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                 if (pte_uffd_wp(pteval)) {
 629                         result = SCAN_PTE_UFFD_WP;
 630                         goto out;
 631                 }
 632                 page = vm_normal_page(vma, address, pteval);
 633                 if (unlikely(!page)) {
 634                         result = SCAN_PAGE_NULL;
 635                         goto out;
 636                 }
 637
 638                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 639
 640                 if (page_mapcount(page) > 1 &&
 641                                 ++shared > khugepaged_max_ptes_shared) {
 642                         result = SCAN_EXCEED_SHARED_PTE;
 643                         goto out;
 644                 }
 645
 646                 if (PageCompound(page)) {
 647                         struct page *p;
 648                         page = compound_head(page);
 649
 650                         /*
 651                          * Check if we have dealt with the compound page
 652                          * already
 653                          */
 654                         list_for_each_entry(p, compound_pagelist, lru) {
 655                                 if (page == p)
 656                                         goto next;
 657                         }
 658                 }
 659
 660                 /*
 661                  * We can do it before isolate_lru_page because the
 662                  * page can't be freed from under us. NOTE: PG_lock
 663                  * is needed to serialize against split_huge_page
 664                  * when invoked from the VM.
 665                  */
 666                 if (!trylock_page(page)) {
 667                         result = SCAN_PAGE_LOCK;
 668                         goto out;
 669                 }
 670
 671                 /*
 672                  * Check if the page has any GUP (or other external) pins.
 673                  *
 674                  * The page table that maps the page has been already unlinked
 675                  * from the page table tree and this process cannot get
 676                  * an additional pin on the page.
 677                  *
 678                  * New pins can come later if the page is shared across fork,
 679                  * but not from this process. The other process cannot write to
 680                  * the page, only trigger CoW.
 681                  */
 682                 if (!is_refcount_suitable(page)) {
 683                         unlock_page(page);
 684                         result = SCAN_PAGE_COUNT;
 685                         goto out;
 686                 }
 687                 if (!pte_write(pteval) && PageSwapCache(page) &&
 688                                 !reuse_swap_page(page, NULL)) {
 689                         /*
 690                          * Page is in the swap cache and cannot be re-used.
 691                          * It cannot be collapsed into a THP.
 692                          */
 693                         unlock_page(page);
 694                         result = SCAN_SWAP_CACHE_PAGE;
 695                         goto out;
 696                 }
 697
 698                 /*
 699                  * Isolate the page to avoid collapsing an hugepage
 700                  * currently in use by the VM.
 701                  */
 702                 if (isolate_lru_page(page)) {
 703                         unlock_page(page);
 704                         result = SCAN_DEL_PAGE_LRU;
 705                         goto out;
 706                 }
 707                 mod_node_page_state(page_pgdat(page),
 708                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 709                                 compound_nr(page));
 710                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 711                 VM_BUG_ON_PAGE(PageLRU(page), page);
 712
 713                 if (PageCompound(page))
 714                         list_add_tail(&page->lru, compound_pagelist);
 715 next:
 716                 /* There should be enough young pte to collapse the page */
 717                 if (pte_young(pteval) ||
 718                     page_is_young(page) || PageReferenced(page) ||
 719                     mmu_notifier_test_young(vma->vm_mm, address))
 720                         referenced++;
 721
 722                 if (pte_write(pteval))
 723                         writable = true;
 724         }
 725
 726         if (unlikely(!writable)) {
 727                 result = SCAN_PAGE_RO;
 728         } else if (unlikely(!referenced)) {
 729                 result = SCAN_LACK_REFERENCED_PAGE;
 730         } else {
 731                 result = SCAN_SUCCEED;
 732                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 733                                                     referenced, writable, result);
 734                 return 1;
 735         }
 736 out:
 737         release_pte_pages(pte, _pte, compound_pagelist);
 738         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 739                                             referenced, writable, result);
 740         return 0;
 741 }
 742
 743 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 744                                       struct vm_area_struct *vma,
 745                                       unsigned long address,
 746                                       spinlock_t *ptl,
 747                                       struct list_head *compound_pagelist)
 748 {
 749         struct page *src_page, *tmp;
 750         pte_t *_pte;
 751         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 752                                 _pte++, page++, address += PAGE_SIZE) {
 753                 pte_t pteval = *_pte;
 754
 755                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 756                         clear_user_highpage(page, address);
 757                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 758                         if (is_zero_pfn(pte_pfn(pteval))) {
 759                                 /*
 760                                  * ptl mostly unnecessary.
 761                                  */
 762                                 spin_lock(ptl);
 763                                 /*
 764                                  * paravirt calls inside pte_clear here are
 765                                  * superfluous.
 766                                  */
 767                                 pte_clear(vma->vm_mm, address, _pte);
 768                                 spin_unlock(ptl);
 769                         }
 770                 } else {
 771                         src_page = pte_page(pteval);
 772                         copy_user_highpage(page, src_page, address, vma);
 773                         if (!PageCompound(src_page))
 774                                 release_pte_page(src_page);
 775                         /*
 776                          * ptl mostly unnecessary, but preempt has to
 777                          * be disabled to update the per-cpu stats
 778                          * inside page_remove_rmap().
 779                          */
 780                         spin_lock(ptl);
 781                         /*
 782                          * paravirt calls inside pte_clear here are
 783                          * superfluous.
 784                          */
 785                         pte_clear(vma->vm_mm, address, _pte);
 786                         page_remove_rmap(src_page, false);
 787                         spin_unlock(ptl);
 788                         free_page_and_swap_cache(src_page);
 789                 }
 790         }
 791
 792         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 793                 list_del(&src_page->lru);
 794                 release_pte_page(src_page);
 795         }
 796 }
 797
 798 static void khugepaged_alloc_sleep(void)
 799 {
 800         DEFINE_WAIT(wait);
 801
 802         add_wait_queue(&khugepaged_wait, &wait);
 803         freezable_schedule_timeout_interruptible(
 804                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 805         remove_wait_queue(&khugepaged_wait, &wait);
 806 }
 807
 808 static int khugepaged_node_load[MAX_NUMNODES];
 809
 810 static bool khugepaged_scan_abort(int nid)
 811 {
 812         int i;
 813
 814         /*
 815          * If node_reclaim_mode is disabled, then no extra effort is made to
 816          * allocate memory locally.
 817          */
 818         if (!node_reclaim_enabled())
 819                 return false;
 820
 821         /* If there is a count for this node already, it must be acceptable */
 822         if (khugepaged_node_load[nid])
 823                 return false;
 824
 825         for (i = 0; i < MAX_NUMNODES; i++) {
 826                 if (!khugepaged_node_load[i])
 827                         continue;
 828                 if (node_distance(nid, i) > node_reclaim_distance)
 829                         return true;
 830         }
 831         return false;
 832 }
 833
 834 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 835 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 836 {
 837         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 838 }
 839
 840 #ifdef CONFIG_NUMA
 841 static int khugepaged_find_target_node(void)
 842 {
 843         static int last_khugepaged_target_node = NUMA_NO_NODE;
 844         int nid, target_node = 0, max_value = 0;
 845
 846         /* find first node with max normal pages hit */
 847         for (nid = 0; nid < MAX_NUMNODES; nid++)
 848                 if (khugepaged_node_load[nid] > max_value) {
 849                         max_value = khugepaged_node_load[nid];
 850                         target_node = nid;
 851                 }
 852
 853         /* do some balance if several nodes have the same hit record */
 854         if (target_node <= last_khugepaged_target_node)
 855                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 856                                 nid++)
 857                         if (max_value == khugepaged_node_load[nid]) {
 858                                 target_node = nid;
 859                                 break;
 860                         }
 861
 862         last_khugepaged_target_node = target_node;
 863         return target_node;
 864 }
 865
 866 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 867 {
 868         if (IS_ERR(*hpage)) {
 869                 if (!*wait)
 870                         return false;
 871
 872                 *wait = false;
 873                 *hpage = NULL;
 874                 khugepaged_alloc_sleep();
 875         } else if (*hpage) {
 876                 put_page(*hpage);
 877                 *hpage = NULL;
 878         }
 879
 880         return true;
 881 }
 882
 883 static struct page *
 884 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 885 {
 886         VM_BUG_ON_PAGE(*hpage, *hpage);
 887
 888         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 889         if (unlikely(!*hpage)) {
 890                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 891                 *hpage = ERR_PTR(-ENOMEM);
 892                 return NULL;
 893         }
 894
 895         prep_transhuge_page(*hpage);
 896         count_vm_event(THP_COLLAPSE_ALLOC);
 897         return *hpage;
 898 }
 899 #else
 900 static int khugepaged_find_target_node(void)
 901 {
 902         return 0;
 903 }
 904
 905 static inline struct page *alloc_khugepaged_hugepage(void)
 906 {
 907         struct page *page;
 908
 909         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 910                            HPAGE_PMD_ORDER);
 911         if (page)
 912                 prep_transhuge_page(page);
 913         return page;
 914 }
 915
 916 static struct page *khugepaged_alloc_hugepage(bool *wait)
 917 {
 918         struct page *hpage;
 919
 920         do {
 921                 hpage = alloc_khugepaged_hugepage();
 922                 if (!hpage) {
 923                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 924                         if (!*wait)
 925                                 return NULL;
 926
 927                         *wait = false;
 928                         khugepaged_alloc_sleep();
 929                 } else
 930                         count_vm_event(THP_COLLAPSE_ALLOC);
 931         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 932
 933         return hpage;
 934 }
 935
 936 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 937 {
 938         /*
 939          * If the hpage allocated earlier was briefly exposed in page cache
 940          * before collapse_file() failed, it is possible that racing lookups
 941          * have not yet completed, and would then be unpleasantly surprised by
 942          * finding the hpage reused for the same mapping at a different offset.
 943          * Just release the previous allocation if there is any danger of that.
 944          */
 945         if (*hpage && page_count(*hpage) > 1) {
 946                 put_page(*hpage);
 947                 *hpage = NULL;
 948         }
 949
 950         if (!*hpage)
 951                 *hpage = khugepaged_alloc_hugepage(wait);
 952
 953         if (unlikely(!*hpage))
 954                 return false;
 955
 956         return true;
 957 }
 958
 959 static struct page *
 960 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 961 {
 962         VM_BUG_ON(!*hpage);
 963
 964         return  *hpage;
 965 }
 966 #endif
 967
 968 /*
 969  * If mmap_lock temporarily dropped, revalidate vma
 970  * before taking mmap_lock.
 971  * Return 0 if succeeds, otherwise return none-zero
 972  * value (scan code).
 973  */
 974
 975 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 976                 struct vm_area_struct **vmap)
 977 {
 978         struct vm_area_struct *vma;
 979         unsigned long hstart, hend;
 980
 981         if (unlikely(khugepaged_test_exit(mm)))
 982                 return SCAN_ANY_PROCESS;
 983
 984         *vmap = vma = find_vma(mm, address);
 985         if (!vma)
 986                 return SCAN_VMA_NULL;
 987
 988         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 989         hend = vma->vm_end & HPAGE_PMD_MASK;
 990         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 991                 return SCAN_ADDRESS_RANGE;
 992         if (!hugepage_vma_check(vma, vma->vm_flags))
 993                 return SCAN_VMA_CHECK;
 994         /* Anon VMA expected */
 995         if (!vma->anon_vma || vma->vm_ops)
 996                 return SCAN_VMA_CHECK;
 997         return 0;
 998 }
 999
1000 /*
1001  * Bring missing pages in from swap, to complete THP collapse.
1002  * Only done if khugepaged_scan_pmd believes it is worthwhile.
1003  *
1004  * Called and returns without pte mapped or spinlocks held,
1005  * but with mmap_lock held to protect against vma changes.
1006  */
1007
1008 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1009                                         struct vm_area_struct *vma,
1010                                         unsigned long haddr, pmd_t *pmd,
1011                                         int referenced)
1012 {
1013         int swapped_in = 0;
1014         vm_fault_t ret = 0;
1015         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1016
1017         for (address = haddr; address < end; address += PAGE_SIZE) {
1018                 struct vm_fault vmf = {
1019                         .vma = vma,
1020                         .address = address,
1021                         .pgoff = linear_page_index(vma, haddr),
1022                         .flags = FAULT_FLAG_ALLOW_RETRY,
1023                         .pmd = pmd,
1024                 };
1025
1026                 vmf.pte = pte_offset_map(pmd, address);
1027                 vmf.orig_pte = *vmf.pte;
1028                 if (!is_swap_pte(vmf.orig_pte)) {
1029                         pte_unmap(vmf.pte);
1030                         continue;
1031                 }
1032                 swapped_in++;
1033                 ret = do_swap_page(&vmf);
1034
1035                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1036                 if (ret & VM_FAULT_RETRY) {
1037                         mmap_read_lock(mm);
1038                         if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1039                                 /* vma is no longer available, don't continue to swapin */
1040                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1041                                 return false;
1042                         }
1043                         /* check if the pmd is still valid */
1044                         if (mm_find_pmd(mm, haddr) != pmd) {
1045                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1046                                 return false;
1047                         }
1048                 }
1049                 if (ret & VM_FAULT_ERROR) {
1050                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1051                         return false;
1052                 }
1053         }
1054
1055         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1056         if (swapped_in)
1057                 lru_add_drain();
1058
1059         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1060         return true;
1061 }
1062
1063 static void collapse_huge_page(struct mm_struct *mm,
1064                                    unsigned long address,
1065                                    struct page **hpage,
1066                                    int node, int referenced, int unmapped)
1067 {
1068         LIST_HEAD(compound_pagelist);
1069         pmd_t *pmd, _pmd;
1070         pte_t *pte;
1071         pgtable_t pgtable;
1072         struct page *new_page;
1073         spinlock_t *pmd_ptl, *pte_ptl;
1074         int isolated = 0, result = 0;
1075         struct vm_area_struct *vma;
1076         struct mmu_notifier_range range;
1077         gfp_t gfp;
1078
1079         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1080
1081         /* Only allocate from the target node */
1082         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1083
1084         /*
1085          * Before allocating the hugepage, release the mmap_lock read lock.
1086          * The allocation can take potentially a long time if it involves
1087          * sync compaction, and we do not need to hold the mmap_lock during
1088          * that. We will recheck the vma after taking it again in write mode.
1089          */
1090         mmap_read_unlock(mm);
1091         new_page = khugepaged_alloc_page(hpage, gfp, node);
1092         if (!new_page) {
1093                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1094                 goto out_nolock;
1095         }
1096
1097         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1098                 result = SCAN_CGROUP_CHARGE_FAIL;
1099                 goto out_nolock;
1100         }
1101         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1102
1103         mmap_read_lock(mm);
1104         result = hugepage_vma_revalidate(mm, address, &vma);
1105         if (result) {
1106                 mmap_read_unlock(mm);
1107                 goto out_nolock;
1108         }
1109
1110         pmd = mm_find_pmd(mm, address);
1111         if (!pmd) {
1112                 result = SCAN_PMD_NULL;
1113                 mmap_read_unlock(mm);
1114                 goto out_nolock;
1115         }
1116
1117         /*
1118          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1119          * If it fails, we release mmap_lock and jump out_nolock.
1120          * Continuing to collapse causes inconsistency.
1121          */
1122         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1123                                                      pmd, referenced)) {
1124                 mmap_read_unlock(mm);
1125                 goto out_nolock;
1126         }
1127
1128         mmap_read_unlock(mm);
1129         /*
1130          * Prevent all access to pagetables with the exception of
1131          * gup_fast later handled by the ptep_clear_flush and the VM
1132          * handled by the anon_vma lock + PG_lock.
1133          */
1134         mmap_write_lock(mm);
1135         result = hugepage_vma_revalidate(mm, address, &vma);
1136         if (result)
1137                 goto out_up_write;
1138         /* check if the pmd is still valid */
1139         if (mm_find_pmd(mm, address) != pmd)
1140                 goto out_up_write;
1141
1142         anon_vma_lock_write(vma->anon_vma);
1143
1144         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1145                                 address, address + HPAGE_PMD_SIZE);
1146         mmu_notifier_invalidate_range_start(&range);
1147
1148         pte = pte_offset_map(pmd, address);
1149         pte_ptl = pte_lockptr(mm, pmd);
1150
1151         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1152         /*
1153          * This removes any huge TLB entry from the CPU so we won't allow
1154          * huge and small TLB entries for the same virtual address to
1155          * avoid the risk of CPU bugs in that area.
1156          *
1157          * Parallel fast GUP is fine since fast GUP will back off when
1158          * it detects PMD is changed.
1159          */
1160         _pmd = pmdp_collapse_flush(vma, address, pmd);
1161         spin_unlock(pmd_ptl);
1162         mmu_notifier_invalidate_range_end(&range);
1163         tlb_remove_table_sync_one();
1164
1165         spin_lock(pte_ptl);
1166         isolated = __collapse_huge_page_isolate(vma, address, pte,
1167                         &compound_pagelist);
1168         spin_unlock(pte_ptl);
1169
1170         if (unlikely(!isolated)) {
1171                 pte_unmap(pte);
1172                 spin_lock(pmd_ptl);
1173                 BUG_ON(!pmd_none(*pmd));
1174                 /*
1175                  * We can only use set_pmd_at when establishing
1176                  * hugepmds and never for establishing regular pmds that
1177                  * points to regular pagetables. Use pmd_populate for that
1178                  */
1179                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1180                 spin_unlock(pmd_ptl);
1181                 anon_vma_unlock_write(vma->anon_vma);
1182                 result = SCAN_FAIL;
1183                 goto out_up_write;
1184         }
1185
1186         /*
1187          * All pages are isolated and locked so anon_vma rmap
1188          * can't run anymore.
1189          */
1190         anon_vma_unlock_write(vma->anon_vma);
1191
1192         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1193                         &compound_pagelist);
1194         pte_unmap(pte);
1195         /*
1196          * spin_lock() below is not the equivalent of smp_wmb(), but
1197          * the smp_wmb() inside __SetPageUptodate() can be reused to
1198          * avoid the copy_huge_page writes to become visible after
1199          * the set_pmd_at() write.
1200          */
1201         __SetPageUptodate(new_page);
1202         pgtable = pmd_pgtable(_pmd);
1203
1204         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1205         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1206
1207         spin_lock(pmd_ptl);
1208         BUG_ON(!pmd_none(*pmd));
1209         page_add_new_anon_rmap(new_page, vma, address, true);
1210         lru_cache_add_inactive_or_unevictable(new_page, vma);
1211         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1212         set_pmd_at(mm, address, pmd, _pmd);
1213         update_mmu_cache_pmd(vma, address, pmd);
1214         spin_unlock(pmd_ptl);
1215
1216         *hpage = NULL;
1217
1218         khugepaged_pages_collapsed++;
1219         result = SCAN_SUCCEED;
1220 out_up_write:
1221         mmap_write_unlock(mm);
1222 out_nolock:
1223         if (!IS_ERR_OR_NULL(*hpage))
1224                 mem_cgroup_uncharge(*hpage);
1225         trace_mm_collapse_huge_page(mm, isolated, result);
1226         return;
1227 }
1228
1229 static int khugepaged_scan_pmd(struct mm_struct *mm,
1230                                struct vm_area_struct *vma,
1231                                unsigned long address,
1232                                struct page **hpage)
1233 {
1234         pmd_t *pmd;
1235         pte_t *pte, *_pte;
1236         int ret = 0, result = 0, referenced = 0;
1237         int none_or_zero = 0, shared = 0;
1238         struct page *page = NULL;
1239         unsigned long _address;
1240         spinlock_t *ptl;
1241         int node = NUMA_NO_NODE, unmapped = 0;
1242         bool writable = false;
1243
1244         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1245
1246         pmd = mm_find_pmd(mm, address);
1247         if (!pmd) {
1248                 result = SCAN_PMD_NULL;
1249                 goto out;
1250         }
1251
1252         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1253         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1254         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1255              _pte++, _address += PAGE_SIZE) {
1256                 pte_t pteval = *_pte;
1257                 if (is_swap_pte(pteval)) {
1258                         if (++unmapped <= khugepaged_max_ptes_swap) {
1259                                 /*
1260                                  * Always be strict with uffd-wp
1261                                  * enabled swap entries.  Please see
1262                                  * comment below for pte_uffd_wp().
1263                                  */
1264                                 if (pte_swp_uffd_wp(pteval)) {
1265                                         result = SCAN_PTE_UFFD_WP;
1266                                         goto out_unmap;
1267                                 }
1268                                 continue;
1269                         } else {
1270                                 result = SCAN_EXCEED_SWAP_PTE;
1271                                 goto out_unmap;
1272                         }
1273                 }
1274                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1275                         if (!userfaultfd_armed(vma) &&
1276                             ++none_or_zero <= khugepaged_max_ptes_none) {
1277                                 continue;
1278                         } else {
1279                                 result = SCAN_EXCEED_NONE_PTE;
1280                                 goto out_unmap;
1281                         }
1282                 }
1283                 if (pte_uffd_wp(pteval)) {
1284                         /*
1285                          * Don't collapse the page if any of the small
1286                          * PTEs are armed with uffd write protection.
1287                          * Here we can also mark the new huge pmd as
1288                          * write protected if any of the small ones is
1289                          * marked but that could bring unknown
1290                          * userfault messages that falls outside of
1291                          * the registered range.  So, just be simple.
1292                          */
1293                         result = SCAN_PTE_UFFD_WP;
1294                         goto out_unmap;
1295                 }
1296                 if (pte_write(pteval))
1297                         writable = true;
1298
1299                 page = vm_normal_page(vma, _address, pteval);
1300                 if (unlikely(!page)) {
1301                         result = SCAN_PAGE_NULL;
1302                         goto out_unmap;
1303                 }
1304
1305                 if (page_mapcount(page) > 1 &&
1306                                 ++shared > khugepaged_max_ptes_shared) {
1307                         result = SCAN_EXCEED_SHARED_PTE;
1308                         goto out_unmap;
1309                 }
1310
1311                 page = compound_head(page);
1312
1313                 /*
1314                  * Record which node the original page is from and save this
1315                  * information to khugepaged_node_load[].
1316                  * Khupaged will allocate hugepage from the node has the max
1317                  * hit record.
1318                  */
1319                 node = page_to_nid(page);
1320                 if (khugepaged_scan_abort(node)) {
1321                         result = SCAN_SCAN_ABORT;
1322                         goto out_unmap;
1323                 }
1324                 khugepaged_node_load[node]++;
1325                 if (!PageLRU(page)) {
1326                         result = SCAN_PAGE_LRU;
1327                         goto out_unmap;
1328                 }
1329                 if (PageLocked(page)) {
1330                         result = SCAN_PAGE_LOCK;
1331                         goto out_unmap;
1332                 }
1333                 if (!PageAnon(page)) {
1334                         result = SCAN_PAGE_ANON;
1335                         goto out_unmap;
1336                 }
1337
1338                 /*
1339                  * Check if the page has any GUP (or other external) pins.
1340                  *
1341                  * Here the check is racy it may see totmal_mapcount > refcount
1342                  * in some cases.
1343                  * For example, one process with one forked child process.
1344                  * The parent has the PMD split due to MADV_DONTNEED, then
1345                  * the child is trying unmap the whole PMD, but khugepaged
1346                  * may be scanning the parent between the child has
1347                  * PageDoubleMap flag cleared and dec the mapcount.  So
1348                  * khugepaged may see total_mapcount > refcount.
1349                  *
1350                  * But such case is ephemeral we could always retry collapse
1351                  * later.  However it may report false positive if the page
1352                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1353                  * will be done again later the risk seems low.
1354                  */
1355                 if (!is_refcount_suitable(page)) {
1356                         result = SCAN_PAGE_COUNT;
1357                         goto out_unmap;
1358                 }
1359                 if (pte_young(pteval) ||
1360                     page_is_young(page) || PageReferenced(page) ||
1361                     mmu_notifier_test_young(vma->vm_mm, address))
1362                         referenced++;
1363         }
1364         if (!writable) {
1365                 result = SCAN_PAGE_RO;
1366         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1367                 result = SCAN_LACK_REFERENCED_PAGE;
1368         } else {
1369                 result = SCAN_SUCCEED;
1370                 ret = 1;
1371         }
1372 out_unmap:
1373         pte_unmap_unlock(pte, ptl);
1374         if (ret) {
1375                 node = khugepaged_find_target_node();
1376                 /* collapse_huge_page will return with the mmap_lock released */
1377                 collapse_huge_page(mm, address, hpage, node,
1378                                 referenced, unmapped);
1379         }
1380 out:
1381         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1382                                      none_or_zero, result, unmapped);
1383         return ret;
1384 }
1385
1386 static void collect_mm_slot(struct mm_slot *mm_slot)
1387 {
1388         struct mm_struct *mm = mm_slot->mm;
1389
1390         lockdep_assert_held(&khugepaged_mm_lock);
1391
1392         if (khugepaged_test_exit(mm)) {
1393                 /* free mm_slot */
1394                 hash_del(&mm_slot->hash);
1395                 list_del(&mm_slot->mm_node);
1396
1397                 /*
1398                  * Not strictly needed because the mm exited already.
1399                  *
1400                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1401                  */
1402
1403                 /* khugepaged_mm_lock actually not necessary for the below */
1404                 free_mm_slot(mm_slot);
1405                 mmdrop(mm);
1406         }
1407 }
1408
1409 #ifdef CONFIG_SHMEM
1410 /*
1411  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1412  * khugepaged should try to collapse the page table.
1413  */
1414 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1415                                          unsigned long addr)
1416 {
1417         struct mm_slot *mm_slot;
1418
1419         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1420
1421         spin_lock(&khugepaged_mm_lock);
1422         mm_slot = get_mm_slot(mm);
1423         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1424                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1425         spin_unlock(&khugepaged_mm_lock);
1426         return 0;
1427 }
1428
1429 /**
1430  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1431  * address haddr.
1432  *
1433  * @mm: process address space where collapse happens
1434  * @addr: THP collapse address
1435  *
1436  * This function checks whether all the PTEs in the PMD are pointing to the
1437  * right THP. If so, retract the page table so the THP can refault in with
1438  * as pmd-mapped.
1439  */
1440 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1441 {
1442         unsigned long haddr = addr & HPAGE_PMD_MASK;
1443         struct vm_area_struct *vma = find_vma(mm, haddr);
1444         struct page *hpage;
1445         pte_t *start_pte, *pte;
1446         pmd_t *pmd, _pmd;
1447         spinlock_t *ptl;
1448         int count = 0;
1449         int i;
1450         struct mmu_notifier_range range;
1451
1452         if (!vma || !vma->vm_file ||
1453             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1454                 return;
1455
1456         /*
1457          * This vm_flags may not have VM_HUGEPAGE if the page was not
1458          * collapsed by this mm. But we can still collapse if the page is
1459          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1460          * will not fail the vma for missing VM_HUGEPAGE
1461          */
1462         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1463                 return;
1464
1465         hpage = find_lock_page(vma->vm_file->f_mapping,
1466                                linear_page_index(vma, haddr));
1467         if (!hpage)
1468                 return;
1469
1470         if (!PageHead(hpage))
1471                 goto drop_hpage;
1472
1473         pmd = mm_find_pmd(mm, haddr);
1474         if (!pmd)
1475                 goto drop_hpage;
1476
1477         /*
1478          * We need to lock the mapping so that from here on, only GUP-fast and
1479          * hardware page walks can access the parts of the page tables that
1480          * we're operating on.
1481          */
1482         i_mmap_lock_write(vma->vm_file->f_mapping);
1483
1484         /*
1485          * This spinlock should be unnecessary: Nobody else should be accessing
1486          * the page tables under spinlock protection here, only
1487          * lockless_pages_from_mm() and the hardware page walker can access page
1488          * tables while all the high-level locks are held in write mode.
1489          */
1490         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1491
1492         /* step 1: check all mapped PTEs are to the right huge page */
1493         for (i = 0, addr = haddr, pte = start_pte;
1494              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1495                 struct page *page;
1496
1497                 /* empty pte, skip */
1498                 if (pte_none(*pte))
1499                         continue;
1500
1501                 /* page swapped out, abort */
1502                 if (!pte_present(*pte))
1503                         goto abort;
1504
1505                 page = vm_normal_page(vma, addr, *pte);
1506
1507                 /*
1508                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1509                  * page table, but the new page will not be a subpage of hpage.
1510                  */
1511                 if (hpage + i != page)
1512                         goto abort;
1513                 count++;
1514         }
1515
1516         /* step 2: adjust rmap */
1517         for (i = 0, addr = haddr, pte = start_pte;
1518              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1519                 struct page *page;
1520
1521                 if (pte_none(*pte))
1522                         continue;
1523                 page = vm_normal_page(vma, addr, *pte);
1524                 page_remove_rmap(page, false);
1525         }
1526
1527         pte_unmap_unlock(start_pte, ptl);
1528
1529         /* step 3: set proper refcount and mm_counters. */
1530         if (count) {
1531                 page_ref_sub(hpage, count);
1532                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1533         }
1534
1535         /* step 4: collapse pmd */
1536         /* we make no change to anon, but protect concurrent anon page lookup */
1537         if (vma->anon_vma)
1538                 anon_vma_lock_write(vma->anon_vma);
1539
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         if (vma->anon_vma)
1550                 anon_vma_unlock_write(vma->anon_vma);
1551         i_mmap_unlock_write(vma->vm_file->f_mapping);
1552
1553 drop_hpage:
1554         unlock_page(hpage);
1555         put_page(hpage);
1556         return;
1557
1558 abort:
1559         pte_unmap_unlock(start_pte, ptl);
1560         i_mmap_unlock_write(vma->vm_file->f_mapping);
1561         goto drop_hpage;
1562 }
1563
1564 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1565 {
1566         struct mm_struct *mm = mm_slot->mm;
1567         int i;
1568
1569         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1570                 return;
1571
1572         if (!mmap_write_trylock(mm))
1573                 return;
1574
1575         if (unlikely(khugepaged_test_exit(mm)))
1576                 goto out;
1577
1578         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1579                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1580
1581 out:
1582         mm_slot->nr_pte_mapped_thp = 0;
1583         mmap_write_unlock(mm);
1584 }
1585
1586 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1587 {
1588         struct vm_area_struct *vma;
1589         struct mm_struct *mm;
1590         unsigned long addr;
1591         pmd_t *pmd, _pmd;
1592
1593         i_mmap_lock_write(mapping);
1594         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1595                 /*
1596                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1597                  * got written to. These VMAs are likely not worth investing
1598                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1599                  * later.
1600                  *
1601                  * Not that vma->anon_vma check is racy: it can be set up after
1602                  * the check but before we took mmap_lock by the fault path.
1603                  * But page lock would prevent establishing any new ptes of the
1604                  * page, so we are safe.
1605                  *
1606                  * An alternative would be drop the check, but check that page
1607                  * table is clear before calling pmdp_collapse_flush() under
1608                  * ptl. It has higher chance to recover THP for the VMA, but
1609                  * has higher cost too. It would also probably require locking
1610                  * the anon_vma.
1611                  */
1612                 if (vma->anon_vma)
1613                         continue;
1614                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1615                 if (addr & ~HPAGE_PMD_MASK)
1616                         continue;
1617                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1618                         continue;
1619                 mm = vma->vm_mm;
1620                 pmd = mm_find_pmd(mm, addr);
1621                 if (!pmd)
1622                         continue;
1623                 /*
1624                  * We need exclusive mmap_lock to retract page table.
1625                  *
1626                  * We use trylock due to lock inversion: we need to acquire
1627                  * mmap_lock while holding page lock. Fault path does it in
1628                  * reverse order. Trylock is a way to avoid deadlock.
1629                  */
1630                 if (mmap_write_trylock(mm)) {
1631                         if (!khugepaged_test_exit(mm)) {
1632                                 struct mmu_notifier_range range;
1633
1634                                 mmu_notifier_range_init(&range,
1635                                                         MMU_NOTIFY_CLEAR, 0,
1636                                                         NULL, mm, addr,
1637                                                         addr + HPAGE_PMD_SIZE);
1638                                 mmu_notifier_invalidate_range_start(&range);
1639                                 /* assume page table is clear */
1640                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1641                                 mm_dec_nr_ptes(mm);
1642                                 tlb_remove_table_sync_one();
1643                                 pte_free(mm, pmd_pgtable(_pmd));
1644                                 mmu_notifier_invalidate_range_end(&range);
1645                         }
1646                         mmap_write_unlock(mm);
1647                 } else {
1648                         /* Try again later */
1649                         khugepaged_add_pte_mapped_thp(mm, addr);
1650                 }
1651         }
1652         i_mmap_unlock_write(mapping);
1653 }
1654
1655 /**
1656  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1657  *
1658  * @mm: process address space where collapse happens
1659  * @file: file that collapse on
1660  * @start: collapse start address
1661  * @hpage: new allocated huge page for collapse
1662  * @node: appointed node the new huge page allocate from
1663  *
1664  * Basic scheme is simple, details are more complex:
1665  *  - allocate and lock a new huge page;
1666  *  - scan page cache replacing old pages with the new one
1667  *    + swap/gup in pages if necessary;
1668  *    + fill in gaps;
1669  *    + keep old pages around in case rollback is required;
1670  *  - if replacing succeeds:
1671  *    + copy data over;
1672  *    + free old pages;
1673  *    + unlock huge page;
1674  *  - if replacing failed;
1675  *    + put all pages back and unfreeze them;
1676  *    + restore gaps in the page cache;
1677  *    + unlock and free huge page;
1678  */
1679 static void collapse_file(struct mm_struct *mm,
1680                 struct file *file, pgoff_t start,
1681                 struct page **hpage, int node)
1682 {
1683         struct address_space *mapping = file->f_mapping;
1684         gfp_t gfp;
1685         struct page *new_page;
1686         pgoff_t index, end = start + HPAGE_PMD_NR;
1687         LIST_HEAD(pagelist);
1688         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1689         int nr_none = 0, result = SCAN_SUCCEED;
1690         bool is_shmem = shmem_file(file);
1691         int nr;
1692
1693         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1694         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1695
1696         /* Only allocate from the target node */
1697         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1698
1699         new_page = khugepaged_alloc_page(hpage, gfp, node);
1700         if (!new_page) {
1701                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1702                 goto out;
1703         }
1704
1705         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1706                 result = SCAN_CGROUP_CHARGE_FAIL;
1707                 goto out;
1708         }
1709         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1710
1711         /* This will be less messy when we use multi-index entries */
1712         do {
1713                 xas_lock_irq(&xas);
1714                 xas_create_range(&xas);
1715                 if (!xas_error(&xas))
1716                         break;
1717                 xas_unlock_irq(&xas);
1718                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1719                         result = SCAN_FAIL;
1720                         goto out;
1721                 }
1722         } while (1);
1723
1724         __SetPageLocked(new_page);
1725         if (is_shmem)
1726                 __SetPageSwapBacked(new_page);
1727         new_page->index = start;
1728         new_page->mapping = mapping;
1729
1730         /*
1731          * At this point the new_page is locked and not up-to-date.
1732          * It's safe to insert it into the page cache, because nobody would
1733          * be able to map it or use it in another way until we unlock it.
1734          */
1735
1736         xas_set(&xas, start);
1737         for (index = start; index < end; index++) {
1738                 struct page *page = xas_next(&xas);
1739
1740                 VM_BUG_ON(index != xas.xa_index);
1741                 if (is_shmem) {
1742                         if (!page) {
1743                                 /*
1744                                  * Stop if extent has been truncated or
1745                                  * hole-punched, and is now completely
1746                                  * empty.
1747                                  */
1748                                 if (index == start) {
1749                                         if (!xas_next_entry(&xas, end - 1)) {
1750                                                 result = SCAN_TRUNCATED;
1751                                                 goto xa_locked;
1752                                         }
1753                                         xas_set(&xas, index);
1754                                 }
1755                                 if (!shmem_charge(mapping->host, 1)) {
1756                                         result = SCAN_FAIL;
1757                                         goto xa_locked;
1758                                 }
1759                                 xas_store(&xas, new_page);
1760                                 nr_none++;
1761                                 continue;
1762                         }
1763
1764                         if (xa_is_value(page) || !PageUptodate(page)) {
1765                                 xas_unlock_irq(&xas);
1766                                 /* swap in or instantiate fallocated page */
1767                                 if (shmem_getpage(mapping->host, index, &page,
1768                                                   SGP_NOALLOC)) {
1769                                         result = SCAN_FAIL;
1770                                         goto xa_unlocked;
1771                                 }
1772                         } else if (trylock_page(page)) {
1773                                 get_page(page);
1774                                 xas_unlock_irq(&xas);
1775                         } else {
1776                                 result = SCAN_PAGE_LOCK;
1777                                 goto xa_locked;
1778                         }
1779                 } else {        /* !is_shmem */
1780                         if (!page || xa_is_value(page)) {
1781                                 xas_unlock_irq(&xas);
1782                                 page_cache_sync_readahead(mapping, &file->f_ra,
1783                                                           file, index,
1784                                                           end - index);
1785                                 /* drain pagevecs to help isolate_lru_page() */
1786                                 lru_add_drain();
1787                                 page = find_lock_page(mapping, index);
1788                                 if (unlikely(page == NULL)) {
1789                                         result = SCAN_FAIL;
1790                                         goto xa_unlocked;
1791                                 }
1792                         } else if (PageDirty(page)) {
1793                                 /*
1794                                  * khugepaged only works on read-only fd,
1795                                  * so this page is dirty because it hasn't
1796                                  * been flushed since first write. There
1797                                  * won't be new dirty pages.
1798                                  *
1799                                  * Trigger async flush here and hope the
1800                                  * writeback is done when khugepaged
1801                                  * revisits this page.
1802                                  *
1803                                  * This is a one-off situation. We are not
1804                                  * forcing writeback in loop.
1805                                  */
1806                                 xas_unlock_irq(&xas);
1807                                 filemap_flush(mapping);
1808                                 result = SCAN_FAIL;
1809                                 goto xa_unlocked;
1810                         } else if (PageWriteback(page)) {
1811                                 xas_unlock_irq(&xas);
1812                                 result = SCAN_FAIL;
1813                                 goto xa_unlocked;
1814                         } else if (trylock_page(page)) {
1815                                 get_page(page);
1816                                 xas_unlock_irq(&xas);
1817                         } else {
1818                                 result = SCAN_PAGE_LOCK;
1819                                 goto xa_locked;
1820                         }
1821                 }
1822
1823                 /*
1824                  * The page must be locked, so we can drop the i_pages lock
1825                  * without racing with truncate.
1826                  */
1827                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1828
1829                 /* make sure the page is up to date */
1830                 if (unlikely(!PageUptodate(page))) {
1831                         result = SCAN_FAIL;
1832                         goto out_unlock;
1833                 }
1834
1835                 /*
1836                  * If file was truncated then extended, or hole-punched, before
1837                  * we locked the first page, then a THP might be there already.
1838                  */
1839                 if (PageTransCompound(page)) {
1840                         result = SCAN_PAGE_COMPOUND;
1841                         goto out_unlock;
1842                 }
1843
1844                 if (page_mapping(page) != mapping) {
1845                         result = SCAN_TRUNCATED;
1846                         goto out_unlock;
1847                 }
1848
1849                 if (!is_shmem && (PageDirty(page) ||
1850                                   PageWriteback(page))) {
1851                         /*
1852                          * khugepaged only works on read-only fd, so this
1853                          * page is dirty because it hasn't been flushed
1854                          * since first write.
1855                          */
1856                         result = SCAN_FAIL;
1857                         goto out_unlock;
1858                 }
1859
1860                 if (isolate_lru_page(page)) {
1861                         result = SCAN_DEL_PAGE_LRU;
1862                         goto out_unlock;
1863                 }
1864
1865                 if (page_has_private(page) &&
1866                     !try_to_release_page(page, GFP_KERNEL)) {
1867                         result = SCAN_PAGE_HAS_PRIVATE;
1868                         putback_lru_page(page);
1869                         goto out_unlock;
1870                 }
1871
1872                 if (page_mapped(page))
1873                         unmap_mapping_pages(mapping, index, 1, false);
1874
1875                 xas_lock_irq(&xas);
1876                 xas_set(&xas, index);
1877
1878                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1879                 VM_BUG_ON_PAGE(page_mapped(page), page);
1880
1881                 /*
1882                  * The page is expected to have page_count() == 3:
1883                  *  - we hold a pin on it;
1884                  *  - one reference from page cache;
1885                  *  - one from isolate_lru_page;
1886                  */
1887                 if (!page_ref_freeze(page, 3)) {
1888                         result = SCAN_PAGE_COUNT;
1889                         xas_unlock_irq(&xas);
1890                         putback_lru_page(page);
1891                         goto out_unlock;
1892                 }
1893
1894                 /*
1895                  * Add the page to the list to be able to undo the collapse if
1896                  * something go wrong.
1897                  */
1898                 list_add_tail(&page->lru, &pagelist);
1899
1900                 /* Finally, replace with the new page. */
1901                 xas_store(&xas, new_page);
1902                 continue;
1903 out_unlock:
1904                 unlock_page(page);
1905                 put_page(page);
1906                 goto xa_unlocked;
1907         }
1908         nr = thp_nr_pages(new_page);
1909
1910         if (is_shmem)
1911                 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1912         else {
1913                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1914                 filemap_nr_thps_inc(mapping);
1915                 /*
1916                  * Paired with smp_mb() in do_dentry_open() to ensure
1917                  * i_writecount is up to date and the update to nr_thps is
1918                  * visible. Ensures the page cache will be truncated if the
1919                  * file is opened writable.
1920                  */
1921                 smp_mb();
1922                 if (inode_is_open_for_write(mapping->host)) {
1923                         result = SCAN_FAIL;
1924                         __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1925                         filemap_nr_thps_dec(mapping);
1926                         goto xa_locked;
1927                 }
1928         }
1929
1930         if (nr_none) {
1931                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1932                 if (is_shmem)
1933                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1934         }
1935
1936 xa_locked:
1937         xas_unlock_irq(&xas);
1938 xa_unlocked:
1939
1940         if (result == SCAN_SUCCEED) {
1941                 struct page *page, *tmp;
1942
1943                 /*
1944                  * Replacing old pages with new one has succeeded, now we
1945                  * need to copy the content and free the old pages.
1946                  */
1947                 index = start;
1948                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1949                         while (index < page->index) {
1950                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1951                                 index++;
1952                         }
1953                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1954                                         page);
1955                         list_del(&page->lru);
1956                         page->mapping = NULL;
1957                         page_ref_unfreeze(page, 1);
1958                         ClearPageActive(page);
1959                         ClearPageUnevictable(page);
1960                         unlock_page(page);
1961                         put_page(page);
1962                         index++;
1963                 }
1964                 while (index < end) {
1965                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1966                         index++;
1967                 }
1968
1969                 SetPageUptodate(new_page);
1970                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1971                 if (is_shmem)
1972                         set_page_dirty(new_page);
1973                 lru_cache_add(new_page);
1974
1975                 /*
1976                  * Remove pte page tables, so we can re-fault the page as huge.
1977                  */
1978                 retract_page_tables(mapping, start);
1979                 *hpage = NULL;
1980
1981                 khugepaged_pages_collapsed++;
1982         } else {
1983                 struct page *page;
1984
1985                 /* Something went wrong: roll back page cache changes */
1986                 xas_lock_irq(&xas);
1987                 mapping->nrpages -= nr_none;
1988
1989                 if (is_shmem)
1990                         shmem_uncharge(mapping->host, nr_none);
1991
1992                 xas_set(&xas, start);
1993                 xas_for_each(&xas, page, end - 1) {
1994                         page = list_first_entry_or_null(&pagelist,
1995                                         struct page, lru);
1996                         if (!page || xas.xa_index < page->index) {
1997                                 if (!nr_none)
1998                                         break;
1999                                 nr_none--;
2000                                 /* Put holes back where they were */
2001                                 xas_store(&xas, NULL);
2002                                 continue;
2003                         }
2004
2005                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2006
2007                         /* Unfreeze the page. */
2008                         list_del(&page->lru);
2009                         page_ref_unfreeze(page, 2);
2010                         xas_store(&xas, page);
2011                         xas_pause(&xas);
2012                         xas_unlock_irq(&xas);
2013                         unlock_page(page);
2014                         putback_lru_page(page);
2015                         xas_lock_irq(&xas);
2016                 }
2017                 VM_BUG_ON(nr_none);
2018                 xas_unlock_irq(&xas);
2019
2020                 new_page->mapping = NULL;
2021         }
2022
2023         unlock_page(new_page);
2024 out:
2025         VM_BUG_ON(!list_empty(&pagelist));
2026         if (!IS_ERR_OR_NULL(*hpage))
2027                 mem_cgroup_uncharge(*hpage);
2028         /* TODO: tracepoints */
2029 }
2030
2031 static void khugepaged_scan_file(struct mm_struct *mm,
2032                 struct file *file, pgoff_t start, struct page **hpage)
2033 {
2034         struct page *page = NULL;
2035         struct address_space *mapping = file->f_mapping;
2036         XA_STATE(xas, &mapping->i_pages, start);
2037         int present, swap;
2038         int node = NUMA_NO_NODE;
2039         int result = SCAN_SUCCEED;
2040
2041         present = 0;
2042         swap = 0;
2043         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2044         rcu_read_lock();
2045         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2046                 if (xas_retry(&xas, page))
2047                         continue;
2048
2049                 if (xa_is_value(page)) {
2050                         if (++swap > khugepaged_max_ptes_swap) {
2051                                 result = SCAN_EXCEED_SWAP_PTE;
2052                                 break;
2053                         }
2054                         continue;
2055                 }
2056
2057                 if (PageTransCompound(page)) {
2058                         result = SCAN_PAGE_COMPOUND;
2059                         break;
2060                 }
2061
2062                 node = page_to_nid(page);
2063                 if (khugepaged_scan_abort(node)) {
2064                         result = SCAN_SCAN_ABORT;
2065                         break;
2066                 }
2067                 khugepaged_node_load[node]++;
2068
2069                 if (!PageLRU(page)) {
2070                         result = SCAN_PAGE_LRU;
2071                         break;
2072                 }
2073
2074                 if (page_count(page) !=
2075                     1 + page_mapcount(page) + page_has_private(page)) {
2076                         result = SCAN_PAGE_COUNT;
2077                         break;
2078                 }
2079
2080                 /*
2081                  * We probably should check if the page is referenced here, but
2082                  * nobody would transfer pte_young() to PageReferenced() for us.
2083                  * And rmap walk here is just too costly...
2084                  */
2085
2086                 present++;
2087
2088                 if (need_resched()) {
2089                         xas_pause(&xas);
2090                         cond_resched_rcu();
2091                 }
2092         }
2093         rcu_read_unlock();
2094
2095         if (result == SCAN_SUCCEED) {
2096                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2097                         result = SCAN_EXCEED_NONE_PTE;
2098                 } else {
2099                         node = khugepaged_find_target_node();
2100                         collapse_file(mm, file, start, hpage, node);
2101                 }
2102         }
2103
2104         /* TODO: tracepoints */
2105 }
2106 #else
2107 static void khugepaged_scan_file(struct mm_struct *mm,
2108                 struct file *file, pgoff_t start, struct page **hpage)
2109 {
2110         BUILD_BUG();
2111 }
2112
2113 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2114 {
2115 }
2116 #endif
2117
2118 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2119                                             struct page **hpage)
2120         __releases(&khugepaged_mm_lock)
2121         __acquires(&khugepaged_mm_lock)
2122 {
2123         struct mm_slot *mm_slot;
2124         struct mm_struct *mm;
2125         struct vm_area_struct *vma;
2126         int progress = 0;
2127
2128         VM_BUG_ON(!pages);
2129         lockdep_assert_held(&khugepaged_mm_lock);
2130
2131         if (khugepaged_scan.mm_slot)
2132                 mm_slot = khugepaged_scan.mm_slot;
2133         else {
2134                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2135                                      struct mm_slot, mm_node);
2136                 khugepaged_scan.address = 0;
2137                 khugepaged_scan.mm_slot = mm_slot;
2138         }
2139         spin_unlock(&khugepaged_mm_lock);
2140         khugepaged_collapse_pte_mapped_thps(mm_slot);
2141
2142         mm = mm_slot->mm;
2143         /*
2144          * Don't wait for semaphore (to avoid long wait times).  Just move to
2145          * the next mm on the list.
2146          */
2147         vma = NULL;
2148         if (unlikely(!mmap_read_trylock(mm)))
2149                 goto breakouterloop_mmap_lock;
2150         if (likely(!khugepaged_test_exit(mm)))
2151                 vma = find_vma(mm, khugepaged_scan.address);
2152
2153         progress++;
2154         for (; vma; vma = vma->vm_next) {
2155                 unsigned long hstart, hend;
2156
2157                 cond_resched();
2158                 if (unlikely(khugepaged_test_exit(mm))) {
2159                         progress++;
2160                         break;
2161                 }
2162                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2163 skip:
2164                         progress++;
2165                         continue;
2166                 }
2167                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2168                 hend = vma->vm_end & HPAGE_PMD_MASK;
2169                 if (hstart >= hend)
2170                         goto skip;
2171                 if (khugepaged_scan.address > hend)
2172                         goto skip;
2173                 if (khugepaged_scan.address < hstart)
2174                         khugepaged_scan.address = hstart;
2175                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2176                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2177                         goto skip;
2178
2179                 while (khugepaged_scan.address < hend) {
2180                         int ret;
2181                         cond_resched();
2182                         if (unlikely(khugepaged_test_exit(mm)))
2183                                 goto breakouterloop;
2184
2185                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2186                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2187                                   hend);
2188                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2189                                 struct file *file = get_file(vma->vm_file);
2190                                 pgoff_t pgoff = linear_page_index(vma,
2191                                                 khugepaged_scan.address);
2192
2193                                 mmap_read_unlock(mm);
2194                                 ret = 1;
2195                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2196                                 fput(file);
2197                         } else {
2198                                 ret = khugepaged_scan_pmd(mm, vma,
2199                                                 khugepaged_scan.address,
2200                                                 hpage);
2201                         }
2202                         /* move to next address */
2203                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2204                         progress += HPAGE_PMD_NR;
2205                         if (ret)
2206                                 /* we released mmap_lock so break loop */
2207                                 goto breakouterloop_mmap_lock;
2208                         if (progress >= pages)
2209                                 goto breakouterloop;
2210                 }
2211         }
2212 breakouterloop:
2213         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2214 breakouterloop_mmap_lock:
2215
2216         spin_lock(&khugepaged_mm_lock);
2217         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2218         /*
2219          * Release the current mm_slot if this mm is about to die, or
2220          * if we scanned all vmas of this mm.
2221          */
2222         if (khugepaged_test_exit(mm) || !vma) {
2223                 /*
2224                  * Make sure that if mm_users is reaching zero while
2225                  * khugepaged runs here, khugepaged_exit will find
2226                  * mm_slot not pointing to the exiting mm.
2227                  */
2228                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2229                         khugepaged_scan.mm_slot = list_entry(
2230                                 mm_slot->mm_node.next,
2231                                 struct mm_slot, mm_node);
2232                         khugepaged_scan.address = 0;
2233                 } else {
2234                         khugepaged_scan.mm_slot = NULL;
2235                         khugepaged_full_scans++;
2236                 }
2237
2238                 collect_mm_slot(mm_slot);
2239         }
2240
2241         return progress;
2242 }
2243
2244 static int khugepaged_has_work(void)
2245 {
2246         return !list_empty(&khugepaged_scan.mm_head) &&
2247                 khugepaged_enabled();
2248 }
2249
2250 static int khugepaged_wait_event(void)
2251 {
2252         return !list_empty(&khugepaged_scan.mm_head) ||
2253                 kthread_should_stop();
2254 }
2255
2256 static void khugepaged_do_scan(void)
2257 {
2258         struct page *hpage = NULL;
2259         unsigned int progress = 0, pass_through_head = 0;
2260         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2261         bool wait = true;
2262
2263         lru_add_drain_all();
2264
2265         while (progress < pages) {
2266                 if (!khugepaged_prealloc_page(&hpage, &wait))
2267                         break;
2268
2269                 cond_resched();
2270
2271                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2272                         break;
2273
2274                 spin_lock(&khugepaged_mm_lock);
2275                 if (!khugepaged_scan.mm_slot)
2276                         pass_through_head++;
2277                 if (khugepaged_has_work() &&
2278                     pass_through_head < 2)
2279                         progress += khugepaged_scan_mm_slot(pages - progress,
2280                                                             &hpage);
2281                 else
2282                         progress = pages;
2283                 spin_unlock(&khugepaged_mm_lock);
2284         }
2285
2286         if (!IS_ERR_OR_NULL(hpage))
2287                 put_page(hpage);
2288 }
2289
2290 static bool khugepaged_should_wakeup(void)
2291 {
2292         return kthread_should_stop() ||
2293                time_after_eq(jiffies, khugepaged_sleep_expire);
2294 }
2295
2296 static void khugepaged_wait_work(void)
2297 {
2298         if (khugepaged_has_work()) {
2299                 const unsigned long scan_sleep_jiffies =
2300                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2301
2302                 if (!scan_sleep_jiffies)
2303                         return;
2304
2305                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2306                 wait_event_freezable_timeout(khugepaged_wait,
2307                                              khugepaged_should_wakeup(),
2308                                              scan_sleep_jiffies);
2309                 return;
2310         }
2311
2312         if (khugepaged_enabled())
2313                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2314 }
2315
2316 static int khugepaged(void *none)
2317 {
2318         struct mm_slot *mm_slot;
2319
2320         set_freezable();
2321         set_user_nice(current, MAX_NICE);
2322
2323         while (!kthread_should_stop()) {
2324                 khugepaged_do_scan();
2325                 khugepaged_wait_work();
2326         }
2327
2328         spin_lock(&khugepaged_mm_lock);
2329         mm_slot = khugepaged_scan.mm_slot;
2330         khugepaged_scan.mm_slot = NULL;
2331         if (mm_slot)
2332                 collect_mm_slot(mm_slot);
2333         spin_unlock(&khugepaged_mm_lock);
2334         return 0;
2335 }
2336
2337 static void set_recommended_min_free_kbytes(void)
2338 {
2339         struct zone *zone;
2340         int nr_zones = 0;
2341         unsigned long recommended_min;
2342
2343         for_each_populated_zone(zone) {
2344                 /*
2345                  * We don't need to worry about fragmentation of
2346                  * ZONE_MOVABLE since it only has movable pages.
2347                  */
2348                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2349                         continue;
2350
2351                 nr_zones++;
2352         }
2353
2354         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2355         recommended_min = pageblock_nr_pages * nr_zones * 2;
2356
2357         /*
2358          * Make sure that on average at least two pageblocks are almost free
2359          * of another type, one for a migratetype to fall back to and a
2360          * second to avoid subsequent fallbacks of other types There are 3
2361          * MIGRATE_TYPES we care about.
2362          */
2363         recommended_min += pageblock_nr_pages * nr_zones *
2364                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2365
2366         /* don't ever allow to reserve more than 5% of the lowmem */
2367         recommended_min = min(recommended_min,
2368                               (unsigned long) nr_free_buffer_pages() / 20);
2369         recommended_min <<= (PAGE_SHIFT-10);
2370
2371         if (recommended_min > min_free_kbytes) {
2372                 if (user_min_free_kbytes >= 0)
2373                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2374                                 min_free_kbytes, recommended_min);
2375
2376                 min_free_kbytes = recommended_min;
2377         }
2378         setup_per_zone_wmarks();
2379 }
2380
2381 int start_stop_khugepaged(void)
2382 {
2383         int err = 0;
2384
2385         mutex_lock(&khugepaged_mutex);
2386         if (khugepaged_enabled()) {
2387                 if (!khugepaged_thread)
2388                         khugepaged_thread = kthread_run(khugepaged, NULL,
2389                                                         "khugepaged");
2390                 if (IS_ERR(khugepaged_thread)) {
2391                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2392                         err = PTR_ERR(khugepaged_thread);
2393                         khugepaged_thread = NULL;
2394                         goto fail;
2395                 }
2396
2397                 if (!list_empty(&khugepaged_scan.mm_head))
2398                         wake_up_interruptible(&khugepaged_wait);
2399
2400                 set_recommended_min_free_kbytes();
2401         } else if (khugepaged_thread) {
2402                 kthread_stop(khugepaged_thread);
2403                 khugepaged_thread = NULL;
2404         }
2405 fail:
2406         mutex_unlock(&khugepaged_mutex);
2407         return err;
2408 }
2409
2410 void khugepaged_min_free_kbytes_update(void)
2411 {
2412         mutex_lock(&khugepaged_mutex);
2413         if (khugepaged_enabled() && khugepaged_thread)
2414                 set_recommended_min_free_kbytes();
2415         mutex_unlock(&khugepaged_mutex);
2416 }