mm/khugepaged.c

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