mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/page_table_check.h>
  20 #include <linux/swapops.h>
  21 #include <linux/shmem_fs.h>
  22
  23 #include <asm/tlb.h>
  24 #include <asm/pgalloc.h>
  25 #include "internal.h"
  26 #include "mm_slot.h"
  27
  28 enum scan_result {
  29         SCAN_FAIL,
  30         SCAN_SUCCEED,
  31         SCAN_PMD_NULL,
  32         SCAN_PMD_NONE,
  33         SCAN_PMD_MAPPED,
  34         SCAN_EXCEED_NONE_PTE,
  35         SCAN_EXCEED_SWAP_PTE,
  36         SCAN_EXCEED_SHARED_PTE,
  37         SCAN_PTE_NON_PRESENT,
  38         SCAN_PTE_UFFD_WP,
  39         SCAN_PTE_MAPPED_HUGEPAGE,
  40         SCAN_PAGE_RO,
  41         SCAN_LACK_REFERENCED_PAGE,
  42         SCAN_PAGE_NULL,
  43         SCAN_SCAN_ABORT,
  44         SCAN_PAGE_COUNT,
  45         SCAN_PAGE_LRU,
  46         SCAN_PAGE_LOCK,
  47         SCAN_PAGE_ANON,
  48         SCAN_PAGE_COMPOUND,
  49         SCAN_ANY_PROCESS,
  50         SCAN_VMA_NULL,
  51         SCAN_VMA_CHECK,
  52         SCAN_ADDRESS_RANGE,
  53         SCAN_DEL_PAGE_LRU,
  54         SCAN_ALLOC_HUGE_PAGE_FAIL,
  55         SCAN_CGROUP_CHARGE_FAIL,
  56         SCAN_TRUNCATED,
  57         SCAN_PAGE_HAS_PRIVATE,
  58 };
  59
  60 #define CREATE_TRACE_POINTS
  61 #include <trace/events/huge_memory.h>
  62
  63 static struct task_struct *khugepaged_thread __read_mostly;
  64 static DEFINE_MUTEX(khugepaged_mutex);
  65
  66 /* default scan 8*512 pte (or vmas) every 30 second */
  67 static unsigned int khugepaged_pages_to_scan __read_mostly;
  68 static unsigned int khugepaged_pages_collapsed;
  69 static unsigned int khugepaged_full_scans;
  70 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  71 /* during fragmentation poll the hugepage allocator once every minute */
  72 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  73 static unsigned long khugepaged_sleep_expire;
  74 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  75 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  76 /*
  77  * default collapse hugepages if there is at least one pte mapped like
  78  * it would have happened if the vma was large enough during page
  79  * fault.
  80  *
  81  * Note that these are only respected if collapse was initiated by khugepaged.
  82  */
  83 static unsigned int khugepaged_max_ptes_none __read_mostly;
  84 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  85 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  86
  87 #define MM_SLOTS_HASH_BITS 10
  88 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  89
  90 static struct kmem_cache *mm_slot_cache __read_mostly;
  91
  92 #define MAX_PTE_MAPPED_THP 8
  93
  94 struct collapse_control {
  95         bool is_khugepaged;
  96
  97         /* Num pages scanned per node */
  98         u32 node_load[MAX_NUMNODES];
  99
 100         /* nodemask for allocation fallback */
 101         nodemask_t alloc_nmask;
 102 };
 103
 104 /**
 105  * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned
 106  * @slot: hash lookup from mm to mm_slot
 107  * @nr_pte_mapped_thp: number of pte mapped THP
 108  * @pte_mapped_thp: address array corresponding pte mapped THP
 109  */
 110 struct khugepaged_mm_slot {
 111         struct mm_slot slot;
 112
 113         /* pte-mapped THP in this mm */
 114         int nr_pte_mapped_thp;
 115         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
 116 };
 117
 118 /**
 119  * struct khugepaged_scan - cursor for scanning
 120  * @mm_head: the head of the mm list to scan
 121  * @mm_slot: the current mm_slot we are scanning
 122  * @address: the next address inside that to be scanned
 123  *
 124  * There is only the one khugepaged_scan instance of this cursor structure.
 125  */
 126 struct khugepaged_scan {
 127         struct list_head mm_head;
 128         struct khugepaged_mm_slot *mm_slot;
 129         unsigned long address;
 130 };
 131
 132 static struct khugepaged_scan khugepaged_scan = {
 133         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 134 };
 135
 136 #ifdef CONFIG_SYSFS
 137 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 138                                          struct kobj_attribute *attr,
 139                                          char *buf)
 140 {
 141         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 142 }
 143
 144 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 145                                           struct kobj_attribute *attr,
 146                                           const char *buf, size_t count)
 147 {
 148         unsigned int msecs;
 149         int err;
 150
 151         err = kstrtouint(buf, 10, &msecs);
 152         if (err)
 153                 return -EINVAL;
 154
 155         khugepaged_scan_sleep_millisecs = msecs;
 156         khugepaged_sleep_expire = 0;
 157         wake_up_interruptible(&khugepaged_wait);
 158
 159         return count;
 160 }
 161 static struct kobj_attribute scan_sleep_millisecs_attr =
 162         __ATTR_RW(scan_sleep_millisecs);
 163
 164 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 165                                           struct kobj_attribute *attr,
 166                                           char *buf)
 167 {
 168         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 169 }
 170
 171 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 172                                            struct kobj_attribute *attr,
 173                                            const char *buf, size_t count)
 174 {
 175         unsigned int msecs;
 176         int err;
 177
 178         err = kstrtouint(buf, 10, &msecs);
 179         if (err)
 180                 return -EINVAL;
 181
 182         khugepaged_alloc_sleep_millisecs = msecs;
 183         khugepaged_sleep_expire = 0;
 184         wake_up_interruptible(&khugepaged_wait);
 185
 186         return count;
 187 }
 188 static struct kobj_attribute alloc_sleep_millisecs_attr =
 189         __ATTR_RW(alloc_sleep_millisecs);
 190
 191 static ssize_t pages_to_scan_show(struct kobject *kobj,
 192                                   struct kobj_attribute *attr,
 193                                   char *buf)
 194 {
 195         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 196 }
 197 static ssize_t pages_to_scan_store(struct kobject *kobj,
 198                                    struct kobj_attribute *attr,
 199                                    const char *buf, size_t count)
 200 {
 201         unsigned int pages;
 202         int err;
 203
 204         err = kstrtouint(buf, 10, &pages);
 205         if (err || !pages)
 206                 return -EINVAL;
 207
 208         khugepaged_pages_to_scan = pages;
 209
 210         return count;
 211 }
 212 static struct kobj_attribute pages_to_scan_attr =
 213         __ATTR_RW(pages_to_scan);
 214
 215 static ssize_t pages_collapsed_show(struct kobject *kobj,
 216                                     struct kobj_attribute *attr,
 217                                     char *buf)
 218 {
 219         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 220 }
 221 static struct kobj_attribute pages_collapsed_attr =
 222         __ATTR_RO(pages_collapsed);
 223
 224 static ssize_t full_scans_show(struct kobject *kobj,
 225                                struct kobj_attribute *attr,
 226                                char *buf)
 227 {
 228         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 229 }
 230 static struct kobj_attribute full_scans_attr =
 231         __ATTR_RO(full_scans);
 232
 233 static ssize_t defrag_show(struct kobject *kobj,
 234                            struct kobj_attribute *attr, char *buf)
 235 {
 236         return single_hugepage_flag_show(kobj, attr, buf,
 237                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 238 }
 239 static ssize_t defrag_store(struct kobject *kobj,
 240                             struct kobj_attribute *attr,
 241                             const char *buf, size_t count)
 242 {
 243         return single_hugepage_flag_store(kobj, attr, buf, count,
 244                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 245 }
 246 static struct kobj_attribute khugepaged_defrag_attr =
 247         __ATTR_RW(defrag);
 248
 249 /*
 250  * max_ptes_none controls if khugepaged should collapse hugepages over
 251  * any unmapped ptes in turn potentially increasing the memory
 252  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 253  * reduce the available free memory in the system as it
 254  * runs. Increasing max_ptes_none will instead potentially reduce the
 255  * free memory in the system during the khugepaged scan.
 256  */
 257 static ssize_t max_ptes_none_show(struct kobject *kobj,
 258                                   struct kobj_attribute *attr,
 259                                   char *buf)
 260 {
 261         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 262 }
 263 static ssize_t max_ptes_none_store(struct kobject *kobj,
 264                                    struct kobj_attribute *attr,
 265                                    const char *buf, size_t count)
 266 {
 267         int err;
 268         unsigned long max_ptes_none;
 269
 270         err = kstrtoul(buf, 10, &max_ptes_none);
 271         if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 272                 return -EINVAL;
 273
 274         khugepaged_max_ptes_none = max_ptes_none;
 275
 276         return count;
 277 }
 278 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 279         __ATTR_RW(max_ptes_none);
 280
 281 static ssize_t max_ptes_swap_show(struct kobject *kobj,
 282                                   struct kobj_attribute *attr,
 283                                   char *buf)
 284 {
 285         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 286 }
 287
 288 static ssize_t max_ptes_swap_store(struct kobject *kobj,
 289                                    struct kobj_attribute *attr,
 290                                    const char *buf, size_t count)
 291 {
 292         int err;
 293         unsigned long max_ptes_swap;
 294
 295         err  = kstrtoul(buf, 10, &max_ptes_swap);
 296         if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 297                 return -EINVAL;
 298
 299         khugepaged_max_ptes_swap = max_ptes_swap;
 300
 301         return count;
 302 }
 303
 304 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 305         __ATTR_RW(max_ptes_swap);
 306
 307 static ssize_t max_ptes_shared_show(struct kobject *kobj,
 308                                     struct kobj_attribute *attr,
 309                                     char *buf)
 310 {
 311         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 312 }
 313
 314 static ssize_t max_ptes_shared_store(struct kobject *kobj,
 315                                      struct kobj_attribute *attr,
 316                                      const char *buf, size_t count)
 317 {
 318         int err;
 319         unsigned long max_ptes_shared;
 320
 321         err  = kstrtoul(buf, 10, &max_ptes_shared);
 322         if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 323                 return -EINVAL;
 324
 325         khugepaged_max_ptes_shared = max_ptes_shared;
 326
 327         return count;
 328 }
 329
 330 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 331         __ATTR_RW(max_ptes_shared);
 332
 333 static struct attribute *khugepaged_attr[] = {
 334         &khugepaged_defrag_attr.attr,
 335         &khugepaged_max_ptes_none_attr.attr,
 336         &khugepaged_max_ptes_swap_attr.attr,
 337         &khugepaged_max_ptes_shared_attr.attr,
 338         &pages_to_scan_attr.attr,
 339         &pages_collapsed_attr.attr,
 340         &full_scans_attr.attr,
 341         &scan_sleep_millisecs_attr.attr,
 342         &alloc_sleep_millisecs_attr.attr,
 343         NULL,
 344 };
 345
 346 struct attribute_group khugepaged_attr_group = {
 347         .attrs = khugepaged_attr,
 348         .name = "khugepaged",
 349 };
 350 #endif /* CONFIG_SYSFS */
 351
 352 int hugepage_madvise(struct vm_area_struct *vma,
 353                      unsigned long *vm_flags, int advice)
 354 {
 355         switch (advice) {
 356         case MADV_HUGEPAGE:
 357 #ifdef CONFIG_S390
 358                 /*
 359                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 360                  * can't handle this properly after s390_enable_sie, so we simply
 361                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 362                  */
 363                 if (mm_has_pgste(vma->vm_mm))
 364                         return 0;
 365 #endif
 366                 *vm_flags &= ~VM_NOHUGEPAGE;
 367                 *vm_flags |= VM_HUGEPAGE;
 368                 /*
 369                  * If the vma become good for khugepaged to scan,
 370                  * register it here without waiting a page fault that
 371                  * may not happen any time soon.
 372                  */
 373                 khugepaged_enter_vma(vma, *vm_flags);
 374                 break;
 375         case MADV_NOHUGEPAGE:
 376                 *vm_flags &= ~VM_HUGEPAGE;
 377                 *vm_flags |= VM_NOHUGEPAGE;
 378                 /*
 379                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 380                  * this vma even if we leave the mm registered in khugepaged if
 381                  * it got registered before VM_NOHUGEPAGE was set.
 382                  */
 383                 break;
 384         }
 385
 386         return 0;
 387 }
 388
 389 int __init khugepaged_init(void)
 390 {
 391         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 392                                           sizeof(struct khugepaged_mm_slot),
 393                                           __alignof__(struct khugepaged_mm_slot),
 394                                           0, NULL);
 395         if (!mm_slot_cache)
 396                 return -ENOMEM;
 397
 398         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 399         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 400         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 401         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 402
 403         return 0;
 404 }
 405
 406 void __init khugepaged_destroy(void)
 407 {
 408         kmem_cache_destroy(mm_slot_cache);
 409 }
 410
 411 static inline int hpage_collapse_test_exit(struct mm_struct *mm)
 412 {
 413         return atomic_read(&mm->mm_users) == 0;
 414 }
 415
 416 void __khugepaged_enter(struct mm_struct *mm)
 417 {
 418         struct khugepaged_mm_slot *mm_slot;
 419         struct mm_slot *slot;
 420         int wakeup;
 421
 422         mm_slot = mm_slot_alloc(mm_slot_cache);
 423         if (!mm_slot)
 424                 return;
 425
 426         slot = &mm_slot->slot;
 427
 428         /* __khugepaged_exit() must not run from under us */
 429         VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 430         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 431                 mm_slot_free(mm_slot_cache, mm_slot);
 432                 return;
 433         }
 434
 435         spin_lock(&khugepaged_mm_lock);
 436         mm_slot_insert(mm_slots_hash, mm, slot);
 437         /*
 438          * Insert just behind the scanning cursor, to let the area settle
 439          * down a little.
 440          */
 441         wakeup = list_empty(&khugepaged_scan.mm_head);
 442         list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 443         spin_unlock(&khugepaged_mm_lock);
 444
 445         mmgrab(mm);
 446         if (wakeup)
 447                 wake_up_interruptible(&khugepaged_wait);
 448 }
 449
 450 void khugepaged_enter_vma(struct vm_area_struct *vma,
 451                           unsigned long vm_flags)
 452 {
 453         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 454             hugepage_flags_enabled()) {
 455                 if (hugepage_vma_check(vma, vm_flags, false, false, true))
 456                         __khugepaged_enter(vma->vm_mm);
 457         }
 458 }
 459
 460 void __khugepaged_exit(struct mm_struct *mm)
 461 {
 462         struct khugepaged_mm_slot *mm_slot;
 463         struct mm_slot *slot;
 464         int free = 0;
 465
 466         spin_lock(&khugepaged_mm_lock);
 467         slot = mm_slot_lookup(mm_slots_hash, mm);
 468         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
 469         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 470                 hash_del(&slot->hash);
 471                 list_del(&slot->mm_node);
 472                 free = 1;
 473         }
 474         spin_unlock(&khugepaged_mm_lock);
 475
 476         if (free) {
 477                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 478                 mm_slot_free(mm_slot_cache, mm_slot);
 479                 mmdrop(mm);
 480         } else if (mm_slot) {
 481                 /*
 482                  * This is required to serialize against
 483                  * hpage_collapse_test_exit() (which is guaranteed to run
 484                  * under mmap sem read mode). Stop here (after we return all
 485                  * pagetables will be destroyed) until khugepaged has finished
 486                  * working on the pagetables under the mmap_lock.
 487                  */
 488                 mmap_write_lock(mm);
 489                 mmap_write_unlock(mm);
 490         }
 491 }
 492
 493 static void release_pte_folio(struct folio *folio)
 494 {
 495         node_stat_mod_folio(folio,
 496                         NR_ISOLATED_ANON + folio_is_file_lru(folio),
 497                         -folio_nr_pages(folio));
 498         folio_unlock(folio);
 499         folio_putback_lru(folio);
 500 }
 501
 502 static void release_pte_page(struct page *page)
 503 {
 504         release_pte_folio(page_folio(page));
 505 }
 506
 507 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 508                 struct list_head *compound_pagelist)
 509 {
 510         struct folio *folio, *tmp;
 511
 512         while (--_pte >= pte) {
 513                 pte_t pteval = *_pte;
 514                 unsigned long pfn;
 515
 516                 if (pte_none(pteval))
 517                         continue;
 518                 pfn = pte_pfn(pteval);
 519                 if (is_zero_pfn(pfn))
 520                         continue;
 521                 folio = pfn_folio(pfn);
 522                 if (folio_test_large(folio))
 523                         continue;
 524                 release_pte_folio(folio);
 525         }
 526
 527         list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 528                 list_del(&folio->lru);
 529                 release_pte_folio(folio);
 530         }
 531 }
 532
 533 static bool is_refcount_suitable(struct page *page)
 534 {
 535         int expected_refcount;
 536
 537         expected_refcount = total_mapcount(page);
 538         if (PageSwapCache(page))
 539                 expected_refcount += compound_nr(page);
 540
 541         return page_count(page) == expected_refcount;
 542 }
 543
 544 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 545                                         unsigned long address,
 546                                         pte_t *pte,
 547                                         struct collapse_control *cc,
 548                                         struct list_head *compound_pagelist)
 549 {
 550         struct page *page = NULL;
 551         pte_t *_pte;
 552         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 553         bool writable = false;
 554
 555         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 556              _pte++, address += PAGE_SIZE) {
 557                 pte_t pteval = *_pte;
 558                 if (pte_none(pteval) || (pte_present(pteval) &&
 559                                 is_zero_pfn(pte_pfn(pteval)))) {
 560                         ++none_or_zero;
 561                         if (!userfaultfd_armed(vma) &&
 562                             (!cc->is_khugepaged ||
 563                              none_or_zero <= khugepaged_max_ptes_none)) {
 564                                 continue;
 565                         } else {
 566                                 result = SCAN_EXCEED_NONE_PTE;
 567                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 568                                 goto out;
 569                         }
 570                 }
 571                 if (!pte_present(pteval)) {
 572                         result = SCAN_PTE_NON_PRESENT;
 573                         goto out;
 574                 }
 575                 if (pte_uffd_wp(pteval)) {
 576                         result = SCAN_PTE_UFFD_WP;
 577                         goto out;
 578                 }
 579                 page = vm_normal_page(vma, address, pteval);
 580                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 581                         result = SCAN_PAGE_NULL;
 582                         goto out;
 583                 }
 584
 585                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 586
 587                 if (page_mapcount(page) > 1) {
 588                         ++shared;
 589                         if (cc->is_khugepaged &&
 590                             shared > khugepaged_max_ptes_shared) {
 591                                 result = SCAN_EXCEED_SHARED_PTE;
 592                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 593                                 goto out;
 594                         }
 595                 }
 596
 597                 if (PageCompound(page)) {
 598                         struct page *p;
 599                         page = compound_head(page);
 600
 601                         /*
 602                          * Check if we have dealt with the compound page
 603                          * already
 604                          */
 605                         list_for_each_entry(p, compound_pagelist, lru) {
 606                                 if (page == p)
 607                                         goto next;
 608                         }
 609                 }
 610
 611                 /*
 612                  * We can do it before isolate_lru_page because the
 613                  * page can't be freed from under us. NOTE: PG_lock
 614                  * is needed to serialize against split_huge_page
 615                  * when invoked from the VM.
 616                  */
 617                 if (!trylock_page(page)) {
 618                         result = SCAN_PAGE_LOCK;
 619                         goto out;
 620                 }
 621
 622                 /*
 623                  * Check if the page has any GUP (or other external) pins.
 624                  *
 625                  * The page table that maps the page has been already unlinked
 626                  * from the page table tree and this process cannot get
 627                  * an additional pin on the page.
 628                  *
 629                  * New pins can come later if the page is shared across fork,
 630                  * but not from this process. The other process cannot write to
 631                  * the page, only trigger CoW.
 632                  */
 633                 if (!is_refcount_suitable(page)) {
 634                         unlock_page(page);
 635                         result = SCAN_PAGE_COUNT;
 636                         goto out;
 637                 }
 638
 639                 /*
 640                  * Isolate the page to avoid collapsing an hugepage
 641                  * currently in use by the VM.
 642                  */
 643                 if (!isolate_lru_page(page)) {
 644                         unlock_page(page);
 645                         result = SCAN_DEL_PAGE_LRU;
 646                         goto out;
 647                 }
 648                 mod_node_page_state(page_pgdat(page),
 649                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 650                                 compound_nr(page));
 651                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 652                 VM_BUG_ON_PAGE(PageLRU(page), page);
 653
 654                 if (PageCompound(page))
 655                         list_add_tail(&page->lru, compound_pagelist);
 656 next:
 657                 /*
 658                  * If collapse was initiated by khugepaged, check that there is
 659                  * enough young pte to justify collapsing the page
 660                  */
 661                 if (cc->is_khugepaged &&
 662                     (pte_young(pteval) || page_is_young(page) ||
 663                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
 664                                                                      address)))
 665                         referenced++;
 666
 667                 if (pte_write(pteval))
 668                         writable = true;
 669         }
 670
 671         if (unlikely(!writable)) {
 672                 result = SCAN_PAGE_RO;
 673         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 674                 result = SCAN_LACK_REFERENCED_PAGE;
 675         } else {
 676                 result = SCAN_SUCCEED;
 677                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 678                                                     referenced, writable, result);
 679                 return result;
 680         }
 681 out:
 682         release_pte_pages(pte, _pte, compound_pagelist);
 683         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 684                                             referenced, writable, result);
 685         return result;
 686 }
 687
 688 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 689                                       struct vm_area_struct *vma,
 690                                       unsigned long address,
 691                                       spinlock_t *ptl,
 692                                       struct list_head *compound_pagelist)
 693 {
 694         struct page *src_page, *tmp;
 695         pte_t *_pte;
 696         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 697                                 _pte++, page++, address += PAGE_SIZE) {
 698                 pte_t pteval = *_pte;
 699
 700                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 701                         clear_user_highpage(page, address);
 702                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 703                         if (is_zero_pfn(pte_pfn(pteval))) {
 704                                 /*
 705                                  * ptl mostly unnecessary.
 706                                  */
 707                                 spin_lock(ptl);
 708                                 ptep_clear(vma->vm_mm, address, _pte);
 709                                 spin_unlock(ptl);
 710                         }
 711                 } else {
 712                         src_page = pte_page(pteval);
 713                         copy_user_highpage(page, src_page, address, vma);
 714                         if (!PageCompound(src_page))
 715                                 release_pte_page(src_page);
 716                         /*
 717                          * ptl mostly unnecessary, but preempt has to
 718                          * be disabled to update the per-cpu stats
 719                          * inside page_remove_rmap().
 720                          */
 721                         spin_lock(ptl);
 722                         ptep_clear(vma->vm_mm, address, _pte);
 723                         page_remove_rmap(src_page, vma, false);
 724                         spin_unlock(ptl);
 725                         free_page_and_swap_cache(src_page);
 726                 }
 727         }
 728
 729         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 730                 list_del(&src_page->lru);
 731                 mod_node_page_state(page_pgdat(src_page),
 732                                     NR_ISOLATED_ANON + page_is_file_lru(src_page),
 733                                     -compound_nr(src_page));
 734                 unlock_page(src_page);
 735                 free_swap_cache(src_page);
 736                 putback_lru_page(src_page);
 737         }
 738 }
 739
 740 static void khugepaged_alloc_sleep(void)
 741 {
 742         DEFINE_WAIT(wait);
 743
 744         add_wait_queue(&khugepaged_wait, &wait);
 745         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 746         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 747         remove_wait_queue(&khugepaged_wait, &wait);
 748 }
 749
 750 struct collapse_control khugepaged_collapse_control = {
 751         .is_khugepaged = true,
 752 };
 753
 754 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 755 {
 756         int i;
 757
 758         /*
 759          * If node_reclaim_mode is disabled, then no extra effort is made to
 760          * allocate memory locally.
 761          */
 762         if (!node_reclaim_enabled())
 763                 return false;
 764
 765         /* If there is a count for this node already, it must be acceptable */
 766         if (cc->node_load[nid])
 767                 return false;
 768
 769         for (i = 0; i < MAX_NUMNODES; i++) {
 770                 if (!cc->node_load[i])
 771                         continue;
 772                 if (node_distance(nid, i) > node_reclaim_distance)
 773                         return true;
 774         }
 775         return false;
 776 }
 777
 778 #define khugepaged_defrag()                                     \
 779         (transparent_hugepage_flags &                           \
 780          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 781
 782 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 783 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 784 {
 785         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 786 }
 787
 788 #ifdef CONFIG_NUMA
 789 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 790 {
 791         int nid, target_node = 0, max_value = 0;
 792
 793         /* find first node with max normal pages hit */
 794         for (nid = 0; nid < MAX_NUMNODES; nid++)
 795                 if (cc->node_load[nid] > max_value) {
 796                         max_value = cc->node_load[nid];
 797                         target_node = nid;
 798                 }
 799
 800         for_each_online_node(nid) {
 801                 if (max_value == cc->node_load[nid])
 802                         node_set(nid, cc->alloc_nmask);
 803         }
 804
 805         return target_node;
 806 }
 807 #else
 808 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 809 {
 810         return 0;
 811 }
 812 #endif
 813
 814 static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node,
 815                                       nodemask_t *nmask)
 816 {
 817         *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask);
 818         if (unlikely(!*hpage)) {
 819                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 820                 return false;
 821         }
 822
 823         prep_transhuge_page(*hpage);
 824         count_vm_event(THP_COLLAPSE_ALLOC);
 825         return true;
 826 }
 827
 828 /*
 829  * If mmap_lock temporarily dropped, revalidate vma
 830  * before taking mmap_lock.
 831  * Returns enum scan_result value.
 832  */
 833
 834 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 835                                    bool expect_anon,
 836                                    struct vm_area_struct **vmap,
 837                                    struct collapse_control *cc)
 838 {
 839         struct vm_area_struct *vma;
 840
 841         if (unlikely(hpage_collapse_test_exit(mm)))
 842                 return SCAN_ANY_PROCESS;
 843
 844         *vmap = vma = find_vma(mm, address);
 845         if (!vma)
 846                 return SCAN_VMA_NULL;
 847
 848         if (!transhuge_vma_suitable(vma, address))
 849                 return SCAN_ADDRESS_RANGE;
 850         if (!hugepage_vma_check(vma, vma->vm_flags, false, false,
 851                                 cc->is_khugepaged))
 852                 return SCAN_VMA_CHECK;
 853         /*
 854          * Anon VMA expected, the address may be unmapped then
 855          * remapped to file after khugepaged reaquired the mmap_lock.
 856          *
 857          * hugepage_vma_check may return true for qualified file
 858          * vmas.
 859          */
 860         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 861                 return SCAN_PAGE_ANON;
 862         return SCAN_SUCCEED;
 863 }
 864
 865 /*
 866  * See pmd_trans_unstable() for how the result may change out from
 867  * underneath us, even if we hold mmap_lock in read.
 868  */
 869 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 870                                    unsigned long address,
 871                                    pmd_t **pmd)
 872 {
 873         pmd_t pmde;
 874
 875         *pmd = mm_find_pmd(mm, address);
 876         if (!*pmd)
 877                 return SCAN_PMD_NULL;
 878
 879         pmde = pmdp_get_lockless(*pmd);
 880
 881 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 882         /* See comments in pmd_none_or_trans_huge_or_clear_bad() */
 883         barrier();
 884 #endif
 885         if (pmd_none(pmde))
 886                 return SCAN_PMD_NONE;
 887         if (!pmd_present(pmde))
 888                 return SCAN_PMD_NULL;
 889         if (pmd_trans_huge(pmde))
 890                 return SCAN_PMD_MAPPED;
 891         if (pmd_devmap(pmde))
 892                 return SCAN_PMD_NULL;
 893         if (pmd_bad(pmde))
 894                 return SCAN_PMD_NULL;
 895         return SCAN_SUCCEED;
 896 }
 897
 898 static int check_pmd_still_valid(struct mm_struct *mm,
 899                                  unsigned long address,
 900                                  pmd_t *pmd)
 901 {
 902         pmd_t *new_pmd;
 903         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 904
 905         if (result != SCAN_SUCCEED)
 906                 return result;
 907         if (new_pmd != pmd)
 908                 return SCAN_FAIL;
 909         return SCAN_SUCCEED;
 910 }
 911
 912 /*
 913  * Bring missing pages in from swap, to complete THP collapse.
 914  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 915  *
 916  * Called and returns without pte mapped or spinlocks held.
 917  * Note that if false is returned, mmap_lock will be released.
 918  */
 919
 920 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 921                                        struct vm_area_struct *vma,
 922                                        unsigned long haddr, pmd_t *pmd,
 923                                        int referenced)
 924 {
 925         int swapped_in = 0;
 926         vm_fault_t ret = 0;
 927         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 928
 929         for (address = haddr; address < end; address += PAGE_SIZE) {
 930                 struct vm_fault vmf = {
 931                         .vma = vma,
 932                         .address = address,
 933                         .pgoff = linear_page_index(vma, haddr),
 934                         .flags = FAULT_FLAG_ALLOW_RETRY,
 935                         .pmd = pmd,
 936                 };
 937
 938                 vmf.pte = pte_offset_map(pmd, address);
 939                 vmf.orig_pte = *vmf.pte;
 940                 if (!is_swap_pte(vmf.orig_pte)) {
 941                         pte_unmap(vmf.pte);
 942                         continue;
 943                 }
 944                 ret = do_swap_page(&vmf);
 945
 946                 /*
 947                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
 948                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
 949                  * we do not retry here and swap entry will remain in pagetable
 950                  * resulting in later failure.
 951                  */
 952                 if (ret & VM_FAULT_RETRY) {
 953                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 954                         /* Likely, but not guaranteed, that page lock failed */
 955                         return SCAN_PAGE_LOCK;
 956                 }
 957                 if (ret & VM_FAULT_ERROR) {
 958                         mmap_read_unlock(mm);
 959                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 960                         return SCAN_FAIL;
 961                 }
 962                 swapped_in++;
 963         }
 964
 965         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
 966         if (swapped_in)
 967                 lru_add_drain();
 968
 969         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 970         return SCAN_SUCCEED;
 971 }
 972
 973 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
 974                               struct collapse_control *cc)
 975 {
 976         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
 977                      GFP_TRANSHUGE);
 978         int node = hpage_collapse_find_target_node(cc);
 979
 980         if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask))
 981                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
 982         if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp)))
 983                 return SCAN_CGROUP_CHARGE_FAIL;
 984         count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC);
 985         return SCAN_SUCCEED;
 986 }
 987
 988 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
 989                               int referenced, int unmapped,
 990                               struct collapse_control *cc)
 991 {
 992         LIST_HEAD(compound_pagelist);
 993         pmd_t *pmd, _pmd;
 994         pte_t *pte;
 995         pgtable_t pgtable;
 996         struct page *hpage;
 997         spinlock_t *pmd_ptl, *pte_ptl;
 998         int result = SCAN_FAIL;
 999         struct vm_area_struct *vma;
1000         struct mmu_notifier_range range;
1001
1002         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1003
1004         /*
1005          * Before allocating the hugepage, release the mmap_lock read lock.
1006          * The allocation can take potentially a long time if it involves
1007          * sync compaction, and we do not need to hold the mmap_lock during
1008          * that. We will recheck the vma after taking it again in write mode.
1009          */
1010         mmap_read_unlock(mm);
1011
1012         result = alloc_charge_hpage(&hpage, mm, cc);
1013         if (result != SCAN_SUCCEED)
1014                 goto out_nolock;
1015
1016         mmap_read_lock(mm);
1017         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1018         if (result != SCAN_SUCCEED) {
1019                 mmap_read_unlock(mm);
1020                 goto out_nolock;
1021         }
1022
1023         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1024         if (result != SCAN_SUCCEED) {
1025                 mmap_read_unlock(mm);
1026                 goto out_nolock;
1027         }
1028
1029         if (unmapped) {
1030                 /*
1031                  * __collapse_huge_page_swapin will return with mmap_lock
1032                  * released when it fails. So we jump out_nolock directly in
1033                  * that case.  Continuing to collapse causes inconsistency.
1034                  */
1035                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1036                                                      referenced);
1037                 if (result != SCAN_SUCCEED)
1038                         goto out_nolock;
1039         }
1040
1041         mmap_read_unlock(mm);
1042         /*
1043          * Prevent all access to pagetables with the exception of
1044          * gup_fast later handled by the ptep_clear_flush and the VM
1045          * handled by the anon_vma lock + PG_lock.
1046          */
1047         mmap_write_lock(mm);
1048         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1049         if (result != SCAN_SUCCEED)
1050                 goto out_up_write;
1051         /* check if the pmd is still valid */
1052         result = check_pmd_still_valid(mm, address, pmd);
1053         if (result != SCAN_SUCCEED)
1054                 goto out_up_write;
1055
1056         anon_vma_lock_write(vma->anon_vma);
1057
1058         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1059                                 address + HPAGE_PMD_SIZE);
1060         mmu_notifier_invalidate_range_start(&range);
1061
1062         pte = pte_offset_map(pmd, address);
1063         pte_ptl = pte_lockptr(mm, pmd);
1064
1065         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1066         /*
1067          * This removes any huge TLB entry from the CPU so we won't allow
1068          * huge and small TLB entries for the same virtual address to
1069          * avoid the risk of CPU bugs in that area.
1070          *
1071          * Parallel fast GUP is fine since fast GUP will back off when
1072          * it detects PMD is changed.
1073          */
1074         _pmd = pmdp_collapse_flush(vma, address, pmd);
1075         spin_unlock(pmd_ptl);
1076         mmu_notifier_invalidate_range_end(&range);
1077         tlb_remove_table_sync_one();
1078
1079         spin_lock(pte_ptl);
1080         result =  __collapse_huge_page_isolate(vma, address, pte, cc,
1081                                                &compound_pagelist);
1082         spin_unlock(pte_ptl);
1083
1084         if (unlikely(result != SCAN_SUCCEED)) {
1085                 pte_unmap(pte);
1086                 spin_lock(pmd_ptl);
1087                 BUG_ON(!pmd_none(*pmd));
1088                 /*
1089                  * We can only use set_pmd_at when establishing
1090                  * hugepmds and never for establishing regular pmds that
1091                  * points to regular pagetables. Use pmd_populate for that
1092                  */
1093                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1094                 spin_unlock(pmd_ptl);
1095                 anon_vma_unlock_write(vma->anon_vma);
1096                 goto out_up_write;
1097         }
1098
1099         /*
1100          * All pages are isolated and locked so anon_vma rmap
1101          * can't run anymore.
1102          */
1103         anon_vma_unlock_write(vma->anon_vma);
1104
1105         __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl,
1106                                   &compound_pagelist);
1107         pte_unmap(pte);
1108         /*
1109          * spin_lock() below is not the equivalent of smp_wmb(), but
1110          * the smp_wmb() inside __SetPageUptodate() can be reused to
1111          * avoid the copy_huge_page writes to become visible after
1112          * the set_pmd_at() write.
1113          */
1114         __SetPageUptodate(hpage);
1115         pgtable = pmd_pgtable(_pmd);
1116
1117         _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1118         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1119
1120         spin_lock(pmd_ptl);
1121         BUG_ON(!pmd_none(*pmd));
1122         page_add_new_anon_rmap(hpage, vma, address);
1123         lru_cache_add_inactive_or_unevictable(hpage, vma);
1124         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1125         set_pmd_at(mm, address, pmd, _pmd);
1126         update_mmu_cache_pmd(vma, address, pmd);
1127         spin_unlock(pmd_ptl);
1128
1129         hpage = NULL;
1130
1131         result = SCAN_SUCCEED;
1132 out_up_write:
1133         mmap_write_unlock(mm);
1134 out_nolock:
1135         if (hpage) {
1136                 mem_cgroup_uncharge(page_folio(hpage));
1137                 put_page(hpage);
1138         }
1139         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1140         return result;
1141 }
1142
1143 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1144                                    struct vm_area_struct *vma,
1145                                    unsigned long address, bool *mmap_locked,
1146                                    struct collapse_control *cc)
1147 {
1148         pmd_t *pmd;
1149         pte_t *pte, *_pte;
1150         int result = SCAN_FAIL, referenced = 0;
1151         int none_or_zero = 0, shared = 0;
1152         struct page *page = NULL;
1153         unsigned long _address;
1154         spinlock_t *ptl;
1155         int node = NUMA_NO_NODE, unmapped = 0;
1156         bool writable = false;
1157
1158         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1159
1160         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1161         if (result != SCAN_SUCCEED)
1162                 goto out;
1163
1164         memset(cc->node_load, 0, sizeof(cc->node_load));
1165         nodes_clear(cc->alloc_nmask);
1166         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1167         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1168              _pte++, _address += PAGE_SIZE) {
1169                 pte_t pteval = *_pte;
1170                 if (is_swap_pte(pteval)) {
1171                         ++unmapped;
1172                         if (!cc->is_khugepaged ||
1173                             unmapped <= khugepaged_max_ptes_swap) {
1174                                 /*
1175                                  * Always be strict with uffd-wp
1176                                  * enabled swap entries.  Please see
1177                                  * comment below for pte_uffd_wp().
1178                                  */
1179                                 if (pte_swp_uffd_wp(pteval)) {
1180                                         result = SCAN_PTE_UFFD_WP;
1181                                         goto out_unmap;
1182                                 }
1183                                 continue;
1184                         } else {
1185                                 result = SCAN_EXCEED_SWAP_PTE;
1186                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1187                                 goto out_unmap;
1188                         }
1189                 }
1190                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1191                         ++none_or_zero;
1192                         if (!userfaultfd_armed(vma) &&
1193                             (!cc->is_khugepaged ||
1194                              none_or_zero <= khugepaged_max_ptes_none)) {
1195                                 continue;
1196                         } else {
1197                                 result = SCAN_EXCEED_NONE_PTE;
1198                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1199                                 goto out_unmap;
1200                         }
1201                 }
1202                 if (pte_uffd_wp(pteval)) {
1203                         /*
1204                          * Don't collapse the page if any of the small
1205                          * PTEs are armed with uffd write protection.
1206                          * Here we can also mark the new huge pmd as
1207                          * write protected if any of the small ones is
1208                          * marked but that could bring unknown
1209                          * userfault messages that falls outside of
1210                          * the registered range.  So, just be simple.
1211                          */
1212                         result = SCAN_PTE_UFFD_WP;
1213                         goto out_unmap;
1214                 }
1215                 if (pte_write(pteval))
1216                         writable = true;
1217
1218                 page = vm_normal_page(vma, _address, pteval);
1219                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1220                         result = SCAN_PAGE_NULL;
1221                         goto out_unmap;
1222                 }
1223
1224                 if (page_mapcount(page) > 1) {
1225                         ++shared;
1226                         if (cc->is_khugepaged &&
1227                             shared > khugepaged_max_ptes_shared) {
1228                                 result = SCAN_EXCEED_SHARED_PTE;
1229                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1230                                 goto out_unmap;
1231                         }
1232                 }
1233
1234                 page = compound_head(page);
1235
1236                 /*
1237                  * Record which node the original page is from and save this
1238                  * information to cc->node_load[].
1239                  * Khugepaged will allocate hugepage from the node has the max
1240                  * hit record.
1241                  */
1242                 node = page_to_nid(page);
1243                 if (hpage_collapse_scan_abort(node, cc)) {
1244                         result = SCAN_SCAN_ABORT;
1245                         goto out_unmap;
1246                 }
1247                 cc->node_load[node]++;
1248                 if (!PageLRU(page)) {
1249                         result = SCAN_PAGE_LRU;
1250                         goto out_unmap;
1251                 }
1252                 if (PageLocked(page)) {
1253                         result = SCAN_PAGE_LOCK;
1254                         goto out_unmap;
1255                 }
1256                 if (!PageAnon(page)) {
1257                         result = SCAN_PAGE_ANON;
1258                         goto out_unmap;
1259                 }
1260
1261                 /*
1262                  * Check if the page has any GUP (or other external) pins.
1263                  *
1264                  * Here the check may be racy:
1265                  * it may see total_mapcount > refcount in some cases?
1266                  * But such case is ephemeral we could always retry collapse
1267                  * later.  However it may report false positive if the page
1268                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1269                  * will be done again later the risk seems low.
1270                  */
1271                 if (!is_refcount_suitable(page)) {
1272                         result = SCAN_PAGE_COUNT;
1273                         goto out_unmap;
1274                 }
1275
1276                 /*
1277                  * If collapse was initiated by khugepaged, check that there is
1278                  * enough young pte to justify collapsing the page
1279                  */
1280                 if (cc->is_khugepaged &&
1281                     (pte_young(pteval) || page_is_young(page) ||
1282                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
1283                                                                      address)))
1284                         referenced++;
1285         }
1286         if (!writable) {
1287                 result = SCAN_PAGE_RO;
1288         } else if (cc->is_khugepaged &&
1289                    (!referenced ||
1290                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1291                 result = SCAN_LACK_REFERENCED_PAGE;
1292         } else {
1293                 result = SCAN_SUCCEED;
1294         }
1295 out_unmap:
1296         pte_unmap_unlock(pte, ptl);
1297         if (result == SCAN_SUCCEED) {
1298                 result = collapse_huge_page(mm, address, referenced,
1299                                             unmapped, cc);
1300                 /* collapse_huge_page will return with the mmap_lock released */
1301                 *mmap_locked = false;
1302         }
1303 out:
1304         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1305                                      none_or_zero, result, unmapped);
1306         return result;
1307 }
1308
1309 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1310 {
1311         struct mm_slot *slot = &mm_slot->slot;
1312         struct mm_struct *mm = slot->mm;
1313
1314         lockdep_assert_held(&khugepaged_mm_lock);
1315
1316         if (hpage_collapse_test_exit(mm)) {
1317                 /* free mm_slot */
1318                 hash_del(&slot->hash);
1319                 list_del(&slot->mm_node);
1320
1321                 /*
1322                  * Not strictly needed because the mm exited already.
1323                  *
1324                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1325                  */
1326
1327                 /* khugepaged_mm_lock actually not necessary for the below */
1328                 mm_slot_free(mm_slot_cache, mm_slot);
1329                 mmdrop(mm);
1330         }
1331 }
1332
1333 #ifdef CONFIG_SHMEM
1334 /*
1335  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1336  * khugepaged should try to collapse the page table.
1337  *
1338  * Note that following race exists:
1339  * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A,
1340  *     emptying the A's ->pte_mapped_thp[] array.
1341  * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and
1342  *     retract_page_tables() finds a VMA in mm_struct A mapping the same extent
1343  *     (at virtual address X) and adds an entry (for X) into mm_struct A's
1344  *     ->pte-mapped_thp[] array.
1345  * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X,
1346  *     sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry
1347  *     (for X) into mm_struct A's ->pte-mapped_thp[] array.
1348  * Thus, it's possible the same address is added multiple times for the same
1349  * mm_struct.  Should this happen, we'll simply attempt
1350  * collapse_pte_mapped_thp() multiple times for the same address, under the same
1351  * exclusive mmap_lock, and assuming the first call is successful, subsequent
1352  * attempts will return quickly (without grabbing any additional locks) when
1353  * a huge pmd is found in find_pmd_or_thp_or_none().  Since this is a cheap
1354  * check, and since this is a rare occurrence, the cost of preventing this
1355  * "multiple-add" is thought to be more expensive than just handling it, should
1356  * it occur.
1357  */
1358 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1359                                           unsigned long addr)
1360 {
1361         struct khugepaged_mm_slot *mm_slot;
1362         struct mm_slot *slot;
1363         bool ret = false;
1364
1365         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1366
1367         spin_lock(&khugepaged_mm_lock);
1368         slot = mm_slot_lookup(mm_slots_hash, mm);
1369         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
1370         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) {
1371                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1372                 ret = true;
1373         }
1374         spin_unlock(&khugepaged_mm_lock);
1375         return ret;
1376 }
1377
1378 /* hpage must be locked, and mmap_lock must be held in write */
1379 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1380                         pmd_t *pmdp, struct page *hpage)
1381 {
1382         struct vm_fault vmf = {
1383                 .vma = vma,
1384                 .address = addr,
1385                 .flags = 0,
1386                 .pmd = pmdp,
1387         };
1388
1389         VM_BUG_ON(!PageTransHuge(hpage));
1390         mmap_assert_write_locked(vma->vm_mm);
1391
1392         if (do_set_pmd(&vmf, hpage))
1393                 return SCAN_FAIL;
1394
1395         get_page(hpage);
1396         return SCAN_SUCCEED;
1397 }
1398
1399 /*
1400  * A note about locking:
1401  * Trying to take the page table spinlocks would be useless here because those
1402  * are only used to synchronize:
1403  *
1404  *  - modifying terminal entries (ones that point to a data page, not to another
1405  *    page table)
1406  *  - installing *new* non-terminal entries
1407  *
1408  * Instead, we need roughly the same kind of protection as free_pgtables() or
1409  * mm_take_all_locks() (but only for a single VMA):
1410  * The mmap lock together with this VMA's rmap locks covers all paths towards
1411  * the page table entries we're messing with here, except for hardware page
1412  * table walks and lockless_pages_from_mm().
1413  */
1414 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1415                                   unsigned long addr, pmd_t *pmdp)
1416 {
1417         pmd_t pmd;
1418         struct mmu_notifier_range range;
1419
1420         mmap_assert_write_locked(mm);
1421         if (vma->vm_file)
1422                 lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem);
1423         /*
1424          * All anon_vmas attached to the VMA have the same root and are
1425          * therefore locked by the same lock.
1426          */
1427         if (vma->anon_vma)
1428                 lockdep_assert_held_write(&vma->anon_vma->root->rwsem);
1429
1430         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr,
1431                                 addr + HPAGE_PMD_SIZE);
1432         mmu_notifier_invalidate_range_start(&range);
1433         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1434         tlb_remove_table_sync_one();
1435         mmu_notifier_invalidate_range_end(&range);
1436         mm_dec_nr_ptes(mm);
1437         page_table_check_pte_clear_range(mm, addr, pmd);
1438         pte_free(mm, pmd_pgtable(pmd));
1439 }
1440
1441 /**
1442  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1443  * address haddr.
1444  *
1445  * @mm: process address space where collapse happens
1446  * @addr: THP collapse address
1447  * @install_pmd: If a huge PMD should be installed
1448  *
1449  * This function checks whether all the PTEs in the PMD are pointing to the
1450  * right THP. If so, retract the page table so the THP can refault in with
1451  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1452  */
1453 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1454                             bool install_pmd)
1455 {
1456         unsigned long haddr = addr & HPAGE_PMD_MASK;
1457         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1458         struct page *hpage;
1459         pte_t *start_pte, *pte;
1460         pmd_t *pmd;
1461         spinlock_t *ptl;
1462         int count = 0, result = SCAN_FAIL;
1463         int i;
1464
1465         mmap_assert_write_locked(mm);
1466
1467         /* Fast check before locking page if already PMD-mapped */
1468         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1469         if (result == SCAN_PMD_MAPPED)
1470                 return result;
1471
1472         if (!vma || !vma->vm_file ||
1473             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1474                 return SCAN_VMA_CHECK;
1475
1476         /*
1477          * If we are here, we've succeeded in replacing all the native pages
1478          * in the page cache with a single hugepage. If a mm were to fault-in
1479          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1480          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1481          * analogously elide sysfs THP settings here.
1482          */
1483         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
1484                 return SCAN_VMA_CHECK;
1485
1486         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1487         if (userfaultfd_wp(vma))
1488                 return SCAN_PTE_UFFD_WP;
1489
1490         hpage = find_lock_page(vma->vm_file->f_mapping,
1491                                linear_page_index(vma, haddr));
1492         if (!hpage)
1493                 return SCAN_PAGE_NULL;
1494
1495         if (!PageHead(hpage)) {
1496                 result = SCAN_FAIL;
1497                 goto drop_hpage;
1498         }
1499
1500         if (compound_order(hpage) != HPAGE_PMD_ORDER) {
1501                 result = SCAN_PAGE_COMPOUND;
1502                 goto drop_hpage;
1503         }
1504
1505         switch (result) {
1506         case SCAN_SUCCEED:
1507                 break;
1508         case SCAN_PMD_NONE:
1509                 /*
1510                  * In MADV_COLLAPSE path, possible race with khugepaged where
1511                  * all pte entries have been removed and pmd cleared.  If so,
1512                  * skip all the pte checks and just update the pmd mapping.
1513                  */
1514                 goto maybe_install_pmd;
1515         default:
1516                 goto drop_hpage;
1517         }
1518
1519         /*
1520          * We need to lock the mapping so that from here on, only GUP-fast and
1521          * hardware page walks can access the parts of the page tables that
1522          * we're operating on.
1523          * See collapse_and_free_pmd().
1524          */
1525         i_mmap_lock_write(vma->vm_file->f_mapping);
1526
1527         /*
1528          * This spinlock should be unnecessary: Nobody else should be accessing
1529          * the page tables under spinlock protection here, only
1530          * lockless_pages_from_mm() and the hardware page walker can access page
1531          * tables while all the high-level locks are held in write mode.
1532          */
1533         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1534         result = SCAN_FAIL;
1535
1536         /* step 1: check all mapped PTEs are to the right huge page */
1537         for (i = 0, addr = haddr, pte = start_pte;
1538              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1539                 struct page *page;
1540
1541                 /* empty pte, skip */
1542                 if (pte_none(*pte))
1543                         continue;
1544
1545                 /* page swapped out, abort */
1546                 if (!pte_present(*pte)) {
1547                         result = SCAN_PTE_NON_PRESENT;
1548                         goto abort;
1549                 }
1550
1551                 page = vm_normal_page(vma, addr, *pte);
1552                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1553                         page = NULL;
1554                 /*
1555                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1556                  * page table, but the new page will not be a subpage of hpage.
1557                  */
1558                 if (hpage + i != page)
1559                         goto abort;
1560                 count++;
1561         }
1562
1563         /* step 2: adjust rmap */
1564         for (i = 0, addr = haddr, pte = start_pte;
1565              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1566                 struct page *page;
1567
1568                 if (pte_none(*pte))
1569                         continue;
1570                 page = vm_normal_page(vma, addr, *pte);
1571                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1572                         goto abort;
1573                 page_remove_rmap(page, vma, false);
1574         }
1575
1576         pte_unmap_unlock(start_pte, ptl);
1577
1578         /* step 3: set proper refcount and mm_counters. */
1579         if (count) {
1580                 page_ref_sub(hpage, count);
1581                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1582         }
1583
1584         /* step 4: remove pte entries */
1585         /* we make no change to anon, but protect concurrent anon page lookup */
1586         if (vma->anon_vma)
1587                 anon_vma_lock_write(vma->anon_vma);
1588
1589         collapse_and_free_pmd(mm, vma, haddr, pmd);
1590
1591         if (vma->anon_vma)
1592                 anon_vma_unlock_write(vma->anon_vma);
1593         i_mmap_unlock_write(vma->vm_file->f_mapping);
1594
1595 maybe_install_pmd:
1596         /* step 5: install pmd entry */
1597         result = install_pmd
1598                         ? set_huge_pmd(vma, haddr, pmd, hpage)
1599                         : SCAN_SUCCEED;
1600
1601 drop_hpage:
1602         unlock_page(hpage);
1603         put_page(hpage);
1604         return result;
1605
1606 abort:
1607         pte_unmap_unlock(start_pte, ptl);
1608         i_mmap_unlock_write(vma->vm_file->f_mapping);
1609         goto drop_hpage;
1610 }
1611
1612 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
1613 {
1614         struct mm_slot *slot = &mm_slot->slot;
1615         struct mm_struct *mm = slot->mm;
1616         int i;
1617
1618         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1619                 return;
1620
1621         if (!mmap_write_trylock(mm))
1622                 return;
1623
1624         if (unlikely(hpage_collapse_test_exit(mm)))
1625                 goto out;
1626
1627         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1628                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false);
1629
1630 out:
1631         mm_slot->nr_pte_mapped_thp = 0;
1632         mmap_write_unlock(mm);
1633 }
1634
1635 static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
1636                                struct mm_struct *target_mm,
1637                                unsigned long target_addr, struct page *hpage,
1638                                struct collapse_control *cc)
1639 {
1640         struct vm_area_struct *vma;
1641         int target_result = SCAN_FAIL;
1642
1643         i_mmap_lock_write(mapping);
1644         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1645                 int result = SCAN_FAIL;
1646                 struct mm_struct *mm = NULL;
1647                 unsigned long addr = 0;
1648                 pmd_t *pmd;
1649                 bool is_target = false;
1650
1651                 /*
1652                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1653                  * got written to. These VMAs are likely not worth investing
1654                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1655                  * later.
1656                  *
1657                  * Note that vma->anon_vma check is racy: it can be set up after
1658                  * the check but before we took mmap_lock by the fault path.
1659                  * But page lock would prevent establishing any new ptes of the
1660                  * page, so we are safe.
1661                  *
1662                  * An alternative would be drop the check, but check that page
1663                  * table is clear before calling pmdp_collapse_flush() under
1664                  * ptl. It has higher chance to recover THP for the VMA, but
1665                  * has higher cost too. It would also probably require locking
1666                  * the anon_vma.
1667                  */
1668                 if (READ_ONCE(vma->anon_vma)) {
1669                         result = SCAN_PAGE_ANON;
1670                         goto next;
1671                 }
1672                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1673                 if (addr & ~HPAGE_PMD_MASK ||
1674                     vma->vm_end < addr + HPAGE_PMD_SIZE) {
1675                         result = SCAN_VMA_CHECK;
1676                         goto next;
1677                 }
1678                 mm = vma->vm_mm;
1679                 is_target = mm == target_mm && addr == target_addr;
1680                 result = find_pmd_or_thp_or_none(mm, addr, &pmd);
1681                 if (result != SCAN_SUCCEED)
1682                         goto next;
1683                 /*
1684                  * We need exclusive mmap_lock to retract page table.
1685                  *
1686                  * We use trylock due to lock inversion: we need to acquire
1687                  * mmap_lock while holding page lock. Fault path does it in
1688                  * reverse order. Trylock is a way to avoid deadlock.
1689                  *
1690                  * Also, it's not MADV_COLLAPSE's job to collapse other
1691                  * mappings - let khugepaged take care of them later.
1692                  */
1693                 result = SCAN_PTE_MAPPED_HUGEPAGE;
1694                 if ((cc->is_khugepaged || is_target) &&
1695                     mmap_write_trylock(mm)) {
1696                         /*
1697                          * Re-check whether we have an ->anon_vma, because
1698                          * collapse_and_free_pmd() requires that either no
1699                          * ->anon_vma exists or the anon_vma is locked.
1700                          * We already checked ->anon_vma above, but that check
1701                          * is racy because ->anon_vma can be populated under the
1702                          * mmap lock in read mode.
1703                          */
1704                         if (vma->anon_vma) {
1705                                 result = SCAN_PAGE_ANON;
1706                                 goto unlock_next;
1707                         }
1708                         /*
1709                          * When a vma is registered with uffd-wp, we can't
1710                          * recycle the pmd pgtable because there can be pte
1711                          * markers installed.  Skip it only, so the rest mm/vma
1712                          * can still have the same file mapped hugely, however
1713                          * it'll always mapped in small page size for uffd-wp
1714                          * registered ranges.
1715                          */
1716                         if (hpage_collapse_test_exit(mm)) {
1717                                 result = SCAN_ANY_PROCESS;
1718                                 goto unlock_next;
1719                         }
1720                         if (userfaultfd_wp(vma)) {
1721                                 result = SCAN_PTE_UFFD_WP;
1722                                 goto unlock_next;
1723                         }
1724                         collapse_and_free_pmd(mm, vma, addr, pmd);
1725                         if (!cc->is_khugepaged && is_target)
1726                                 result = set_huge_pmd(vma, addr, pmd, hpage);
1727                         else
1728                                 result = SCAN_SUCCEED;
1729
1730 unlock_next:
1731                         mmap_write_unlock(mm);
1732                         goto next;
1733                 }
1734                 /*
1735                  * Calling context will handle target mm/addr. Otherwise, let
1736                  * khugepaged try again later.
1737                  */
1738                 if (!is_target) {
1739                         khugepaged_add_pte_mapped_thp(mm, addr);
1740                         continue;
1741                 }
1742 next:
1743                 if (is_target)
1744                         target_result = result;
1745         }
1746         i_mmap_unlock_write(mapping);
1747         return target_result;
1748 }
1749
1750 /**
1751  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1752  *
1753  * @mm: process address space where collapse happens
1754  * @addr: virtual collapse start address
1755  * @file: file that collapse on
1756  * @start: collapse start address
1757  * @cc: collapse context and scratchpad
1758  *
1759  * Basic scheme is simple, details are more complex:
1760  *  - allocate and lock a new huge page;
1761  *  - scan page cache replacing old pages with the new one
1762  *    + swap/gup in pages if necessary;
1763  *    + fill in gaps;
1764  *    + keep old pages around in case rollback is required;
1765  *  - if replacing succeeds:
1766  *    + copy data over;
1767  *    + free old pages;
1768  *    + unlock huge page;
1769  *  - if replacing failed;
1770  *    + put all pages back and unfreeze them;
1771  *    + restore gaps in the page cache;
1772  *    + unlock and free huge page;
1773  */
1774 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1775                          struct file *file, pgoff_t start,
1776                          struct collapse_control *cc)
1777 {
1778         struct address_space *mapping = file->f_mapping;
1779         struct page *hpage;
1780         pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1781         LIST_HEAD(pagelist);
1782         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1783         int nr_none = 0, result = SCAN_SUCCEED;
1784         bool is_shmem = shmem_file(file);
1785         int nr = 0;
1786
1787         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1788         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1789
1790         result = alloc_charge_hpage(&hpage, mm, cc);
1791         if (result != SCAN_SUCCEED)
1792                 goto out;
1793
1794         /*
1795          * Ensure we have slots for all the pages in the range.  This is
1796          * almost certainly a no-op because most of the pages must be present
1797          */
1798         do {
1799                 xas_lock_irq(&xas);
1800                 xas_create_range(&xas);
1801                 if (!xas_error(&xas))
1802                         break;
1803                 xas_unlock_irq(&xas);
1804                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1805                         result = SCAN_FAIL;
1806                         goto out;
1807                 }
1808         } while (1);
1809
1810         __SetPageLocked(hpage);
1811         if (is_shmem)
1812                 __SetPageSwapBacked(hpage);
1813         hpage->index = start;
1814         hpage->mapping = mapping;
1815
1816         /*
1817          * At this point the hpage is locked and not up-to-date.
1818          * It's safe to insert it into the page cache, because nobody would
1819          * be able to map it or use it in another way until we unlock it.
1820          */
1821
1822         xas_set(&xas, start);
1823         for (index = start; index < end; index++) {
1824                 struct page *page = xas_next(&xas);
1825                 struct folio *folio;
1826
1827                 VM_BUG_ON(index != xas.xa_index);
1828                 if (is_shmem) {
1829                         if (!page) {
1830                                 /*
1831                                  * Stop if extent has been truncated or
1832                                  * hole-punched, and is now completely
1833                                  * empty.
1834                                  */
1835                                 if (index == start) {
1836                                         if (!xas_next_entry(&xas, end - 1)) {
1837                                                 result = SCAN_TRUNCATED;
1838                                                 goto xa_locked;
1839                                         }
1840                                         xas_set(&xas, index);
1841                                 }
1842                                 if (!shmem_charge(mapping->host, 1)) {
1843                                         result = SCAN_FAIL;
1844                                         goto xa_locked;
1845                                 }
1846                                 xas_store(&xas, hpage);
1847                                 nr_none++;
1848                                 continue;
1849                         }
1850
1851                         if (xa_is_value(page) || !PageUptodate(page)) {
1852                                 xas_unlock_irq(&xas);
1853                                 /* swap in or instantiate fallocated page */
1854                                 if (shmem_get_folio(mapping->host, index,
1855                                                 &folio, SGP_NOALLOC)) {
1856                                         result = SCAN_FAIL;
1857                                         goto xa_unlocked;
1858                                 }
1859                                 page = folio_file_page(folio, index);
1860                         } else if (trylock_page(page)) {
1861                                 get_page(page);
1862                                 xas_unlock_irq(&xas);
1863                         } else {
1864                                 result = SCAN_PAGE_LOCK;
1865                                 goto xa_locked;
1866                         }
1867                 } else {        /* !is_shmem */
1868                         if (!page || xa_is_value(page)) {
1869                                 xas_unlock_irq(&xas);
1870                                 page_cache_sync_readahead(mapping, &file->f_ra,
1871                                                           file, index,
1872                                                           end - index);
1873                                 /* drain pagevecs to help isolate_lru_page() */
1874                                 lru_add_drain();
1875                                 page = find_lock_page(mapping, index);
1876                                 if (unlikely(page == NULL)) {
1877                                         result = SCAN_FAIL;
1878                                         goto xa_unlocked;
1879                                 }
1880                         } else if (PageDirty(page)) {
1881                                 /*
1882                                  * khugepaged only works on read-only fd,
1883                                  * so this page is dirty because it hasn't
1884                                  * been flushed since first write. There
1885                                  * won't be new dirty pages.
1886                                  *
1887                                  * Trigger async flush here and hope the
1888                                  * writeback is done when khugepaged
1889                                  * revisits this page.
1890                                  *
1891                                  * This is a one-off situation. We are not
1892                                  * forcing writeback in loop.
1893                                  */
1894                                 xas_unlock_irq(&xas);
1895                                 filemap_flush(mapping);
1896                                 result = SCAN_FAIL;
1897                                 goto xa_unlocked;
1898                         } else if (PageWriteback(page)) {
1899                                 xas_unlock_irq(&xas);
1900                                 result = SCAN_FAIL;
1901                                 goto xa_unlocked;
1902                         } else if (trylock_page(page)) {
1903                                 get_page(page);
1904                                 xas_unlock_irq(&xas);
1905                         } else {
1906                                 result = SCAN_PAGE_LOCK;
1907                                 goto xa_locked;
1908                         }
1909                 }
1910
1911                 /*
1912                  * The page must be locked, so we can drop the i_pages lock
1913                  * without racing with truncate.
1914                  */
1915                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1916
1917                 /* make sure the page is up to date */
1918                 if (unlikely(!PageUptodate(page))) {
1919                         result = SCAN_FAIL;
1920                         goto out_unlock;
1921                 }
1922
1923                 /*
1924                  * If file was truncated then extended, or hole-punched, before
1925                  * we locked the first page, then a THP might be there already.
1926                  * This will be discovered on the first iteration.
1927                  */
1928                 if (PageTransCompound(page)) {
1929                         struct page *head = compound_head(page);
1930
1931                         result = compound_order(head) == HPAGE_PMD_ORDER &&
1932                                         head->index == start
1933                                         /* Maybe PMD-mapped */
1934                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1935                                         : SCAN_PAGE_COMPOUND;
1936                         goto out_unlock;
1937                 }
1938
1939                 folio = page_folio(page);
1940
1941                 if (folio_mapping(folio) != mapping) {
1942                         result = SCAN_TRUNCATED;
1943                         goto out_unlock;
1944                 }
1945
1946                 if (!is_shmem && (folio_test_dirty(folio) ||
1947                                   folio_test_writeback(folio))) {
1948                         /*
1949                          * khugepaged only works on read-only fd, so this
1950                          * page is dirty because it hasn't been flushed
1951                          * since first write.
1952                          */
1953                         result = SCAN_FAIL;
1954                         goto out_unlock;
1955                 }
1956
1957                 if (!folio_isolate_lru(folio)) {
1958                         result = SCAN_DEL_PAGE_LRU;
1959                         goto out_unlock;
1960                 }
1961
1962                 if (folio_has_private(folio) &&
1963                     !filemap_release_folio(folio, GFP_KERNEL)) {
1964                         result = SCAN_PAGE_HAS_PRIVATE;
1965                         folio_putback_lru(folio);
1966                         goto out_unlock;
1967                 }
1968
1969                 if (folio_mapped(folio))
1970                         try_to_unmap(folio,
1971                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1972
1973                 xas_lock_irq(&xas);
1974                 xas_set(&xas, index);
1975
1976                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1977
1978                 /*
1979                  * The page is expected to have page_count() == 3:
1980                  *  - we hold a pin on it;
1981                  *  - one reference from page cache;
1982                  *  - one from isolate_lru_page;
1983                  */
1984                 if (!page_ref_freeze(page, 3)) {
1985                         result = SCAN_PAGE_COUNT;
1986                         xas_unlock_irq(&xas);
1987                         putback_lru_page(page);
1988                         goto out_unlock;
1989                 }
1990
1991                 /*
1992                  * Add the page to the list to be able to undo the collapse if
1993                  * something go wrong.
1994                  */
1995                 list_add_tail(&page->lru, &pagelist);
1996
1997                 /* Finally, replace with the new page. */
1998                 xas_store(&xas, hpage);
1999                 continue;
2000 out_unlock:
2001                 unlock_page(page);
2002                 put_page(page);
2003                 goto xa_unlocked;
2004         }
2005         nr = thp_nr_pages(hpage);
2006
2007         if (is_shmem)
2008                 __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr);
2009         else {
2010                 __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr);
2011                 filemap_nr_thps_inc(mapping);
2012                 /*
2013                  * Paired with smp_mb() in do_dentry_open() to ensure
2014                  * i_writecount is up to date and the update to nr_thps is
2015                  * visible. Ensures the page cache will be truncated if the
2016                  * file is opened writable.
2017                  */
2018                 smp_mb();
2019                 if (inode_is_open_for_write(mapping->host)) {
2020                         result = SCAN_FAIL;
2021                         __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr);
2022                         filemap_nr_thps_dec(mapping);
2023                         goto xa_locked;
2024                 }
2025         }
2026
2027         if (nr_none) {
2028                 __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none);
2029                 /* nr_none is always 0 for non-shmem. */
2030                 __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none);
2031         }
2032
2033         /* Join all the small entries into a single multi-index entry */
2034         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2035         xas_store(&xas, hpage);
2036 xa_locked:
2037         xas_unlock_irq(&xas);
2038 xa_unlocked:
2039
2040         /*
2041          * If collapse is successful, flush must be done now before copying.
2042          * If collapse is unsuccessful, does flush actually need to be done?
2043          * Do it anyway, to clear the state.
2044          */
2045         try_to_unmap_flush();
2046
2047         if (result == SCAN_SUCCEED) {
2048                 struct page *page, *tmp;
2049                 struct folio *folio;
2050
2051                 /*
2052                  * Replacing old pages with new one has succeeded, now we
2053                  * need to copy the content and free the old pages.
2054                  */
2055                 index = start;
2056                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2057                         while (index < page->index) {
2058                                 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2059                                 index++;
2060                         }
2061                         copy_highpage(hpage + (page->index % HPAGE_PMD_NR),
2062                                       page);
2063                         list_del(&page->lru);
2064                         page->mapping = NULL;
2065                         page_ref_unfreeze(page, 1);
2066                         ClearPageActive(page);
2067                         ClearPageUnevictable(page);
2068                         unlock_page(page);
2069                         put_page(page);
2070                         index++;
2071                 }
2072                 while (index < end) {
2073                         clear_highpage(hpage + (index % HPAGE_PMD_NR));
2074                         index++;
2075                 }
2076
2077                 folio = page_folio(hpage);
2078                 folio_mark_uptodate(folio);
2079                 folio_ref_add(folio, HPAGE_PMD_NR - 1);
2080
2081                 if (is_shmem)
2082                         folio_mark_dirty(folio);
2083                 folio_add_lru(folio);
2084
2085                 /*
2086                  * Remove pte page tables, so we can re-fault the page as huge.
2087                  */
2088                 result = retract_page_tables(mapping, start, mm, addr, hpage,
2089                                              cc);
2090                 unlock_page(hpage);
2091                 hpage = NULL;
2092         } else {
2093                 struct page *page;
2094
2095                 /* Something went wrong: roll back page cache changes */
2096                 xas_lock_irq(&xas);
2097                 if (nr_none) {
2098                         mapping->nrpages -= nr_none;
2099                         shmem_uncharge(mapping->host, nr_none);
2100                 }
2101
2102                 xas_set(&xas, start);
2103                 xas_for_each(&xas, page, end - 1) {
2104                         page = list_first_entry_or_null(&pagelist,
2105                                         struct page, lru);
2106                         if (!page || xas.xa_index < page->index) {
2107                                 if (!nr_none)
2108                                         break;
2109                                 nr_none--;
2110                                 /* Put holes back where they were */
2111                                 xas_store(&xas, NULL);
2112                                 continue;
2113                         }
2114
2115                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2116
2117                         /* Unfreeze the page. */
2118                         list_del(&page->lru);
2119                         page_ref_unfreeze(page, 2);
2120                         xas_store(&xas, page);
2121                         xas_pause(&xas);
2122                         xas_unlock_irq(&xas);
2123                         unlock_page(page);
2124                         putback_lru_page(page);
2125                         xas_lock_irq(&xas);
2126                 }
2127                 VM_BUG_ON(nr_none);
2128                 xas_unlock_irq(&xas);
2129
2130                 hpage->mapping = NULL;
2131         }
2132
2133         if (hpage)
2134                 unlock_page(hpage);
2135 out:
2136         VM_BUG_ON(!list_empty(&pagelist));
2137         if (hpage) {
2138                 mem_cgroup_uncharge(page_folio(hpage));
2139                 put_page(hpage);
2140         }
2141
2142         trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2143         return result;
2144 }
2145
2146 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2147                                     struct file *file, pgoff_t start,
2148                                     struct collapse_control *cc)
2149 {
2150         struct page *page = NULL;
2151         struct address_space *mapping = file->f_mapping;
2152         XA_STATE(xas, &mapping->i_pages, start);
2153         int present, swap;
2154         int node = NUMA_NO_NODE;
2155         int result = SCAN_SUCCEED;
2156
2157         present = 0;
2158         swap = 0;
2159         memset(cc->node_load, 0, sizeof(cc->node_load));
2160         nodes_clear(cc->alloc_nmask);
2161         rcu_read_lock();
2162         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2163                 if (xas_retry(&xas, page))
2164                         continue;
2165
2166                 if (xa_is_value(page)) {
2167                         ++swap;
2168                         if (cc->is_khugepaged &&
2169                             swap > khugepaged_max_ptes_swap) {
2170                                 result = SCAN_EXCEED_SWAP_PTE;
2171                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2172                                 break;
2173                         }
2174                         continue;
2175                 }
2176
2177                 /*
2178                  * TODO: khugepaged should compact smaller compound pages
2179                  * into a PMD sized page
2180                  */
2181                 if (PageTransCompound(page)) {
2182                         struct page *head = compound_head(page);
2183
2184                         result = compound_order(head) == HPAGE_PMD_ORDER &&
2185                                         head->index == start
2186                                         /* Maybe PMD-mapped */
2187                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2188                                         : SCAN_PAGE_COMPOUND;
2189                         /*
2190                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2191                          * by the caller won't touch the page cache, and so
2192                          * it's safe to skip LRU and refcount checks before
2193                          * returning.
2194                          */
2195                         break;
2196                 }
2197
2198                 node = page_to_nid(page);
2199                 if (hpage_collapse_scan_abort(node, cc)) {
2200                         result = SCAN_SCAN_ABORT;
2201                         break;
2202                 }
2203                 cc->node_load[node]++;
2204
2205                 if (!PageLRU(page)) {
2206                         result = SCAN_PAGE_LRU;
2207                         break;
2208                 }
2209
2210                 if (page_count(page) !=
2211                     1 + page_mapcount(page) + page_has_private(page)) {
2212                         result = SCAN_PAGE_COUNT;
2213                         break;
2214                 }
2215
2216                 /*
2217                  * We probably should check if the page is referenced here, but
2218                  * nobody would transfer pte_young() to PageReferenced() for us.
2219                  * And rmap walk here is just too costly...
2220                  */
2221
2222                 present++;
2223
2224                 if (need_resched()) {
2225                         xas_pause(&xas);
2226                         cond_resched_rcu();
2227                 }
2228         }
2229         rcu_read_unlock();
2230
2231         if (result == SCAN_SUCCEED) {
2232                 if (cc->is_khugepaged &&
2233                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2234                         result = SCAN_EXCEED_NONE_PTE;
2235                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2236                 } else {
2237                         result = collapse_file(mm, addr, file, start, cc);
2238                 }
2239         }
2240
2241         trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2242         return result;
2243 }
2244 #else
2245 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2246                                     struct file *file, pgoff_t start,
2247                                     struct collapse_control *cc)
2248 {
2249         BUILD_BUG();
2250 }
2251
2252 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
2253 {
2254 }
2255
2256 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
2257                                           unsigned long addr)
2258 {
2259         return false;
2260 }
2261 #endif
2262
2263 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2264                                             struct collapse_control *cc)
2265         __releases(&khugepaged_mm_lock)
2266         __acquires(&khugepaged_mm_lock)
2267 {
2268         struct vma_iterator vmi;
2269         struct khugepaged_mm_slot *mm_slot;
2270         struct mm_slot *slot;
2271         struct mm_struct *mm;
2272         struct vm_area_struct *vma;
2273         int progress = 0;
2274
2275         VM_BUG_ON(!pages);
2276         lockdep_assert_held(&khugepaged_mm_lock);
2277         *result = SCAN_FAIL;
2278
2279         if (khugepaged_scan.mm_slot) {
2280                 mm_slot = khugepaged_scan.mm_slot;
2281                 slot = &mm_slot->slot;
2282         } else {
2283                 slot = list_entry(khugepaged_scan.mm_head.next,
2284                                      struct mm_slot, mm_node);
2285                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2286                 khugepaged_scan.address = 0;
2287                 khugepaged_scan.mm_slot = mm_slot;
2288         }
2289         spin_unlock(&khugepaged_mm_lock);
2290         khugepaged_collapse_pte_mapped_thps(mm_slot);
2291
2292         mm = slot->mm;
2293         /*
2294          * Don't wait for semaphore (to avoid long wait times).  Just move to
2295          * the next mm on the list.
2296          */
2297         vma = NULL;
2298         if (unlikely(!mmap_read_trylock(mm)))
2299                 goto breakouterloop_mmap_lock;
2300
2301         progress++;
2302         if (unlikely(hpage_collapse_test_exit(mm)))
2303                 goto breakouterloop;
2304
2305         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2306         for_each_vma(vmi, vma) {
2307                 unsigned long hstart, hend;
2308
2309                 cond_resched();
2310                 if (unlikely(hpage_collapse_test_exit(mm))) {
2311                         progress++;
2312                         break;
2313                 }
2314                 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) {
2315 skip:
2316                         progress++;
2317                         continue;
2318                 }
2319                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2320                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2321                 if (khugepaged_scan.address > hend)
2322                         goto skip;
2323                 if (khugepaged_scan.address < hstart)
2324                         khugepaged_scan.address = hstart;
2325                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2326
2327                 while (khugepaged_scan.address < hend) {
2328                         bool mmap_locked = true;
2329
2330                         cond_resched();
2331                         if (unlikely(hpage_collapse_test_exit(mm)))
2332                                 goto breakouterloop;
2333
2334                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2335                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2336                                   hend);
2337                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2338                                 struct file *file = get_file(vma->vm_file);
2339                                 pgoff_t pgoff = linear_page_index(vma,
2340                                                 khugepaged_scan.address);
2341
2342                                 mmap_read_unlock(mm);
2343                                 *result = hpage_collapse_scan_file(mm,
2344                                                                    khugepaged_scan.address,
2345                                                                    file, pgoff, cc);
2346                                 mmap_locked = false;
2347                                 fput(file);
2348                         } else {
2349                                 *result = hpage_collapse_scan_pmd(mm, vma,
2350                                                                   khugepaged_scan.address,
2351                                                                   &mmap_locked,
2352                                                                   cc);
2353                         }
2354                         switch (*result) {
2355                         case SCAN_PTE_MAPPED_HUGEPAGE: {
2356                                 pmd_t *pmd;
2357
2358                                 *result = find_pmd_or_thp_or_none(mm,
2359                                                                   khugepaged_scan.address,
2360                                                                   &pmd);
2361                                 if (*result != SCAN_SUCCEED)
2362                                         break;
2363                                 if (!khugepaged_add_pte_mapped_thp(mm,
2364                                                                    khugepaged_scan.address))
2365                                         break;
2366                         } fallthrough;
2367                         case SCAN_SUCCEED:
2368                                 ++khugepaged_pages_collapsed;
2369                                 break;
2370                         default:
2371                                 break;
2372                         }
2373
2374                         /* move to next address */
2375                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2376                         progress += HPAGE_PMD_NR;
2377                         if (!mmap_locked)
2378                                 /*
2379                                  * We released mmap_lock so break loop.  Note
2380                                  * that we drop mmap_lock before all hugepage
2381                                  * allocations, so if allocation fails, we are
2382                                  * guaranteed to break here and report the
2383                                  * correct result back to caller.
2384                                  */
2385                                 goto breakouterloop_mmap_lock;
2386                         if (progress >= pages)
2387                                 goto breakouterloop;
2388                 }
2389         }
2390 breakouterloop:
2391         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2392 breakouterloop_mmap_lock:
2393
2394         spin_lock(&khugepaged_mm_lock);
2395         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2396         /*
2397          * Release the current mm_slot if this mm is about to die, or
2398          * if we scanned all vmas of this mm.
2399          */
2400         if (hpage_collapse_test_exit(mm) || !vma) {
2401                 /*
2402                  * Make sure that if mm_users is reaching zero while
2403                  * khugepaged runs here, khugepaged_exit will find
2404                  * mm_slot not pointing to the exiting mm.
2405                  */
2406                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2407                         slot = list_entry(slot->mm_node.next,
2408                                           struct mm_slot, mm_node);
2409                         khugepaged_scan.mm_slot =
2410                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2411                         khugepaged_scan.address = 0;
2412                 } else {
2413                         khugepaged_scan.mm_slot = NULL;
2414                         khugepaged_full_scans++;
2415                 }
2416
2417                 collect_mm_slot(mm_slot);
2418         }
2419
2420         return progress;
2421 }
2422
2423 static int khugepaged_has_work(void)
2424 {
2425         return !list_empty(&khugepaged_scan.mm_head) &&
2426                 hugepage_flags_enabled();
2427 }
2428
2429 static int khugepaged_wait_event(void)
2430 {
2431         return !list_empty(&khugepaged_scan.mm_head) ||
2432                 kthread_should_stop();
2433 }
2434
2435 static void khugepaged_do_scan(struct collapse_control *cc)
2436 {
2437         unsigned int progress = 0, pass_through_head = 0;
2438         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2439         bool wait = true;
2440         int result = SCAN_SUCCEED;
2441
2442         lru_add_drain_all();
2443
2444         while (true) {
2445                 cond_resched();
2446
2447                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2448                         break;
2449
2450                 spin_lock(&khugepaged_mm_lock);
2451                 if (!khugepaged_scan.mm_slot)
2452                         pass_through_head++;
2453                 if (khugepaged_has_work() &&
2454                     pass_through_head < 2)
2455                         progress += khugepaged_scan_mm_slot(pages - progress,
2456                                                             &result, cc);
2457                 else
2458                         progress = pages;
2459                 spin_unlock(&khugepaged_mm_lock);
2460
2461                 if (progress >= pages)
2462                         break;
2463
2464                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2465                         /*
2466                          * If fail to allocate the first time, try to sleep for
2467                          * a while.  When hit again, cancel the scan.
2468                          */
2469                         if (!wait)
2470                                 break;
2471                         wait = false;
2472                         khugepaged_alloc_sleep();
2473                 }
2474         }
2475 }
2476
2477 static bool khugepaged_should_wakeup(void)
2478 {
2479         return kthread_should_stop() ||
2480                time_after_eq(jiffies, khugepaged_sleep_expire);
2481 }
2482
2483 static void khugepaged_wait_work(void)
2484 {
2485         if (khugepaged_has_work()) {
2486                 const unsigned long scan_sleep_jiffies =
2487                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2488
2489                 if (!scan_sleep_jiffies)
2490                         return;
2491
2492                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2493                 wait_event_freezable_timeout(khugepaged_wait,
2494                                              khugepaged_should_wakeup(),
2495                                              scan_sleep_jiffies);
2496                 return;
2497         }
2498
2499         if (hugepage_flags_enabled())
2500                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2501 }
2502
2503 static int khugepaged(void *none)
2504 {
2505         struct khugepaged_mm_slot *mm_slot;
2506
2507         set_freezable();
2508         set_user_nice(current, MAX_NICE);
2509
2510         while (!kthread_should_stop()) {
2511                 khugepaged_do_scan(&khugepaged_collapse_control);
2512                 khugepaged_wait_work();
2513         }
2514
2515         spin_lock(&khugepaged_mm_lock);
2516         mm_slot = khugepaged_scan.mm_slot;
2517         khugepaged_scan.mm_slot = NULL;
2518         if (mm_slot)
2519                 collect_mm_slot(mm_slot);
2520         spin_unlock(&khugepaged_mm_lock);
2521         return 0;
2522 }
2523
2524 static void set_recommended_min_free_kbytes(void)
2525 {
2526         struct zone *zone;
2527         int nr_zones = 0;
2528         unsigned long recommended_min;
2529
2530         if (!hugepage_flags_enabled()) {
2531                 calculate_min_free_kbytes();
2532                 goto update_wmarks;
2533         }
2534
2535         for_each_populated_zone(zone) {
2536                 /*
2537                  * We don't need to worry about fragmentation of
2538                  * ZONE_MOVABLE since it only has movable pages.
2539                  */
2540                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2541                         continue;
2542
2543                 nr_zones++;
2544         }
2545
2546         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2547         recommended_min = pageblock_nr_pages * nr_zones * 2;
2548
2549         /*
2550          * Make sure that on average at least two pageblocks are almost free
2551          * of another type, one for a migratetype to fall back to and a
2552          * second to avoid subsequent fallbacks of other types There are 3
2553          * MIGRATE_TYPES we care about.
2554          */
2555         recommended_min += pageblock_nr_pages * nr_zones *
2556                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2557
2558         /* don't ever allow to reserve more than 5% of the lowmem */
2559         recommended_min = min(recommended_min,
2560                               (unsigned long) nr_free_buffer_pages() / 20);
2561         recommended_min <<= (PAGE_SHIFT-10);
2562
2563         if (recommended_min > min_free_kbytes) {
2564                 if (user_min_free_kbytes >= 0)
2565                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2566                                 min_free_kbytes, recommended_min);
2567
2568                 min_free_kbytes = recommended_min;
2569         }
2570
2571 update_wmarks:
2572         setup_per_zone_wmarks();
2573 }
2574
2575 int start_stop_khugepaged(void)
2576 {
2577         int err = 0;
2578
2579         mutex_lock(&khugepaged_mutex);
2580         if (hugepage_flags_enabled()) {
2581                 if (!khugepaged_thread)
2582                         khugepaged_thread = kthread_run(khugepaged, NULL,
2583                                                         "khugepaged");
2584                 if (IS_ERR(khugepaged_thread)) {
2585                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2586                         err = PTR_ERR(khugepaged_thread);
2587                         khugepaged_thread = NULL;
2588                         goto fail;
2589                 }
2590
2591                 if (!list_empty(&khugepaged_scan.mm_head))
2592                         wake_up_interruptible(&khugepaged_wait);
2593         } else if (khugepaged_thread) {
2594                 kthread_stop(khugepaged_thread);
2595                 khugepaged_thread = NULL;
2596         }
2597         set_recommended_min_free_kbytes();
2598 fail:
2599         mutex_unlock(&khugepaged_mutex);
2600         return err;
2601 }
2602
2603 void khugepaged_min_free_kbytes_update(void)
2604 {
2605         mutex_lock(&khugepaged_mutex);
2606         if (hugepage_flags_enabled() && khugepaged_thread)
2607                 set_recommended_min_free_kbytes();
2608         mutex_unlock(&khugepaged_mutex);
2609 }
2610
2611 bool current_is_khugepaged(void)
2612 {
2613         return kthread_func(current) == khugepaged;
2614 }
2615
2616 static int madvise_collapse_errno(enum scan_result r)
2617 {
2618         /*
2619          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2620          * actionable feedback to caller, so they may take an appropriate
2621          * fallback measure depending on the nature of the failure.
2622          */
2623         switch (r) {
2624         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2625                 return -ENOMEM;
2626         case SCAN_CGROUP_CHARGE_FAIL:
2627                 return -EBUSY;
2628         /* Resource temporary unavailable - trying again might succeed */
2629         case SCAN_PAGE_COUNT:
2630         case SCAN_PAGE_LOCK:
2631         case SCAN_PAGE_LRU:
2632         case SCAN_DEL_PAGE_LRU:
2633                 return -EAGAIN;
2634         /*
2635          * Other: Trying again likely not to succeed / error intrinsic to
2636          * specified memory range. khugepaged likely won't be able to collapse
2637          * either.
2638          */
2639         default:
2640                 return -EINVAL;
2641         }
2642 }
2643
2644 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2645                      unsigned long start, unsigned long end)
2646 {
2647         struct collapse_control *cc;
2648         struct mm_struct *mm = vma->vm_mm;
2649         unsigned long hstart, hend, addr;
2650         int thps = 0, last_fail = SCAN_FAIL;
2651         bool mmap_locked = true;
2652
2653         BUG_ON(vma->vm_start > start);
2654         BUG_ON(vma->vm_end < end);
2655
2656         *prev = vma;
2657
2658         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
2659                 return -EINVAL;
2660
2661         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2662         if (!cc)
2663                 return -ENOMEM;
2664         cc->is_khugepaged = false;
2665
2666         mmgrab(mm);
2667         lru_add_drain_all();
2668
2669         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2670         hend = end & HPAGE_PMD_MASK;
2671
2672         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2673                 int result = SCAN_FAIL;
2674
2675                 if (!mmap_locked) {
2676                         cond_resched();
2677                         mmap_read_lock(mm);
2678                         mmap_locked = true;
2679                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2680                                                          cc);
2681                         if (result  != SCAN_SUCCEED) {
2682                                 last_fail = result;
2683                                 goto out_nolock;
2684                         }
2685
2686                         hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2687                 }
2688                 mmap_assert_locked(mm);
2689                 memset(cc->node_load, 0, sizeof(cc->node_load));
2690                 nodes_clear(cc->alloc_nmask);
2691                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2692                         struct file *file = get_file(vma->vm_file);
2693                         pgoff_t pgoff = linear_page_index(vma, addr);
2694
2695                         mmap_read_unlock(mm);
2696                         mmap_locked = false;
2697                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2698                                                           cc);
2699                         fput(file);
2700                 } else {
2701                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2702                                                          &mmap_locked, cc);
2703                 }
2704                 if (!mmap_locked)
2705                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2706
2707 handle_result:
2708                 switch (result) {
2709                 case SCAN_SUCCEED:
2710                 case SCAN_PMD_MAPPED:
2711                         ++thps;
2712                         break;
2713                 case SCAN_PTE_MAPPED_HUGEPAGE:
2714                         BUG_ON(mmap_locked);
2715                         BUG_ON(*prev);
2716                         mmap_write_lock(mm);
2717                         result = collapse_pte_mapped_thp(mm, addr, true);
2718                         mmap_write_unlock(mm);
2719                         goto handle_result;
2720                 /* Whitelisted set of results where continuing OK */
2721                 case SCAN_PMD_NULL:
2722                 case SCAN_PTE_NON_PRESENT:
2723                 case SCAN_PTE_UFFD_WP:
2724                 case SCAN_PAGE_RO:
2725                 case SCAN_LACK_REFERENCED_PAGE:
2726                 case SCAN_PAGE_NULL:
2727                 case SCAN_PAGE_COUNT:
2728                 case SCAN_PAGE_LOCK:
2729                 case SCAN_PAGE_COMPOUND:
2730                 case SCAN_PAGE_LRU:
2731                 case SCAN_DEL_PAGE_LRU:
2732                         last_fail = result;
2733                         break;
2734                 default:
2735                         last_fail = result;
2736                         /* Other error, exit */
2737                         goto out_maybelock;
2738                 }
2739         }
2740
2741 out_maybelock:
2742         /* Caller expects us to hold mmap_lock on return */
2743         if (!mmap_locked)
2744                 mmap_read_lock(mm);
2745 out_nolock:
2746         mmap_assert_locked(mm);
2747         mmdrop(mm);
2748         kfree(cc);
2749
2750         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2751                         : madvise_collapse_errno(last_fail);
2752 }