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/page_table_check.h>
  20 #include <linux/swapops.h>
  21 #include <linux/shmem_fs.h>
  22 #include <linux/ksm.h>
  23
  24 #include <asm/tlb.h>
  25 #include <asm/pgalloc.h>
  26 #include "internal.h"
  27 #include "mm_slot.h"
  28
  29 enum scan_result {
  30         SCAN_FAIL,
  31         SCAN_SUCCEED,
  32         SCAN_PMD_NULL,
  33         SCAN_PMD_NONE,
  34         SCAN_PMD_MAPPED,
  35         SCAN_EXCEED_NONE_PTE,
  36         SCAN_EXCEED_SWAP_PTE,
  37         SCAN_EXCEED_SHARED_PTE,
  38         SCAN_PTE_NON_PRESENT,
  39         SCAN_PTE_UFFD_WP,
  40         SCAN_PTE_MAPPED_HUGEPAGE,
  41         SCAN_PAGE_RO,
  42         SCAN_LACK_REFERENCED_PAGE,
  43         SCAN_PAGE_NULL,
  44         SCAN_SCAN_ABORT,
  45         SCAN_PAGE_COUNT,
  46         SCAN_PAGE_LRU,
  47         SCAN_PAGE_LOCK,
  48         SCAN_PAGE_ANON,
  49         SCAN_PAGE_COMPOUND,
  50         SCAN_ANY_PROCESS,
  51         SCAN_VMA_NULL,
  52         SCAN_VMA_CHECK,
  53         SCAN_ADDRESS_RANGE,
  54         SCAN_DEL_PAGE_LRU,
  55         SCAN_ALLOC_HUGE_PAGE_FAIL,
  56         SCAN_CGROUP_CHARGE_FAIL,
  57         SCAN_TRUNCATED,
  58         SCAN_PAGE_HAS_PRIVATE,
  59         SCAN_STORE_FAILED,
  60         SCAN_COPY_MC,
  61         SCAN_PAGE_FILLED,
  62 };
  63
  64 #define CREATE_TRACE_POINTS
  65 #include <trace/events/huge_memory.h>
  66
  67 static struct task_struct *khugepaged_thread __read_mostly;
  68 static DEFINE_MUTEX(khugepaged_mutex);
  69
  70 /* default scan 8*512 pte (or vmas) every 30 second */
  71 static unsigned int khugepaged_pages_to_scan __read_mostly;
  72 static unsigned int khugepaged_pages_collapsed;
  73 static unsigned int khugepaged_full_scans;
  74 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  75 /* during fragmentation poll the hugepage allocator once every minute */
  76 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  77 static unsigned long khugepaged_sleep_expire;
  78 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  79 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  80 /*
  81  * default collapse hugepages if there is at least one pte mapped like
  82  * it would have happened if the vma was large enough during page
  83  * fault.
  84  *
  85  * Note that these are only respected if collapse was initiated by khugepaged.
  86  */
  87 static unsigned int khugepaged_max_ptes_none __read_mostly;
  88 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  89 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  90
  91 #define MM_SLOTS_HASH_BITS 10
  92 static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  93
  94 static struct kmem_cache *mm_slot_cache __read_mostly;
  95
  96 struct collapse_control {
  97         bool is_khugepaged;
  98
  99         /* Num pages scanned per node */
 100         u32 node_load[MAX_NUMNODES];
 101
 102         /* nodemask for allocation fallback */
 103         nodemask_t alloc_nmask;
 104 };
 105
 106 /**
 107  * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned
 108  * @slot: hash lookup from mm to mm_slot
 109  */
 110 struct khugepaged_mm_slot {
 111         struct mm_slot slot;
 112 };
 113
 114 /**
 115  * struct khugepaged_scan - cursor for scanning
 116  * @mm_head: the head of the mm list to scan
 117  * @mm_slot: the current mm_slot we are scanning
 118  * @address: the next address inside that to be scanned
 119  *
 120  * There is only the one khugepaged_scan instance of this cursor structure.
 121  */
 122 struct khugepaged_scan {
 123         struct list_head mm_head;
 124         struct khugepaged_mm_slot *mm_slot;
 125         unsigned long address;
 126 };
 127
 128 static struct khugepaged_scan khugepaged_scan = {
 129         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 130 };
 131
 132 #ifdef CONFIG_SYSFS
 133 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 134                                          struct kobj_attribute *attr,
 135                                          char *buf)
 136 {
 137         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 138 }
 139
 140 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 141                                           struct kobj_attribute *attr,
 142                                           const char *buf, size_t count)
 143 {
 144         unsigned int msecs;
 145         int err;
 146
 147         err = kstrtouint(buf, 10, &msecs);
 148         if (err)
 149                 return -EINVAL;
 150
 151         khugepaged_scan_sleep_millisecs = msecs;
 152         khugepaged_sleep_expire = 0;
 153         wake_up_interruptible(&khugepaged_wait);
 154
 155         return count;
 156 }
 157 static struct kobj_attribute scan_sleep_millisecs_attr =
 158         __ATTR_RW(scan_sleep_millisecs);
 159
 160 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 161                                           struct kobj_attribute *attr,
 162                                           char *buf)
 163 {
 164         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 165 }
 166
 167 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 168                                            struct kobj_attribute *attr,
 169                                            const char *buf, size_t count)
 170 {
 171         unsigned int msecs;
 172         int err;
 173
 174         err = kstrtouint(buf, 10, &msecs);
 175         if (err)
 176                 return -EINVAL;
 177
 178         khugepaged_alloc_sleep_millisecs = msecs;
 179         khugepaged_sleep_expire = 0;
 180         wake_up_interruptible(&khugepaged_wait);
 181
 182         return count;
 183 }
 184 static struct kobj_attribute alloc_sleep_millisecs_attr =
 185         __ATTR_RW(alloc_sleep_millisecs);
 186
 187 static ssize_t pages_to_scan_show(struct kobject *kobj,
 188                                   struct kobj_attribute *attr,
 189                                   char *buf)
 190 {
 191         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 192 }
 193 static ssize_t pages_to_scan_store(struct kobject *kobj,
 194                                    struct kobj_attribute *attr,
 195                                    const char *buf, size_t count)
 196 {
 197         unsigned int pages;
 198         int err;
 199
 200         err = kstrtouint(buf, 10, &pages);
 201         if (err || !pages)
 202                 return -EINVAL;
 203
 204         khugepaged_pages_to_scan = pages;
 205
 206         return count;
 207 }
 208 static struct kobj_attribute pages_to_scan_attr =
 209         __ATTR_RW(pages_to_scan);
 210
 211 static ssize_t pages_collapsed_show(struct kobject *kobj,
 212                                     struct kobj_attribute *attr,
 213                                     char *buf)
 214 {
 215         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 216 }
 217 static struct kobj_attribute pages_collapsed_attr =
 218         __ATTR_RO(pages_collapsed);
 219
 220 static ssize_t full_scans_show(struct kobject *kobj,
 221                                struct kobj_attribute *attr,
 222                                char *buf)
 223 {
 224         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 225 }
 226 static struct kobj_attribute full_scans_attr =
 227         __ATTR_RO(full_scans);
 228
 229 static ssize_t defrag_show(struct kobject *kobj,
 230                            struct kobj_attribute *attr, char *buf)
 231 {
 232         return single_hugepage_flag_show(kobj, attr, buf,
 233                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 234 }
 235 static ssize_t defrag_store(struct kobject *kobj,
 236                             struct kobj_attribute *attr,
 237                             const char *buf, size_t count)
 238 {
 239         return single_hugepage_flag_store(kobj, attr, buf, count,
 240                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 241 }
 242 static struct kobj_attribute khugepaged_defrag_attr =
 243         __ATTR_RW(defrag);
 244
 245 /*
 246  * max_ptes_none controls if khugepaged should collapse hugepages over
 247  * any unmapped ptes in turn potentially increasing the memory
 248  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 249  * reduce the available free memory in the system as it
 250  * runs. Increasing max_ptes_none will instead potentially reduce the
 251  * free memory in the system during the khugepaged scan.
 252  */
 253 static ssize_t max_ptes_none_show(struct kobject *kobj,
 254                                   struct kobj_attribute *attr,
 255                                   char *buf)
 256 {
 257         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 258 }
 259 static ssize_t max_ptes_none_store(struct kobject *kobj,
 260                                    struct kobj_attribute *attr,
 261                                    const char *buf, size_t count)
 262 {
 263         int err;
 264         unsigned long max_ptes_none;
 265
 266         err = kstrtoul(buf, 10, &max_ptes_none);
 267         if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 268                 return -EINVAL;
 269
 270         khugepaged_max_ptes_none = max_ptes_none;
 271
 272         return count;
 273 }
 274 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 275         __ATTR_RW(max_ptes_none);
 276
 277 static ssize_t max_ptes_swap_show(struct kobject *kobj,
 278                                   struct kobj_attribute *attr,
 279                                   char *buf)
 280 {
 281         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 282 }
 283
 284 static ssize_t max_ptes_swap_store(struct kobject *kobj,
 285                                    struct kobj_attribute *attr,
 286                                    const char *buf, size_t count)
 287 {
 288         int err;
 289         unsigned long max_ptes_swap;
 290
 291         err  = kstrtoul(buf, 10, &max_ptes_swap);
 292         if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 293                 return -EINVAL;
 294
 295         khugepaged_max_ptes_swap = max_ptes_swap;
 296
 297         return count;
 298 }
 299
 300 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 301         __ATTR_RW(max_ptes_swap);
 302
 303 static ssize_t max_ptes_shared_show(struct kobject *kobj,
 304                                     struct kobj_attribute *attr,
 305                                     char *buf)
 306 {
 307         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 308 }
 309
 310 static ssize_t max_ptes_shared_store(struct kobject *kobj,
 311                                      struct kobj_attribute *attr,
 312                                      const char *buf, size_t count)
 313 {
 314         int err;
 315         unsigned long max_ptes_shared;
 316
 317         err  = kstrtoul(buf, 10, &max_ptes_shared);
 318         if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 319                 return -EINVAL;
 320
 321         khugepaged_max_ptes_shared = max_ptes_shared;
 322
 323         return count;
 324 }
 325
 326 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 327         __ATTR_RW(max_ptes_shared);
 328
 329 static struct attribute *khugepaged_attr[] = {
 330         &khugepaged_defrag_attr.attr,
 331         &khugepaged_max_ptes_none_attr.attr,
 332         &khugepaged_max_ptes_swap_attr.attr,
 333         &khugepaged_max_ptes_shared_attr.attr,
 334         &pages_to_scan_attr.attr,
 335         &pages_collapsed_attr.attr,
 336         &full_scans_attr.attr,
 337         &scan_sleep_millisecs_attr.attr,
 338         &alloc_sleep_millisecs_attr.attr,
 339         NULL,
 340 };
 341
 342 struct attribute_group khugepaged_attr_group = {
 343         .attrs = khugepaged_attr,
 344         .name = "khugepaged",
 345 };
 346 #endif /* CONFIG_SYSFS */
 347
 348 int hugepage_madvise(struct vm_area_struct *vma,
 349                      unsigned long *vm_flags, int advice)
 350 {
 351         switch (advice) {
 352         case MADV_HUGEPAGE:
 353 #ifdef CONFIG_S390
 354                 /*
 355                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 356                  * can't handle this properly after s390_enable_sie, so we simply
 357                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 358                  */
 359                 if (mm_has_pgste(vma->vm_mm))
 360                         return 0;
 361 #endif
 362                 *vm_flags &= ~VM_NOHUGEPAGE;
 363                 *vm_flags |= VM_HUGEPAGE;
 364                 /*
 365                  * If the vma become good for khugepaged to scan,
 366                  * register it here without waiting a page fault that
 367                  * may not happen any time soon.
 368                  */
 369                 khugepaged_enter_vma(vma, *vm_flags);
 370                 break;
 371         case MADV_NOHUGEPAGE:
 372                 *vm_flags &= ~VM_HUGEPAGE;
 373                 *vm_flags |= VM_NOHUGEPAGE;
 374                 /*
 375                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 376                  * this vma even if we leave the mm registered in khugepaged if
 377                  * it got registered before VM_NOHUGEPAGE was set.
 378                  */
 379                 break;
 380         }
 381
 382         return 0;
 383 }
 384
 385 int __init khugepaged_init(void)
 386 {
 387         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 388                                           sizeof(struct khugepaged_mm_slot),
 389                                           __alignof__(struct khugepaged_mm_slot),
 390                                           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 int hpage_collapse_test_exit(struct mm_struct *mm)
 408 {
 409         return atomic_read(&mm->mm_users) == 0;
 410 }
 411
 412 void __khugepaged_enter(struct mm_struct *mm)
 413 {
 414         struct khugepaged_mm_slot *mm_slot;
 415         struct mm_slot *slot;
 416         int wakeup;
 417
 418         /* __khugepaged_exit() must not run from under us */
 419         VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 420         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags)))
 421                 return;
 422
 423         mm_slot = mm_slot_alloc(mm_slot_cache);
 424         if (!mm_slot)
 425                 return;
 426
 427         slot = &mm_slot->slot;
 428
 429         spin_lock(&khugepaged_mm_lock);
 430         mm_slot_insert(mm_slots_hash, mm, slot);
 431         /*
 432          * Insert just behind the scanning cursor, to let the area settle
 433          * down a little.
 434          */
 435         wakeup = list_empty(&khugepaged_scan.mm_head);
 436         list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 437         spin_unlock(&khugepaged_mm_lock);
 438
 439         mmgrab(mm);
 440         if (wakeup)
 441                 wake_up_interruptible(&khugepaged_wait);
 442 }
 443
 444 void khugepaged_enter_vma(struct vm_area_struct *vma,
 445                           unsigned long vm_flags)
 446 {
 447         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 448             hugepage_flags_enabled()) {
 449                 if (hugepage_vma_check(vma, vm_flags, false, false, true))
 450                         __khugepaged_enter(vma->vm_mm);
 451         }
 452 }
 453
 454 void __khugepaged_exit(struct mm_struct *mm)
 455 {
 456         struct khugepaged_mm_slot *mm_slot;
 457         struct mm_slot *slot;
 458         int free = 0;
 459
 460         spin_lock(&khugepaged_mm_lock);
 461         slot = mm_slot_lookup(mm_slots_hash, mm);
 462         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
 463         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 464                 hash_del(&slot->hash);
 465                 list_del(&slot->mm_node);
 466                 free = 1;
 467         }
 468         spin_unlock(&khugepaged_mm_lock);
 469
 470         if (free) {
 471                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 472                 mm_slot_free(mm_slot_cache, mm_slot);
 473                 mmdrop(mm);
 474         } else if (mm_slot) {
 475                 /*
 476                  * This is required to serialize against
 477                  * hpage_collapse_test_exit() (which is guaranteed to run
 478                  * under mmap sem read mode). Stop here (after we return all
 479                  * pagetables will be destroyed) until khugepaged has finished
 480                  * working on the pagetables under the mmap_lock.
 481                  */
 482                 mmap_write_lock(mm);
 483                 mmap_write_unlock(mm);
 484         }
 485 }
 486
 487 static void release_pte_folio(struct folio *folio)
 488 {
 489         node_stat_mod_folio(folio,
 490                         NR_ISOLATED_ANON + folio_is_file_lru(folio),
 491                         -folio_nr_pages(folio));
 492         folio_unlock(folio);
 493         folio_putback_lru(folio);
 494 }
 495
 496 static void release_pte_page(struct page *page)
 497 {
 498         release_pte_folio(page_folio(page));
 499 }
 500
 501 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 502                 struct list_head *compound_pagelist)
 503 {
 504         struct folio *folio, *tmp;
 505
 506         while (--_pte >= pte) {
 507                 pte_t pteval = ptep_get(_pte);
 508                 unsigned long pfn;
 509
 510                 if (pte_none(pteval))
 511                         continue;
 512                 pfn = pte_pfn(pteval);
 513                 if (is_zero_pfn(pfn))
 514                         continue;
 515                 folio = pfn_folio(pfn);
 516                 if (folio_test_large(folio))
 517                         continue;
 518                 release_pte_folio(folio);
 519         }
 520
 521         list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 522                 list_del(&folio->lru);
 523                 release_pte_folio(folio);
 524         }
 525 }
 526
 527 static bool is_refcount_suitable(struct page *page)
 528 {
 529         int expected_refcount;
 530
 531         expected_refcount = total_mapcount(page);
 532         if (PageSwapCache(page))
 533                 expected_refcount += compound_nr(page);
 534
 535         return page_count(page) == expected_refcount;
 536 }
 537
 538 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 539                                         unsigned long address,
 540                                         pte_t *pte,
 541                                         struct collapse_control *cc,
 542                                         struct list_head *compound_pagelist)
 543 {
 544         struct page *page = NULL;
 545         pte_t *_pte;
 546         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 547         bool writable = false;
 548
 549         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 550              _pte++, address += PAGE_SIZE) {
 551                 pte_t pteval = ptep_get(_pte);
 552                 if (pte_none(pteval) || (pte_present(pteval) &&
 553                                 is_zero_pfn(pte_pfn(pteval)))) {
 554                         ++none_or_zero;
 555                         if (!userfaultfd_armed(vma) &&
 556                             (!cc->is_khugepaged ||
 557                              none_or_zero <= khugepaged_max_ptes_none)) {
 558                                 continue;
 559                         } else {
 560                                 result = SCAN_EXCEED_NONE_PTE;
 561                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 562                                 goto out;
 563                         }
 564                 }
 565                 if (!pte_present(pteval)) {
 566                         result = SCAN_PTE_NON_PRESENT;
 567                         goto out;
 568                 }
 569                 if (pte_uffd_wp(pteval)) {
 570                         result = SCAN_PTE_UFFD_WP;
 571                         goto out;
 572                 }
 573                 page = vm_normal_page(vma, address, pteval);
 574                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 575                         result = SCAN_PAGE_NULL;
 576                         goto out;
 577                 }
 578
 579                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 580
 581                 if (page_mapcount(page) > 1) {
 582                         ++shared;
 583                         if (cc->is_khugepaged &&
 584                             shared > khugepaged_max_ptes_shared) {
 585                                 result = SCAN_EXCEED_SHARED_PTE;
 586                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 587                                 goto out;
 588                         }
 589                 }
 590
 591                 if (PageCompound(page)) {
 592                         struct page *p;
 593                         page = compound_head(page);
 594
 595                         /*
 596                          * Check if we have dealt with the compound page
 597                          * already
 598                          */
 599                         list_for_each_entry(p, compound_pagelist, lru) {
 600                                 if (page == p)
 601                                         goto next;
 602                         }
 603                 }
 604
 605                 /*
 606                  * We can do it before isolate_lru_page because the
 607                  * page can't be freed from under us. NOTE: PG_lock
 608                  * is needed to serialize against split_huge_page
 609                  * when invoked from the VM.
 610                  */
 611                 if (!trylock_page(page)) {
 612                         result = SCAN_PAGE_LOCK;
 613                         goto out;
 614                 }
 615
 616                 /*
 617                  * Check if the page has any GUP (or other external) pins.
 618                  *
 619                  * The page table that maps the page has been already unlinked
 620                  * from the page table tree and this process cannot get
 621                  * an additional pin on the page.
 622                  *
 623                  * New pins can come later if the page is shared across fork,
 624                  * but not from this process. The other process cannot write to
 625                  * the page, only trigger CoW.
 626                  */
 627                 if (!is_refcount_suitable(page)) {
 628                         unlock_page(page);
 629                         result = SCAN_PAGE_COUNT;
 630                         goto out;
 631                 }
 632
 633                 /*
 634                  * Isolate the page to avoid collapsing an hugepage
 635                  * currently in use by the VM.
 636                  */
 637                 if (!isolate_lru_page(page)) {
 638                         unlock_page(page);
 639                         result = SCAN_DEL_PAGE_LRU;
 640                         goto out;
 641                 }
 642                 mod_node_page_state(page_pgdat(page),
 643                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 644                                 compound_nr(page));
 645                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 646                 VM_BUG_ON_PAGE(PageLRU(page), page);
 647
 648                 if (PageCompound(page))
 649                         list_add_tail(&page->lru, compound_pagelist);
 650 next:
 651                 /*
 652                  * If collapse was initiated by khugepaged, check that there is
 653                  * enough young pte to justify collapsing the page
 654                  */
 655                 if (cc->is_khugepaged &&
 656                     (pte_young(pteval) || page_is_young(page) ||
 657                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
 658                                                                      address)))
 659                         referenced++;
 660
 661                 if (pte_write(pteval))
 662                         writable = true;
 663         }
 664
 665         if (unlikely(!writable)) {
 666                 result = SCAN_PAGE_RO;
 667         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 668                 result = SCAN_LACK_REFERENCED_PAGE;
 669         } else {
 670                 result = SCAN_SUCCEED;
 671                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 672                                                     referenced, writable, result);
 673                 return result;
 674         }
 675 out:
 676         release_pte_pages(pte, _pte, compound_pagelist);
 677         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 678                                             referenced, writable, result);
 679         return result;
 680 }
 681
 682 static void __collapse_huge_page_copy_succeeded(pte_t *pte,
 683                                                 struct vm_area_struct *vma,
 684                                                 unsigned long address,
 685                                                 spinlock_t *ptl,
 686                                                 struct list_head *compound_pagelist)
 687 {
 688         struct page *src_page;
 689         struct page *tmp;
 690         pte_t *_pte;
 691         pte_t pteval;
 692
 693         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 694              _pte++, address += PAGE_SIZE) {
 695                 pteval = ptep_get(_pte);
 696                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 697                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 698                         if (is_zero_pfn(pte_pfn(pteval))) {
 699                                 /*
 700                                  * ptl mostly unnecessary.
 701                                  */
 702                                 spin_lock(ptl);
 703                                 ptep_clear(vma->vm_mm, address, _pte);
 704                                 spin_unlock(ptl);
 705                                 ksm_might_unmap_zero_page(vma->vm_mm, pteval);
 706                         }
 707                 } else {
 708                         src_page = pte_page(pteval);
 709                         if (!PageCompound(src_page))
 710                                 release_pte_page(src_page);
 711                         /*
 712                          * ptl mostly unnecessary, but preempt has to
 713                          * be disabled to update the per-cpu stats
 714                          * inside page_remove_rmap().
 715                          */
 716                         spin_lock(ptl);
 717                         ptep_clear(vma->vm_mm, address, _pte);
 718                         page_remove_rmap(src_page, vma, false);
 719                         spin_unlock(ptl);
 720                         free_page_and_swap_cache(src_page);
 721                 }
 722         }
 723
 724         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 725                 list_del(&src_page->lru);
 726                 mod_node_page_state(page_pgdat(src_page),
 727                                     NR_ISOLATED_ANON + page_is_file_lru(src_page),
 728                                     -compound_nr(src_page));
 729                 unlock_page(src_page);
 730                 free_swap_cache(src_page);
 731                 putback_lru_page(src_page);
 732         }
 733 }
 734
 735 static void __collapse_huge_page_copy_failed(pte_t *pte,
 736                                              pmd_t *pmd,
 737                                              pmd_t orig_pmd,
 738                                              struct vm_area_struct *vma,
 739                                              struct list_head *compound_pagelist)
 740 {
 741         spinlock_t *pmd_ptl;
 742
 743         /*
 744          * Re-establish the PMD to point to the original page table
 745          * entry. Restoring PMD needs to be done prior to releasing
 746          * pages. Since pages are still isolated and locked here,
 747          * acquiring anon_vma_lock_write is unnecessary.
 748          */
 749         pmd_ptl = pmd_lock(vma->vm_mm, pmd);
 750         pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));
 751         spin_unlock(pmd_ptl);
 752         /*
 753          * Release both raw and compound pages isolated
 754          * in __collapse_huge_page_isolate.
 755          */
 756         release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);
 757 }
 758
 759 /*
 760  * __collapse_huge_page_copy - attempts to copy memory contents from raw
 761  * pages to a hugepage. Cleans up the raw pages if copying succeeds;
 762  * otherwise restores the original page table and releases isolated raw pages.
 763  * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.
 764  *
 765  * @pte: starting of the PTEs to copy from
 766  * @page: the new hugepage to copy contents to
 767  * @pmd: pointer to the new hugepage's PMD
 768  * @orig_pmd: the original raw pages' PMD
 769  * @vma: the original raw pages' virtual memory area
 770  * @address: starting address to copy
 771  * @ptl: lock on raw pages' PTEs
 772  * @compound_pagelist: list that stores compound pages
 773  */
 774 static int __collapse_huge_page_copy(pte_t *pte,
 775                                      struct page *page,
 776                                      pmd_t *pmd,
 777                                      pmd_t orig_pmd,
 778                                      struct vm_area_struct *vma,
 779                                      unsigned long address,
 780                                      spinlock_t *ptl,
 781                                      struct list_head *compound_pagelist)
 782 {
 783         struct page *src_page;
 784         pte_t *_pte;
 785         pte_t pteval;
 786         unsigned long _address;
 787         int result = SCAN_SUCCEED;
 788
 789         /*
 790          * Copying pages' contents is subject to memory poison at any iteration.
 791          */
 792         for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR;
 793              _pte++, page++, _address += PAGE_SIZE) {
 794                 pteval = ptep_get(_pte);
 795                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 796                         clear_user_highpage(page, _address);
 797                         continue;
 798                 }
 799                 src_page = pte_page(pteval);
 800                 if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) {
 801                         result = SCAN_COPY_MC;
 802                         break;
 803                 }
 804         }
 805
 806         if (likely(result == SCAN_SUCCEED))
 807                 __collapse_huge_page_copy_succeeded(pte, vma, address, ptl,
 808                                                     compound_pagelist);
 809         else
 810                 __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,
 811                                                  compound_pagelist);
 812
 813         return result;
 814 }
 815
 816 static void khugepaged_alloc_sleep(void)
 817 {
 818         DEFINE_WAIT(wait);
 819
 820         add_wait_queue(&khugepaged_wait, &wait);
 821         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 822         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 823         remove_wait_queue(&khugepaged_wait, &wait);
 824 }
 825
 826 struct collapse_control khugepaged_collapse_control = {
 827         .is_khugepaged = true,
 828 };
 829
 830 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 831 {
 832         int i;
 833
 834         /*
 835          * If node_reclaim_mode is disabled, then no extra effort is made to
 836          * allocate memory locally.
 837          */
 838         if (!node_reclaim_enabled())
 839                 return false;
 840
 841         /* If there is a count for this node already, it must be acceptable */
 842         if (cc->node_load[nid])
 843                 return false;
 844
 845         for (i = 0; i < MAX_NUMNODES; i++) {
 846                 if (!cc->node_load[i])
 847                         continue;
 848                 if (node_distance(nid, i) > node_reclaim_distance)
 849                         return true;
 850         }
 851         return false;
 852 }
 853
 854 #define khugepaged_defrag()                                     \
 855         (transparent_hugepage_flags &                           \
 856          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 857
 858 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 859 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 860 {
 861         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 862 }
 863
 864 #ifdef CONFIG_NUMA
 865 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 866 {
 867         int nid, target_node = 0, max_value = 0;
 868
 869         /* find first node with max normal pages hit */
 870         for (nid = 0; nid < MAX_NUMNODES; nid++)
 871                 if (cc->node_load[nid] > max_value) {
 872                         max_value = cc->node_load[nid];
 873                         target_node = nid;
 874                 }
 875
 876         for_each_online_node(nid) {
 877                 if (max_value == cc->node_load[nid])
 878                         node_set(nid, cc->alloc_nmask);
 879         }
 880
 881         return target_node;
 882 }
 883 #else
 884 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 885 {
 886         return 0;
 887 }
 888 #endif
 889
 890 static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node,
 891                                       nodemask_t *nmask)
 892 {
 893         *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask);
 894         if (unlikely(!*hpage)) {
 895                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 896                 return false;
 897         }
 898
 899         folio_prep_large_rmappable((struct folio *)*hpage);
 900         count_vm_event(THP_COLLAPSE_ALLOC);
 901         return true;
 902 }
 903
 904 /*
 905  * If mmap_lock temporarily dropped, revalidate vma
 906  * before taking mmap_lock.
 907  * Returns enum scan_result value.
 908  */
 909
 910 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 911                                    bool expect_anon,
 912                                    struct vm_area_struct **vmap,
 913                                    struct collapse_control *cc)
 914 {
 915         struct vm_area_struct *vma;
 916
 917         if (unlikely(hpage_collapse_test_exit(mm)))
 918                 return SCAN_ANY_PROCESS;
 919
 920         *vmap = vma = find_vma(mm, address);
 921         if (!vma)
 922                 return SCAN_VMA_NULL;
 923
 924         if (!transhuge_vma_suitable(vma, address))
 925                 return SCAN_ADDRESS_RANGE;
 926         if (!hugepage_vma_check(vma, vma->vm_flags, false, false,
 927                                 cc->is_khugepaged))
 928                 return SCAN_VMA_CHECK;
 929         /*
 930          * Anon VMA expected, the address may be unmapped then
 931          * remapped to file after khugepaged reaquired the mmap_lock.
 932          *
 933          * hugepage_vma_check may return true for qualified file
 934          * vmas.
 935          */
 936         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 937                 return SCAN_PAGE_ANON;
 938         return SCAN_SUCCEED;
 939 }
 940
 941 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 942                                    unsigned long address,
 943                                    pmd_t **pmd)
 944 {
 945         pmd_t pmde;
 946
 947         *pmd = mm_find_pmd(mm, address);
 948         if (!*pmd)
 949                 return SCAN_PMD_NULL;
 950
 951         pmde = pmdp_get_lockless(*pmd);
 952         if (pmd_none(pmde))
 953                 return SCAN_PMD_NONE;
 954         if (!pmd_present(pmde))
 955                 return SCAN_PMD_NULL;
 956         if (pmd_trans_huge(pmde))
 957                 return SCAN_PMD_MAPPED;
 958         if (pmd_devmap(pmde))
 959                 return SCAN_PMD_NULL;
 960         if (pmd_bad(pmde))
 961                 return SCAN_PMD_NULL;
 962         return SCAN_SUCCEED;
 963 }
 964
 965 static int check_pmd_still_valid(struct mm_struct *mm,
 966                                  unsigned long address,
 967                                  pmd_t *pmd)
 968 {
 969         pmd_t *new_pmd;
 970         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 971
 972         if (result != SCAN_SUCCEED)
 973                 return result;
 974         if (new_pmd != pmd)
 975                 return SCAN_FAIL;
 976         return SCAN_SUCCEED;
 977 }
 978
 979 /*
 980  * Bring missing pages in from swap, to complete THP collapse.
 981  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 982  *
 983  * Called and returns without pte mapped or spinlocks held.
 984  * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.
 985  */
 986 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 987                                        struct vm_area_struct *vma,
 988                                        unsigned long haddr, pmd_t *pmd,
 989                                        int referenced)
 990 {
 991         int swapped_in = 0;
 992         vm_fault_t ret = 0;
 993         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 994         int result;
 995         pte_t *pte = NULL;
 996         spinlock_t *ptl;
 997
 998         for (address = haddr; address < end; address += PAGE_SIZE) {
 999                 struct vm_fault vmf = {
1000                         .vma = vma,
1001                         .address = address,
1002                         .pgoff = linear_page_index(vma, address),
1003                         .flags = FAULT_FLAG_ALLOW_RETRY,
1004                         .pmd = pmd,
1005                 };
1006
1007                 if (!pte++) {
1008                         pte = pte_offset_map_nolock(mm, pmd, address, &ptl);
1009                         if (!pte) {
1010                                 mmap_read_unlock(mm);
1011                                 result = SCAN_PMD_NULL;
1012                                 goto out;
1013                         }
1014                 }
1015
1016                 vmf.orig_pte = ptep_get_lockless(pte);
1017                 if (!is_swap_pte(vmf.orig_pte))
1018                         continue;
1019
1020                 vmf.pte = pte;
1021                 vmf.ptl = ptl;
1022                 ret = do_swap_page(&vmf);
1023                 /* Which unmaps pte (after perhaps re-checking the entry) */
1024                 pte = NULL;
1025
1026                 /*
1027                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1028                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1029                  * we do not retry here and swap entry will remain in pagetable
1030                  * resulting in later failure.
1031                  */
1032                 if (ret & VM_FAULT_RETRY) {
1033                         /* Likely, but not guaranteed, that page lock failed */
1034                         result = SCAN_PAGE_LOCK;
1035                         goto out;
1036                 }
1037                 if (ret & VM_FAULT_ERROR) {
1038                         mmap_read_unlock(mm);
1039                         result = SCAN_FAIL;
1040                         goto out;
1041                 }
1042                 swapped_in++;
1043         }
1044
1045         if (pte)
1046                 pte_unmap(pte);
1047
1048         /* Drain LRU cache to remove extra pin on the swapped in pages */
1049         if (swapped_in)
1050                 lru_add_drain();
1051
1052         result = SCAN_SUCCEED;
1053 out:
1054         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);
1055         return result;
1056 }
1057
1058 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
1059                               struct collapse_control *cc)
1060 {
1061         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
1062                      GFP_TRANSHUGE);
1063         int node = hpage_collapse_find_target_node(cc);
1064         struct folio *folio;
1065
1066         if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask))
1067                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
1068
1069         folio = page_folio(*hpage);
1070         if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {
1071                 folio_put(folio);
1072                 *hpage = NULL;
1073                 return SCAN_CGROUP_CHARGE_FAIL;
1074         }
1075         count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC);
1076
1077         return SCAN_SUCCEED;
1078 }
1079
1080 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
1081                               int referenced, int unmapped,
1082                               struct collapse_control *cc)
1083 {
1084         LIST_HEAD(compound_pagelist);
1085         pmd_t *pmd, _pmd;
1086         pte_t *pte;
1087         pgtable_t pgtable;
1088         struct page *hpage;
1089         spinlock_t *pmd_ptl, *pte_ptl;
1090         int result = SCAN_FAIL;
1091         struct vm_area_struct *vma;
1092         struct mmu_notifier_range range;
1093
1094         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1095
1096         /*
1097          * Before allocating the hugepage, release the mmap_lock read lock.
1098          * The allocation can take potentially a long time if it involves
1099          * sync compaction, and we do not need to hold the mmap_lock during
1100          * that. We will recheck the vma after taking it again in write mode.
1101          */
1102         mmap_read_unlock(mm);
1103
1104         result = alloc_charge_hpage(&hpage, mm, cc);
1105         if (result != SCAN_SUCCEED)
1106                 goto out_nolock;
1107
1108         mmap_read_lock(mm);
1109         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1110         if (result != SCAN_SUCCEED) {
1111                 mmap_read_unlock(mm);
1112                 goto out_nolock;
1113         }
1114
1115         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1116         if (result != SCAN_SUCCEED) {
1117                 mmap_read_unlock(mm);
1118                 goto out_nolock;
1119         }
1120
1121         if (unmapped) {
1122                 /*
1123                  * __collapse_huge_page_swapin will return with mmap_lock
1124                  * released when it fails. So we jump out_nolock directly in
1125                  * that case.  Continuing to collapse causes inconsistency.
1126                  */
1127                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1128                                                      referenced);
1129                 if (result != SCAN_SUCCEED)
1130                         goto out_nolock;
1131         }
1132
1133         mmap_read_unlock(mm);
1134         /*
1135          * Prevent all access to pagetables with the exception of
1136          * gup_fast later handled by the ptep_clear_flush and the VM
1137          * handled by the anon_vma lock + PG_lock.
1138          */
1139         mmap_write_lock(mm);
1140         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1141         if (result != SCAN_SUCCEED)
1142                 goto out_up_write;
1143         /* check if the pmd is still valid */
1144         result = check_pmd_still_valid(mm, address, pmd);
1145         if (result != SCAN_SUCCEED)
1146                 goto out_up_write;
1147
1148         vma_start_write(vma);
1149         anon_vma_lock_write(vma->anon_vma);
1150
1151         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1152                                 address + HPAGE_PMD_SIZE);
1153         mmu_notifier_invalidate_range_start(&range);
1154
1155         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1156         /*
1157          * This removes any huge TLB entry from the CPU so we won't allow
1158          * huge and small TLB entries for the same virtual address to
1159          * avoid the risk of CPU bugs in that area.
1160          *
1161          * Parallel fast GUP is fine since fast GUP will back off when
1162          * it detects PMD is changed.
1163          */
1164         _pmd = pmdp_collapse_flush(vma, address, pmd);
1165         spin_unlock(pmd_ptl);
1166         mmu_notifier_invalidate_range_end(&range);
1167         tlb_remove_table_sync_one();
1168
1169         pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
1170         if (pte) {
1171                 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1172                                                       &compound_pagelist);
1173                 spin_unlock(pte_ptl);
1174         } else {
1175                 result = SCAN_PMD_NULL;
1176         }
1177
1178         if (unlikely(result != SCAN_SUCCEED)) {
1179                 if (pte)
1180                         pte_unmap(pte);
1181                 spin_lock(pmd_ptl);
1182                 BUG_ON(!pmd_none(*pmd));
1183                 /*
1184                  * We can only use set_pmd_at when establishing
1185                  * hugepmds and never for establishing regular pmds that
1186                  * points to regular pagetables. Use pmd_populate for that
1187                  */
1188                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1189                 spin_unlock(pmd_ptl);
1190                 anon_vma_unlock_write(vma->anon_vma);
1191                 goto out_up_write;
1192         }
1193
1194         /*
1195          * All pages are isolated and locked so anon_vma rmap
1196          * can't run anymore.
1197          */
1198         anon_vma_unlock_write(vma->anon_vma);
1199
1200         result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd,
1201                                            vma, address, pte_ptl,
1202                                            &compound_pagelist);
1203         pte_unmap(pte);
1204         if (unlikely(result != SCAN_SUCCEED))
1205                 goto out_up_write;
1206
1207         /*
1208          * spin_lock() below is not the equivalent of smp_wmb(), but
1209          * the smp_wmb() inside __SetPageUptodate() can be reused to
1210          * avoid the copy_huge_page writes to become visible after
1211          * the set_pmd_at() write.
1212          */
1213         __SetPageUptodate(hpage);
1214         pgtable = pmd_pgtable(_pmd);
1215
1216         _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1217         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1218
1219         spin_lock(pmd_ptl);
1220         BUG_ON(!pmd_none(*pmd));
1221         page_add_new_anon_rmap(hpage, vma, address);
1222         lru_cache_add_inactive_or_unevictable(hpage, vma);
1223         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1224         set_pmd_at(mm, address, pmd, _pmd);
1225         update_mmu_cache_pmd(vma, address, pmd);
1226         spin_unlock(pmd_ptl);
1227
1228         hpage = NULL;
1229
1230         result = SCAN_SUCCEED;
1231 out_up_write:
1232         mmap_write_unlock(mm);
1233 out_nolock:
1234         if (hpage)
1235                 put_page(hpage);
1236         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1237         return result;
1238 }
1239
1240 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1241                                    struct vm_area_struct *vma,
1242                                    unsigned long address, bool *mmap_locked,
1243                                    struct collapse_control *cc)
1244 {
1245         pmd_t *pmd;
1246         pte_t *pte, *_pte;
1247         int result = SCAN_FAIL, referenced = 0;
1248         int none_or_zero = 0, shared = 0;
1249         struct page *page = NULL;
1250         unsigned long _address;
1251         spinlock_t *ptl;
1252         int node = NUMA_NO_NODE, unmapped = 0;
1253         bool writable = false;
1254
1255         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1256
1257         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1258         if (result != SCAN_SUCCEED)
1259                 goto out;
1260
1261         memset(cc->node_load, 0, sizeof(cc->node_load));
1262         nodes_clear(cc->alloc_nmask);
1263         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1264         if (!pte) {
1265                 result = SCAN_PMD_NULL;
1266                 goto out;
1267         }
1268
1269         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1270              _pte++, _address += PAGE_SIZE) {
1271                 pte_t pteval = ptep_get(_pte);
1272                 if (is_swap_pte(pteval)) {
1273                         ++unmapped;
1274                         if (!cc->is_khugepaged ||
1275                             unmapped <= khugepaged_max_ptes_swap) {
1276                                 /*
1277                                  * Always be strict with uffd-wp
1278                                  * enabled swap entries.  Please see
1279                                  * comment below for pte_uffd_wp().
1280                                  */
1281                                 if (pte_swp_uffd_wp_any(pteval)) {
1282                                         result = SCAN_PTE_UFFD_WP;
1283                                         goto out_unmap;
1284                                 }
1285                                 continue;
1286                         } else {
1287                                 result = SCAN_EXCEED_SWAP_PTE;
1288                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1289                                 goto out_unmap;
1290                         }
1291                 }
1292                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1293                         ++none_or_zero;
1294                         if (!userfaultfd_armed(vma) &&
1295                             (!cc->is_khugepaged ||
1296                              none_or_zero <= khugepaged_max_ptes_none)) {
1297                                 continue;
1298                         } else {
1299                                 result = SCAN_EXCEED_NONE_PTE;
1300                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1301                                 goto out_unmap;
1302                         }
1303                 }
1304                 if (pte_uffd_wp(pteval)) {
1305                         /*
1306                          * Don't collapse the page if any of the small
1307                          * PTEs are armed with uffd write protection.
1308                          * Here we can also mark the new huge pmd as
1309                          * write protected if any of the small ones is
1310                          * marked but that could bring unknown
1311                          * userfault messages that falls outside of
1312                          * the registered range.  So, just be simple.
1313                          */
1314                         result = SCAN_PTE_UFFD_WP;
1315                         goto out_unmap;
1316                 }
1317                 if (pte_write(pteval))
1318                         writable = true;
1319
1320                 page = vm_normal_page(vma, _address, pteval);
1321                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1322                         result = SCAN_PAGE_NULL;
1323                         goto out_unmap;
1324                 }
1325
1326                 if (page_mapcount(page) > 1) {
1327                         ++shared;
1328                         if (cc->is_khugepaged &&
1329                             shared > khugepaged_max_ptes_shared) {
1330                                 result = SCAN_EXCEED_SHARED_PTE;
1331                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1332                                 goto out_unmap;
1333                         }
1334                 }
1335
1336                 page = compound_head(page);
1337
1338                 /*
1339                  * Record which node the original page is from and save this
1340                  * information to cc->node_load[].
1341                  * Khugepaged will allocate hugepage from the node has the max
1342                  * hit record.
1343                  */
1344                 node = page_to_nid(page);
1345                 if (hpage_collapse_scan_abort(node, cc)) {
1346                         result = SCAN_SCAN_ABORT;
1347                         goto out_unmap;
1348                 }
1349                 cc->node_load[node]++;
1350                 if (!PageLRU(page)) {
1351                         result = SCAN_PAGE_LRU;
1352                         goto out_unmap;
1353                 }
1354                 if (PageLocked(page)) {
1355                         result = SCAN_PAGE_LOCK;
1356                         goto out_unmap;
1357                 }
1358                 if (!PageAnon(page)) {
1359                         result = SCAN_PAGE_ANON;
1360                         goto out_unmap;
1361                 }
1362
1363                 /*
1364                  * Check if the page has any GUP (or other external) pins.
1365                  *
1366                  * Here the check may be racy:
1367                  * it may see total_mapcount > refcount in some cases?
1368                  * But such case is ephemeral we could always retry collapse
1369                  * later.  However it may report false positive if the page
1370                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1371                  * will be done again later the risk seems low.
1372                  */
1373                 if (!is_refcount_suitable(page)) {
1374                         result = SCAN_PAGE_COUNT;
1375                         goto out_unmap;
1376                 }
1377
1378                 /*
1379                  * If collapse was initiated by khugepaged, check that there is
1380                  * enough young pte to justify collapsing the page
1381                  */
1382                 if (cc->is_khugepaged &&
1383                     (pte_young(pteval) || page_is_young(page) ||
1384                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
1385                                                                      address)))
1386                         referenced++;
1387         }
1388         if (!writable) {
1389                 result = SCAN_PAGE_RO;
1390         } else if (cc->is_khugepaged &&
1391                    (!referenced ||
1392                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1393                 result = SCAN_LACK_REFERENCED_PAGE;
1394         } else {
1395                 result = SCAN_SUCCEED;
1396         }
1397 out_unmap:
1398         pte_unmap_unlock(pte, ptl);
1399         if (result == SCAN_SUCCEED) {
1400                 result = collapse_huge_page(mm, address, referenced,
1401                                             unmapped, cc);
1402                 /* collapse_huge_page will return with the mmap_lock released */
1403                 *mmap_locked = false;
1404         }
1405 out:
1406         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1407                                      none_or_zero, result, unmapped);
1408         return result;
1409 }
1410
1411 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1412 {
1413         struct mm_slot *slot = &mm_slot->slot;
1414         struct mm_struct *mm = slot->mm;
1415
1416         lockdep_assert_held(&khugepaged_mm_lock);
1417
1418         if (hpage_collapse_test_exit(mm)) {
1419                 /* free mm_slot */
1420                 hash_del(&slot->hash);
1421                 list_del(&slot->mm_node);
1422
1423                 /*
1424                  * Not strictly needed because the mm exited already.
1425                  *
1426                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1427                  */
1428
1429                 /* khugepaged_mm_lock actually not necessary for the below */
1430                 mm_slot_free(mm_slot_cache, mm_slot);
1431                 mmdrop(mm);
1432         }
1433 }
1434
1435 #ifdef CONFIG_SHMEM
1436 /* hpage must be locked, and mmap_lock must be held */
1437 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1438                         pmd_t *pmdp, struct page *hpage)
1439 {
1440         struct vm_fault vmf = {
1441                 .vma = vma,
1442                 .address = addr,
1443                 .flags = 0,
1444                 .pmd = pmdp,
1445         };
1446
1447         VM_BUG_ON(!PageTransHuge(hpage));
1448         mmap_assert_locked(vma->vm_mm);
1449
1450         if (do_set_pmd(&vmf, hpage))
1451                 return SCAN_FAIL;
1452
1453         get_page(hpage);
1454         return SCAN_SUCCEED;
1455 }
1456
1457 /**
1458  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1459  * address haddr.
1460  *
1461  * @mm: process address space where collapse happens
1462  * @addr: THP collapse address
1463  * @install_pmd: If a huge PMD should be installed
1464  *
1465  * This function checks whether all the PTEs in the PMD are pointing to the
1466  * right THP. If so, retract the page table so the THP can refault in with
1467  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1468  */
1469 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1470                             bool install_pmd)
1471 {
1472         struct mmu_notifier_range range;
1473         bool notified = false;
1474         unsigned long haddr = addr & HPAGE_PMD_MASK;
1475         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1476         struct page *hpage;
1477         pte_t *start_pte, *pte;
1478         pmd_t *pmd, pgt_pmd;
1479         spinlock_t *pml = NULL, *ptl;
1480         int nr_ptes = 0, result = SCAN_FAIL;
1481         int i;
1482
1483         mmap_assert_locked(mm);
1484
1485         /* First check VMA found, in case page tables are being torn down */
1486         if (!vma || !vma->vm_file ||
1487             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1488                 return SCAN_VMA_CHECK;
1489
1490         /* Fast check before locking page if already PMD-mapped */
1491         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1492         if (result == SCAN_PMD_MAPPED)
1493                 return result;
1494
1495         /*
1496          * If we are here, we've succeeded in replacing all the native pages
1497          * in the page cache with a single hugepage. If a mm were to fault-in
1498          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1499          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1500          * analogously elide sysfs THP settings here.
1501          */
1502         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
1503                 return SCAN_VMA_CHECK;
1504
1505         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1506         if (userfaultfd_wp(vma))
1507                 return SCAN_PTE_UFFD_WP;
1508
1509         hpage = find_lock_page(vma->vm_file->f_mapping,
1510                                linear_page_index(vma, haddr));
1511         if (!hpage)
1512                 return SCAN_PAGE_NULL;
1513
1514         if (!PageHead(hpage)) {
1515                 result = SCAN_FAIL;
1516                 goto drop_hpage;
1517         }
1518
1519         if (compound_order(hpage) != HPAGE_PMD_ORDER) {
1520                 result = SCAN_PAGE_COMPOUND;
1521                 goto drop_hpage;
1522         }
1523
1524         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1525         switch (result) {
1526         case SCAN_SUCCEED:
1527                 break;
1528         case SCAN_PMD_NONE:
1529                 /*
1530                  * All pte entries have been removed and pmd cleared.
1531                  * Skip all the pte checks and just update the pmd mapping.
1532                  */
1533                 goto maybe_install_pmd;
1534         default:
1535                 goto drop_hpage;
1536         }
1537
1538         result = SCAN_FAIL;
1539         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1540         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1541                 goto drop_hpage;
1542
1543         /* step 1: check all mapped PTEs are to the right huge page */
1544         for (i = 0, addr = haddr, pte = start_pte;
1545              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1546                 struct page *page;
1547                 pte_t ptent = ptep_get(pte);
1548
1549                 /* empty pte, skip */
1550                 if (pte_none(ptent))
1551                         continue;
1552
1553                 /* page swapped out, abort */
1554                 if (!pte_present(ptent)) {
1555                         result = SCAN_PTE_NON_PRESENT;
1556                         goto abort;
1557                 }
1558
1559                 page = vm_normal_page(vma, addr, ptent);
1560                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1561                         page = NULL;
1562                 /*
1563                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1564                  * page table, but the new page will not be a subpage of hpage.
1565                  */
1566                 if (hpage + i != page)
1567                         goto abort;
1568         }
1569
1570         pte_unmap_unlock(start_pte, ptl);
1571         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1572                                 haddr, haddr + HPAGE_PMD_SIZE);
1573         mmu_notifier_invalidate_range_start(&range);
1574         notified = true;
1575
1576         /*
1577          * pmd_lock covers a wider range than ptl, and (if split from mm's
1578          * page_table_lock) ptl nests inside pml. The less time we hold pml,
1579          * the better; but userfaultfd's mfill_atomic_pte() on a private VMA
1580          * inserts a valid as-if-COWed PTE without even looking up page cache.
1581          * So page lock of hpage does not protect from it, so we must not drop
1582          * ptl before pgt_pmd is removed, so uffd private needs pml taken now.
1583          */
1584         if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
1585                 pml = pmd_lock(mm, pmd);
1586
1587         start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);
1588         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1589                 goto abort;
1590         if (!pml)
1591                 spin_lock(ptl);
1592         else if (ptl != pml)
1593                 spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1594
1595         /* step 2: clear page table and adjust rmap */
1596         for (i = 0, addr = haddr, pte = start_pte;
1597              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1598                 struct page *page;
1599                 pte_t ptent = ptep_get(pte);
1600
1601                 if (pte_none(ptent))
1602                         continue;
1603                 /*
1604                  * We dropped ptl after the first scan, to do the mmu_notifier:
1605                  * page lock stops more PTEs of the hpage being faulted in, but
1606                  * does not stop write faults COWing anon copies from existing
1607                  * PTEs; and does not stop those being swapped out or migrated.
1608                  */
1609                 if (!pte_present(ptent)) {
1610                         result = SCAN_PTE_NON_PRESENT;
1611                         goto abort;
1612                 }
1613                 page = vm_normal_page(vma, addr, ptent);
1614                 if (hpage + i != page)
1615                         goto abort;
1616
1617                 /*
1618                  * Must clear entry, or a racing truncate may re-remove it.
1619                  * TLB flush can be left until pmdp_collapse_flush() does it.
1620                  * PTE dirty? Shmem page is already dirty; file is read-only.
1621                  */
1622                 ptep_clear(mm, addr, pte);
1623                 page_remove_rmap(page, vma, false);
1624                 nr_ptes++;
1625         }
1626
1627         pte_unmap(start_pte);
1628         if (!pml)
1629                 spin_unlock(ptl);
1630
1631         /* step 3: set proper refcount and mm_counters. */
1632         if (nr_ptes) {
1633                 page_ref_sub(hpage, nr_ptes);
1634                 add_mm_counter(mm, mm_counter_file(hpage), -nr_ptes);
1635         }
1636
1637         /* step 4: remove empty page table */
1638         if (!pml) {
1639                 pml = pmd_lock(mm, pmd);
1640                 if (ptl != pml)
1641                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1642         }
1643         pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1644         pmdp_get_lockless_sync();
1645         if (ptl != pml)
1646                 spin_unlock(ptl);
1647         spin_unlock(pml);
1648
1649         mmu_notifier_invalidate_range_end(&range);
1650
1651         mm_dec_nr_ptes(mm);
1652         page_table_check_pte_clear_range(mm, haddr, pgt_pmd);
1653         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1654
1655 maybe_install_pmd:
1656         /* step 5: install pmd entry */
1657         result = install_pmd
1658                         ? set_huge_pmd(vma, haddr, pmd, hpage)
1659                         : SCAN_SUCCEED;
1660         goto drop_hpage;
1661 abort:
1662         if (nr_ptes) {
1663                 flush_tlb_mm(mm);
1664                 page_ref_sub(hpage, nr_ptes);
1665                 add_mm_counter(mm, mm_counter_file(hpage), -nr_ptes);
1666         }
1667         if (start_pte)
1668                 pte_unmap_unlock(start_pte, ptl);
1669         if (pml && pml != ptl)
1670                 spin_unlock(pml);
1671         if (notified)
1672                 mmu_notifier_invalidate_range_end(&range);
1673 drop_hpage:
1674         unlock_page(hpage);
1675         put_page(hpage);
1676         return result;
1677 }
1678
1679 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1680 {
1681         struct vm_area_struct *vma;
1682
1683         i_mmap_lock_read(mapping);
1684         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1685                 struct mmu_notifier_range range;
1686                 struct mm_struct *mm;
1687                 unsigned long addr;
1688                 pmd_t *pmd, pgt_pmd;
1689                 spinlock_t *pml;
1690                 spinlock_t *ptl;
1691                 bool skipped_uffd = false;
1692
1693                 /*
1694                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1695                  * got written to. These VMAs are likely not worth removing
1696                  * page tables from, as PMD-mapping is likely to be split later.
1697                  */
1698                 if (READ_ONCE(vma->anon_vma))
1699                         continue;
1700
1701                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1702                 if (addr & ~HPAGE_PMD_MASK ||
1703                     vma->vm_end < addr + HPAGE_PMD_SIZE)
1704                         continue;
1705
1706                 mm = vma->vm_mm;
1707                 if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)
1708                         continue;
1709
1710                 if (hpage_collapse_test_exit(mm))
1711                         continue;
1712                 /*
1713                  * When a vma is registered with uffd-wp, we cannot recycle
1714                  * the page table because there may be pte markers installed.
1715                  * Other vmas can still have the same file mapped hugely, but
1716                  * skip this one: it will always be mapped in small page size
1717                  * for uffd-wp registered ranges.
1718                  */
1719                 if (userfaultfd_wp(vma))
1720                         continue;
1721
1722                 /* PTEs were notified when unmapped; but now for the PMD? */
1723                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1724                                         addr, addr + HPAGE_PMD_SIZE);
1725                 mmu_notifier_invalidate_range_start(&range);
1726
1727                 pml = pmd_lock(mm, pmd);
1728                 ptl = pte_lockptr(mm, pmd);
1729                 if (ptl != pml)
1730                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1731
1732                 /*
1733                  * Huge page lock is still held, so normally the page table
1734                  * must remain empty; and we have already skipped anon_vma
1735                  * and userfaultfd_wp() vmas.  But since the mmap_lock is not
1736                  * held, it is still possible for a racing userfaultfd_ioctl()
1737                  * to have inserted ptes or markers.  Now that we hold ptlock,
1738                  * repeating the anon_vma check protects from one category,
1739                  * and repeating the userfaultfd_wp() check from another.
1740                  */
1741                 if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) {
1742                         skipped_uffd = true;
1743                 } else {
1744                         pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);
1745                         pmdp_get_lockless_sync();
1746                 }
1747
1748                 if (ptl != pml)
1749                         spin_unlock(ptl);
1750                 spin_unlock(pml);
1751
1752                 mmu_notifier_invalidate_range_end(&range);
1753
1754                 if (!skipped_uffd) {
1755                         mm_dec_nr_ptes(mm);
1756                         page_table_check_pte_clear_range(mm, addr, pgt_pmd);
1757                         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1758                 }
1759         }
1760         i_mmap_unlock_read(mapping);
1761 }
1762
1763 /**
1764  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1765  *
1766  * @mm: process address space where collapse happens
1767  * @addr: virtual collapse start address
1768  * @file: file that collapse on
1769  * @start: collapse start address
1770  * @cc: collapse context and scratchpad
1771  *
1772  * Basic scheme is simple, details are more complex:
1773  *  - allocate and lock a new huge page;
1774  *  - scan page cache, locking old pages
1775  *    + swap/gup in pages if necessary;
1776  *  - copy data to new page
1777  *  - handle shmem holes
1778  *    + re-validate that holes weren't filled by someone else
1779  *    + check for userfaultfd
1780  *  - finalize updates to the page cache;
1781  *  - if replacing succeeds:
1782  *    + unlock huge page;
1783  *    + free old pages;
1784  *  - if replacing failed;
1785  *    + unlock old pages
1786  *    + unlock and free huge page;
1787  */
1788 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1789                          struct file *file, pgoff_t start,
1790                          struct collapse_control *cc)
1791 {
1792         struct address_space *mapping = file->f_mapping;
1793         struct page *hpage;
1794         struct page *page;
1795         struct page *tmp;
1796         struct folio *folio;
1797         pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1798         LIST_HEAD(pagelist);
1799         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1800         int nr_none = 0, result = SCAN_SUCCEED;
1801         bool is_shmem = shmem_file(file);
1802         int nr = 0;
1803
1804         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1805         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1806
1807         result = alloc_charge_hpage(&hpage, mm, cc);
1808         if (result != SCAN_SUCCEED)
1809                 goto out;
1810
1811         __SetPageLocked(hpage);
1812         if (is_shmem)
1813                 __SetPageSwapBacked(hpage);
1814         hpage->index = start;
1815         hpage->mapping = mapping;
1816
1817         /*
1818          * Ensure we have slots for all the pages in the range.  This is
1819          * almost certainly a no-op because most of the pages must be present
1820          */
1821         do {
1822                 xas_lock_irq(&xas);
1823                 xas_create_range(&xas);
1824                 if (!xas_error(&xas))
1825                         break;
1826                 xas_unlock_irq(&xas);
1827                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1828                         result = SCAN_FAIL;
1829                         goto rollback;
1830                 }
1831         } while (1);
1832
1833         for (index = start; index < end; index++) {
1834                 xas_set(&xas, index);
1835                 page = xas_load(&xas);
1836
1837                 VM_BUG_ON(index != xas.xa_index);
1838                 if (is_shmem) {
1839                         if (!page) {
1840                                 /*
1841                                  * Stop if extent has been truncated or
1842                                  * hole-punched, and is now completely
1843                                  * empty.
1844                                  */
1845                                 if (index == start) {
1846                                         if (!xas_next_entry(&xas, end - 1)) {
1847                                                 result = SCAN_TRUNCATED;
1848                                                 goto xa_locked;
1849                                         }
1850                                 }
1851                                 nr_none++;
1852                                 continue;
1853                         }
1854
1855                         if (xa_is_value(page) || !PageUptodate(page)) {
1856                                 xas_unlock_irq(&xas);
1857                                 /* swap in or instantiate fallocated page */
1858                                 if (shmem_get_folio(mapping->host, index,
1859                                                 &folio, SGP_NOALLOC)) {
1860                                         result = SCAN_FAIL;
1861                                         goto xa_unlocked;
1862                                 }
1863                                 /* drain lru cache to help isolate_lru_page() */
1864                                 lru_add_drain();
1865                                 page = folio_file_page(folio, index);
1866                         } else if (trylock_page(page)) {
1867                                 get_page(page);
1868                                 xas_unlock_irq(&xas);
1869                         } else {
1870                                 result = SCAN_PAGE_LOCK;
1871                                 goto xa_locked;
1872                         }
1873                 } else {        /* !is_shmem */
1874                         if (!page || xa_is_value(page)) {
1875                                 xas_unlock_irq(&xas);
1876                                 page_cache_sync_readahead(mapping, &file->f_ra,
1877                                                           file, index,
1878                                                           end - index);
1879                                 /* drain lru cache to help isolate_lru_page() */
1880                                 lru_add_drain();
1881                                 page = find_lock_page(mapping, index);
1882                                 if (unlikely(page == NULL)) {
1883                                         result = SCAN_FAIL;
1884                                         goto xa_unlocked;
1885                                 }
1886                         } else if (PageDirty(page)) {
1887                                 /*
1888                                  * khugepaged only works on read-only fd,
1889                                  * so this page is dirty because it hasn't
1890                                  * been flushed since first write. There
1891                                  * won't be new dirty pages.
1892                                  *
1893                                  * Trigger async flush here and hope the
1894                                  * writeback is done when khugepaged
1895                                  * revisits this page.
1896                                  *
1897                                  * This is a one-off situation. We are not
1898                                  * forcing writeback in loop.
1899                                  */
1900                                 xas_unlock_irq(&xas);
1901                                 filemap_flush(mapping);
1902                                 result = SCAN_FAIL;
1903                                 goto xa_unlocked;
1904                         } else if (PageWriteback(page)) {
1905                                 xas_unlock_irq(&xas);
1906                                 result = SCAN_FAIL;
1907                                 goto xa_unlocked;
1908                         } else if (trylock_page(page)) {
1909                                 get_page(page);
1910                                 xas_unlock_irq(&xas);
1911                         } else {
1912                                 result = SCAN_PAGE_LOCK;
1913                                 goto xa_locked;
1914                         }
1915                 }
1916
1917                 /*
1918                  * The page must be locked, so we can drop the i_pages lock
1919                  * without racing with truncate.
1920                  */
1921                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1922
1923                 /* make sure the page is up to date */
1924                 if (unlikely(!PageUptodate(page))) {
1925                         result = SCAN_FAIL;
1926                         goto out_unlock;
1927                 }
1928
1929                 /*
1930                  * If file was truncated then extended, or hole-punched, before
1931                  * we locked the first page, then a THP might be there already.
1932                  * This will be discovered on the first iteration.
1933                  */
1934                 if (PageTransCompound(page)) {
1935                         struct page *head = compound_head(page);
1936
1937                         result = compound_order(head) == HPAGE_PMD_ORDER &&
1938                                         head->index == start
1939                                         /* Maybe PMD-mapped */
1940                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1941                                         : SCAN_PAGE_COMPOUND;
1942                         goto out_unlock;
1943                 }
1944
1945                 folio = page_folio(page);
1946
1947                 if (folio_mapping(folio) != mapping) {
1948                         result = SCAN_TRUNCATED;
1949                         goto out_unlock;
1950                 }
1951
1952                 if (!is_shmem && (folio_test_dirty(folio) ||
1953                                   folio_test_writeback(folio))) {
1954                         /*
1955                          * khugepaged only works on read-only fd, so this
1956                          * page is dirty because it hasn't been flushed
1957                          * since first write.
1958                          */
1959                         result = SCAN_FAIL;
1960                         goto out_unlock;
1961                 }
1962
1963                 if (!folio_isolate_lru(folio)) {
1964                         result = SCAN_DEL_PAGE_LRU;
1965                         goto out_unlock;
1966                 }
1967
1968                 if (!filemap_release_folio(folio, GFP_KERNEL)) {
1969                         result = SCAN_PAGE_HAS_PRIVATE;
1970                         folio_putback_lru(folio);
1971                         goto out_unlock;
1972                 }
1973
1974                 if (folio_mapped(folio))
1975                         try_to_unmap(folio,
1976                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1977
1978                 xas_lock_irq(&xas);
1979
1980                 VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page);
1981
1982                 /*
1983                  * We control three references to the page:
1984                  *  - we hold a pin on it;
1985                  *  - one reference from page cache;
1986                  *  - one from isolate_lru_page;
1987                  * If those are the only references, then any new usage of the
1988                  * page will have to fetch it from the page cache. That requires
1989                  * locking the page to handle truncate, so any new usage will be
1990                  * blocked until we unlock page after collapse/during rollback.
1991                  */
1992                 if (page_count(page) != 3) {
1993                         result = SCAN_PAGE_COUNT;
1994                         xas_unlock_irq(&xas);
1995                         putback_lru_page(page);
1996                         goto out_unlock;
1997                 }
1998
1999                 /*
2000                  * Accumulate the pages that are being collapsed.
2001                  */
2002                 list_add_tail(&page->lru, &pagelist);
2003                 continue;
2004 out_unlock:
2005                 unlock_page(page);
2006                 put_page(page);
2007                 goto xa_unlocked;
2008         }
2009
2010         if (!is_shmem) {
2011                 filemap_nr_thps_inc(mapping);
2012                 /*
2013                  * Paired with smp_mb() in do_dentry_open() to ensure
2014                  * i_writecount is up to date and the update to nr_thps is
2015                  * visible. Ensures the page cache will be truncated if the
2016                  * file is opened writable.
2017                  */
2018                 smp_mb();
2019                 if (inode_is_open_for_write(mapping->host)) {
2020                         result = SCAN_FAIL;
2021                         filemap_nr_thps_dec(mapping);
2022                 }
2023         }
2024
2025 xa_locked:
2026         xas_unlock_irq(&xas);
2027 xa_unlocked:
2028
2029         /*
2030          * If collapse is successful, flush must be done now before copying.
2031          * If collapse is unsuccessful, does flush actually need to be done?
2032          * Do it anyway, to clear the state.
2033          */
2034         try_to_unmap_flush();
2035
2036         if (result == SCAN_SUCCEED && nr_none &&
2037             !shmem_charge(mapping->host, nr_none))
2038                 result = SCAN_FAIL;
2039         if (result != SCAN_SUCCEED) {
2040                 nr_none = 0;
2041                 goto rollback;
2042         }
2043
2044         /*
2045          * The old pages are locked, so they won't change anymore.
2046          */
2047         index = start;
2048         list_for_each_entry(page, &pagelist, lru) {
2049                 while (index < page->index) {
2050                         clear_highpage(hpage + (index % HPAGE_PMD_NR));
2051                         index++;
2052                 }
2053                 if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) {
2054                         result = SCAN_COPY_MC;
2055                         goto rollback;
2056                 }
2057                 index++;
2058         }
2059         while (index < end) {
2060                 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2061                 index++;
2062         }
2063
2064         if (nr_none) {
2065                 struct vm_area_struct *vma;
2066                 int nr_none_check = 0;
2067
2068                 i_mmap_lock_read(mapping);
2069                 xas_lock_irq(&xas);
2070
2071                 xas_set(&xas, start);
2072                 for (index = start; index < end; index++) {
2073                         if (!xas_next(&xas)) {
2074                                 xas_store(&xas, XA_RETRY_ENTRY);
2075                                 if (xas_error(&xas)) {
2076                                         result = SCAN_STORE_FAILED;
2077                                         goto immap_locked;
2078                                 }
2079                                 nr_none_check++;
2080                         }
2081                 }
2082
2083                 if (nr_none != nr_none_check) {
2084                         result = SCAN_PAGE_FILLED;
2085                         goto immap_locked;
2086                 }
2087
2088                 /*
2089                  * If userspace observed a missing page in a VMA with a MODE_MISSING
2090                  * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that
2091                  * page. If so, we need to roll back to avoid suppressing such an
2092                  * event. Since wp/minor userfaultfds don't give userspace any
2093                  * guarantees that the kernel doesn't fill a missing page with a zero
2094                  * page, so they don't matter here.
2095                  *
2096                  * Any userfaultfds registered after this point will not be able to
2097                  * observe any missing pages due to the previously inserted retry
2098                  * entries.
2099                  */
2100                 vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {
2101                         if (userfaultfd_missing(vma)) {
2102                                 result = SCAN_EXCEED_NONE_PTE;
2103                                 goto immap_locked;
2104                         }
2105                 }
2106
2107 immap_locked:
2108                 i_mmap_unlock_read(mapping);
2109                 if (result != SCAN_SUCCEED) {
2110                         xas_set(&xas, start);
2111                         for (index = start; index < end; index++) {
2112                                 if (xas_next(&xas) == XA_RETRY_ENTRY)
2113                                         xas_store(&xas, NULL);
2114                         }
2115
2116                         xas_unlock_irq(&xas);
2117                         goto rollback;
2118                 }
2119         } else {
2120                 xas_lock_irq(&xas);
2121         }
2122
2123         nr = thp_nr_pages(hpage);
2124         if (is_shmem)
2125                 __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr);
2126         else
2127                 __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr);
2128
2129         if (nr_none) {
2130                 __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none);
2131                 /* nr_none is always 0 for non-shmem. */
2132                 __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none);
2133         }
2134
2135         /*
2136          * Mark hpage as uptodate before inserting it into the page cache so
2137          * that it isn't mistaken for an fallocated but unwritten page.
2138          */
2139         folio = page_folio(hpage);
2140         folio_mark_uptodate(folio);
2141         folio_ref_add(folio, HPAGE_PMD_NR - 1);
2142
2143         if (is_shmem)
2144                 folio_mark_dirty(folio);
2145         folio_add_lru(folio);
2146
2147         /* Join all the small entries into a single multi-index entry. */
2148         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2149         xas_store(&xas, hpage);
2150         WARN_ON_ONCE(xas_error(&xas));
2151         xas_unlock_irq(&xas);
2152
2153         /*
2154          * Remove pte page tables, so we can re-fault the page as huge.
2155          * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().
2156          */
2157         retract_page_tables(mapping, start);
2158         if (cc && !cc->is_khugepaged)
2159                 result = SCAN_PTE_MAPPED_HUGEPAGE;
2160         unlock_page(hpage);
2161
2162         /*
2163          * The collapse has succeeded, so free the old pages.
2164          */
2165         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2166                 list_del(&page->lru);
2167                 page->mapping = NULL;
2168                 ClearPageActive(page);
2169                 ClearPageUnevictable(page);
2170                 unlock_page(page);
2171                 folio_put_refs(page_folio(page), 3);
2172         }
2173
2174         goto out;
2175
2176 rollback:
2177         /* Something went wrong: roll back page cache changes */
2178         if (nr_none) {
2179                 xas_lock_irq(&xas);
2180                 mapping->nrpages -= nr_none;
2181                 xas_unlock_irq(&xas);
2182                 shmem_uncharge(mapping->host, nr_none);
2183         }
2184
2185         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2186                 list_del(&page->lru);
2187                 unlock_page(page);
2188                 putback_lru_page(page);
2189                 put_page(page);
2190         }
2191         /*
2192          * Undo the updates of filemap_nr_thps_inc for non-SHMEM
2193          * file only. This undo is not needed unless failure is
2194          * due to SCAN_COPY_MC.
2195          */
2196         if (!is_shmem && result == SCAN_COPY_MC) {
2197                 filemap_nr_thps_dec(mapping);
2198                 /*
2199                  * Paired with smp_mb() in do_dentry_open() to
2200                  * ensure the update to nr_thps is visible.
2201                  */
2202                 smp_mb();
2203         }
2204
2205         hpage->mapping = NULL;
2206
2207         unlock_page(hpage);
2208         put_page(hpage);
2209 out:
2210         VM_BUG_ON(!list_empty(&pagelist));
2211         trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2212         return result;
2213 }
2214
2215 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2216                                     struct file *file, pgoff_t start,
2217                                     struct collapse_control *cc)
2218 {
2219         struct page *page = NULL;
2220         struct address_space *mapping = file->f_mapping;
2221         XA_STATE(xas, &mapping->i_pages, start);
2222         int present, swap;
2223         int node = NUMA_NO_NODE;
2224         int result = SCAN_SUCCEED;
2225
2226         present = 0;
2227         swap = 0;
2228         memset(cc->node_load, 0, sizeof(cc->node_load));
2229         nodes_clear(cc->alloc_nmask);
2230         rcu_read_lock();
2231         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2232                 if (xas_retry(&xas, page))
2233                         continue;
2234
2235                 if (xa_is_value(page)) {
2236                         ++swap;
2237                         if (cc->is_khugepaged &&
2238                             swap > khugepaged_max_ptes_swap) {
2239                                 result = SCAN_EXCEED_SWAP_PTE;
2240                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2241                                 break;
2242                         }
2243                         continue;
2244                 }
2245
2246                 /*
2247                  * TODO: khugepaged should compact smaller compound pages
2248                  * into a PMD sized page
2249                  */
2250                 if (PageTransCompound(page)) {
2251                         struct page *head = compound_head(page);
2252
2253                         result = compound_order(head) == HPAGE_PMD_ORDER &&
2254                                         head->index == start
2255                                         /* Maybe PMD-mapped */
2256                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2257                                         : SCAN_PAGE_COMPOUND;
2258                         /*
2259                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2260                          * by the caller won't touch the page cache, and so
2261                          * it's safe to skip LRU and refcount checks before
2262                          * returning.
2263                          */
2264                         break;
2265                 }
2266
2267                 node = page_to_nid(page);
2268                 if (hpage_collapse_scan_abort(node, cc)) {
2269                         result = SCAN_SCAN_ABORT;
2270                         break;
2271                 }
2272                 cc->node_load[node]++;
2273
2274                 if (!PageLRU(page)) {
2275                         result = SCAN_PAGE_LRU;
2276                         break;
2277                 }
2278
2279                 if (page_count(page) !=
2280                     1 + page_mapcount(page) + page_has_private(page)) {
2281                         result = SCAN_PAGE_COUNT;
2282                         break;
2283                 }
2284
2285                 /*
2286                  * We probably should check if the page is referenced here, but
2287                  * nobody would transfer pte_young() to PageReferenced() for us.
2288                  * And rmap walk here is just too costly...
2289                  */
2290
2291                 present++;
2292
2293                 if (need_resched()) {
2294                         xas_pause(&xas);
2295                         cond_resched_rcu();
2296                 }
2297         }
2298         rcu_read_unlock();
2299
2300         if (result == SCAN_SUCCEED) {
2301                 if (cc->is_khugepaged &&
2302                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2303                         result = SCAN_EXCEED_NONE_PTE;
2304                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2305                 } else {
2306                         result = collapse_file(mm, addr, file, start, cc);
2307                 }
2308         }
2309
2310         trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2311         return result;
2312 }
2313 #else
2314 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2315                                     struct file *file, pgoff_t start,
2316                                     struct collapse_control *cc)
2317 {
2318         BUILD_BUG();
2319 }
2320 #endif
2321
2322 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2323                                             struct collapse_control *cc)
2324         __releases(&khugepaged_mm_lock)
2325         __acquires(&khugepaged_mm_lock)
2326 {
2327         struct vma_iterator vmi;
2328         struct khugepaged_mm_slot *mm_slot;
2329         struct mm_slot *slot;
2330         struct mm_struct *mm;
2331         struct vm_area_struct *vma;
2332         int progress = 0;
2333
2334         VM_BUG_ON(!pages);
2335         lockdep_assert_held(&khugepaged_mm_lock);
2336         *result = SCAN_FAIL;
2337
2338         if (khugepaged_scan.mm_slot) {
2339                 mm_slot = khugepaged_scan.mm_slot;
2340                 slot = &mm_slot->slot;
2341         } else {
2342                 slot = list_entry(khugepaged_scan.mm_head.next,
2343                                      struct mm_slot, mm_node);
2344                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2345                 khugepaged_scan.address = 0;
2346                 khugepaged_scan.mm_slot = mm_slot;
2347         }
2348         spin_unlock(&khugepaged_mm_lock);
2349
2350         mm = slot->mm;
2351         /*
2352          * Don't wait for semaphore (to avoid long wait times).  Just move to
2353          * the next mm on the list.
2354          */
2355         vma = NULL;
2356         if (unlikely(!mmap_read_trylock(mm)))
2357                 goto breakouterloop_mmap_lock;
2358
2359         progress++;
2360         if (unlikely(hpage_collapse_test_exit(mm)))
2361                 goto breakouterloop;
2362
2363         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2364         for_each_vma(vmi, vma) {
2365                 unsigned long hstart, hend;
2366
2367                 cond_resched();
2368                 if (unlikely(hpage_collapse_test_exit(mm))) {
2369                         progress++;
2370                         break;
2371                 }
2372                 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) {
2373 skip:
2374                         progress++;
2375                         continue;
2376                 }
2377                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2378                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2379                 if (khugepaged_scan.address > hend)
2380                         goto skip;
2381                 if (khugepaged_scan.address < hstart)
2382                         khugepaged_scan.address = hstart;
2383                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2384
2385                 while (khugepaged_scan.address < hend) {
2386                         bool mmap_locked = true;
2387
2388                         cond_resched();
2389                         if (unlikely(hpage_collapse_test_exit(mm)))
2390                                 goto breakouterloop;
2391
2392                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2393                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2394                                   hend);
2395                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2396                                 struct file *file = get_file(vma->vm_file);
2397                                 pgoff_t pgoff = linear_page_index(vma,
2398                                                 khugepaged_scan.address);
2399
2400                                 mmap_read_unlock(mm);
2401                                 mmap_locked = false;
2402                                 *result = hpage_collapse_scan_file(mm,
2403                                         khugepaged_scan.address, file, pgoff, cc);
2404                                 fput(file);
2405                                 if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {
2406                                         mmap_read_lock(mm);
2407                                         if (hpage_collapse_test_exit(mm))
2408                                                 goto breakouterloop;
2409                                         *result = collapse_pte_mapped_thp(mm,
2410                                                 khugepaged_scan.address, false);
2411                                         if (*result == SCAN_PMD_MAPPED)
2412                                                 *result = SCAN_SUCCEED;
2413                                         mmap_read_unlock(mm);
2414                                 }
2415                         } else {
2416                                 *result = hpage_collapse_scan_pmd(mm, vma,
2417                                         khugepaged_scan.address, &mmap_locked, cc);
2418                         }
2419
2420                         if (*result == SCAN_SUCCEED)
2421                                 ++khugepaged_pages_collapsed;
2422
2423                         /* move to next address */
2424                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2425                         progress += HPAGE_PMD_NR;
2426                         if (!mmap_locked)
2427                                 /*
2428                                  * We released mmap_lock so break loop.  Note
2429                                  * that we drop mmap_lock before all hugepage
2430                                  * allocations, so if allocation fails, we are
2431                                  * guaranteed to break here and report the
2432                                  * correct result back to caller.
2433                                  */
2434                                 goto breakouterloop_mmap_lock;
2435                         if (progress >= pages)
2436                                 goto breakouterloop;
2437                 }
2438         }
2439 breakouterloop:
2440         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2441 breakouterloop_mmap_lock:
2442
2443         spin_lock(&khugepaged_mm_lock);
2444         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2445         /*
2446          * Release the current mm_slot if this mm is about to die, or
2447          * if we scanned all vmas of this mm.
2448          */
2449         if (hpage_collapse_test_exit(mm) || !vma) {
2450                 /*
2451                  * Make sure that if mm_users is reaching zero while
2452                  * khugepaged runs here, khugepaged_exit will find
2453                  * mm_slot not pointing to the exiting mm.
2454                  */
2455                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2456                         slot = list_entry(slot->mm_node.next,
2457                                           struct mm_slot, mm_node);
2458                         khugepaged_scan.mm_slot =
2459                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2460                         khugepaged_scan.address = 0;
2461                 } else {
2462                         khugepaged_scan.mm_slot = NULL;
2463                         khugepaged_full_scans++;
2464                 }
2465
2466                 collect_mm_slot(mm_slot);
2467         }
2468
2469         return progress;
2470 }
2471
2472 static int khugepaged_has_work(void)
2473 {
2474         return !list_empty(&khugepaged_scan.mm_head) &&
2475                 hugepage_flags_enabled();
2476 }
2477
2478 static int khugepaged_wait_event(void)
2479 {
2480         return !list_empty(&khugepaged_scan.mm_head) ||
2481                 kthread_should_stop();
2482 }
2483
2484 static void khugepaged_do_scan(struct collapse_control *cc)
2485 {
2486         unsigned int progress = 0, pass_through_head = 0;
2487         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2488         bool wait = true;
2489         int result = SCAN_SUCCEED;
2490
2491         lru_add_drain_all();
2492
2493         while (true) {
2494                 cond_resched();
2495
2496                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2497                         break;
2498
2499                 spin_lock(&khugepaged_mm_lock);
2500                 if (!khugepaged_scan.mm_slot)
2501                         pass_through_head++;
2502                 if (khugepaged_has_work() &&
2503                     pass_through_head < 2)
2504                         progress += khugepaged_scan_mm_slot(pages - progress,
2505                                                             &result, cc);
2506                 else
2507                         progress = pages;
2508                 spin_unlock(&khugepaged_mm_lock);
2509
2510                 if (progress >= pages)
2511                         break;
2512
2513                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2514                         /*
2515                          * If fail to allocate the first time, try to sleep for
2516                          * a while.  When hit again, cancel the scan.
2517                          */
2518                         if (!wait)
2519                                 break;
2520                         wait = false;
2521                         khugepaged_alloc_sleep();
2522                 }
2523         }
2524 }
2525
2526 static bool khugepaged_should_wakeup(void)
2527 {
2528         return kthread_should_stop() ||
2529                time_after_eq(jiffies, khugepaged_sleep_expire);
2530 }
2531
2532 static void khugepaged_wait_work(void)
2533 {
2534         if (khugepaged_has_work()) {
2535                 const unsigned long scan_sleep_jiffies =
2536                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2537
2538                 if (!scan_sleep_jiffies)
2539                         return;
2540
2541                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2542                 wait_event_freezable_timeout(khugepaged_wait,
2543                                              khugepaged_should_wakeup(),
2544                                              scan_sleep_jiffies);
2545                 return;
2546         }
2547
2548         if (hugepage_flags_enabled())
2549                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2550 }
2551
2552 static int khugepaged(void *none)
2553 {
2554         struct khugepaged_mm_slot *mm_slot;
2555
2556         set_freezable();
2557         set_user_nice(current, MAX_NICE);
2558
2559         while (!kthread_should_stop()) {
2560                 khugepaged_do_scan(&khugepaged_collapse_control);
2561                 khugepaged_wait_work();
2562         }
2563
2564         spin_lock(&khugepaged_mm_lock);
2565         mm_slot = khugepaged_scan.mm_slot;
2566         khugepaged_scan.mm_slot = NULL;
2567         if (mm_slot)
2568                 collect_mm_slot(mm_slot);
2569         spin_unlock(&khugepaged_mm_lock);
2570         return 0;
2571 }
2572
2573 static void set_recommended_min_free_kbytes(void)
2574 {
2575         struct zone *zone;
2576         int nr_zones = 0;
2577         unsigned long recommended_min;
2578
2579         if (!hugepage_flags_enabled()) {
2580                 calculate_min_free_kbytes();
2581                 goto update_wmarks;
2582         }
2583
2584         for_each_populated_zone(zone) {
2585                 /*
2586                  * We don't need to worry about fragmentation of
2587                  * ZONE_MOVABLE since it only has movable pages.
2588                  */
2589                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2590                         continue;
2591
2592                 nr_zones++;
2593         }
2594
2595         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2596         recommended_min = pageblock_nr_pages * nr_zones * 2;
2597
2598         /*
2599          * Make sure that on average at least two pageblocks are almost free
2600          * of another type, one for a migratetype to fall back to and a
2601          * second to avoid subsequent fallbacks of other types There are 3
2602          * MIGRATE_TYPES we care about.
2603          */
2604         recommended_min += pageblock_nr_pages * nr_zones *
2605                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2606
2607         /* don't ever allow to reserve more than 5% of the lowmem */
2608         recommended_min = min(recommended_min,
2609                               (unsigned long) nr_free_buffer_pages() / 20);
2610         recommended_min <<= (PAGE_SHIFT-10);
2611
2612         if (recommended_min > min_free_kbytes) {
2613                 if (user_min_free_kbytes >= 0)
2614                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2615                                 min_free_kbytes, recommended_min);
2616
2617                 min_free_kbytes = recommended_min;
2618         }
2619
2620 update_wmarks:
2621         setup_per_zone_wmarks();
2622 }
2623
2624 int start_stop_khugepaged(void)
2625 {
2626         int err = 0;
2627
2628         mutex_lock(&khugepaged_mutex);
2629         if (hugepage_flags_enabled()) {
2630                 if (!khugepaged_thread)
2631                         khugepaged_thread = kthread_run(khugepaged, NULL,
2632                                                         "khugepaged");
2633                 if (IS_ERR(khugepaged_thread)) {
2634                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2635                         err = PTR_ERR(khugepaged_thread);
2636                         khugepaged_thread = NULL;
2637                         goto fail;
2638                 }
2639
2640                 if (!list_empty(&khugepaged_scan.mm_head))
2641                         wake_up_interruptible(&khugepaged_wait);
2642         } else if (khugepaged_thread) {
2643                 kthread_stop(khugepaged_thread);
2644                 khugepaged_thread = NULL;
2645         }
2646         set_recommended_min_free_kbytes();
2647 fail:
2648         mutex_unlock(&khugepaged_mutex);
2649         return err;
2650 }
2651
2652 void khugepaged_min_free_kbytes_update(void)
2653 {
2654         mutex_lock(&khugepaged_mutex);
2655         if (hugepage_flags_enabled() && khugepaged_thread)
2656                 set_recommended_min_free_kbytes();
2657         mutex_unlock(&khugepaged_mutex);
2658 }
2659
2660 bool current_is_khugepaged(void)
2661 {
2662         return kthread_func(current) == khugepaged;
2663 }
2664
2665 static int madvise_collapse_errno(enum scan_result r)
2666 {
2667         /*
2668          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2669          * actionable feedback to caller, so they may take an appropriate
2670          * fallback measure depending on the nature of the failure.
2671          */
2672         switch (r) {
2673         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2674                 return -ENOMEM;
2675         case SCAN_CGROUP_CHARGE_FAIL:
2676         case SCAN_EXCEED_NONE_PTE:
2677                 return -EBUSY;
2678         /* Resource temporary unavailable - trying again might succeed */
2679         case SCAN_PAGE_COUNT:
2680         case SCAN_PAGE_LOCK:
2681         case SCAN_PAGE_LRU:
2682         case SCAN_DEL_PAGE_LRU:
2683         case SCAN_PAGE_FILLED:
2684                 return -EAGAIN;
2685         /*
2686          * Other: Trying again likely not to succeed / error intrinsic to
2687          * specified memory range. khugepaged likely won't be able to collapse
2688          * either.
2689          */
2690         default:
2691                 return -EINVAL;
2692         }
2693 }
2694
2695 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2696                      unsigned long start, unsigned long end)
2697 {
2698         struct collapse_control *cc;
2699         struct mm_struct *mm = vma->vm_mm;
2700         unsigned long hstart, hend, addr;
2701         int thps = 0, last_fail = SCAN_FAIL;
2702         bool mmap_locked = true;
2703
2704         BUG_ON(vma->vm_start > start);
2705         BUG_ON(vma->vm_end < end);
2706
2707         *prev = vma;
2708
2709         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
2710                 return -EINVAL;
2711
2712         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2713         if (!cc)
2714                 return -ENOMEM;
2715         cc->is_khugepaged = false;
2716
2717         mmgrab(mm);
2718         lru_add_drain_all();
2719
2720         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2721         hend = end & HPAGE_PMD_MASK;
2722
2723         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2724                 int result = SCAN_FAIL;
2725
2726                 if (!mmap_locked) {
2727                         cond_resched();
2728                         mmap_read_lock(mm);
2729                         mmap_locked = true;
2730                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2731                                                          cc);
2732                         if (result  != SCAN_SUCCEED) {
2733                                 last_fail = result;
2734                                 goto out_nolock;
2735                         }
2736
2737                         hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2738                 }
2739                 mmap_assert_locked(mm);
2740                 memset(cc->node_load, 0, sizeof(cc->node_load));
2741                 nodes_clear(cc->alloc_nmask);
2742                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2743                         struct file *file = get_file(vma->vm_file);
2744                         pgoff_t pgoff = linear_page_index(vma, addr);
2745
2746                         mmap_read_unlock(mm);
2747                         mmap_locked = false;
2748                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2749                                                           cc);
2750                         fput(file);
2751                 } else {
2752                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2753                                                          &mmap_locked, cc);
2754                 }
2755                 if (!mmap_locked)
2756                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2757
2758 handle_result:
2759                 switch (result) {
2760                 case SCAN_SUCCEED:
2761                 case SCAN_PMD_MAPPED:
2762                         ++thps;
2763                         break;
2764                 case SCAN_PTE_MAPPED_HUGEPAGE:
2765                         BUG_ON(mmap_locked);
2766                         BUG_ON(*prev);
2767                         mmap_read_lock(mm);
2768                         result = collapse_pte_mapped_thp(mm, addr, true);
2769                         mmap_read_unlock(mm);
2770                         goto handle_result;
2771                 /* Whitelisted set of results where continuing OK */
2772                 case SCAN_PMD_NULL:
2773                 case SCAN_PTE_NON_PRESENT:
2774                 case SCAN_PTE_UFFD_WP:
2775                 case SCAN_PAGE_RO:
2776                 case SCAN_LACK_REFERENCED_PAGE:
2777                 case SCAN_PAGE_NULL:
2778                 case SCAN_PAGE_COUNT:
2779                 case SCAN_PAGE_LOCK:
2780                 case SCAN_PAGE_COMPOUND:
2781                 case SCAN_PAGE_LRU:
2782                 case SCAN_DEL_PAGE_LRU:
2783                         last_fail = result;
2784                         break;
2785                 default:
2786                         last_fail = result;
2787                         /* Other error, exit */
2788                         goto out_maybelock;
2789                 }
2790         }
2791
2792 out_maybelock:
2793         /* Caller expects us to hold mmap_lock on return */
2794         if (!mmap_locked)
2795                 mmap_read_lock(mm);
2796 out_nolock:
2797         mmap_assert_locked(mm);
2798         mmdrop(mm);
2799         kfree(cc);
2800
2801         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2802                         : madvise_collapse_errno(last_fail);
2803 }