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/rcupdate_wait.h>
  21 #include <linux/swapops.h>
  22 #include <linux/shmem_fs.h>
  23 #include <linux/ksm.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28 #include "mm_slot.h"
  29
  30 enum scan_result {
  31         SCAN_FAIL,
  32         SCAN_SUCCEED,
  33         SCAN_PMD_NULL,
  34         SCAN_PMD_NONE,
  35         SCAN_PMD_MAPPED,
  36         SCAN_EXCEED_NONE_PTE,
  37         SCAN_EXCEED_SWAP_PTE,
  38         SCAN_EXCEED_SHARED_PTE,
  39         SCAN_PTE_NON_PRESENT,
  40         SCAN_PTE_UFFD_WP,
  41         SCAN_PTE_MAPPED_HUGEPAGE,
  42         SCAN_PAGE_RO,
  43         SCAN_LACK_REFERENCED_PAGE,
  44         SCAN_PAGE_NULL,
  45         SCAN_SCAN_ABORT,
  46         SCAN_PAGE_COUNT,
  47         SCAN_PAGE_LRU,
  48         SCAN_PAGE_LOCK,
  49         SCAN_PAGE_ANON,
  50         SCAN_PAGE_COMPOUND,
  51         SCAN_ANY_PROCESS,
  52         SCAN_VMA_NULL,
  53         SCAN_VMA_CHECK,
  54         SCAN_ADDRESS_RANGE,
  55         SCAN_DEL_PAGE_LRU,
  56         SCAN_ALLOC_HUGE_PAGE_FAIL,
  57         SCAN_CGROUP_CHARGE_FAIL,
  58         SCAN_TRUNCATED,
  59         SCAN_PAGE_HAS_PRIVATE,
  60         SCAN_STORE_FAILED,
  61         SCAN_COPY_MC,
  62         SCAN_PAGE_FILLED,
  63 };
  64
  65 #define CREATE_TRACE_POINTS
  66 #include <trace/events/huge_memory.h>
  67
  68 static struct task_struct *khugepaged_thread __read_mostly;
  69 static DEFINE_MUTEX(khugepaged_mutex);
  70
  71 /* default scan 8*512 pte (or vmas) every 30 second */
  72 static unsigned int khugepaged_pages_to_scan __read_mostly;
  73 static unsigned int khugepaged_pages_collapsed;
  74 static unsigned int khugepaged_full_scans;
  75 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  76 /* during fragmentation poll the hugepage allocator once every minute */
  77 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  78 static unsigned long khugepaged_sleep_expire;
  79 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  80 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  81 /*
  82  * default collapse hugepages if there is at least one pte mapped like
  83  * it would have happened if the vma was large enough during page
  84  * fault.
  85  *
  86  * Note that these are only respected if collapse was initiated by khugepaged.
  87  */
  88 static unsigned int khugepaged_max_ptes_none __read_mostly;
  89 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  90 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  91
  92 #define MM_SLOTS_HASH_BITS 10
  93 static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  94
  95 static struct kmem_cache *mm_slot_cache __ro_after_init;
  96
  97 struct collapse_control {
  98         bool is_khugepaged;
  99
 100         /* Num pages scanned per node */
 101         u32 node_load[MAX_NUMNODES];
 102
 103         /* nodemask for allocation fallback */
 104         nodemask_t alloc_nmask;
 105 };
 106
 107 /**
 108  * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned
 109  * @slot: hash lookup from mm to mm_slot
 110  */
 111 struct khugepaged_mm_slot {
 112         struct mm_slot slot;
 113 };
 114
 115 /**
 116  * struct khugepaged_scan - cursor for scanning
 117  * @mm_head: the head of the mm list to scan
 118  * @mm_slot: the current mm_slot we are scanning
 119  * @address: the next address inside that to be scanned
 120  *
 121  * There is only the one khugepaged_scan instance of this cursor structure.
 122  */
 123 struct khugepaged_scan {
 124         struct list_head mm_head;
 125         struct khugepaged_mm_slot *mm_slot;
 126         unsigned long address;
 127 };
 128
 129 static struct khugepaged_scan khugepaged_scan = {
 130         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 131 };
 132
 133 #ifdef CONFIG_SYSFS
 134 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 135                                          struct kobj_attribute *attr,
 136                                          char *buf)
 137 {
 138         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 139 }
 140
 141 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 142                                           struct kobj_attribute *attr,
 143                                           const char *buf, size_t count)
 144 {
 145         unsigned int msecs;
 146         int err;
 147
 148         err = kstrtouint(buf, 10, &msecs);
 149         if (err)
 150                 return -EINVAL;
 151
 152         khugepaged_scan_sleep_millisecs = msecs;
 153         khugepaged_sleep_expire = 0;
 154         wake_up_interruptible(&khugepaged_wait);
 155
 156         return count;
 157 }
 158 static struct kobj_attribute scan_sleep_millisecs_attr =
 159         __ATTR_RW(scan_sleep_millisecs);
 160
 161 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 162                                           struct kobj_attribute *attr,
 163                                           char *buf)
 164 {
 165         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 166 }
 167
 168 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 169                                            struct kobj_attribute *attr,
 170                                            const char *buf, size_t count)
 171 {
 172         unsigned int msecs;
 173         int err;
 174
 175         err = kstrtouint(buf, 10, &msecs);
 176         if (err)
 177                 return -EINVAL;
 178
 179         khugepaged_alloc_sleep_millisecs = msecs;
 180         khugepaged_sleep_expire = 0;
 181         wake_up_interruptible(&khugepaged_wait);
 182
 183         return count;
 184 }
 185 static struct kobj_attribute alloc_sleep_millisecs_attr =
 186         __ATTR_RW(alloc_sleep_millisecs);
 187
 188 static ssize_t pages_to_scan_show(struct kobject *kobj,
 189                                   struct kobj_attribute *attr,
 190                                   char *buf)
 191 {
 192         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 193 }
 194 static ssize_t pages_to_scan_store(struct kobject *kobj,
 195                                    struct kobj_attribute *attr,
 196                                    const char *buf, size_t count)
 197 {
 198         unsigned int pages;
 199         int err;
 200
 201         err = kstrtouint(buf, 10, &pages);
 202         if (err || !pages)
 203                 return -EINVAL;
 204
 205         khugepaged_pages_to_scan = pages;
 206
 207         return count;
 208 }
 209 static struct kobj_attribute pages_to_scan_attr =
 210         __ATTR_RW(pages_to_scan);
 211
 212 static ssize_t pages_collapsed_show(struct kobject *kobj,
 213                                     struct kobj_attribute *attr,
 214                                     char *buf)
 215 {
 216         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 217 }
 218 static struct kobj_attribute pages_collapsed_attr =
 219         __ATTR_RO(pages_collapsed);
 220
 221 static ssize_t full_scans_show(struct kobject *kobj,
 222                                struct kobj_attribute *attr,
 223                                char *buf)
 224 {
 225         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 226 }
 227 static struct kobj_attribute full_scans_attr =
 228         __ATTR_RO(full_scans);
 229
 230 static ssize_t defrag_show(struct kobject *kobj,
 231                            struct kobj_attribute *attr, char *buf)
 232 {
 233         return single_hugepage_flag_show(kobj, attr, buf,
 234                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 235 }
 236 static ssize_t defrag_store(struct kobject *kobj,
 237                             struct kobj_attribute *attr,
 238                             const char *buf, size_t count)
 239 {
 240         return single_hugepage_flag_store(kobj, attr, buf, count,
 241                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 242 }
 243 static struct kobj_attribute khugepaged_defrag_attr =
 244         __ATTR_RW(defrag);
 245
 246 /*
 247  * max_ptes_none controls if khugepaged should collapse hugepages over
 248  * any unmapped ptes in turn potentially increasing the memory
 249  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 250  * reduce the available free memory in the system as it
 251  * runs. Increasing max_ptes_none will instead potentially reduce the
 252  * free memory in the system during the khugepaged scan.
 253  */
 254 static ssize_t max_ptes_none_show(struct kobject *kobj,
 255                                   struct kobj_attribute *attr,
 256                                   char *buf)
 257 {
 258         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 259 }
 260 static ssize_t max_ptes_none_store(struct kobject *kobj,
 261                                    struct kobj_attribute *attr,
 262                                    const char *buf, size_t count)
 263 {
 264         int err;
 265         unsigned long max_ptes_none;
 266
 267         err = kstrtoul(buf, 10, &max_ptes_none);
 268         if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 269                 return -EINVAL;
 270
 271         khugepaged_max_ptes_none = max_ptes_none;
 272
 273         return count;
 274 }
 275 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 276         __ATTR_RW(max_ptes_none);
 277
 278 static ssize_t max_ptes_swap_show(struct kobject *kobj,
 279                                   struct kobj_attribute *attr,
 280                                   char *buf)
 281 {
 282         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 283 }
 284
 285 static ssize_t max_ptes_swap_store(struct kobject *kobj,
 286                                    struct kobj_attribute *attr,
 287                                    const char *buf, size_t count)
 288 {
 289         int err;
 290         unsigned long max_ptes_swap;
 291
 292         err  = kstrtoul(buf, 10, &max_ptes_swap);
 293         if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 294                 return -EINVAL;
 295
 296         khugepaged_max_ptes_swap = max_ptes_swap;
 297
 298         return count;
 299 }
 300
 301 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 302         __ATTR_RW(max_ptes_swap);
 303
 304 static ssize_t max_ptes_shared_show(struct kobject *kobj,
 305                                     struct kobj_attribute *attr,
 306                                     char *buf)
 307 {
 308         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 309 }
 310
 311 static ssize_t max_ptes_shared_store(struct kobject *kobj,
 312                                      struct kobj_attribute *attr,
 313                                      const char *buf, size_t count)
 314 {
 315         int err;
 316         unsigned long max_ptes_shared;
 317
 318         err  = kstrtoul(buf, 10, &max_ptes_shared);
 319         if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 320                 return -EINVAL;
 321
 322         khugepaged_max_ptes_shared = max_ptes_shared;
 323
 324         return count;
 325 }
 326
 327 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 328         __ATTR_RW(max_ptes_shared);
 329
 330 static struct attribute *khugepaged_attr[] = {
 331         &khugepaged_defrag_attr.attr,
 332         &khugepaged_max_ptes_none_attr.attr,
 333         &khugepaged_max_ptes_swap_attr.attr,
 334         &khugepaged_max_ptes_shared_attr.attr,
 335         &pages_to_scan_attr.attr,
 336         &pages_collapsed_attr.attr,
 337         &full_scans_attr.attr,
 338         &scan_sleep_millisecs_attr.attr,
 339         &alloc_sleep_millisecs_attr.attr,
 340         NULL,
 341 };
 342
 343 struct attribute_group khugepaged_attr_group = {
 344         .attrs = khugepaged_attr,
 345         .name = "khugepaged",
 346 };
 347 #endif /* CONFIG_SYSFS */
 348
 349 int hugepage_madvise(struct vm_area_struct *vma,
 350                      unsigned long *vm_flags, int advice)
 351 {
 352         switch (advice) {
 353         case MADV_HUGEPAGE:
 354 #ifdef CONFIG_S390
 355                 /*
 356                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 357                  * can't handle this properly after s390_enable_sie, so we simply
 358                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 359                  */
 360                 if (mm_has_pgste(vma->vm_mm))
 361                         return 0;
 362 #endif
 363                 *vm_flags &= ~VM_NOHUGEPAGE;
 364                 *vm_flags |= VM_HUGEPAGE;
 365                 /*
 366                  * If the vma become good for khugepaged to scan,
 367                  * register it here without waiting a page fault that
 368                  * may not happen any time soon.
 369                  */
 370                 khugepaged_enter_vma(vma, *vm_flags);
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct khugepaged_mm_slot),
 390                                           __alignof__(struct khugepaged_mm_slot),
 391                                           0, NULL);
 392         if (!mm_slot_cache)
 393                 return -ENOMEM;
 394
 395         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 396         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 397         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 398         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 399
 400         return 0;
 401 }
 402
 403 void __init khugepaged_destroy(void)
 404 {
 405         kmem_cache_destroy(mm_slot_cache);
 406 }
 407
 408 static inline int hpage_collapse_test_exit(struct mm_struct *mm)
 409 {
 410         return atomic_read(&mm->mm_users) == 0;
 411 }
 412
 413 static inline int hpage_collapse_test_exit_or_disable(struct mm_struct *mm)
 414 {
 415         return hpage_collapse_test_exit(mm) ||
 416                test_bit(MMF_DISABLE_THP, &mm->flags);
 417 }
 418
 419 void __khugepaged_enter(struct mm_struct *mm)
 420 {
 421         struct khugepaged_mm_slot *mm_slot;
 422         struct mm_slot *slot;
 423         int wakeup;
 424
 425         /* __khugepaged_exit() must not run from under us */
 426         VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 427         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags)))
 428                 return;
 429
 430         mm_slot = mm_slot_alloc(mm_slot_cache);
 431         if (!mm_slot)
 432                 return;
 433
 434         slot = &mm_slot->slot;
 435
 436         spin_lock(&khugepaged_mm_lock);
 437         mm_slot_insert(mm_slots_hash, mm, slot);
 438         /*
 439          * Insert just behind the scanning cursor, to let the area settle
 440          * down a little.
 441          */
 442         wakeup = list_empty(&khugepaged_scan.mm_head);
 443         list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 444         spin_unlock(&khugepaged_mm_lock);
 445
 446         mmgrab(mm);
 447         if (wakeup)
 448                 wake_up_interruptible(&khugepaged_wait);
 449 }
 450
 451 void khugepaged_enter_vma(struct vm_area_struct *vma,
 452                           unsigned long vm_flags)
 453 {
 454         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 455             hugepage_flags_enabled()) {
 456                 if (thp_vma_allowable_order(vma, vm_flags, false, false, true,
 457                                             PMD_ORDER))
 458                         __khugepaged_enter(vma->vm_mm);
 459         }
 460 }
 461
 462 void __khugepaged_exit(struct mm_struct *mm)
 463 {
 464         struct khugepaged_mm_slot *mm_slot;
 465         struct mm_slot *slot;
 466         int free = 0;
 467
 468         spin_lock(&khugepaged_mm_lock);
 469         slot = mm_slot_lookup(mm_slots_hash, mm);
 470         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
 471         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 472                 hash_del(&slot->hash);
 473                 list_del(&slot->mm_node);
 474                 free = 1;
 475         }
 476         spin_unlock(&khugepaged_mm_lock);
 477
 478         if (free) {
 479                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 480                 mm_slot_free(mm_slot_cache, mm_slot);
 481                 mmdrop(mm);
 482         } else if (mm_slot) {
 483                 /*
 484                  * This is required to serialize against
 485                  * hpage_collapse_test_exit() (which is guaranteed to run
 486                  * under mmap sem read mode). Stop here (after we return all
 487                  * pagetables will be destroyed) until khugepaged has finished
 488                  * working on the pagetables under the mmap_lock.
 489                  */
 490                 mmap_write_lock(mm);
 491                 mmap_write_unlock(mm);
 492         }
 493 }
 494
 495 static void release_pte_folio(struct folio *folio)
 496 {
 497         node_stat_mod_folio(folio,
 498                         NR_ISOLATED_ANON + folio_is_file_lru(folio),
 499                         -folio_nr_pages(folio));
 500         folio_unlock(folio);
 501         folio_putback_lru(folio);
 502 }
 503
 504 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 505                 struct list_head *compound_pagelist)
 506 {
 507         struct folio *folio, *tmp;
 508
 509         while (--_pte >= pte) {
 510                 pte_t pteval = ptep_get(_pte);
 511                 unsigned long pfn;
 512
 513                 if (pte_none(pteval))
 514                         continue;
 515                 pfn = pte_pfn(pteval);
 516                 if (is_zero_pfn(pfn))
 517                         continue;
 518                 folio = pfn_folio(pfn);
 519                 if (folio_test_large(folio))
 520                         continue;
 521                 release_pte_folio(folio);
 522         }
 523
 524         list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 525                 list_del(&folio->lru);
 526                 release_pte_folio(folio);
 527         }
 528 }
 529
 530 static bool is_refcount_suitable(struct folio *folio)
 531 {
 532         int expected_refcount;
 533
 534         expected_refcount = folio_mapcount(folio);
 535         if (folio_test_swapcache(folio))
 536                 expected_refcount += folio_nr_pages(folio);
 537
 538         return folio_ref_count(folio) == expected_refcount;
 539 }
 540
 541 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 542                                         unsigned long address,
 543                                         pte_t *pte,
 544                                         struct collapse_control *cc,
 545                                         struct list_head *compound_pagelist)
 546 {
 547         struct page *page = NULL;
 548         struct folio *folio = NULL;
 549         pte_t *_pte;
 550         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 551         bool writable = false;
 552
 553         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 554              _pte++, address += PAGE_SIZE) {
 555                 pte_t pteval = ptep_get(_pte);
 556                 if (pte_none(pteval) || (pte_present(pteval) &&
 557                                 is_zero_pfn(pte_pfn(pteval)))) {
 558                         ++none_or_zero;
 559                         if (!userfaultfd_armed(vma) &&
 560                             (!cc->is_khugepaged ||
 561                              none_or_zero <= khugepaged_max_ptes_none)) {
 562                                 continue;
 563                         } else {
 564                                 result = SCAN_EXCEED_NONE_PTE;
 565                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 566                                 goto out;
 567                         }
 568                 }
 569                 if (!pte_present(pteval)) {
 570                         result = SCAN_PTE_NON_PRESENT;
 571                         goto out;
 572                 }
 573                 if (pte_uffd_wp(pteval)) {
 574                         result = SCAN_PTE_UFFD_WP;
 575                         goto out;
 576                 }
 577                 page = vm_normal_page(vma, address, pteval);
 578                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 579                         result = SCAN_PAGE_NULL;
 580                         goto out;
 581                 }
 582
 583                 folio = page_folio(page);
 584                 VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio);
 585
 586                 if (page_mapcount(page) > 1) {
 587                         ++shared;
 588                         if (cc->is_khugepaged &&
 589                             shared > khugepaged_max_ptes_shared) {
 590                                 result = SCAN_EXCEED_SHARED_PTE;
 591                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 592                                 goto out;
 593                         }
 594                 }
 595
 596                 if (folio_test_large(folio)) {
 597                         struct folio *f;
 598
 599                         /*
 600                          * Check if we have dealt with the compound page
 601                          * already
 602                          */
 603                         list_for_each_entry(f, compound_pagelist, lru) {
 604                                 if (folio == f)
 605                                         goto next;
 606                         }
 607                 }
 608
 609                 /*
 610                  * We can do it before isolate_lru_page because the
 611                  * page can't be freed from under us. NOTE: PG_lock
 612                  * is needed to serialize against split_huge_page
 613                  * when invoked from the VM.
 614                  */
 615                 if (!folio_trylock(folio)) {
 616                         result = SCAN_PAGE_LOCK;
 617                         goto out;
 618                 }
 619
 620                 /*
 621                  * Check if the page has any GUP (or other external) pins.
 622                  *
 623                  * The page table that maps the page has been already unlinked
 624                  * from the page table tree and this process cannot get
 625                  * an additional pin on the page.
 626                  *
 627                  * New pins can come later if the page is shared across fork,
 628                  * but not from this process. The other process cannot write to
 629                  * the page, only trigger CoW.
 630                  */
 631                 if (!is_refcount_suitable(folio)) {
 632                         folio_unlock(folio);
 633                         result = SCAN_PAGE_COUNT;
 634                         goto out;
 635                 }
 636
 637                 /*
 638                  * Isolate the page to avoid collapsing an hugepage
 639                  * currently in use by the VM.
 640                  */
 641                 if (!folio_isolate_lru(folio)) {
 642                         folio_unlock(folio);
 643                         result = SCAN_DEL_PAGE_LRU;
 644                         goto out;
 645                 }
 646                 node_stat_mod_folio(folio,
 647                                 NR_ISOLATED_ANON + folio_is_file_lru(folio),
 648                                 folio_nr_pages(folio));
 649                 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 650                 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 651
 652                 if (folio_test_large(folio))
 653                         list_add_tail(&folio->lru, compound_pagelist);
 654 next:
 655                 /*
 656                  * If collapse was initiated by khugepaged, check that there is
 657                  * enough young pte to justify collapsing the page
 658                  */
 659                 if (cc->is_khugepaged &&
 660                     (pte_young(pteval) || folio_test_young(folio) ||
 661                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
 662                                                                      address)))
 663                         referenced++;
 664
 665                 if (pte_write(pteval))
 666                         writable = true;
 667         }
 668
 669         if (unlikely(!writable)) {
 670                 result = SCAN_PAGE_RO;
 671         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 672                 result = SCAN_LACK_REFERENCED_PAGE;
 673         } else {
 674                 result = SCAN_SUCCEED;
 675                 trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 676                                                     referenced, writable, result);
 677                 return result;
 678         }
 679 out:
 680         release_pte_pages(pte, _pte, compound_pagelist);
 681         trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 682                                             referenced, writable, result);
 683         return result;
 684 }
 685
 686 static void __collapse_huge_page_copy_succeeded(pte_t *pte,
 687                                                 struct vm_area_struct *vma,
 688                                                 unsigned long address,
 689                                                 spinlock_t *ptl,
 690                                                 struct list_head *compound_pagelist)
 691 {
 692         struct folio *src, *tmp;
 693         pte_t *_pte;
 694         pte_t pteval;
 695
 696         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 697              _pte++, address += PAGE_SIZE) {
 698                 pteval = ptep_get(_pte);
 699                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 700                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 701                         if (is_zero_pfn(pte_pfn(pteval))) {
 702                                 /*
 703                                  * ptl mostly unnecessary.
 704                                  */
 705                                 spin_lock(ptl);
 706                                 ptep_clear(vma->vm_mm, address, _pte);
 707                                 spin_unlock(ptl);
 708                                 ksm_might_unmap_zero_page(vma->vm_mm, pteval);
 709                         }
 710                 } else {
 711                         struct page *src_page = pte_page(pteval);
 712
 713                         src = page_folio(src_page);
 714                         if (!folio_test_large(src))
 715                                 release_pte_folio(src);
 716                         /*
 717                          * ptl mostly unnecessary, but preempt has to
 718                          * be disabled to update the per-cpu stats
 719                          * inside folio_remove_rmap_pte().
 720                          */
 721                         spin_lock(ptl);
 722                         ptep_clear(vma->vm_mm, address, _pte);
 723                         folio_remove_rmap_pte(src, src_page, vma);
 724                         spin_unlock(ptl);
 725                         free_page_and_swap_cache(src_page);
 726                 }
 727         }
 728
 729         list_for_each_entry_safe(src, tmp, compound_pagelist, lru) {
 730                 list_del(&src->lru);
 731                 node_stat_sub_folio(src, NR_ISOLATED_ANON +
 732                                 folio_is_file_lru(src));
 733                 folio_unlock(src);
 734                 free_swap_cache(src);
 735                 folio_putback_lru(src);
 736         }
 737 }
 738
 739 static void __collapse_huge_page_copy_failed(pte_t *pte,
 740                                              pmd_t *pmd,
 741                                              pmd_t orig_pmd,
 742                                              struct vm_area_struct *vma,
 743                                              struct list_head *compound_pagelist)
 744 {
 745         spinlock_t *pmd_ptl;
 746
 747         /*
 748          * Re-establish the PMD to point to the original page table
 749          * entry. Restoring PMD needs to be done prior to releasing
 750          * pages. Since pages are still isolated and locked here,
 751          * acquiring anon_vma_lock_write is unnecessary.
 752          */
 753         pmd_ptl = pmd_lock(vma->vm_mm, pmd);
 754         pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));
 755         spin_unlock(pmd_ptl);
 756         /*
 757          * Release both raw and compound pages isolated
 758          * in __collapse_huge_page_isolate.
 759          */
 760         release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);
 761 }
 762
 763 /*
 764  * __collapse_huge_page_copy - attempts to copy memory contents from raw
 765  * pages to a hugepage. Cleans up the raw pages if copying succeeds;
 766  * otherwise restores the original page table and releases isolated raw pages.
 767  * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.
 768  *
 769  * @pte: starting of the PTEs to copy from
 770  * @page: the new hugepage to copy contents to
 771  * @pmd: pointer to the new hugepage's PMD
 772  * @orig_pmd: the original raw pages' PMD
 773  * @vma: the original raw pages' virtual memory area
 774  * @address: starting address to copy
 775  * @ptl: lock on raw pages' PTEs
 776  * @compound_pagelist: list that stores compound pages
 777  */
 778 static int __collapse_huge_page_copy(pte_t *pte,
 779                                      struct page *page,
 780                                      pmd_t *pmd,
 781                                      pmd_t orig_pmd,
 782                                      struct vm_area_struct *vma,
 783                                      unsigned long address,
 784                                      spinlock_t *ptl,
 785                                      struct list_head *compound_pagelist)
 786 {
 787         struct page *src_page;
 788         pte_t *_pte;
 789         pte_t pteval;
 790         unsigned long _address;
 791         int result = SCAN_SUCCEED;
 792
 793         /*
 794          * Copying pages' contents is subject to memory poison at any iteration.
 795          */
 796         for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR;
 797              _pte++, page++, _address += PAGE_SIZE) {
 798                 pteval = ptep_get(_pte);
 799                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 800                         clear_user_highpage(page, _address);
 801                         continue;
 802                 }
 803                 src_page = pte_page(pteval);
 804                 if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) {
 805                         result = SCAN_COPY_MC;
 806                         break;
 807                 }
 808         }
 809
 810         if (likely(result == SCAN_SUCCEED))
 811                 __collapse_huge_page_copy_succeeded(pte, vma, address, ptl,
 812                                                     compound_pagelist);
 813         else
 814                 __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,
 815                                                  compound_pagelist);
 816
 817         return result;
 818 }
 819
 820 static void khugepaged_alloc_sleep(void)
 821 {
 822         DEFINE_WAIT(wait);
 823
 824         add_wait_queue(&khugepaged_wait, &wait);
 825         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 826         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 827         remove_wait_queue(&khugepaged_wait, &wait);
 828 }
 829
 830 struct collapse_control khugepaged_collapse_control = {
 831         .is_khugepaged = true,
 832 };
 833
 834 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 835 {
 836         int i;
 837
 838         /*
 839          * If node_reclaim_mode is disabled, then no extra effort is made to
 840          * allocate memory locally.
 841          */
 842         if (!node_reclaim_enabled())
 843                 return false;
 844
 845         /* If there is a count for this node already, it must be acceptable */
 846         if (cc->node_load[nid])
 847                 return false;
 848
 849         for (i = 0; i < MAX_NUMNODES; i++) {
 850                 if (!cc->node_load[i])
 851                         continue;
 852                 if (node_distance(nid, i) > node_reclaim_distance)
 853                         return true;
 854         }
 855         return false;
 856 }
 857
 858 #define khugepaged_defrag()                                     \
 859         (transparent_hugepage_flags &                           \
 860          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 861
 862 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 863 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 864 {
 865         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 866 }
 867
 868 #ifdef CONFIG_NUMA
 869 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 870 {
 871         int nid, target_node = 0, max_value = 0;
 872
 873         /* find first node with max normal pages hit */
 874         for (nid = 0; nid < MAX_NUMNODES; nid++)
 875                 if (cc->node_load[nid] > max_value) {
 876                         max_value = cc->node_load[nid];
 877                         target_node = nid;
 878                 }
 879
 880         for_each_online_node(nid) {
 881                 if (max_value == cc->node_load[nid])
 882                         node_set(nid, cc->alloc_nmask);
 883         }
 884
 885         return target_node;
 886 }
 887 #else
 888 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 889 {
 890         return 0;
 891 }
 892 #endif
 893
 894 static bool hpage_collapse_alloc_folio(struct folio **folio, gfp_t gfp, int node,
 895                                       nodemask_t *nmask)
 896 {
 897         *folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, nmask);
 898
 899         if (unlikely(!*folio)) {
 900                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 901                 return false;
 902         }
 903
 904         count_vm_event(THP_COLLAPSE_ALLOC);
 905         return true;
 906 }
 907
 908 /*
 909  * If mmap_lock temporarily dropped, revalidate vma
 910  * before taking mmap_lock.
 911  * Returns enum scan_result value.
 912  */
 913
 914 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 915                                    bool expect_anon,
 916                                    struct vm_area_struct **vmap,
 917                                    struct collapse_control *cc)
 918 {
 919         struct vm_area_struct *vma;
 920
 921         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
 922                 return SCAN_ANY_PROCESS;
 923
 924         *vmap = vma = find_vma(mm, address);
 925         if (!vma)
 926                 return SCAN_VMA_NULL;
 927
 928         if (!thp_vma_suitable_order(vma, address, PMD_ORDER))
 929                 return SCAN_ADDRESS_RANGE;
 930         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false,
 931                                      cc->is_khugepaged, PMD_ORDER))
 932                 return SCAN_VMA_CHECK;
 933         /*
 934          * Anon VMA expected, the address may be unmapped then
 935          * remapped to file after khugepaged reaquired the mmap_lock.
 936          *
 937          * thp_vma_allowable_order may return true for qualified file
 938          * vmas.
 939          */
 940         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 941                 return SCAN_PAGE_ANON;
 942         return SCAN_SUCCEED;
 943 }
 944
 945 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 946                                    unsigned long address,
 947                                    pmd_t **pmd)
 948 {
 949         pmd_t pmde;
 950
 951         *pmd = mm_find_pmd(mm, address);
 952         if (!*pmd)
 953                 return SCAN_PMD_NULL;
 954
 955         pmde = pmdp_get_lockless(*pmd);
 956         if (pmd_none(pmde))
 957                 return SCAN_PMD_NONE;
 958         if (!pmd_present(pmde))
 959                 return SCAN_PMD_NULL;
 960         if (pmd_trans_huge(pmde))
 961                 return SCAN_PMD_MAPPED;
 962         if (pmd_devmap(pmde))
 963                 return SCAN_PMD_NULL;
 964         if (pmd_bad(pmde))
 965                 return SCAN_PMD_NULL;
 966         return SCAN_SUCCEED;
 967 }
 968
 969 static int check_pmd_still_valid(struct mm_struct *mm,
 970                                  unsigned long address,
 971                                  pmd_t *pmd)
 972 {
 973         pmd_t *new_pmd;
 974         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 975
 976         if (result != SCAN_SUCCEED)
 977                 return result;
 978         if (new_pmd != pmd)
 979                 return SCAN_FAIL;
 980         return SCAN_SUCCEED;
 981 }
 982
 983 /*
 984  * Bring missing pages in from swap, to complete THP collapse.
 985  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 986  *
 987  * Called and returns without pte mapped or spinlocks held.
 988  * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.
 989  */
 990 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 991                                        struct vm_area_struct *vma,
 992                                        unsigned long haddr, pmd_t *pmd,
 993                                        int referenced)
 994 {
 995         int swapped_in = 0;
 996         vm_fault_t ret = 0;
 997         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 998         int result;
 999         pte_t *pte = NULL;
1000         spinlock_t *ptl;
1001
1002         for (address = haddr; address < end; address += PAGE_SIZE) {
1003                 struct vm_fault vmf = {
1004                         .vma = vma,
1005                         .address = address,
1006                         .pgoff = linear_page_index(vma, address),
1007                         .flags = FAULT_FLAG_ALLOW_RETRY,
1008                         .pmd = pmd,
1009                 };
1010
1011                 if (!pte++) {
1012                         pte = pte_offset_map_nolock(mm, pmd, address, &ptl);
1013                         if (!pte) {
1014                                 mmap_read_unlock(mm);
1015                                 result = SCAN_PMD_NULL;
1016                                 goto out;
1017                         }
1018                 }
1019
1020                 vmf.orig_pte = ptep_get_lockless(pte);
1021                 if (!is_swap_pte(vmf.orig_pte))
1022                         continue;
1023
1024                 vmf.pte = pte;
1025                 vmf.ptl = ptl;
1026                 ret = do_swap_page(&vmf);
1027                 /* Which unmaps pte (after perhaps re-checking the entry) */
1028                 pte = NULL;
1029
1030                 /*
1031                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1032                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1033                  * we do not retry here and swap entry will remain in pagetable
1034                  * resulting in later failure.
1035                  */
1036                 if (ret & VM_FAULT_RETRY) {
1037                         /* Likely, but not guaranteed, that page lock failed */
1038                         result = SCAN_PAGE_LOCK;
1039                         goto out;
1040                 }
1041                 if (ret & VM_FAULT_ERROR) {
1042                         mmap_read_unlock(mm);
1043                         result = SCAN_FAIL;
1044                         goto out;
1045                 }
1046                 swapped_in++;
1047         }
1048
1049         if (pte)
1050                 pte_unmap(pte);
1051
1052         /* Drain LRU cache to remove extra pin on the swapped in pages */
1053         if (swapped_in)
1054                 lru_add_drain();
1055
1056         result = SCAN_SUCCEED;
1057 out:
1058         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);
1059         return result;
1060 }
1061
1062 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
1063                               struct collapse_control *cc)
1064 {
1065         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
1066                      GFP_TRANSHUGE);
1067         int node = hpage_collapse_find_target_node(cc);
1068         struct folio *folio;
1069
1070         if (!hpage_collapse_alloc_folio(&folio, gfp, node, &cc->alloc_nmask)) {
1071                 *hpage = NULL;
1072                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
1073         }
1074
1075         if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {
1076                 folio_put(folio);
1077                 *hpage = NULL;
1078                 return SCAN_CGROUP_CHARGE_FAIL;
1079         }
1080
1081         count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1);
1082
1083         *hpage = folio_page(folio, 0);
1084         return SCAN_SUCCEED;
1085 }
1086
1087 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
1088                               int referenced, int unmapped,
1089                               struct collapse_control *cc)
1090 {
1091         LIST_HEAD(compound_pagelist);
1092         pmd_t *pmd, _pmd;
1093         pte_t *pte;
1094         pgtable_t pgtable;
1095         struct folio *folio;
1096         struct page *hpage;
1097         spinlock_t *pmd_ptl, *pte_ptl;
1098         int result = SCAN_FAIL;
1099         struct vm_area_struct *vma;
1100         struct mmu_notifier_range range;
1101
1102         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1103
1104         /*
1105          * Before allocating the hugepage, release the mmap_lock read lock.
1106          * The allocation can take potentially a long time if it involves
1107          * sync compaction, and we do not need to hold the mmap_lock during
1108          * that. We will recheck the vma after taking it again in write mode.
1109          */
1110         mmap_read_unlock(mm);
1111
1112         result = alloc_charge_hpage(&hpage, mm, cc);
1113         if (result != SCAN_SUCCEED)
1114                 goto out_nolock;
1115
1116         mmap_read_lock(mm);
1117         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1118         if (result != SCAN_SUCCEED) {
1119                 mmap_read_unlock(mm);
1120                 goto out_nolock;
1121         }
1122
1123         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1124         if (result != SCAN_SUCCEED) {
1125                 mmap_read_unlock(mm);
1126                 goto out_nolock;
1127         }
1128
1129         if (unmapped) {
1130                 /*
1131                  * __collapse_huge_page_swapin will return with mmap_lock
1132                  * released when it fails. So we jump out_nolock directly in
1133                  * that case.  Continuing to collapse causes inconsistency.
1134                  */
1135                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1136                                                      referenced);
1137                 if (result != SCAN_SUCCEED)
1138                         goto out_nolock;
1139         }
1140
1141         mmap_read_unlock(mm);
1142         /*
1143          * Prevent all access to pagetables with the exception of
1144          * gup_fast later handled by the ptep_clear_flush and the VM
1145          * handled by the anon_vma lock + PG_lock.
1146          *
1147          * UFFDIO_MOVE is prevented to race as well thanks to the
1148          * mmap_lock.
1149          */
1150         mmap_write_lock(mm);
1151         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1152         if (result != SCAN_SUCCEED)
1153                 goto out_up_write;
1154         /* check if the pmd is still valid */
1155         result = check_pmd_still_valid(mm, address, pmd);
1156         if (result != SCAN_SUCCEED)
1157                 goto out_up_write;
1158
1159         vma_start_write(vma);
1160         anon_vma_lock_write(vma->anon_vma);
1161
1162         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1163                                 address + HPAGE_PMD_SIZE);
1164         mmu_notifier_invalidate_range_start(&range);
1165
1166         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1167         /*
1168          * This removes any huge TLB entry from the CPU so we won't allow
1169          * huge and small TLB entries for the same virtual address to
1170          * avoid the risk of CPU bugs in that area.
1171          *
1172          * Parallel fast GUP is fine since fast GUP will back off when
1173          * it detects PMD is changed.
1174          */
1175         _pmd = pmdp_collapse_flush(vma, address, pmd);
1176         spin_unlock(pmd_ptl);
1177         mmu_notifier_invalidate_range_end(&range);
1178         tlb_remove_table_sync_one();
1179
1180         pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
1181         if (pte) {
1182                 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1183                                                       &compound_pagelist);
1184                 spin_unlock(pte_ptl);
1185         } else {
1186                 result = SCAN_PMD_NULL;
1187         }
1188
1189         if (unlikely(result != SCAN_SUCCEED)) {
1190                 if (pte)
1191                         pte_unmap(pte);
1192                 spin_lock(pmd_ptl);
1193                 BUG_ON(!pmd_none(*pmd));
1194                 /*
1195                  * We can only use set_pmd_at when establishing
1196                  * hugepmds and never for establishing regular pmds that
1197                  * points to regular pagetables. Use pmd_populate for that
1198                  */
1199                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1200                 spin_unlock(pmd_ptl);
1201                 anon_vma_unlock_write(vma->anon_vma);
1202                 goto out_up_write;
1203         }
1204
1205         /*
1206          * All pages are isolated and locked so anon_vma rmap
1207          * can't run anymore.
1208          */
1209         anon_vma_unlock_write(vma->anon_vma);
1210
1211         result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd,
1212                                            vma, address, pte_ptl,
1213                                            &compound_pagelist);
1214         pte_unmap(pte);
1215         if (unlikely(result != SCAN_SUCCEED))
1216                 goto out_up_write;
1217
1218         folio = page_folio(hpage);
1219         /*
1220          * The smp_wmb() inside __folio_mark_uptodate() ensures the
1221          * copy_huge_page writes become visible before the set_pmd_at()
1222          * write.
1223          */
1224         __folio_mark_uptodate(folio);
1225         pgtable = pmd_pgtable(_pmd);
1226
1227         _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1228         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1229
1230         spin_lock(pmd_ptl);
1231         BUG_ON(!pmd_none(*pmd));
1232         folio_add_new_anon_rmap(folio, vma, address);
1233         folio_add_lru_vma(folio, vma);
1234         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1235         set_pmd_at(mm, address, pmd, _pmd);
1236         update_mmu_cache_pmd(vma, address, pmd);
1237         spin_unlock(pmd_ptl);
1238
1239         hpage = NULL;
1240
1241         result = SCAN_SUCCEED;
1242 out_up_write:
1243         mmap_write_unlock(mm);
1244 out_nolock:
1245         if (hpage)
1246                 put_page(hpage);
1247         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1248         return result;
1249 }
1250
1251 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1252                                    struct vm_area_struct *vma,
1253                                    unsigned long address, bool *mmap_locked,
1254                                    struct collapse_control *cc)
1255 {
1256         pmd_t *pmd;
1257         pte_t *pte, *_pte;
1258         int result = SCAN_FAIL, referenced = 0;
1259         int none_or_zero = 0, shared = 0;
1260         struct page *page = NULL;
1261         struct folio *folio = NULL;
1262         unsigned long _address;
1263         spinlock_t *ptl;
1264         int node = NUMA_NO_NODE, unmapped = 0;
1265         bool writable = false;
1266
1267         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1268
1269         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1270         if (result != SCAN_SUCCEED)
1271                 goto out;
1272
1273         memset(cc->node_load, 0, sizeof(cc->node_load));
1274         nodes_clear(cc->alloc_nmask);
1275         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1276         if (!pte) {
1277                 result = SCAN_PMD_NULL;
1278                 goto out;
1279         }
1280
1281         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1282              _pte++, _address += PAGE_SIZE) {
1283                 pte_t pteval = ptep_get(_pte);
1284                 if (is_swap_pte(pteval)) {
1285                         ++unmapped;
1286                         if (!cc->is_khugepaged ||
1287                             unmapped <= khugepaged_max_ptes_swap) {
1288                                 /*
1289                                  * Always be strict with uffd-wp
1290                                  * enabled swap entries.  Please see
1291                                  * comment below for pte_uffd_wp().
1292                                  */
1293                                 if (pte_swp_uffd_wp_any(pteval)) {
1294                                         result = SCAN_PTE_UFFD_WP;
1295                                         goto out_unmap;
1296                                 }
1297                                 continue;
1298                         } else {
1299                                 result = SCAN_EXCEED_SWAP_PTE;
1300                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1301                                 goto out_unmap;
1302                         }
1303                 }
1304                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1305                         ++none_or_zero;
1306                         if (!userfaultfd_armed(vma) &&
1307                             (!cc->is_khugepaged ||
1308                              none_or_zero <= khugepaged_max_ptes_none)) {
1309                                 continue;
1310                         } else {
1311                                 result = SCAN_EXCEED_NONE_PTE;
1312                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1313                                 goto out_unmap;
1314                         }
1315                 }
1316                 if (pte_uffd_wp(pteval)) {
1317                         /*
1318                          * Don't collapse the page if any of the small
1319                          * PTEs are armed with uffd write protection.
1320                          * Here we can also mark the new huge pmd as
1321                          * write protected if any of the small ones is
1322                          * marked but that could bring unknown
1323                          * userfault messages that falls outside of
1324                          * the registered range.  So, just be simple.
1325                          */
1326                         result = SCAN_PTE_UFFD_WP;
1327                         goto out_unmap;
1328                 }
1329                 if (pte_write(pteval))
1330                         writable = true;
1331
1332                 page = vm_normal_page(vma, _address, pteval);
1333                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1334                         result = SCAN_PAGE_NULL;
1335                         goto out_unmap;
1336                 }
1337
1338                 if (page_mapcount(page) > 1) {
1339                         ++shared;
1340                         if (cc->is_khugepaged &&
1341                             shared > khugepaged_max_ptes_shared) {
1342                                 result = SCAN_EXCEED_SHARED_PTE;
1343                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1344                                 goto out_unmap;
1345                         }
1346                 }
1347
1348                 folio = page_folio(page);
1349                 /*
1350                  * Record which node the original page is from and save this
1351                  * information to cc->node_load[].
1352                  * Khugepaged will allocate hugepage from the node has the max
1353                  * hit record.
1354                  */
1355                 node = folio_nid(folio);
1356                 if (hpage_collapse_scan_abort(node, cc)) {
1357                         result = SCAN_SCAN_ABORT;
1358                         goto out_unmap;
1359                 }
1360                 cc->node_load[node]++;
1361                 if (!folio_test_lru(folio)) {
1362                         result = SCAN_PAGE_LRU;
1363                         goto out_unmap;
1364                 }
1365                 if (folio_test_locked(folio)) {
1366                         result = SCAN_PAGE_LOCK;
1367                         goto out_unmap;
1368                 }
1369                 if (!folio_test_anon(folio)) {
1370                         result = SCAN_PAGE_ANON;
1371                         goto out_unmap;
1372                 }
1373
1374                 /*
1375                  * Check if the page has any GUP (or other external) pins.
1376                  *
1377                  * Here the check may be racy:
1378                  * it may see total_mapcount > refcount in some cases?
1379                  * But such case is ephemeral we could always retry collapse
1380                  * later.  However it may report false positive if the page
1381                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1382                  * will be done again later the risk seems low.
1383                  */
1384                 if (!is_refcount_suitable(folio)) {
1385                         result = SCAN_PAGE_COUNT;
1386                         goto out_unmap;
1387                 }
1388
1389                 /*
1390                  * If collapse was initiated by khugepaged, check that there is
1391                  * enough young pte to justify collapsing the page
1392                  */
1393                 if (cc->is_khugepaged &&
1394                     (pte_young(pteval) || folio_test_young(folio) ||
1395                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
1396                                                                      address)))
1397                         referenced++;
1398         }
1399         if (!writable) {
1400                 result = SCAN_PAGE_RO;
1401         } else if (cc->is_khugepaged &&
1402                    (!referenced ||
1403                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1404                 result = SCAN_LACK_REFERENCED_PAGE;
1405         } else {
1406                 result = SCAN_SUCCEED;
1407         }
1408 out_unmap:
1409         pte_unmap_unlock(pte, ptl);
1410         if (result == SCAN_SUCCEED) {
1411                 result = collapse_huge_page(mm, address, referenced,
1412                                             unmapped, cc);
1413                 /* collapse_huge_page will return with the mmap_lock released */
1414                 *mmap_locked = false;
1415         }
1416 out:
1417         trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced,
1418                                      none_or_zero, result, unmapped);
1419         return result;
1420 }
1421
1422 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1423 {
1424         struct mm_slot *slot = &mm_slot->slot;
1425         struct mm_struct *mm = slot->mm;
1426
1427         lockdep_assert_held(&khugepaged_mm_lock);
1428
1429         if (hpage_collapse_test_exit(mm)) {
1430                 /* free mm_slot */
1431                 hash_del(&slot->hash);
1432                 list_del(&slot->mm_node);
1433
1434                 /*
1435                  * Not strictly needed because the mm exited already.
1436                  *
1437                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1438                  */
1439
1440                 /* khugepaged_mm_lock actually not necessary for the below */
1441                 mm_slot_free(mm_slot_cache, mm_slot);
1442                 mmdrop(mm);
1443         }
1444 }
1445
1446 #ifdef CONFIG_SHMEM
1447 /* hpage must be locked, and mmap_lock must be held */
1448 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1449                         pmd_t *pmdp, struct page *hpage)
1450 {
1451         struct vm_fault vmf = {
1452                 .vma = vma,
1453                 .address = addr,
1454                 .flags = 0,
1455                 .pmd = pmdp,
1456         };
1457
1458         VM_BUG_ON(!PageTransHuge(hpage));
1459         mmap_assert_locked(vma->vm_mm);
1460
1461         if (do_set_pmd(&vmf, hpage))
1462                 return SCAN_FAIL;
1463
1464         get_page(hpage);
1465         return SCAN_SUCCEED;
1466 }
1467
1468 /**
1469  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1470  * address haddr.
1471  *
1472  * @mm: process address space where collapse happens
1473  * @addr: THP collapse address
1474  * @install_pmd: If a huge PMD should be installed
1475  *
1476  * This function checks whether all the PTEs in the PMD are pointing to the
1477  * right THP. If so, retract the page table so the THP can refault in with
1478  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1479  */
1480 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1481                             bool install_pmd)
1482 {
1483         struct mmu_notifier_range range;
1484         bool notified = false;
1485         unsigned long haddr = addr & HPAGE_PMD_MASK;
1486         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1487         struct folio *folio;
1488         pte_t *start_pte, *pte;
1489         pmd_t *pmd, pgt_pmd;
1490         spinlock_t *pml = NULL, *ptl;
1491         int nr_ptes = 0, result = SCAN_FAIL;
1492         int i;
1493
1494         mmap_assert_locked(mm);
1495
1496         /* First check VMA found, in case page tables are being torn down */
1497         if (!vma || !vma->vm_file ||
1498             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1499                 return SCAN_VMA_CHECK;
1500
1501         /* Fast check before locking page if already PMD-mapped */
1502         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1503         if (result == SCAN_PMD_MAPPED)
1504                 return result;
1505
1506         /*
1507          * If we are here, we've succeeded in replacing all the native pages
1508          * in the page cache with a single hugepage. If a mm were to fault-in
1509          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1510          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1511          * analogously elide sysfs THP settings here.
1512          */
1513         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false,
1514                                      PMD_ORDER))
1515                 return SCAN_VMA_CHECK;
1516
1517         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1518         if (userfaultfd_wp(vma))
1519                 return SCAN_PTE_UFFD_WP;
1520
1521         folio = filemap_lock_folio(vma->vm_file->f_mapping,
1522                                linear_page_index(vma, haddr));
1523         if (IS_ERR(folio))
1524                 return SCAN_PAGE_NULL;
1525
1526         if (folio_order(folio) != HPAGE_PMD_ORDER) {
1527                 result = SCAN_PAGE_COMPOUND;
1528                 goto drop_folio;
1529         }
1530
1531         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1532         switch (result) {
1533         case SCAN_SUCCEED:
1534                 break;
1535         case SCAN_PMD_NONE:
1536                 /*
1537                  * All pte entries have been removed and pmd cleared.
1538                  * Skip all the pte checks and just update the pmd mapping.
1539                  */
1540                 goto maybe_install_pmd;
1541         default:
1542                 goto drop_folio;
1543         }
1544
1545         result = SCAN_FAIL;
1546         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1547         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1548                 goto drop_folio;
1549
1550         /* step 1: check all mapped PTEs are to the right huge page */
1551         for (i = 0, addr = haddr, pte = start_pte;
1552              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1553                 struct page *page;
1554                 pte_t ptent = ptep_get(pte);
1555
1556                 /* empty pte, skip */
1557                 if (pte_none(ptent))
1558                         continue;
1559
1560                 /* page swapped out, abort */
1561                 if (!pte_present(ptent)) {
1562                         result = SCAN_PTE_NON_PRESENT;
1563                         goto abort;
1564                 }
1565
1566                 page = vm_normal_page(vma, addr, ptent);
1567                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1568                         page = NULL;
1569                 /*
1570                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1571                  * page table, but the new page will not be a subpage of hpage.
1572                  */
1573                 if (folio_page(folio, i) != page)
1574                         goto abort;
1575         }
1576
1577         pte_unmap_unlock(start_pte, ptl);
1578         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1579                                 haddr, haddr + HPAGE_PMD_SIZE);
1580         mmu_notifier_invalidate_range_start(&range);
1581         notified = true;
1582
1583         /*
1584          * pmd_lock covers a wider range than ptl, and (if split from mm's
1585          * page_table_lock) ptl nests inside pml. The less time we hold pml,
1586          * the better; but userfaultfd's mfill_atomic_pte() on a private VMA
1587          * inserts a valid as-if-COWed PTE without even looking up page cache.
1588          * So page lock of folio does not protect from it, so we must not drop
1589          * ptl before pgt_pmd is removed, so uffd private needs pml taken now.
1590          */
1591         if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
1592                 pml = pmd_lock(mm, pmd);
1593
1594         start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);
1595         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1596                 goto abort;
1597         if (!pml)
1598                 spin_lock(ptl);
1599         else if (ptl != pml)
1600                 spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1601
1602         /* step 2: clear page table and adjust rmap */
1603         for (i = 0, addr = haddr, pte = start_pte;
1604              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1605                 struct page *page;
1606                 pte_t ptent = ptep_get(pte);
1607
1608                 if (pte_none(ptent))
1609                         continue;
1610                 /*
1611                  * We dropped ptl after the first scan, to do the mmu_notifier:
1612                  * page lock stops more PTEs of the folio being faulted in, but
1613                  * does not stop write faults COWing anon copies from existing
1614                  * PTEs; and does not stop those being swapped out or migrated.
1615                  */
1616                 if (!pte_present(ptent)) {
1617                         result = SCAN_PTE_NON_PRESENT;
1618                         goto abort;
1619                 }
1620                 page = vm_normal_page(vma, addr, ptent);
1621                 if (folio_page(folio, i) != page)
1622                         goto abort;
1623
1624                 /*
1625                  * Must clear entry, or a racing truncate may re-remove it.
1626                  * TLB flush can be left until pmdp_collapse_flush() does it.
1627                  * PTE dirty? Shmem page is already dirty; file is read-only.
1628                  */
1629                 ptep_clear(mm, addr, pte);
1630                 folio_remove_rmap_pte(folio, page, vma);
1631                 nr_ptes++;
1632         }
1633
1634         pte_unmap(start_pte);
1635         if (!pml)
1636                 spin_unlock(ptl);
1637
1638         /* step 3: set proper refcount and mm_counters. */
1639         if (nr_ptes) {
1640                 folio_ref_sub(folio, nr_ptes);
1641                 add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
1642         }
1643
1644         /* step 4: remove empty page table */
1645         if (!pml) {
1646                 pml = pmd_lock(mm, pmd);
1647                 if (ptl != pml)
1648                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1649         }
1650         pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1651         pmdp_get_lockless_sync();
1652         if (ptl != pml)
1653                 spin_unlock(ptl);
1654         spin_unlock(pml);
1655
1656         mmu_notifier_invalidate_range_end(&range);
1657
1658         mm_dec_nr_ptes(mm);
1659         page_table_check_pte_clear_range(mm, haddr, pgt_pmd);
1660         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1661
1662 maybe_install_pmd:
1663         /* step 5: install pmd entry */
1664         result = install_pmd
1665                         ? set_huge_pmd(vma, haddr, pmd, &folio->page)
1666                         : SCAN_SUCCEED;
1667         goto drop_folio;
1668 abort:
1669         if (nr_ptes) {
1670                 flush_tlb_mm(mm);
1671                 folio_ref_sub(folio, nr_ptes);
1672                 add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
1673         }
1674         if (start_pte)
1675                 pte_unmap_unlock(start_pte, ptl);
1676         if (pml && pml != ptl)
1677                 spin_unlock(pml);
1678         if (notified)
1679                 mmu_notifier_invalidate_range_end(&range);
1680 drop_folio:
1681         folio_unlock(folio);
1682         folio_put(folio);
1683         return result;
1684 }
1685
1686 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1687 {
1688         struct vm_area_struct *vma;
1689
1690         i_mmap_lock_read(mapping);
1691         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1692                 struct mmu_notifier_range range;
1693                 struct mm_struct *mm;
1694                 unsigned long addr;
1695                 pmd_t *pmd, pgt_pmd;
1696                 spinlock_t *pml;
1697                 spinlock_t *ptl;
1698                 bool skipped_uffd = false;
1699
1700                 /*
1701                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1702                  * got written to. These VMAs are likely not worth removing
1703                  * page tables from, as PMD-mapping is likely to be split later.
1704                  */
1705                 if (READ_ONCE(vma->anon_vma))
1706                         continue;
1707
1708                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1709                 if (addr & ~HPAGE_PMD_MASK ||
1710                     vma->vm_end < addr + HPAGE_PMD_SIZE)
1711                         continue;
1712
1713                 mm = vma->vm_mm;
1714                 if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)
1715                         continue;
1716
1717                 if (hpage_collapse_test_exit(mm))
1718                         continue;
1719                 /*
1720                  * When a vma is registered with uffd-wp, we cannot recycle
1721                  * the page table because there may be pte markers installed.
1722                  * Other vmas can still have the same file mapped hugely, but
1723                  * skip this one: it will always be mapped in small page size
1724                  * for uffd-wp registered ranges.
1725                  */
1726                 if (userfaultfd_wp(vma))
1727                         continue;
1728
1729                 /* PTEs were notified when unmapped; but now for the PMD? */
1730                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1731                                         addr, addr + HPAGE_PMD_SIZE);
1732                 mmu_notifier_invalidate_range_start(&range);
1733
1734                 pml = pmd_lock(mm, pmd);
1735                 ptl = pte_lockptr(mm, pmd);
1736                 if (ptl != pml)
1737                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1738
1739                 /*
1740                  * Huge page lock is still held, so normally the page table
1741                  * must remain empty; and we have already skipped anon_vma
1742                  * and userfaultfd_wp() vmas.  But since the mmap_lock is not
1743                  * held, it is still possible for a racing userfaultfd_ioctl()
1744                  * to have inserted ptes or markers.  Now that we hold ptlock,
1745                  * repeating the anon_vma check protects from one category,
1746                  * and repeating the userfaultfd_wp() check from another.
1747                  */
1748                 if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) {
1749                         skipped_uffd = true;
1750                 } else {
1751                         pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);
1752                         pmdp_get_lockless_sync();
1753                 }
1754
1755                 if (ptl != pml)
1756                         spin_unlock(ptl);
1757                 spin_unlock(pml);
1758
1759                 mmu_notifier_invalidate_range_end(&range);
1760
1761                 if (!skipped_uffd) {
1762                         mm_dec_nr_ptes(mm);
1763                         page_table_check_pte_clear_range(mm, addr, pgt_pmd);
1764                         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1765                 }
1766         }
1767         i_mmap_unlock_read(mapping);
1768 }
1769
1770 /**
1771  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1772  *
1773  * @mm: process address space where collapse happens
1774  * @addr: virtual collapse start address
1775  * @file: file that collapse on
1776  * @start: collapse start address
1777  * @cc: collapse context and scratchpad
1778  *
1779  * Basic scheme is simple, details are more complex:
1780  *  - allocate and lock a new huge page;
1781  *  - scan page cache, locking old pages
1782  *    + swap/gup in pages if necessary;
1783  *  - copy data to new page
1784  *  - handle shmem holes
1785  *    + re-validate that holes weren't filled by someone else
1786  *    + check for userfaultfd
1787  *  - finalize updates to the page cache;
1788  *  - if replacing succeeds:
1789  *    + unlock huge page;
1790  *    + free old pages;
1791  *  - if replacing failed;
1792  *    + unlock old pages
1793  *    + unlock and free huge page;
1794  */
1795 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1796                          struct file *file, pgoff_t start,
1797                          struct collapse_control *cc)
1798 {
1799         struct address_space *mapping = file->f_mapping;
1800         struct page *hpage;
1801         struct page *page;
1802         struct page *tmp;
1803         struct folio *folio;
1804         pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1805         LIST_HEAD(pagelist);
1806         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1807         int nr_none = 0, result = SCAN_SUCCEED;
1808         bool is_shmem = shmem_file(file);
1809         int nr = 0;
1810
1811         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1812         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1813
1814         result = alloc_charge_hpage(&hpage, mm, cc);
1815         if (result != SCAN_SUCCEED)
1816                 goto out;
1817
1818         __SetPageLocked(hpage);
1819         if (is_shmem)
1820                 __SetPageSwapBacked(hpage);
1821         hpage->index = start;
1822         hpage->mapping = mapping;
1823
1824         /*
1825          * Ensure we have slots for all the pages in the range.  This is
1826          * almost certainly a no-op because most of the pages must be present
1827          */
1828         do {
1829                 xas_lock_irq(&xas);
1830                 xas_create_range(&xas);
1831                 if (!xas_error(&xas))
1832                         break;
1833                 xas_unlock_irq(&xas);
1834                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1835                         result = SCAN_FAIL;
1836                         goto rollback;
1837                 }
1838         } while (1);
1839
1840         for (index = start; index < end; index++) {
1841                 xas_set(&xas, index);
1842                 page = xas_load(&xas);
1843
1844                 VM_BUG_ON(index != xas.xa_index);
1845                 if (is_shmem) {
1846                         if (!page) {
1847                                 /*
1848                                  * Stop if extent has been truncated or
1849                                  * hole-punched, and is now completely
1850                                  * empty.
1851                                  */
1852                                 if (index == start) {
1853                                         if (!xas_next_entry(&xas, end - 1)) {
1854                                                 result = SCAN_TRUNCATED;
1855                                                 goto xa_locked;
1856                                         }
1857                                 }
1858                                 nr_none++;
1859                                 continue;
1860                         }
1861
1862                         if (xa_is_value(page) || !PageUptodate(page)) {
1863                                 xas_unlock_irq(&xas);
1864                                 /* swap in or instantiate fallocated page */
1865                                 if (shmem_get_folio(mapping->host, index,
1866                                                 &folio, SGP_NOALLOC)) {
1867                                         result = SCAN_FAIL;
1868                                         goto xa_unlocked;
1869                                 }
1870                                 /* drain lru cache to help isolate_lru_page() */
1871                                 lru_add_drain();
1872                                 page = folio_file_page(folio, index);
1873                         } else if (trylock_page(page)) {
1874                                 get_page(page);
1875                                 xas_unlock_irq(&xas);
1876                         } else {
1877                                 result = SCAN_PAGE_LOCK;
1878                                 goto xa_locked;
1879                         }
1880                 } else {        /* !is_shmem */
1881                         if (!page || xa_is_value(page)) {
1882                                 xas_unlock_irq(&xas);
1883                                 page_cache_sync_readahead(mapping, &file->f_ra,
1884                                                           file, index,
1885                                                           end - index);
1886                                 /* drain lru cache to help isolate_lru_page() */
1887                                 lru_add_drain();
1888                                 page = find_lock_page(mapping, index);
1889                                 if (unlikely(page == NULL)) {
1890                                         result = SCAN_FAIL;
1891                                         goto xa_unlocked;
1892                                 }
1893                         } else if (PageDirty(page)) {
1894                                 /*
1895                                  * khugepaged only works on read-only fd,
1896                                  * so this page is dirty because it hasn't
1897                                  * been flushed since first write. There
1898                                  * won't be new dirty pages.
1899                                  *
1900                                  * Trigger async flush here and hope the
1901                                  * writeback is done when khugepaged
1902                                  * revisits this page.
1903                                  *
1904                                  * This is a one-off situation. We are not
1905                                  * forcing writeback in loop.
1906                                  */
1907                                 xas_unlock_irq(&xas);
1908                                 filemap_flush(mapping);
1909                                 result = SCAN_FAIL;
1910                                 goto xa_unlocked;
1911                         } else if (PageWriteback(page)) {
1912                                 xas_unlock_irq(&xas);
1913                                 result = SCAN_FAIL;
1914                                 goto xa_unlocked;
1915                         } else if (trylock_page(page)) {
1916                                 get_page(page);
1917                                 xas_unlock_irq(&xas);
1918                         } else {
1919                                 result = SCAN_PAGE_LOCK;
1920                                 goto xa_locked;
1921                         }
1922                 }
1923
1924                 /*
1925                  * The page must be locked, so we can drop the i_pages lock
1926                  * without racing with truncate.
1927                  */
1928                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1929
1930                 /* make sure the page is up to date */
1931                 if (unlikely(!PageUptodate(page))) {
1932                         result = SCAN_FAIL;
1933                         goto out_unlock;
1934                 }
1935
1936                 /*
1937                  * If file was truncated then extended, or hole-punched, before
1938                  * we locked the first page, then a THP might be there already.
1939                  * This will be discovered on the first iteration.
1940                  */
1941                 if (PageTransCompound(page)) {
1942                         struct page *head = compound_head(page);
1943
1944                         result = compound_order(head) == HPAGE_PMD_ORDER &&
1945                                         head->index == start
1946                                         /* Maybe PMD-mapped */
1947                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1948                                         : SCAN_PAGE_COMPOUND;
1949                         goto out_unlock;
1950                 }
1951
1952                 folio = page_folio(page);
1953
1954                 if (folio_mapping(folio) != mapping) {
1955                         result = SCAN_TRUNCATED;
1956                         goto out_unlock;
1957                 }
1958
1959                 if (!is_shmem && (folio_test_dirty(folio) ||
1960                                   folio_test_writeback(folio))) {
1961                         /*
1962                          * khugepaged only works on read-only fd, so this
1963                          * page is dirty because it hasn't been flushed
1964                          * since first write.
1965                          */
1966                         result = SCAN_FAIL;
1967                         goto out_unlock;
1968                 }
1969
1970                 if (!folio_isolate_lru(folio)) {
1971                         result = SCAN_DEL_PAGE_LRU;
1972                         goto out_unlock;
1973                 }
1974
1975                 if (!filemap_release_folio(folio, GFP_KERNEL)) {
1976                         result = SCAN_PAGE_HAS_PRIVATE;
1977                         folio_putback_lru(folio);
1978                         goto out_unlock;
1979                 }
1980
1981                 if (folio_mapped(folio))
1982                         try_to_unmap(folio,
1983                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1984
1985                 xas_lock_irq(&xas);
1986
1987                 VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page);
1988
1989                 /*
1990                  * We control three references to the page:
1991                  *  - we hold a pin on it;
1992                  *  - one reference from page cache;
1993                  *  - one from isolate_lru_page;
1994                  * If those are the only references, then any new usage of the
1995                  * page will have to fetch it from the page cache. That requires
1996                  * locking the page to handle truncate, so any new usage will be
1997                  * blocked until we unlock page after collapse/during rollback.
1998                  */
1999                 if (page_count(page) != 3) {
2000                         result = SCAN_PAGE_COUNT;
2001                         xas_unlock_irq(&xas);
2002                         putback_lru_page(page);
2003                         goto out_unlock;
2004                 }
2005
2006                 /*
2007                  * Accumulate the pages that are being collapsed.
2008                  */
2009                 list_add_tail(&page->lru, &pagelist);
2010                 continue;
2011 out_unlock:
2012                 unlock_page(page);
2013                 put_page(page);
2014                 goto xa_unlocked;
2015         }
2016
2017         if (!is_shmem) {
2018                 filemap_nr_thps_inc(mapping);
2019                 /*
2020                  * Paired with smp_mb() in do_dentry_open() to ensure
2021                  * i_writecount is up to date and the update to nr_thps is
2022                  * visible. Ensures the page cache will be truncated if the
2023                  * file is opened writable.
2024                  */
2025                 smp_mb();
2026                 if (inode_is_open_for_write(mapping->host)) {
2027                         result = SCAN_FAIL;
2028                         filemap_nr_thps_dec(mapping);
2029                 }
2030         }
2031
2032 xa_locked:
2033         xas_unlock_irq(&xas);
2034 xa_unlocked:
2035
2036         /*
2037          * If collapse is successful, flush must be done now before copying.
2038          * If collapse is unsuccessful, does flush actually need to be done?
2039          * Do it anyway, to clear the state.
2040          */
2041         try_to_unmap_flush();
2042
2043         if (result == SCAN_SUCCEED && nr_none &&
2044             !shmem_charge(mapping->host, nr_none))
2045                 result = SCAN_FAIL;
2046         if (result != SCAN_SUCCEED) {
2047                 nr_none = 0;
2048                 goto rollback;
2049         }
2050
2051         /*
2052          * The old pages are locked, so they won't change anymore.
2053          */
2054         index = start;
2055         list_for_each_entry(page, &pagelist, lru) {
2056                 while (index < page->index) {
2057                         clear_highpage(hpage + (index % HPAGE_PMD_NR));
2058                         index++;
2059                 }
2060                 if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) {
2061                         result = SCAN_COPY_MC;
2062                         goto rollback;
2063                 }
2064                 index++;
2065         }
2066         while (index < end) {
2067                 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2068                 index++;
2069         }
2070
2071         if (nr_none) {
2072                 struct vm_area_struct *vma;
2073                 int nr_none_check = 0;
2074
2075                 i_mmap_lock_read(mapping);
2076                 xas_lock_irq(&xas);
2077
2078                 xas_set(&xas, start);
2079                 for (index = start; index < end; index++) {
2080                         if (!xas_next(&xas)) {
2081                                 xas_store(&xas, XA_RETRY_ENTRY);
2082                                 if (xas_error(&xas)) {
2083                                         result = SCAN_STORE_FAILED;
2084                                         goto immap_locked;
2085                                 }
2086                                 nr_none_check++;
2087                         }
2088                 }
2089
2090                 if (nr_none != nr_none_check) {
2091                         result = SCAN_PAGE_FILLED;
2092                         goto immap_locked;
2093                 }
2094
2095                 /*
2096                  * If userspace observed a missing page in a VMA with a MODE_MISSING
2097                  * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that
2098                  * page. If so, we need to roll back to avoid suppressing such an
2099                  * event. Since wp/minor userfaultfds don't give userspace any
2100                  * guarantees that the kernel doesn't fill a missing page with a zero
2101                  * page, so they don't matter here.
2102                  *
2103                  * Any userfaultfds registered after this point will not be able to
2104                  * observe any missing pages due to the previously inserted retry
2105                  * entries.
2106                  */
2107                 vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {
2108                         if (userfaultfd_missing(vma)) {
2109                                 result = SCAN_EXCEED_NONE_PTE;
2110                                 goto immap_locked;
2111                         }
2112                 }
2113
2114 immap_locked:
2115                 i_mmap_unlock_read(mapping);
2116                 if (result != SCAN_SUCCEED) {
2117                         xas_set(&xas, start);
2118                         for (index = start; index < end; index++) {
2119                                 if (xas_next(&xas) == XA_RETRY_ENTRY)
2120                                         xas_store(&xas, NULL);
2121                         }
2122
2123                         xas_unlock_irq(&xas);
2124                         goto rollback;
2125                 }
2126         } else {
2127                 xas_lock_irq(&xas);
2128         }
2129
2130         folio = page_folio(hpage);
2131         nr = folio_nr_pages(folio);
2132         if (is_shmem)
2133                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
2134         else
2135                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, nr);
2136
2137         if (nr_none) {
2138                 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr_none);
2139                 /* nr_none is always 0 for non-shmem. */
2140                 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr_none);
2141         }
2142
2143         /*
2144          * Mark hpage as uptodate before inserting it into the page cache so
2145          * that it isn't mistaken for an fallocated but unwritten page.
2146          */
2147         folio_mark_uptodate(folio);
2148         folio_ref_add(folio, HPAGE_PMD_NR - 1);
2149
2150         if (is_shmem)
2151                 folio_mark_dirty(folio);
2152         folio_add_lru(folio);
2153
2154         /* Join all the small entries into a single multi-index entry. */
2155         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2156         xas_store(&xas, folio);
2157         WARN_ON_ONCE(xas_error(&xas));
2158         xas_unlock_irq(&xas);
2159
2160         /*
2161          * Remove pte page tables, so we can re-fault the page as huge.
2162          * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().
2163          */
2164         retract_page_tables(mapping, start);
2165         if (cc && !cc->is_khugepaged)
2166                 result = SCAN_PTE_MAPPED_HUGEPAGE;
2167         folio_unlock(folio);
2168
2169         /*
2170          * The collapse has succeeded, so free the old pages.
2171          */
2172         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2173                 list_del(&page->lru);
2174                 page->mapping = NULL;
2175                 ClearPageActive(page);
2176                 ClearPageUnevictable(page);
2177                 unlock_page(page);
2178                 folio_put_refs(page_folio(page), 3);
2179         }
2180
2181         goto out;
2182
2183 rollback:
2184         /* Something went wrong: roll back page cache changes */
2185         if (nr_none) {
2186                 xas_lock_irq(&xas);
2187                 mapping->nrpages -= nr_none;
2188                 xas_unlock_irq(&xas);
2189                 shmem_uncharge(mapping->host, nr_none);
2190         }
2191
2192         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2193                 list_del(&page->lru);
2194                 unlock_page(page);
2195                 putback_lru_page(page);
2196                 put_page(page);
2197         }
2198         /*
2199          * Undo the updates of filemap_nr_thps_inc for non-SHMEM
2200          * file only. This undo is not needed unless failure is
2201          * due to SCAN_COPY_MC.
2202          */
2203         if (!is_shmem && result == SCAN_COPY_MC) {
2204                 filemap_nr_thps_dec(mapping);
2205                 /*
2206                  * Paired with smp_mb() in do_dentry_open() to
2207                  * ensure the update to nr_thps is visible.
2208                  */
2209                 smp_mb();
2210         }
2211
2212         hpage->mapping = NULL;
2213
2214         unlock_page(hpage);
2215         put_page(hpage);
2216 out:
2217         VM_BUG_ON(!list_empty(&pagelist));
2218         trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2219         return result;
2220 }
2221
2222 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2223                                     struct file *file, pgoff_t start,
2224                                     struct collapse_control *cc)
2225 {
2226         struct page *page = NULL;
2227         struct address_space *mapping = file->f_mapping;
2228         XA_STATE(xas, &mapping->i_pages, start);
2229         int present, swap;
2230         int node = NUMA_NO_NODE;
2231         int result = SCAN_SUCCEED;
2232
2233         present = 0;
2234         swap = 0;
2235         memset(cc->node_load, 0, sizeof(cc->node_load));
2236         nodes_clear(cc->alloc_nmask);
2237         rcu_read_lock();
2238         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2239                 if (xas_retry(&xas, page))
2240                         continue;
2241
2242                 if (xa_is_value(page)) {
2243                         ++swap;
2244                         if (cc->is_khugepaged &&
2245                             swap > khugepaged_max_ptes_swap) {
2246                                 result = SCAN_EXCEED_SWAP_PTE;
2247                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2248                                 break;
2249                         }
2250                         continue;
2251                 }
2252
2253                 /*
2254                  * TODO: khugepaged should compact smaller compound pages
2255                  * into a PMD sized page
2256                  */
2257                 if (PageTransCompound(page)) {
2258                         struct page *head = compound_head(page);
2259
2260                         result = compound_order(head) == HPAGE_PMD_ORDER &&
2261                                         head->index == start
2262                                         /* Maybe PMD-mapped */
2263                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2264                                         : SCAN_PAGE_COMPOUND;
2265                         /*
2266                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2267                          * by the caller won't touch the page cache, and so
2268                          * it's safe to skip LRU and refcount checks before
2269                          * returning.
2270                          */
2271                         break;
2272                 }
2273
2274                 node = page_to_nid(page);
2275                 if (hpage_collapse_scan_abort(node, cc)) {
2276                         result = SCAN_SCAN_ABORT;
2277                         break;
2278                 }
2279                 cc->node_load[node]++;
2280
2281                 if (!PageLRU(page)) {
2282                         result = SCAN_PAGE_LRU;
2283                         break;
2284                 }
2285
2286                 if (page_count(page) !=
2287                     1 + page_mapcount(page) + page_has_private(page)) {
2288                         result = SCAN_PAGE_COUNT;
2289                         break;
2290                 }
2291
2292                 /*
2293                  * We probably should check if the page is referenced here, but
2294                  * nobody would transfer pte_young() to PageReferenced() for us.
2295                  * And rmap walk here is just too costly...
2296                  */
2297
2298                 present++;
2299
2300                 if (need_resched()) {
2301                         xas_pause(&xas);
2302                         cond_resched_rcu();
2303                 }
2304         }
2305         rcu_read_unlock();
2306
2307         if (result == SCAN_SUCCEED) {
2308                 if (cc->is_khugepaged &&
2309                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2310                         result = SCAN_EXCEED_NONE_PTE;
2311                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2312                 } else {
2313                         result = collapse_file(mm, addr, file, start, cc);
2314                 }
2315         }
2316
2317         trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2318         return result;
2319 }
2320 #else
2321 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2322                                     struct file *file, pgoff_t start,
2323                                     struct collapse_control *cc)
2324 {
2325         BUILD_BUG();
2326 }
2327 #endif
2328
2329 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2330                                             struct collapse_control *cc)
2331         __releases(&khugepaged_mm_lock)
2332         __acquires(&khugepaged_mm_lock)
2333 {
2334         struct vma_iterator vmi;
2335         struct khugepaged_mm_slot *mm_slot;
2336         struct mm_slot *slot;
2337         struct mm_struct *mm;
2338         struct vm_area_struct *vma;
2339         int progress = 0;
2340
2341         VM_BUG_ON(!pages);
2342         lockdep_assert_held(&khugepaged_mm_lock);
2343         *result = SCAN_FAIL;
2344
2345         if (khugepaged_scan.mm_slot) {
2346                 mm_slot = khugepaged_scan.mm_slot;
2347                 slot = &mm_slot->slot;
2348         } else {
2349                 slot = list_entry(khugepaged_scan.mm_head.next,
2350                                      struct mm_slot, mm_node);
2351                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2352                 khugepaged_scan.address = 0;
2353                 khugepaged_scan.mm_slot = mm_slot;
2354         }
2355         spin_unlock(&khugepaged_mm_lock);
2356
2357         mm = slot->mm;
2358         /*
2359          * Don't wait for semaphore (to avoid long wait times).  Just move to
2360          * the next mm on the list.
2361          */
2362         vma = NULL;
2363         if (unlikely(!mmap_read_trylock(mm)))
2364                 goto breakouterloop_mmap_lock;
2365
2366         progress++;
2367         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2368                 goto breakouterloop;
2369
2370         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2371         for_each_vma(vmi, vma) {
2372                 unsigned long hstart, hend;
2373
2374                 cond_resched();
2375                 if (unlikely(hpage_collapse_test_exit_or_disable(mm))) {
2376                         progress++;
2377                         break;
2378                 }
2379                 if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false,
2380                                              true, PMD_ORDER)) {
2381 skip:
2382                         progress++;
2383                         continue;
2384                 }
2385                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2386                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2387                 if (khugepaged_scan.address > hend)
2388                         goto skip;
2389                 if (khugepaged_scan.address < hstart)
2390                         khugepaged_scan.address = hstart;
2391                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2392
2393                 while (khugepaged_scan.address < hend) {
2394                         bool mmap_locked = true;
2395
2396                         cond_resched();
2397                         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2398                                 goto breakouterloop;
2399
2400                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2401                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2402                                   hend);
2403                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2404                                 struct file *file = get_file(vma->vm_file);
2405                                 pgoff_t pgoff = linear_page_index(vma,
2406                                                 khugepaged_scan.address);
2407
2408                                 mmap_read_unlock(mm);
2409                                 mmap_locked = false;
2410                                 *result = hpage_collapse_scan_file(mm,
2411                                         khugepaged_scan.address, file, pgoff, cc);
2412                                 fput(file);
2413                                 if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {
2414                                         mmap_read_lock(mm);
2415                                         if (hpage_collapse_test_exit_or_disable(mm))
2416                                                 goto breakouterloop;
2417                                         *result = collapse_pte_mapped_thp(mm,
2418                                                 khugepaged_scan.address, false);
2419                                         if (*result == SCAN_PMD_MAPPED)
2420                                                 *result = SCAN_SUCCEED;
2421                                         mmap_read_unlock(mm);
2422                                 }
2423                         } else {
2424                                 *result = hpage_collapse_scan_pmd(mm, vma,
2425                                         khugepaged_scan.address, &mmap_locked, cc);
2426                         }
2427
2428                         if (*result == SCAN_SUCCEED)
2429                                 ++khugepaged_pages_collapsed;
2430
2431                         /* move to next address */
2432                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2433                         progress += HPAGE_PMD_NR;
2434                         if (!mmap_locked)
2435                                 /*
2436                                  * We released mmap_lock so break loop.  Note
2437                                  * that we drop mmap_lock before all hugepage
2438                                  * allocations, so if allocation fails, we are
2439                                  * guaranteed to break here and report the
2440                                  * correct result back to caller.
2441                                  */
2442                                 goto breakouterloop_mmap_lock;
2443                         if (progress >= pages)
2444                                 goto breakouterloop;
2445                 }
2446         }
2447 breakouterloop:
2448         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2449 breakouterloop_mmap_lock:
2450
2451         spin_lock(&khugepaged_mm_lock);
2452         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2453         /*
2454          * Release the current mm_slot if this mm is about to die, or
2455          * if we scanned all vmas of this mm.
2456          */
2457         if (hpage_collapse_test_exit(mm) || !vma) {
2458                 /*
2459                  * Make sure that if mm_users is reaching zero while
2460                  * khugepaged runs here, khugepaged_exit will find
2461                  * mm_slot not pointing to the exiting mm.
2462                  */
2463                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2464                         slot = list_entry(slot->mm_node.next,
2465                                           struct mm_slot, mm_node);
2466                         khugepaged_scan.mm_slot =
2467                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2468                         khugepaged_scan.address = 0;
2469                 } else {
2470                         khugepaged_scan.mm_slot = NULL;
2471                         khugepaged_full_scans++;
2472                 }
2473
2474                 collect_mm_slot(mm_slot);
2475         }
2476
2477         return progress;
2478 }
2479
2480 static int khugepaged_has_work(void)
2481 {
2482         return !list_empty(&khugepaged_scan.mm_head) &&
2483                 hugepage_flags_enabled();
2484 }
2485
2486 static int khugepaged_wait_event(void)
2487 {
2488         return !list_empty(&khugepaged_scan.mm_head) ||
2489                 kthread_should_stop();
2490 }
2491
2492 static void khugepaged_do_scan(struct collapse_control *cc)
2493 {
2494         unsigned int progress = 0, pass_through_head = 0;
2495         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2496         bool wait = true;
2497         int result = SCAN_SUCCEED;
2498
2499         lru_add_drain_all();
2500
2501         while (true) {
2502                 cond_resched();
2503
2504                 if (unlikely(kthread_should_stop()))
2505                         break;
2506
2507                 spin_lock(&khugepaged_mm_lock);
2508                 if (!khugepaged_scan.mm_slot)
2509                         pass_through_head++;
2510                 if (khugepaged_has_work() &&
2511                     pass_through_head < 2)
2512                         progress += khugepaged_scan_mm_slot(pages - progress,
2513                                                             &result, cc);
2514                 else
2515                         progress = pages;
2516                 spin_unlock(&khugepaged_mm_lock);
2517
2518                 if (progress >= pages)
2519                         break;
2520
2521                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2522                         /*
2523                          * If fail to allocate the first time, try to sleep for
2524                          * a while.  When hit again, cancel the scan.
2525                          */
2526                         if (!wait)
2527                                 break;
2528                         wait = false;
2529                         khugepaged_alloc_sleep();
2530                 }
2531         }
2532 }
2533
2534 static bool khugepaged_should_wakeup(void)
2535 {
2536         return kthread_should_stop() ||
2537                time_after_eq(jiffies, khugepaged_sleep_expire);
2538 }
2539
2540 static void khugepaged_wait_work(void)
2541 {
2542         if (khugepaged_has_work()) {
2543                 const unsigned long scan_sleep_jiffies =
2544                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2545
2546                 if (!scan_sleep_jiffies)
2547                         return;
2548
2549                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2550                 wait_event_freezable_timeout(khugepaged_wait,
2551                                              khugepaged_should_wakeup(),
2552                                              scan_sleep_jiffies);
2553                 return;
2554         }
2555
2556         if (hugepage_flags_enabled())
2557                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2558 }
2559
2560 static int khugepaged(void *none)
2561 {
2562         struct khugepaged_mm_slot *mm_slot;
2563
2564         set_freezable();
2565         set_user_nice(current, MAX_NICE);
2566
2567         while (!kthread_should_stop()) {
2568                 khugepaged_do_scan(&khugepaged_collapse_control);
2569                 khugepaged_wait_work();
2570         }
2571
2572         spin_lock(&khugepaged_mm_lock);
2573         mm_slot = khugepaged_scan.mm_slot;
2574         khugepaged_scan.mm_slot = NULL;
2575         if (mm_slot)
2576                 collect_mm_slot(mm_slot);
2577         spin_unlock(&khugepaged_mm_lock);
2578         return 0;
2579 }
2580
2581 static void set_recommended_min_free_kbytes(void)
2582 {
2583         struct zone *zone;
2584         int nr_zones = 0;
2585         unsigned long recommended_min;
2586
2587         if (!hugepage_flags_enabled()) {
2588                 calculate_min_free_kbytes();
2589                 goto update_wmarks;
2590         }
2591
2592         for_each_populated_zone(zone) {
2593                 /*
2594                  * We don't need to worry about fragmentation of
2595                  * ZONE_MOVABLE since it only has movable pages.
2596                  */
2597                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2598                         continue;
2599
2600                 nr_zones++;
2601         }
2602
2603         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2604         recommended_min = pageblock_nr_pages * nr_zones * 2;
2605
2606         /*
2607          * Make sure that on average at least two pageblocks are almost free
2608          * of another type, one for a migratetype to fall back to and a
2609          * second to avoid subsequent fallbacks of other types There are 3
2610          * MIGRATE_TYPES we care about.
2611          */
2612         recommended_min += pageblock_nr_pages * nr_zones *
2613                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2614
2615         /* don't ever allow to reserve more than 5% of the lowmem */
2616         recommended_min = min(recommended_min,
2617                               (unsigned long) nr_free_buffer_pages() / 20);
2618         recommended_min <<= (PAGE_SHIFT-10);
2619
2620         if (recommended_min > min_free_kbytes) {
2621                 if (user_min_free_kbytes >= 0)
2622                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2623                                 min_free_kbytes, recommended_min);
2624
2625                 min_free_kbytes = recommended_min;
2626         }
2627
2628 update_wmarks:
2629         setup_per_zone_wmarks();
2630 }
2631
2632 int start_stop_khugepaged(void)
2633 {
2634         int err = 0;
2635
2636         mutex_lock(&khugepaged_mutex);
2637         if (hugepage_flags_enabled()) {
2638                 if (!khugepaged_thread)
2639                         khugepaged_thread = kthread_run(khugepaged, NULL,
2640                                                         "khugepaged");
2641                 if (IS_ERR(khugepaged_thread)) {
2642                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2643                         err = PTR_ERR(khugepaged_thread);
2644                         khugepaged_thread = NULL;
2645                         goto fail;
2646                 }
2647
2648                 if (!list_empty(&khugepaged_scan.mm_head))
2649                         wake_up_interruptible(&khugepaged_wait);
2650         } else if (khugepaged_thread) {
2651                 kthread_stop(khugepaged_thread);
2652                 khugepaged_thread = NULL;
2653         }
2654         set_recommended_min_free_kbytes();
2655 fail:
2656         mutex_unlock(&khugepaged_mutex);
2657         return err;
2658 }
2659
2660 void khugepaged_min_free_kbytes_update(void)
2661 {
2662         mutex_lock(&khugepaged_mutex);
2663         if (hugepage_flags_enabled() && khugepaged_thread)
2664                 set_recommended_min_free_kbytes();
2665         mutex_unlock(&khugepaged_mutex);
2666 }
2667
2668 bool current_is_khugepaged(void)
2669 {
2670         return kthread_func(current) == khugepaged;
2671 }
2672
2673 static int madvise_collapse_errno(enum scan_result r)
2674 {
2675         /*
2676          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2677          * actionable feedback to caller, so they may take an appropriate
2678          * fallback measure depending on the nature of the failure.
2679          */
2680         switch (r) {
2681         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2682                 return -ENOMEM;
2683         case SCAN_CGROUP_CHARGE_FAIL:
2684         case SCAN_EXCEED_NONE_PTE:
2685                 return -EBUSY;
2686         /* Resource temporary unavailable - trying again might succeed */
2687         case SCAN_PAGE_COUNT:
2688         case SCAN_PAGE_LOCK:
2689         case SCAN_PAGE_LRU:
2690         case SCAN_DEL_PAGE_LRU:
2691         case SCAN_PAGE_FILLED:
2692                 return -EAGAIN;
2693         /*
2694          * Other: Trying again likely not to succeed / error intrinsic to
2695          * specified memory range. khugepaged likely won't be able to collapse
2696          * either.
2697          */
2698         default:
2699                 return -EINVAL;
2700         }
2701 }
2702
2703 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2704                      unsigned long start, unsigned long end)
2705 {
2706         struct collapse_control *cc;
2707         struct mm_struct *mm = vma->vm_mm;
2708         unsigned long hstart, hend, addr;
2709         int thps = 0, last_fail = SCAN_FAIL;
2710         bool mmap_locked = true;
2711
2712         BUG_ON(vma->vm_start > start);
2713         BUG_ON(vma->vm_end < end);
2714
2715         *prev = vma;
2716
2717         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false,
2718                                      PMD_ORDER))
2719                 return -EINVAL;
2720
2721         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2722         if (!cc)
2723                 return -ENOMEM;
2724         cc->is_khugepaged = false;
2725
2726         mmgrab(mm);
2727         lru_add_drain_all();
2728
2729         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2730         hend = end & HPAGE_PMD_MASK;
2731
2732         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2733                 int result = SCAN_FAIL;
2734
2735                 if (!mmap_locked) {
2736                         cond_resched();
2737                         mmap_read_lock(mm);
2738                         mmap_locked = true;
2739                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2740                                                          cc);
2741                         if (result  != SCAN_SUCCEED) {
2742                                 last_fail = result;
2743                                 goto out_nolock;
2744                         }
2745
2746                         hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2747                 }
2748                 mmap_assert_locked(mm);
2749                 memset(cc->node_load, 0, sizeof(cc->node_load));
2750                 nodes_clear(cc->alloc_nmask);
2751                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2752                         struct file *file = get_file(vma->vm_file);
2753                         pgoff_t pgoff = linear_page_index(vma, addr);
2754
2755                         mmap_read_unlock(mm);
2756                         mmap_locked = false;
2757                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2758                                                           cc);
2759                         fput(file);
2760                 } else {
2761                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2762                                                          &mmap_locked, cc);
2763                 }
2764                 if (!mmap_locked)
2765                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2766
2767 handle_result:
2768                 switch (result) {
2769                 case SCAN_SUCCEED:
2770                 case SCAN_PMD_MAPPED:
2771                         ++thps;
2772                         break;
2773                 case SCAN_PTE_MAPPED_HUGEPAGE:
2774                         BUG_ON(mmap_locked);
2775                         BUG_ON(*prev);
2776                         mmap_read_lock(mm);
2777                         result = collapse_pte_mapped_thp(mm, addr, true);
2778                         mmap_read_unlock(mm);
2779                         goto handle_result;
2780                 /* Whitelisted set of results where continuing OK */
2781                 case SCAN_PMD_NULL:
2782                 case SCAN_PTE_NON_PRESENT:
2783                 case SCAN_PTE_UFFD_WP:
2784                 case SCAN_PAGE_RO:
2785                 case SCAN_LACK_REFERENCED_PAGE:
2786                 case SCAN_PAGE_NULL:
2787                 case SCAN_PAGE_COUNT:
2788                 case SCAN_PAGE_LOCK:
2789                 case SCAN_PAGE_COMPOUND:
2790                 case SCAN_PAGE_LRU:
2791                 case SCAN_DEL_PAGE_LRU:
2792                         last_fail = result;
2793                         break;
2794                 default:
2795                         last_fail = result;
2796                         /* Other error, exit */
2797                         goto out_maybelock;
2798                 }
2799         }
2800
2801 out_maybelock:
2802         /* Caller expects us to hold mmap_lock on return */
2803         if (!mmap_locked)
2804                 mmap_read_lock(mm);
2805 out_nolock:
2806         mmap_assert_locked(mm);
2807         mmdrop(mm);
2808         kfree(cc);
2809
2810         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2811                         : madvise_collapse_errno(last_fail);
2812 }