mm/khugepaged.c

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