mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_EXCEED_SWAP_PTE,
  32         SCAN_EXCEED_SHARED_PTE,
  33         SCAN_PTE_NON_PRESENT,
  34         SCAN_PTE_UFFD_WP,
  35         SCAN_PAGE_RO,
  36         SCAN_LACK_REFERENCED_PAGE,
  37         SCAN_PAGE_NULL,
  38         SCAN_SCAN_ABORT,
  39         SCAN_PAGE_COUNT,
  40         SCAN_PAGE_LRU,
  41         SCAN_PAGE_LOCK,
  42         SCAN_PAGE_ANON,
  43         SCAN_PAGE_COMPOUND,
  44         SCAN_ANY_PROCESS,
  45         SCAN_VMA_NULL,
  46         SCAN_VMA_CHECK,
  47         SCAN_ADDRESS_RANGE,
  48         SCAN_SWAP_CACHE_PAGE,
  49         SCAN_DEL_PAGE_LRU,
  50         SCAN_ALLOC_HUGE_PAGE_FAIL,
  51         SCAN_CGROUP_CHARGE_FAIL,
  52         SCAN_TRUNCATED,
  53         SCAN_PAGE_HAS_PRIVATE,
  54 };
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/huge_memory.h>
  58
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61
  62 /* default scan 8*512 pte (or vmas) every 30 second */
  63 static unsigned int khugepaged_pages_to_scan __read_mostly;
  64 static unsigned int khugepaged_pages_collapsed;
  65 static unsigned int khugepaged_full_scans;
  66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  67 /* during fragmentation poll the hugepage allocator once every minute */
  68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  69 static unsigned long khugepaged_sleep_expire;
  70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  72 /*
  73  * default collapse hugepages if there is at least one pte mapped like
  74  * it would have happened if the vma was large enough during page
  75  * fault.
  76  */
  77 static unsigned int khugepaged_max_ptes_none __read_mostly;
  78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  80
  81 #define MM_SLOTS_HASH_BITS 10
  82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  83
  84 static struct kmem_cache *mm_slot_cache __read_mostly;
  85
  86 #define MAX_PTE_MAPPED_THP 8
  87
  88 /**
  89  * struct mm_slot - hash lookup from mm to mm_slot
  90  * @hash: hash collision list
  91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  92  * @mm: the mm that this information is valid for
  93  * @nr_pte_mapped_thp: number of pte mapped THP
  94  * @pte_mapped_thp: address array corresponding pte mapped THP
  95  */
  96 struct mm_slot {
  97         struct hlist_node hash;
  98         struct list_head mm_node;
  99         struct mm_struct *mm;
 100
 101         /* pte-mapped THP in this mm */
 102         int nr_pte_mapped_thp;
 103         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 104 };
 105
 106 /**
 107  * struct khugepaged_scan - cursor for scanning
 108  * @mm_head: the head of the mm list to scan
 109  * @mm_slot: the current mm_slot we are scanning
 110  * @address: the next address inside that to be scanned
 111  *
 112  * There is only the one khugepaged_scan instance of this cursor structure.
 113  */
 114 struct khugepaged_scan {
 115         struct list_head mm_head;
 116         struct mm_slot *mm_slot;
 117         unsigned long address;
 118 };
 119
 120 static struct khugepaged_scan khugepaged_scan = {
 121         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 122 };
 123
 124 #ifdef CONFIG_SYSFS
 125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 126                                          struct kobj_attribute *attr,
 127                                          char *buf)
 128 {
 129         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 130 }
 131
 132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           const char *buf, size_t count)
 135 {
 136         unsigned int msecs;
 137         int err;
 138
 139         err = kstrtouint(buf, 10, &msecs);
 140         if (err)
 141                 return -EINVAL;
 142
 143         khugepaged_scan_sleep_millisecs = msecs;
 144         khugepaged_sleep_expire = 0;
 145         wake_up_interruptible(&khugepaged_wait);
 146
 147         return count;
 148 }
 149 static struct kobj_attribute scan_sleep_millisecs_attr =
 150         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 151                scan_sleep_millisecs_store);
 152
 153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 154                                           struct kobj_attribute *attr,
 155                                           char *buf)
 156 {
 157         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 158 }
 159
 160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 161                                            struct kobj_attribute *attr,
 162                                            const char *buf, size_t count)
 163 {
 164         unsigned int msecs;
 165         int err;
 166
 167         err = kstrtouint(buf, 10, &msecs);
 168         if (err)
 169                 return -EINVAL;
 170
 171         khugepaged_alloc_sleep_millisecs = msecs;
 172         khugepaged_sleep_expire = 0;
 173         wake_up_interruptible(&khugepaged_wait);
 174
 175         return count;
 176 }
 177 static struct kobj_attribute alloc_sleep_millisecs_attr =
 178         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 179                alloc_sleep_millisecs_store);
 180
 181 static ssize_t pages_to_scan_show(struct kobject *kobj,
 182                                   struct kobj_attribute *attr,
 183                                   char *buf)
 184 {
 185         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 186 }
 187 static ssize_t pages_to_scan_store(struct kobject *kobj,
 188                                    struct kobj_attribute *attr,
 189                                    const char *buf, size_t count)
 190 {
 191         unsigned int pages;
 192         int err;
 193
 194         err = kstrtouint(buf, 10, &pages);
 195         if (err || !pages)
 196                 return -EINVAL;
 197
 198         khugepaged_pages_to_scan = pages;
 199
 200         return count;
 201 }
 202 static struct kobj_attribute pages_to_scan_attr =
 203         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 204                pages_to_scan_store);
 205
 206 static ssize_t pages_collapsed_show(struct kobject *kobj,
 207                                     struct kobj_attribute *attr,
 208                                     char *buf)
 209 {
 210         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 211 }
 212 static struct kobj_attribute pages_collapsed_attr =
 213         __ATTR_RO(pages_collapsed);
 214
 215 static ssize_t full_scans_show(struct kobject *kobj,
 216                                struct kobj_attribute *attr,
 217                                char *buf)
 218 {
 219         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 220 }
 221 static struct kobj_attribute full_scans_attr =
 222         __ATTR_RO(full_scans);
 223
 224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 225                                       struct kobj_attribute *attr, char *buf)
 226 {
 227         return single_hugepage_flag_show(kobj, attr, buf,
 228                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 229 }
 230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 231                                        struct kobj_attribute *attr,
 232                                        const char *buf, size_t count)
 233 {
 234         return single_hugepage_flag_store(kobj, attr, buf, count,
 235                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 236 }
 237 static struct kobj_attribute khugepaged_defrag_attr =
 238         __ATTR(defrag, 0644, khugepaged_defrag_show,
 239                khugepaged_defrag_store);
 240
 241 /*
 242  * max_ptes_none controls if khugepaged should collapse hugepages over
 243  * any unmapped ptes in turn potentially increasing the memory
 244  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 245  * reduce the available free memory in the system as it
 246  * runs. Increasing max_ptes_none will instead potentially reduce the
 247  * free memory in the system during the khugepaged scan.
 248  */
 249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 250                                              struct kobj_attribute *attr,
 251                                              char *buf)
 252 {
 253         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 254 }
 255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 256                                               struct kobj_attribute *attr,
 257                                               const char *buf, size_t count)
 258 {
 259         int err;
 260         unsigned long max_ptes_none;
 261
 262         err = kstrtoul(buf, 10, &max_ptes_none);
 263         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 264                 return -EINVAL;
 265
 266         khugepaged_max_ptes_none = max_ptes_none;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 271         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 272                khugepaged_max_ptes_none_store);
 273
 274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 275                                              struct kobj_attribute *attr,
 276                                              char *buf)
 277 {
 278         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 279 }
 280
 281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 282                                               struct kobj_attribute *attr,
 283                                               const char *buf, size_t count)
 284 {
 285         int err;
 286         unsigned long max_ptes_swap;
 287
 288         err  = kstrtoul(buf, 10, &max_ptes_swap);
 289         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 290                 return -EINVAL;
 291
 292         khugepaged_max_ptes_swap = max_ptes_swap;
 293
 294         return count;
 295 }
 296
 297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 298         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 299                khugepaged_max_ptes_swap_store);
 300
 301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
 302                                                struct kobj_attribute *attr,
 303                                                char *buf)
 304 {
 305         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 306 }
 307
 308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
 309                                               struct kobj_attribute *attr,
 310                                               const char *buf, size_t count)
 311 {
 312         int err;
 313         unsigned long max_ptes_shared;
 314
 315         err  = kstrtoul(buf, 10, &max_ptes_shared);
 316         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
 317                 return -EINVAL;
 318
 319         khugepaged_max_ptes_shared = max_ptes_shared;
 320
 321         return count;
 322 }
 323
 324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 325         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
 326                khugepaged_max_ptes_shared_store);
 327
 328 static struct attribute *khugepaged_attr[] = {
 329         &khugepaged_defrag_attr.attr,
 330         &khugepaged_max_ptes_none_attr.attr,
 331         &khugepaged_max_ptes_swap_attr.attr,
 332         &khugepaged_max_ptes_shared_attr.attr,
 333         &pages_to_scan_attr.attr,
 334         &pages_collapsed_attr.attr,
 335         &full_scans_attr.attr,
 336         &scan_sleep_millisecs_attr.attr,
 337         &alloc_sleep_millisecs_attr.attr,
 338         NULL,
 339 };
 340
 341 struct attribute_group khugepaged_attr_group = {
 342         .attrs = khugepaged_attr,
 343         .name = "khugepaged",
 344 };
 345 #endif /* CONFIG_SYSFS */
 346
 347 int hugepage_madvise(struct vm_area_struct *vma,
 348                      unsigned long *vm_flags, int advice)
 349 {
 350         switch (advice) {
 351         case MADV_HUGEPAGE:
 352 #ifdef CONFIG_S390
 353                 /*
 354                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 355                  * can't handle this properly after s390_enable_sie, so we simply
 356                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 357                  */
 358                 if (mm_has_pgste(vma->vm_mm))
 359                         return 0;
 360 #endif
 361                 *vm_flags &= ~VM_NOHUGEPAGE;
 362                 *vm_flags |= VM_HUGEPAGE;
 363                 /*
 364                  * If the vma become good for khugepaged to scan,
 365                  * register it here without waiting a page fault that
 366                  * may not happen any time soon.
 367                  */
 368                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 369                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 370                         return -ENOMEM;
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct mm_slot),
 390                                           __alignof__(struct mm_slot), 0, NULL);
 391         if (!mm_slot_cache)
 392                 return -ENOMEM;
 393
 394         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 395         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 396         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 397         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 398
 399         return 0;
 400 }
 401
 402 void __init khugepaged_destroy(void)
 403 {
 404         kmem_cache_destroy(mm_slot_cache);
 405 }
 406
 407 static inline struct mm_slot *alloc_mm_slot(void)
 408 {
 409         if (!mm_slot_cache)     /* initialization failed */
 410                 return NULL;
 411         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 412 }
 413
 414 static inline void free_mm_slot(struct mm_slot *mm_slot)
 415 {
 416         kmem_cache_free(mm_slot_cache, mm_slot);
 417 }
 418
 419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 420 {
 421         struct mm_slot *mm_slot;
 422
 423         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 424                 if (mm == mm_slot->mm)
 425                         return mm_slot;
 426
 427         return NULL;
 428 }
 429
 430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 431                                     struct mm_slot *mm_slot)
 432 {
 433         mm_slot->mm = mm;
 434         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 435 }
 436
 437 static inline int khugepaged_test_exit(struct mm_struct *mm)
 438 {
 439         return atomic_read(&mm->mm_users) == 0;
 440 }
 441
 442 static bool hugepage_vma_check(struct vm_area_struct *vma,
 443                                unsigned long vm_flags)
 444 {
 445         if (!transhuge_vma_enabled(vma, vm_flags))
 446                 return false;
 447
 448         if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
 449                                 vma->vm_pgoff, HPAGE_PMD_NR))
 450                 return false;
 451
 452         /* Enabled via shmem mount options or sysfs settings. */
 453         if (shmem_file(vma->vm_file))
 454                 return shmem_huge_enabled(vma);
 455
 456         /* THP settings require madvise. */
 457         if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
 458                 return false;
 459
 460         /* Only regular file is valid */
 461         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
 462             (vm_flags & VM_EXEC)) {
 463                 struct inode *inode = vma->vm_file->f_inode;
 464
 465                 return !inode_is_open_for_write(inode) &&
 466                         S_ISREG(inode->i_mode);
 467         }
 468
 469         if (!vma->anon_vma || vma->vm_ops)
 470                 return false;
 471         if (vma_is_temporary_stack(vma))
 472                 return false;
 473         return !(vm_flags & VM_NO_KHUGEPAGED);
 474 }
 475
 476 int __khugepaged_enter(struct mm_struct *mm)
 477 {
 478         struct mm_slot *mm_slot;
 479         int wakeup;
 480
 481         mm_slot = alloc_mm_slot();
 482         if (!mm_slot)
 483                 return -ENOMEM;
 484
 485         /* __khugepaged_exit() must not run from under us */
 486         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 487         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 488                 free_mm_slot(mm_slot);
 489                 return 0;
 490         }
 491
 492         spin_lock(&khugepaged_mm_lock);
 493         insert_to_mm_slots_hash(mm, mm_slot);
 494         /*
 495          * Insert just behind the scanning cursor, to let the area settle
 496          * down a little.
 497          */
 498         wakeup = list_empty(&khugepaged_scan.mm_head);
 499         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 500         spin_unlock(&khugepaged_mm_lock);
 501
 502         mmgrab(mm);
 503         if (wakeup)
 504                 wake_up_interruptible(&khugepaged_wait);
 505
 506         return 0;
 507 }
 508
 509 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 510                                unsigned long vm_flags)
 511 {
 512         unsigned long hstart, hend;
 513
 514         /*
 515          * khugepaged only supports read-only files for non-shmem files.
 516          * khugepaged does not yet work on special mappings. And
 517          * file-private shmem THP is not supported.
 518          */
 519         if (!hugepage_vma_check(vma, vm_flags))
 520                 return 0;
 521
 522         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 523         hend = vma->vm_end & HPAGE_PMD_MASK;
 524         if (hstart < hend)
 525                 return khugepaged_enter(vma, vm_flags);
 526         return 0;
 527 }
 528
 529 void __khugepaged_exit(struct mm_struct *mm)
 530 {
 531         struct mm_slot *mm_slot;
 532         int free = 0;
 533
 534         spin_lock(&khugepaged_mm_lock);
 535         mm_slot = get_mm_slot(mm);
 536         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 537                 hash_del(&mm_slot->hash);
 538                 list_del(&mm_slot->mm_node);
 539                 free = 1;
 540         }
 541         spin_unlock(&khugepaged_mm_lock);
 542
 543         if (free) {
 544                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 545                 free_mm_slot(mm_slot);
 546                 mmdrop(mm);
 547         } else if (mm_slot) {
 548                 /*
 549                  * This is required to serialize against
 550                  * khugepaged_test_exit() (which is guaranteed to run
 551                  * under mmap sem read mode). Stop here (after we
 552                  * return all pagetables will be destroyed) until
 553                  * khugepaged has finished working on the pagetables
 554                  * under the mmap_lock.
 555                  */
 556                 mmap_write_lock(mm);
 557                 mmap_write_unlock(mm);
 558         }
 559 }
 560
 561 static void release_pte_page(struct page *page)
 562 {
 563         mod_node_page_state(page_pgdat(page),
 564                         NR_ISOLATED_ANON + page_is_file_lru(page),
 565                         -compound_nr(page));
 566         unlock_page(page);
 567         putback_lru_page(page);
 568 }
 569
 570 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 571                 struct list_head *compound_pagelist)
 572 {
 573         struct page *page, *tmp;
 574
 575         while (--_pte >= pte) {
 576                 pte_t pteval = *_pte;
 577
 578                 page = pte_page(pteval);
 579                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 580                                 !PageCompound(page))
 581                         release_pte_page(page);
 582         }
 583
 584         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 585                 list_del(&page->lru);
 586                 release_pte_page(page);
 587         }
 588 }
 589
 590 static bool is_refcount_suitable(struct page *page)
 591 {
 592         int expected_refcount;
 593
 594         expected_refcount = total_mapcount(page);
 595         if (PageSwapCache(page))
 596                 expected_refcount += compound_nr(page);
 597
 598         return page_count(page) == expected_refcount;
 599 }
 600
 601 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 602                                         unsigned long address,
 603                                         pte_t *pte,
 604                                         struct list_head *compound_pagelist)
 605 {
 606         struct page *page = NULL;
 607         pte_t *_pte;
 608         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
 609         bool writable = false;
 610
 611         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 612              _pte++, address += PAGE_SIZE) {
 613                 pte_t pteval = *_pte;
 614                 if (pte_none(pteval) || (pte_present(pteval) &&
 615                                 is_zero_pfn(pte_pfn(pteval)))) {
 616                         if (!userfaultfd_armed(vma) &&
 617                             ++none_or_zero <= khugepaged_max_ptes_none) {
 618                                 continue;
 619                         } else {
 620                                 result = SCAN_EXCEED_NONE_PTE;
 621                                 goto out;
 622                         }
 623                 }
 624                 if (!pte_present(pteval)) {
 625                         result = SCAN_PTE_NON_PRESENT;
 626                         goto out;
 627                 }
 628                 page = vm_normal_page(vma, address, pteval);
 629                 if (unlikely(!page)) {
 630                         result = SCAN_PAGE_NULL;
 631                         goto out;
 632                 }
 633
 634                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 635
 636                 if (page_mapcount(page) > 1 &&
 637                                 ++shared > khugepaged_max_ptes_shared) {
 638                         result = SCAN_EXCEED_SHARED_PTE;
 639                         goto out;
 640                 }
 641
 642                 if (PageCompound(page)) {
 643                         struct page *p;
 644                         page = compound_head(page);
 645
 646                         /*
 647                          * Check if we have dealt with the compound page
 648                          * already
 649                          */
 650                         list_for_each_entry(p, compound_pagelist, lru) {
 651                                 if (page == p)
 652                                         goto next;
 653                         }
 654                 }
 655
 656                 /*
 657                  * We can do it before isolate_lru_page because the
 658                  * page can't be freed from under us. NOTE: PG_lock
 659                  * is needed to serialize against split_huge_page
 660                  * when invoked from the VM.
 661                  */
 662                 if (!trylock_page(page)) {
 663                         result = SCAN_PAGE_LOCK;
 664                         goto out;
 665                 }
 666
 667                 /*
 668                  * Check if the page has any GUP (or other external) pins.
 669                  *
 670                  * The page table that maps the page has been already unlinked
 671                  * from the page table tree and this process cannot get
 672                  * an additional pin on the page.
 673                  *
 674                  * New pins can come later if the page is shared across fork,
 675                  * but not from this process. The other process cannot write to
 676                  * the page, only trigger CoW.
 677                  */
 678                 if (!is_refcount_suitable(page)) {
 679                         unlock_page(page);
 680                         result = SCAN_PAGE_COUNT;
 681                         goto out;
 682                 }
 683                 if (!pte_write(pteval) && PageSwapCache(page) &&
 684                                 !reuse_swap_page(page, NULL)) {
 685                         /*
 686                          * Page is in the swap cache and cannot be re-used.
 687                          * It cannot be collapsed into a THP.
 688                          */
 689                         unlock_page(page);
 690                         result = SCAN_SWAP_CACHE_PAGE;
 691                         goto out;
 692                 }
 693
 694                 /*
 695                  * Isolate the page to avoid collapsing an hugepage
 696                  * currently in use by the VM.
 697                  */
 698                 if (isolate_lru_page(page)) {
 699                         unlock_page(page);
 700                         result = SCAN_DEL_PAGE_LRU;
 701                         goto out;
 702                 }
 703                 mod_node_page_state(page_pgdat(page),
 704                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 705                                 compound_nr(page));
 706                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 707                 VM_BUG_ON_PAGE(PageLRU(page), page);
 708
 709                 if (PageCompound(page))
 710                         list_add_tail(&page->lru, compound_pagelist);
 711 next:
 712                 /* There should be enough young pte to collapse the page */
 713                 if (pte_young(pteval) ||
 714                     page_is_young(page) || PageReferenced(page) ||
 715                     mmu_notifier_test_young(vma->vm_mm, address))
 716                         referenced++;
 717
 718                 if (pte_write(pteval))
 719                         writable = true;
 720         }
 721
 722         if (unlikely(!writable)) {
 723                 result = SCAN_PAGE_RO;
 724         } else if (unlikely(!referenced)) {
 725                 result = SCAN_LACK_REFERENCED_PAGE;
 726         } else {
 727                 result = SCAN_SUCCEED;
 728                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 729                                                     referenced, writable, result);
 730                 return 1;
 731         }
 732 out:
 733         release_pte_pages(pte, _pte, compound_pagelist);
 734         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 735                                             referenced, writable, result);
 736         return 0;
 737 }
 738
 739 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 740                                       struct vm_area_struct *vma,
 741                                       unsigned long address,
 742                                       spinlock_t *ptl,
 743                                       struct list_head *compound_pagelist)
 744 {
 745         struct page *src_page, *tmp;
 746         pte_t *_pte;
 747         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 748                                 _pte++, page++, address += PAGE_SIZE) {
 749                 pte_t pteval = *_pte;
 750
 751                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 752                         clear_user_highpage(page, address);
 753                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 754                         if (is_zero_pfn(pte_pfn(pteval))) {
 755                                 /*
 756                                  * ptl mostly unnecessary.
 757                                  */
 758                                 spin_lock(ptl);
 759                                 /*
 760                                  * paravirt calls inside pte_clear here are
 761                                  * superfluous.
 762                                  */
 763                                 pte_clear(vma->vm_mm, address, _pte);
 764                                 spin_unlock(ptl);
 765                         }
 766                 } else {
 767                         src_page = pte_page(pteval);
 768                         copy_user_highpage(page, src_page, address, vma);
 769                         if (!PageCompound(src_page))
 770                                 release_pte_page(src_page);
 771                         /*
 772                          * ptl mostly unnecessary, but preempt has to
 773                          * be disabled to update the per-cpu stats
 774                          * inside page_remove_rmap().
 775                          */
 776                         spin_lock(ptl);
 777                         /*
 778                          * paravirt calls inside pte_clear here are
 779                          * superfluous.
 780                          */
 781                         pte_clear(vma->vm_mm, address, _pte);
 782                         page_remove_rmap(src_page, false);
 783                         spin_unlock(ptl);
 784                         free_page_and_swap_cache(src_page);
 785                 }
 786         }
 787
 788         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 789                 list_del(&src_page->lru);
 790                 release_pte_page(src_page);
 791         }
 792 }
 793
 794 static void khugepaged_alloc_sleep(void)
 795 {
 796         DEFINE_WAIT(wait);
 797
 798         add_wait_queue(&khugepaged_wait, &wait);
 799         freezable_schedule_timeout_interruptible(
 800                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 801         remove_wait_queue(&khugepaged_wait, &wait);
 802 }
 803
 804 static int khugepaged_node_load[MAX_NUMNODES];
 805
 806 static bool khugepaged_scan_abort(int nid)
 807 {
 808         int i;
 809
 810         /*
 811          * If node_reclaim_mode is disabled, then no extra effort is made to
 812          * allocate memory locally.
 813          */
 814         if (!node_reclaim_enabled())
 815                 return false;
 816
 817         /* If there is a count for this node already, it must be acceptable */
 818         if (khugepaged_node_load[nid])
 819                 return false;
 820
 821         for (i = 0; i < MAX_NUMNODES; i++) {
 822                 if (!khugepaged_node_load[i])
 823                         continue;
 824                 if (node_distance(nid, i) > node_reclaim_distance)
 825                         return true;
 826         }
 827         return false;
 828 }
 829
 830 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 831 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 832 {
 833         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 834 }
 835
 836 #ifdef CONFIG_NUMA
 837 static int khugepaged_find_target_node(void)
 838 {
 839         static int last_khugepaged_target_node = NUMA_NO_NODE;
 840         int nid, target_node = 0, max_value = 0;
 841
 842         /* find first node with max normal pages hit */
 843         for (nid = 0; nid < MAX_NUMNODES; nid++)
 844                 if (khugepaged_node_load[nid] > max_value) {
 845                         max_value = khugepaged_node_load[nid];
 846                         target_node = nid;
 847                 }
 848
 849         /* do some balance if several nodes have the same hit record */
 850         if (target_node <= last_khugepaged_target_node)
 851                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 852                                 nid++)
 853                         if (max_value == khugepaged_node_load[nid]) {
 854                                 target_node = nid;
 855                                 break;
 856                         }
 857
 858         last_khugepaged_target_node = target_node;
 859         return target_node;
 860 }
 861
 862 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 863 {
 864         if (IS_ERR(*hpage)) {
 865                 if (!*wait)
 866                         return false;
 867
 868                 *wait = false;
 869                 *hpage = NULL;
 870                 khugepaged_alloc_sleep();
 871         } else if (*hpage) {
 872                 put_page(*hpage);
 873                 *hpage = NULL;
 874         }
 875
 876         return true;
 877 }
 878
 879 static struct page *
 880 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 881 {
 882         VM_BUG_ON_PAGE(*hpage, *hpage);
 883
 884         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 885         if (unlikely(!*hpage)) {
 886                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 887                 *hpage = ERR_PTR(-ENOMEM);
 888                 return NULL;
 889         }
 890
 891         prep_transhuge_page(*hpage);
 892         count_vm_event(THP_COLLAPSE_ALLOC);
 893         return *hpage;
 894 }
 895 #else
 896 static int khugepaged_find_target_node(void)
 897 {
 898         return 0;
 899 }
 900
 901 static inline struct page *alloc_khugepaged_hugepage(void)
 902 {
 903         struct page *page;
 904
 905         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 906                            HPAGE_PMD_ORDER);
 907         if (page)
 908                 prep_transhuge_page(page);
 909         return page;
 910 }
 911
 912 static struct page *khugepaged_alloc_hugepage(bool *wait)
 913 {
 914         struct page *hpage;
 915
 916         do {
 917                 hpage = alloc_khugepaged_hugepage();
 918                 if (!hpage) {
 919                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 920                         if (!*wait)
 921                                 return NULL;
 922
 923                         *wait = false;
 924                         khugepaged_alloc_sleep();
 925                 } else
 926                         count_vm_event(THP_COLLAPSE_ALLOC);
 927         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 928
 929         return hpage;
 930 }
 931
 932 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 933 {
 934         /*
 935          * If the hpage allocated earlier was briefly exposed in page cache
 936          * before collapse_file() failed, it is possible that racing lookups
 937          * have not yet completed, and would then be unpleasantly surprised by
 938          * finding the hpage reused for the same mapping at a different offset.
 939          * Just release the previous allocation if there is any danger of that.
 940          */
 941         if (*hpage && page_count(*hpage) > 1) {
 942                 put_page(*hpage);
 943                 *hpage = NULL;
 944         }
 945
 946         if (!*hpage)
 947                 *hpage = khugepaged_alloc_hugepage(wait);
 948
 949         if (unlikely(!*hpage))
 950                 return false;
 951
 952         return true;
 953 }
 954
 955 static struct page *
 956 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 957 {
 958         VM_BUG_ON(!*hpage);
 959
 960         return  *hpage;
 961 }
 962 #endif
 963
 964 /*
 965  * If mmap_lock temporarily dropped, revalidate vma
 966  * before taking mmap_lock.
 967  * Return 0 if succeeds, otherwise return none-zero
 968  * value (scan code).
 969  */
 970
 971 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 972                 struct vm_area_struct **vmap)
 973 {
 974         struct vm_area_struct *vma;
 975         unsigned long hstart, hend;
 976
 977         if (unlikely(khugepaged_test_exit(mm)))
 978                 return SCAN_ANY_PROCESS;
 979
 980         *vmap = vma = find_vma(mm, address);
 981         if (!vma)
 982                 return SCAN_VMA_NULL;
 983
 984         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 985         hend = vma->vm_end & HPAGE_PMD_MASK;
 986         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 987                 return SCAN_ADDRESS_RANGE;
 988         if (!hugepage_vma_check(vma, vma->vm_flags))
 989                 return SCAN_VMA_CHECK;
 990         /* Anon VMA expected */
 991         if (!vma->anon_vma || vma->vm_ops)
 992                 return SCAN_VMA_CHECK;
 993         return 0;
 994 }
 995
 996 /*
 997  * Bring missing pages in from swap, to complete THP collapse.
 998  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 999  *
1000  * Called and returns without pte mapped or spinlocks held,
1001  * but with mmap_lock held to protect against vma changes.
1002  */
1003
1004 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1005                                         struct vm_area_struct *vma,
1006                                         unsigned long haddr, pmd_t *pmd,
1007                                         int referenced)
1008 {
1009         int swapped_in = 0;
1010         vm_fault_t ret = 0;
1011         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1012
1013         for (address = haddr; address < end; address += PAGE_SIZE) {
1014                 struct vm_fault vmf = {
1015                         .vma = vma,
1016                         .address = address,
1017                         .pgoff = linear_page_index(vma, haddr),
1018                         .flags = FAULT_FLAG_ALLOW_RETRY,
1019                         .pmd = pmd,
1020                 };
1021
1022                 vmf.pte = pte_offset_map(pmd, address);
1023                 vmf.orig_pte = *vmf.pte;
1024                 if (!is_swap_pte(vmf.orig_pte)) {
1025                         pte_unmap(vmf.pte);
1026                         continue;
1027                 }
1028                 swapped_in++;
1029                 ret = do_swap_page(&vmf);
1030
1031                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1032                 if (ret & VM_FAULT_RETRY) {
1033                         mmap_read_lock(mm);
1034                         if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1035                                 /* vma is no longer available, don't continue to swapin */
1036                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1037                                 return false;
1038                         }
1039                         /* check if the pmd is still valid */
1040                         if (mm_find_pmd(mm, haddr) != pmd) {
1041                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1042                                 return false;
1043                         }
1044                 }
1045                 if (ret & VM_FAULT_ERROR) {
1046                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1047                         return false;
1048                 }
1049         }
1050
1051         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1052         if (swapped_in)
1053                 lru_add_drain();
1054
1055         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1056         return true;
1057 }
1058
1059 static void collapse_huge_page(struct mm_struct *mm,
1060                                    unsigned long address,
1061                                    struct page **hpage,
1062                                    int node, int referenced, int unmapped)
1063 {
1064         LIST_HEAD(compound_pagelist);
1065         pmd_t *pmd, _pmd;
1066         pte_t *pte;
1067         pgtable_t pgtable;
1068         struct page *new_page;
1069         spinlock_t *pmd_ptl, *pte_ptl;
1070         int isolated = 0, result = 0;
1071         struct vm_area_struct *vma;
1072         struct mmu_notifier_range range;
1073         gfp_t gfp;
1074
1075         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1076
1077         /* Only allocate from the target node */
1078         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1079
1080         /*
1081          * Before allocating the hugepage, release the mmap_lock read lock.
1082          * The allocation can take potentially a long time if it involves
1083          * sync compaction, and we do not need to hold the mmap_lock during
1084          * that. We will recheck the vma after taking it again in write mode.
1085          */
1086         mmap_read_unlock(mm);
1087         new_page = khugepaged_alloc_page(hpage, gfp, node);
1088         if (!new_page) {
1089                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1090                 goto out_nolock;
1091         }
1092
1093         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1094                 result = SCAN_CGROUP_CHARGE_FAIL;
1095                 goto out_nolock;
1096         }
1097         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1098
1099         mmap_read_lock(mm);
1100         result = hugepage_vma_revalidate(mm, address, &vma);
1101         if (result) {
1102                 mmap_read_unlock(mm);
1103                 goto out_nolock;
1104         }
1105
1106         pmd = mm_find_pmd(mm, address);
1107         if (!pmd) {
1108                 result = SCAN_PMD_NULL;
1109                 mmap_read_unlock(mm);
1110                 goto out_nolock;
1111         }
1112
1113         /*
1114          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1115          * If it fails, we release mmap_lock and jump out_nolock.
1116          * Continuing to collapse causes inconsistency.
1117          */
1118         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1119                                                      pmd, referenced)) {
1120                 mmap_read_unlock(mm);
1121                 goto out_nolock;
1122         }
1123
1124         mmap_read_unlock(mm);
1125         /*
1126          * Prevent all access to pagetables with the exception of
1127          * gup_fast later handled by the ptep_clear_flush and the VM
1128          * handled by the anon_vma lock + PG_lock.
1129          */
1130         mmap_write_lock(mm);
1131         result = hugepage_vma_revalidate(mm, address, &vma);
1132         if (result)
1133                 goto out_up_write;
1134         /* check if the pmd is still valid */
1135         if (mm_find_pmd(mm, address) != pmd)
1136                 goto out_up_write;
1137
1138         anon_vma_lock_write(vma->anon_vma);
1139
1140         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1141                                 address, address + HPAGE_PMD_SIZE);
1142         mmu_notifier_invalidate_range_start(&range);
1143
1144         pte = pte_offset_map(pmd, address);
1145         pte_ptl = pte_lockptr(mm, pmd);
1146
1147         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1148         /*
1149          * This removes any huge TLB entry from the CPU so we won't allow
1150          * huge and small TLB entries for the same virtual address to
1151          * avoid the risk of CPU bugs in that area.
1152          *
1153          * Parallel fast GUP is fine since fast GUP will back off when
1154          * it detects PMD is changed.
1155          */
1156         _pmd = pmdp_collapse_flush(vma, address, pmd);
1157         spin_unlock(pmd_ptl);
1158         mmu_notifier_invalidate_range_end(&range);
1159         tlb_remove_table_sync_one();
1160
1161         spin_lock(pte_ptl);
1162         isolated = __collapse_huge_page_isolate(vma, address, pte,
1163                         &compound_pagelist);
1164         spin_unlock(pte_ptl);
1165
1166         if (unlikely(!isolated)) {
1167                 pte_unmap(pte);
1168                 spin_lock(pmd_ptl);
1169                 BUG_ON(!pmd_none(*pmd));
1170                 /*
1171                  * We can only use set_pmd_at when establishing
1172                  * hugepmds and never for establishing regular pmds that
1173                  * points to regular pagetables. Use pmd_populate for that
1174                  */
1175                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1176                 spin_unlock(pmd_ptl);
1177                 anon_vma_unlock_write(vma->anon_vma);
1178                 result = SCAN_FAIL;
1179                 goto out_up_write;
1180         }
1181
1182         /*
1183          * All pages are isolated and locked so anon_vma rmap
1184          * can't run anymore.
1185          */
1186         anon_vma_unlock_write(vma->anon_vma);
1187
1188         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1189                         &compound_pagelist);
1190         pte_unmap(pte);
1191         /*
1192          * spin_lock() below is not the equivalent of smp_wmb(), but
1193          * the smp_wmb() inside __SetPageUptodate() can be reused to
1194          * avoid the copy_huge_page writes to become visible after
1195          * the set_pmd_at() write.
1196          */
1197         __SetPageUptodate(new_page);
1198         pgtable = pmd_pgtable(_pmd);
1199
1200         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1201         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1202
1203         spin_lock(pmd_ptl);
1204         BUG_ON(!pmd_none(*pmd));
1205         page_add_new_anon_rmap(new_page, vma, address, true);
1206         lru_cache_add_inactive_or_unevictable(new_page, vma);
1207         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1208         set_pmd_at(mm, address, pmd, _pmd);
1209         update_mmu_cache_pmd(vma, address, pmd);
1210         spin_unlock(pmd_ptl);
1211
1212         *hpage = NULL;
1213
1214         khugepaged_pages_collapsed++;
1215         result = SCAN_SUCCEED;
1216 out_up_write:
1217         mmap_write_unlock(mm);
1218 out_nolock:
1219         if (!IS_ERR_OR_NULL(*hpage))
1220                 mem_cgroup_uncharge(*hpage);
1221         trace_mm_collapse_huge_page(mm, isolated, result);
1222         return;
1223 }
1224
1225 static int khugepaged_scan_pmd(struct mm_struct *mm,
1226                                struct vm_area_struct *vma,
1227                                unsigned long address,
1228                                struct page **hpage)
1229 {
1230         pmd_t *pmd;
1231         pte_t *pte, *_pte;
1232         int ret = 0, result = 0, referenced = 0;
1233         int none_or_zero = 0, shared = 0;
1234         struct page *page = NULL;
1235         unsigned long _address;
1236         spinlock_t *ptl;
1237         int node = NUMA_NO_NODE, unmapped = 0;
1238         bool writable = false;
1239
1240         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1241
1242         pmd = mm_find_pmd(mm, address);
1243         if (!pmd) {
1244                 result = SCAN_PMD_NULL;
1245                 goto out;
1246         }
1247
1248         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1249         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1250         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1251              _pte++, _address += PAGE_SIZE) {
1252                 pte_t pteval = *_pte;
1253                 if (is_swap_pte(pteval)) {
1254                         if (++unmapped <= khugepaged_max_ptes_swap) {
1255                                 /*
1256                                  * Always be strict with uffd-wp
1257                                  * enabled swap entries.  Please see
1258                                  * comment below for pte_uffd_wp().
1259                                  */
1260                                 if (pte_swp_uffd_wp(pteval)) {
1261                                         result = SCAN_PTE_UFFD_WP;
1262                                         goto out_unmap;
1263                                 }
1264                                 continue;
1265                         } else {
1266                                 result = SCAN_EXCEED_SWAP_PTE;
1267                                 goto out_unmap;
1268                         }
1269                 }
1270                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1271                         if (!userfaultfd_armed(vma) &&
1272                             ++none_or_zero <= khugepaged_max_ptes_none) {
1273                                 continue;
1274                         } else {
1275                                 result = SCAN_EXCEED_NONE_PTE;
1276                                 goto out_unmap;
1277                         }
1278                 }
1279                 if (pte_uffd_wp(pteval)) {
1280                         /*
1281                          * Don't collapse the page if any of the small
1282                          * PTEs are armed with uffd write protection.
1283                          * Here we can also mark the new huge pmd as
1284                          * write protected if any of the small ones is
1285                          * marked but that could bring unknown
1286                          * userfault messages that falls outside of
1287                          * the registered range.  So, just be simple.
1288                          */
1289                         result = SCAN_PTE_UFFD_WP;
1290                         goto out_unmap;
1291                 }
1292                 if (pte_write(pteval))
1293                         writable = true;
1294
1295                 page = vm_normal_page(vma, _address, pteval);
1296                 if (unlikely(!page)) {
1297                         result = SCAN_PAGE_NULL;
1298                         goto out_unmap;
1299                 }
1300
1301                 if (page_mapcount(page) > 1 &&
1302                                 ++shared > khugepaged_max_ptes_shared) {
1303                         result = SCAN_EXCEED_SHARED_PTE;
1304                         goto out_unmap;
1305                 }
1306
1307                 page = compound_head(page);
1308
1309                 /*
1310                  * Record which node the original page is from and save this
1311                  * information to khugepaged_node_load[].
1312                  * Khupaged will allocate hugepage from the node has the max
1313                  * hit record.
1314                  */
1315                 node = page_to_nid(page);
1316                 if (khugepaged_scan_abort(node)) {
1317                         result = SCAN_SCAN_ABORT;
1318                         goto out_unmap;
1319                 }
1320                 khugepaged_node_load[node]++;
1321                 if (!PageLRU(page)) {
1322                         result = SCAN_PAGE_LRU;
1323                         goto out_unmap;
1324                 }
1325                 if (PageLocked(page)) {
1326                         result = SCAN_PAGE_LOCK;
1327                         goto out_unmap;
1328                 }
1329                 if (!PageAnon(page)) {
1330                         result = SCAN_PAGE_ANON;
1331                         goto out_unmap;
1332                 }
1333
1334                 /*
1335                  * Check if the page has any GUP (or other external) pins.
1336                  *
1337                  * Here the check is racy it may see totmal_mapcount > refcount
1338                  * in some cases.
1339                  * For example, one process with one forked child process.
1340                  * The parent has the PMD split due to MADV_DONTNEED, then
1341                  * the child is trying unmap the whole PMD, but khugepaged
1342                  * may be scanning the parent between the child has
1343                  * PageDoubleMap flag cleared and dec the mapcount.  So
1344                  * khugepaged may see total_mapcount > refcount.
1345                  *
1346                  * But such case is ephemeral we could always retry collapse
1347                  * later.  However it may report false positive if the page
1348                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1349                  * will be done again later the risk seems low.
1350                  */
1351                 if (!is_refcount_suitable(page)) {
1352                         result = SCAN_PAGE_COUNT;
1353                         goto out_unmap;
1354                 }
1355                 if (pte_young(pteval) ||
1356                     page_is_young(page) || PageReferenced(page) ||
1357                     mmu_notifier_test_young(vma->vm_mm, address))
1358                         referenced++;
1359         }
1360         if (!writable) {
1361                 result = SCAN_PAGE_RO;
1362         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1363                 result = SCAN_LACK_REFERENCED_PAGE;
1364         } else {
1365                 result = SCAN_SUCCEED;
1366                 ret = 1;
1367         }
1368 out_unmap:
1369         pte_unmap_unlock(pte, ptl);
1370         if (ret) {
1371                 node = khugepaged_find_target_node();
1372                 /* collapse_huge_page will return with the mmap_lock released */
1373                 collapse_huge_page(mm, address, hpage, node,
1374                                 referenced, unmapped);
1375         }
1376 out:
1377         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1378                                      none_or_zero, result, unmapped);
1379         return ret;
1380 }
1381
1382 static void collect_mm_slot(struct mm_slot *mm_slot)
1383 {
1384         struct mm_struct *mm = mm_slot->mm;
1385
1386         lockdep_assert_held(&khugepaged_mm_lock);
1387
1388         if (khugepaged_test_exit(mm)) {
1389                 /* free mm_slot */
1390                 hash_del(&mm_slot->hash);
1391                 list_del(&mm_slot->mm_node);
1392
1393                 /*
1394                  * Not strictly needed because the mm exited already.
1395                  *
1396                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1397                  */
1398
1399                 /* khugepaged_mm_lock actually not necessary for the below */
1400                 free_mm_slot(mm_slot);
1401                 mmdrop(mm);
1402         }
1403 }
1404
1405 #ifdef CONFIG_SHMEM
1406 /*
1407  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1408  * khugepaged should try to collapse the page table.
1409  */
1410 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1411                                          unsigned long addr)
1412 {
1413         struct mm_slot *mm_slot;
1414
1415         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1416
1417         spin_lock(&khugepaged_mm_lock);
1418         mm_slot = get_mm_slot(mm);
1419         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1420                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1421         spin_unlock(&khugepaged_mm_lock);
1422         return 0;
1423 }
1424
1425 /**
1426  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1427  * address haddr.
1428  *
1429  * @mm: process address space where collapse happens
1430  * @addr: THP collapse address
1431  *
1432  * This function checks whether all the PTEs in the PMD are pointing to the
1433  * right THP. If so, retract the page table so the THP can refault in with
1434  * as pmd-mapped.
1435  */
1436 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1437 {
1438         unsigned long haddr = addr & HPAGE_PMD_MASK;
1439         struct vm_area_struct *vma = find_vma(mm, haddr);
1440         struct page *hpage;
1441         pte_t *start_pte, *pte;
1442         pmd_t *pmd, _pmd;
1443         spinlock_t *ptl;
1444         int count = 0;
1445         int i;
1446         struct mmu_notifier_range range;
1447
1448         if (!vma || !vma->vm_file ||
1449             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1450                 return;
1451
1452         /*
1453          * This vm_flags may not have VM_HUGEPAGE if the page was not
1454          * collapsed by this mm. But we can still collapse if the page is
1455          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1456          * will not fail the vma for missing VM_HUGEPAGE
1457          */
1458         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1459                 return;
1460
1461         /*
1462          * Symmetry with retract_page_tables(): Exclude MAP_PRIVATE mappings
1463          * that got written to. Without this, we'd have to also lock the
1464          * anon_vma if one exists.
1465          */
1466         if (vma->anon_vma)
1467                 return;
1468
1469         hpage = find_lock_page(vma->vm_file->f_mapping,
1470                                linear_page_index(vma, haddr));
1471         if (!hpage)
1472                 return;
1473
1474         if (!PageHead(hpage))
1475                 goto drop_hpage;
1476
1477         pmd = mm_find_pmd(mm, haddr);
1478         if (!pmd)
1479                 goto drop_hpage;
1480
1481         /*
1482          * We need to lock the mapping so that from here on, only GUP-fast and
1483          * hardware page walks can access the parts of the page tables that
1484          * we're operating on.
1485          */
1486         i_mmap_lock_write(vma->vm_file->f_mapping);
1487
1488         /*
1489          * This spinlock should be unnecessary: Nobody else should be accessing
1490          * the page tables under spinlock protection here, only
1491          * lockless_pages_from_mm() and the hardware page walker can access page
1492          * tables while all the high-level locks are held in write mode.
1493          */
1494         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1495
1496         /* step 1: check all mapped PTEs are to the right huge page */
1497         for (i = 0, addr = haddr, pte = start_pte;
1498              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1499                 struct page *page;
1500
1501                 /* empty pte, skip */
1502                 if (pte_none(*pte))
1503                         continue;
1504
1505                 /* page swapped out, abort */
1506                 if (!pte_present(*pte))
1507                         goto abort;
1508
1509                 page = vm_normal_page(vma, addr, *pte);
1510
1511                 /*
1512                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1513                  * page table, but the new page will not be a subpage of hpage.
1514                  */
1515                 if (hpage + i != page)
1516                         goto abort;
1517                 count++;
1518         }
1519
1520         /* step 2: adjust rmap */
1521         for (i = 0, addr = haddr, pte = start_pte;
1522              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1523                 struct page *page;
1524
1525                 if (pte_none(*pte))
1526                         continue;
1527                 page = vm_normal_page(vma, addr, *pte);
1528                 page_remove_rmap(page, false);
1529         }
1530
1531         pte_unmap_unlock(start_pte, ptl);
1532
1533         /* step 3: set proper refcount and mm_counters. */
1534         if (count) {
1535                 page_ref_sub(hpage, count);
1536                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1537         }
1538
1539         /* step 4: collapse pmd */
1540         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1541                                 haddr + HPAGE_PMD_SIZE);
1542         mmu_notifier_invalidate_range_start(&range);
1543         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1544         mm_dec_nr_ptes(mm);
1545         tlb_remove_table_sync_one();
1546         mmu_notifier_invalidate_range_end(&range);
1547         pte_free(mm, pmd_pgtable(_pmd));
1548
1549         i_mmap_unlock_write(vma->vm_file->f_mapping);
1550
1551 drop_hpage:
1552         unlock_page(hpage);
1553         put_page(hpage);
1554         return;
1555
1556 abort:
1557         pte_unmap_unlock(start_pte, ptl);
1558         i_mmap_unlock_write(vma->vm_file->f_mapping);
1559         goto drop_hpage;
1560 }
1561
1562 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1563 {
1564         struct mm_struct *mm = mm_slot->mm;
1565         int i;
1566
1567         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1568                 return;
1569
1570         if (!mmap_write_trylock(mm))
1571                 return;
1572
1573         if (unlikely(khugepaged_test_exit(mm)))
1574                 goto out;
1575
1576         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1577                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1578
1579 out:
1580         mm_slot->nr_pte_mapped_thp = 0;
1581         mmap_write_unlock(mm);
1582 }
1583
1584 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1585 {
1586         struct vm_area_struct *vma;
1587         struct mm_struct *mm;
1588         unsigned long addr;
1589         pmd_t *pmd, _pmd;
1590
1591         i_mmap_lock_write(mapping);
1592         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1593                 /*
1594                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1595                  * got written to. These VMAs are likely not worth investing
1596                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1597                  * later.
1598                  *
1599                  * Not that vma->anon_vma check is racy: it can be set up after
1600                  * the check but before we took mmap_lock by the fault path.
1601                  * But page lock would prevent establishing any new ptes of the
1602                  * page, so we are safe.
1603                  *
1604                  * An alternative would be drop the check, but check that page
1605                  * table is clear before calling pmdp_collapse_flush() under
1606                  * ptl. It has higher chance to recover THP for the VMA, but
1607                  * has higher cost too. It would also probably require locking
1608                  * the anon_vma.
1609                  */
1610                 if (vma->anon_vma)
1611                         continue;
1612                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1613                 if (addr & ~HPAGE_PMD_MASK)
1614                         continue;
1615                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1616                         continue;
1617                 mm = vma->vm_mm;
1618                 pmd = mm_find_pmd(mm, addr);
1619                 if (!pmd)
1620                         continue;
1621                 /*
1622                  * We need exclusive mmap_lock to retract page table.
1623                  *
1624                  * We use trylock due to lock inversion: we need to acquire
1625                  * mmap_lock while holding page lock. Fault path does it in
1626                  * reverse order. Trylock is a way to avoid deadlock.
1627                  */
1628                 if (mmap_write_trylock(mm)) {
1629                         if (!khugepaged_test_exit(mm)) {
1630                                 struct mmu_notifier_range range;
1631
1632                                 mmu_notifier_range_init(&range,
1633                                                         MMU_NOTIFY_CLEAR, 0,
1634                                                         NULL, mm, addr,
1635                                                         addr + HPAGE_PMD_SIZE);
1636                                 mmu_notifier_invalidate_range_start(&range);
1637                                 /* assume page table is clear */
1638                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1639                                 mm_dec_nr_ptes(mm);
1640                                 tlb_remove_table_sync_one();
1641                                 pte_free(mm, pmd_pgtable(_pmd));
1642                                 mmu_notifier_invalidate_range_end(&range);
1643                         }
1644                         mmap_write_unlock(mm);
1645                 } else {
1646                         /* Try again later */
1647                         khugepaged_add_pte_mapped_thp(mm, addr);
1648                 }
1649         }
1650         i_mmap_unlock_write(mapping);
1651 }
1652
1653 /**
1654  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1655  *
1656  * @mm: process address space where collapse happens
1657  * @file: file that collapse on
1658  * @start: collapse start address
1659  * @hpage: new allocated huge page for collapse
1660  * @node: appointed node the new huge page allocate from
1661  *
1662  * Basic scheme is simple, details are more complex:
1663  *  - allocate and lock a new huge page;
1664  *  - scan page cache replacing old pages with the new one
1665  *    + swap/gup in pages if necessary;
1666  *    + fill in gaps;
1667  *    + keep old pages around in case rollback is required;
1668  *  - if replacing succeeds:
1669  *    + copy data over;
1670  *    + free old pages;
1671  *    + unlock huge page;
1672  *  - if replacing failed;
1673  *    + put all pages back and unfreeze them;
1674  *    + restore gaps in the page cache;
1675  *    + unlock and free huge page;
1676  */
1677 static void collapse_file(struct mm_struct *mm,
1678                 struct file *file, pgoff_t start,
1679                 struct page **hpage, int node)
1680 {
1681         struct address_space *mapping = file->f_mapping;
1682         gfp_t gfp;
1683         struct page *new_page;
1684         pgoff_t index, end = start + HPAGE_PMD_NR;
1685         LIST_HEAD(pagelist);
1686         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1687         int nr_none = 0, result = SCAN_SUCCEED;
1688         bool is_shmem = shmem_file(file);
1689         int nr;
1690
1691         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1692         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1693
1694         /* Only allocate from the target node */
1695         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1696
1697         new_page = khugepaged_alloc_page(hpage, gfp, node);
1698         if (!new_page) {
1699                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1700                 goto out;
1701         }
1702
1703         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1704                 result = SCAN_CGROUP_CHARGE_FAIL;
1705                 goto out;
1706         }
1707         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1708
1709         /* This will be less messy when we use multi-index entries */
1710         do {
1711                 xas_lock_irq(&xas);
1712                 xas_create_range(&xas);
1713                 if (!xas_error(&xas))
1714                         break;
1715                 xas_unlock_irq(&xas);
1716                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1717                         result = SCAN_FAIL;
1718                         goto out;
1719                 }
1720         } while (1);
1721
1722         __SetPageLocked(new_page);
1723         if (is_shmem)
1724                 __SetPageSwapBacked(new_page);
1725         new_page->index = start;
1726         new_page->mapping = mapping;
1727
1728         /*
1729          * At this point the new_page is locked and not up-to-date.
1730          * It's safe to insert it into the page cache, because nobody would
1731          * be able to map it or use it in another way until we unlock it.
1732          */
1733
1734         xas_set(&xas, start);
1735         for (index = start; index < end; index++) {
1736                 struct page *page = xas_next(&xas);
1737
1738                 VM_BUG_ON(index != xas.xa_index);
1739                 if (is_shmem) {
1740                         if (!page) {
1741                                 /*
1742                                  * Stop if extent has been truncated or
1743                                  * hole-punched, and is now completely
1744                                  * empty.
1745                                  */
1746                                 if (index == start) {
1747                                         if (!xas_next_entry(&xas, end - 1)) {
1748                                                 result = SCAN_TRUNCATED;
1749                                                 goto xa_locked;
1750                                         }
1751                                         xas_set(&xas, index);
1752                                 }
1753                                 if (!shmem_charge(mapping->host, 1)) {
1754                                         result = SCAN_FAIL;
1755                                         goto xa_locked;
1756                                 }
1757                                 xas_store(&xas, new_page);
1758                                 nr_none++;
1759                                 continue;
1760                         }
1761
1762                         if (xa_is_value(page) || !PageUptodate(page)) {
1763                                 xas_unlock_irq(&xas);
1764                                 /* swap in or instantiate fallocated page */
1765                                 if (shmem_getpage(mapping->host, index, &page,
1766                                                   SGP_NOALLOC)) {
1767                                         result = SCAN_FAIL;
1768                                         goto xa_unlocked;
1769                                 }
1770                         } else if (trylock_page(page)) {
1771                                 get_page(page);
1772                                 xas_unlock_irq(&xas);
1773                         } else {
1774                                 result = SCAN_PAGE_LOCK;
1775                                 goto xa_locked;
1776                         }
1777                 } else {        /* !is_shmem */
1778                         if (!page || xa_is_value(page)) {
1779                                 xas_unlock_irq(&xas);
1780                                 page_cache_sync_readahead(mapping, &file->f_ra,
1781                                                           file, index,
1782                                                           end - index);
1783                                 /* drain pagevecs to help isolate_lru_page() */
1784                                 lru_add_drain();
1785                                 page = find_lock_page(mapping, index);
1786                                 if (unlikely(page == NULL)) {
1787                                         result = SCAN_FAIL;
1788                                         goto xa_unlocked;
1789                                 }
1790                         } else if (PageDirty(page)) {
1791                                 /*
1792                                  * khugepaged only works on read-only fd,
1793                                  * so this page is dirty because it hasn't
1794                                  * been flushed since first write. There
1795                                  * won't be new dirty pages.
1796                                  *
1797                                  * Trigger async flush here and hope the
1798                                  * writeback is done when khugepaged
1799                                  * revisits this page.
1800                                  *
1801                                  * This is a one-off situation. We are not
1802                                  * forcing writeback in loop.
1803                                  */
1804                                 xas_unlock_irq(&xas);
1805                                 filemap_flush(mapping);
1806                                 result = SCAN_FAIL;
1807                                 goto xa_unlocked;
1808                         } else if (PageWriteback(page)) {
1809                                 xas_unlock_irq(&xas);
1810                                 result = SCAN_FAIL;
1811                                 goto xa_unlocked;
1812                         } else if (trylock_page(page)) {
1813                                 get_page(page);
1814                                 xas_unlock_irq(&xas);
1815                         } else {
1816                                 result = SCAN_PAGE_LOCK;
1817                                 goto xa_locked;
1818                         }
1819                 }
1820
1821                 /*
1822                  * The page must be locked, so we can drop the i_pages lock
1823                  * without racing with truncate.
1824                  */
1825                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1826
1827                 /* make sure the page is up to date */
1828                 if (unlikely(!PageUptodate(page))) {
1829                         result = SCAN_FAIL;
1830                         goto out_unlock;
1831                 }
1832
1833                 /*
1834                  * If file was truncated then extended, or hole-punched, before
1835                  * we locked the first page, then a THP might be there already.
1836                  */
1837                 if (PageTransCompound(page)) {
1838                         result = SCAN_PAGE_COMPOUND;
1839                         goto out_unlock;
1840                 }
1841
1842                 if (page_mapping(page) != mapping) {
1843                         result = SCAN_TRUNCATED;
1844                         goto out_unlock;
1845                 }
1846
1847                 if (!is_shmem && (PageDirty(page) ||
1848                                   PageWriteback(page))) {
1849                         /*
1850                          * khugepaged only works on read-only fd, so this
1851                          * page is dirty because it hasn't been flushed
1852                          * since first write.
1853                          */
1854                         result = SCAN_FAIL;
1855                         goto out_unlock;
1856                 }
1857
1858                 if (isolate_lru_page(page)) {
1859                         result = SCAN_DEL_PAGE_LRU;
1860                         goto out_unlock;
1861                 }
1862
1863                 if (page_has_private(page) &&
1864                     !try_to_release_page(page, GFP_KERNEL)) {
1865                         result = SCAN_PAGE_HAS_PRIVATE;
1866                         putback_lru_page(page);
1867                         goto out_unlock;
1868                 }
1869
1870                 if (page_mapped(page))
1871                         unmap_mapping_pages(mapping, index, 1, false);
1872
1873                 xas_lock_irq(&xas);
1874                 xas_set(&xas, index);
1875
1876                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1877                 VM_BUG_ON_PAGE(page_mapped(page), page);
1878
1879                 /*
1880                  * The page is expected to have page_count() == 3:
1881                  *  - we hold a pin on it;
1882                  *  - one reference from page cache;
1883                  *  - one from isolate_lru_page;
1884                  */
1885                 if (!page_ref_freeze(page, 3)) {
1886                         result = SCAN_PAGE_COUNT;
1887                         xas_unlock_irq(&xas);
1888                         putback_lru_page(page);
1889                         goto out_unlock;
1890                 }
1891
1892                 /*
1893                  * Add the page to the list to be able to undo the collapse if
1894                  * something go wrong.
1895                  */
1896                 list_add_tail(&page->lru, &pagelist);
1897
1898                 /* Finally, replace with the new page. */
1899                 xas_store(&xas, new_page);
1900                 continue;
1901 out_unlock:
1902                 unlock_page(page);
1903                 put_page(page);
1904                 goto xa_unlocked;
1905         }
1906         nr = thp_nr_pages(new_page);
1907
1908         if (is_shmem)
1909                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1910         else {
1911                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1912                 filemap_nr_thps_inc(mapping);
1913                 /*
1914                  * Paired with smp_mb() in do_dentry_open() to ensure
1915                  * i_writecount is up to date and the update to nr_thps is
1916                  * visible. Ensures the page cache will be truncated if the
1917                  * file is opened writable.
1918                  */
1919                 smp_mb();
1920                 if (inode_is_open_for_write(mapping->host)) {
1921                         result = SCAN_FAIL;
1922                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1923                         filemap_nr_thps_dec(mapping);
1924                         goto xa_locked;
1925                 }
1926         }
1927
1928         if (nr_none) {
1929                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1930                 if (is_shmem)
1931                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1932         }
1933
1934 xa_locked:
1935         xas_unlock_irq(&xas);
1936 xa_unlocked:
1937
1938         if (result == SCAN_SUCCEED) {
1939                 struct page *page, *tmp;
1940
1941                 /*
1942                  * Replacing old pages with new one has succeeded, now we
1943                  * need to copy the content and free the old pages.
1944                  */
1945                 index = start;
1946                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1947                         while (index < page->index) {
1948                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1949                                 index++;
1950                         }
1951                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1952                                         page);
1953                         list_del(&page->lru);
1954                         page->mapping = NULL;
1955                         page_ref_unfreeze(page, 1);
1956                         ClearPageActive(page);
1957                         ClearPageUnevictable(page);
1958                         unlock_page(page);
1959                         put_page(page);
1960                         index++;
1961                 }
1962                 while (index < end) {
1963                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1964                         index++;
1965                 }
1966
1967                 SetPageUptodate(new_page);
1968                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1969                 if (is_shmem)
1970                         set_page_dirty(new_page);
1971                 lru_cache_add(new_page);
1972
1973                 /*
1974                  * Remove pte page tables, so we can re-fault the page as huge.
1975                  */
1976                 retract_page_tables(mapping, start);
1977                 *hpage = NULL;
1978
1979                 khugepaged_pages_collapsed++;
1980         } else {
1981                 struct page *page;
1982
1983                 /* Something went wrong: roll back page cache changes */
1984                 xas_lock_irq(&xas);
1985                 mapping->nrpages -= nr_none;
1986
1987                 if (is_shmem)
1988                         shmem_uncharge(mapping->host, nr_none);
1989
1990                 xas_set(&xas, start);
1991                 xas_for_each(&xas, page, end - 1) {
1992                         page = list_first_entry_or_null(&pagelist,
1993                                         struct page, lru);
1994                         if (!page || xas.xa_index < page->index) {
1995                                 if (!nr_none)
1996                                         break;
1997                                 nr_none--;
1998                                 /* Put holes back where they were */
1999                                 xas_store(&xas, NULL);
2000                                 continue;
2001                         }
2002
2003                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2004
2005                         /* Unfreeze the page. */
2006                         list_del(&page->lru);
2007                         page_ref_unfreeze(page, 2);
2008                         xas_store(&xas, page);
2009                         xas_pause(&xas);
2010                         xas_unlock_irq(&xas);
2011                         unlock_page(page);
2012                         putback_lru_page(page);
2013                         xas_lock_irq(&xas);
2014                 }
2015                 VM_BUG_ON(nr_none);
2016                 xas_unlock_irq(&xas);
2017
2018                 new_page->mapping = NULL;
2019         }
2020
2021         unlock_page(new_page);
2022 out:
2023         VM_BUG_ON(!list_empty(&pagelist));
2024         if (!IS_ERR_OR_NULL(*hpage))
2025                 mem_cgroup_uncharge(*hpage);
2026         /* TODO: tracepoints */
2027 }
2028
2029 static void khugepaged_scan_file(struct mm_struct *mm,
2030                 struct file *file, pgoff_t start, struct page **hpage)
2031 {
2032         struct page *page = NULL;
2033         struct address_space *mapping = file->f_mapping;
2034         XA_STATE(xas, &mapping->i_pages, start);
2035         int present, swap;
2036         int node = NUMA_NO_NODE;
2037         int result = SCAN_SUCCEED;
2038
2039         present = 0;
2040         swap = 0;
2041         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2042         rcu_read_lock();
2043         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2044                 if (xas_retry(&xas, page))
2045                         continue;
2046
2047                 if (xa_is_value(page)) {
2048                         if (++swap > khugepaged_max_ptes_swap) {
2049                                 result = SCAN_EXCEED_SWAP_PTE;
2050                                 break;
2051                         }
2052                         continue;
2053                 }
2054
2055                 if (PageTransCompound(page)) {
2056                         result = SCAN_PAGE_COMPOUND;
2057                         break;
2058                 }
2059
2060                 node = page_to_nid(page);
2061                 if (khugepaged_scan_abort(node)) {
2062                         result = SCAN_SCAN_ABORT;
2063                         break;
2064                 }
2065                 khugepaged_node_load[node]++;
2066
2067                 if (!PageLRU(page)) {
2068                         result = SCAN_PAGE_LRU;
2069                         break;
2070                 }
2071
2072                 if (page_count(page) !=
2073                     1 + page_mapcount(page) + page_has_private(page)) {
2074                         result = SCAN_PAGE_COUNT;
2075                         break;
2076                 }
2077
2078                 /*
2079                  * We probably should check if the page is referenced here, but
2080                  * nobody would transfer pte_young() to PageReferenced() for us.
2081                  * And rmap walk here is just too costly...
2082                  */
2083
2084                 present++;
2085
2086                 if (need_resched()) {
2087                         xas_pause(&xas);
2088                         cond_resched_rcu();
2089                 }
2090         }
2091         rcu_read_unlock();
2092
2093         if (result == SCAN_SUCCEED) {
2094                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2095                         result = SCAN_EXCEED_NONE_PTE;
2096                 } else {
2097                         node = khugepaged_find_target_node();
2098                         collapse_file(mm, file, start, hpage, node);
2099                 }
2100         }
2101
2102         /* TODO: tracepoints */
2103 }
2104 #else
2105 static void khugepaged_scan_file(struct mm_struct *mm,
2106                 struct file *file, pgoff_t start, struct page **hpage)
2107 {
2108         BUILD_BUG();
2109 }
2110
2111 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2112 {
2113 }
2114 #endif
2115
2116 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2117                                             struct page **hpage)
2118         __releases(&khugepaged_mm_lock)
2119         __acquires(&khugepaged_mm_lock)
2120 {
2121         struct mm_slot *mm_slot;
2122         struct mm_struct *mm;
2123         struct vm_area_struct *vma;
2124         int progress = 0;
2125
2126         VM_BUG_ON(!pages);
2127         lockdep_assert_held(&khugepaged_mm_lock);
2128
2129         if (khugepaged_scan.mm_slot)
2130                 mm_slot = khugepaged_scan.mm_slot;
2131         else {
2132                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2133                                      struct mm_slot, mm_node);
2134                 khugepaged_scan.address = 0;
2135                 khugepaged_scan.mm_slot = mm_slot;
2136         }
2137         spin_unlock(&khugepaged_mm_lock);
2138         khugepaged_collapse_pte_mapped_thps(mm_slot);
2139
2140         mm = mm_slot->mm;
2141         /*
2142          * Don't wait for semaphore (to avoid long wait times).  Just move to
2143          * the next mm on the list.
2144          */
2145         vma = NULL;
2146         if (unlikely(!mmap_read_trylock(mm)))
2147                 goto breakouterloop_mmap_lock;
2148         if (likely(!khugepaged_test_exit(mm)))
2149                 vma = find_vma(mm, khugepaged_scan.address);
2150
2151         progress++;
2152         for (; vma; vma = vma->vm_next) {
2153                 unsigned long hstart, hend;
2154
2155                 cond_resched();
2156                 if (unlikely(khugepaged_test_exit(mm))) {
2157                         progress++;
2158                         break;
2159                 }
2160                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2161 skip:
2162                         progress++;
2163                         continue;
2164                 }
2165                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2166                 hend = vma->vm_end & HPAGE_PMD_MASK;
2167                 if (hstart >= hend)
2168                         goto skip;
2169                 if (khugepaged_scan.address > hend)
2170                         goto skip;
2171                 if (khugepaged_scan.address < hstart)
2172                         khugepaged_scan.address = hstart;
2173                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2174                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2175                         goto skip;
2176
2177                 while (khugepaged_scan.address < hend) {
2178                         int ret;
2179                         cond_resched();
2180                         if (unlikely(khugepaged_test_exit(mm)))
2181                                 goto breakouterloop;
2182
2183                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2184                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2185                                   hend);
2186                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2187                                 struct file *file = get_file(vma->vm_file);
2188                                 pgoff_t pgoff = linear_page_index(vma,
2189                                                 khugepaged_scan.address);
2190
2191                                 mmap_read_unlock(mm);
2192                                 ret = 1;
2193                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2194                                 fput(file);
2195                         } else {
2196                                 ret = khugepaged_scan_pmd(mm, vma,
2197                                                 khugepaged_scan.address,
2198                                                 hpage);
2199                         }
2200                         /* move to next address */
2201                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2202                         progress += HPAGE_PMD_NR;
2203                         if (ret)
2204                                 /* we released mmap_lock so break loop */
2205                                 goto breakouterloop_mmap_lock;
2206                         if (progress >= pages)
2207                                 goto breakouterloop;
2208                 }
2209         }
2210 breakouterloop:
2211         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2212 breakouterloop_mmap_lock:
2213
2214         spin_lock(&khugepaged_mm_lock);
2215         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2216         /*
2217          * Release the current mm_slot if this mm is about to die, or
2218          * if we scanned all vmas of this mm.
2219          */
2220         if (khugepaged_test_exit(mm) || !vma) {
2221                 /*
2222                  * Make sure that if mm_users is reaching zero while
2223                  * khugepaged runs here, khugepaged_exit will find
2224                  * mm_slot not pointing to the exiting mm.
2225                  */
2226                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2227                         khugepaged_scan.mm_slot = list_entry(
2228                                 mm_slot->mm_node.next,
2229                                 struct mm_slot, mm_node);
2230                         khugepaged_scan.address = 0;
2231                 } else {
2232                         khugepaged_scan.mm_slot = NULL;
2233                         khugepaged_full_scans++;
2234                 }
2235
2236                 collect_mm_slot(mm_slot);
2237         }
2238
2239         return progress;
2240 }
2241
2242 static int khugepaged_has_work(void)
2243 {
2244         return !list_empty(&khugepaged_scan.mm_head) &&
2245                 khugepaged_enabled();
2246 }
2247
2248 static int khugepaged_wait_event(void)
2249 {
2250         return !list_empty(&khugepaged_scan.mm_head) ||
2251                 kthread_should_stop();
2252 }
2253
2254 static void khugepaged_do_scan(void)
2255 {
2256         struct page *hpage = NULL;
2257         unsigned int progress = 0, pass_through_head = 0;
2258         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2259         bool wait = true;
2260
2261         lru_add_drain_all();
2262
2263         while (progress < pages) {
2264                 if (!khugepaged_prealloc_page(&hpage, &wait))
2265                         break;
2266
2267                 cond_resched();
2268
2269                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2270                         break;
2271
2272                 spin_lock(&khugepaged_mm_lock);
2273                 if (!khugepaged_scan.mm_slot)
2274                         pass_through_head++;
2275                 if (khugepaged_has_work() &&
2276                     pass_through_head < 2)
2277                         progress += khugepaged_scan_mm_slot(pages - progress,
2278                                                             &hpage);
2279                 else
2280                         progress = pages;
2281                 spin_unlock(&khugepaged_mm_lock);
2282         }
2283
2284         if (!IS_ERR_OR_NULL(hpage))
2285                 put_page(hpage);
2286 }
2287
2288 static bool khugepaged_should_wakeup(void)
2289 {
2290         return kthread_should_stop() ||
2291                time_after_eq(jiffies, khugepaged_sleep_expire);
2292 }
2293
2294 static void khugepaged_wait_work(void)
2295 {
2296         if (khugepaged_has_work()) {
2297                 const unsigned long scan_sleep_jiffies =
2298                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2299
2300                 if (!scan_sleep_jiffies)
2301                         return;
2302
2303                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2304                 wait_event_freezable_timeout(khugepaged_wait,
2305                                              khugepaged_should_wakeup(),
2306                                              scan_sleep_jiffies);
2307                 return;
2308         }
2309
2310         if (khugepaged_enabled())
2311                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2312 }
2313
2314 static int khugepaged(void *none)
2315 {
2316         struct mm_slot *mm_slot;
2317
2318         set_freezable();
2319         set_user_nice(current, MAX_NICE);
2320
2321         while (!kthread_should_stop()) {
2322                 khugepaged_do_scan();
2323                 khugepaged_wait_work();
2324         }
2325
2326         spin_lock(&khugepaged_mm_lock);
2327         mm_slot = khugepaged_scan.mm_slot;
2328         khugepaged_scan.mm_slot = NULL;
2329         if (mm_slot)
2330                 collect_mm_slot(mm_slot);
2331         spin_unlock(&khugepaged_mm_lock);
2332         return 0;
2333 }
2334
2335 static void set_recommended_min_free_kbytes(void)
2336 {
2337         struct zone *zone;
2338         int nr_zones = 0;
2339         unsigned long recommended_min;
2340
2341         for_each_populated_zone(zone) {
2342                 /*
2343                  * We don't need to worry about fragmentation of
2344                  * ZONE_MOVABLE since it only has movable pages.
2345                  */
2346                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2347                         continue;
2348
2349                 nr_zones++;
2350         }
2351
2352         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2353         recommended_min = pageblock_nr_pages * nr_zones * 2;
2354
2355         /*
2356          * Make sure that on average at least two pageblocks are almost free
2357          * of another type, one for a migratetype to fall back to and a
2358          * second to avoid subsequent fallbacks of other types There are 3
2359          * MIGRATE_TYPES we care about.
2360          */
2361         recommended_min += pageblock_nr_pages * nr_zones *
2362                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2363
2364         /* don't ever allow to reserve more than 5% of the lowmem */
2365         recommended_min = min(recommended_min,
2366                               (unsigned long) nr_free_buffer_pages() / 20);
2367         recommended_min <<= (PAGE_SHIFT-10);
2368
2369         if (recommended_min > min_free_kbytes) {
2370                 if (user_min_free_kbytes >= 0)
2371                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2372                                 min_free_kbytes, recommended_min);
2373
2374                 min_free_kbytes = recommended_min;
2375         }
2376         setup_per_zone_wmarks();
2377 }
2378
2379 int start_stop_khugepaged(void)
2380 {
2381         int err = 0;
2382
2383         mutex_lock(&khugepaged_mutex);
2384         if (khugepaged_enabled()) {
2385                 if (!khugepaged_thread)
2386                         khugepaged_thread = kthread_run(khugepaged, NULL,
2387                                                         "khugepaged");
2388                 if (IS_ERR(khugepaged_thread)) {
2389                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2390                         err = PTR_ERR(khugepaged_thread);
2391                         khugepaged_thread = NULL;
2392                         goto fail;
2393                 }
2394
2395                 if (!list_empty(&khugepaged_scan.mm_head))
2396                         wake_up_interruptible(&khugepaged_wait);
2397
2398                 set_recommended_min_free_kbytes();
2399         } else if (khugepaged_thread) {
2400                 kthread_stop(khugepaged_thread);
2401                 khugepaged_thread = NULL;
2402         }
2403 fail:
2404         mutex_unlock(&khugepaged_mutex);
2405         return err;
2406 }
2407
2408 void khugepaged_min_free_kbytes_update(void)
2409 {
2410         mutex_lock(&khugepaged_mutex);
2411         if (khugepaged_enabled() && khugepaged_thread)
2412                 set_recommended_min_free_kbytes();
2413         mutex_unlock(&khugepaged_mutex);
2414 }