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