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_PTE_NON_PRESENT,
  32         SCAN_PAGE_RO,
  33         SCAN_LACK_REFERENCED_PAGE,
  34         SCAN_PAGE_NULL,
  35         SCAN_SCAN_ABORT,
  36         SCAN_PAGE_COUNT,
  37         SCAN_PAGE_LRU,
  38         SCAN_PAGE_LOCK,
  39         SCAN_PAGE_ANON,
  40         SCAN_PAGE_COMPOUND,
  41         SCAN_ANY_PROCESS,
  42         SCAN_VMA_NULL,
  43         SCAN_VMA_CHECK,
  44         SCAN_ADDRESS_RANGE,
  45         SCAN_SWAP_CACHE_PAGE,
  46         SCAN_DEL_PAGE_LRU,
  47         SCAN_ALLOC_HUGE_PAGE_FAIL,
  48         SCAN_CGROUP_CHARGE_FAIL,
  49         SCAN_EXCEED_SWAP_PTE,
  50         SCAN_TRUNCATED,
  51         SCAN_PAGE_HAS_PRIVATE,
  52 };
  53
  54 #define CREATE_TRACE_POINTS
  55 #include <trace/events/huge_memory.h>
  56
  57 static struct task_struct *khugepaged_thread __read_mostly;
  58 static DEFINE_MUTEX(khugepaged_mutex);
  59
  60 /* default scan 8*512 pte (or vmas) every 30 second */
  61 static unsigned int khugepaged_pages_to_scan __read_mostly;
  62 static unsigned int khugepaged_pages_collapsed;
  63 static unsigned int khugepaged_full_scans;
  64 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  65 /* during fragmentation poll the hugepage allocator once every minute */
  66 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  67 static unsigned long khugepaged_sleep_expire;
  68 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  69 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  70 /*
  71  * default collapse hugepages if there is at least one pte mapped like
  72  * it would have happened if the vma was large enough during page
  73  * fault.
  74  */
  75 static unsigned int khugepaged_max_ptes_none __read_mostly;
  76 static unsigned int khugepaged_max_ptes_swap __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 struct attribute *khugepaged_attr[] = {
 297         &khugepaged_defrag_attr.attr,
 298         &khugepaged_max_ptes_none_attr.attr,
 299         &pages_to_scan_attr.attr,
 300         &pages_collapsed_attr.attr,
 301         &full_scans_attr.attr,
 302         &scan_sleep_millisecs_attr.attr,
 303         &alloc_sleep_millisecs_attr.attr,
 304         &khugepaged_max_ptes_swap_attr.attr,
 305         NULL,
 306 };
 307
 308 struct attribute_group khugepaged_attr_group = {
 309         .attrs = khugepaged_attr,
 310         .name = "khugepaged",
 311 };
 312 #endif /* CONFIG_SYSFS */
 313
 314 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 315
 316 int hugepage_madvise(struct vm_area_struct *vma,
 317                      unsigned long *vm_flags, int advice)
 318 {
 319         switch (advice) {
 320         case MADV_HUGEPAGE:
 321 #ifdef CONFIG_S390
 322                 /*
 323                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 324                  * can't handle this properly after s390_enable_sie, so we simply
 325                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 326                  */
 327                 if (mm_has_pgste(vma->vm_mm))
 328                         return 0;
 329 #endif
 330                 *vm_flags &= ~VM_NOHUGEPAGE;
 331                 *vm_flags |= VM_HUGEPAGE;
 332                 /*
 333                  * If the vma become good for khugepaged to scan,
 334                  * register it here without waiting a page fault that
 335                  * may not happen any time soon.
 336                  */
 337                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 338                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 339                         return -ENOMEM;
 340                 break;
 341         case MADV_NOHUGEPAGE:
 342                 *vm_flags &= ~VM_HUGEPAGE;
 343                 *vm_flags |= VM_NOHUGEPAGE;
 344                 /*
 345                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 346                  * this vma even if we leave the mm registered in khugepaged if
 347                  * it got registered before VM_NOHUGEPAGE was set.
 348                  */
 349                 break;
 350         }
 351
 352         return 0;
 353 }
 354
 355 int __init khugepaged_init(void)
 356 {
 357         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 358                                           sizeof(struct mm_slot),
 359                                           __alignof__(struct mm_slot), 0, NULL);
 360         if (!mm_slot_cache)
 361                 return -ENOMEM;
 362
 363         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 364         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 365         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 366
 367         return 0;
 368 }
 369
 370 void __init khugepaged_destroy(void)
 371 {
 372         kmem_cache_destroy(mm_slot_cache);
 373 }
 374
 375 static inline struct mm_slot *alloc_mm_slot(void)
 376 {
 377         if (!mm_slot_cache)     /* initialization failed */
 378                 return NULL;
 379         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 380 }
 381
 382 static inline void free_mm_slot(struct mm_slot *mm_slot)
 383 {
 384         kmem_cache_free(mm_slot_cache, mm_slot);
 385 }
 386
 387 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 388 {
 389         struct mm_slot *mm_slot;
 390
 391         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 392                 if (mm == mm_slot->mm)
 393                         return mm_slot;
 394
 395         return NULL;
 396 }
 397
 398 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 399                                     struct mm_slot *mm_slot)
 400 {
 401         mm_slot->mm = mm;
 402         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 403 }
 404
 405 static inline int khugepaged_test_exit(struct mm_struct *mm)
 406 {
 407         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
 408 }
 409
 410 static bool hugepage_vma_check(struct vm_area_struct *vma,
 411                                unsigned long vm_flags)
 412 {
 413         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 414             (vm_flags & VM_NOHUGEPAGE) ||
 415             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 416                 return false;
 417
 418         if (shmem_file(vma->vm_file) ||
 419             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 420              vma->vm_file &&
 421              (vm_flags & VM_DENYWRITE))) {
 422                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 423                         return false;
 424                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 425                                 HPAGE_PMD_NR);
 426         }
 427         if (!vma->anon_vma || vma->vm_ops)
 428                 return false;
 429         if (is_vma_temporary_stack(vma))
 430                 return false;
 431         return !(vm_flags & VM_NO_KHUGEPAGED);
 432 }
 433
 434 int __khugepaged_enter(struct mm_struct *mm)
 435 {
 436         struct mm_slot *mm_slot;
 437         int wakeup;
 438
 439         mm_slot = alloc_mm_slot();
 440         if (!mm_slot)
 441                 return -ENOMEM;
 442
 443         /* __khugepaged_exit() must not run from under us */
 444         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 445         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 446                 free_mm_slot(mm_slot);
 447                 return 0;
 448         }
 449
 450         spin_lock(&khugepaged_mm_lock);
 451         insert_to_mm_slots_hash(mm, mm_slot);
 452         /*
 453          * Insert just behind the scanning cursor, to let the area settle
 454          * down a little.
 455          */
 456         wakeup = list_empty(&khugepaged_scan.mm_head);
 457         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 458         spin_unlock(&khugepaged_mm_lock);
 459
 460         mmgrab(mm);
 461         if (wakeup)
 462                 wake_up_interruptible(&khugepaged_wait);
 463
 464         return 0;
 465 }
 466
 467 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 468                                unsigned long vm_flags)
 469 {
 470         unsigned long hstart, hend;
 471
 472         /*
 473          * khugepaged only supports read-only files for non-shmem files.
 474          * khugepaged does not yet work on special mappings. And
 475          * file-private shmem THP is not supported.
 476          */
 477         if (!hugepage_vma_check(vma, vm_flags))
 478                 return 0;
 479
 480         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 481         hend = vma->vm_end & HPAGE_PMD_MASK;
 482         if (hstart < hend)
 483                 return khugepaged_enter(vma, vm_flags);
 484         return 0;
 485 }
 486
 487 void __khugepaged_exit(struct mm_struct *mm)
 488 {
 489         struct mm_slot *mm_slot;
 490         int free = 0;
 491
 492         spin_lock(&khugepaged_mm_lock);
 493         mm_slot = get_mm_slot(mm);
 494         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 495                 hash_del(&mm_slot->hash);
 496                 list_del(&mm_slot->mm_node);
 497                 free = 1;
 498         }
 499         spin_unlock(&khugepaged_mm_lock);
 500
 501         if (free) {
 502                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 503                 free_mm_slot(mm_slot);
 504                 mmdrop(mm);
 505         } else if (mm_slot) {
 506                 /*
 507                  * This is required to serialize against
 508                  * khugepaged_test_exit() (which is guaranteed to run
 509                  * under mmap sem read mode). Stop here (after we
 510                  * return all pagetables will be destroyed) until
 511                  * khugepaged has finished working on the pagetables
 512                  * under the mmap_sem.
 513                  */
 514                 down_write(&mm->mmap_sem);
 515                 up_write(&mm->mmap_sem);
 516         }
 517 }
 518
 519 static void release_pte_page(struct page *page)
 520 {
 521         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 522         unlock_page(page);
 523         putback_lru_page(page);
 524 }
 525
 526 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 527 {
 528         while (--_pte >= pte) {
 529                 pte_t pteval = *_pte;
 530                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 531                         release_pte_page(pte_page(pteval));
 532         }
 533 }
 534
 535 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 536                                         unsigned long address,
 537                                         pte_t *pte)
 538 {
 539         struct page *page = NULL;
 540         pte_t *_pte;
 541         int none_or_zero = 0, result = 0, referenced = 0;
 542         bool writable = false;
 543
 544         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 545              _pte++, address += PAGE_SIZE) {
 546                 pte_t pteval = *_pte;
 547                 if (pte_none(pteval) || (pte_present(pteval) &&
 548                                 is_zero_pfn(pte_pfn(pteval)))) {
 549                         if (!userfaultfd_armed(vma) &&
 550                             ++none_or_zero <= khugepaged_max_ptes_none) {
 551                                 continue;
 552                         } else {
 553                                 result = SCAN_EXCEED_NONE_PTE;
 554                                 goto out;
 555                         }
 556                 }
 557                 if (!pte_present(pteval)) {
 558                         result = SCAN_PTE_NON_PRESENT;
 559                         goto out;
 560                 }
 561                 page = vm_normal_page(vma, address, pteval);
 562                 if (unlikely(!page)) {
 563                         result = SCAN_PAGE_NULL;
 564                         goto out;
 565                 }
 566
 567                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 568                 if (PageCompound(page)) {
 569                         result = SCAN_PAGE_COMPOUND;
 570                         goto out;
 571                 }
 572
 573                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 574
 575                 /*
 576                  * We can do it before isolate_lru_page because the
 577                  * page can't be freed from under us. NOTE: PG_lock
 578                  * is needed to serialize against split_huge_page
 579                  * when invoked from the VM.
 580                  */
 581                 if (!trylock_page(page)) {
 582                         result = SCAN_PAGE_LOCK;
 583                         goto out;
 584                 }
 585
 586                 /*
 587                  * cannot use mapcount: can't collapse if there's a gup pin.
 588                  * The page must only be referenced by the scanned process
 589                  * and page swap cache.
 590                  */
 591                 if (page_count(page) != 1 + PageSwapCache(page)) {
 592                         unlock_page(page);
 593                         result = SCAN_PAGE_COUNT;
 594                         goto out;
 595                 }
 596                 if (pte_write(pteval)) {
 597                         writable = true;
 598                 } else {
 599                         if (PageSwapCache(page) &&
 600                             !reuse_swap_page(page, NULL)) {
 601                                 unlock_page(page);
 602                                 result = SCAN_SWAP_CACHE_PAGE;
 603                                 goto out;
 604                         }
 605                         /*
 606                          * Page is not in the swap cache. It can be collapsed
 607                          * into a THP.
 608                          */
 609                 }
 610
 611                 /*
 612                  * Isolate the page to avoid collapsing an hugepage
 613                  * currently in use by the VM.
 614                  */
 615                 if (isolate_lru_page(page)) {
 616                         unlock_page(page);
 617                         result = SCAN_DEL_PAGE_LRU;
 618                         goto out;
 619                 }
 620                 inc_node_page_state(page,
 621                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 622                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 623                 VM_BUG_ON_PAGE(PageLRU(page), page);
 624
 625                 /* There should be enough young pte to collapse the page */
 626                 if (pte_young(pteval) ||
 627                     page_is_young(page) || PageReferenced(page) ||
 628                     mmu_notifier_test_young(vma->vm_mm, address))
 629                         referenced++;
 630         }
 631
 632         if (unlikely(!writable)) {
 633                 result = SCAN_PAGE_RO;
 634         } else if (unlikely(!referenced)) {
 635                 result = SCAN_LACK_REFERENCED_PAGE;
 636         } else {
 637                 result = SCAN_SUCCEED;
 638                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 639                                                     referenced, writable, result);
 640                 return 1;
 641         }
 642 out:
 643         release_pte_pages(pte, _pte);
 644         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 645                                             referenced, writable, result);
 646         return 0;
 647 }
 648
 649 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 650                                       struct vm_area_struct *vma,
 651                                       unsigned long address,
 652                                       spinlock_t *ptl)
 653 {
 654         pte_t *_pte;
 655         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 656                                 _pte++, page++, address += PAGE_SIZE) {
 657                 pte_t pteval = *_pte;
 658                 struct page *src_page;
 659
 660                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 661                         clear_user_highpage(page, address);
 662                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 663                         if (is_zero_pfn(pte_pfn(pteval))) {
 664                                 /*
 665                                  * ptl mostly unnecessary.
 666                                  */
 667                                 spin_lock(ptl);
 668                                 /*
 669                                  * paravirt calls inside pte_clear here are
 670                                  * superfluous.
 671                                  */
 672                                 pte_clear(vma->vm_mm, address, _pte);
 673                                 spin_unlock(ptl);
 674                         }
 675                 } else {
 676                         src_page = pte_page(pteval);
 677                         copy_user_highpage(page, src_page, address, vma);
 678                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 679                         release_pte_page(src_page);
 680                         /*
 681                          * ptl mostly unnecessary, but preempt has to
 682                          * be disabled to update the per-cpu stats
 683                          * inside page_remove_rmap().
 684                          */
 685                         spin_lock(ptl);
 686                         /*
 687                          * paravirt calls inside pte_clear here are
 688                          * superfluous.
 689                          */
 690                         pte_clear(vma->vm_mm, address, _pte);
 691                         page_remove_rmap(src_page, false);
 692                         spin_unlock(ptl);
 693                         free_page_and_swap_cache(src_page);
 694                 }
 695         }
 696 }
 697
 698 static void khugepaged_alloc_sleep(void)
 699 {
 700         DEFINE_WAIT(wait);
 701
 702         add_wait_queue(&khugepaged_wait, &wait);
 703         freezable_schedule_timeout_interruptible(
 704                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 705         remove_wait_queue(&khugepaged_wait, &wait);
 706 }
 707
 708 static int khugepaged_node_load[MAX_NUMNODES];
 709
 710 static bool khugepaged_scan_abort(int nid)
 711 {
 712         int i;
 713
 714         /*
 715          * If node_reclaim_mode is disabled, then no extra effort is made to
 716          * allocate memory locally.
 717          */
 718         if (!node_reclaim_mode)
 719                 return false;
 720
 721         /* If there is a count for this node already, it must be acceptable */
 722         if (khugepaged_node_load[nid])
 723                 return false;
 724
 725         for (i = 0; i < MAX_NUMNODES; i++) {
 726                 if (!khugepaged_node_load[i])
 727                         continue;
 728                 if (node_distance(nid, i) > node_reclaim_distance)
 729                         return true;
 730         }
 731         return false;
 732 }
 733
 734 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 735 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 736 {
 737         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 738 }
 739
 740 #ifdef CONFIG_NUMA
 741 static int khugepaged_find_target_node(void)
 742 {
 743         static int last_khugepaged_target_node = NUMA_NO_NODE;
 744         int nid, target_node = 0, max_value = 0;
 745
 746         /* find first node with max normal pages hit */
 747         for (nid = 0; nid < MAX_NUMNODES; nid++)
 748                 if (khugepaged_node_load[nid] > max_value) {
 749                         max_value = khugepaged_node_load[nid];
 750                         target_node = nid;
 751                 }
 752
 753         /* do some balance if several nodes have the same hit record */
 754         if (target_node <= last_khugepaged_target_node)
 755                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 756                                 nid++)
 757                         if (max_value == khugepaged_node_load[nid]) {
 758                                 target_node = nid;
 759                                 break;
 760                         }
 761
 762         last_khugepaged_target_node = target_node;
 763         return target_node;
 764 }
 765
 766 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 767 {
 768         if (IS_ERR(*hpage)) {
 769                 if (!*wait)
 770                         return false;
 771
 772                 *wait = false;
 773                 *hpage = NULL;
 774                 khugepaged_alloc_sleep();
 775         } else if (*hpage) {
 776                 put_page(*hpage);
 777                 *hpage = NULL;
 778         }
 779
 780         return true;
 781 }
 782
 783 static struct page *
 784 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 785 {
 786         VM_BUG_ON_PAGE(*hpage, *hpage);
 787
 788         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 789         if (unlikely(!*hpage)) {
 790                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 791                 *hpage = ERR_PTR(-ENOMEM);
 792                 return NULL;
 793         }
 794
 795         prep_transhuge_page(*hpage);
 796         count_vm_event(THP_COLLAPSE_ALLOC);
 797         return *hpage;
 798 }
 799 #else
 800 static int khugepaged_find_target_node(void)
 801 {
 802         return 0;
 803 }
 804
 805 static inline struct page *alloc_khugepaged_hugepage(void)
 806 {
 807         struct page *page;
 808
 809         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 810                            HPAGE_PMD_ORDER);
 811         if (page)
 812                 prep_transhuge_page(page);
 813         return page;
 814 }
 815
 816 static struct page *khugepaged_alloc_hugepage(bool *wait)
 817 {
 818         struct page *hpage;
 819
 820         do {
 821                 hpage = alloc_khugepaged_hugepage();
 822                 if (!hpage) {
 823                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 824                         if (!*wait)
 825                                 return NULL;
 826
 827                         *wait = false;
 828                         khugepaged_alloc_sleep();
 829                 } else
 830                         count_vm_event(THP_COLLAPSE_ALLOC);
 831         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 832
 833         return hpage;
 834 }
 835
 836 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 837 {
 838         /*
 839          * If the hpage allocated earlier was briefly exposed in page cache
 840          * before collapse_file() failed, it is possible that racing lookups
 841          * have not yet completed, and would then be unpleasantly surprised by
 842          * finding the hpage reused for the same mapping at a different offset.
 843          * Just release the previous allocation if there is any danger of that.
 844          */
 845         if (*hpage && page_count(*hpage) > 1) {
 846                 put_page(*hpage);
 847                 *hpage = NULL;
 848         }
 849
 850         if (!*hpage)
 851                 *hpage = khugepaged_alloc_hugepage(wait);
 852
 853         if (unlikely(!*hpage))
 854                 return false;
 855
 856         return true;
 857 }
 858
 859 static struct page *
 860 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 861 {
 862         VM_BUG_ON(!*hpage);
 863
 864         return  *hpage;
 865 }
 866 #endif
 867
 868 /*
 869  * If mmap_sem temporarily dropped, revalidate vma
 870  * before taking mmap_sem.
 871  * Return 0 if succeeds, otherwise return none-zero
 872  * value (scan code).
 873  */
 874
 875 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 876                 struct vm_area_struct **vmap)
 877 {
 878         struct vm_area_struct *vma;
 879         unsigned long hstart, hend;
 880
 881         if (unlikely(khugepaged_test_exit(mm)))
 882                 return SCAN_ANY_PROCESS;
 883
 884         *vmap = vma = find_vma(mm, address);
 885         if (!vma)
 886                 return SCAN_VMA_NULL;
 887
 888         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 889         hend = vma->vm_end & HPAGE_PMD_MASK;
 890         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 891                 return SCAN_ADDRESS_RANGE;
 892         if (!hugepage_vma_check(vma, vma->vm_flags))
 893                 return SCAN_VMA_CHECK;
 894         /* Anon VMA expected */
 895         if (!vma->anon_vma || vma->vm_ops)
 896                 return SCAN_VMA_CHECK;
 897         return 0;
 898 }
 899
 900 /*
 901  * Bring missing pages in from swap, to complete THP collapse.
 902  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 903  *
 904  * Called and returns without pte mapped or spinlocks held,
 905  * but with mmap_sem held to protect against vma changes.
 906  */
 907
 908 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 909                                         struct vm_area_struct *vma,
 910                                         unsigned long address, pmd_t *pmd,
 911                                         int referenced)
 912 {
 913         int swapped_in = 0;
 914         vm_fault_t ret = 0;
 915         struct vm_fault vmf = {
 916                 .vma = vma,
 917                 .address = address,
 918                 .flags = FAULT_FLAG_ALLOW_RETRY,
 919                 .pmd = pmd,
 920                 .pgoff = linear_page_index(vma, address),
 921         };
 922
 923         /* we only decide to swapin, if there is enough young ptes */
 924         if (referenced < HPAGE_PMD_NR/2) {
 925                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 926                 return false;
 927         }
 928         vmf.pte = pte_offset_map(pmd, address);
 929         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 930                         vmf.pte++, vmf.address += PAGE_SIZE) {
 931                 vmf.orig_pte = *vmf.pte;
 932                 if (!is_swap_pte(vmf.orig_pte))
 933                         continue;
 934                 swapped_in++;
 935                 ret = do_swap_page(&vmf);
 936
 937                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 938                 if (ret & VM_FAULT_RETRY) {
 939                         down_read(&mm->mmap_sem);
 940                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 941                                 /* vma is no longer available, don't continue to swapin */
 942                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 943                                 return false;
 944                         }
 945                         /* check if the pmd is still valid */
 946                         if (mm_find_pmd(mm, address) != pmd) {
 947                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 948                                 return false;
 949                         }
 950                 }
 951                 if (ret & VM_FAULT_ERROR) {
 952                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 953                         return false;
 954                 }
 955                 /* pte is unmapped now, we need to map it */
 956                 vmf.pte = pte_offset_map(pmd, vmf.address);
 957         }
 958         vmf.pte--;
 959         pte_unmap(vmf.pte);
 960         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 961         return true;
 962 }
 963
 964 static void collapse_huge_page(struct mm_struct *mm,
 965                                    unsigned long address,
 966                                    struct page **hpage,
 967                                    int node, int referenced)
 968 {
 969         pmd_t *pmd, _pmd;
 970         pte_t *pte;
 971         pgtable_t pgtable;
 972         struct page *new_page;
 973         spinlock_t *pmd_ptl, *pte_ptl;
 974         int isolated = 0, result = 0;
 975         struct mem_cgroup *memcg;
 976         struct vm_area_struct *vma;
 977         struct mmu_notifier_range range;
 978         gfp_t gfp;
 979
 980         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 981
 982         /* Only allocate from the target node */
 983         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 984
 985         /*
 986          * Before allocating the hugepage, release the mmap_sem read lock.
 987          * The allocation can take potentially a long time if it involves
 988          * sync compaction, and we do not need to hold the mmap_sem during
 989          * that. We will recheck the vma after taking it again in write mode.
 990          */
 991         up_read(&mm->mmap_sem);
 992         new_page = khugepaged_alloc_page(hpage, gfp, node);
 993         if (!new_page) {
 994                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 995                 goto out_nolock;
 996         }
 997
 998         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 999                 result = SCAN_CGROUP_CHARGE_FAIL;
1000                 goto out_nolock;
1001         }
1002
1003         down_read(&mm->mmap_sem);
1004         result = hugepage_vma_revalidate(mm, address, &vma);
1005         if (result) {
1006                 mem_cgroup_cancel_charge(new_page, memcg, true);
1007                 up_read(&mm->mmap_sem);
1008                 goto out_nolock;
1009         }
1010
1011         pmd = mm_find_pmd(mm, address);
1012         if (!pmd) {
1013                 result = SCAN_PMD_NULL;
1014                 mem_cgroup_cancel_charge(new_page, memcg, true);
1015                 up_read(&mm->mmap_sem);
1016                 goto out_nolock;
1017         }
1018
1019         /*
1020          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1021          * If it fails, we release mmap_sem and jump out_nolock.
1022          * Continuing to collapse causes inconsistency.
1023          */
1024         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1025                 mem_cgroup_cancel_charge(new_page, memcg, true);
1026                 up_read(&mm->mmap_sem);
1027                 goto out_nolock;
1028         }
1029
1030         up_read(&mm->mmap_sem);
1031         /*
1032          * Prevent all access to pagetables with the exception of
1033          * gup_fast later handled by the ptep_clear_flush and the VM
1034          * handled by the anon_vma lock + PG_lock.
1035          */
1036         down_write(&mm->mmap_sem);
1037         result = hugepage_vma_revalidate(mm, address, &vma);
1038         if (result)
1039                 goto out;
1040         /* check if the pmd is still valid */
1041         if (mm_find_pmd(mm, address) != pmd)
1042                 goto out;
1043
1044         anon_vma_lock_write(vma->anon_vma);
1045
1046         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1047                                 address, address + HPAGE_PMD_SIZE);
1048         mmu_notifier_invalidate_range_start(&range);
1049
1050         pte = pte_offset_map(pmd, address);
1051         pte_ptl = pte_lockptr(mm, pmd);
1052
1053         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1054         /*
1055          * After this gup_fast can't run anymore. This also removes
1056          * any huge TLB entry from the CPU so we won't allow
1057          * huge and small TLB entries for the same virtual address
1058          * to avoid the risk of CPU bugs in that area.
1059          */
1060         _pmd = pmdp_collapse_flush(vma, address, pmd);
1061         spin_unlock(pmd_ptl);
1062         mmu_notifier_invalidate_range_end(&range);
1063         tlb_remove_table_sync_one();
1064
1065         spin_lock(pte_ptl);
1066         isolated = __collapse_huge_page_isolate(vma, address, pte);
1067         spin_unlock(pte_ptl);
1068
1069         if (unlikely(!isolated)) {
1070                 pte_unmap(pte);
1071                 spin_lock(pmd_ptl);
1072                 BUG_ON(!pmd_none(*pmd));
1073                 /*
1074                  * We can only use set_pmd_at when establishing
1075                  * hugepmds and never for establishing regular pmds that
1076                  * points to regular pagetables. Use pmd_populate for that
1077                  */
1078                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1079                 spin_unlock(pmd_ptl);
1080                 anon_vma_unlock_write(vma->anon_vma);
1081                 result = SCAN_FAIL;
1082                 goto out;
1083         }
1084
1085         /*
1086          * All pages are isolated and locked so anon_vma rmap
1087          * can't run anymore.
1088          */
1089         anon_vma_unlock_write(vma->anon_vma);
1090
1091         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1092         pte_unmap(pte);
1093         __SetPageUptodate(new_page);
1094         pgtable = pmd_pgtable(_pmd);
1095
1096         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1097         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1098
1099         /*
1100          * spin_lock() below is not the equivalent of smp_wmb(), so
1101          * this is needed to avoid the copy_huge_page writes to become
1102          * visible after the set_pmd_at() write.
1103          */
1104         smp_wmb();
1105
1106         spin_lock(pmd_ptl);
1107         BUG_ON(!pmd_none(*pmd));
1108         page_add_new_anon_rmap(new_page, vma, address, true);
1109         mem_cgroup_commit_charge(new_page, memcg, false, true);
1110         count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1111         lru_cache_add_active_or_unevictable(new_page, vma);
1112         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1113         set_pmd_at(mm, address, pmd, _pmd);
1114         update_mmu_cache_pmd(vma, address, pmd);
1115         spin_unlock(pmd_ptl);
1116
1117         *hpage = NULL;
1118
1119         khugepaged_pages_collapsed++;
1120         result = SCAN_SUCCEED;
1121 out_up_write:
1122         up_write(&mm->mmap_sem);
1123 out_nolock:
1124         trace_mm_collapse_huge_page(mm, isolated, result);
1125         return;
1126 out:
1127         mem_cgroup_cancel_charge(new_page, memcg, true);
1128         goto out_up_write;
1129 }
1130
1131 static int khugepaged_scan_pmd(struct mm_struct *mm,
1132                                struct vm_area_struct *vma,
1133                                unsigned long address,
1134                                struct page **hpage)
1135 {
1136         pmd_t *pmd;
1137         pte_t *pte, *_pte;
1138         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1139         struct page *page = NULL;
1140         unsigned long _address;
1141         spinlock_t *ptl;
1142         int node = NUMA_NO_NODE, unmapped = 0;
1143         bool writable = false;
1144
1145         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1146
1147         pmd = mm_find_pmd(mm, address);
1148         if (!pmd) {
1149                 result = SCAN_PMD_NULL;
1150                 goto out;
1151         }
1152
1153         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1154         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1155         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1156              _pte++, _address += PAGE_SIZE) {
1157                 pte_t pteval = *_pte;
1158                 if (is_swap_pte(pteval)) {
1159                         if (++unmapped <= khugepaged_max_ptes_swap) {
1160                                 continue;
1161                         } else {
1162                                 result = SCAN_EXCEED_SWAP_PTE;
1163                                 goto out_unmap;
1164                         }
1165                 }
1166                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1167                         if (!userfaultfd_armed(vma) &&
1168                             ++none_or_zero <= khugepaged_max_ptes_none) {
1169                                 continue;
1170                         } else {
1171                                 result = SCAN_EXCEED_NONE_PTE;
1172                                 goto out_unmap;
1173                         }
1174                 }
1175                 if (!pte_present(pteval)) {
1176                         result = SCAN_PTE_NON_PRESENT;
1177                         goto out_unmap;
1178                 }
1179                 if (pte_write(pteval))
1180                         writable = true;
1181
1182                 page = vm_normal_page(vma, _address, pteval);
1183                 if (unlikely(!page)) {
1184                         result = SCAN_PAGE_NULL;
1185                         goto out_unmap;
1186                 }
1187
1188                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1189                 if (PageCompound(page)) {
1190                         result = SCAN_PAGE_COMPOUND;
1191                         goto out_unmap;
1192                 }
1193
1194                 /*
1195                  * Record which node the original page is from and save this
1196                  * information to khugepaged_node_load[].
1197                  * Khupaged will allocate hugepage from the node has the max
1198                  * hit record.
1199                  */
1200                 node = page_to_nid(page);
1201                 if (khugepaged_scan_abort(node)) {
1202                         result = SCAN_SCAN_ABORT;
1203                         goto out_unmap;
1204                 }
1205                 khugepaged_node_load[node]++;
1206                 if (!PageLRU(page)) {
1207                         result = SCAN_PAGE_LRU;
1208                         goto out_unmap;
1209                 }
1210                 if (PageLocked(page)) {
1211                         result = SCAN_PAGE_LOCK;
1212                         goto out_unmap;
1213                 }
1214                 if (!PageAnon(page)) {
1215                         result = SCAN_PAGE_ANON;
1216                         goto out_unmap;
1217                 }
1218
1219                 /*
1220                  * cannot use mapcount: can't collapse if there's a gup pin.
1221                  * The page must only be referenced by the scanned process
1222                  * and page swap cache.
1223                  */
1224                 if (page_count(page) != 1 + PageSwapCache(page)) {
1225                         result = SCAN_PAGE_COUNT;
1226                         goto out_unmap;
1227                 }
1228                 if (pte_young(pteval) ||
1229                     page_is_young(page) || PageReferenced(page) ||
1230                     mmu_notifier_test_young(vma->vm_mm, address))
1231                         referenced++;
1232         }
1233         if (writable) {
1234                 if (referenced) {
1235                         result = SCAN_SUCCEED;
1236                         ret = 1;
1237                 } else {
1238                         result = SCAN_LACK_REFERENCED_PAGE;
1239                 }
1240         } else {
1241                 result = SCAN_PAGE_RO;
1242         }
1243 out_unmap:
1244         pte_unmap_unlock(pte, ptl);
1245         if (ret) {
1246                 node = khugepaged_find_target_node();
1247                 /* collapse_huge_page will return with the mmap_sem released */
1248                 collapse_huge_page(mm, address, hpage, node, referenced);
1249         }
1250 out:
1251         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1252                                      none_or_zero, result, unmapped);
1253         return ret;
1254 }
1255
1256 static void collect_mm_slot(struct mm_slot *mm_slot)
1257 {
1258         struct mm_struct *mm = mm_slot->mm;
1259
1260         lockdep_assert_held(&khugepaged_mm_lock);
1261
1262         if (khugepaged_test_exit(mm)) {
1263                 /* free mm_slot */
1264                 hash_del(&mm_slot->hash);
1265                 list_del(&mm_slot->mm_node);
1266
1267                 /*
1268                  * Not strictly needed because the mm exited already.
1269                  *
1270                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1271                  */
1272
1273                 /* khugepaged_mm_lock actually not necessary for the below */
1274                 free_mm_slot(mm_slot);
1275                 mmdrop(mm);
1276         }
1277 }
1278
1279 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1280 /*
1281  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1282  * khugepaged should try to collapse the page table.
1283  */
1284 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1285                                          unsigned long addr)
1286 {
1287         struct mm_slot *mm_slot;
1288
1289         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1290
1291         spin_lock(&khugepaged_mm_lock);
1292         mm_slot = get_mm_slot(mm);
1293         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1294                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1295         spin_unlock(&khugepaged_mm_lock);
1296         return 0;
1297 }
1298
1299 /**
1300  * Try to collapse a pte-mapped THP for mm at address haddr.
1301  *
1302  * This function checks whether all the PTEs in the PMD are pointing to the
1303  * right THP. If so, retract the page table so the THP can refault in with
1304  * as pmd-mapped.
1305  */
1306 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1307 {
1308         unsigned long haddr = addr & HPAGE_PMD_MASK;
1309         struct vm_area_struct *vma = find_vma(mm, haddr);
1310         struct page *hpage;
1311         pte_t *start_pte, *pte;
1312         pmd_t *pmd, _pmd;
1313         spinlock_t *ptl;
1314         int count = 0;
1315         int i;
1316         struct mmu_notifier_range range;
1317
1318         if (!vma || !vma->vm_file ||
1319             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1320                 return;
1321
1322         /*
1323          * This vm_flags may not have VM_HUGEPAGE if the page was not
1324          * collapsed by this mm. But we can still collapse if the page is
1325          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1326          * will not fail the vma for missing VM_HUGEPAGE
1327          */
1328         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1329                 return;
1330
1331         hpage = find_lock_page(vma->vm_file->f_mapping,
1332                                linear_page_index(vma, haddr));
1333         if (!hpage)
1334                 return;
1335
1336         if (!PageHead(hpage))
1337                 goto drop_hpage;
1338
1339         pmd = mm_find_pmd(mm, haddr);
1340         if (!pmd)
1341                 goto drop_hpage;
1342
1343         /*
1344          * We need to lock the mapping so that from here on, only GUP-fast and
1345          * hardware page walks can access the parts of the page tables that
1346          * we're operating on.
1347          */
1348         i_mmap_lock_write(vma->vm_file->f_mapping);
1349
1350         /*
1351          * This spinlock should be unnecessary: Nobody else should be accessing
1352          * the page tables under spinlock protection here, only
1353          * lockless_pages_from_mm() and the hardware page walker can access page
1354          * tables while all the high-level locks are held in write mode.
1355          */
1356         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1357
1358         /* step 1: check all mapped PTEs are to the right huge page */
1359         for (i = 0, addr = haddr, pte = start_pte;
1360              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1361                 struct page *page;
1362
1363                 /* empty pte, skip */
1364                 if (pte_none(*pte))
1365                         continue;
1366
1367                 /* page swapped out, abort */
1368                 if (!pte_present(*pte))
1369                         goto abort;
1370
1371                 page = vm_normal_page(vma, addr, *pte);
1372
1373                 /*
1374                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1375                  * page table, but the new page will not be a subpage of hpage.
1376                  */
1377                 if (hpage + i != page)
1378                         goto abort;
1379                 count++;
1380         }
1381
1382         /* step 2: adjust rmap */
1383         for (i = 0, addr = haddr, pte = start_pte;
1384              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1385                 struct page *page;
1386
1387                 if (pte_none(*pte))
1388                         continue;
1389                 page = vm_normal_page(vma, addr, *pte);
1390                 page_remove_rmap(page, false);
1391         }
1392
1393         pte_unmap_unlock(start_pte, ptl);
1394
1395         /* step 3: set proper refcount and mm_counters. */
1396         if (count) {
1397                 page_ref_sub(hpage, count);
1398                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1399         }
1400
1401         /* step 4: collapse pmd */
1402         /* we make no change to anon, but protect concurrent anon page lookup */
1403         if (vma->anon_vma)
1404                 anon_vma_lock_write(vma->anon_vma);
1405
1406         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1407                                 haddr + HPAGE_PMD_SIZE);
1408         mmu_notifier_invalidate_range_start(&range);
1409         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1410         mm_dec_nr_ptes(mm);
1411         tlb_remove_table_sync_one();
1412         mmu_notifier_invalidate_range_end(&range);
1413         pte_free(mm, pmd_pgtable(_pmd));
1414
1415         if (vma->anon_vma)
1416                 anon_vma_unlock_write(vma->anon_vma);
1417         i_mmap_unlock_write(vma->vm_file->f_mapping);
1418
1419 drop_hpage:
1420         unlock_page(hpage);
1421         put_page(hpage);
1422         return;
1423
1424 abort:
1425         pte_unmap_unlock(start_pte, ptl);
1426         i_mmap_unlock_write(vma->vm_file->f_mapping);
1427         goto drop_hpage;
1428 }
1429
1430 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1431 {
1432         struct mm_struct *mm = mm_slot->mm;
1433         int i;
1434
1435         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1436                 return 0;
1437
1438         if (!down_write_trylock(&mm->mmap_sem))
1439                 return -EBUSY;
1440
1441         if (unlikely(khugepaged_test_exit(mm)))
1442                 goto out;
1443
1444         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1445                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1446
1447 out:
1448         mm_slot->nr_pte_mapped_thp = 0;
1449         up_write(&mm->mmap_sem);
1450         return 0;
1451 }
1452
1453 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1454 {
1455         struct vm_area_struct *vma;
1456         struct mm_struct *mm;
1457         unsigned long addr;
1458         pmd_t *pmd, _pmd;
1459
1460         i_mmap_lock_write(mapping);
1461         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1462                 /*
1463                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1464                  * got written to. These VMAs are likely not worth investing
1465                  * down_write(mmap_sem) as PMD-mapping is likely to be split
1466                  * later.
1467                  *
1468                  * Not that vma->anon_vma check is racy: it can be set up after
1469                  * the check but before we took mmap_sem by the fault path.
1470                  * But page lock would prevent establishing any new ptes of the
1471                  * page, so we are safe.
1472                  *
1473                  * An alternative would be drop the check, but check that page
1474                  * table is clear before calling pmdp_collapse_flush() under
1475                  * ptl. It has higher chance to recover THP for the VMA, but
1476                  * has higher cost too. It would also probably require locking
1477                  * the anon_vma.
1478                  */
1479                 if (vma->anon_vma)
1480                         continue;
1481                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1482                 if (addr & ~HPAGE_PMD_MASK)
1483                         continue;
1484                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1485                         continue;
1486                 mm = vma->vm_mm;
1487                 pmd = mm_find_pmd(mm, addr);
1488                 if (!pmd)
1489                         continue;
1490                 /*
1491                  * We need exclusive mmap_sem to retract page table.
1492                  *
1493                  * We use trylock due to lock inversion: we need to acquire
1494                  * mmap_sem while holding page lock. Fault path does it in
1495                  * reverse order. Trylock is a way to avoid deadlock.
1496                  */
1497                 if (down_write_trylock(&mm->mmap_sem)) {
1498                         if (!khugepaged_test_exit(mm)) {
1499                                 struct mmu_notifier_range range;
1500
1501                                 mmu_notifier_range_init(&range,
1502                                                         MMU_NOTIFY_CLEAR, 0,
1503                                                         NULL, mm, addr,
1504                                                         addr + HPAGE_PMD_SIZE);
1505                                 mmu_notifier_invalidate_range_start(&range);
1506                                 /* assume page table is clear */
1507                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1508                                 mm_dec_nr_ptes(mm);
1509                                 tlb_remove_table_sync_one();
1510                                 pte_free(mm, pmd_pgtable(_pmd));
1511                                 mmu_notifier_invalidate_range_end(&range);
1512                         }
1513                         up_write(&mm->mmap_sem);
1514                 } else {
1515                         /* Try again later */
1516                         khugepaged_add_pte_mapped_thp(mm, addr);
1517                 }
1518         }
1519         i_mmap_unlock_write(mapping);
1520 }
1521
1522 /**
1523  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1524  *
1525  * Basic scheme is simple, details are more complex:
1526  *  - allocate and lock a new huge page;
1527  *  - scan page cache replacing old pages with the new one
1528  *    + swap/gup in pages if necessary;
1529  *    + fill in gaps;
1530  *    + keep old pages around in case rollback is required;
1531  *  - if replacing succeeds:
1532  *    + copy data over;
1533  *    + free old pages;
1534  *    + unlock huge page;
1535  *  - if replacing failed;
1536  *    + put all pages back and unfreeze them;
1537  *    + restore gaps in the page cache;
1538  *    + unlock and free huge page;
1539  */
1540 static void collapse_file(struct mm_struct *mm,
1541                 struct file *file, pgoff_t start,
1542                 struct page **hpage, int node)
1543 {
1544         struct address_space *mapping = file->f_mapping;
1545         gfp_t gfp;
1546         struct page *new_page;
1547         struct mem_cgroup *memcg;
1548         pgoff_t index, end = start + HPAGE_PMD_NR;
1549         LIST_HEAD(pagelist);
1550         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1551         int nr_none = 0, result = SCAN_SUCCEED;
1552         bool is_shmem = shmem_file(file);
1553
1554         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1555         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1556
1557         /* Only allocate from the target node */
1558         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1559
1560         new_page = khugepaged_alloc_page(hpage, gfp, node);
1561         if (!new_page) {
1562                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1563                 goto out;
1564         }
1565
1566         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1567                 result = SCAN_CGROUP_CHARGE_FAIL;
1568                 goto out;
1569         }
1570
1571         /* This will be less messy when we use multi-index entries */
1572         do {
1573                 xas_lock_irq(&xas);
1574                 xas_create_range(&xas);
1575                 if (!xas_error(&xas))
1576                         break;
1577                 xas_unlock_irq(&xas);
1578                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1579                         mem_cgroup_cancel_charge(new_page, memcg, true);
1580                         result = SCAN_FAIL;
1581                         goto out;
1582                 }
1583         } while (1);
1584
1585         __SetPageLocked(new_page);
1586         if (is_shmem)
1587                 __SetPageSwapBacked(new_page);
1588         new_page->index = start;
1589         new_page->mapping = mapping;
1590
1591         /*
1592          * At this point the new_page is locked and not up-to-date.
1593          * It's safe to insert it into the page cache, because nobody would
1594          * be able to map it or use it in another way until we unlock it.
1595          */
1596
1597         xas_set(&xas, start);
1598         for (index = start; index < end; index++) {
1599                 struct page *page = xas_next(&xas);
1600
1601                 VM_BUG_ON(index != xas.xa_index);
1602                 if (is_shmem) {
1603                         if (!page) {
1604                                 /*
1605                                  * Stop if extent has been truncated or
1606                                  * hole-punched, and is now completely
1607                                  * empty.
1608                                  */
1609                                 if (index == start) {
1610                                         if (!xas_next_entry(&xas, end - 1)) {
1611                                                 result = SCAN_TRUNCATED;
1612                                                 goto xa_locked;
1613                                         }
1614                                         xas_set(&xas, index);
1615                                 }
1616                                 if (!shmem_charge(mapping->host, 1)) {
1617                                         result = SCAN_FAIL;
1618                                         goto xa_locked;
1619                                 }
1620                                 xas_store(&xas, new_page);
1621                                 nr_none++;
1622                                 continue;
1623                         }
1624
1625                         if (xa_is_value(page) || !PageUptodate(page)) {
1626                                 xas_unlock_irq(&xas);
1627                                 /* swap in or instantiate fallocated page */
1628                                 if (shmem_getpage(mapping->host, index, &page,
1629                                                   SGP_NOHUGE)) {
1630                                         result = SCAN_FAIL;
1631                                         goto xa_unlocked;
1632                                 }
1633                         } else if (trylock_page(page)) {
1634                                 get_page(page);
1635                                 xas_unlock_irq(&xas);
1636                         } else {
1637                                 result = SCAN_PAGE_LOCK;
1638                                 goto xa_locked;
1639                         }
1640                 } else {        /* !is_shmem */
1641                         if (!page || xa_is_value(page)) {
1642                                 xas_unlock_irq(&xas);
1643                                 page_cache_sync_readahead(mapping, &file->f_ra,
1644                                                           file, index,
1645                                                           end - index);
1646                                 /* drain pagevecs to help isolate_lru_page() */
1647                                 lru_add_drain();
1648                                 page = find_lock_page(mapping, index);
1649                                 if (unlikely(page == NULL)) {
1650                                         result = SCAN_FAIL;
1651                                         goto xa_unlocked;
1652                                 }
1653                         } else if (trylock_page(page)) {
1654                                 get_page(page);
1655                                 xas_unlock_irq(&xas);
1656                         } else {
1657                                 result = SCAN_PAGE_LOCK;
1658                                 goto xa_locked;
1659                         }
1660                 }
1661
1662                 /*
1663                  * The page must be locked, so we can drop the i_pages lock
1664                  * without racing with truncate.
1665                  */
1666                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1667
1668                 /* make sure the page is up to date */
1669                 if (unlikely(!PageUptodate(page))) {
1670                         result = SCAN_FAIL;
1671                         goto out_unlock;
1672                 }
1673
1674                 /*
1675                  * If file was truncated then extended, or hole-punched, before
1676                  * we locked the first page, then a THP might be there already.
1677                  */
1678                 if (PageTransCompound(page)) {
1679                         result = SCAN_PAGE_COMPOUND;
1680                         goto out_unlock;
1681                 }
1682
1683                 if (page_mapping(page) != mapping) {
1684                         result = SCAN_TRUNCATED;
1685                         goto out_unlock;
1686                 }
1687
1688                 if (!is_shmem && PageDirty(page)) {
1689                         /*
1690                          * khugepaged only works on read-only fd, so this
1691                          * page is dirty because it hasn't been flushed
1692                          * since first write.
1693                          */
1694                         result = SCAN_FAIL;
1695                         goto out_unlock;
1696                 }
1697
1698                 if (isolate_lru_page(page)) {
1699                         result = SCAN_DEL_PAGE_LRU;
1700                         goto out_unlock;
1701                 }
1702
1703                 if (page_has_private(page) &&
1704                     !try_to_release_page(page, GFP_KERNEL)) {
1705                         result = SCAN_PAGE_HAS_PRIVATE;
1706                         putback_lru_page(page);
1707                         goto out_unlock;
1708                 }
1709
1710                 if (page_mapped(page))
1711                         unmap_mapping_pages(mapping, index, 1, false);
1712
1713                 xas_lock_irq(&xas);
1714                 xas_set(&xas, index);
1715
1716                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1717                 VM_BUG_ON_PAGE(page_mapped(page), page);
1718
1719                 /*
1720                  * The page is expected to have page_count() == 3:
1721                  *  - we hold a pin on it;
1722                  *  - one reference from page cache;
1723                  *  - one from isolate_lru_page;
1724                  */
1725                 if (!page_ref_freeze(page, 3)) {
1726                         result = SCAN_PAGE_COUNT;
1727                         xas_unlock_irq(&xas);
1728                         putback_lru_page(page);
1729                         goto out_unlock;
1730                 }
1731
1732                 /*
1733                  * Add the page to the list to be able to undo the collapse if
1734                  * something go wrong.
1735                  */
1736                 list_add_tail(&page->lru, &pagelist);
1737
1738                 /* Finally, replace with the new page. */
1739                 xas_store(&xas, new_page);
1740                 continue;
1741 out_unlock:
1742                 unlock_page(page);
1743                 put_page(page);
1744                 goto xa_unlocked;
1745         }
1746
1747         if (is_shmem)
1748                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1749         else {
1750                 __inc_node_page_state(new_page, NR_FILE_THPS);
1751                 filemap_nr_thps_inc(mapping);
1752         }
1753
1754         if (nr_none) {
1755                 struct zone *zone = page_zone(new_page);
1756
1757                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1758                 if (is_shmem)
1759                         __mod_node_page_state(zone->zone_pgdat,
1760                                               NR_SHMEM, nr_none);
1761         }
1762
1763 xa_locked:
1764         xas_unlock_irq(&xas);
1765 xa_unlocked:
1766
1767         if (result == SCAN_SUCCEED) {
1768                 struct page *page, *tmp;
1769
1770                 /*
1771                  * Replacing old pages with new one has succeeded, now we
1772                  * need to copy the content and free the old pages.
1773                  */
1774                 index = start;
1775                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1776                         while (index < page->index) {
1777                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1778                                 index++;
1779                         }
1780                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1781                                         page);
1782                         list_del(&page->lru);
1783                         page->mapping = NULL;
1784                         page_ref_unfreeze(page, 1);
1785                         ClearPageActive(page);
1786                         ClearPageUnevictable(page);
1787                         unlock_page(page);
1788                         put_page(page);
1789                         index++;
1790                 }
1791                 while (index < end) {
1792                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1793                         index++;
1794                 }
1795
1796                 SetPageUptodate(new_page);
1797                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1798                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1799
1800                 if (is_shmem) {
1801                         set_page_dirty(new_page);
1802                         lru_cache_add_anon(new_page);
1803                 } else {
1804                         lru_cache_add_file(new_page);
1805                 }
1806                 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1807
1808                 /*
1809                  * Remove pte page tables, so we can re-fault the page as huge.
1810                  */
1811                 retract_page_tables(mapping, start);
1812                 *hpage = NULL;
1813
1814                 khugepaged_pages_collapsed++;
1815         } else {
1816                 struct page *page;
1817
1818                 /* Something went wrong: roll back page cache changes */
1819                 xas_lock_irq(&xas);
1820                 mapping->nrpages -= nr_none;
1821
1822                 if (is_shmem)
1823                         shmem_uncharge(mapping->host, nr_none);
1824
1825                 xas_set(&xas, start);
1826                 xas_for_each(&xas, page, end - 1) {
1827                         page = list_first_entry_or_null(&pagelist,
1828                                         struct page, lru);
1829                         if (!page || xas.xa_index < page->index) {
1830                                 if (!nr_none)
1831                                         break;
1832                                 nr_none--;
1833                                 /* Put holes back where they were */
1834                                 xas_store(&xas, NULL);
1835                                 continue;
1836                         }
1837
1838                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1839
1840                         /* Unfreeze the page. */
1841                         list_del(&page->lru);
1842                         page_ref_unfreeze(page, 2);
1843                         xas_store(&xas, page);
1844                         xas_pause(&xas);
1845                         xas_unlock_irq(&xas);
1846                         unlock_page(page);
1847                         putback_lru_page(page);
1848                         xas_lock_irq(&xas);
1849                 }
1850                 VM_BUG_ON(nr_none);
1851                 xas_unlock_irq(&xas);
1852
1853                 mem_cgroup_cancel_charge(new_page, memcg, true);
1854                 new_page->mapping = NULL;
1855         }
1856
1857         unlock_page(new_page);
1858 out:
1859         VM_BUG_ON(!list_empty(&pagelist));
1860         /* TODO: tracepoints */
1861 }
1862
1863 static void khugepaged_scan_file(struct mm_struct *mm,
1864                 struct file *file, pgoff_t start, struct page **hpage)
1865 {
1866         struct page *page = NULL;
1867         struct address_space *mapping = file->f_mapping;
1868         XA_STATE(xas, &mapping->i_pages, start);
1869         int present, swap;
1870         int node = NUMA_NO_NODE;
1871         int result = SCAN_SUCCEED;
1872
1873         present = 0;
1874         swap = 0;
1875         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1876         rcu_read_lock();
1877         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1878                 if (xas_retry(&xas, page))
1879                         continue;
1880
1881                 if (xa_is_value(page)) {
1882                         if (++swap > khugepaged_max_ptes_swap) {
1883                                 result = SCAN_EXCEED_SWAP_PTE;
1884                                 break;
1885                         }
1886                         continue;
1887                 }
1888
1889                 if (PageTransCompound(page)) {
1890                         result = SCAN_PAGE_COMPOUND;
1891                         break;
1892                 }
1893
1894                 node = page_to_nid(page);
1895                 if (khugepaged_scan_abort(node)) {
1896                         result = SCAN_SCAN_ABORT;
1897                         break;
1898                 }
1899                 khugepaged_node_load[node]++;
1900
1901                 if (!PageLRU(page)) {
1902                         result = SCAN_PAGE_LRU;
1903                         break;
1904                 }
1905
1906                 if (page_count(page) !=
1907                     1 + page_mapcount(page) + page_has_private(page)) {
1908                         result = SCAN_PAGE_COUNT;
1909                         break;
1910                 }
1911
1912                 /*
1913                  * We probably should check if the page is referenced here, but
1914                  * nobody would transfer pte_young() to PageReferenced() for us.
1915                  * And rmap walk here is just too costly...
1916                  */
1917
1918                 present++;
1919
1920                 if (need_resched()) {
1921                         xas_pause(&xas);
1922                         cond_resched_rcu();
1923                 }
1924         }
1925         rcu_read_unlock();
1926
1927         if (result == SCAN_SUCCEED) {
1928                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1929                         result = SCAN_EXCEED_NONE_PTE;
1930                 } else {
1931                         node = khugepaged_find_target_node();
1932                         collapse_file(mm, file, start, hpage, node);
1933                 }
1934         }
1935
1936         /* TODO: tracepoints */
1937 }
1938 #else
1939 static void khugepaged_scan_file(struct mm_struct *mm,
1940                 struct file *file, pgoff_t start, struct page **hpage)
1941 {
1942         BUILD_BUG();
1943 }
1944
1945 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1946 {
1947         return 0;
1948 }
1949 #endif
1950
1951 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1952                                             struct page **hpage)
1953         __releases(&khugepaged_mm_lock)
1954         __acquires(&khugepaged_mm_lock)
1955 {
1956         struct mm_slot *mm_slot;
1957         struct mm_struct *mm;
1958         struct vm_area_struct *vma;
1959         int progress = 0;
1960
1961         VM_BUG_ON(!pages);
1962         lockdep_assert_held(&khugepaged_mm_lock);
1963
1964         if (khugepaged_scan.mm_slot)
1965                 mm_slot = khugepaged_scan.mm_slot;
1966         else {
1967                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1968                                      struct mm_slot, mm_node);
1969                 khugepaged_scan.address = 0;
1970                 khugepaged_scan.mm_slot = mm_slot;
1971         }
1972         spin_unlock(&khugepaged_mm_lock);
1973         khugepaged_collapse_pte_mapped_thps(mm_slot);
1974
1975         mm = mm_slot->mm;
1976         /*
1977          * Don't wait for semaphore (to avoid long wait times).  Just move to
1978          * the next mm on the list.
1979          */
1980         vma = NULL;
1981         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1982                 goto breakouterloop_mmap_sem;
1983         if (likely(!khugepaged_test_exit(mm)))
1984                 vma = find_vma(mm, khugepaged_scan.address);
1985
1986         progress++;
1987         for (; vma; vma = vma->vm_next) {
1988                 unsigned long hstart, hend;
1989
1990                 cond_resched();
1991                 if (unlikely(khugepaged_test_exit(mm))) {
1992                         progress++;
1993                         break;
1994                 }
1995                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1996 skip:
1997                         progress++;
1998                         continue;
1999                 }
2000                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2001                 hend = vma->vm_end & HPAGE_PMD_MASK;
2002                 if (hstart >= hend)
2003                         goto skip;
2004                 if (khugepaged_scan.address > hend)
2005                         goto skip;
2006                 if (khugepaged_scan.address < hstart)
2007                         khugepaged_scan.address = hstart;
2008                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2009
2010                 while (khugepaged_scan.address < hend) {
2011                         int ret;
2012                         cond_resched();
2013                         if (unlikely(khugepaged_test_exit(mm)))
2014                                 goto breakouterloop;
2015
2016                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2017                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2018                                   hend);
2019                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2020                                 struct file *file;
2021                                 pgoff_t pgoff = linear_page_index(vma,
2022                                                 khugepaged_scan.address);
2023
2024                                 if (shmem_file(vma->vm_file)
2025                                     && !shmem_huge_enabled(vma))
2026                                         goto skip;
2027                                 file = get_file(vma->vm_file);
2028                                 up_read(&mm->mmap_sem);
2029                                 ret = 1;
2030                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2031                                 fput(file);
2032                         } else {
2033                                 ret = khugepaged_scan_pmd(mm, vma,
2034                                                 khugepaged_scan.address,
2035                                                 hpage);
2036                         }
2037                         /* move to next address */
2038                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2039                         progress += HPAGE_PMD_NR;
2040                         if (ret)
2041                                 /* we released mmap_sem so break loop */
2042                                 goto breakouterloop_mmap_sem;
2043                         if (progress >= pages)
2044                                 goto breakouterloop;
2045                 }
2046         }
2047 breakouterloop:
2048         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2049 breakouterloop_mmap_sem:
2050
2051         spin_lock(&khugepaged_mm_lock);
2052         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2053         /*
2054          * Release the current mm_slot if this mm is about to die, or
2055          * if we scanned all vmas of this mm.
2056          */
2057         if (khugepaged_test_exit(mm) || !vma) {
2058                 /*
2059                  * Make sure that if mm_users is reaching zero while
2060                  * khugepaged runs here, khugepaged_exit will find
2061                  * mm_slot not pointing to the exiting mm.
2062                  */
2063                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2064                         khugepaged_scan.mm_slot = list_entry(
2065                                 mm_slot->mm_node.next,
2066                                 struct mm_slot, mm_node);
2067                         khugepaged_scan.address = 0;
2068                 } else {
2069                         khugepaged_scan.mm_slot = NULL;
2070                         khugepaged_full_scans++;
2071                 }
2072
2073                 collect_mm_slot(mm_slot);
2074         }
2075
2076         return progress;
2077 }
2078
2079 static int khugepaged_has_work(void)
2080 {
2081         return !list_empty(&khugepaged_scan.mm_head) &&
2082                 khugepaged_enabled();
2083 }
2084
2085 static int khugepaged_wait_event(void)
2086 {
2087         return !list_empty(&khugepaged_scan.mm_head) ||
2088                 kthread_should_stop();
2089 }
2090
2091 static void khugepaged_do_scan(void)
2092 {
2093         struct page *hpage = NULL;
2094         unsigned int progress = 0, pass_through_head = 0;
2095         unsigned int pages = khugepaged_pages_to_scan;
2096         bool wait = true;
2097
2098         barrier(); /* write khugepaged_pages_to_scan to local stack */
2099
2100         while (progress < pages) {
2101                 if (!khugepaged_prealloc_page(&hpage, &wait))
2102                         break;
2103
2104                 cond_resched();
2105
2106                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2107                         break;
2108
2109                 spin_lock(&khugepaged_mm_lock);
2110                 if (!khugepaged_scan.mm_slot)
2111                         pass_through_head++;
2112                 if (khugepaged_has_work() &&
2113                     pass_through_head < 2)
2114                         progress += khugepaged_scan_mm_slot(pages - progress,
2115                                                             &hpage);
2116                 else
2117                         progress = pages;
2118                 spin_unlock(&khugepaged_mm_lock);
2119         }
2120
2121         if (!IS_ERR_OR_NULL(hpage))
2122                 put_page(hpage);
2123 }
2124
2125 static bool khugepaged_should_wakeup(void)
2126 {
2127         return kthread_should_stop() ||
2128                time_after_eq(jiffies, khugepaged_sleep_expire);
2129 }
2130
2131 static void khugepaged_wait_work(void)
2132 {
2133         if (khugepaged_has_work()) {
2134                 const unsigned long scan_sleep_jiffies =
2135                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2136
2137                 if (!scan_sleep_jiffies)
2138                         return;
2139
2140                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2141                 wait_event_freezable_timeout(khugepaged_wait,
2142                                              khugepaged_should_wakeup(),
2143                                              scan_sleep_jiffies);
2144                 return;
2145         }
2146
2147         if (khugepaged_enabled())
2148                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2149 }
2150
2151 static int khugepaged(void *none)
2152 {
2153         struct mm_slot *mm_slot;
2154
2155         set_freezable();
2156         set_user_nice(current, MAX_NICE);
2157
2158         while (!kthread_should_stop()) {
2159                 khugepaged_do_scan();
2160                 khugepaged_wait_work();
2161         }
2162
2163         spin_lock(&khugepaged_mm_lock);
2164         mm_slot = khugepaged_scan.mm_slot;
2165         khugepaged_scan.mm_slot = NULL;
2166         if (mm_slot)
2167                 collect_mm_slot(mm_slot);
2168         spin_unlock(&khugepaged_mm_lock);
2169         return 0;
2170 }
2171
2172 static void set_recommended_min_free_kbytes(void)
2173 {
2174         struct zone *zone;
2175         int nr_zones = 0;
2176         unsigned long recommended_min;
2177
2178         for_each_populated_zone(zone) {
2179                 /*
2180                  * We don't need to worry about fragmentation of
2181                  * ZONE_MOVABLE since it only has movable pages.
2182                  */
2183                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2184                         continue;
2185
2186                 nr_zones++;
2187         }
2188
2189         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2190         recommended_min = pageblock_nr_pages * nr_zones * 2;
2191
2192         /*
2193          * Make sure that on average at least two pageblocks are almost free
2194          * of another type, one for a migratetype to fall back to and a
2195          * second to avoid subsequent fallbacks of other types There are 3
2196          * MIGRATE_TYPES we care about.
2197          */
2198         recommended_min += pageblock_nr_pages * nr_zones *
2199                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2200
2201         /* don't ever allow to reserve more than 5% of the lowmem */
2202         recommended_min = min(recommended_min,
2203                               (unsigned long) nr_free_buffer_pages() / 20);
2204         recommended_min <<= (PAGE_SHIFT-10);
2205
2206         if (recommended_min > min_free_kbytes) {
2207                 if (user_min_free_kbytes >= 0)
2208                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2209                                 min_free_kbytes, recommended_min);
2210
2211                 min_free_kbytes = recommended_min;
2212         }
2213         setup_per_zone_wmarks();
2214 }
2215
2216 int start_stop_khugepaged(void)
2217 {
2218         int err = 0;
2219
2220         mutex_lock(&khugepaged_mutex);
2221         if (khugepaged_enabled()) {
2222                 if (!khugepaged_thread)
2223                         khugepaged_thread = kthread_run(khugepaged, NULL,
2224                                                         "khugepaged");
2225                 if (IS_ERR(khugepaged_thread)) {
2226                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2227                         err = PTR_ERR(khugepaged_thread);
2228                         khugepaged_thread = NULL;
2229                         goto fail;
2230                 }
2231
2232                 if (!list_empty(&khugepaged_scan.mm_head))
2233                         wake_up_interruptible(&khugepaged_wait);
2234
2235                 set_recommended_min_free_kbytes();
2236         } else if (khugepaged_thread) {
2237                 kthread_stop(khugepaged_thread);
2238                 khugepaged_thread = NULL;
2239         }
2240 fail:
2241         mutex_unlock(&khugepaged_mutex);
2242         return err;
2243 }
2244
2245 void khugepaged_min_free_kbytes_update(void)
2246 {
2247         mutex_lock(&khugepaged_mutex);
2248         if (khugepaged_enabled() && khugepaged_thread)
2249                 set_recommended_min_free_kbytes();
2250         mutex_unlock(&khugepaged_mutex);
2251 }