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         /* Last target selected in hpage_collapse_find_target_node() */
 101         int last_target_node;
 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_page(struct page *page)
 494 {
 495         mod_node_page_state(page_pgdat(page),
 496                         NR_ISOLATED_ANON + page_is_file_lru(page),
 497                         -compound_nr(page));
 498         unlock_page(page);
 499         putback_lru_page(page);
 500 }
 501
 502 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 503                 struct list_head *compound_pagelist)
 504 {
 505         struct page *page, *tmp;
 506
 507         while (--_pte >= pte) {
 508                 pte_t pteval = *_pte;
 509
 510                 page = pte_page(pteval);
 511                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 512                                 !PageCompound(page))
 513                         release_pte_page(page);
 514         }
 515
 516         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 517                 list_del(&page->lru);
 518                 release_pte_page(page);
 519         }
 520 }
 521
 522 static bool is_refcount_suitable(struct page *page)
 523 {
 524         int expected_refcount;
 525
 526         expected_refcount = total_mapcount(page);
 527         if (PageSwapCache(page))
 528                 expected_refcount += compound_nr(page);
 529
 530         return page_count(page) == expected_refcount;
 531 }
 532
 533 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 534                                         unsigned long address,
 535                                         pte_t *pte,
 536                                         struct collapse_control *cc,
 537                                         struct list_head *compound_pagelist)
 538 {
 539         struct page *page = NULL;
 540         pte_t *_pte;
 541         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 542         bool writable = false;
 543
 544         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 545              _pte++, address += PAGE_SIZE) {
 546                 pte_t pteval = *_pte;
 547                 if (pte_none(pteval) || (pte_present(pteval) &&
 548                                 is_zero_pfn(pte_pfn(pteval)))) {
 549                         ++none_or_zero;
 550                         if (!userfaultfd_armed(vma) &&
 551                             (!cc->is_khugepaged ||
 552                              none_or_zero <= khugepaged_max_ptes_none)) {
 553                                 continue;
 554                         } else {
 555                                 result = SCAN_EXCEED_NONE_PTE;
 556                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 557                                 goto out;
 558                         }
 559                 }
 560                 if (!pte_present(pteval)) {
 561                         result = SCAN_PTE_NON_PRESENT;
 562                         goto out;
 563                 }
 564                 page = vm_normal_page(vma, address, pteval);
 565                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 566                         result = SCAN_PAGE_NULL;
 567                         goto out;
 568                 }
 569
 570                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 571
 572                 if (page_mapcount(page) > 1) {
 573                         ++shared;
 574                         if (cc->is_khugepaged &&
 575                             shared > khugepaged_max_ptes_shared) {
 576                                 result = SCAN_EXCEED_SHARED_PTE;
 577                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 578                                 goto out;
 579                         }
 580                 }
 581
 582                 if (PageCompound(page)) {
 583                         struct page *p;
 584                         page = compound_head(page);
 585
 586                         /*
 587                          * Check if we have dealt with the compound page
 588                          * already
 589                          */
 590                         list_for_each_entry(p, compound_pagelist, lru) {
 591                                 if (page == p)
 592                                         goto next;
 593                         }
 594                 }
 595
 596                 /*
 597                  * We can do it before isolate_lru_page because the
 598                  * page can't be freed from under us. NOTE: PG_lock
 599                  * is needed to serialize against split_huge_page
 600                  * when invoked from the VM.
 601                  */
 602                 if (!trylock_page(page)) {
 603                         result = SCAN_PAGE_LOCK;
 604                         goto out;
 605                 }
 606
 607                 /*
 608                  * Check if the page has any GUP (or other external) pins.
 609                  *
 610                  * The page table that maps the page has been already unlinked
 611                  * from the page table tree and this process cannot get
 612                  * an additional pin on the page.
 613                  *
 614                  * New pins can come later if the page is shared across fork,
 615                  * but not from this process. The other process cannot write to
 616                  * the page, only trigger CoW.
 617                  */
 618                 if (!is_refcount_suitable(page)) {
 619                         unlock_page(page);
 620                         result = SCAN_PAGE_COUNT;
 621                         goto out;
 622                 }
 623
 624                 /*
 625                  * Isolate the page to avoid collapsing an hugepage
 626                  * currently in use by the VM.
 627                  */
 628                 if (isolate_lru_page(page)) {
 629                         unlock_page(page);
 630                         result = SCAN_DEL_PAGE_LRU;
 631                         goto out;
 632                 }
 633                 mod_node_page_state(page_pgdat(page),
 634                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 635                                 compound_nr(page));
 636                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 637                 VM_BUG_ON_PAGE(PageLRU(page), page);
 638
 639                 if (PageCompound(page))
 640                         list_add_tail(&page->lru, compound_pagelist);
 641 next:
 642                 /*
 643                  * If collapse was initiated by khugepaged, check that there is
 644                  * enough young pte to justify collapsing the page
 645                  */
 646                 if (cc->is_khugepaged &&
 647                     (pte_young(pteval) || page_is_young(page) ||
 648                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
 649                                                                      address)))
 650                         referenced++;
 651
 652                 if (pte_write(pteval))
 653                         writable = true;
 654         }
 655
 656         if (unlikely(!writable)) {
 657                 result = SCAN_PAGE_RO;
 658         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 659                 result = SCAN_LACK_REFERENCED_PAGE;
 660         } else {
 661                 result = SCAN_SUCCEED;
 662                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 663                                                     referenced, writable, result);
 664                 return result;
 665         }
 666 out:
 667         release_pte_pages(pte, _pte, compound_pagelist);
 668         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 669                                             referenced, writable, result);
 670         return result;
 671 }
 672
 673 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 674                                       struct vm_area_struct *vma,
 675                                       unsigned long address,
 676                                       spinlock_t *ptl,
 677                                       struct list_head *compound_pagelist)
 678 {
 679         struct page *src_page, *tmp;
 680         pte_t *_pte;
 681         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 682                                 _pte++, page++, address += PAGE_SIZE) {
 683                 pte_t pteval = *_pte;
 684
 685                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 686                         clear_user_highpage(page, address);
 687                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 688                         if (is_zero_pfn(pte_pfn(pteval))) {
 689                                 /*
 690                                  * ptl mostly unnecessary.
 691                                  */
 692                                 spin_lock(ptl);
 693                                 ptep_clear(vma->vm_mm, address, _pte);
 694                                 spin_unlock(ptl);
 695                         }
 696                 } else {
 697                         src_page = pte_page(pteval);
 698                         copy_user_highpage(page, src_page, address, vma);
 699                         if (!PageCompound(src_page))
 700                                 release_pte_page(src_page);
 701                         /*
 702                          * ptl mostly unnecessary, but preempt has to
 703                          * be disabled to update the per-cpu stats
 704                          * inside page_remove_rmap().
 705                          */
 706                         spin_lock(ptl);
 707                         ptep_clear(vma->vm_mm, address, _pte);
 708                         page_remove_rmap(src_page, vma, false);
 709                         spin_unlock(ptl);
 710                         free_page_and_swap_cache(src_page);
 711                 }
 712         }
 713
 714         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 715                 list_del(&src_page->lru);
 716                 mod_node_page_state(page_pgdat(src_page),
 717                                     NR_ISOLATED_ANON + page_is_file_lru(src_page),
 718                                     -compound_nr(src_page));
 719                 unlock_page(src_page);
 720                 free_swap_cache(src_page);
 721                 putback_lru_page(src_page);
 722         }
 723 }
 724
 725 static void khugepaged_alloc_sleep(void)
 726 {
 727         DEFINE_WAIT(wait);
 728
 729         add_wait_queue(&khugepaged_wait, &wait);
 730         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 731         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 732         remove_wait_queue(&khugepaged_wait, &wait);
 733 }
 734
 735 struct collapse_control khugepaged_collapse_control = {
 736         .is_khugepaged = true,
 737         .last_target_node = NUMA_NO_NODE,
 738 };
 739
 740 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 741 {
 742         int i;
 743
 744         /*
 745          * If node_reclaim_mode is disabled, then no extra effort is made to
 746          * allocate memory locally.
 747          */
 748         if (!node_reclaim_enabled())
 749                 return false;
 750
 751         /* If there is a count for this node already, it must be acceptable */
 752         if (cc->node_load[nid])
 753                 return false;
 754
 755         for (i = 0; i < MAX_NUMNODES; i++) {
 756                 if (!cc->node_load[i])
 757                         continue;
 758                 if (node_distance(nid, i) > node_reclaim_distance)
 759                         return true;
 760         }
 761         return false;
 762 }
 763
 764 #define khugepaged_defrag()                                     \
 765         (transparent_hugepage_flags &                           \
 766          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 767
 768 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 769 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 770 {
 771         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 772 }
 773
 774 #ifdef CONFIG_NUMA
 775 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 776 {
 777         int nid, target_node = 0, max_value = 0;
 778
 779         /* find first node with max normal pages hit */
 780         for (nid = 0; nid < MAX_NUMNODES; nid++)
 781                 if (cc->node_load[nid] > max_value) {
 782                         max_value = cc->node_load[nid];
 783                         target_node = nid;
 784                 }
 785
 786         /* do some balance if several nodes have the same hit record */
 787         if (target_node <= cc->last_target_node)
 788                 for (nid = cc->last_target_node + 1; nid < MAX_NUMNODES;
 789                      nid++)
 790                         if (max_value == cc->node_load[nid]) {
 791                                 target_node = nid;
 792                                 break;
 793                         }
 794
 795         cc->last_target_node = target_node;
 796         return target_node;
 797 }
 798 #else
 799 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 800 {
 801         return 0;
 802 }
 803 #endif
 804
 805 static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node)
 806 {
 807         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 808         if (unlikely(!*hpage)) {
 809                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 810                 return false;
 811         }
 812
 813         prep_transhuge_page(*hpage);
 814         count_vm_event(THP_COLLAPSE_ALLOC);
 815         return true;
 816 }
 817
 818 /*
 819  * If mmap_lock temporarily dropped, revalidate vma
 820  * before taking mmap_lock.
 821  * Returns enum scan_result value.
 822  */
 823
 824 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 825                                    bool expect_anon,
 826                                    struct vm_area_struct **vmap,
 827                                    struct collapse_control *cc)
 828 {
 829         struct vm_area_struct *vma;
 830
 831         if (unlikely(hpage_collapse_test_exit(mm)))
 832                 return SCAN_ANY_PROCESS;
 833
 834         *vmap = vma = find_vma(mm, address);
 835         if (!vma)
 836                 return SCAN_VMA_NULL;
 837
 838         if (!transhuge_vma_suitable(vma, address))
 839                 return SCAN_ADDRESS_RANGE;
 840         if (!hugepage_vma_check(vma, vma->vm_flags, false, false,
 841                                 cc->is_khugepaged))
 842                 return SCAN_VMA_CHECK;
 843         /*
 844          * Anon VMA expected, the address may be unmapped then
 845          * remapped to file after khugepaged reaquired the mmap_lock.
 846          *
 847          * hugepage_vma_check may return true for qualified file
 848          * vmas.
 849          */
 850         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 851                 return SCAN_PAGE_ANON;
 852         return SCAN_SUCCEED;
 853 }
 854
 855 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 856                                    unsigned long address,
 857                                    pmd_t **pmd)
 858 {
 859         pmd_t pmde;
 860
 861         *pmd = mm_find_pmd(mm, address);
 862         if (!*pmd)
 863                 return SCAN_PMD_NULL;
 864
 865         pmde = pmd_read_atomic(*pmd);
 866
 867 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 868         /* See comments in pmd_none_or_trans_huge_or_clear_bad() */
 869         barrier();
 870 #endif
 871         if (pmd_none(pmde))
 872                 return SCAN_PMD_NONE;
 873         if (pmd_trans_huge(pmde))
 874                 return SCAN_PMD_MAPPED;
 875         if (pmd_bad(pmde))
 876                 return SCAN_PMD_NULL;
 877         return SCAN_SUCCEED;
 878 }
 879
 880 static int check_pmd_still_valid(struct mm_struct *mm,
 881                                  unsigned long address,
 882                                  pmd_t *pmd)
 883 {
 884         pmd_t *new_pmd;
 885         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 886
 887         if (result != SCAN_SUCCEED)
 888                 return result;
 889         if (new_pmd != pmd)
 890                 return SCAN_FAIL;
 891         return SCAN_SUCCEED;
 892 }
 893
 894 /*
 895  * Bring missing pages in from swap, to complete THP collapse.
 896  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 897  *
 898  * Called and returns without pte mapped or spinlocks held.
 899  * Note that if false is returned, mmap_lock will be released.
 900  */
 901
 902 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 903                                        struct vm_area_struct *vma,
 904                                        unsigned long haddr, pmd_t *pmd,
 905                                        int referenced)
 906 {
 907         int swapped_in = 0;
 908         vm_fault_t ret = 0;
 909         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 910
 911         for (address = haddr; address < end; address += PAGE_SIZE) {
 912                 struct vm_fault vmf = {
 913                         .vma = vma,
 914                         .address = address,
 915                         .pgoff = linear_page_index(vma, haddr),
 916                         .flags = FAULT_FLAG_ALLOW_RETRY,
 917                         .pmd = pmd,
 918                 };
 919
 920                 vmf.pte = pte_offset_map(pmd, address);
 921                 vmf.orig_pte = *vmf.pte;
 922                 if (!is_swap_pte(vmf.orig_pte)) {
 923                         pte_unmap(vmf.pte);
 924                         continue;
 925                 }
 926                 ret = do_swap_page(&vmf);
 927
 928                 /*
 929                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
 930                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
 931                  * we do not retry here and swap entry will remain in pagetable
 932                  * resulting in later failure.
 933                  */
 934                 if (ret & VM_FAULT_RETRY) {
 935                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 936                         /* Likely, but not guaranteed, that page lock failed */
 937                         return SCAN_PAGE_LOCK;
 938                 }
 939                 if (ret & VM_FAULT_ERROR) {
 940                         mmap_read_unlock(mm);
 941                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 942                         return SCAN_FAIL;
 943                 }
 944                 swapped_in++;
 945         }
 946
 947         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
 948         if (swapped_in)
 949                 lru_add_drain();
 950
 951         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 952         return SCAN_SUCCEED;
 953 }
 954
 955 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
 956                               struct collapse_control *cc)
 957 {
 958         /* Only allocate from the target node */
 959         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
 960                      GFP_TRANSHUGE) | __GFP_THISNODE;
 961         int node = hpage_collapse_find_target_node(cc);
 962
 963         if (!hpage_collapse_alloc_page(hpage, gfp, node))
 964                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
 965         if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp)))
 966                 return SCAN_CGROUP_CHARGE_FAIL;
 967         count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC);
 968         return SCAN_SUCCEED;
 969 }
 970
 971 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
 972                               int referenced, int unmapped,
 973                               struct collapse_control *cc)
 974 {
 975         LIST_HEAD(compound_pagelist);
 976         pmd_t *pmd, _pmd;
 977         pte_t *pte;
 978         pgtable_t pgtable;
 979         struct page *hpage;
 980         spinlock_t *pmd_ptl, *pte_ptl;
 981         int result = SCAN_FAIL;
 982         struct vm_area_struct *vma;
 983         struct mmu_notifier_range range;
 984
 985         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 986
 987         /*
 988          * Before allocating the hugepage, release the mmap_lock read lock.
 989          * The allocation can take potentially a long time if it involves
 990          * sync compaction, and we do not need to hold the mmap_lock during
 991          * that. We will recheck the vma after taking it again in write mode.
 992          */
 993         mmap_read_unlock(mm);
 994
 995         result = alloc_charge_hpage(&hpage, mm, cc);
 996         if (result != SCAN_SUCCEED)
 997                 goto out_nolock;
 998
 999         mmap_read_lock(mm);
1000         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1001         if (result != SCAN_SUCCEED) {
1002                 mmap_read_unlock(mm);
1003                 goto out_nolock;
1004         }
1005
1006         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1007         if (result != SCAN_SUCCEED) {
1008                 mmap_read_unlock(mm);
1009                 goto out_nolock;
1010         }
1011
1012         if (unmapped) {
1013                 /*
1014                  * __collapse_huge_page_swapin will return with mmap_lock
1015                  * released when it fails. So we jump out_nolock directly in
1016                  * that case.  Continuing to collapse causes inconsistency.
1017                  */
1018                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1019                                                      referenced);
1020                 if (result != SCAN_SUCCEED)
1021                         goto out_nolock;
1022         }
1023
1024         mmap_read_unlock(mm);
1025         /*
1026          * Prevent all access to pagetables with the exception of
1027          * gup_fast later handled by the ptep_clear_flush and the VM
1028          * handled by the anon_vma lock + PG_lock.
1029          */
1030         mmap_write_lock(mm);
1031         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1032         if (result != SCAN_SUCCEED)
1033                 goto out_up_write;
1034         /* check if the pmd is still valid */
1035         result = check_pmd_still_valid(mm, address, pmd);
1036         if (result != SCAN_SUCCEED)
1037                 goto out_up_write;
1038
1039         anon_vma_lock_write(vma->anon_vma);
1040
1041         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1042                                 address, address + HPAGE_PMD_SIZE);
1043         mmu_notifier_invalidate_range_start(&range);
1044
1045         pte = pte_offset_map(pmd, address);
1046         pte_ptl = pte_lockptr(mm, pmd);
1047
1048         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1049         /*
1050          * This removes any huge TLB entry from the CPU so we won't allow
1051          * huge and small TLB entries for the same virtual address to
1052          * avoid the risk of CPU bugs in that area.
1053          *
1054          * Parallel fast GUP is fine since fast GUP will back off when
1055          * it detects PMD is changed.
1056          */
1057         _pmd = pmdp_collapse_flush(vma, address, pmd);
1058         spin_unlock(pmd_ptl);
1059         mmu_notifier_invalidate_range_end(&range);
1060
1061         spin_lock(pte_ptl);
1062         result =  __collapse_huge_page_isolate(vma, address, pte, cc,
1063                                                &compound_pagelist);
1064         spin_unlock(pte_ptl);
1065
1066         if (unlikely(result != SCAN_SUCCEED)) {
1067                 pte_unmap(pte);
1068                 spin_lock(pmd_ptl);
1069                 BUG_ON(!pmd_none(*pmd));
1070                 /*
1071                  * We can only use set_pmd_at when establishing
1072                  * hugepmds and never for establishing regular pmds that
1073                  * points to regular pagetables. Use pmd_populate for that
1074                  */
1075                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1076                 spin_unlock(pmd_ptl);
1077                 anon_vma_unlock_write(vma->anon_vma);
1078                 goto out_up_write;
1079         }
1080
1081         /*
1082          * All pages are isolated and locked so anon_vma rmap
1083          * can't run anymore.
1084          */
1085         anon_vma_unlock_write(vma->anon_vma);
1086
1087         __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl,
1088                                   &compound_pagelist);
1089         pte_unmap(pte);
1090         /*
1091          * spin_lock() below is not the equivalent of smp_wmb(), but
1092          * the smp_wmb() inside __SetPageUptodate() can be reused to
1093          * avoid the copy_huge_page writes to become visible after
1094          * the set_pmd_at() write.
1095          */
1096         __SetPageUptodate(hpage);
1097         pgtable = pmd_pgtable(_pmd);
1098
1099         _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1100         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1101
1102         spin_lock(pmd_ptl);
1103         BUG_ON(!pmd_none(*pmd));
1104         page_add_new_anon_rmap(hpage, vma, address);
1105         lru_cache_add_inactive_or_unevictable(hpage, vma);
1106         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1107         set_pmd_at(mm, address, pmd, _pmd);
1108         update_mmu_cache_pmd(vma, address, pmd);
1109         spin_unlock(pmd_ptl);
1110
1111         hpage = NULL;
1112
1113         result = SCAN_SUCCEED;
1114 out_up_write:
1115         mmap_write_unlock(mm);
1116 out_nolock:
1117         if (hpage) {
1118                 mem_cgroup_uncharge(page_folio(hpage));
1119                 put_page(hpage);
1120         }
1121         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1122         return result;
1123 }
1124
1125 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1126                                    struct vm_area_struct *vma,
1127                                    unsigned long address, bool *mmap_locked,
1128                                    struct collapse_control *cc)
1129 {
1130         pmd_t *pmd;
1131         pte_t *pte, *_pte;
1132         int result = SCAN_FAIL, referenced = 0;
1133         int none_or_zero = 0, shared = 0;
1134         struct page *page = NULL;
1135         unsigned long _address;
1136         spinlock_t *ptl;
1137         int node = NUMA_NO_NODE, unmapped = 0;
1138         bool writable = false;
1139
1140         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1141
1142         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1143         if (result != SCAN_SUCCEED)
1144                 goto out;
1145
1146         memset(cc->node_load, 0, sizeof(cc->node_load));
1147         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1148         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1149              _pte++, _address += PAGE_SIZE) {
1150                 pte_t pteval = *_pte;
1151                 if (is_swap_pte(pteval)) {
1152                         ++unmapped;
1153                         if (!cc->is_khugepaged ||
1154                             unmapped <= khugepaged_max_ptes_swap) {
1155                                 /*
1156                                  * Always be strict with uffd-wp
1157                                  * enabled swap entries.  Please see
1158                                  * comment below for pte_uffd_wp().
1159                                  */
1160                                 if (pte_swp_uffd_wp(pteval)) {
1161                                         result = SCAN_PTE_UFFD_WP;
1162                                         goto out_unmap;
1163                                 }
1164                                 continue;
1165                         } else {
1166                                 result = SCAN_EXCEED_SWAP_PTE;
1167                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1168                                 goto out_unmap;
1169                         }
1170                 }
1171                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1172                         ++none_or_zero;
1173                         if (!userfaultfd_armed(vma) &&
1174                             (!cc->is_khugepaged ||
1175                              none_or_zero <= khugepaged_max_ptes_none)) {
1176                                 continue;
1177                         } else {
1178                                 result = SCAN_EXCEED_NONE_PTE;
1179                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1180                                 goto out_unmap;
1181                         }
1182                 }
1183                 if (pte_uffd_wp(pteval)) {
1184                         /*
1185                          * Don't collapse the page if any of the small
1186                          * PTEs are armed with uffd write protection.
1187                          * Here we can also mark the new huge pmd as
1188                          * write protected if any of the small ones is
1189                          * marked but that could bring unknown
1190                          * userfault messages that falls outside of
1191                          * the registered range.  So, just be simple.
1192                          */
1193                         result = SCAN_PTE_UFFD_WP;
1194                         goto out_unmap;
1195                 }
1196                 if (pte_write(pteval))
1197                         writable = true;
1198
1199                 page = vm_normal_page(vma, _address, pteval);
1200                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1201                         result = SCAN_PAGE_NULL;
1202                         goto out_unmap;
1203                 }
1204
1205                 if (page_mapcount(page) > 1) {
1206                         ++shared;
1207                         if (cc->is_khugepaged &&
1208                             shared > khugepaged_max_ptes_shared) {
1209                                 result = SCAN_EXCEED_SHARED_PTE;
1210                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1211                                 goto out_unmap;
1212                         }
1213                 }
1214
1215                 page = compound_head(page);
1216
1217                 /*
1218                  * Record which node the original page is from and save this
1219                  * information to cc->node_load[].
1220                  * Khugepaged will allocate hugepage from the node has the max
1221                  * hit record.
1222                  */
1223                 node = page_to_nid(page);
1224                 if (hpage_collapse_scan_abort(node, cc)) {
1225                         result = SCAN_SCAN_ABORT;
1226                         goto out_unmap;
1227                 }
1228                 cc->node_load[node]++;
1229                 if (!PageLRU(page)) {
1230                         result = SCAN_PAGE_LRU;
1231                         goto out_unmap;
1232                 }
1233                 if (PageLocked(page)) {
1234                         result = SCAN_PAGE_LOCK;
1235                         goto out_unmap;
1236                 }
1237                 if (!PageAnon(page)) {
1238                         result = SCAN_PAGE_ANON;
1239                         goto out_unmap;
1240                 }
1241
1242                 /*
1243                  * Check if the page has any GUP (or other external) pins.
1244                  *
1245                  * Here the check is racy it may see total_mapcount > refcount
1246                  * in some cases.
1247                  * For example, one process with one forked child process.
1248                  * The parent has the PMD split due to MADV_DONTNEED, then
1249                  * the child is trying unmap the whole PMD, but khugepaged
1250                  * may be scanning the parent between the child has
1251                  * PageDoubleMap flag cleared and dec the mapcount.  So
1252                  * khugepaged may see total_mapcount > refcount.
1253                  *
1254                  * But such case is ephemeral we could always retry collapse
1255                  * later.  However it may report false positive if the page
1256                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1257                  * will be done again later the risk seems low.
1258                  */
1259                 if (!is_refcount_suitable(page)) {
1260                         result = SCAN_PAGE_COUNT;
1261                         goto out_unmap;
1262                 }
1263
1264                 /*
1265                  * If collapse was initiated by khugepaged, check that there is
1266                  * enough young pte to justify collapsing the page
1267                  */
1268                 if (cc->is_khugepaged &&
1269                     (pte_young(pteval) || page_is_young(page) ||
1270                      PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
1271                                                                      address)))
1272                         referenced++;
1273         }
1274         if (!writable) {
1275                 result = SCAN_PAGE_RO;
1276         } else if (cc->is_khugepaged &&
1277                    (!referenced ||
1278                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1279                 result = SCAN_LACK_REFERENCED_PAGE;
1280         } else {
1281                 result = SCAN_SUCCEED;
1282         }
1283 out_unmap:
1284         pte_unmap_unlock(pte, ptl);
1285         if (result == SCAN_SUCCEED) {
1286                 result = collapse_huge_page(mm, address, referenced,
1287                                             unmapped, cc);
1288                 /* collapse_huge_page will return with the mmap_lock released */
1289                 *mmap_locked = false;
1290         }
1291 out:
1292         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1293                                      none_or_zero, result, unmapped);
1294         return result;
1295 }
1296
1297 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1298 {
1299         struct mm_slot *slot = &mm_slot->slot;
1300         struct mm_struct *mm = slot->mm;
1301
1302         lockdep_assert_held(&khugepaged_mm_lock);
1303
1304         if (hpage_collapse_test_exit(mm)) {
1305                 /* free mm_slot */
1306                 hash_del(&slot->hash);
1307                 list_del(&slot->mm_node);
1308
1309                 /*
1310                  * Not strictly needed because the mm exited already.
1311                  *
1312                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1313                  */
1314
1315                 /* khugepaged_mm_lock actually not necessary for the below */
1316                 mm_slot_free(mm_slot_cache, mm_slot);
1317                 mmdrop(mm);
1318         }
1319 }
1320
1321 #ifdef CONFIG_SHMEM
1322 /*
1323  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1324  * khugepaged should try to collapse the page table.
1325  *
1326  * Note that following race exists:
1327  * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A,
1328  *     emptying the A's ->pte_mapped_thp[] array.
1329  * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and
1330  *     retract_page_tables() finds a VMA in mm_struct A mapping the same extent
1331  *     (at virtual address X) and adds an entry (for X) into mm_struct A's
1332  *     ->pte-mapped_thp[] array.
1333  * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X,
1334  *     sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry
1335  *     (for X) into mm_struct A's ->pte-mapped_thp[] array.
1336  * Thus, it's possible the same address is added multiple times for the same
1337  * mm_struct.  Should this happen, we'll simply attempt
1338  * collapse_pte_mapped_thp() multiple times for the same address, under the same
1339  * exclusive mmap_lock, and assuming the first call is successful, subsequent
1340  * attempts will return quickly (without grabbing any additional locks) when
1341  * a huge pmd is found in find_pmd_or_thp_or_none().  Since this is a cheap
1342  * check, and since this is a rare occurrence, the cost of preventing this
1343  * "multiple-add" is thought to be more expensive than just handling it, should
1344  * it occur.
1345  */
1346 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1347                                           unsigned long addr)
1348 {
1349         struct khugepaged_mm_slot *mm_slot;
1350         struct mm_slot *slot;
1351         bool ret = false;
1352
1353         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1354
1355         spin_lock(&khugepaged_mm_lock);
1356         slot = mm_slot_lookup(mm_slots_hash, mm);
1357         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
1358         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) {
1359                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1360                 ret = true;
1361         }
1362         spin_unlock(&khugepaged_mm_lock);
1363         return ret;
1364 }
1365
1366 /* hpage must be locked, and mmap_lock must be held in write */
1367 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1368                         pmd_t *pmdp, struct page *hpage)
1369 {
1370         struct vm_fault vmf = {
1371                 .vma = vma,
1372                 .address = addr,
1373                 .flags = 0,
1374                 .pmd = pmdp,
1375         };
1376
1377         VM_BUG_ON(!PageTransHuge(hpage));
1378         mmap_assert_write_locked(vma->vm_mm);
1379
1380         if (do_set_pmd(&vmf, hpage))
1381                 return SCAN_FAIL;
1382
1383         get_page(hpage);
1384         return SCAN_SUCCEED;
1385 }
1386
1387 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1388                                   unsigned long addr, pmd_t *pmdp)
1389 {
1390         spinlock_t *ptl;
1391         pmd_t pmd;
1392
1393         mmap_assert_write_locked(mm);
1394         ptl = pmd_lock(vma->vm_mm, pmdp);
1395         pmd = pmdp_collapse_flush(vma, addr, pmdp);
1396         spin_unlock(ptl);
1397         mm_dec_nr_ptes(mm);
1398         page_table_check_pte_clear_range(mm, addr, pmd);
1399         pte_free(mm, pmd_pgtable(pmd));
1400 }
1401
1402 /**
1403  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1404  * address haddr.
1405  *
1406  * @mm: process address space where collapse happens
1407  * @addr: THP collapse address
1408  * @install_pmd: If a huge PMD should be installed
1409  *
1410  * This function checks whether all the PTEs in the PMD are pointing to the
1411  * right THP. If so, retract the page table so the THP can refault in with
1412  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1413  */
1414 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1415                             bool install_pmd)
1416 {
1417         unsigned long haddr = addr & HPAGE_PMD_MASK;
1418         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1419         struct page *hpage;
1420         pte_t *start_pte, *pte;
1421         pmd_t *pmd;
1422         spinlock_t *ptl;
1423         int count = 0, result = SCAN_FAIL;
1424         int i;
1425
1426         mmap_assert_write_locked(mm);
1427
1428         /* Fast check before locking page if already PMD-mapped */
1429         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1430         if (result == SCAN_PMD_MAPPED)
1431                 return result;
1432
1433         if (!vma || !vma->vm_file ||
1434             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1435                 return SCAN_VMA_CHECK;
1436
1437         /*
1438          * If we are here, we've succeeded in replacing all the native pages
1439          * in the page cache with a single hugepage. If a mm were to fault-in
1440          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1441          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1442          * analogously elide sysfs THP settings here.
1443          */
1444         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
1445                 return SCAN_VMA_CHECK;
1446
1447         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1448         if (userfaultfd_wp(vma))
1449                 return SCAN_PTE_UFFD_WP;
1450
1451         hpage = find_lock_page(vma->vm_file->f_mapping,
1452                                linear_page_index(vma, haddr));
1453         if (!hpage)
1454                 return SCAN_PAGE_NULL;
1455
1456         if (!PageHead(hpage)) {
1457                 result = SCAN_FAIL;
1458                 goto drop_hpage;
1459         }
1460
1461         if (compound_order(hpage) != HPAGE_PMD_ORDER) {
1462                 result = SCAN_PAGE_COMPOUND;
1463                 goto drop_hpage;
1464         }
1465
1466         switch (result) {
1467         case SCAN_SUCCEED:
1468                 break;
1469         case SCAN_PMD_NONE:
1470                 /*
1471                  * In MADV_COLLAPSE path, possible race with khugepaged where
1472                  * all pte entries have been removed and pmd cleared.  If so,
1473                  * skip all the pte checks and just update the pmd mapping.
1474                  */
1475                 goto maybe_install_pmd;
1476         default:
1477                 goto drop_hpage;
1478         }
1479
1480         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1481         result = SCAN_FAIL;
1482
1483         /* step 1: check all mapped PTEs are to the right huge page */
1484         for (i = 0, addr = haddr, pte = start_pte;
1485              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1486                 struct page *page;
1487
1488                 /* empty pte, skip */
1489                 if (pte_none(*pte))
1490                         continue;
1491
1492                 /* page swapped out, abort */
1493                 if (!pte_present(*pte)) {
1494                         result = SCAN_PTE_NON_PRESENT;
1495                         goto abort;
1496                 }
1497
1498                 page = vm_normal_page(vma, addr, *pte);
1499                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1500                         page = NULL;
1501                 /*
1502                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1503                  * page table, but the new page will not be a subpage of hpage.
1504                  */
1505                 if (hpage + i != page)
1506                         goto abort;
1507                 count++;
1508         }
1509
1510         /* step 2: adjust rmap */
1511         for (i = 0, addr = haddr, pte = start_pte;
1512              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1513                 struct page *page;
1514
1515                 if (pte_none(*pte))
1516                         continue;
1517                 page = vm_normal_page(vma, addr, *pte);
1518                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1519                         goto abort;
1520                 page_remove_rmap(page, vma, false);
1521         }
1522
1523         pte_unmap_unlock(start_pte, ptl);
1524
1525         /* step 3: set proper refcount and mm_counters. */
1526         if (count) {
1527                 page_ref_sub(hpage, count);
1528                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1529         }
1530
1531         /* step 4: remove pte entries */
1532         collapse_and_free_pmd(mm, vma, haddr, pmd);
1533
1534 maybe_install_pmd:
1535         /* step 5: install pmd entry */
1536         result = install_pmd
1537                         ? set_huge_pmd(vma, haddr, pmd, hpage)
1538                         : SCAN_SUCCEED;
1539
1540 drop_hpage:
1541         unlock_page(hpage);
1542         put_page(hpage);
1543         return result;
1544
1545 abort:
1546         pte_unmap_unlock(start_pte, ptl);
1547         goto drop_hpage;
1548 }
1549
1550 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
1551 {
1552         struct mm_slot *slot = &mm_slot->slot;
1553         struct mm_struct *mm = slot->mm;
1554         int i;
1555
1556         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1557                 return;
1558
1559         if (!mmap_write_trylock(mm))
1560                 return;
1561
1562         if (unlikely(hpage_collapse_test_exit(mm)))
1563                 goto out;
1564
1565         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1566                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false);
1567
1568 out:
1569         mm_slot->nr_pte_mapped_thp = 0;
1570         mmap_write_unlock(mm);
1571 }
1572
1573 static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
1574                                struct mm_struct *target_mm,
1575                                unsigned long target_addr, struct page *hpage,
1576                                struct collapse_control *cc)
1577 {
1578         struct vm_area_struct *vma;
1579         int target_result = SCAN_FAIL;
1580
1581         i_mmap_lock_write(mapping);
1582         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1583                 int result = SCAN_FAIL;
1584                 struct mm_struct *mm = NULL;
1585                 unsigned long addr = 0;
1586                 pmd_t *pmd;
1587                 bool is_target = false;
1588
1589                 /*
1590                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1591                  * got written to. These VMAs are likely not worth investing
1592                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1593                  * later.
1594                  *
1595                  * Note that vma->anon_vma check is racy: it can be set up after
1596                  * the check but before we took mmap_lock by the fault path.
1597                  * But page lock would prevent establishing any new ptes of the
1598                  * page, so we are safe.
1599                  *
1600                  * An alternative would be drop the check, but check that page
1601                  * table is clear before calling pmdp_collapse_flush() under
1602                  * ptl. It has higher chance to recover THP for the VMA, but
1603                  * has higher cost too.
1604                  */
1605                 if (vma->anon_vma) {
1606                         result = SCAN_PAGE_ANON;
1607                         goto next;
1608                 }
1609                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1610                 if (addr & ~HPAGE_PMD_MASK ||
1611                     vma->vm_end < addr + HPAGE_PMD_SIZE) {
1612                         result = SCAN_VMA_CHECK;
1613                         goto next;
1614                 }
1615                 mm = vma->vm_mm;
1616                 is_target = mm == target_mm && addr == target_addr;
1617                 result = find_pmd_or_thp_or_none(mm, addr, &pmd);
1618                 if (result != SCAN_SUCCEED)
1619                         goto next;
1620                 /*
1621                  * We need exclusive mmap_lock to retract page table.
1622                  *
1623                  * We use trylock due to lock inversion: we need to acquire
1624                  * mmap_lock while holding page lock. Fault path does it in
1625                  * reverse order. Trylock is a way to avoid deadlock.
1626                  *
1627                  * Also, it's not MADV_COLLAPSE's job to collapse other
1628                  * mappings - let khugepaged take care of them later.
1629                  */
1630                 result = SCAN_PTE_MAPPED_HUGEPAGE;
1631                 if ((cc->is_khugepaged || is_target) &&
1632                     mmap_write_trylock(mm)) {
1633                         /*
1634                          * When a vma is registered with uffd-wp, we can't
1635                          * recycle the pmd pgtable because there can be pte
1636                          * markers installed.  Skip it only, so the rest mm/vma
1637                          * can still have the same file mapped hugely, however
1638                          * it'll always mapped in small page size for uffd-wp
1639                          * registered ranges.
1640                          */
1641                         if (hpage_collapse_test_exit(mm)) {
1642                                 result = SCAN_ANY_PROCESS;
1643                                 goto unlock_next;
1644                         }
1645                         if (userfaultfd_wp(vma)) {
1646                                 result = SCAN_PTE_UFFD_WP;
1647                                 goto unlock_next;
1648                         }
1649                         collapse_and_free_pmd(mm, vma, addr, pmd);
1650                         if (!cc->is_khugepaged && is_target)
1651                                 result = set_huge_pmd(vma, addr, pmd, hpage);
1652                         else
1653                                 result = SCAN_SUCCEED;
1654
1655 unlock_next:
1656                         mmap_write_unlock(mm);
1657                         goto next;
1658                 }
1659                 /*
1660                  * Calling context will handle target mm/addr. Otherwise, let
1661                  * khugepaged try again later.
1662                  */
1663                 if (!is_target) {
1664                         khugepaged_add_pte_mapped_thp(mm, addr);
1665                         continue;
1666                 }
1667 next:
1668                 if (is_target)
1669                         target_result = result;
1670         }
1671         i_mmap_unlock_write(mapping);
1672         return target_result;
1673 }
1674
1675 /**
1676  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1677  *
1678  * @mm: process address space where collapse happens
1679  * @addr: virtual collapse start address
1680  * @file: file that collapse on
1681  * @start: collapse start address
1682  * @cc: collapse context and scratchpad
1683  *
1684  * Basic scheme is simple, details are more complex:
1685  *  - allocate and lock a new huge page;
1686  *  - scan page cache replacing old pages with the new one
1687  *    + swap/gup in pages if necessary;
1688  *    + fill in gaps;
1689  *    + keep old pages around in case rollback is required;
1690  *  - if replacing succeeds:
1691  *    + copy data over;
1692  *    + free old pages;
1693  *    + unlock huge page;
1694  *  - if replacing failed;
1695  *    + put all pages back and unfreeze them;
1696  *    + restore gaps in the page cache;
1697  *    + unlock and free huge page;
1698  */
1699 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1700                          struct file *file, pgoff_t start,
1701                          struct collapse_control *cc)
1702 {
1703         struct address_space *mapping = file->f_mapping;
1704         struct page *hpage;
1705         pgoff_t index, end = start + HPAGE_PMD_NR;
1706         LIST_HEAD(pagelist);
1707         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1708         int nr_none = 0, result = SCAN_SUCCEED;
1709         bool is_shmem = shmem_file(file);
1710         int nr;
1711
1712         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1713         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1714
1715         result = alloc_charge_hpage(&hpage, mm, cc);
1716         if (result != SCAN_SUCCEED)
1717                 goto out;
1718
1719         /*
1720          * Ensure we have slots for all the pages in the range.  This is
1721          * almost certainly a no-op because most of the pages must be present
1722          */
1723         do {
1724                 xas_lock_irq(&xas);
1725                 xas_create_range(&xas);
1726                 if (!xas_error(&xas))
1727                         break;
1728                 xas_unlock_irq(&xas);
1729                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1730                         result = SCAN_FAIL;
1731                         goto out;
1732                 }
1733         } while (1);
1734
1735         __SetPageLocked(hpage);
1736         if (is_shmem)
1737                 __SetPageSwapBacked(hpage);
1738         hpage->index = start;
1739         hpage->mapping = mapping;
1740
1741         /*
1742          * At this point the hpage is locked and not up-to-date.
1743          * It's safe to insert it into the page cache, because nobody would
1744          * be able to map it or use it in another way until we unlock it.
1745          */
1746
1747         xas_set(&xas, start);
1748         for (index = start; index < end; index++) {
1749                 struct page *page = xas_next(&xas);
1750
1751                 VM_BUG_ON(index != xas.xa_index);
1752                 if (is_shmem) {
1753                         if (!page) {
1754                                 /*
1755                                  * Stop if extent has been truncated or
1756                                  * hole-punched, and is now completely
1757                                  * empty.
1758                                  */
1759                                 if (index == start) {
1760                                         if (!xas_next_entry(&xas, end - 1)) {
1761                                                 result = SCAN_TRUNCATED;
1762                                                 goto xa_locked;
1763                                         }
1764                                         xas_set(&xas, index);
1765                                 }
1766                                 if (!shmem_charge(mapping->host, 1)) {
1767                                         result = SCAN_FAIL;
1768                                         goto xa_locked;
1769                                 }
1770                                 xas_store(&xas, hpage);
1771                                 nr_none++;
1772                                 continue;
1773                         }
1774
1775                         if (xa_is_value(page) || !PageUptodate(page)) {
1776                                 struct folio *folio;
1777
1778                                 xas_unlock_irq(&xas);
1779                                 /* swap in or instantiate fallocated page */
1780                                 if (shmem_get_folio(mapping->host, index,
1781                                                 &folio, SGP_NOALLOC)) {
1782                                         result = SCAN_FAIL;
1783                                         goto xa_unlocked;
1784                                 }
1785                                 page = folio_file_page(folio, index);
1786                         } else if (trylock_page(page)) {
1787                                 get_page(page);
1788                                 xas_unlock_irq(&xas);
1789                         } else {
1790                                 result = SCAN_PAGE_LOCK;
1791                                 goto xa_locked;
1792                         }
1793                 } else {        /* !is_shmem */
1794                         if (!page || xa_is_value(page)) {
1795                                 xas_unlock_irq(&xas);
1796                                 page_cache_sync_readahead(mapping, &file->f_ra,
1797                                                           file, index,
1798                                                           end - index);
1799                                 /* drain pagevecs to help isolate_lru_page() */
1800                                 lru_add_drain();
1801                                 page = find_lock_page(mapping, index);
1802                                 if (unlikely(page == NULL)) {
1803                                         result = SCAN_FAIL;
1804                                         goto xa_unlocked;
1805                                 }
1806                         } else if (PageDirty(page)) {
1807                                 /*
1808                                  * khugepaged only works on read-only fd,
1809                                  * so this page is dirty because it hasn't
1810                                  * been flushed since first write. There
1811                                  * won't be new dirty pages.
1812                                  *
1813                                  * Trigger async flush here and hope the
1814                                  * writeback is done when khugepaged
1815                                  * revisits this page.
1816                                  *
1817                                  * This is a one-off situation. We are not
1818                                  * forcing writeback in loop.
1819                                  */
1820                                 xas_unlock_irq(&xas);
1821                                 filemap_flush(mapping);
1822                                 result = SCAN_FAIL;
1823                                 goto xa_unlocked;
1824                         } else if (PageWriteback(page)) {
1825                                 xas_unlock_irq(&xas);
1826                                 result = SCAN_FAIL;
1827                                 goto xa_unlocked;
1828                         } else if (trylock_page(page)) {
1829                                 get_page(page);
1830                                 xas_unlock_irq(&xas);
1831                         } else {
1832                                 result = SCAN_PAGE_LOCK;
1833                                 goto xa_locked;
1834                         }
1835                 }
1836
1837                 /*
1838                  * The page must be locked, so we can drop the i_pages lock
1839                  * without racing with truncate.
1840                  */
1841                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1842
1843                 /* make sure the page is up to date */
1844                 if (unlikely(!PageUptodate(page))) {
1845                         result = SCAN_FAIL;
1846                         goto out_unlock;
1847                 }
1848
1849                 /*
1850                  * If file was truncated then extended, or hole-punched, before
1851                  * we locked the first page, then a THP might be there already.
1852                  * This will be discovered on the first iteration.
1853                  */
1854                 if (PageTransCompound(page)) {
1855                         struct page *head = compound_head(page);
1856
1857                         result = compound_order(head) == HPAGE_PMD_ORDER &&
1858                                         head->index == start
1859                                         /* Maybe PMD-mapped */
1860                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1861                                         : SCAN_PAGE_COMPOUND;
1862                         goto out_unlock;
1863                 }
1864
1865                 if (page_mapping(page) != mapping) {
1866                         result = SCAN_TRUNCATED;
1867                         goto out_unlock;
1868                 }
1869
1870                 if (!is_shmem && (PageDirty(page) ||
1871                                   PageWriteback(page))) {
1872                         /*
1873                          * khugepaged only works on read-only fd, so this
1874                          * page is dirty because it hasn't been flushed
1875                          * since first write.
1876                          */
1877                         result = SCAN_FAIL;
1878                         goto out_unlock;
1879                 }
1880
1881                 if (isolate_lru_page(page)) {
1882                         result = SCAN_DEL_PAGE_LRU;
1883                         goto out_unlock;
1884                 }
1885
1886                 if (page_has_private(page) &&
1887                     !try_to_release_page(page, GFP_KERNEL)) {
1888                         result = SCAN_PAGE_HAS_PRIVATE;
1889                         putback_lru_page(page);
1890                         goto out_unlock;
1891                 }
1892
1893                 if (page_mapped(page))
1894                         try_to_unmap(page_folio(page),
1895                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1896
1897                 xas_lock_irq(&xas);
1898                 xas_set(&xas, index);
1899
1900                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1901
1902                 /*
1903                  * The page is expected to have page_count() == 3:
1904                  *  - we hold a pin on it;
1905                  *  - one reference from page cache;
1906                  *  - one from isolate_lru_page;
1907                  */
1908                 if (!page_ref_freeze(page, 3)) {
1909                         result = SCAN_PAGE_COUNT;
1910                         xas_unlock_irq(&xas);
1911                         putback_lru_page(page);
1912                         goto out_unlock;
1913                 }
1914
1915                 /*
1916                  * Add the page to the list to be able to undo the collapse if
1917                  * something go wrong.
1918                  */
1919                 list_add_tail(&page->lru, &pagelist);
1920
1921                 /* Finally, replace with the new page. */
1922                 xas_store(&xas, hpage);
1923                 continue;
1924 out_unlock:
1925                 unlock_page(page);
1926                 put_page(page);
1927                 goto xa_unlocked;
1928         }
1929         nr = thp_nr_pages(hpage);
1930
1931         if (is_shmem)
1932                 __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr);
1933         else {
1934                 __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr);
1935                 filemap_nr_thps_inc(mapping);
1936                 /*
1937                  * Paired with smp_mb() in do_dentry_open() to ensure
1938                  * i_writecount is up to date and the update to nr_thps is
1939                  * visible. Ensures the page cache will be truncated if the
1940                  * file is opened writable.
1941                  */
1942                 smp_mb();
1943                 if (inode_is_open_for_write(mapping->host)) {
1944                         result = SCAN_FAIL;
1945                         __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr);
1946                         filemap_nr_thps_dec(mapping);
1947                         goto xa_locked;
1948                 }
1949         }
1950
1951         if (nr_none) {
1952                 __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none);
1953                 /* nr_none is always 0 for non-shmem. */
1954                 __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none);
1955         }
1956
1957         /* Join all the small entries into a single multi-index entry */
1958         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1959         xas_store(&xas, hpage);
1960 xa_locked:
1961         xas_unlock_irq(&xas);
1962 xa_unlocked:
1963
1964         /*
1965          * If collapse is successful, flush must be done now before copying.
1966          * If collapse is unsuccessful, does flush actually need to be done?
1967          * Do it anyway, to clear the state.
1968          */
1969         try_to_unmap_flush();
1970
1971         if (result == SCAN_SUCCEED) {
1972                 struct page *page, *tmp;
1973
1974                 /*
1975                  * Replacing old pages with new one has succeeded, now we
1976                  * need to copy the content and free the old pages.
1977                  */
1978                 index = start;
1979                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1980                         while (index < page->index) {
1981                                 clear_highpage(hpage + (index % HPAGE_PMD_NR));
1982                                 index++;
1983                         }
1984                         copy_highpage(hpage + (page->index % HPAGE_PMD_NR),
1985                                       page);
1986                         list_del(&page->lru);
1987                         page->mapping = NULL;
1988                         page_ref_unfreeze(page, 1);
1989                         ClearPageActive(page);
1990                         ClearPageUnevictable(page);
1991                         unlock_page(page);
1992                         put_page(page);
1993                         index++;
1994                 }
1995                 while (index < end) {
1996                         clear_highpage(hpage + (index % HPAGE_PMD_NR));
1997                         index++;
1998                 }
1999
2000                 SetPageUptodate(hpage);
2001                 page_ref_add(hpage, HPAGE_PMD_NR - 1);
2002                 if (is_shmem)
2003                         set_page_dirty(hpage);
2004                 lru_cache_add(hpage);
2005
2006                 /*
2007                  * Remove pte page tables, so we can re-fault the page as huge.
2008                  */
2009                 result = retract_page_tables(mapping, start, mm, addr, hpage,
2010                                              cc);
2011                 unlock_page(hpage);
2012                 hpage = NULL;
2013         } else {
2014                 struct page *page;
2015
2016                 /* Something went wrong: roll back page cache changes */
2017                 xas_lock_irq(&xas);
2018                 if (nr_none) {
2019                         mapping->nrpages -= nr_none;
2020                         shmem_uncharge(mapping->host, nr_none);
2021                 }
2022
2023                 xas_set(&xas, start);
2024                 xas_for_each(&xas, page, end - 1) {
2025                         page = list_first_entry_or_null(&pagelist,
2026                                         struct page, lru);
2027                         if (!page || xas.xa_index < page->index) {
2028                                 if (!nr_none)
2029                                         break;
2030                                 nr_none--;
2031                                 /* Put holes back where they were */
2032                                 xas_store(&xas, NULL);
2033                                 continue;
2034                         }
2035
2036                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2037
2038                         /* Unfreeze the page. */
2039                         list_del(&page->lru);
2040                         page_ref_unfreeze(page, 2);
2041                         xas_store(&xas, page);
2042                         xas_pause(&xas);
2043                         xas_unlock_irq(&xas);
2044                         unlock_page(page);
2045                         putback_lru_page(page);
2046                         xas_lock_irq(&xas);
2047                 }
2048                 VM_BUG_ON(nr_none);
2049                 xas_unlock_irq(&xas);
2050
2051                 hpage->mapping = NULL;
2052         }
2053
2054         if (hpage)
2055                 unlock_page(hpage);
2056 out:
2057         VM_BUG_ON(!list_empty(&pagelist));
2058         if (hpage) {
2059                 mem_cgroup_uncharge(page_folio(hpage));
2060                 put_page(hpage);
2061         }
2062         /* TODO: tracepoints */
2063         return result;
2064 }
2065
2066 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2067                                     struct file *file, pgoff_t start,
2068                                     struct collapse_control *cc)
2069 {
2070         struct page *page = NULL;
2071         struct address_space *mapping = file->f_mapping;
2072         XA_STATE(xas, &mapping->i_pages, start);
2073         int present, swap;
2074         int node = NUMA_NO_NODE;
2075         int result = SCAN_SUCCEED;
2076
2077         present = 0;
2078         swap = 0;
2079         memset(cc->node_load, 0, sizeof(cc->node_load));
2080         rcu_read_lock();
2081         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2082                 if (xas_retry(&xas, page))
2083                         continue;
2084
2085                 if (xa_is_value(page)) {
2086                         ++swap;
2087                         if (cc->is_khugepaged &&
2088                             swap > khugepaged_max_ptes_swap) {
2089                                 result = SCAN_EXCEED_SWAP_PTE;
2090                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2091                                 break;
2092                         }
2093                         continue;
2094                 }
2095
2096                 /*
2097                  * TODO: khugepaged should compact smaller compound pages
2098                  * into a PMD sized page
2099                  */
2100                 if (PageTransCompound(page)) {
2101                         struct page *head = compound_head(page);
2102
2103                         result = compound_order(head) == HPAGE_PMD_ORDER &&
2104                                         head->index == start
2105                                         /* Maybe PMD-mapped */
2106                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2107                                         : SCAN_PAGE_COMPOUND;
2108                         /*
2109                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2110                          * by the caller won't touch the page cache, and so
2111                          * it's safe to skip LRU and refcount checks before
2112                          * returning.
2113                          */
2114                         break;
2115                 }
2116
2117                 node = page_to_nid(page);
2118                 if (hpage_collapse_scan_abort(node, cc)) {
2119                         result = SCAN_SCAN_ABORT;
2120                         break;
2121                 }
2122                 cc->node_load[node]++;
2123
2124                 if (!PageLRU(page)) {
2125                         result = SCAN_PAGE_LRU;
2126                         break;
2127                 }
2128
2129                 if (page_count(page) !=
2130                     1 + page_mapcount(page) + page_has_private(page)) {
2131                         result = SCAN_PAGE_COUNT;
2132                         break;
2133                 }
2134
2135                 /*
2136                  * We probably should check if the page is referenced here, but
2137                  * nobody would transfer pte_young() to PageReferenced() for us.
2138                  * And rmap walk here is just too costly...
2139                  */
2140
2141                 present++;
2142
2143                 if (need_resched()) {
2144                         xas_pause(&xas);
2145                         cond_resched_rcu();
2146                 }
2147         }
2148         rcu_read_unlock();
2149
2150         if (result == SCAN_SUCCEED) {
2151                 if (cc->is_khugepaged &&
2152                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2153                         result = SCAN_EXCEED_NONE_PTE;
2154                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2155                 } else {
2156                         result = collapse_file(mm, addr, file, start, cc);
2157                 }
2158         }
2159
2160         trace_mm_khugepaged_scan_file(mm, page, file->f_path.dentry->d_iname,
2161                                       present, swap, result);
2162         return result;
2163 }
2164 #else
2165 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2166                                     struct file *file, pgoff_t start,
2167                                     struct collapse_control *cc)
2168 {
2169         BUILD_BUG();
2170 }
2171
2172 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
2173 {
2174 }
2175
2176 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
2177                                           unsigned long addr)
2178 {
2179         return false;
2180 }
2181 #endif
2182
2183 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2184                                             struct collapse_control *cc)
2185         __releases(&khugepaged_mm_lock)
2186         __acquires(&khugepaged_mm_lock)
2187 {
2188         struct vma_iterator vmi;
2189         struct khugepaged_mm_slot *mm_slot;
2190         struct mm_slot *slot;
2191         struct mm_struct *mm;
2192         struct vm_area_struct *vma;
2193         int progress = 0;
2194
2195         VM_BUG_ON(!pages);
2196         lockdep_assert_held(&khugepaged_mm_lock);
2197         *result = SCAN_FAIL;
2198
2199         if (khugepaged_scan.mm_slot) {
2200                 mm_slot = khugepaged_scan.mm_slot;
2201                 slot = &mm_slot->slot;
2202         } else {
2203                 slot = list_entry(khugepaged_scan.mm_head.next,
2204                                      struct mm_slot, mm_node);
2205                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2206                 khugepaged_scan.address = 0;
2207                 khugepaged_scan.mm_slot = mm_slot;
2208         }
2209         spin_unlock(&khugepaged_mm_lock);
2210         khugepaged_collapse_pte_mapped_thps(mm_slot);
2211
2212         mm = slot->mm;
2213         /*
2214          * Don't wait for semaphore (to avoid long wait times).  Just move to
2215          * the next mm on the list.
2216          */
2217         vma = NULL;
2218         if (unlikely(!mmap_read_trylock(mm)))
2219                 goto breakouterloop_mmap_lock;
2220
2221         progress++;
2222         if (unlikely(hpage_collapse_test_exit(mm)))
2223                 goto breakouterloop;
2224
2225         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2226         for_each_vma(vmi, vma) {
2227                 unsigned long hstart, hend;
2228
2229                 cond_resched();
2230                 if (unlikely(hpage_collapse_test_exit(mm))) {
2231                         progress++;
2232                         break;
2233                 }
2234                 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) {
2235 skip:
2236                         progress++;
2237                         continue;
2238                 }
2239                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2240                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2241                 if (khugepaged_scan.address > hend)
2242                         goto skip;
2243                 if (khugepaged_scan.address < hstart)
2244                         khugepaged_scan.address = hstart;
2245                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2246
2247                 while (khugepaged_scan.address < hend) {
2248                         bool mmap_locked = true;
2249
2250                         cond_resched();
2251                         if (unlikely(hpage_collapse_test_exit(mm)))
2252                                 goto breakouterloop;
2253
2254                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2255                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2256                                   hend);
2257                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2258                                 struct file *file = get_file(vma->vm_file);
2259                                 pgoff_t pgoff = linear_page_index(vma,
2260                                                 khugepaged_scan.address);
2261
2262                                 mmap_read_unlock(mm);
2263                                 *result = hpage_collapse_scan_file(mm,
2264                                                                    khugepaged_scan.address,
2265                                                                    file, pgoff, cc);
2266                                 mmap_locked = false;
2267                                 fput(file);
2268                         } else {
2269                                 *result = hpage_collapse_scan_pmd(mm, vma,
2270                                                                   khugepaged_scan.address,
2271                                                                   &mmap_locked,
2272                                                                   cc);
2273                         }
2274                         switch (*result) {
2275                         case SCAN_PTE_MAPPED_HUGEPAGE: {
2276                                 pmd_t *pmd;
2277
2278                                 *result = find_pmd_or_thp_or_none(mm,
2279                                                                   khugepaged_scan.address,
2280                                                                   &pmd);
2281                                 if (*result != SCAN_SUCCEED)
2282                                         break;
2283                                 if (!khugepaged_add_pte_mapped_thp(mm,
2284                                                                    khugepaged_scan.address))
2285                                         break;
2286                         } fallthrough;
2287                         case SCAN_SUCCEED:
2288                                 ++khugepaged_pages_collapsed;
2289                                 break;
2290                         default:
2291                                 break;
2292                         }
2293
2294                         /* move to next address */
2295                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2296                         progress += HPAGE_PMD_NR;
2297                         if (!mmap_locked)
2298                                 /*
2299                                  * We released mmap_lock so break loop.  Note
2300                                  * that we drop mmap_lock before all hugepage
2301                                  * allocations, so if allocation fails, we are
2302                                  * guaranteed to break here and report the
2303                                  * correct result back to caller.
2304                                  */
2305                                 goto breakouterloop_mmap_lock;
2306                         if (progress >= pages)
2307                                 goto breakouterloop;
2308                 }
2309         }
2310 breakouterloop:
2311         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2312 breakouterloop_mmap_lock:
2313
2314         spin_lock(&khugepaged_mm_lock);
2315         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2316         /*
2317          * Release the current mm_slot if this mm is about to die, or
2318          * if we scanned all vmas of this mm.
2319          */
2320         if (hpage_collapse_test_exit(mm) || !vma) {
2321                 /*
2322                  * Make sure that if mm_users is reaching zero while
2323                  * khugepaged runs here, khugepaged_exit will find
2324                  * mm_slot not pointing to the exiting mm.
2325                  */
2326                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2327                         slot = list_entry(slot->mm_node.next,
2328                                           struct mm_slot, mm_node);
2329                         khugepaged_scan.mm_slot =
2330                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2331                         khugepaged_scan.address = 0;
2332                 } else {
2333                         khugepaged_scan.mm_slot = NULL;
2334                         khugepaged_full_scans++;
2335                 }
2336
2337                 collect_mm_slot(mm_slot);
2338         }
2339
2340         return progress;
2341 }
2342
2343 static int khugepaged_has_work(void)
2344 {
2345         return !list_empty(&khugepaged_scan.mm_head) &&
2346                 hugepage_flags_enabled();
2347 }
2348
2349 static int khugepaged_wait_event(void)
2350 {
2351         return !list_empty(&khugepaged_scan.mm_head) ||
2352                 kthread_should_stop();
2353 }
2354
2355 static void khugepaged_do_scan(struct collapse_control *cc)
2356 {
2357         unsigned int progress = 0, pass_through_head = 0;
2358         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2359         bool wait = true;
2360         int result = SCAN_SUCCEED;
2361
2362         lru_add_drain_all();
2363
2364         while (true) {
2365                 cond_resched();
2366
2367                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2368                         break;
2369
2370                 spin_lock(&khugepaged_mm_lock);
2371                 if (!khugepaged_scan.mm_slot)
2372                         pass_through_head++;
2373                 if (khugepaged_has_work() &&
2374                     pass_through_head < 2)
2375                         progress += khugepaged_scan_mm_slot(pages - progress,
2376                                                             &result, cc);
2377                 else
2378                         progress = pages;
2379                 spin_unlock(&khugepaged_mm_lock);
2380
2381                 if (progress >= pages)
2382                         break;
2383
2384                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2385                         /*
2386                          * If fail to allocate the first time, try to sleep for
2387                          * a while.  When hit again, cancel the scan.
2388                          */
2389                         if (!wait)
2390                                 break;
2391                         wait = false;
2392                         khugepaged_alloc_sleep();
2393                 }
2394         }
2395 }
2396
2397 static bool khugepaged_should_wakeup(void)
2398 {
2399         return kthread_should_stop() ||
2400                time_after_eq(jiffies, khugepaged_sleep_expire);
2401 }
2402
2403 static void khugepaged_wait_work(void)
2404 {
2405         if (khugepaged_has_work()) {
2406                 const unsigned long scan_sleep_jiffies =
2407                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2408
2409                 if (!scan_sleep_jiffies)
2410                         return;
2411
2412                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2413                 wait_event_freezable_timeout(khugepaged_wait,
2414                                              khugepaged_should_wakeup(),
2415                                              scan_sleep_jiffies);
2416                 return;
2417         }
2418
2419         if (hugepage_flags_enabled())
2420                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2421 }
2422
2423 static int khugepaged(void *none)
2424 {
2425         struct khugepaged_mm_slot *mm_slot;
2426
2427         set_freezable();
2428         set_user_nice(current, MAX_NICE);
2429
2430         while (!kthread_should_stop()) {
2431                 khugepaged_do_scan(&khugepaged_collapse_control);
2432                 khugepaged_wait_work();
2433         }
2434
2435         spin_lock(&khugepaged_mm_lock);
2436         mm_slot = khugepaged_scan.mm_slot;
2437         khugepaged_scan.mm_slot = NULL;
2438         if (mm_slot)
2439                 collect_mm_slot(mm_slot);
2440         spin_unlock(&khugepaged_mm_lock);
2441         return 0;
2442 }
2443
2444 static void set_recommended_min_free_kbytes(void)
2445 {
2446         struct zone *zone;
2447         int nr_zones = 0;
2448         unsigned long recommended_min;
2449
2450         if (!hugepage_flags_enabled()) {
2451                 calculate_min_free_kbytes();
2452                 goto update_wmarks;
2453         }
2454
2455         for_each_populated_zone(zone) {
2456                 /*
2457                  * We don't need to worry about fragmentation of
2458                  * ZONE_MOVABLE since it only has movable pages.
2459                  */
2460                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2461                         continue;
2462
2463                 nr_zones++;
2464         }
2465
2466         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2467         recommended_min = pageblock_nr_pages * nr_zones * 2;
2468
2469         /*
2470          * Make sure that on average at least two pageblocks are almost free
2471          * of another type, one for a migratetype to fall back to and a
2472          * second to avoid subsequent fallbacks of other types There are 3
2473          * MIGRATE_TYPES we care about.
2474          */
2475         recommended_min += pageblock_nr_pages * nr_zones *
2476                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2477
2478         /* don't ever allow to reserve more than 5% of the lowmem */
2479         recommended_min = min(recommended_min,
2480                               (unsigned long) nr_free_buffer_pages() / 20);
2481         recommended_min <<= (PAGE_SHIFT-10);
2482
2483         if (recommended_min > min_free_kbytes) {
2484                 if (user_min_free_kbytes >= 0)
2485                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2486                                 min_free_kbytes, recommended_min);
2487
2488                 min_free_kbytes = recommended_min;
2489         }
2490
2491 update_wmarks:
2492         setup_per_zone_wmarks();
2493 }
2494
2495 int start_stop_khugepaged(void)
2496 {
2497         int err = 0;
2498
2499         mutex_lock(&khugepaged_mutex);
2500         if (hugepage_flags_enabled()) {
2501                 if (!khugepaged_thread)
2502                         khugepaged_thread = kthread_run(khugepaged, NULL,
2503                                                         "khugepaged");
2504                 if (IS_ERR(khugepaged_thread)) {
2505                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2506                         err = PTR_ERR(khugepaged_thread);
2507                         khugepaged_thread = NULL;
2508                         goto fail;
2509                 }
2510
2511                 if (!list_empty(&khugepaged_scan.mm_head))
2512                         wake_up_interruptible(&khugepaged_wait);
2513         } else if (khugepaged_thread) {
2514                 kthread_stop(khugepaged_thread);
2515                 khugepaged_thread = NULL;
2516         }
2517         set_recommended_min_free_kbytes();
2518 fail:
2519         mutex_unlock(&khugepaged_mutex);
2520         return err;
2521 }
2522
2523 void khugepaged_min_free_kbytes_update(void)
2524 {
2525         mutex_lock(&khugepaged_mutex);
2526         if (hugepage_flags_enabled() && khugepaged_thread)
2527                 set_recommended_min_free_kbytes();
2528         mutex_unlock(&khugepaged_mutex);
2529 }
2530
2531 static int madvise_collapse_errno(enum scan_result r)
2532 {
2533         /*
2534          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2535          * actionable feedback to caller, so they may take an appropriate
2536          * fallback measure depending on the nature of the failure.
2537          */
2538         switch (r) {
2539         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2540                 return -ENOMEM;
2541         case SCAN_CGROUP_CHARGE_FAIL:
2542                 return -EBUSY;
2543         /* Resource temporary unavailable - trying again might succeed */
2544         case SCAN_PAGE_LOCK:
2545         case SCAN_PAGE_LRU:
2546         case SCAN_DEL_PAGE_LRU:
2547                 return -EAGAIN;
2548         /*
2549          * Other: Trying again likely not to succeed / error intrinsic to
2550          * specified memory range. khugepaged likely won't be able to collapse
2551          * either.
2552          */
2553         default:
2554                 return -EINVAL;
2555         }
2556 }
2557
2558 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2559                      unsigned long start, unsigned long end)
2560 {
2561         struct collapse_control *cc;
2562         struct mm_struct *mm = vma->vm_mm;
2563         unsigned long hstart, hend, addr;
2564         int thps = 0, last_fail = SCAN_FAIL;
2565         bool mmap_locked = true;
2566
2567         BUG_ON(vma->vm_start > start);
2568         BUG_ON(vma->vm_end < end);
2569
2570         *prev = vma;
2571
2572         if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
2573                 return -EINVAL;
2574
2575         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2576         if (!cc)
2577                 return -ENOMEM;
2578         cc->is_khugepaged = false;
2579         cc->last_target_node = NUMA_NO_NODE;
2580
2581         mmgrab(mm);
2582         lru_add_drain_all();
2583
2584         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2585         hend = end & HPAGE_PMD_MASK;
2586
2587         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2588                 int result = SCAN_FAIL;
2589
2590                 if (!mmap_locked) {
2591                         cond_resched();
2592                         mmap_read_lock(mm);
2593                         mmap_locked = true;
2594                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2595                                                          cc);
2596                         if (result  != SCAN_SUCCEED) {
2597                                 last_fail = result;
2598                                 goto out_nolock;
2599                         }
2600
2601                         hend = vma->vm_end & HPAGE_PMD_MASK;
2602                 }
2603                 mmap_assert_locked(mm);
2604                 memset(cc->node_load, 0, sizeof(cc->node_load));
2605                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2606                         struct file *file = get_file(vma->vm_file);
2607                         pgoff_t pgoff = linear_page_index(vma, addr);
2608
2609                         mmap_read_unlock(mm);
2610                         mmap_locked = false;
2611                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2612                                                           cc);
2613                         fput(file);
2614                 } else {
2615                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2616                                                          &mmap_locked, cc);
2617                 }
2618                 if (!mmap_locked)
2619                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2620
2621 handle_result:
2622                 switch (result) {
2623                 case SCAN_SUCCEED:
2624                 case SCAN_PMD_MAPPED:
2625                         ++thps;
2626                         break;
2627                 case SCAN_PTE_MAPPED_HUGEPAGE:
2628                         BUG_ON(mmap_locked);
2629                         BUG_ON(*prev);
2630                         mmap_write_lock(mm);
2631                         result = collapse_pte_mapped_thp(mm, addr, true);
2632                         mmap_write_unlock(mm);
2633                         goto handle_result;
2634                 /* Whitelisted set of results where continuing OK */
2635                 case SCAN_PMD_NULL:
2636                 case SCAN_PTE_NON_PRESENT:
2637                 case SCAN_PTE_UFFD_WP:
2638                 case SCAN_PAGE_RO:
2639                 case SCAN_LACK_REFERENCED_PAGE:
2640                 case SCAN_PAGE_NULL:
2641                 case SCAN_PAGE_COUNT:
2642                 case SCAN_PAGE_LOCK:
2643                 case SCAN_PAGE_COMPOUND:
2644                 case SCAN_PAGE_LRU:
2645                 case SCAN_DEL_PAGE_LRU:
2646                         last_fail = result;
2647                         break;
2648                 default:
2649                         last_fail = result;
2650                         /* Other error, exit */
2651                         goto out_maybelock;
2652                 }
2653         }
2654
2655 out_maybelock:
2656         /* Caller expects us to hold mmap_lock on return */
2657         if (!mmap_locked)
2658                 mmap_read_lock(mm);
2659 out_nolock:
2660         mmap_assert_locked(mm);
2661         mmdrop(mm);
2662         kfree(cc);
2663
2664         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2665                         : madvise_collapse_errno(last_fail);
2666 }