reclaim_mode is disabled, then no extra effort is made to
 795          * allocate memory locally.
 796          */
 797         if (!node_reclaim_mode)
 798                 return false;
 799
 800         /* If there is a count for this node already, it must be acceptable */
 801         if (khugepaged_node_load[nid])
 802                 return false;
 803
 804         for (i = 0; i < MAX_NUMNODES; i++) {
 805                 if (!khugepaged_node_load[i])
 806                         continue;
 807                 if (node_distance(nid, i) > node_reclaim_distance)
 808                         return true;
 809         }
 810         return false;
 811 }
 812
 813 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 814 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 815 {
 816         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 817 }
 818
 819 #ifdef CONFIG_NUMA
 820 static int khugepaged_find_target_node(void)
 821 {
 822         static int last_khugepaged_target_node = NUMA_NO_NODE;
 823         int nid, target_node = 0, max_value = 0;
 824
 825         /* find first node with max normal pages hit */
 826         for (nid = 0; nid < MAX_NUMNODES; nid++)
 827                 if (khugepaged_node_load[nid] > max_value) {
 828                         max_value = khugepaged_node_load[nid];
 829                         target_node = nid;
 830                 }
 831
 832         /* do some balance if several nodes have the same hit record */
 833         if (target_node <= last_khugepaged_target_node)
 834                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 835                                 nid++)
 836                         if (max_value == khugepaged_node_load[nid]) {
 837                                 target_node = nid;
 838                                 break;
 839                         }
 840
 841         last_khugepaged_target_node = target_node;
 842         return target_node;
 843 }
 844
 845 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 846 {
 847         if (IS_ERR(*hpage)) {
 848                 if (!*wait)
 849                         return false;
 850
 851                 *wait = false;
 852                 *hpage = NULL;
 853                 khugepaged_alloc_sleep();
 854         } else if (*hpage) {
 855                 put_page(*hpage);
 856                 *hpage = NULL;
 857         }
 858
 859         return true;
 860 }
 861
 862 static struct page *
 863 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 864 {
 865         VM_BUG_ON_PAGE(*hpage, *hpage);
 866
 867         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 868         if (unlikely(!*hpage)) {
 869                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 870                 *hpage = ERR_PTR(-ENOMEM);
 871                 return NULL;
 872         }
 873
 874         prep_transhuge_page(*hpage);
 875         count_vm_event(THP_COLLAPSE_ALLOC);
 876         return *hpage;
 877 }
 878 #else
 879 static int khugepaged_find_target_node(void)
 880 {
 881         return 0;
 882 }
 883
 884 static inline struct page *alloc_khugepaged_hugepage(void)
 885 {
 886         struct page *page;
 887
 888         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 889                            HPAGE_PMD_ORDER);
 890         if (page)
 891                 prep_transhuge_page(page);
 892         return page;
 893 }
 894
 895 static struct page *khugepaged_alloc_hugepage(bool *wait)
 896 {
 897         struct page *hpage;
 898
 899         do {
 900                 hpage = alloc_khugepaged_hugepage();
 901                 if (!hpage) {
 902                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 903                         if (!*wait)
 904                                 return NULL;
 905
 906                         *wait = false;
 907                         khugepaged_alloc_sleep();
 908                 } else
 909                         count_vm_event(THP_COLLAPSE_ALLOC);
 910         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 911
 912         return hpage;
 913 }
 914
 915 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 916 {
 917         if (!*hpage)
 918                 *hpage = khugepaged_alloc_hugepage(wait);
 919
 920         if (unlikely(!*hpage))
 921                 return false;
 922
 923         return true;
 924 }
 925
 926 static struct page *
 927 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 928 {
 929         VM_BUG_ON(!*hpage);
 930
 931         return  *hpage;
 932 }
 933 #endif
 934
 935 /*
 936  * If mmap_lock temporarily dropped, revalidate vma
 937  * before taking mmap_lock.
 938  * Return 0 if succeeds, otherwise return none-zero
 939  * value (scan code).
 940  */
 941
 942 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 943                 struct vm_area_struct **vmap)
 944 {
 945         struct vm_area_struct *vma;
 946         unsigned long hstart, hend;
 947
 948         if (unlikely(khugepaged_test_exit(mm)))
 949                 return SCAN_ANY_PROCESS;
 950
 951         *vmap = vma = find_vma(mm, address);
 952         if (!vma)
 953                 return SCAN_VMA_NULL;
 954
 955         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 956         hend = vma->vm_end & HPAGE_PMD_MASK;
 957         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 958                 return SCAN_ADDRESS_RANGE;
 959         if (!hugepage_vma_check(vma, vma->vm_flags))
 960                 return SCAN_VMA_CHECK;
 961         return 0;
 962 }
 963
 964 /*
 965  * Bring missing pages in from swap, to complete THP collapse.
 966  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 967  *
 968  * Called and returns without pte mapped or spinlocks held,
 969  * but with mmap_lock held to protect against vma changes.
 970  */
 971
 972 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 973                                         struct vm_area_struct *vma,
 974                                         unsigned long address, pmd_t *pmd,
 975                                         int referenced)
 976 {
 977         int swapped_in = 0;
 978         vm_fault_t ret = 0;
 979         struct vm_fault vmf = {
 980                 .vma = vma,
 981                 .address = address,
 982                 .flags = FAULT_FLAG_ALLOW_RETRY,
 983                 .pmd = pmd,
 984                 .pgoff = linear_page_index(vma, address),
 985         };
 986
 987         vmf.pte = pte_offset_map(pmd, address);
 988         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 989                         vmf.pte++, vmf.address += PAGE_SIZE) {
 990                 vmf.orig_pte = *vmf.pte;
 991                 if (!is_swap_pte(vmf.orig_pte))
 992                         continue;
 993                 swapped_in++;
 994                 ret = do_swap_page(&vmf);
 995
 996                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
 997                 if (ret & VM_FAULT_RETRY) {
 998                         mmap_read_lock(mm);
 999                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1000                                 /* vma is no longer available, don't continue to swapin */
1001                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1002                                 return false;
1003                         }
1004                         /* check if the pmd is still valid */
1005                         if (mm_find_pmd(mm, address) != pmd) {
1006                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1007                                 return false;
1008                         }
1009                 }
1010                 if (ret & VM_FAULT_ERROR) {
1011                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1012                         return false;
1013                 }
1014                 /* pte is unmapped now, we need to map it */
1015                 vmf.pte = pte_offset_map(pmd, vmf.address);
1016         }
1017         vmf.pte--;
1018         pte_unmap(vmf.pte);
1019
1020         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1021         if (swapped_in)
1022                 lru_add_drain();
1023
1024         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1025         return true;
1026 }
1027
1028 static void collapse_huge_page(struct mm_struct *mm,
1029                                    unsigned long address,
1030                                    struct page **hpage,
1031                                    int node, int referenced, int unmapped)
1032 {
1033         LIST_HEAD(compound_pagelist);
1034         pmd_t *pmd, _pmd;
1035         pte_t *pte;
1036         pgtable_t pgtable;
1037         struct page *new_page;
1038         spinlock_t *pmd_ptl, *pte_ptl;
1039         int isolated = 0, result = 0;
1040         struct vm_area_struct *vma;
1041         struct mmu_notifier_range range;
1042         gfp_t gfp;
1043
1044         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1045
1046         /* Only allocate from the target node */
1047         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1048
1049         /*
1050          * Before allocating the hugepage, release the mmap_lock read lock.
1051          * The allocation can take potentially a long time if it involves
1052          * sync compaction, and we do not need to hold the mmap_lock during
1053          * that. We will recheck the vma after taking it again in write mode.
1054          */
1055         mmap_read_unlock(mm);
1056         new_page = khugepaged_alloc_page(hpage, gfp, node);
1057         if (!new_page) {
1058                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1059                 goto out_nolock;
1060         }
1061
1062         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1063                 result = SCAN_CGROUP_CHARGE_FAIL;
1064                 goto out_nolock;
1065         }
1066         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1067
1068         mmap_read_lock(mm);
1069         result = hugepage_vma_revalidate(mm, address, &vma);
1070         if (result) {
1071                 mmap_read_unlock(mm);
1072                 goto out_nolock;
1073         }
1074
1075         pmd = mm_find_pmd(mm, address);
1076         if (!pmd) {
1077                 result = SCAN_PMD_NULL;
1078                 mmap_read_unlock(mm);
1079                 goto out_nolock;
1080         }
1081
1082         /*
1083          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1084          * If it fails, we release mmap_lock and jump out_nolock.
1085          * Continuing to collapse causes inconsistency.
1086          */
1087         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1088                                                      pmd, referenced)) {
1089                 mmap_read_unlock(mm);
1090                 goto out_nolock;
1091         }
1092
1093         mmap_read_unlock(mm);
1094         /*
1095          * Prevent all access to pagetables with the exception of
1096          * gup_fast later handled by the ptep_clear_flush and the VM
1097          * handled by the anon_vma lock + PG_lock.
1098          */
1099         mmap_write_lock(mm);
1100         result = SCAN_ANY_PROCESS;
1101         if (!mmget_still_valid(mm))
1102                 goto out;
1103         result = hugepage_vma_revalidate(mm, address, &vma);
1104         if (result)
1105                 goto out;
1106         /* check if the pmd is still valid */
1107         if (mm_find_pmd(mm, address) != pmd)
1108                 goto out;
1109
1110         anon_vma_lock_write(vma->anon_vma);
1111
1112         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1113                                 address, address + HPAGE_PMD_SIZE);
1114         mmu_notifier_invalidate_range_start(&range);
1115
1116         pte = pte_offset_map(pmd, address);
1117         pte_ptl = pte_lockptr(mm, pmd);
1118
1119         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1120         /*
1121          * After this gup_fast can't run anymore. This also removes
1122          * any huge TLB entry from the CPU so we won't allow
1123          * huge and small TLB entries for the same virtual address
1124          * to avoid the risk of CPU bugs in that area.
1125          */
1126         _pmd = pmdp_collapse_flush(vma, address, pmd);
1127         spin_unlock(pmd_ptl);
1128         mmu_notifier_invalidate_range_end(&range);
1129
1130         spin_lock(pte_ptl);
1131         isolated = __collapse_huge_page_isolate(vma, address, pte,
1132                         &compound_pagelist);
1133         spin_unlock(pte_ptl);
1134
1135         if (unlikely(!isolated)) {
1136                 pte_unmap(pte);
1137                 spin_lock(pmd_ptl);
1138                 BUG_ON(!pmd_none(*pmd));
1139                 /*
1140                  * We can only use set_pmd_at when establishing
1141                  * hugepmds and never for establishing regular pmds that
1142                  * points to regular pagetables. Use pmd_populate for that
1143                  */
1144                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1145                 spin_unlock(pmd_ptl);
1146                 anon_vma_unlock_write(vma->anon_vma);
1147                 result = SCAN_FAIL;
1148                 goto out;
1149         }
1150
1151         /*
1152          * All pages are isolated and locked so anon_vma rmap
1153          * can't run anymore.
1154          */
1155         anon_vma_unlock_write(vma->anon_vma);
1156
1157         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1158                         &compound_pagelist);
1159         pte_unmap(pte);
1160         __SetPageUptodate(new_page);
1161         pgtable = pmd_pgtable(_pmd);
1162
1163         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1164         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1165
1166         /*
1167          * spin_lock() below is not the equivalent of smp_wmb(), so
1168          * this is needed to avoid the copy_huge_page writes to become
1169          * visible after the set_pmd_at() write.
1170          */
1171         smp_wmb();
1172
1173         spin_lock(pmd_ptl);
1174         BUG_ON(!pmd_none(*pmd));
1175         page_add_new_anon_rmap(new_page, vma, address, true);
1176         lru_cache_add_active_or_unevictable(new_page, vma);
1177         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1178         set_pmd_at(mm, address, pmd, _pmd);
1179         update_mmu_cache_pmd(vma, address, pmd);
1180         spin_unlock(pmd_ptl);
1181
1182         *hpage = NULL;
1183
1184         khugepaged_pages_collapsed++;
1185         result = SCAN_SUCCEED;
1186 out_up_write:
1187         mmap_write_unlock(mm);
1188 out_nolock:
1189         if (!IS_ERR_OR_NULL(*hpage))
1190                 mem_cgroup_uncharge(*hpage);
1191         trace_mm_collapse_huge_page(mm, isolated, result);
1192         return;
1193 out:
1194         goto out_up_write;
1195 }
1196
1197 static int khugepaged_scan_pmd(struct mm_struct *mm,
1198                                struct vm_area_struct *vma,
1199                                unsigned long address,
1200                                struct page **hpage)
1201 {
1202         pmd_t *pmd;
1203         pte_t *pte, *_pte;
1204         int ret = 0, result = 0, referenced = 0;
1205         int none_or_zero = 0, shared = 0;
1206         struct page *page = NULL;
1207         unsigned long _address;
1208         spinlock_t *ptl;
1209         int node = NUMA_NO_NODE, unmapped = 0;
1210         bool writable = false;
1211
1212         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1213
1214         pmd = mm_find_pmd(mm, address);
1215         if (!pmd) {
1216                 result = SCAN_PMD_NULL;
1217                 goto out;
1218         }
1219
1220         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1221         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1222         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1223              _pte++, _address += PAGE_SIZE) {
1224                 pte_t pteval = *_pte;
1225                 if (is_swap_pte(pteval)) {
1226                         if (++unmapped <= khugepaged_max_ptes_swap) {
1227                                 /*
1228                                  * Always be strict with uffd-wp
1229                                  * enabled swap entries.  Please see
1230                                  * comment below for pte_uffd_wp().
1231                                  */
1232                                 if (pte_swp_uffd_wp(pteval)) {
1233                                         result = SCAN_PTE_UFFD_WP;
1234                                         goto out_unmap;
1235                                 }
1236                                 continue;
1237                         } else {
1238                                 result = SCAN_EXCEED_SWAP_PTE;
1239                                 goto out_unmap;
1240                         }
1241                 }
1242                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1243                         if (!userfaultfd_armed(vma) &&
1244                             ++none_or_zero <= khugepaged_max_ptes_none) {
1245                                 continue;
1246                         } else {
1247                                 result = SCAN_EXCEED_NONE_PTE;
1248                                 goto out_unmap;
1249                         }
1250                 }
1251                 if (!pte_present(pteval)) {
1252                         result = SCAN_PTE_NON_PRESENT;
1253                         goto out_unmap;
1254                 }
1255                 if (pte_uffd_wp(pteval)) {
1256                         /*
1257                          * Don't collapse the page if any of the small
1258                          * PTEs are armed with uffd write protection.
1259                          * Here we can also mark the new huge pmd as
1260                          * write protected if any of the small ones is
1261                          * marked but that could bring uknown
1262                          * userfault messages that falls outside of
1263                          * the registered range.  So, just be simple.
1264                          */
1265                         result = SCAN_PTE_UFFD_WP;
1266                         goto out_unmap;
1267                 }
1268                 if (pte_write(pteval))
1269                         writable = true;
1270
1271                 page = vm_normal_page(vma, _address, pteval);
1272                 if (unlikely(!page)) {
1273                         result = SCAN_PAGE_NULL;
1274                         goto out_unmap;
1275                 }
1276
1277                 if (page_mapcount(page) > 1 &&
1278                                 ++shared > khugepaged_max_ptes_shared) {
1279                         result = SCAN_EXCEED_SHARED_PTE;
1280                         goto out_unmap;
1281                 }
1282
1283                 page = compound_head(page);
1284
1285                 /*
1286                  * Record which node the original page is from and save this
1287                  * information to khugepaged_node_load[].
1288                  * Khupaged will allocate hugepage from the node has the max
1289                  * hit record.
1290                  */
1291                 node = page_to_nid(page);
1292                 if (khugepaged_scan_abort(node)) {
1293                         result = SCAN_SCAN_ABORT;
1294                         goto out_unmap;
1295                 }
1296                 khugepaged_node_load[node]++;
1297                 if (!PageLRU(page)) {
1298                         result = SCAN_PAGE_LRU;
1299                         goto out_unmap;
1300                 }
1301                 if (PageLocked(page)) {
1302                         result = SCAN_PAGE_LOCK;
1303                         goto out_unmap;
1304                 }
1305                 if (!PageAnon(page)) {
1306                         result = SCAN_PAGE_ANON;
1307                         goto out_unmap;
1308                 }
1309
1310                 /*
1311                  * Check if the page has any GUP (or other external) pins.
1312                  *
1313                  * Here the check is racy it may see totmal_mapcount > refcount
1314                  * in some cases.
1315                  * For example, one process with one forked child process.
1316                  * The parent has the PMD split due to MADV_DONTNEED, then
1317                  * the child is trying unmap the whole PMD, but khugepaged
1318                  * may be scanning the parent between the child has
1319                  * PageDoubleMap flag cleared and dec the mapcount.  So
1320                  * khugepaged may see total_mapcount > refcount.
1321                  *
1322                  * But such case is ephemeral we could always retry collapse
1323                  * later.  However it may report false positive if the page
1324                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1325                  * will be done again later the risk seems low.
1326                  */
1327                 if (!is_refcount_suitable(page)) {
1328                         result = SCAN_PAGE_COUNT;
1329                         goto out_unmap;
1330                 }
1331                 if (pte_young(pteval) ||
1332                     page_is_young(page) || PageReferenced(page) ||
1333                     mmu_notifier_test_young(vma->vm_mm, address))
1334                         referenced++;
1335         }
1336         if (!writable) {
1337                 result = SCAN_PAGE_RO;
1338         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1339                 result = SCAN_LACK_REFERENCED_PAGE;
1340         } else {
1341                 result = SCAN_SUCCEED;
1342                 ret = 1;
1343         }
1344 out_unmap:
1345         pte_unmap_unlock(pte, ptl);
1346         if (ret) {
1347                 node = khugepaged_find_target_node();
1348                 /* collapse_huge_page will return with the mmap_lock released */
1349                 collapse_huge_page(mm, address, hpage, node,
1350                                 referenced, unmapped);
1351         }
1352 out:
1353         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1354                                      none_or_zero, result, unmapped);
1355         return ret;
1356 }
1357
1358 static void collect_mm_slot(struct mm_slot *mm_slot)
1359 {
1360         struct mm_struct *mm = mm_slot->mm;
1361
1362         lockdep_assert_held(&khugepaged_mm_lock);
1363
1364         if (khugepaged_test_exit(mm)) {
1365                 /* free mm_slot */
1366                 hash_del(&mm_slot->hash);
1367                 list_del(&mm_slot->mm_node);
1368
1369                 /*
1370                  * Not strictly needed because the mm exited already.
1371                  *
1372                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1373                  */
1374
1375                 /* khugepaged_mm_lock actually not necessary for the below */
1376                 free_mm_slot(mm_slot);
1377                 mmdrop(mm);
1378         }
1379 }
1380
1381 #ifdef CONFIG_SHMEM
1382 /*
1383  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1384  * khugepaged should try to collapse the page table.
1385  */
1386 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1387                                          unsigned long addr)
1388 {
1389         struct mm_slot *mm_slot;
1390
1391         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1392
1393         spin_lock(&khugepaged_mm_lock);
1394         mm_slot = get_mm_slot(mm);
1395         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1396                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1397         spin_unlock(&khugepaged_mm_lock);
1398         return 0;
1399 }
1400
1401 /**
1402  * Try to collapse a pte-mapped THP for mm at address haddr.
1403  *
1404  * This function checks whether all the PTEs in the PMD are pointing to the
1405  * right THP. If so, retract the page table so the THP can refault in with
1406  * as pmd-mapped.
1407  */
1408 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1409 {
1410         unsigned long haddr = addr & HPAGE_PMD_MASK;
1411         struct vm_area_struct *vma = find_vma(mm, haddr);
1412         struct page *hpage = NULL;
1413         pte_t *start_pte, *pte;
1414         pmd_t *pmd, _pmd;
1415         spinlock_t *ptl;
1416         int count = 0;
1417         int i;
1418
1419         if (!vma || !vma->vm_file ||
1420             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1421                 return;
1422
1423         /*
1424          * This vm_flags may not have VM_HUGEPAGE if the page was not
1425          * collapsed by this mm. But we can still collapse if the page is
1426          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1427          * will not fail the vma for missing VM_HUGEPAGE
1428          */
1429         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1430                 return;
1431
1432         pmd = mm_find_pmd(mm, haddr);
1433         if (!pmd)
1434                 return;
1435
1436         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1437
1438         /* step 1: check all mapped PTEs are to the right huge page */
1439         for (i = 0, addr = haddr, pte = start_pte;
1440              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1441                 struct page *page;
1442
1443                 /* empty pte, skip */
1444                 if (pte_none(*pte))
1445                         continue;
1446
1447                 /* page swapped out, abort */
1448                 if (!pte_present(*pte))
1449                         goto abort;
1450
1451                 page = vm_normal_page(vma, addr, *pte);
1452
1453                 if (!page || !PageCompound(page))
1454                         goto abort;
1455
1456                 if (!hpage) {
1457                         hpage = compound_head(page);
1458                         /*
1459                          * The mapping of the THP should not change.
1460                          *
1461                          * Note that uprobe, debugger, or MAP_PRIVATE may
1462                          * change the page table, but the new page will
1463                          * not pass PageCompound() check.
1464                          */
1465                         if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1466                                 goto abort;
1467                 }
1468
1469                 /*
1470                  * Confirm the page maps to the correct subpage.
1471                  *
1472                  * Note that uprobe, debugger, or MAP_PRIVATE may change
1473                  * the page table, but the new page will not pass
1474                  * PageCompound() check.
1475                  */
1476                 if (WARN_ON(hpage + i != page))
1477                         goto abort;
1478                 count++;
1479         }
1480
1481         /* step 2: adjust rmap */
1482         for (i = 0, addr = haddr, pte = start_pte;
1483              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1484                 struct page *page;
1485
1486                 if (pte_none(*pte))
1487                         continue;
1488                 page = vm_normal_page(vma, addr, *pte);
1489                 page_remove_rmap(page, false);
1490         }
1491
1492         pte_unmap_unlock(start_pte, ptl);
1493
1494         /* step 3: set proper refcount and mm_counters. */
1495         if (hpage) {
1496                 page_ref_sub(hpage, count);
1497                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1498         }
1499
1500         /* step 4: collapse pmd */
1501         ptl = pmd_lock(vma->vm_mm, pmd);
1502         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1503         spin_unlock(ptl);
1504         mm_dec_nr_ptes(mm);
1505         pte_free(mm, pmd_pgtable(_pmd));
1506         return;
1507
1508 abort:
1509         pte_unmap_unlock(start_pte, ptl);
1510 }
1511
1512 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1513 {
1514         struct mm_struct *mm = mm_slot->mm;
1515         int i;
1516
1517         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1518                 return 0;
1519
1520         if (!mmap_write_trylock(mm))
1521                 return -EBUSY;
1522
1523         if (unlikely(khugepaged_test_exit(mm)))
1524                 goto out;
1525
1526         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1527                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1528
1529 out:
1530         mm_slot->nr_pte_mapped_thp = 0;
1531         mmap_write_unlock(mm);
1532         return 0;
1533 }
1534
1535 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1536 {
1537         struct vm_area_struct *vma;
1538         unsigned long addr;
1539         pmd_t *pmd, _pmd;
1540
1541         i_mmap_lock_write(mapping);
1542         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1543                 /*
1544                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1545                  * got written to. These VMAs are likely not worth investing
1546                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1547                  * later.
1548                  *
1549                  * Not that vma->anon_vma check is racy: it can be set up after
1550                  * the check but before we took mmap_lock by the fault path.
1551                  * But page lock would prevent establishing any new ptes of the
1552                  * page, so we are safe.
1553                  *
1554                  * An alternative would be drop the check, but check that page
1555                  * table is clear before calling pmdp_collapse_flush() under
1556                  * ptl. It has higher chance to recover THP for the VMA, but
1557                  * has higher cost too.
1558                  */
1559                 if (vma->anon_vma)
1560                         continue;
1561                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1562                 if (addr & ~HPAGE_PMD_MASK)
1563                         continue;
1564                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1565                         continue;
1566                 pmd = mm_find_pmd(vma->vm_mm, addr);
1567                 if (!pmd)
1568                         continue;
1569                 /*
1570                  * We need exclusive mmap_lock to retract page table.
1571                  *
1572                  * We use trylock due to lock inversion: we need to acquire
1573                  * mmap_lock while holding page lock. Fault path does it in
1574                  * reverse order. Trylock is a way to avoid deadlock.
1575                  */
1576                 if (mmap_write_trylock(vma->vm_mm)) {
1577                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1578                         /* assume page table is clear */
1579                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1580                         spin_unlock(ptl);
1581                         mmap_write_unlock(vma->vm_mm);
1582                         mm_dec_nr_ptes(vma->vm_mm);
1583                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1584                 } else {
1585                         /* Try again later */
1586                         khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1587                 }
1588         }
1589         i_mmap_unlock_write(mapping);
1590 }
1591
1592 /**
1593  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1594  *
1595  * Basic scheme is simple, details are more complex:
1596  *  - allocate and lock a new huge page;
1597  *  - scan page cache replacing old pages with the new one
1598  *    + swap/gup in pages if necessary;
1599  *    + fill in gaps;
1600  *    + keep old pages around in case rollback is required;
1601  *  - if replacing succeeds:
1602  *    + copy data over;
1603  *    + free old pages;
1604  *    + unlock huge page;
1605  *  - if replacing failed;
1606  *    + put all pages back and unfreeze them;
1607  *    + restore gaps in the page cache;
1608  *    + unlock and free huge page;
1609  */
1610 static void collapse_file(struct mm_struct *mm,
1611                 struct file *file, pgoff_t start,
1612                 struct page **hpage, int node)
1613 {
1614         struct address_space *mapping = file->f_mapping;
1615         gfp_t gfp;
1616         struct page *new_page;
1617         pgoff_t index, end = start + HPAGE_PMD_NR;
1618         LIST_HEAD(pagelist);
1619         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1620         int nr_none = 0, result = SCAN_SUCCEED;
1621         bool is_shmem = shmem_file(file);
1622
1623         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1624         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1625
1626         /* Only allocate from the target node */
1627         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1628
1629         new_page = khugepaged_alloc_page(hpage, gfp, node);
1630         if (!new_page) {
1631                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1632                 goto out;
1633         }
1634
1635         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1636                 result = SCAN_CGROUP_CHARGE_FAIL;
1637                 goto out;
1638         }
1639         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1640
1641         /* This will be less messy when we use multi-index entries */
1642         do {
1643                 xas_lock_irq(&xas);
1644                 xas_create_range(&xas);
1645                 if (!xas_error(&xas))
1646                         break;
1647                 xas_unlock_irq(&xas);
1648                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1649                         result = SCAN_FAIL;
1650                         goto out;
1651                 }
1652         } while (1);
1653
1654         __SetPageLocked(new_page);
1655         if (is_shmem)
1656                 __SetPageSwapBacked(new_page);
1657         new_page->index = start;
1658         new_page->mapping = mapping;
1659
1660         /*
1661          * At this point the new_page is locked and not up-to-date.
1662          * It's safe to insert it into the page cache, because nobody would
1663          * be able to map it or use it in another way until we unlock it.
1664          */
1665
1666         xas_set(&xas, start);
1667         for (index = start; index < end; index++) {
1668                 struct page *page = xas_next(&xas);
1669
1670                 VM_BUG_ON(index != xas.xa_index);
1671                 if (is_shmem) {
1672                         if (!page) {
1673                                 /*
1674                                  * Stop if extent has been truncated or
1675                                  * hole-punched, and is now completely
1676                                  * empty.
1677                                  */
1678                                 if (index == start) {
1679                                         if (!xas_next_entry(&xas, end - 1)) {
1680                                                 result = SCAN_TRUNCATED;
1681                                                 goto xa_locked;
1682                                         }
1683                                         xas_set(&xas, index);
1684                                 }
1685                                 if (!shmem_charge(mapping->host, 1)) {
1686                                         result = SCAN_FAIL;
1687                                         goto xa_locked;
1688                                 }
1689                                 xas_store(&xas, new_page);
1690                                 nr_none++;
1691                                 continue;
1692                         }
1693
1694                         if (xa_is_value(page) || !PageUptodate(page)) {
1695                                 xas_unlock_irq(&xas);
1696                                 /* swap in or instantiate fallocated page */
1697                                 if (shmem_getpage(mapping->host, index, &page,
1698                                                   SGP_NOHUGE)) {
1699                                         result = SCAN_FAIL;
1700                                         goto xa_unlocked;
1701                                 }
1702                         } else if (trylock_page(page)) {
1703                                 get_page(page);
1704                                 xas_unlock_irq(&xas);
1705                         } else {
1706                                 result = SCAN_PAGE_LOCK;
1707                                 goto xa_locked;
1708                         }
1709                 } else {        /* !is_shmem */
1710                         if (!page || xa_is_value(page)) {
1711                                 xas_unlock_irq(&xas);
1712                                 page_cache_sync_readahead(mapping, &file->f_ra,
1713                                                           file, index,
1714                                                           PAGE_SIZE);
1715                                 /* drain pagevecs to help isolate_lru_page() */
1716                                 lru_add_drain();
1717                                 page = find_lock_page(mapping, index);
1718                                 if (unlikely(page == NULL)) {
1719                                         result = SCAN_FAIL;
1720                                         goto xa_unlocked;
1721                                 }
1722                         } else if (PageDirty(page)) {
1723                                 /*
1724                                  * khugepaged only works on read-only fd,
1725                                  * so this page is dirty because it hasn't
1726                                  * been flushed since first write. There
1727                                  * won't be new dirty pages.
1728                                  *
1729                                  * Trigger async flush here and hope the
1730                                  * writeback is done when khugepaged
1731                                  * revisits this page.
1732                                  *
1733                                  * This is a one-off situation. We are not
1734                                  * forcing writeback in loop.
1735                                  */
1736                                 xas_unlock_irq(&xas);
1737                                 filemap_flush(mapping);
1738                                 result = SCAN_FAIL;
1739                                 goto xa_unlocked;
1740                         } else if (trylock_page(page)) {
1741                                 get_page(page);
1742                                 xas_unlock_irq(&xas);
1743                         } else {
1744                                 result = SCAN_PAGE_LOCK;
1745                                 goto xa_locked;
1746                         }
1747                 }
1748
1749                 /*
1750                  * The page must be locked, so we can drop the i_pages lock
1751                  * without racing with truncate.
1752                  */
1753                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1754
1755                 /* make sure the page is up to date */
1756                 if (unlikely(!PageUptodate(page))) {
1757                         result = SCAN_FAIL;
1758                         goto out_unlock;
1759                 }
1760
1761                 /*
1762                  * If file was truncated then extended, or hole-punched, before
1763                  * we locked the first page, then a THP might be there already.
1764                  */
1765                 if (PageTransCompound(page)) {
1766                         result = SCAN_PAGE_COMPOUND;
1767                         goto out_unlock;
1768                 }
1769
1770                 if (page_mapping(page) != mapping) {
1771                         result = SCAN_TRUNCATED;
1772                         goto out_unlock;
1773                 }
1774
1775                 if (!is_shmem && PageDirty(page)) {
1776                         /*
1777                          * khugepaged only works on read-only fd, so this
1778                          * page is dirty because it hasn't been flushed
1779                          * since first write.
1780                          */
1781                         result = SCAN_FAIL;
1782                         goto out_unlock;
1783                 }
1784
1785                 if (isolate_lru_page(page)) {
1786                         result = SCAN_DEL_PAGE_LRU;
1787                         goto out_unlock;
1788                 }
1789
1790                 if (page_has_private(page) &&
1791                     !try_to_release_page(page, GFP_KERNEL)) {
1792                         result = SCAN_PAGE_HAS_PRIVATE;
1793                         putback_lru_page(page);
1794                         goto out_unlock;
1795                 }
1796
1797                 if (page_mapped(page))
1798                         unmap_mapping_pages(mapping, index, 1, false);
1799
1800                 xas_lock_irq(&xas);
1801                 xas_set(&xas, index);
1802
1803                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1804                 VM_BUG_ON_PAGE(page_mapped(page), page);
1805
1806                 /*
1807                  * The page is expected to have page_count() == 3:
1808                  *  - we hold a pin on it;
1809                  *  - one reference from page cache;
1810                  *  - one from isolate_lru_page;
1811                  */
1812                 if (!page_ref_freeze(page, 3)) {
1813                         result = SCAN_PAGE_COUNT;
1814                         xas_unlock_irq(&xas);
1815                         putback_lru_page(page);
1816                         goto out_unlock;
1817                 }
1818
1819                 /*
1820                  * Add the page to the list to be able to undo the collapse if
1821                  * something go wrong.
1822                  */
1823                 list_add_tail(&page->lru, &pagelist);
1824
1825                 /* Finally, replace with the new page. */
1826                 xas_store(&xas, new_page);
1827                 continue;
1828 out_unlock:
1829                 unlock_page(page);
1830                 put_page(page);
1831                 goto xa_unlocked;
1832         }
1833
1834         if (is_shmem)
1835                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1836         else {
1837                 __inc_node_page_state(new_page, NR_FILE_THPS);
1838                 filemap_nr_thps_inc(mapping);
1839         }
1840
1841         if (nr_none) {
1842                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1843                 if (is_shmem)
1844                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1845         }
1846
1847 xa_locked:
1848         xas_unlock_irq(&xas);
1849 xa_unlocked:
1850
1851         if (result == SCAN_SUCCEED) {
1852                 struct page *page, *tmp;
1853
1854                 /*
1855                  * Replacing old pages with new one has succeeded, now we
1856                  * need to copy the content and free the old pages.
1857                  */
1858                 index = start;
1859                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1860                         while (index < page->index) {
1861                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1862                                 index++;
1863                         }
1864                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1865                                         page);
1866                         list_del(&page->lru);
1867                         page->mapping = NULL;
1868                         page_ref_unfreeze(page, 1);
1869                         ClearPageActive(page);
1870                         ClearPageUnevictable(page);
1871                         unlock_page(page);
1872                         put_page(page);
1873                         index++;
1874                 }
1875                 while (index < end) {
1876                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1877                         index++;
1878                 }
1879
1880                 SetPageUptodate(new_page);
1881                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1882                 if (is_shmem)
1883                         set_page_dirty(new_page);
1884                 lru_cache_add(new_page);
1885
1886                 /*
1887                  * Remove pte page tables, so we can re-fault the page as huge.
1888                  */
1889                 retract_page_tables(mapping, start);
1890                 *hpage = NULL;
1891
1892                 khugepaged_pages_collapsed++;
1893         } else {
1894                 struct page *page;
1895
1896                 /* Something went wrong: roll back page cache changes */
1897                 xas_lock_irq(&xas);
1898                 mapping->nrpages -= nr_none;
1899
1900                 if (is_shmem)
1901                         shmem_uncharge(mapping->host, nr_none);
1902
1903                 xas_set(&xas, start);
1904                 xas_for_each(&xas, page, end - 1) {
1905                         page = list_first_entry_or_null(&pagelist,
1906                                         struct page, lru);
1907                         if (!page || xas.xa_index < page->index) {
1908                                 if (!nr_none)
1909                                         break;
1910                                 nr_none--;
1911                                 /* Put holes back where they were */
1912                                 xas_store(&xas, NULL);
1913                                 continue;
1914                         }
1915
1916                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1917
1918                         /* Unfreeze the page. */
1919                         list_del(&page->lru);
1920                         page_ref_unfreeze(page, 2);
1921                         xas_store(&xas, page);
1922                         xas_pause(&xas);
1923                         xas_unlock_irq(&xas);
1924                         unlock_page(page);
1925                         putback_lru_page(page);
1926                         xas_lock_irq(&xas);
1927                 }
1928                 VM_BUG_ON(nr_none);
1929                 xas_unlock_irq(&xas);
1930
1931                 new_page->mapping = NULL;
1932         }
1933
1934         unlock_page(new_page);
1935 out:
1936         VM_BUG_ON(!list_empty(&pagelist));
1937         if (!IS_ERR_OR_NULL(*hpage))
1938                 mem_cgroup_uncharge(*hpage);
1939         /* TODO: tracepoints */
1940 }
1941
1942 static void khugepaged_scan_file(struct mm_struct *mm,
1943                 struct file *file, pgoff_t start, struct page **hpage)
1944 {
1945         struct page *page = NULL;
1946         struct address_space *mapping = file->f_mapping;
1947         XA_STATE(xas, &mapping->i_pages, start);
1948         int present, swap;
1949         int node = NUMA_NO_NODE;
1950         int result = SCAN_SUCCEED;
1951
1952         present = 0;
1953         swap = 0;
1954         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1955         rcu_read_lock();
1956         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1957                 if (xas_retry(&xas, page))
1958                         continue;
1959
1960                 if (xa_is_value(page)) {
1961                         if (++swap > khugepaged_max_ptes_swap) {
1962                                 result = SCAN_EXCEED_SWAP_PTE;
1963                                 break;
1964                         }
1965                         continue;
1966                 }
1967
1968                 if (PageTransCompound(page)) {
1969                         result = SCAN_PAGE_COMPOUND;
1970                         break;
1971                 }
1972
1973                 node = page_to_nid(page);
1974                 if (khugepaged_scan_abort(node)) {
1975                         result = SCAN_SCAN_ABORT;
1976                         break;
1977                 }
1978                 khugepaged_node_load[node]++;
1979
1980                 if (!PageLRU(page)) {
1981                         result = SCAN_PAGE_LRU;
1982                         break;
1983                 }
1984
1985                 if (page_count(page) !=
1986                     1 + page_mapcount(page) + page_has_private(page)) {
1987                         result = SCAN_PAGE_COUNT;
1988                         break;
1989                 }
1990
1991                 /*
1992                  * We probably should check if the page is referenced here, but
1993                  * nobody would transfer pte_young() to PageReferenced() for us.
1994                  * And rmap walk here is just too costly...
1995                  */
1996
1997                 present++;
1998
1999                 if (need_resched()) {
2000                         xas_pause(&xas);
2001                         cond_resched_rcu();
2002                 }
2003         }
2004         rcu_read_unlock();
2005
2006         if (result == SCAN_SUCCEED) {
2007                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2008                         result = SCAN_EXCEED_NONE_PTE;
2009                 } else {
2010                         node = khugepaged_find_target_node();
2011                         collapse_file(mm, file, start, hpage, node);
2012                 }
2013         }
2014
2015         /* TODO: tracepoints */
2016 }
2017 #else
2018 static void khugepaged_scan_file(struct mm_struct *mm,
2019                 struct file *file, pgoff_t start, struct page **hpage)
2020 {
2021         BUILD_BUG();
2022 }
2023
2024 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2025 {
2026         return 0;
2027 }
2028 #endif
2029
2030 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2031                                             struct page **hpage)
2032         __releases(&khugepaged_mm_lock)
2033         __acquires(&khugepaged_mm_lock)
2034 {
2035         struct mm_slot *mm_slot;
2036         struct mm_struct *mm;
2037         struct vm_area_struct *vma;
2038         int progress = 0;
2039
2040         VM_BUG_ON(!pages);
2041         lockdep_assert_held(&khugepaged_mm_lock);
2042
2043         if (khugepaged_scan.mm_slot)
2044                 mm_slot = khugepaged_scan.mm_slot;
2045         else {
2046                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2047                                      struct mm_slot, mm_node);
2048                 khugepaged_scan.address = 0;
2049                 khugepaged_scan.mm_slot = mm_slot;
2050         }
2051         spin_unlock(&khugepaged_mm_lock);
2052         khugepaged_collapse_pte_mapped_thps(mm_slot);
2053
2054         mm = mm_slot->mm;
2055         /*
2056          * Don't wait for semaphore (to avoid long wait times).  Just move to
2057          * the next mm on the list.
2058          */
2059         vma = NULL;
2060         if (unlikely(!mmap_read_trylock(mm)))
2061                 goto breakouterloop_mmap_lock;
2062         if (likely(!khugepaged_test_exit(mm)))
2063                 vma = find_vma(mm, khugepaged_scan.address);
2064
2065         progress++;
2066         for (; vma; vma = vma->vm_next) {
2067                 unsigned long hstart, hend;
2068
2069                 cond_resched();
2070                 if (unlikely(khugepaged_test_exit(mm))) {
2071                         progress++;
2072                         break;
2073                 }
2074                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2075 skip:
2076                         progress++;
2077                         continue;
2078                 }
2079                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2080                 hend = vma->vm_end & HPAGE_PMD_MASK;
2081                 if (hstart >= hend)
2082                         goto skip;
2083                 if (khugepaged_scan.address > hend)
2084                         goto skip;
2085                 if (khugepaged_scan.address < hstart)
2086                         khugepaged_scan.address = hstart;
2087                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2088                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2089                         goto skip;
2090
2091                 while (khugepaged_scan.address < hend) {
2092                         int ret;
2093                         cond_resched();
2094                         if (unlikely(khugepaged_test_exit(mm)))
2095                                 goto breakouterloop;
2096
2097                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2098                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2099                                   hend);
2100                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2101                                 struct file *file = get_file(vma->vm_file);
2102                                 pgoff_t pgoff = linear_page_index(vma,
2103                                                 khugepaged_scan.address);
2104
2105                                 mmap_read_unlock(mm);
2106                                 ret = 1;
2107                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2108                                 fput(file);
2109                         } else {
2110                                 ret = khugepaged_scan_pmd(mm, vma,
2111                                                 khugepaged_scan.address,
2112                                                 hpage);
2113                         }
2114                         /* move to next address */
2115                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2116                         progress += HPAGE_PMD_NR;
2117                         if (ret)
2118                                 /* we released mmap_lock so break loop */
2119                                 goto breakouterloop_mmap_lock;
2120                         if (progress >= pages)
2121                                 goto breakouterloop;
2122                 }
2123         }
2124 breakouterloop:
2125         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2126 breakouterloop_mmap_lock:
2127
2128         spin_lock(&khugepaged_mm_lock);
2129         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2130         /*
2131          * Release the current mm_slot if this mm is about to die, or
2132          * if we scanned all vmas of this mm.
2133          */
2134         if (khugepaged_test_exit(mm) || !vma) {
2135                 /*
2136                  * Make sure that if mm_users is reaching zero while
2137                  * khugepaged runs here, khugepaged_exit will find
2138                  * mm_slot not pointing to the exiting mm.
2139                  */
2140                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2141                         khugepaged_scan.mm_slot = list_entry(
2142                                 mm_slot->mm_node.next,
2143                                 struct mm_slot, mm_node);
2144                         khugepaged_scan.address = 0;
2145                 } else {
2146                         khugepaged_scan.mm_slot = NULL;
2147                         khugepaged_full_scans++;
2148                 }
2149
2150                 collect_mm_slot(mm_slot);
2151         }
2152
2153         return progress;
2154 }
2155
2156 static int khugepaged_has_work(void)
2157 {
2158         return !list_empty(&khugepaged_scan.mm_head) &&
2159                 khugepaged_enabled();
2160 }
2161
2162 static int khugepaged_wait_event(void)
2163 {
2164         return !list_empty(&khugepaged_scan.mm_head) ||
2165                 kthread_should_stop();
2166 }
2167
2168 static void khugepaged_do_scan(void)
2169 {
2170         struct page *hpage = NULL;
2171         unsigned int progress = 0, pass_through_head = 0;
2172         unsigned int pages = khugepaged_pages_to_scan;
2173         bool wait = true;
2174
2175         barrier(); /* write khugepaged_pages_to_scan to local stack */
2176
2177         lru_add_drain_all();
2178
2179         while (progress < pages) {
2180                 if (!khugepaged_prealloc_page(&hpage, &wait))
2181                         break;
2182
2183                 cond_resched();
2184
2185                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2186                         break;
2187
2188                 spin_lock(&khugepaged_mm_lock);
2189                 if (!khugepaged_scan.mm_slot)
2190                         pass_through_head++;
2191                 if (khugepaged_has_work() &&
2192                     pass_through_head < 2)
2193                         progress += khugepaged_scan_mm_slot(pages - progress,
2194                                                             &hpage);
2195                 else
2196                         progress = pages;
2197                 spin_unlock(&khugepaged_mm_lock);
2198         }
2199
2200         if (!IS_ERR_OR_NULL(hpage))
2201                 put_page(hpage);
2202 }
2203
2204 static bool khugepaged_should_wakeup(void)
2205 {
2206         return kthread_should_stop() ||
2207                time_after_eq(jiffies, khugepaged_sleep_expire);
2208 }
2209
2210 static void khugepaged_wait_work(void)
2211 {
2212         if (khugepaged_has_work()) {
2213                 const unsigned long scan_sleep_jiffies =
2214                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2215
2216                 if (!scan_sleep_jiffies)
2217                         return;
2218
2219                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2220                 wait_event_freezable_timeout(khugepaged_wait,
2221                                              khugepaged_should_wakeup(),
2222                                              scan_sleep_jiffies);
2223                 return;
2224         }
2225
2226         if (khugepaged_enabled())
2227                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2228 }
2229
2230 static int khugepaged(void *none)
2231 {
2232         struct mm_slot *mm_slot;
2233
2234         set_freezable();
2235         set_user_nice(current, MAX_NICE);
2236
2237         while (!kthread_should_stop()) {
2238                 khugepaged_do_scan();
2239                 khugepaged_wait_work();
2240         }
2241
2242         spin_lock(&khugepaged_mm_lock);
2243         mm_slot = khugepaged_scan.mm_slot;
2244         khugepaged_scan.mm_slot = NULL;
2245         if (mm_slot)
2246                 collect_mm_slot(mm_slot);
2247         spin_unlock(&khugepaged_mm_lock);
2248         return 0;
2249 }
2250
2251 static void set_recommended_min_free_kbytes(void)
2252 {
2253         struct zone *zone;
2254         int nr_zones = 0;
2255         unsigned long recommended_min;
2256
2257         for_each_populated_zone(zone) {
2258                 /*
2259                  * We don't need to worry about fragmentation of
2260                  * ZONE_MOVABLE since it only has movable pages.
2261                  */
2262                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2263                         continue;
2264
2265                 nr_zones++;
2266         }
2267
2268         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2269         recommended_min = pageblock_nr_pages * nr_zones * 2;
2270
2271         /*
2272          * Make sure that on average at least two pageblocks are almost free
2273          * of another type, one for a migratetype to fall back to and a
2274          * second to avoid subsequent fallbacks of other types There are 3
2275          * MIGRATE_TYPES we care about.
2276          */
2277         recommended_min += pageblock_nr_pages * nr_zones *
2278                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2279
2280         /* don't ever allow to reserve more than 5% of the lowmem */
2281         recommended_min = min(recommended_min,
2282                               (unsigned long) nr_free_buffer_pages() / 20);
2283         recommended_min <<= (PAGE_SHIFT-10);
2284
2285         if (recommended_min > min_free_kbytes) {
2286                 if (user_min_free_kbytes >= 0)
2287                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2288                                 min_free_kbytes, recommended_min);
2289
2290                 min_free_kbytes = recommended_min;
2291         }
2292         setup_per_zone_wmarks();
2293 }
2294
2295 int start_stop_khugepaged(void)
2296 {
2297         static struct task_struct *khugepaged_thread __read_mostly;
2298         static DEFINE_MUTEX(khugepaged_mutex);
2299         int err = 0;
2300
2301         mutex_lock(&khugepaged_mutex);
2302         if (khugepaged_enabled()) {
2303                 if (!khugepaged_thread)
2304                         khugepaged_thread = kthread_run(khugepaged, NULL,
2305                                                         "khugepaged");
2306                 if (IS_ERR(khugepaged_thread)) {
2307                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2308                         err = PTR_ERR(khugepaged_thread);
2309                         khugepaged_thread = NULL;
2310                         goto fail;
2311                 }
2312
2313                 if (!list_empty(&khugepaged_scan.mm_head))
2314                         wake_up_interruptible(&khugepaged_wait);
2315
2316                 set_recommended_min_free_kbytes();
2317         } else if (khugepaged_thread) {
2318                 kthread_stop(khugepaged_thread);
2319                 khugepaged_thread = NULL;
2320         }
2321 fail:
2322         mutex_unlock(&khugepaged_mutex);
2323         return err;
2324 }