mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_EXCEED_SWAP_PTE,
  32         SCAN_EXCEED_SHARED_PTE,
  33         SCAN_PTE_NON_PRESENT,
  34         SCAN_PTE_UFFD_WP,
  35         SCAN_PAGE_RO,
  36         SCAN_LACK_REFERENCED_PAGE,
  37         SCAN_PAGE_NULL,
  38         SCAN_SCAN_ABORT,
  39         SCAN_PAGE_COUNT,
  40         SCAN_PAGE_LRU,
  41         SCAN_PAGE_LOCK,
  42         SCAN_PAGE_ANON,
  43         SCAN_PAGE_COMPOUND,
  44         SCAN_ANY_PROCESS,
  45         SCAN_VMA_NULL,
  46         SCAN_VMA_CHECK,
  47         SCAN_ADDRESS_RANGE,
  48         SCAN_SWAP_CACHE_PAGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 151                scan_sleep_millisecs_store);
 152
 153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 154                                           struct kobj_attribute *attr,
 155                                           char *buf)
 156 {
 157         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 158 }
 159
 160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 161                                            struct kobj_attribute *attr,
 162                                            const char *buf, size_t count)
 163 {
 164         unsigned int msecs;
 165         int err;
 166
 167         err = kstrtouint(buf, 10, &msecs);
 168         if (err)
 169                 return -EINVAL;
 170
 171         khugepaged_alloc_sleep_millisecs = msecs;
 172         khugepaged_sleep_expire = 0;
 173         wake_up_interruptible(&khugepaged_wait);
 174
 175         return count;
 176 }
 177 static struct kobj_attribute alloc_sleep_millisecs_attr =
 178         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 179                alloc_sleep_millisecs_store);
 180
 181 static ssize_t pages_to_scan_show(struct kobject *kobj,
 182                                   struct kobj_attribute *attr,
 183                                   char *buf)
 184 {
 185         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 186 }
 187 static ssize_t pages_to_scan_store(struct kobject *kobj,
 188                                    struct kobj_attribute *attr,
 189                                    const char *buf, size_t count)
 190 {
 191         unsigned int pages;
 192         int err;
 193
 194         err = kstrtouint(buf, 10, &pages);
 195         if (err || !pages)
 196                 return -EINVAL;
 197
 198         khugepaged_pages_to_scan = pages;
 199
 200         return count;
 201 }
 202 static struct kobj_attribute pages_to_scan_attr =
 203         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 204                pages_to_scan_store);
 205
 206 static ssize_t pages_collapsed_show(struct kobject *kobj,
 207                                     struct kobj_attribute *attr,
 208                                     char *buf)
 209 {
 210         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 211 }
 212 static struct kobj_attribute pages_collapsed_attr =
 213         __ATTR_RO(pages_collapsed);
 214
 215 static ssize_t full_scans_show(struct kobject *kobj,
 216                                struct kobj_attribute *attr,
 217                                char *buf)
 218 {
 219         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 220 }
 221 static struct kobj_attribute full_scans_attr =
 222         __ATTR_RO(full_scans);
 223
 224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 225                                       struct kobj_attribute *attr, char *buf)
 226 {
 227         return single_hugepage_flag_show(kobj, attr, buf,
 228                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 229 }
 230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 231                                        struct kobj_attribute *attr,
 232                                        const char *buf, size_t count)
 233 {
 234         return single_hugepage_flag_store(kobj, attr, buf, count,
 235                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 236 }
 237 static struct kobj_attribute khugepaged_defrag_attr =
 238         __ATTR(defrag, 0644, khugepaged_defrag_show,
 239                khugepaged_defrag_store);
 240
 241 /*
 242  * max_ptes_none controls if khugepaged should collapse hugepages over
 243  * any unmapped ptes in turn potentially increasing the memory
 244  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 245  * reduce the available free memory in the system as it
 246  * runs. Increasing max_ptes_none will instead potentially reduce the
 247  * free memory in the system during the khugepaged scan.
 248  */
 249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 250                                              struct kobj_attribute *attr,
 251                                              char *buf)
 252 {
 253         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 254 }
 255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 256                                               struct kobj_attribute *attr,
 257                                               const char *buf, size_t count)
 258 {
 259         int err;
 260         unsigned long max_ptes_none;
 261
 262         err = kstrtoul(buf, 10, &max_ptes_none);
 263         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 264                 return -EINVAL;
 265
 266         khugepaged_max_ptes_none = max_ptes_none;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 271         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 272                khugepaged_max_ptes_none_store);
 273
 274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 275                                              struct kobj_attribute *attr,
 276                                              char *buf)
 277 {
 278         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 279 }
 280
 281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 282                                               struct kobj_attribute *attr,
 283                                               const char *buf, size_t count)
 284 {
 285         int err;
 286         unsigned long max_ptes_swap;
 287
 288         err  = kstrtoul(buf, 10, &max_ptes_swap);
 289         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 290                 return -EINVAL;
 291
 292         khugepaged_max_ptes_swap = max_ptes_swap;
 293
 294         return count;
 295 }
 296
 297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 298         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 299                khugepaged_max_ptes_swap_store);
 300
 301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 302                                                struct kobj_attribute *attr,
 303                                                char *buf)
 304 {
 305         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 306 }
 307
 308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 309                                               struct kobj_attribute *attr,
 310                                               const char *buf, size_t count)
 311 {
 312         int err;
 313         unsigned long max_ptes_shared;
 314
 315         err  = kstrtoul(buf, 10, &max_ptes_shared);
 316         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 317                 return -EINVAL;
 318
 319         khugepaged_max_ptes_shared = max_ptes_shared;
 320
 321         return count;
 322 }
 323
 324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 325         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 326                khugepaged_max_ptes_shared_store);
 327
 328 static struct attribute *khugepaged_attr[] = {
 329         &khugepaged_defrag_attr.attr,
 330         &khugepaged_max_ptes_none_attr.attr,
 331         &khugepaged_max_ptes_swap_attr.attr,
 332         &khugepaged_max_ptes_shared_attr.attr,
 333         &pages_to_scan_attr.attr,
 334         &pages_collapsed_attr.attr,
 335         &full_scans_attr.attr,
 336         &scan_sleep_millisecs_attr.attr,
 337         &alloc_sleep_millisecs_attr.attr,
 338         NULL,
 339 };
 340
 341 struct attribute_group khugepaged_attr_group = {
 342         .attrs = khugepaged_attr,
 343         .name = "khugepaged",
 344 };
 345 #endif /* CONFIG_SYSFS */
 346
 347 int hugepage_madvise(struct vm_area_struct *vma,
 348                      unsigned long *vm_flags, int advice)
 349 {
 350         switch (advice) {
 351         case MADV_HUGEPAGE:
 352 #ifdef CONFIG_S390
 353                 /*
 354                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 355                  * can't handle this properly after s390_enable_sie, so we simply
 356                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 357                  */
 358                 if (mm_has_pgste(vma->vm_mm))
 359                         return 0;
 360 #endif
 361                 *vm_flags &= ~VM_NOHUGEPAGE;
 362                 *vm_flags |= VM_HUGEPAGE;
 363                 /*
 364                  * If the vma become good for khugepaged to scan,
 365                  * register it here without waiting a page fault that
 366                  * may not happen any time soon.
 367                  */
 368                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 369                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 370                         return -ENOMEM;
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct mm_slot),
 390                                           __alignof__(struct mm_slot), 0, NULL);
 391         if (!mm_slot_cache)
 392                 return -ENOMEM;
 393
 394         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 395         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 396         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 397         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 398
 399         return 0;
 400 }
 401
 402 void __init khugepaged_destroy(void)
 403 {
 404         kmem_cache_destroy(mm_slot_cache);
 405 }
 406
 407 static inline struct mm_slot *alloc_mm_slot(void)
 408 {
 409         if (!mm_slot_cache)     /* initialization failed */
 410                 return NULL;
 411         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 412 }
 413
 414 static inline void free_mm_slot(struct mm_slot *mm_slot)
 415 {
 416         kmem_cache_free(mm_slot_cache, mm_slot);
 417 }
 418
 419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 420 {
 421         struct mm_slot *mm_slot;
 422
 423         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 424                 if (mm == mm_slot->mm)
 425                         return mm_slot;
 426
 427         return NULL;
 428 }
 429
 430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 431                                     struct mm_slot *mm_slot)
 432 {
 433         mm_slot->mm = mm;
 434         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 435 }
 436
 437 static inline int khugepaged_test_exit(struct mm_struct *mm)
 438 {
 439         return atomic_read(&mm->mm_users) == 0;
 440 }
 441
 442 static bool hugepage_vma_check(struct vm_area_struct *vma,
 443                                unsigned long vm_flags)
 444 {
 445         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 446             (vm_flags & VM_NOHUGEPAGE) ||
 447             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 448                 return false;
 449
 450         if (shmem_file(vma->vm_file) ||
 451             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 452              vma->vm_file &&
 453              (vm_flags & VM_DENYWRITE))) {
 454                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 455                                 HPAGE_PMD_NR);
 456         }
 457         if (!vma->anon_vma || vma->vm_ops)
 458                 return false;
 459         if (vma_is_temporary_stack(vma))
 460                 return false;
 461         return !(vm_flags & VM_NO_KHUGEPAGED);
 462 }
 463
 464 int __khugepaged_enter(struct mm_struct *mm)
 465 {
 466         struct mm_slot *mm_slot;
 467         int wakeup;
 468
 469         mm_slot = alloc_mm_slot();
 470         if (!mm_slot)
 471                 return -ENOMEM;
 472
 473         /* __khugepaged_exit() must not run from under us */
 474         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 475         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 476                 free_mm_slot(mm_slot);
 477                 return 0;
 478         }
 479
 480         spin_lock(&khugepaged_mm_lock);
 481         insert_to_mm_slots_hash(mm, mm_slot);
 482         /*
 483          * Insert just behind the scanning cursor, to let the area settle
 484          * down a little.
 485          */
 486         wakeup = list_empty(&khugepaged_scan.mm_head);
 487         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 488         spin_unlock(&khugepaged_mm_lock);
 489
 490         mmgrab(mm);
 491         if (wakeup)
 492                 wake_up_interruptible(&khugepaged_wait);
 493
 494         return 0;
 495 }
 496
 497 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 498                                unsigned long vm_flags)
 499 {
 500         unsigned long hstart, hend;
 501
 502         /*
 503          * khugepaged only supports read-only files for non-shmem files.
 504          * khugepaged does not yet work on special mappings. And
 505          * file-private shmem THP is not supported.
 506          */
 507         if (!hugepage_vma_check(vma, vm_flags))
 508                 return 0;
 509
 510         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 511         hend = vma->vm_end & HPAGE_PMD_MASK;
 512         if (hstart < hend)
 513                 return khugepaged_enter(vma, vm_flags);
 514         return 0;
 515 }
 516
 517 void __khugepaged_exit(struct mm_struct *mm)
 518 {
 519         struct mm_slot *mm_slot;
 520         int free = 0;
 521
 522         spin_lock(&khugepaged_mm_lock);
 523         mm_slot = get_mm_slot(mm);
 524         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 525                 hash_del(&mm_slot->hash);
 526                 list_del(&mm_slot->mm_node);
 527                 free = 1;
 528         }
 529         spin_unlock(&khugepaged_mm_lock);
 530
 531         if (free) {
 532                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 533                 free_mm_slot(mm_slot);
 534                 mmdrop(mm);
 535         } else if (mm_slot) {
 536                 /*
 537                  * This is required to serialize against
 538                  * khugepaged_test_exit() (which is guaranteed to run
 539                  * under mmap sem read mode). Stop here (after we
 540                  * return all pagetables will be destroyed) until
 541                  * khugepaged has finished working on the pagetables
 542                  * under the mmap_lock.
 543                  */
 544                 mmap_write_lock(mm);
 545                 mmap_write_unlock(mm);
 546         }
 547 }
 548
 549 static void release_pte_page(struct page *page)
 550 {
 551         mod_node_page_state(page_pgdat(page),
 552                         NR_ISOLATED_ANON + page_is_file_lru(page),
 553                         -compound_nr(page));
 554         unlock_page(page);
 555         putback_lru_page(page);
 556 }
 557
 558 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 559                 struct list_head *compound_pagelist)
 560 {
 561         struct page *page, *tmp;
 562
 563         while (--_pte >= pte) {
 564                 pte_t pteval = *_pte;
 565
 566                 page = pte_page(pteval);
 567                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 568                                 !PageCompound(page))
 569                         release_pte_page(page);
 570         }
 571
 572         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 573                 list_del(&page->lru);
 574                 release_pte_page(page);
 575         }
 576 }
 577
 578 static bool is_refcount_suitable(struct page *page)
 579 {
 580         int expected_refcount;
 581
 582         expected_refcount = total_mapcount(page);
 583         if (PageSwapCache(page))
 584                 expected_refcount += compound_nr(page);
 585
 586         return page_count(page) == expected_refcount;
 587 }
 588
 589 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 590                                         unsigned long address,
 591                                         pte_t *pte,
 592                                         struct list_head *compound_pagelist)
 593 {
 594         struct page *page = NULL;
 595         pte_t *_pte;
 596         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 597         bool writable = false;
 598
 599         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 600              _pte++, address += PAGE_SIZE) {
 601                 pte_t pteval = *_pte;
 602                 if (pte_none(pteval) || (pte_present(pteval) &&
 603                                 is_zero_pfn(pte_pfn(pteval)))) {
 604                         if (!userfaultfd_armed(vma) &&
 605                             ++none_or_zero <= khugepaged_max_ptes_none) {
 606                                 continue;
 607                         } else {
 608                                 result = SCAN_EXCEED_NONE_PTE;
 609                                 goto out;
 610                         }
 611                 }
 612                 if (!pte_present(pteval)) {
 613                         result = SCAN_PTE_NON_PRESENT;
 614                         goto out;
 615                 }
 616                 page = vm_normal_page(vma, address, pteval);
 617                 if (unlikely(!page)) {
 618                         result = SCAN_PAGE_NULL;
 619                         goto out;
 620                 }
 621
 622                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 623
 624                 if (page_mapcount(page) > 1 &&
 625                                 ++shared > khugepaged_max_ptes_shared) {
 626                         result = SCAN_EXCEED_SHARED_PTE;
 627                         goto out;
 628                 }
 629
 630                 if (PageCompound(page)) {
 631                         struct page *p;
 632                         page = compound_head(page);
 633
 634                         /*
 635                          * Check if we have dealt with the compound page
 636                          * already
 637                          */
 638                         list_for_each_entry(p, compound_pagelist, lru) {
 639                                 if (page == p)
 640                                         goto next;
 641                         }
 642                 }
 643
 644                 /*
 645                  * We can do it before isolate_lru_page because the
 646                  * page can't be freed from under us. NOTE: PG_lock
 647                  * is needed to serialize against split_huge_page
 648                  * when invoked from the VM.
 649                  */
 650                 if (!trylock_page(page)) {
 651                         result = SCAN_PAGE_LOCK;
 652                         goto out;
 653                 }
 654
 655                 /*
 656                  * Check if the page has any GUP (or other external) pins.
 657                  *
 658                  * The page table that maps the page has been already unlinked
 659                  * from the page table tree and this process cannot get
 660                  * an additinal pin on the page.
 661                  *
 662                  * New pins can come later if the page is shared across fork,
 663                  * but not from this process. The other process cannot write to
 664                  * the page, only trigger CoW.
 665                  */
 666                 if (!is_refcount_suitable(page)) {
 667                         unlock_page(page);
 668                         result = SCAN_PAGE_COUNT;
 669                         goto out;
 670                 }
 671                 if (!pte_write(pteval) && PageSwapCache(page) &&
 672                                 !reuse_swap_page(page, NULL)) {
 673                         /*
 674                          * Page is in the swap cache and cannot be re-used.
 675                          * It cannot be collapsed into a THP.
 676                          */
 677                         unlock_page(page);
 678                         result = SCAN_SWAP_CACHE_PAGE;
 679                         goto out;
 680                 }
 681
 682                 /*
 683                  * Isolate the page to avoid collapsing an hugepage
 684                  * currently in use by the VM.
 685                  */
 686                 if (isolate_lru_page(page)) {
 687                         unlock_page(page);
 688                         result = SCAN_DEL_PAGE_LRU;
 689                         goto out;
 690                 }
 691                 mod_node_page_state(page_pgdat(page),
 692                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 693                                 compound_nr(page));
 694                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 695                 VM_BUG_ON_PAGE(PageLRU(page), page);
 696
 697                 if (PageCompound(page))
 698                         list_add_tail(&page->lru, compound_pagelist);
 699 next:
 700                 /* There should be enough young pte to collapse the page */
 701                 if (pte_young(pteval) ||
 702                     page_is_young(page) || PageReferenced(page) ||
 703                     mmu_notifier_test_young(vma->vm_mm, address))
 704                         referenced++;
 705
 706                 if (pte_write(pteval))
 707                         writable = true;
 708         }
 709         if (likely(writable)) {
 710                 if (likely(referenced)) {
 711                         result = SCAN_SUCCEED;
 712                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 713                                                             referenced, writable, result);
 714                         return 1;
 715                 }
 716         } else {
 717                 result = SCAN_PAGE_RO;
 718         }
 719
 720 out:
 721         release_pte_pages(pte, _pte, compound_pagelist);
 722         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 723                                             referenced, writable, result);
 724         return 0;
 725 }
 726
 727 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 728                                       struct vm_area_struct *vma,
 729                                       unsigned long address,
 730                                       spinlock_t *ptl,
 731                                       struct list_head *compound_pagelist)
 732 {
 733         struct page *src_page, *tmp;
 734         pte_t *_pte;
 735         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 736                                 _pte++, page++, address += PAGE_SIZE) {
 737                 pte_t pteval = *_pte;
 738
 739                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 740                         clear_user_highpage(page, address);
 741                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 742                         if (is_zero_pfn(pte_pfn(pteval))) {
 743                                 /*
 744                                  * ptl mostly unnecessary.
 745                                  */
 746                                 spin_lock(ptl);
 747                                 /*
 748                                  * paravirt calls inside pte_clear here are
 749                                  * superfluous.
 750                                  */
 751                                 pte_clear(vma->vm_mm, address, _pte);
 752                                 spin_unlock(ptl);
 753                         }
 754                 } else {
 755                         src_page = pte_page(pteval);
 756                         copy_user_highpage(page, src_page, address, vma);
 757                         if (!PageCompound(src_page))
 758                                 release_pte_page(src_page);
 759                         /*
 760                          * ptl mostly unnecessary, but preempt has to
 761                          * be disabled to update the per-cpu stats
 762                          * inside page_remove_rmap().
 763                          */
 764                         spin_lock(ptl);
 765                         /*
 766                          * paravirt calls inside pte_clear here are
 767                          * superfluous.
 768                          */
 769                         pte_clear(vma->vm_mm, address, _pte);
 770                         page_remove_rmap(src_page, false);
 771                         spin_unlock(ptl);
 772                         free_page_and_swap_cache(src_page);
 773                 }
 774         }
 775
 776         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 777                 list_del(&src_page->lru);
 778                 release_pte_page(src_page);
 779         }
 780 }
 781
 782 static void khugepaged_alloc_sleep(void)
 783 {
 784         DEFINE_WAIT(wait);
 785
 786         add_wait_queue(&khugepaged_wait, &wait);
 787         freezable_schedule_timeout_interruptible(
 788                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 789         remove_wait_queue(&khugepaged_wait, &wait);
 790 }
 791
 792 static int khugepaged_node_load[MAX_NUMNODES];
 793
 794 static bool khugepaged_scan_abort(int nid)
 795 {
 796         int i;
 797
 798         /*
 799          * If node_reclaim_mode is disabled, then no extra effort is made to
 800          * allocate memory locally.
 801          */
 802         if (!node_reclaim_mode)
 803                 return false;
 804
 805         /* If there is a count for this node already, it must be acceptable */
 806         if (khugepaged_node_load[nid])
 807                 return false;
 808
 809         for (i = 0; i < MAX_NUMNODES; i++) {
 810                 if (!khugepaged_node_load[i])
 811                         continue;
 812                 if (node_distance(nid, i) > node_reclaim_distance)
 813                         return true;
 814         }
 815         return false;
 816 }
 817
 818 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 819 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 820 {
 821         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 822 }
 823
 824 #ifdef CONFIG_NUMA
 825 static int khugepaged_find_target_node(void)
 826 {
 827         static int last_khugepaged_target_node = NUMA_NO_NODE;
 828         int nid, target_node = 0, max_value = 0;
 829
 830         /* find first node with max normal pages hit */
 831         for (nid = 0; nid < MAX_NUMNODES; nid++)
 832                 if (khugepaged_node_load[nid] > max_value) {
 833                         max_value = khugepaged_node_load[nid];
 834                         target_node = nid;
 835                 }
 836
 837         /* do some balance if several nodes have the same hit record */
 838         if (target_node <= last_khugepaged_target_node)
 839                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 840                                 nid++)
 841                         if (max_value == khugepaged_node_load[nid]) {
 842                                 target_node = nid;
 843                                 break;
 844                         }
 845
 846         last_khugepaged_target_node = target_node;
 847         return target_node;
 848 }
 849
 850 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 851 {
 852         if (IS_ERR(*hpage)) {
 853                 if (!*wait)
 854                         return false;
 855
 856                 *wait = false;
 857                 *hpage = NULL;
 858                 khugepaged_alloc_sleep();
 859         } else if (*hpage) {
 860                 put_page(*hpage);
 861                 *hpage = NULL;
 862         }
 863
 864         return true;
 865 }
 866
 867 static struct page *
 868 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 869 {
 870         VM_BUG_ON_PAGE(*hpage, *hpage);
 871
 872         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 873         if (unlikely(!*hpage)) {
 874                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 875                 *hpage = ERR_PTR(-ENOMEM);
 876                 return NULL;
 877         }
 878
 879         prep_transhuge_page(*hpage);
 880         count_vm_event(THP_COLLAPSE_ALLOC);
 881         return *hpage;
 882 }
 883 #else
 884 static int khugepaged_find_target_node(void)
 885 {
 886         return 0;
 887 }
 888
 889 static inline struct page *alloc_khugepaged_hugepage(void)
 890 {
 891         struct page *page;
 892
 893         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 894                            HPAGE_PMD_ORDER);
 895         if (page)
 896                 prep_transhuge_page(page);
 897         return page;
 898 }
 899
 900 static struct page *khugepaged_alloc_hugepage(bool *wait)
 901 {
 902         struct page *hpage;
 903
 904         do {
 905                 hpage = alloc_khugepaged_hugepage();
 906                 if (!hpage) {
 907                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 908                         if (!*wait)
 909                                 return NULL;
 910
 911                         *wait = false;
 912                         khugepaged_alloc_sleep();
 913                 } else
 914                         count_vm_event(THP_COLLAPSE_ALLOC);
 915         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 916
 917         return hpage;
 918 }
 919
 920 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 921 {
 922         /*
 923          * If the hpage allocated earlier was briefly exposed in page cache
 924          * before collapse_file() failed, it is possible that racing lookups
 925          * have not yet completed, and would then be unpleasantly surprised by
 926          * finding the hpage reused for the same mapping at a different offset.
 927          * Just release the previous allocation if there is any danger of that.
 928          */
 929         if (*hpage && page_count(*hpage) > 1) {
 930                 put_page(*hpage);
 931                 *hpage = NULL;
 932         }
 933
 934         if (!*hpage)
 935                 *hpage = khugepaged_alloc_hugepage(wait);
 936
 937         if (unlikely(!*hpage))
 938                 return false;
 939
 940         return true;
 941 }
 942
 943 static struct page *
 944 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 945 {
 946         VM_BUG_ON(!*hpage);
 947
 948         return  *hpage;
 949 }
 950 #endif
 951
 952 /*
 953  * If mmap_lock temporarily dropped, revalidate vma
 954  * before taking mmap_lock.
 955  * Return 0 if succeeds, otherwise return none-zero
 956  * value (scan code).
 957  */
 958
 959 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 960                 struct vm_area_struct **vmap)
 961 {
 962         struct vm_area_struct *vma;
 963         unsigned long hstart, hend;
 964
 965         if (unlikely(khugepaged_test_exit(mm)))
 966                 return SCAN_ANY_PROCESS;
 967
 968         *vmap = vma = find_vma(mm, address);
 969         if (!vma)
 970                 return SCAN_VMA_NULL;
 971
 972         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 973         hend = vma->vm_end & HPAGE_PMD_MASK;
 974         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 975                 return SCAN_ADDRESS_RANGE;
 976         if (!hugepage_vma_check(vma, vma->vm_flags))
 977                 return SCAN_VMA_CHECK;
 978         /* Anon VMA expected */
 979         if (!vma->anon_vma || vma->vm_ops)
 980                 return SCAN_VMA_CHECK;
 981         return 0;
 982 }
 983
 984 /*
 985  * Bring missing pages in from swap, to complete THP collapse.
 986  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 987  *
 988  * Called and returns without pte mapped or spinlocks held,
 989  * but with mmap_lock held to protect against vma changes.
 990  */
 991
 992 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 993                                         struct vm_area_struct *vma,
 994                                         unsigned long address, pmd_t *pmd,
 995                                         int referenced)
 996 {
 997         int swapped_in = 0;
 998         vm_fault_t ret = 0;
 999         struct vm_fault vmf = {
1000                 .vma = vma,
1001                 .address = address,
1002                 .flags = FAULT_FLAG_ALLOW_RETRY,
1003                 .pmd = pmd,
1004                 .pgoff = linear_page_index(vma, address),
1005         };
1006
1007         vmf.pte = pte_offset_map(pmd, address);
1008         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1009                         vmf.pte++, vmf.address += PAGE_SIZE) {
1010                 vmf.orig_pte = *vmf.pte;
1011                 if (!is_swap_pte(vmf.orig_pte))
1012                         continue;
1013                 swapped_in++;
1014                 ret = do_swap_page(&vmf);
1015
1016                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1017                 if (ret & VM_FAULT_RETRY) {
1018                         mmap_read_lock(mm);
1019                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1020                                 /* vma is no longer available, don't continue to swapin */
1021                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1022                                 return false;
1023                         }
1024                         /* check if the pmd is still valid */
1025                         if (mm_find_pmd(mm, address) != pmd) {
1026                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1027                                 return false;
1028                         }
1029                 }
1030                 if (ret & VM_FAULT_ERROR) {
1031                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1032                         return false;
1033                 }
1034                 /* pte is unmapped now, we need to map it */
1035                 vmf.pte = pte_offset_map(pmd, vmf.address);
1036         }
1037         vmf.pte--;
1038         pte_unmap(vmf.pte);
1039
1040         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1041         if (swapped_in)
1042                 lru_add_drain();
1043
1044         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1045         return true;
1046 }
1047
1048 static void collapse_huge_page(struct mm_struct *mm,
1049                                    unsigned long address,
1050                                    struct page **hpage,
1051                                    int node, int referenced, int unmapped)
1052 {
1053         LIST_HEAD(compound_pagelist);
1054         pmd_t *pmd, _pmd;
1055         pte_t *pte;
1056         pgtable_t pgtable;
1057         struct page *new_page;
1058         spinlock_t *pmd_ptl, *pte_ptl;
1059         int isolated = 0, result = 0;
1060         struct vm_area_struct *vma;
1061         struct mmu_notifier_range range;
1062         gfp_t gfp;
1063
1064         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1065
1066         /* Only allocate from the target node */
1067         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1068
1069         /*
1070          * Before allocating the hugepage, release the mmap_lock read lock.
1071          * The allocation can take potentially a long time if it involves
1072          * sync compaction, and we do not need to hold the mmap_lock during
1073          * that. We will recheck the vma after taking it again in write mode.
1074          */
1075         mmap_read_unlock(mm);
1076         new_page = khugepaged_alloc_page(hpage, gfp, node);
1077         if (!new_page) {
1078                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1079                 goto out_nolock;
1080         }
1081
1082         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1083                 result = SCAN_CGROUP_CHARGE_FAIL;
1084                 goto out_nolock;
1085         }
1086         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1087
1088         mmap_read_lock(mm);
1089         result = hugepage_vma_revalidate(mm, address, &vma);
1090         if (result) {
1091                 mmap_read_unlock(mm);
1092                 goto out_nolock;
1093         }
1094
1095         pmd = mm_find_pmd(mm, address);
1096         if (!pmd) {
1097                 result = SCAN_PMD_NULL;
1098                 mmap_read_unlock(mm);
1099                 goto out_nolock;
1100         }
1101
1102         /*
1103          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1104          * If it fails, we release mmap_lock and jump out_nolock.
1105          * Continuing to collapse causes inconsistency.
1106          */
1107         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1108                                                      pmd, referenced)) {
1109                 mmap_read_unlock(mm);
1110                 goto out_nolock;
1111         }
1112
1113         mmap_read_unlock(mm);
1114         /*
1115          * Prevent all access to pagetables with the exception of
1116          * gup_fast later handled by the ptep_clear_flush and the VM
1117          * handled by the anon_vma lock + PG_lock.
1118          */
1119         mmap_write_lock(mm);
1120         result = hugepage_vma_revalidate(mm, address, &vma);
1121         if (result)
1122                 goto out;
1123         /* check if the pmd is still valid */
1124         if (mm_find_pmd(mm, address) != pmd)
1125                 goto out;
1126
1127         anon_vma_lock_write(vma->anon_vma);
1128
1129         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1130                                 address, address + HPAGE_PMD_SIZE);
1131         mmu_notifier_invalidate_range_start(&range);
1132
1133         pte = pte_offset_map(pmd, address);
1134         pte_ptl = pte_lockptr(mm, pmd);
1135
1136         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1137         /*
1138          * After this gup_fast can't run anymore. This also removes
1139          * any huge TLB entry from the CPU so we won't allow
1140          * huge and small TLB entries for the same virtual address
1141          * to avoid the risk of CPU bugs in that area.
1142          */
1143         _pmd = pmdp_collapse_flush(vma, address, pmd);
1144         spin_unlock(pmd_ptl);
1145         mmu_notifier_invalidate_range_end(&range);
1146
1147         spin_lock(pte_ptl);
1148         isolated = __collapse_huge_page_isolate(vma, address, pte,
1149                         &compound_pagelist);
1150         spin_unlock(pte_ptl);
1151
1152         if (unlikely(!isolated)) {
1153                 pte_unmap(pte);
1154                 spin_lock(pmd_ptl);
1155                 BUG_ON(!pmd_none(*pmd));
1156                 /*
1157                  * We can only use set_pmd_at when establishing
1158                  * hugepmds and never for establishing regular pmds that
1159                  * points to regular pagetables. Use pmd_populate for that
1160                  */
1161                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1162                 spin_unlock(pmd_ptl);
1163                 anon_vma_unlock_write(vma->anon_vma);
1164                 result = SCAN_FAIL;
1165                 goto out;
1166         }
1167
1168         /*
1169          * All pages are isolated and locked so anon_vma rmap
1170          * can't run anymore.
1171          */
1172         anon_vma_unlock_write(vma->anon_vma);
1173
1174         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1175                         &compound_pagelist);
1176         pte_unmap(pte);
1177         __SetPageUptodate(new_page);
1178         pgtable = pmd_pgtable(_pmd);
1179
1180         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1181         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1182
1183         /*
1184          * spin_lock() below is not the equivalent of smp_wmb(), so
1185          * this is needed to avoid the copy_huge_page writes to become
1186          * visible after the set_pmd_at() write.
1187          */
1188         smp_wmb();
1189
1190         spin_lock(pmd_ptl);
1191         BUG_ON(!pmd_none(*pmd));
1192         page_add_new_anon_rmap(new_page, vma, address, true);
1193         lru_cache_add_inactive_or_unevictable(new_page, vma);
1194         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1195         set_pmd_at(mm, address, pmd, _pmd);
1196         update_mmu_cache_pmd(vma, address, pmd);
1197         spin_unlock(pmd_ptl);
1198
1199         *hpage = NULL;
1200
1201         khugepaged_pages_collapsed++;
1202         result = SCAN_SUCCEED;
1203 out_up_write:
1204         mmap_write_unlock(mm);
1205 out_nolock:
1206         if (!IS_ERR_OR_NULL(*hpage))
1207                 mem_cgroup_uncharge(*hpage);
1208         trace_mm_collapse_huge_page(mm, isolated, result);
1209         return;
1210 out:
1211         goto out_up_write;
1212 }
1213
1214 static int khugepaged_scan_pmd(struct mm_struct *mm,
1215                                struct vm_area_struct *vma,
1216                                unsigned long address,
1217                                struct page **hpage)
1218 {
1219         pmd_t *pmd;
1220         pte_t *pte, *_pte;
1221         int ret = 0, result = 0, referenced = 0;
1222         int none_or_zero = 0, shared = 0;
1223         struct page *page = NULL;
1224         unsigned long _address;
1225         spinlock_t *ptl;
1226         int node = NUMA_NO_NODE, unmapped = 0;
1227         bool writable = false;
1228
1229         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1230
1231         pmd = mm_find_pmd(mm, address);
1232         if (!pmd) {
1233                 result = SCAN_PMD_NULL;
1234                 goto out;
1235         }
1236
1237         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1238         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1239         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1240              _pte++, _address += PAGE_SIZE) {
1241                 pte_t pteval = *_pte;
1242                 if (is_swap_pte(pteval)) {
1243                         if (++unmapped <= khugepaged_max_ptes_swap) {
1244                                 /*
1245                                  * Always be strict with uffd-wp
1246                                  * enabled swap entries.  Please see
1247                                  * comment below for pte_uffd_wp().
1248                                  */
1249                                 if (pte_swp_uffd_wp(pteval)) {
1250                                         result = SCAN_PTE_UFFD_WP;
1251                                         goto out_unmap;
1252                                 }
1253                                 continue;
1254                         } else {
1255                                 result = SCAN_EXCEED_SWAP_PTE;
1256                                 goto out_unmap;
1257                         }
1258                 }
1259                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1260                         if (!userfaultfd_armed(vma) &&
1261                             ++none_or_zero <= khugepaged_max_ptes_none) {
1262                                 continue;
1263                         } else {
1264                                 result = SCAN_EXCEED_NONE_PTE;
1265                                 goto out_unmap;
1266                         }
1267                 }
1268                 if (!pte_present(pteval)) {
1269                         result = SCAN_PTE_NON_PRESENT;
1270                         goto out_unmap;
1271                 }
1272                 if (pte_uffd_wp(pteval)) {
1273                         /*
1274                          * Don't collapse the page if any of the small
1275                          * PTEs are armed with uffd write protection.
1276                          * Here we can also mark the new huge pmd as
1277                          * write protected if any of the small ones is
1278                          * marked but that could bring unknown
1279                          * userfault messages that falls outside of
1280                          * the registered range.  So, just be simple.
1281                          */
1282                         result = SCAN_PTE_UFFD_WP;
1283                         goto out_unmap;
1284                 }
1285                 if (pte_write(pteval))
1286                         writable = true;
1287
1288                 page = vm_normal_page(vma, _address, pteval);
1289                 if (unlikely(!page)) {
1290                         result = SCAN_PAGE_NULL;
1291                         goto out_unmap;
1292                 }
1293
1294                 if (page_mapcount(page) > 1 &&
1295                                 ++shared > khugepaged_max_ptes_shared) {
1296                         result = SCAN_EXCEED_SHARED_PTE;
1297                         goto out_unmap;
1298                 }
1299
1300                 page = compound_head(page);
1301
1302                 /*
1303                  * Record which node the original page is from and save this
1304                  * information to khugepaged_node_load[].
1305                  * Khupaged will allocate hugepage from the node has the max
1306                  * hit record.
1307                  */
1308                 node = page_to_nid(page);
1309                 if (khugepaged_scan_abort(node)) {
1310                         result = SCAN_SCAN_ABORT;
1311                         goto out_unmap;
1312                 }
1313                 khugepaged_node_load[node]++;
1314                 if (!PageLRU(page)) {
1315                         result = SCAN_PAGE_LRU;
1316                         goto out_unmap;
1317                 }
1318                 if (PageLocked(page)) {
1319                         result = SCAN_PAGE_LOCK;
1320                         goto out_unmap;
1321                 }
1322                 if (!PageAnon(page)) {
1323                         result = SCAN_PAGE_ANON;
1324                         goto out_unmap;
1325                 }
1326
1327                 /*
1328                  * Check if the page has any GUP (or other external) pins.
1329                  *
1330                  * Here the check is racy it may see totmal_mapcount > refcount
1331                  * in some cases.
1332                  * For example, one process with one forked child process.
1333                  * The parent has the PMD split due to MADV_DONTNEED, then
1334                  * the child is trying unmap the whole PMD, but khugepaged
1335                  * may be scanning the parent between the child has
1336                  * PageDoubleMap flag cleared and dec the mapcount.  So
1337                  * khugepaged may see total_mapcount > refcount.
1338                  *
1339                  * But such case is ephemeral we could always retry collapse
1340                  * later.  However it may report false positive if the page
1341                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1342                  * will be done again later the risk seems low.
1343                  */
1344                 if (!is_refcount_suitable(page)) {
1345                         result = SCAN_PAGE_COUNT;
1346                         goto out_unmap;
1347                 }
1348                 if (pte_young(pteval) ||
1349                     page_is_young(page) || PageReferenced(page) ||
1350                     mmu_notifier_test_young(vma->vm_mm, address))
1351                         referenced++;
1352         }
1353         if (!writable) {
1354                 result = SCAN_PAGE_RO;
1355         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1356                 result = SCAN_LACK_REFERENCED_PAGE;
1357         } else {
1358                 result = SCAN_SUCCEED;
1359                 ret = 1;
1360         }
1361 out_unmap:
1362         pte_unmap_unlock(pte, ptl);
1363         if (ret) {
1364                 node = khugepaged_find_target_node();
1365                 /* collapse_huge_page will return with the mmap_lock released */
1366                 collapse_huge_page(mm, address, hpage, node,
1367                                 referenced, unmapped);
1368         }
1369 out:
1370         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1371                                      none_or_zero, result, unmapped);
1372         return ret;
1373 }
1374
1375 static void collect_mm_slot(struct mm_slot *mm_slot)
1376 {
1377         struct mm_struct *mm = mm_slot->mm;
1378
1379         lockdep_assert_held(&khugepaged_mm_lock);
1380
1381         if (khugepaged_test_exit(mm)) {
1382                 /* free mm_slot */
1383                 hash_del(&mm_slot->hash);
1384                 list_del(&mm_slot->mm_node);
1385
1386                 /*
1387                  * Not strictly needed because the mm exited already.
1388                  *
1389                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1390                  */
1391
1392                 /* khugepaged_mm_lock actually not necessary for the below */
1393                 free_mm_slot(mm_slot);
1394                 mmdrop(mm);
1395         }
1396 }
1397
1398 #ifdef CONFIG_SHMEM
1399 /*
1400  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1401  * khugepaged should try to collapse the page table.
1402  */
1403 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1404                                          unsigned long addr)
1405 {
1406         struct mm_slot *mm_slot;
1407
1408         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1409
1410         spin_lock(&khugepaged_mm_lock);
1411         mm_slot = get_mm_slot(mm);
1412         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1413                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1414         spin_unlock(&khugepaged_mm_lock);
1415         return 0;
1416 }
1417
1418 /**
1419  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1420  * address haddr.
1421  *
1422  * @mm: process address space where collapse happens
1423  * @addr: THP collapse address
1424  *
1425  * This function checks whether all the PTEs in the PMD are pointing to the
1426  * right THP. If so, retract the page table so the THP can refault in with
1427  * as pmd-mapped.
1428  */
1429 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1430 {
1431         unsigned long haddr = addr & HPAGE_PMD_MASK;
1432         struct vm_area_struct *vma = find_vma(mm, haddr);
1433         struct page *hpage;
1434         pte_t *start_pte, *pte;
1435         pmd_t *pmd, _pmd;
1436         spinlock_t *ptl;
1437         int count = 0;
1438         int i;
1439
1440         if (!vma || !vma->vm_file ||
1441             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1442                 return;
1443
1444         /*
1445          * This vm_flags may not have VM_HUGEPAGE if the page was not
1446          * collapsed by this mm. But we can still collapse if the page is
1447          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1448          * will not fail the vma for missing VM_HUGEPAGE
1449          */
1450         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1451                 return;
1452
1453         hpage = find_lock_page(vma->vm_file->f_mapping,
1454                                linear_page_index(vma, haddr));
1455         if (!hpage)
1456                 return;
1457
1458         if (!PageHead(hpage))
1459                 goto drop_hpage;
1460
1461         pmd = mm_find_pmd(mm, haddr);
1462         if (!pmd)
1463                 goto drop_hpage;
1464
1465         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1466
1467         /* step 1: check all mapped PTEs are to the right huge page */
1468         for (i = 0, addr = haddr, pte = start_pte;
1469              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1470                 struct page *page;
1471
1472                 /* empty pte, skip */
1473                 if (pte_none(*pte))
1474                         continue;
1475
1476                 /* page swapped out, abort */
1477                 if (!pte_present(*pte))
1478                         goto abort;
1479
1480                 page = vm_normal_page(vma, addr, *pte);
1481
1482                 /*
1483                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1484                  * page table, but the new page will not be a subpage of hpage.
1485                  */
1486                 if (hpage + i != page)
1487                         goto abort;
1488                 count++;
1489         }
1490
1491         /* step 2: adjust rmap */
1492         for (i = 0, addr = haddr, pte = start_pte;
1493              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1494                 struct page *page;
1495
1496                 if (pte_none(*pte))
1497                         continue;
1498                 page = vm_normal_page(vma, addr, *pte);
1499                 page_remove_rmap(page, false);
1500         }
1501
1502         pte_unmap_unlock(start_pte, ptl);
1503
1504         /* step 3: set proper refcount and mm_counters. */
1505         if (count) {
1506                 page_ref_sub(hpage, count);
1507                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1508         }
1509
1510         /* step 4: collapse pmd */
1511         ptl = pmd_lock(vma->vm_mm, pmd);
1512         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1513         spin_unlock(ptl);
1514         mm_dec_nr_ptes(mm);
1515         pte_free(mm, pmd_pgtable(_pmd));
1516
1517 drop_hpage:
1518         unlock_page(hpage);
1519         put_page(hpage);
1520         return;
1521
1522 abort:
1523         pte_unmap_unlock(start_pte, ptl);
1524         goto drop_hpage;
1525 }
1526
1527 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1528 {
1529         struct mm_struct *mm = mm_slot->mm;
1530         int i;
1531
1532         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1533                 return 0;
1534
1535         if (!mmap_write_trylock(mm))
1536                 return -EBUSY;
1537
1538         if (unlikely(khugepaged_test_exit(mm)))
1539                 goto out;
1540
1541         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1542                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1543
1544 out:
1545         mm_slot->nr_pte_mapped_thp = 0;
1546         mmap_write_unlock(mm);
1547         return 0;
1548 }
1549
1550 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1551 {
1552         struct vm_area_struct *vma;
1553         struct mm_struct *mm;
1554         unsigned long addr;
1555         pmd_t *pmd, _pmd;
1556
1557         i_mmap_lock_write(mapping);
1558         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1559                 /*
1560                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1561                  * got written to. These VMAs are likely not worth investing
1562                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1563                  * later.
1564                  *
1565                  * Not that vma->anon_vma check is racy: it can be set up after
1566                  * the check but before we took mmap_lock by the fault path.
1567                  * But page lock would prevent establishing any new ptes of the
1568                  * page, so we are safe.
1569                  *
1570                  * An alternative would be drop the check, but check that page
1571                  * table is clear before calling pmdp_collapse_flush() under
1572                  * ptl. It has higher chance to recover THP for the VMA, but
1573                  * has higher cost too.
1574                  */
1575                 if (vma->anon_vma)
1576                         continue;
1577                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1578                 if (addr & ~HPAGE_PMD_MASK)
1579                         continue;
1580                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1581                         continue;
1582                 mm = vma->vm_mm;
1583                 pmd = mm_find_pmd(mm, addr);
1584                 if (!pmd)
1585                         continue;
1586                 /*
1587                  * We need exclusive mmap_lock to retract page table.
1588                  *
1589                  * We use trylock due to lock inversion: we need to acquire
1590                  * mmap_lock while holding page lock. Fault path does it in
1591                  * reverse order. Trylock is a way to avoid deadlock.
1592                  */
1593                 if (mmap_write_trylock(mm)) {
1594                         if (!khugepaged_test_exit(mm)) {
1595                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1596                                 /* assume page table is clear */
1597                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1598                                 spin_unlock(ptl);
1599                                 mm_dec_nr_ptes(mm);
1600                                 pte_free(mm, pmd_pgtable(_pmd));
1601                         }
1602                         mmap_write_unlock(mm);
1603                 } else {
1604                         /* Try again later */
1605                         khugepaged_add_pte_mapped_thp(mm, addr);
1606                 }
1607         }
1608         i_mmap_unlock_write(mapping);
1609 }
1610
1611 /**
1612  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1613  *
1614  * @mm: process address space where collapse happens
1615  * @file: file that collapse on
1616  * @start: collapse start address
1617  * @hpage: new allocated huge page for collapse
1618  * @node: appointed node the new huge page allocate from
1619  *
1620  * Basic scheme is simple, details are more complex:
1621  *  - allocate and lock a new huge page;
1622  *  - scan page cache replacing old pages with the new one
1623  *    + swap/gup in pages if necessary;
1624  *    + fill in gaps;
1625  *    + keep old pages around in case rollback is required;
1626  *  - if replacing succeeds:
1627  *    + copy data over;
1628  *    + free old pages;
1629  *    + unlock huge page;
1630  *  - if replacing failed;
1631  *    + put all pages back and unfreeze them;
1632  *    + restore gaps in the page cache;
1633  *    + unlock and free huge page;
1634  */
1635 static void collapse_file(struct mm_struct *mm,
1636                 struct file *file, pgoff_t start,
1637                 struct page **hpage, int node)
1638 {
1639         struct address_space *mapping = file->f_mapping;
1640         gfp_t gfp;
1641         struct page *new_page;
1642         pgoff_t index, end = start + HPAGE_PMD_NR;
1643         LIST_HEAD(pagelist);
1644         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1645         int nr_none = 0, result = SCAN_SUCCEED;
1646         bool is_shmem = shmem_file(file);
1647
1648         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1649         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1650
1651         /* Only allocate from the target node */
1652         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1653
1654         new_page = khugepaged_alloc_page(hpage, gfp, node);
1655         if (!new_page) {
1656                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1657                 goto out;
1658         }
1659
1660         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1661                 result = SCAN_CGROUP_CHARGE_FAIL;
1662                 goto out;
1663         }
1664         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1665
1666         /* This will be less messy when we use multi-index entries */
1667         do {
1668                 xas_lock_irq(&xas);
1669                 xas_create_range(&xas);
1670                 if (!xas_error(&xas))
1671                         break;
1672                 xas_unlock_irq(&xas);
1673                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1674                         result = SCAN_FAIL;
1675                         goto out;
1676                 }
1677         } while (1);
1678
1679         __SetPageLocked(new_page);
1680         if (is_shmem)
1681                 __SetPageSwapBacked(new_page);
1682         new_page->index = start;
1683         new_page->mapping = mapping;
1684
1685         /*
1686          * At this point the new_page is locked and not up-to-date.
1687          * It's safe to insert it into the page cache, because nobody would
1688          * be able to map it or use it in another way until we unlock it.
1689          */
1690
1691         xas_set(&xas, start);
1692         for (index = start; index < end; index++) {
1693                 struct page *page = xas_next(&xas);
1694
1695                 VM_BUG_ON(index != xas.xa_index);
1696                 if (is_shmem) {
1697                         if (!page) {
1698                                 /*
1699                                  * Stop if extent has been truncated or
1700                                  * hole-punched, and is now completely
1701                                  * empty.
1702                                  */
1703                                 if (index == start) {
1704                                         if (!xas_next_entry(&xas, end - 1)) {
1705                                                 result = SCAN_TRUNCATED;
1706                                                 goto xa_locked;
1707                                         }
1708                                         xas_set(&xas, index);
1709                                 }
1710                                 if (!shmem_charge(mapping->host, 1)) {
1711                                         result = SCAN_FAIL;
1712                                         goto xa_locked;
1713                                 }
1714                                 xas_store(&xas, new_page);
1715                                 nr_none++;
1716                                 continue;
1717                         }
1718
1719                         if (xa_is_value(page) || !PageUptodate(page)) {
1720                                 xas_unlock_irq(&xas);
1721                                 /* swap in or instantiate fallocated page */
1722                                 if (shmem_getpage(mapping->host, index, &page,
1723                                                   SGP_NOHUGE)) {
1724                                         result = SCAN_FAIL;
1725                                         goto xa_unlocked;
1726                                 }
1727                         } else if (trylock_page(page)) {
1728                                 get_page(page);
1729                                 xas_unlock_irq(&xas);
1730                         } else {
1731                                 result = SCAN_PAGE_LOCK;
1732                                 goto xa_locked;
1733                         }
1734                 } else {        /* !is_shmem */
1735                         if (!page || xa_is_value(page)) {
1736                                 xas_unlock_irq(&xas);
1737                                 page_cache_sync_readahead(mapping, &file->f_ra,
1738                                                           file, index,
1739                                                           end - index);
1740                                 /* drain pagevecs to help isolate_lru_page() */
1741                                 lru_add_drain();
1742                                 page = find_lock_page(mapping, index);
1743                                 if (unlikely(page == NULL)) {
1744                                         result = SCAN_FAIL;
1745                                         goto xa_unlocked;
1746                                 }
1747                         } else if (PageDirty(page)) {
1748                                 /*
1749                                  * khugepaged only works on read-only fd,
1750                                  * so this page is dirty because it hasn't
1751                                  * been flushed since first write. There
1752                                  * won't be new dirty pages.
1753                                  *
1754                                  * Trigger async flush here and hope the
1755                                  * writeback is done when khugepaged
1756                                  * revisits this page.
1757                                  *
1758                                  * This is a one-off situation. We are not
1759                                  * forcing writeback in loop.
1760                                  */
1761                                 xas_unlock_irq(&xas);
1762                                 filemap_flush(mapping);
1763                                 result = SCAN_FAIL;
1764                                 goto xa_unlocked;
1765                         } else if (trylock_page(page)) {
1766                                 get_page(page);
1767                                 xas_unlock_irq(&xas);
1768                         } else {
1769                                 result = SCAN_PAGE_LOCK;
1770                                 goto xa_locked;
1771                         }
1772                 }
1773
1774                 /*
1775                  * The page must be locked, so we can drop the i_pages lock
1776                  * without racing with truncate.
1777                  */
1778                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1779
1780                 /* make sure the page is up to date */
1781                 if (unlikely(!PageUptodate(page))) {
1782                         result = SCAN_FAIL;
1783                         goto out_unlock;
1784                 }
1785
1786                 /*
1787                  * If file was truncated then extended, or hole-punched, before
1788                  * we locked the first page, then a THP might be there already.
1789                  */
1790                 if (PageTransCompound(page)) {
1791                         result = SCAN_PAGE_COMPOUND;
1792                         goto out_unlock;
1793                 }
1794
1795                 if (page_mapping(page) != mapping) {
1796                         result = SCAN_TRUNCATED;
1797                         goto out_unlock;
1798                 }
1799
1800                 if (!is_shmem && PageDirty(page)) {
1801                         /*
1802                          * khugepaged only works on read-only fd, so this
1803                          * page is dirty because it hasn't been flushed
1804                          * since first write.
1805                          */
1806                         result = SCAN_FAIL;
1807                         goto out_unlock;
1808                 }
1809
1810                 if (isolate_lru_page(page)) {
1811                         result = SCAN_DEL_PAGE_LRU;
1812                         goto out_unlock;
1813                 }
1814
1815                 if (page_has_private(page) &&
1816                     !try_to_release_page(page, GFP_KERNEL)) {
1817                         result = SCAN_PAGE_HAS_PRIVATE;
1818                         putback_lru_page(page);
1819                         goto out_unlock;
1820                 }
1821
1822                 if (page_mapped(page))
1823                         unmap_mapping_pages(mapping, index, 1, false);
1824
1825                 xas_lock_irq(&xas);
1826                 xas_set(&xas, index);
1827
1828                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1829                 VM_BUG_ON_PAGE(page_mapped(page), page);
1830
1831                 /*
1832                  * The page is expected to have page_count() == 3:
1833                  *  - we hold a pin on it;
1834                  *  - one reference from page cache;
1835                  *  - one from isolate_lru_page;
1836                  */
1837                 if (!page_ref_freeze(page, 3)) {
1838                         result = SCAN_PAGE_COUNT;
1839                         xas_unlock_irq(&xas);
1840                         putback_lru_page(page);
1841                         goto out_unlock;
1842                 }
1843
1844                 /*
1845                  * Add the page to the list to be able to undo the collapse if
1846                  * something go wrong.
1847                  */
1848                 list_add_tail(&page->lru, &pagelist);
1849
1850                 /* Finally, replace with the new page. */
1851                 xas_store(&xas, new_page);
1852                 continue;
1853 out_unlock:
1854                 unlock_page(page);
1855                 put_page(page);
1856                 goto xa_unlocked;
1857         }
1858
1859         if (is_shmem)
1860                 __inc_lruvec_page_state(new_page, NR_SHMEM_THPS);
1861         else {
1862                 __inc_lruvec_page_state(new_page, NR_FILE_THPS);
1863                 filemap_nr_thps_inc(mapping);
1864         }
1865
1866         if (nr_none) {
1867                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1868                 if (is_shmem)
1869                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1870         }
1871
1872 xa_locked:
1873         xas_unlock_irq(&xas);
1874 xa_unlocked:
1875
1876         if (result == SCAN_SUCCEED) {
1877                 struct page *page, *tmp;
1878
1879                 /*
1880                  * Replacing old pages with new one has succeeded, now we
1881                  * need to copy the content and free the old pages.
1882                  */
1883                 index = start;
1884                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1885                         while (index < page->index) {
1886                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1887                                 index++;
1888                         }
1889                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1890                                         page);
1891                         list_del(&page->lru);
1892                         page->mapping = NULL;
1893                         page_ref_unfreeze(page, 1);
1894                         ClearPageActive(page);
1895                         ClearPageUnevictable(page);
1896                         unlock_page(page);
1897                         put_page(page);
1898                         index++;
1899                 }
1900                 while (index < end) {
1901                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1902                         index++;
1903                 }
1904
1905                 SetPageUptodate(new_page);
1906                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1907                 if (is_shmem)
1908                         set_page_dirty(new_page);
1909                 lru_cache_add(new_page);
1910
1911                 /*
1912                  * Remove pte page tables, so we can re-fault the page as huge.
1913                  */
1914                 retract_page_tables(mapping, start);
1915                 *hpage = NULL;
1916
1917                 khugepaged_pages_collapsed++;
1918         } else {
1919                 struct page *page;
1920
1921                 /* Something went wrong: roll back page cache changes */
1922                 xas_lock_irq(&xas);
1923                 mapping->nrpages -= nr_none;
1924
1925                 if (is_shmem)
1926                         shmem_uncharge(mapping->host, nr_none);
1927
1928                 xas_set(&xas, start);
1929                 xas_for_each(&xas, page, end - 1) {
1930                         page = list_first_entry_or_null(&pagelist,
1931                                         struct page, lru);
1932                         if (!page || xas.xa_index < page->index) {
1933                                 if (!nr_none)
1934                                         break;
1935                                 nr_none--;
1936                                 /* Put holes back where they were */
1937                                 xas_store(&xas, NULL);
1938                                 continue;
1939                         }
1940
1941                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1942
1943                         /* Unfreeze the page. */
1944                         list_del(&page->lru);
1945                         page_ref_unfreeze(page, 2);
1946                         xas_store(&xas, page);
1947                         xas_pause(&xas);
1948                         xas_unlock_irq(&xas);
1949                         unlock_page(page);
1950                         putback_lru_page(page);
1951                         xas_lock_irq(&xas);
1952                 }
1953                 VM_BUG_ON(nr_none);
1954                 xas_unlock_irq(&xas);
1955
1956                 new_page->mapping = NULL;
1957         }
1958
1959         unlock_page(new_page);
1960 out:
1961         VM_BUG_ON(!list_empty(&pagelist));
1962         if (!IS_ERR_OR_NULL(*hpage))
1963                 mem_cgroup_uncharge(*hpage);
1964         /* TODO: tracepoints */
1965 }
1966
1967 static void khugepaged_scan_file(struct mm_struct *mm,
1968                 struct file *file, pgoff_t start, struct page **hpage)
1969 {
1970         struct page *page = NULL;
1971         struct address_space *mapping = file->f_mapping;
1972         XA_STATE(xas, &mapping->i_pages, start);
1973         int present, swap;
1974         int node = NUMA_NO_NODE;
1975         int result = SCAN_SUCCEED;
1976
1977         present = 0;
1978         swap = 0;
1979         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1980         rcu_read_lock();
1981         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1982                 if (xas_retry(&xas, page))
1983                         continue;
1984
1985                 if (xa_is_value(page)) {
1986                         if (++swap > khugepaged_max_ptes_swap) {
1987                                 result = SCAN_EXCEED_SWAP_PTE;
1988                                 break;
1989                         }
1990                         continue;
1991                 }
1992
1993                 if (PageTransCompound(page)) {
1994                         result = SCAN_PAGE_COMPOUND;
1995                         break;
1996                 }
1997
1998                 node = page_to_nid(page);
1999                 if (khugepaged_scan_abort(node)) {
2000                         result = SCAN_SCAN_ABORT;
2001                         break;
2002                 }
2003                 khugepaged_node_load[node]++;
2004
2005                 if (!PageLRU(page)) {
2006                         result = SCAN_PAGE_LRU;
2007                         break;
2008                 }
2009
2010                 if (page_count(page) !=
2011                     1 + page_mapcount(page) + page_has_private(page)) {
2012                         result = SCAN_PAGE_COUNT;
2013                         break;
2014                 }
2015
2016                 /*
2017                  * We probably should check if the page is referenced here, but
2018                  * nobody would transfer pte_young() to PageReferenced() for us.
2019                  * And rmap walk here is just too costly...
2020                  */
2021
2022                 present++;
2023
2024                 if (need_resched()) {
2025                         xas_pause(&xas);
2026                         cond_resched_rcu();
2027                 }
2028         }
2029         rcu_read_unlock();
2030
2031         if (result == SCAN_SUCCEED) {
2032                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2033                         result = SCAN_EXCEED_NONE_PTE;
2034                 } else {
2035                         node = khugepaged_find_target_node();
2036                         collapse_file(mm, file, start, hpage, node);
2037                 }
2038         }
2039
2040         /* TODO: tracepoints */
2041 }
2042 #else
2043 static void khugepaged_scan_file(struct mm_struct *mm,
2044                 struct file *file, pgoff_t start, struct page **hpage)
2045 {
2046         BUILD_BUG();
2047 }
2048
2049 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2050 {
2051         return 0;
2052 }
2053 #endif
2054
2055 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2056                                             struct page **hpage)
2057         __releases(&khugepaged_mm_lock)
2058         __acquires(&khugepaged_mm_lock)
2059 {
2060         struct mm_slot *mm_slot;
2061         struct mm_struct *mm;
2062         struct vm_area_struct *vma;
2063         int progress = 0;
2064
2065         VM_BUG_ON(!pages);
2066         lockdep_assert_held(&khugepaged_mm_lock);
2067
2068         if (khugepaged_scan.mm_slot)
2069                 mm_slot = khugepaged_scan.mm_slot;
2070         else {
2071                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2072                                      struct mm_slot, mm_node);
2073                 khugepaged_scan.address = 0;
2074                 khugepaged_scan.mm_slot = mm_slot;
2075         }
2076         spin_unlock(&khugepaged_mm_lock);
2077         khugepaged_collapse_pte_mapped_thps(mm_slot);
2078
2079         mm = mm_slot->mm;
2080         /*
2081          * Don't wait for semaphore (to avoid long wait times).  Just move to
2082          * the next mm on the list.
2083          */
2084         vma = NULL;
2085         if (unlikely(!mmap_read_trylock(mm)))
2086                 goto breakouterloop_mmap_lock;
2087         if (likely(!khugepaged_test_exit(mm)))
2088                 vma = find_vma(mm, khugepaged_scan.address);
2089
2090         progress++;
2091         for (; vma; vma = vma->vm_next) {
2092                 unsigned long hstart, hend;
2093
2094                 cond_resched();
2095                 if (unlikely(khugepaged_test_exit(mm))) {
2096                         progress++;
2097                         break;
2098                 }
2099                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2100 skip:
2101                         progress++;
2102                         continue;
2103                 }
2104                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2105                 hend = vma->vm_end & HPAGE_PMD_MASK;
2106                 if (hstart >= hend)
2107                         goto skip;
2108                 if (khugepaged_scan.address > hend)
2109                         goto skip;
2110                 if (khugepaged_scan.address < hstart)
2111                         khugepaged_scan.address = hstart;
2112                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2113                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2114                         goto skip;
2115
2116                 while (khugepaged_scan.address < hend) {
2117                         int ret;
2118                         cond_resched();
2119                         if (unlikely(khugepaged_test_exit(mm)))
2120                                 goto breakouterloop;
2121
2122                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2123                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2124                                   hend);
2125                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2126                                 struct file *file = get_file(vma->vm_file);
2127                                 pgoff_t pgoff = linear_page_index(vma,
2128                                                 khugepaged_scan.address);
2129
2130                                 mmap_read_unlock(mm);
2131                                 ret = 1;
2132                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2133                                 fput(file);
2134                         } else {
2135                                 ret = khugepaged_scan_pmd(mm, vma,
2136                                                 khugepaged_scan.address,
2137                                                 hpage);
2138                         }
2139                         /* move to next address */
2140                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2141                         progress += HPAGE_PMD_NR;
2142                         if (ret)
2143                                 /* we released mmap_lock so break loop */
2144                                 goto breakouterloop_mmap_lock;
2145                         if (progress >= pages)
2146                                 goto breakouterloop;
2147                 }
2148         }
2149 breakouterloop:
2150         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2151 breakouterloop_mmap_lock:
2152
2153         spin_lock(&khugepaged_mm_lock);
2154         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2155         /*
2156          * Release the current mm_slot if this mm is about to die, or
2157          * if we scanned all vmas of this mm.
2158          */
2159         if (khugepaged_test_exit(mm) || !vma) {
2160                 /*
2161                  * Make sure that if mm_users is reaching zero while
2162                  * khugepaged runs here, khugepaged_exit will find
2163                  * mm_slot not pointing to the exiting mm.
2164                  */
2165                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2166                         khugepaged_scan.mm_slot = list_entry(
2167                                 mm_slot->mm_node.next,
2168                                 struct mm_slot, mm_node);
2169                         khugepaged_scan.address = 0;
2170                 } else {
2171                         khugepaged_scan.mm_slot = NULL;
2172                         khugepaged_full_scans++;
2173                 }
2174
2175                 collect_mm_slot(mm_slot);
2176         }
2177
2178         return progress;
2179 }
2180
2181 static int khugepaged_has_work(void)
2182 {
2183         return !list_empty(&khugepaged_scan.mm_head) &&
2184                 khugepaged_enabled();
2185 }
2186
2187 static int khugepaged_wait_event(void)
2188 {
2189         return !list_empty(&khugepaged_scan.mm_head) ||
2190                 kthread_should_stop();
2191 }
2192
2193 static void khugepaged_do_scan(void)
2194 {
2195         struct page *hpage = NULL;
2196         unsigned int progress = 0, pass_through_head = 0;
2197         unsigned int pages = khugepaged_pages_to_scan;
2198         bool wait = true;
2199
2200         barrier(); /* write khugepaged_pages_to_scan to local stack */
2201
2202         lru_add_drain_all();
2203
2204         while (progress < pages) {
2205                 if (!khugepaged_prealloc_page(&hpage, &wait))
2206                         break;
2207
2208                 cond_resched();
2209
2210                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2211                         break;
2212
2213                 spin_lock(&khugepaged_mm_lock);
2214                 if (!khugepaged_scan.mm_slot)
2215                         pass_through_head++;
2216                 if (khugepaged_has_work() &&
2217                     pass_through_head < 2)
2218                         progress += khugepaged_scan_mm_slot(pages - progress,
2219                                                             &hpage);
2220                 else
2221                         progress = pages;
2222                 spin_unlock(&khugepaged_mm_lock);
2223         }
2224
2225         if (!IS_ERR_OR_NULL(hpage))
2226                 put_page(hpage);
2227 }
2228
2229 static bool khugepaged_should_wakeup(void)
2230 {
2231         return kthread_should_stop() ||
2232                time_after_eq(jiffies, khugepaged_sleep_expire);
2233 }
2234
2235 static void khugepaged_wait_work(void)
2236 {
2237         if (khugepaged_has_work()) {
2238                 const unsigned long scan_sleep_jiffies =
2239                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2240
2241                 if (!scan_sleep_jiffies)
2242                         return;
2243
2244                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2245                 wait_event_freezable_timeout(khugepaged_wait,
2246                                              khugepaged_should_wakeup(),
2247                                              scan_sleep_jiffies);
2248                 return;
2249         }
2250
2251         if (khugepaged_enabled())
2252                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2253 }
2254
2255 static int khugepaged(void *none)
2256 {
2257         struct mm_slot *mm_slot;
2258
2259         set_freezable();
2260         set_user_nice(current, MAX_NICE);
2261
2262         while (!kthread_should_stop()) {
2263                 khugepaged_do_scan();
2264                 khugepaged_wait_work();
2265         }
2266
2267         spin_lock(&khugepaged_mm_lock);
2268         mm_slot = khugepaged_scan.mm_slot;
2269         khugepaged_scan.mm_slot = NULL;
2270         if (mm_slot)
2271                 collect_mm_slot(mm_slot);
2272         spin_unlock(&khugepaged_mm_lock);
2273         return 0;
2274 }
2275
2276 static void set_recommended_min_free_kbytes(void)
2277 {
2278         struct zone *zone;
2279         int nr_zones = 0;
2280         unsigned long recommended_min;
2281
2282         for_each_populated_zone(zone) {
2283                 /*
2284                  * We don't need to worry about fragmentation of
2285                  * ZONE_MOVABLE since it only has movable pages.
2286                  */
2287                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2288                         continue;
2289
2290                 nr_zones++;
2291         }
2292
2293         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2294         recommended_min = pageblock_nr_pages * nr_zones * 2;
2295
2296         /*
2297          * Make sure that on average at least two pageblocks are almost free
2298          * of another type, one for a migratetype to fall back to and a
2299          * second to avoid subsequent fallbacks of other types There are 3
2300          * MIGRATE_TYPES we care about.
2301          */
2302         recommended_min += pageblock_nr_pages * nr_zones *
2303                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2304
2305         /* don't ever allow to reserve more than 5% of the lowmem */
2306         recommended_min = min(recommended_min,
2307                               (unsigned long) nr_free_buffer_pages() / 20);
2308         recommended_min <<= (PAGE_SHIFT-10);
2309
2310         if (recommended_min > min_free_kbytes) {
2311                 if (user_min_free_kbytes >= 0)
2312                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2313                                 min_free_kbytes, recommended_min);
2314
2315                 min_free_kbytes = recommended_min;
2316         }
2317         setup_per_zone_wmarks();
2318 }
2319
2320 int start_stop_khugepaged(void)
2321 {
2322         int err = 0;
2323
2324         mutex_lock(&khugepaged_mutex);
2325         if (khugepaged_enabled()) {
2326                 if (!khugepaged_thread)
2327                         khugepaged_thread = kthread_run(khugepaged, NULL,
2328                                                         "khugepaged");
2329                 if (IS_ERR(khugepaged_thread)) {
2330                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2331                         err = PTR_ERR(khugepaged_thread);
2332                         khugepaged_thread = NULL;
2333                         goto fail;
2334                 }
2335
2336                 if (!list_empty(&khugepaged_scan.mm_head))
2337                         wake_up_interruptible(&khugepaged_wait);
2338
2339                 set_recommended_min_free_kbytes();
2340         } else if (khugepaged_thread) {
2341                 kthread_stop(khugepaged_thread);
2342                 khugepaged_thread = NULL;
2343         }
2344 fail:
2345         mutex_unlock(&khugepaged_mutex);
2346         return err;
2347 }
2348
2349 void khugepaged_min_free_kbytes_update(void)
2350 {
2351         mutex_lock(&khugepaged_mutex);
2352         if (khugepaged_enabled() && khugepaged_thread)
2353                 set_recommended_min_free_kbytes();
2354         mutex_unlock(&khugepaged_mutex);
2355 }