mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/kthread.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33 #include <linux/shmem_fs.h>
  34
  35 #include <asm/tlb.h>
  36 #include <asm/pgalloc.h>
  37 #include "internal.h"
  38
  39 enum scan_result {
  40         SCAN_FAIL,
  41         SCAN_SUCCEED,
  42         SCAN_PMD_NULL,
  43         SCAN_EXCEED_NONE_PTE,
  44         SCAN_PTE_NON_PRESENT,
  45         SCAN_PAGE_RO,
  46         SCAN_NO_REFERENCED_PAGE,
  47         SCAN_PAGE_NULL,
  48         SCAN_SCAN_ABORT,
  49         SCAN_PAGE_COUNT,
  50         SCAN_PAGE_LRU,
  51         SCAN_PAGE_LOCK,
  52         SCAN_PAGE_ANON,
  53         SCAN_PAGE_COMPOUND,
  54         SCAN_ANY_PROCESS,
  55         SCAN_VMA_NULL,
  56         SCAN_VMA_CHECK,
  57         SCAN_ADDRESS_RANGE,
  58         SCAN_SWAP_CACHE_PAGE,
  59         SCAN_DEL_PAGE_LRU,
  60         SCAN_ALLOC_HUGE_PAGE_FAIL,
  61         SCAN_CGROUP_CHARGE_FAIL,
  62         SCAN_EXCEED_SWAP_PTE
  63 };
  64
  65 #define CREATE_TRACE_POINTS
  66 #include <trace/events/huge_memory.h>
  67
  68 /*
  69  * By default transparent hugepage support is disabled in order that avoid
  70  * to risk increase the memory footprint of applications without a guaranteed
  71  * benefit. When transparent hugepage support is enabled, is for all mappings,
  72  * and khugepaged scans all mappings.
  73  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  74  * for all hugepage allocations.
  75  */
  76 unsigned long transparent_hugepage_flags __read_mostly =
  77 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  78         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  79 #endif
  80 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  81         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  82 #endif
  83         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  84         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  85         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  86
  87 /* default scan 8*512 pte (or vmas) every 30 second */
  88 static unsigned int khugepaged_pages_to_scan __read_mostly;
  89 static unsigned int khugepaged_pages_collapsed;
  90 static unsigned int khugepaged_full_scans;
  91 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  92 /* during fragmentation poll the hugepage allocator once every minute */
  93 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  94 static unsigned long khugepaged_sleep_expire;
  95 static struct task_struct *khugepaged_thread __read_mostly;
  96 static DEFINE_MUTEX(khugepaged_mutex);
  97 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  98 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  99 /*
 100  * default collapse hugepages if there is at least one pte mapped like
 101  * it would have happened if the vma was large enough during page
 102  * fault.
 103  */
 104 static unsigned int khugepaged_max_ptes_none __read_mostly;
 105 static unsigned int khugepaged_max_ptes_swap __read_mostly;
 106
 107 static int khugepaged(void *none);
 108 static int khugepaged_slab_init(void);
 109 static void khugepaged_slab_exit(void);
 110
 111 #define MM_SLOTS_HASH_BITS 10
 112 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 113
 114 static struct kmem_cache *mm_slot_cache __read_mostly;
 115
 116 /**
 117  * struct mm_slot - hash lookup from mm to mm_slot
 118  * @hash: hash collision list
 119  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 120  * @mm: the mm that this information is valid for
 121  */
 122 struct mm_slot {
 123         struct hlist_node hash;
 124         struct list_head mm_node;
 125         struct mm_struct *mm;
 126 };
 127
 128 /**
 129  * struct khugepaged_scan - cursor for scanning
 130  * @mm_head: the head of the mm list to scan
 131  * @mm_slot: the current mm_slot we are scanning
 132  * @address: the next address inside that to be scanned
 133  *
 134  * There is only the one khugepaged_scan instance of this cursor structure.
 135  */
 136 struct khugepaged_scan {
 137         struct list_head mm_head;
 138         struct mm_slot *mm_slot;
 139         unsigned long address;
 140 };
 141 static struct khugepaged_scan khugepaged_scan = {
 142         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 143 };
 144
 145 static struct shrinker deferred_split_shrinker;
 146
 147 static void set_recommended_min_free_kbytes(void)
 148 {
 149         struct zone *zone;
 150         int nr_zones = 0;
 151         unsigned long recommended_min;
 152
 153         for_each_populated_zone(zone)
 154                 nr_zones++;
 155
 156         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 157         recommended_min = pageblock_nr_pages * nr_zones * 2;
 158
 159         /*
 160          * Make sure that on average at least two pageblocks are almost free
 161          * of another type, one for a migratetype to fall back to and a
 162          * second to avoid subsequent fallbacks of other types There are 3
 163          * MIGRATE_TYPES we care about.
 164          */
 165         recommended_min += pageblock_nr_pages * nr_zones *
 166                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 167
 168         /* don't ever allow to reserve more than 5% of the lowmem */
 169         recommended_min = min(recommended_min,
 170                               (unsigned long) nr_free_buffer_pages() / 20);
 171         recommended_min <<= (PAGE_SHIFT-10);
 172
 173         if (recommended_min > min_free_kbytes) {
 174                 if (user_min_free_kbytes >= 0)
 175                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
 176                                 min_free_kbytes, recommended_min);
 177
 178                 min_free_kbytes = recommended_min;
 179         }
 180         setup_per_zone_wmarks();
 181 }
 182
 183 static int start_stop_khugepaged(void)
 184 {
 185         int err = 0;
 186         if (khugepaged_enabled()) {
 187                 if (!khugepaged_thread)
 188                         khugepaged_thread = kthread_run(khugepaged, NULL,
 189                                                         "khugepaged");
 190                 if (IS_ERR(khugepaged_thread)) {
 191                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 192                         err = PTR_ERR(khugepaged_thread);
 193                         khugepaged_thread = NULL;
 194                         goto fail;
 195                 }
 196
 197                 if (!list_empty(&khugepaged_scan.mm_head))
 198                         wake_up_interruptible(&khugepaged_wait);
 199
 200                 set_recommended_min_free_kbytes();
 201         } else if (khugepaged_thread) {
 202                 kthread_stop(khugepaged_thread);
 203                 khugepaged_thread = NULL;
 204         }
 205 fail:
 206         return err;
 207 }
 208
 209 static atomic_t huge_zero_refcount;
 210 struct page *huge_zero_page __read_mostly;
 211
 212 struct page *get_huge_zero_page(void)
 213 {
 214         struct page *zero_page;
 215 retry:
 216         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 217                 return READ_ONCE(huge_zero_page);
 218
 219         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 220                         HPAGE_PMD_ORDER);
 221         if (!zero_page) {
 222                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 223                 return NULL;
 224         }
 225         count_vm_event(THP_ZERO_PAGE_ALLOC);
 226         preempt_disable();
 227         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 228                 preempt_enable();
 229                 __free_pages(zero_page, compound_order(zero_page));
 230                 goto retry;
 231         }
 232
 233         /* We take additional reference here. It will be put back by shrinker */
 234         atomic_set(&huge_zero_refcount, 2);
 235         preempt_enable();
 236         return READ_ONCE(huge_zero_page);
 237 }
 238
 239 void put_huge_zero_page(void)
 240 {
 241         /*
 242          * Counter should never go to zero here. Only shrinker can put
 243          * last reference.
 244          */
 245         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 246 }
 247
 248 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 249                                         struct shrink_control *sc)
 250 {
 251         /* we can free zero page only if last reference remains */
 252         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 253 }
 254
 255 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 256                                        struct shrink_control *sc)
 257 {
 258         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 259                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 260                 BUG_ON(zero_page == NULL);
 261                 __free_pages(zero_page, compound_order(zero_page));
 262                 return HPAGE_PMD_NR;
 263         }
 264
 265         return 0;
 266 }
 267
 268 static struct shrinker huge_zero_page_shrinker = {
 269         .count_objects = shrink_huge_zero_page_count,
 270         .scan_objects = shrink_huge_zero_page_scan,
 271         .seeks = DEFAULT_SEEKS,
 272 };
 273
 274 #ifdef CONFIG_SYSFS
 275
 276 static ssize_t triple_flag_store(struct kobject *kobj,
 277                                  struct kobj_attribute *attr,
 278                                  const char *buf, size_t count,
 279                                  enum transparent_hugepage_flag enabled,
 280                                  enum transparent_hugepage_flag deferred,
 281                                  enum transparent_hugepage_flag req_madv)
 282 {
 283         if (!memcmp("defer", buf,
 284                     min(sizeof("defer")-1, count))) {
 285                 if (enabled == deferred)
 286                         return -EINVAL;
 287                 clear_bit(enabled, &transparent_hugepage_flags);
 288                 clear_bit(req_madv, &transparent_hugepage_flags);
 289                 set_bit(deferred, &transparent_hugepage_flags);
 290         } else if (!memcmp("always", buf,
 291                     min(sizeof("always")-1, count))) {
 292                 clear_bit(deferred, &transparent_hugepage_flags);
 293                 clear_bit(req_madv, &transparent_hugepage_flags);
 294                 set_bit(enabled, &transparent_hugepage_flags);
 295         } else if (!memcmp("madvise", buf,
 296                            min(sizeof("madvise")-1, count))) {
 297                 clear_bit(enabled, &transparent_hugepage_flags);
 298                 clear_bit(deferred, &transparent_hugepage_flags);
 299                 set_bit(req_madv, &transparent_hugepage_flags);
 300         } else if (!memcmp("never", buf,
 301                            min(sizeof("never")-1, count))) {
 302                 clear_bit(enabled, &transparent_hugepage_flags);
 303                 clear_bit(req_madv, &transparent_hugepage_flags);
 304                 clear_bit(deferred, &transparent_hugepage_flags);
 305         } else
 306                 return -EINVAL;
 307
 308         return count;
 309 }
 310
 311 static ssize_t enabled_show(struct kobject *kobj,
 312                             struct kobj_attribute *attr, char *buf)
 313 {
 314         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 315                 return sprintf(buf, "[always] madvise never\n");
 316         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 317                 return sprintf(buf, "always [madvise] never\n");
 318         else
 319                 return sprintf(buf, "always madvise [never]\n");
 320 }
 321
 322 static ssize_t enabled_store(struct kobject *kobj,
 323                              struct kobj_attribute *attr,
 324                              const char *buf, size_t count)
 325 {
 326         ssize_t ret;
 327
 328         ret = triple_flag_store(kobj, attr, buf, count,
 329                                 TRANSPARENT_HUGEPAGE_FLAG,
 330                                 TRANSPARENT_HUGEPAGE_FLAG,
 331                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 332
 333         if (ret > 0) {
 334                 int err;
 335
 336                 mutex_lock(&khugepaged_mutex);
 337                 err = start_stop_khugepaged();
 338                 mutex_unlock(&khugepaged_mutex);
 339
 340                 if (err)
 341                         ret = err;
 342         }
 343
 344         return ret;
 345 }
 346 static struct kobj_attribute enabled_attr =
 347         __ATTR(enabled, 0644, enabled_show, enabled_store);
 348
 349 static ssize_t single_flag_show(struct kobject *kobj,
 350                                 struct kobj_attribute *attr, char *buf,
 351                                 enum transparent_hugepage_flag flag)
 352 {
 353         return sprintf(buf, "%d\n",
 354                        !!test_bit(flag, &transparent_hugepage_flags));
 355 }
 356
 357 static ssize_t single_flag_store(struct kobject *kobj,
 358                                  struct kobj_attribute *attr,
 359                                  const char *buf, size_t count,
 360                                  enum transparent_hugepage_flag flag)
 361 {
 362         unsigned long value;
 363         int ret;
 364
 365         ret = kstrtoul(buf, 10, &value);
 366         if (ret < 0)
 367                 return ret;
 368         if (value > 1)
 369                 return -EINVAL;
 370
 371         if (value)
 372                 set_bit(flag, &transparent_hugepage_flags);
 373         else
 374                 clear_bit(flag, &transparent_hugepage_flags);
 375
 376         return count;
 377 }
 378
 379 /*
 380  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 381  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 382  * memory just to allocate one more hugepage.
 383  */
 384 static ssize_t defrag_show(struct kobject *kobj,
 385                            struct kobj_attribute *attr, char *buf)
 386 {
 387         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 388                 return sprintf(buf, "[always] defer madvise never\n");
 389         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 390                 return sprintf(buf, "always [defer] madvise never\n");
 391         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 392                 return sprintf(buf, "always defer [madvise] never\n");
 393         else
 394                 return sprintf(buf, "always defer madvise [never]\n");
 395
 396 }
 397 static ssize_t defrag_store(struct kobject *kobj,
 398                             struct kobj_attribute *attr,
 399                             const char *buf, size_t count)
 400 {
 401         return triple_flag_store(kobj, attr, buf, count,
 402                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 403                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 404                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 405 }
 406 static struct kobj_attribute defrag_attr =
 407         __ATTR(defrag, 0644, defrag_show, defrag_store);
 408
 409 static ssize_t use_zero_page_show(struct kobject *kobj,
 410                 struct kobj_attribute *attr, char *buf)
 411 {
 412         return single_flag_show(kobj, attr, buf,
 413                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 414 }
 415 static ssize_t use_zero_page_store(struct kobject *kobj,
 416                 struct kobj_attribute *attr, const char *buf, size_t count)
 417 {
 418         return single_flag_store(kobj, attr, buf, count,
 419                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 420 }
 421 static struct kobj_attribute use_zero_page_attr =
 422         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 423 #ifdef CONFIG_DEBUG_VM
 424 static ssize_t debug_cow_show(struct kobject *kobj,
 425                                 struct kobj_attribute *attr, char *buf)
 426 {
 427         return single_flag_show(kobj, attr, buf,
 428                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 429 }
 430 static ssize_t debug_cow_store(struct kobject *kobj,
 431                                struct kobj_attribute *attr,
 432                                const char *buf, size_t count)
 433 {
 434         return single_flag_store(kobj, attr, buf, count,
 435                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 436 }
 437 static struct kobj_attribute debug_cow_attr =
 438         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 439 #endif /* CONFIG_DEBUG_VM */
 440
 441 static struct attribute *hugepage_attr[] = {
 442         &enabled_attr.attr,
 443         &defrag_attr.attr,
 444         &use_zero_page_attr.attr,
 445 #ifdef CONFIG_DEBUG_VM
 446         &debug_cow_attr.attr,
 447 #endif
 448         NULL,
 449 };
 450
 451 static struct attribute_group hugepage_attr_group = {
 452         .attrs = hugepage_attr,
 453 };
 454
 455 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 456                                          struct kobj_attribute *attr,
 457                                          char *buf)
 458 {
 459         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 460 }
 461
 462 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 463                                           struct kobj_attribute *attr,
 464                                           const char *buf, size_t count)
 465 {
 466         unsigned long msecs;
 467         int err;
 468
 469         err = kstrtoul(buf, 10, &msecs);
 470         if (err || msecs > UINT_MAX)
 471                 return -EINVAL;
 472
 473         khugepaged_scan_sleep_millisecs = msecs;
 474         khugepaged_sleep_expire = 0;
 475         wake_up_interruptible(&khugepaged_wait);
 476
 477         return count;
 478 }
 479 static struct kobj_attribute scan_sleep_millisecs_attr =
 480         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 481                scan_sleep_millisecs_store);
 482
 483 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 484                                           struct kobj_attribute *attr,
 485                                           char *buf)
 486 {
 487         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 488 }
 489
 490 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 491                                            struct kobj_attribute *attr,
 492                                            const char *buf, size_t count)
 493 {
 494         unsigned long msecs;
 495         int err;
 496
 497         err = kstrtoul(buf, 10, &msecs);
 498         if (err || msecs > UINT_MAX)
 499                 return -EINVAL;
 500
 501         khugepaged_alloc_sleep_millisecs = msecs;
 502         khugepaged_sleep_expire = 0;
 503         wake_up_interruptible(&khugepaged_wait);
 504
 505         return count;
 506 }
 507 static struct kobj_attribute alloc_sleep_millisecs_attr =
 508         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 509                alloc_sleep_millisecs_store);
 510
 511 static ssize_t pages_to_scan_show(struct kobject *kobj,
 512                                   struct kobj_attribute *attr,
 513                                   char *buf)
 514 {
 515         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 516 }
 517 static ssize_t pages_to_scan_store(struct kobject *kobj,
 518                                    struct kobj_attribute *attr,
 519                                    const char *buf, size_t count)
 520 {
 521         int err;
 522         unsigned long pages;
 523
 524         err = kstrtoul(buf, 10, &pages);
 525         if (err || !pages || pages > UINT_MAX)
 526                 return -EINVAL;
 527
 528         khugepaged_pages_to_scan = pages;
 529
 530         return count;
 531 }
 532 static struct kobj_attribute pages_to_scan_attr =
 533         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 534                pages_to_scan_store);
 535
 536 static ssize_t pages_collapsed_show(struct kobject *kobj,
 537                                     struct kobj_attribute *attr,
 538                                     char *buf)
 539 {
 540         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 541 }
 542 static struct kobj_attribute pages_collapsed_attr =
 543         __ATTR_RO(pages_collapsed);
 544
 545 static ssize_t full_scans_show(struct kobject *kobj,
 546                                struct kobj_attribute *attr,
 547                                char *buf)
 548 {
 549         return sprintf(buf, "%u\n", khugepaged_full_scans);
 550 }
 551 static struct kobj_attribute full_scans_attr =
 552         __ATTR_RO(full_scans);
 553
 554 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 555                                       struct kobj_attribute *attr, char *buf)
 556 {
 557         return single_flag_show(kobj, attr, buf,
 558                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 559 }
 560 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 561                                        struct kobj_attribute *attr,
 562                                        const char *buf, size_t count)
 563 {
 564         return single_flag_store(kobj, attr, buf, count,
 565                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 566 }
 567 static struct kobj_attribute khugepaged_defrag_attr =
 568         __ATTR(defrag, 0644, khugepaged_defrag_show,
 569                khugepaged_defrag_store);
 570
 571 /*
 572  * max_ptes_none controls if khugepaged should collapse hugepages over
 573  * any unmapped ptes in turn potentially increasing the memory
 574  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 575  * reduce the available free memory in the system as it
 576  * runs. Increasing max_ptes_none will instead potentially reduce the
 577  * free memory in the system during the khugepaged scan.
 578  */
 579 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 580                                              struct kobj_attribute *attr,
 581                                              char *buf)
 582 {
 583         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 584 }
 585 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 586                                               struct kobj_attribute *attr,
 587                                               const char *buf, size_t count)
 588 {
 589         int err;
 590         unsigned long max_ptes_none;
 591
 592         err = kstrtoul(buf, 10, &max_ptes_none);
 593         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 594                 return -EINVAL;
 595
 596         khugepaged_max_ptes_none = max_ptes_none;
 597
 598         return count;
 599 }
 600 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 601         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 602                khugepaged_max_ptes_none_store);
 603
 604 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 605                                              struct kobj_attribute *attr,
 606                                              char *buf)
 607 {
 608         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 609 }
 610
 611 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 612                                               struct kobj_attribute *attr,
 613                                               const char *buf, size_t count)
 614 {
 615         int err;
 616         unsigned long max_ptes_swap;
 617
 618         err  = kstrtoul(buf, 10, &max_ptes_swap);
 619         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 620                 return -EINVAL;
 621
 622         khugepaged_max_ptes_swap = max_ptes_swap;
 623
 624         return count;
 625 }
 626
 627 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 628         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 629                khugepaged_max_ptes_swap_store);
 630
 631 static struct attribute *khugepaged_attr[] = {
 632         &khugepaged_defrag_attr.attr,
 633         &khugepaged_max_ptes_none_attr.attr,
 634         &pages_to_scan_attr.attr,
 635         &pages_collapsed_attr.attr,
 636         &full_scans_attr.attr,
 637         &scan_sleep_millisecs_attr.attr,
 638         &alloc_sleep_millisecs_attr.attr,
 639         &khugepaged_max_ptes_swap_attr.attr,
 640         NULL,
 641 };
 642
 643 static struct attribute_group khugepaged_attr_group = {
 644         .attrs = khugepaged_attr,
 645         .name = "khugepaged",
 646 };
 647
 648 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 649 {
 650         int err;
 651
 652         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 653         if (unlikely(!*hugepage_kobj)) {
 654                 pr_err("failed to create transparent hugepage kobject\n");
 655                 return -ENOMEM;
 656         }
 657
 658         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 659         if (err) {
 660                 pr_err("failed to register transparent hugepage group\n");
 661                 goto delete_obj;
 662         }
 663
 664         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 665         if (err) {
 666                 pr_err("failed to register transparent hugepage group\n");
 667                 goto remove_hp_group;
 668         }
 669
 670         return 0;
 671
 672 remove_hp_group:
 673         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 674 delete_obj:
 675         kobject_put(*hugepage_kobj);
 676         return err;
 677 }
 678
 679 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 680 {
 681         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 682         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 683         kobject_put(hugepage_kobj);
 684 }
 685 #else
 686 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 687 {
 688         return 0;
 689 }
 690
 691 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 692 {
 693 }
 694 #endif /* CONFIG_SYSFS */
 695
 696 static int __init hugepage_init(void)
 697 {
 698         int err;
 699         struct kobject *hugepage_kobj;
 700
 701         if (!has_transparent_hugepage()) {
 702                 transparent_hugepage_flags = 0;
 703                 return -EINVAL;
 704         }
 705
 706         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 707         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 708         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 709         /*
 710          * hugepages can't be allocated by the buddy allocator
 711          */
 712         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 713         /*
 714          * we use page->mapping and page->index in second tail page
 715          * as list_head: assuming THP order >= 2
 716          */
 717         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 718
 719         err = hugepage_init_sysfs(&hugepage_kobj);
 720         if (err)
 721                 goto err_sysfs;
 722
 723         err = khugepaged_slab_init();
 724         if (err)
 725                 goto err_slab;
 726
 727         err = register_shrinker(&huge_zero_page_shrinker);
 728         if (err)
 729                 goto err_hzp_shrinker;
 730         err = register_shrinker(&deferred_split_shrinker);
 731         if (err)
 732                 goto err_split_shrinker;
 733
 734         /*
 735          * By default disable transparent hugepages on smaller systems,
 736          * where the extra memory used could hurt more than TLB overhead
 737          * is likely to save.  The admin can still enable it through /sys.
 738          */
 739         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 740                 transparent_hugepage_flags = 0;
 741                 return 0;
 742         }
 743
 744         err = start_stop_khugepaged();
 745         if (err)
 746                 goto err_khugepaged;
 747
 748         return 0;
 749 err_khugepaged:
 750         unregister_shrinker(&deferred_split_shrinker);
 751 err_split_shrinker:
 752         unregister_shrinker(&huge_zero_page_shrinker);
 753 err_hzp_shrinker:
 754         khugepaged_slab_exit();
 755 err_slab:
 756         hugepage_exit_sysfs(hugepage_kobj);
 757 err_sysfs:
 758         return err;
 759 }
 760 subsys_initcall(hugepage_init);
 761
 762 static int __init setup_transparent_hugepage(char *str)
 763 {
 764         int ret = 0;
 765         if (!str)
 766                 goto out;
 767         if (!strcmp(str, "always")) {
 768                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 769                         &transparent_hugepage_flags);
 770                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 771                           &transparent_hugepage_flags);
 772                 ret = 1;
 773         } else if (!strcmp(str, "madvise")) {
 774                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 775                           &transparent_hugepage_flags);
 776                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 777                         &transparent_hugepage_flags);
 778                 ret = 1;
 779         } else if (!strcmp(str, "never")) {
 780                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 781                           &transparent_hugepage_flags);
 782                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 783                           &transparent_hugepage_flags);
 784                 ret = 1;
 785         }
 786 out:
 787         if (!ret)
 788                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 789         return ret;
 790 }
 791 __setup("transparent_hugepage=", setup_transparent_hugepage);
 792
 793 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 794 {
 795         if (likely(vma->vm_flags & VM_WRITE))
 796                 pmd = pmd_mkwrite(pmd);
 797         return pmd;
 798 }
 799
 800 static inline struct list_head *page_deferred_list(struct page *page)
 801 {
 802         /*
 803          * ->lru in the tail pages is occupied by compound_head.
 804          * Let's use ->mapping + ->index in the second tail page as list_head.
 805          */
 806         return (struct list_head *)&page[2].mapping;
 807 }
 808
 809 void prep_transhuge_page(struct page *page)
 810 {
 811         /*
 812          * we use page->mapping and page->indexlru in second tail page
 813          * as list_head: assuming THP order >= 2
 814          */
 815
 816         INIT_LIST_HEAD(page_deferred_list(page));
 817         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 818 }
 819
 820 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 821                 gfp_t gfp)
 822 {
 823         struct vm_area_struct *vma = fe->vma;
 824         struct mem_cgroup *memcg;
 825         pgtable_t pgtable;
 826         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 827
 828         VM_BUG_ON_PAGE(!PageCompound(page), page);
 829
 830         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 831                 put_page(page);
 832                 count_vm_event(THP_FAULT_FALLBACK);
 833                 return VM_FAULT_FALLBACK;
 834         }
 835
 836         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 837         if (unlikely(!pgtable)) {
 838                 mem_cgroup_cancel_charge(page, memcg, true);
 839                 put_page(page);
 840                 return VM_FAULT_OOM;
 841         }
 842
 843         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 844         /*
 845          * The memory barrier inside __SetPageUptodate makes sure that
 846          * clear_huge_page writes become visible before the set_pmd_at()
 847          * write.
 848          */
 849         __SetPageUptodate(page);
 850
 851         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 852         if (unlikely(!pmd_none(*fe->pmd))) {
 853                 spin_unlock(fe->ptl);
 854                 mem_cgroup_cancel_charge(page, memcg, true);
 855                 put_page(page);
 856                 pte_free(vma->vm_mm, pgtable);
 857         } else {
 858                 pmd_t entry;
 859
 860                 /* Deliver the page fault to userland */
 861                 if (userfaultfd_missing(vma)) {
 862                         int ret;
 863
 864                         spin_unlock(fe->ptl);
 865                         mem_cgroup_cancel_charge(page, memcg, true);
 866                         put_page(page);
 867                         pte_free(vma->vm_mm, pgtable);
 868                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 869                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 870                         return ret;
 871                 }
 872
 873                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 874                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 875                 page_add_new_anon_rmap(page, vma, haddr, true);
 876                 mem_cgroup_commit_charge(page, memcg, false, true);
 877                 lru_cache_add_active_or_unevictable(page, vma);
 878                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 879                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 880                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 881                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 882                 spin_unlock(fe->ptl);
 883                 count_vm_event(THP_FAULT_ALLOC);
 884         }
 885
 886         return 0;
 887 }
 888
 889 /*
 890  * If THP is set to always then directly reclaim/compact as necessary
 891  * If set to defer then do no reclaim and defer to khugepaged
 892  * If set to madvise and the VMA is flagged then directly reclaim/compact
 893  */
 894 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 895 {
 896         gfp_t reclaim_flags = 0;
 897
 898         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) &&
 899             (vma->vm_flags & VM_HUGEPAGE))
 900                 reclaim_flags = __GFP_DIRECT_RECLAIM;
 901         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 902                 reclaim_flags = __GFP_KSWAPD_RECLAIM;
 903         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 904                 reclaim_flags = __GFP_DIRECT_RECLAIM;
 905
 906         return GFP_TRANSHUGE | reclaim_flags;
 907 }
 908
 909 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 910 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 911 {
 912         return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
 913 }
 914
 915 /* Caller must hold page table lock. */
 916 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 917                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 918                 struct page *zero_page)
 919 {
 920         pmd_t entry;
 921         if (!pmd_none(*pmd))
 922                 return false;
 923         entry = mk_pmd(zero_page, vma->vm_page_prot);
 924         entry = pmd_mkhuge(entry);
 925         if (pgtable)
 926                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 927         set_pmd_at(mm, haddr, pmd, entry);
 928         atomic_long_inc(&mm->nr_ptes);
 929         return true;
 930 }
 931
 932 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 933 {
 934         struct vm_area_struct *vma = fe->vma;
 935         gfp_t gfp;
 936         struct page *page;
 937         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 938
 939         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 940                 return VM_FAULT_FALLBACK;
 941         if (unlikely(anon_vma_prepare(vma)))
 942                 return VM_FAULT_OOM;
 943         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 944                 return VM_FAULT_OOM;
 945         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 946                         !mm_forbids_zeropage(vma->vm_mm) &&
 947                         transparent_hugepage_use_zero_page()) {
 948                 pgtable_t pgtable;
 949                 struct page *zero_page;
 950                 bool set;
 951                 int ret;
 952                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 953                 if (unlikely(!pgtable))
 954                         return VM_FAULT_OOM;
 955                 zero_page = get_huge_zero_page();
 956                 if (unlikely(!zero_page)) {
 957                         pte_free(vma->vm_mm, pgtable);
 958                         count_vm_event(THP_FAULT_FALLBACK);
 959                         return VM_FAULT_FALLBACK;
 960                 }
 961                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 962                 ret = 0;
 963                 set = false;
 964                 if (pmd_none(*fe->pmd)) {
 965                         if (userfaultfd_missing(vma)) {
 966                                 spin_unlock(fe->ptl);
 967                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 968                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 969                         } else {
 970                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 971                                                    haddr, fe->pmd, zero_page);
 972                                 spin_unlock(fe->ptl);
 973                                 set = true;
 974                         }
 975                 } else
 976                         spin_unlock(fe->ptl);
 977                 if (!set) {
 978                         pte_free(vma->vm_mm, pgtable);
 979                         put_huge_zero_page();
 980                 }
 981                 return ret;
 982         }
 983         gfp = alloc_hugepage_direct_gfpmask(vma);
 984         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 985         if (unlikely(!page)) {
 986                 count_vm_event(THP_FAULT_FALLBACK);
 987                 return VM_FAULT_FALLBACK;
 988         }
 989         prep_transhuge_page(page);
 990         return __do_huge_pmd_anonymous_page(fe, page, gfp);
 991 }
 992
 993 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 994                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 995 {
 996         struct mm_struct *mm = vma->vm_mm;
 997         pmd_t entry;
 998         spinlock_t *ptl;
 999
1000         ptl = pmd_lock(mm, pmd);
1001         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1002         if (pfn_t_devmap(pfn))
1003                 entry = pmd_mkdevmap(entry);
1004         if (write) {
1005                 entry = pmd_mkyoung(pmd_mkdirty(entry));
1006                 entry = maybe_pmd_mkwrite(entry, vma);
1007         }
1008         set_pmd_at(mm, addr, pmd, entry);
1009         update_mmu_cache_pmd(vma, addr, pmd);
1010         spin_unlock(ptl);
1011 }
1012
1013 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1014                         pmd_t *pmd, pfn_t pfn, bool write)
1015 {
1016         pgprot_t pgprot = vma->vm_page_prot;
1017         /*
1018          * If we had pmd_special, we could avoid all these restrictions,
1019          * but we need to be consistent with PTEs and architectures that
1020          * can't support a 'special' bit.
1021          */
1022         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1023         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1024                                                 (VM_PFNMAP|VM_MIXEDMAP));
1025         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1026         BUG_ON(!pfn_t_devmap(pfn));
1027
1028         if (addr < vma->vm_start || addr >= vma->vm_end)
1029                 return VM_FAULT_SIGBUS;
1030         if (track_pfn_insert(vma, &pgprot, pfn))
1031                 return VM_FAULT_SIGBUS;
1032         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
1033         return VM_FAULT_NOPAGE;
1034 }
1035 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1036
1037 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1038                 pmd_t *pmd)
1039 {
1040         pmd_t _pmd;
1041
1042         /*
1043          * We should set the dirty bit only for FOLL_WRITE but for now
1044          * the dirty bit in the pmd is meaningless.  And if the dirty
1045          * bit will become meaningful and we'll only set it with
1046          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1047          * set the young bit, instead of the current set_pmd_at.
1048          */
1049         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1050         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1051                                 pmd, _pmd,  1))
1052                 update_mmu_cache_pmd(vma, addr, pmd);
1053 }
1054
1055 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1056                 pmd_t *pmd, int flags)
1057 {
1058         unsigned long pfn = pmd_pfn(*pmd);
1059         struct mm_struct *mm = vma->vm_mm;
1060         struct dev_pagemap *pgmap;
1061         struct page *page;
1062
1063         assert_spin_locked(pmd_lockptr(mm, pmd));
1064
1065         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1066                 return NULL;
1067
1068         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1069                 /* pass */;
1070         else
1071                 return NULL;
1072
1073         if (flags & FOLL_TOUCH)
1074                 touch_pmd(vma, addr, pmd);
1075
1076         /*
1077          * device mapped pages can only be returned if the
1078          * caller will manage the page reference count.
1079          */
1080         if (!(flags & FOLL_GET))
1081                 return ERR_PTR(-EEXIST);
1082
1083         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1084         pgmap = get_dev_pagemap(pfn, NULL);
1085         if (!pgmap)
1086                 return ERR_PTR(-EFAULT);
1087         page = pfn_to_page(pfn);
1088         get_page(page);
1089         put_dev_pagemap(pgmap);
1090
1091         return page;
1092 }
1093
1094 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1095                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1096                   struct vm_area_struct *vma)
1097 {
1098         spinlock_t *dst_ptl, *src_ptl;
1099         struct page *src_page;
1100         pmd_t pmd;
1101         pgtable_t pgtable = NULL;
1102         int ret = -ENOMEM;
1103
1104         /* Skip if can be re-fill on fault */
1105         if (!vma_is_anonymous(vma))
1106                 return 0;
1107
1108         pgtable = pte_alloc_one(dst_mm, addr);
1109         if (unlikely(!pgtable))
1110                 goto out;
1111
1112         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1113         src_ptl = pmd_lockptr(src_mm, src_pmd);
1114         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1115
1116         ret = -EAGAIN;
1117         pmd = *src_pmd;
1118         if (unlikely(!pmd_trans_huge(pmd))) {
1119                 pte_free(dst_mm, pgtable);
1120                 goto out_unlock;
1121         }
1122         /*
1123          * When page table lock is held, the huge zero pmd should not be
1124          * under splitting since we don't split the page itself, only pmd to
1125          * a page table.
1126          */
1127         if (is_huge_zero_pmd(pmd)) {
1128                 struct page *zero_page;
1129                 /*
1130                  * get_huge_zero_page() will never allocate a new page here,
1131                  * since we already have a zero page to copy. It just takes a
1132                  * reference.
1133                  */
1134                 zero_page = get_huge_zero_page();
1135                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1136                                 zero_page);
1137                 ret = 0;
1138                 goto out_unlock;
1139         }
1140
1141         src_page = pmd_page(pmd);
1142         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1143         get_page(src_page);
1144         page_dup_rmap(src_page, true);
1145         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1146         atomic_long_inc(&dst_mm->nr_ptes);
1147         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1148
1149         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1150         pmd = pmd_mkold(pmd_wrprotect(pmd));
1151         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1152
1153         ret = 0;
1154 out_unlock:
1155         spin_unlock(src_ptl);
1156         spin_unlock(dst_ptl);
1157 out:
1158         return ret;
1159 }
1160
1161 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
1162 {
1163         pmd_t entry;
1164         unsigned long haddr;
1165
1166         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
1167         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1168                 goto unlock;
1169
1170         entry = pmd_mkyoung(orig_pmd);
1171         haddr = fe->address & HPAGE_PMD_MASK;
1172         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
1173                                 fe->flags & FAULT_FLAG_WRITE))
1174                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
1175
1176 unlock:
1177         spin_unlock(fe->ptl);
1178 }
1179
1180 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
1181                 struct page *page)
1182 {
1183         struct vm_area_struct *vma = fe->vma;
1184         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1185         struct mem_cgroup *memcg;
1186         pgtable_t pgtable;
1187         pmd_t _pmd;
1188         int ret = 0, i;
1189         struct page **pages;
1190         unsigned long mmun_start;       /* For mmu_notifiers */
1191         unsigned long mmun_end;         /* For mmu_notifiers */
1192
1193         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1194                         GFP_KERNEL);
1195         if (unlikely(!pages)) {
1196                 ret |= VM_FAULT_OOM;
1197                 goto out;
1198         }
1199
1200         for (i = 0; i < HPAGE_PMD_NR; i++) {
1201                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1202                                                __GFP_OTHER_NODE, vma,
1203                                                fe->address, page_to_nid(page));
1204                 if (unlikely(!pages[i] ||
1205                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1206                                      GFP_KERNEL, &memcg, false))) {
1207                         if (pages[i])
1208                                 put_page(pages[i]);
1209                         while (--i >= 0) {
1210                                 memcg = (void *)page_private(pages[i]);
1211                                 set_page_private(pages[i], 0);
1212                                 mem_cgroup_cancel_charge(pages[i], memcg,
1213                                                 false);
1214                                 put_page(pages[i]);
1215                         }
1216                         kfree(pages);
1217                         ret |= VM_FAULT_OOM;
1218                         goto out;
1219                 }
1220                 set_page_private(pages[i], (unsigned long)memcg);
1221         }
1222
1223         for (i = 0; i < HPAGE_PMD_NR; i++) {
1224                 copy_user_highpage(pages[i], page + i,
1225                                    haddr + PAGE_SIZE * i, vma);
1226                 __SetPageUptodate(pages[i]);
1227                 cond_resched();
1228         }
1229
1230         mmun_start = haddr;
1231         mmun_end   = haddr + HPAGE_PMD_SIZE;
1232         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1233
1234         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1235         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1236                 goto out_free_pages;
1237         VM_BUG_ON_PAGE(!PageHead(page), page);
1238
1239         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1240         /* leave pmd empty until pte is filled */
1241
1242         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
1243         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1244
1245         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1246                 pte_t entry;
1247                 entry = mk_pte(pages[i], vma->vm_page_prot);
1248                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1249                 memcg = (void *)page_private(pages[i]);
1250                 set_page_private(pages[i], 0);
1251                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
1252                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1253                 lru_cache_add_active_or_unevictable(pages[i], vma);
1254                 fe->pte = pte_offset_map(&_pmd, haddr);
1255                 VM_BUG_ON(!pte_none(*fe->pte));
1256                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
1257                 pte_unmap(fe->pte);
1258         }
1259         kfree(pages);
1260
1261         smp_wmb(); /* make pte visible before pmd */
1262         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
1263         page_remove_rmap(page, true);
1264         spin_unlock(fe->ptl);
1265
1266         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1267
1268         ret |= VM_FAULT_WRITE;
1269         put_page(page);
1270
1271 out:
1272         return ret;
1273
1274 out_free_pages:
1275         spin_unlock(fe->ptl);
1276         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1277         for (i = 0; i < HPAGE_PMD_NR; i++) {
1278                 memcg = (void *)page_private(pages[i]);
1279                 set_page_private(pages[i], 0);
1280                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1281                 put_page(pages[i]);
1282         }
1283         kfree(pages);
1284         goto out;
1285 }
1286
1287 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1288 {
1289         struct vm_area_struct *vma = fe->vma;
1290         struct page *page = NULL, *new_page;
1291         struct mem_cgroup *memcg;
1292         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1293         unsigned long mmun_start;       /* For mmu_notifiers */
1294         unsigned long mmun_end;         /* For mmu_notifiers */
1295         gfp_t huge_gfp;                 /* for allocation and charge */
1296         int ret = 0;
1297
1298         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1299         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1300         if (is_huge_zero_pmd(orig_pmd))
1301                 goto alloc;
1302         spin_lock(fe->ptl);
1303         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1304                 goto out_unlock;
1305
1306         page = pmd_page(orig_pmd);
1307         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1308         /*
1309          * We can only reuse the page if nobody else maps the huge page or it's
1310          * part.
1311          */
1312         if (page_trans_huge_mapcount(page, NULL) == 1) {
1313                 pmd_t entry;
1314                 entry = pmd_mkyoung(orig_pmd);
1315                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1316                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1317                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1318                 ret |= VM_FAULT_WRITE;
1319                 goto out_unlock;
1320         }
1321         get_page(page);
1322         spin_unlock(fe->ptl);
1323 alloc:
1324         if (transparent_hugepage_enabled(vma) &&
1325             !transparent_hugepage_debug_cow()) {
1326                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1327                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1328         } else
1329                 new_page = NULL;
1330
1331         if (likely(new_page)) {
1332                 prep_transhuge_page(new_page);
1333         } else {
1334                 if (!page) {
1335                         split_huge_pmd(vma, fe->pmd, fe->address);
1336                         ret |= VM_FAULT_FALLBACK;
1337                 } else {
1338                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1339                         if (ret & VM_FAULT_OOM) {
1340                                 split_huge_pmd(vma, fe->pmd, fe->address);
1341                                 ret |= VM_FAULT_FALLBACK;
1342                         }
1343                         put_page(page);
1344                 }
1345                 count_vm_event(THP_FAULT_FALLBACK);
1346                 goto out;
1347         }
1348
1349         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1350                                         huge_gfp, &memcg, true))) {
1351                 put_page(new_page);
1352                 split_huge_pmd(vma, fe->pmd, fe->address);
1353                 if (page)
1354                         put_page(page);
1355                 ret |= VM_FAULT_FALLBACK;
1356                 count_vm_event(THP_FAULT_FALLBACK);
1357                 goto out;
1358         }
1359
1360         count_vm_event(THP_FAULT_ALLOC);
1361
1362         if (!page)
1363                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1364         else
1365                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1366         __SetPageUptodate(new_page);
1367
1368         mmun_start = haddr;
1369         mmun_end   = haddr + HPAGE_PMD_SIZE;
1370         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1371
1372         spin_lock(fe->ptl);
1373         if (page)
1374                 put_page(page);
1375         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1376                 spin_unlock(fe->ptl);
1377                 mem_cgroup_cancel_charge(new_page, memcg, true);
1378                 put_page(new_page);
1379                 goto out_mn;
1380         } else {
1381                 pmd_t entry;
1382                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1383                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1384                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1385                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1386                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1387                 lru_cache_add_active_or_unevictable(new_page, vma);
1388                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1389                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1390                 if (!page) {
1391                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1392                         put_huge_zero_page();
1393                 } else {
1394                         VM_BUG_ON_PAGE(!PageHead(page), page);
1395                         page_remove_rmap(page, true);
1396                         put_page(page);
1397                 }
1398                 ret |= VM_FAULT_WRITE;
1399         }
1400         spin_unlock(fe->ptl);
1401 out_mn:
1402         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1403 out:
1404         return ret;
1405 out_unlock:
1406         spin_unlock(fe->ptl);
1407         return ret;
1408 }
1409
1410 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1411                                    unsigned long addr,
1412                                    pmd_t *pmd,
1413                                    unsigned int flags)
1414 {
1415         struct mm_struct *mm = vma->vm_mm;
1416         struct page *page = NULL;
1417
1418         assert_spin_locked(pmd_lockptr(mm, pmd));
1419
1420         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1421                 goto out;
1422
1423         /* Avoid dumping huge zero page */
1424         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1425                 return ERR_PTR(-EFAULT);
1426
1427         /* Full NUMA hinting faults to serialise migration in fault paths */
1428         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1429                 goto out;
1430
1431         page = pmd_page(*pmd);
1432         VM_BUG_ON_PAGE(!PageHead(page), page);
1433         if (flags & FOLL_TOUCH)
1434                 touch_pmd(vma, addr, pmd);
1435         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1436                 /*
1437                  * We don't mlock() pte-mapped THPs. This way we can avoid
1438                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1439                  *
1440                  * In most cases the pmd is the only mapping of the page as we
1441                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1442                  * writable private mappings in populate_vma_page_range().
1443                  *
1444                  * The only scenario when we have the page shared here is if we
1445                  * mlocking read-only mapping shared over fork(). We skip
1446                  * mlocking such pages.
1447                  */
1448                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1449                                 page->mapping && trylock_page(page)) {
1450                         lru_add_drain();
1451                         if (page->mapping)
1452                                 mlock_vma_page(page);
1453                         unlock_page(page);
1454                 }
1455         }
1456         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1457         VM_BUG_ON_PAGE(!PageCompound(page), page);
1458         if (flags & FOLL_GET)
1459                 get_page(page);
1460
1461 out:
1462         return page;
1463 }
1464
1465 /* NUMA hinting page fault entry point for trans huge pmds */
1466 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1467 {
1468         struct vm_area_struct *vma = fe->vma;
1469         struct anon_vma *anon_vma = NULL;
1470         struct page *page;
1471         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1472         int page_nid = -1, this_nid = numa_node_id();
1473         int target_nid, last_cpupid = -1;
1474         bool page_locked;
1475         bool migrated = false;
1476         bool was_writable;
1477         int flags = 0;
1478
1479         /* A PROT_NONE fault should not end up here */
1480         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1481
1482         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1483         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1484                 goto out_unlock;
1485
1486         /*
1487          * If there are potential migrations, wait for completion and retry
1488          * without disrupting NUMA hinting information. Do not relock and
1489          * check_same as the page may no longer be mapped.
1490          */
1491         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1492                 page = pmd_page(*fe->pmd);
1493                 spin_unlock(fe->ptl);
1494                 wait_on_page_locked(page);
1495                 goto out;
1496         }
1497
1498         page = pmd_page(pmd);
1499         BUG_ON(is_huge_zero_page(page));
1500         page_nid = page_to_nid(page);
1501         last_cpupid = page_cpupid_last(page);
1502         count_vm_numa_event(NUMA_HINT_FAULTS);
1503         if (page_nid == this_nid) {
1504                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1505                 flags |= TNF_FAULT_LOCAL;
1506         }
1507
1508         /* See similar comment in do_numa_page for explanation */
1509         if (!(vma->vm_flags & VM_WRITE))
1510                 flags |= TNF_NO_GROUP;
1511
1512         /*
1513          * Acquire the page lock to serialise THP migrations but avoid dropping
1514          * page_table_lock if at all possible
1515          */
1516         page_locked = trylock_page(page);
1517         target_nid = mpol_misplaced(page, vma, haddr);
1518         if (target_nid == -1) {
1519                 /* If the page was locked, there are no parallel migrations */
1520                 if (page_locked)
1521                         goto clear_pmdnuma;
1522         }
1523
1524         /* Migration could have started since the pmd_trans_migrating check */
1525         if (!page_locked) {
1526                 spin_unlock(fe->ptl);
1527                 wait_on_page_locked(page);
1528                 page_nid = -1;
1529                 goto out;
1530         }
1531
1532         /*
1533          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1534          * to serialises splits
1535          */
1536         get_page(page);
1537         spin_unlock(fe->ptl);
1538         anon_vma = page_lock_anon_vma_read(page);
1539
1540         /* Confirm the PMD did not change while page_table_lock was released */
1541         spin_lock(fe->ptl);
1542         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1543                 unlock_page(page);
1544                 put_page(page);
1545                 page_nid = -1;
1546                 goto out_unlock;
1547         }
1548
1549         /* Bail if we fail to protect against THP splits for any reason */
1550         if (unlikely(!anon_vma)) {
1551                 put_page(page);
1552                 page_nid = -1;
1553                 goto clear_pmdnuma;
1554         }
1555
1556         /*
1557          * Migrate the THP to the requested node, returns with page unlocked
1558          * and access rights restored.
1559          */
1560         spin_unlock(fe->ptl);
1561         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1562                                 fe->pmd, pmd, fe->address, page, target_nid);
1563         if (migrated) {
1564                 flags |= TNF_MIGRATED;
1565                 page_nid = target_nid;
1566         } else
1567                 flags |= TNF_MIGRATE_FAIL;
1568
1569         goto out;
1570 clear_pmdnuma:
1571         BUG_ON(!PageLocked(page));
1572         was_writable = pmd_write(pmd);
1573         pmd = pmd_modify(pmd, vma->vm_page_prot);
1574         pmd = pmd_mkyoung(pmd);
1575         if (was_writable)
1576                 pmd = pmd_mkwrite(pmd);
1577         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1578         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1579         unlock_page(page);
1580 out_unlock:
1581         spin_unlock(fe->ptl);
1582
1583 out:
1584         if (anon_vma)
1585                 page_unlock_anon_vma_read(anon_vma);
1586
1587         if (page_nid != -1)
1588                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1589
1590         return 0;
1591 }
1592
1593 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1594                 pmd_t *pmd, unsigned long addr, unsigned long next)
1595
1596 {
1597         spinlock_t *ptl;
1598         pmd_t orig_pmd;
1599         struct page *page;
1600         struct mm_struct *mm = tlb->mm;
1601         int ret = 0;
1602
1603         ptl = pmd_trans_huge_lock(pmd, vma);
1604         if (!ptl)
1605                 goto out_unlocked;
1606
1607         orig_pmd = *pmd;
1608         if (is_huge_zero_pmd(orig_pmd)) {
1609                 ret = 1;
1610                 goto out;
1611         }
1612
1613         page = pmd_page(orig_pmd);
1614         /*
1615          * If other processes are mapping this page, we couldn't discard
1616          * the page unless they all do MADV_FREE so let's skip the page.
1617          */
1618         if (page_mapcount(page) != 1)
1619                 goto out;
1620
1621         if (!trylock_page(page))
1622                 goto out;
1623
1624         /*
1625          * If user want to discard part-pages of THP, split it so MADV_FREE
1626          * will deactivate only them.
1627          */
1628         if (next - addr != HPAGE_PMD_SIZE) {
1629                 get_page(page);
1630                 spin_unlock(ptl);
1631                 split_huge_page(page);
1632                 put_page(page);
1633                 unlock_page(page);
1634                 goto out_unlocked;
1635         }
1636
1637         if (PageDirty(page))
1638                 ClearPageDirty(page);
1639         unlock_page(page);
1640
1641         if (PageActive(page))
1642                 deactivate_page(page);
1643
1644         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1645                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1646                         tlb->fullmm);
1647                 orig_pmd = pmd_mkold(orig_pmd);
1648                 orig_pmd = pmd_mkclean(orig_pmd);
1649
1650                 set_pmd_at(mm, addr, pmd, orig_pmd);
1651                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1652         }
1653         ret = 1;
1654 out:
1655         spin_unlock(ptl);
1656 out_unlocked:
1657         return ret;
1658 }
1659
1660 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1661                  pmd_t *pmd, unsigned long addr)
1662 {
1663         pmd_t orig_pmd;
1664         spinlock_t *ptl;
1665
1666         ptl = __pmd_trans_huge_lock(pmd, vma);
1667         if (!ptl)
1668                 return 0;
1669         /*
1670          * For architectures like ppc64 we look at deposited pgtable
1671          * when calling pmdp_huge_get_and_clear. So do the
1672          * pgtable_trans_huge_withdraw after finishing pmdp related
1673          * operations.
1674          */
1675         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1676                         tlb->fullmm);
1677         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1678         if (vma_is_dax(vma)) {
1679                 spin_unlock(ptl);
1680                 if (is_huge_zero_pmd(orig_pmd))
1681                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1682         } else if (is_huge_zero_pmd(orig_pmd)) {
1683                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1684                 atomic_long_dec(&tlb->mm->nr_ptes);
1685                 spin_unlock(ptl);
1686                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1687         } else {
1688                 struct page *page = pmd_page(orig_pmd);
1689                 page_remove_rmap(page, true);
1690                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1691                 VM_BUG_ON_PAGE(!PageHead(page), page);
1692                 if (PageAnon(page)) {
1693                         pgtable_t pgtable;
1694                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1695                         pte_free(tlb->mm, pgtable);
1696                         atomic_long_dec(&tlb->mm->nr_ptes);
1697                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1698                 } else {
1699                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1700                 }
1701                 spin_unlock(ptl);
1702                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1703         }
1704         return 1;
1705 }
1706
1707 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1708                   unsigned long new_addr, unsigned long old_end,
1709                   pmd_t *old_pmd, pmd_t *new_pmd)
1710 {
1711         spinlock_t *old_ptl, *new_ptl;
1712         pmd_t pmd;
1713         struct mm_struct *mm = vma->vm_mm;
1714
1715         if ((old_addr & ~HPAGE_PMD_MASK) ||
1716             (new_addr & ~HPAGE_PMD_MASK) ||
1717             old_end - old_addr < HPAGE_PMD_SIZE)
1718                 return false;
1719
1720         /*
1721          * The destination pmd shouldn't be established, free_pgtables()
1722          * should have release it.
1723          */
1724         if (WARN_ON(!pmd_none(*new_pmd))) {
1725                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1726                 return false;
1727         }
1728
1729         /*
1730          * We don't have to worry about the ordering of src and dst
1731          * ptlocks because exclusive mmap_sem prevents deadlock.
1732          */
1733         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1734         if (old_ptl) {
1735                 new_ptl = pmd_lockptr(mm, new_pmd);
1736                 if (new_ptl != old_ptl)
1737                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1738                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1739                 VM_BUG_ON(!pmd_none(*new_pmd));
1740
1741                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1742                                 vma_is_anonymous(vma)) {
1743                         pgtable_t pgtable;
1744                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1745                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1746                 }
1747                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1748                 if (new_ptl != old_ptl)
1749                         spin_unlock(new_ptl);
1750                 spin_unlock(old_ptl);
1751                 return true;
1752         }
1753         return false;
1754 }
1755
1756 /*
1757  * Returns
1758  *  - 0 if PMD could not be locked
1759  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1760  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1761  */
1762 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1763                 unsigned long addr, pgprot_t newprot, int prot_numa)
1764 {
1765         struct mm_struct *mm = vma->vm_mm;
1766         spinlock_t *ptl;
1767         int ret = 0;
1768
1769         ptl = __pmd_trans_huge_lock(pmd, vma);
1770         if (ptl) {
1771                 pmd_t entry;
1772                 bool preserve_write = prot_numa && pmd_write(*pmd);
1773                 ret = 1;
1774
1775                 /*
1776                  * Avoid trapping faults against the zero page. The read-only
1777                  * data is likely to be read-cached on the local CPU and
1778                  * local/remote hits to the zero page are not interesting.
1779                  */
1780                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1781                         spin_unlock(ptl);
1782                         return ret;
1783                 }
1784
1785                 if (!prot_numa || !pmd_protnone(*pmd)) {
1786                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1787                         entry = pmd_modify(entry, newprot);
1788                         if (preserve_write)
1789                                 entry = pmd_mkwrite(entry);
1790                         ret = HPAGE_PMD_NR;
1791                         set_pmd_at(mm, addr, pmd, entry);
1792                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1793                                         pmd_write(entry));
1794                 }
1795                 spin_unlock(ptl);
1796         }
1797
1798         return ret;
1799 }
1800
1801 /*
1802  * Returns true if a given pmd maps a thp, false otherwise.
1803  *
1804  * Note that if it returns true, this routine returns without unlocking page
1805  * table lock. So callers must unlock it.
1806  */
1807 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1808 {
1809         spinlock_t *ptl;
1810         ptl = pmd_lock(vma->vm_mm, pmd);
1811         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1812                 return ptl;
1813         spin_unlock(ptl);
1814         return NULL;
1815 }
1816
1817 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1818
1819 int hugepage_madvise(struct vm_area_struct *vma,
1820                      unsigned long *vm_flags, int advice)
1821 {
1822         switch (advice) {
1823         case MADV_HUGEPAGE:
1824 #ifdef CONFIG_S390
1825                 /*
1826                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1827                  * can't handle this properly after s390_enable_sie, so we simply
1828                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1829                  */
1830                 if (mm_has_pgste(vma->vm_mm))
1831                         return 0;
1832 #endif
1833                 /*
1834                  * Be somewhat over-protective like KSM for now!
1835                  */
1836                 if (*vm_flags & VM_NO_THP)
1837                         return -EINVAL;
1838                 *vm_flags &= ~VM_NOHUGEPAGE;
1839                 *vm_flags |= VM_HUGEPAGE;
1840                 /*
1841                  * If the vma become good for khugepaged to scan,
1842                  * register it here without waiting a page fault that
1843                  * may not happen any time soon.
1844                  */
1845                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1846                         return -ENOMEM;
1847                 break;
1848         case MADV_NOHUGEPAGE:
1849                 /*
1850                  * Be somewhat over-protective like KSM for now!
1851                  */
1852                 if (*vm_flags & VM_NO_THP)
1853                         return -EINVAL;
1854                 *vm_flags &= ~VM_HUGEPAGE;
1855                 *vm_flags |= VM_NOHUGEPAGE;
1856                 /*
1857                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1858                  * this vma even if we leave the mm registered in khugepaged if
1859                  * it got registered before VM_NOHUGEPAGE was set.
1860                  */
1861                 break;
1862         }
1863
1864         return 0;
1865 }
1866
1867 static int __init khugepaged_slab_init(void)
1868 {
1869         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1870                                           sizeof(struct mm_slot),
1871                                           __alignof__(struct mm_slot), 0, NULL);
1872         if (!mm_slot_cache)
1873                 return -ENOMEM;
1874
1875         return 0;
1876 }
1877
1878 static void __init khugepaged_slab_exit(void)
1879 {
1880         kmem_cache_destroy(mm_slot_cache);
1881 }
1882
1883 static inline struct mm_slot *alloc_mm_slot(void)
1884 {
1885         if (!mm_slot_cache)     /* initialization failed */
1886                 return NULL;
1887         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1888 }
1889
1890 static inline void free_mm_slot(struct mm_slot *mm_slot)
1891 {
1892         kmem_cache_free(mm_slot_cache, mm_slot);
1893 }
1894
1895 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1896 {
1897         struct mm_slot *mm_slot;
1898
1899         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1900                 if (mm == mm_slot->mm)
1901                         return mm_slot;
1902
1903         return NULL;
1904 }
1905
1906 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1907                                     struct mm_slot *mm_slot)
1908 {
1909         mm_slot->mm = mm;
1910         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1911 }
1912
1913 static inline int khugepaged_test_exit(struct mm_struct *mm)
1914 {
1915         return atomic_read(&mm->mm_users) == 0;
1916 }
1917
1918 int __khugepaged_enter(struct mm_struct *mm)
1919 {
1920         struct mm_slot *mm_slot;
1921         int wakeup;
1922
1923         mm_slot = alloc_mm_slot();
1924         if (!mm_slot)
1925                 return -ENOMEM;
1926
1927         /* __khugepaged_exit() must not run from under us */
1928         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1929         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1930                 free_mm_slot(mm_slot);
1931                 return 0;
1932         }
1933
1934         spin_lock(&khugepaged_mm_lock);
1935         insert_to_mm_slots_hash(mm, mm_slot);
1936         /*
1937          * Insert just behind the scanning cursor, to let the area settle
1938          * down a little.
1939          */
1940         wakeup = list_empty(&khugepaged_scan.mm_head);
1941         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1942         spin_unlock(&khugepaged_mm_lock);
1943
1944         atomic_inc(&mm->mm_count);
1945         if (wakeup)
1946                 wake_up_interruptible(&khugepaged_wait);
1947
1948         return 0;
1949 }
1950
1951 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1952                                unsigned long vm_flags)
1953 {
1954         unsigned long hstart, hend;
1955         if (!vma->anon_vma)
1956                 /*
1957                  * Not yet faulted in so we will register later in the
1958                  * page fault if needed.
1959                  */
1960                 return 0;
1961         if (vma->vm_ops || (vm_flags & VM_NO_THP))
1962                 /* khugepaged not yet working on file or special mappings */
1963                 return 0;
1964         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1965         hend = vma->vm_end & HPAGE_PMD_MASK;
1966         if (hstart < hend)
1967                 return khugepaged_enter(vma, vm_flags);
1968         return 0;
1969 }
1970
1971 void __khugepaged_exit(struct mm_struct *mm)
1972 {
1973         struct mm_slot *mm_slot;
1974         int free = 0;
1975
1976         spin_lock(&khugepaged_mm_lock);
1977         mm_slot = get_mm_slot(mm);
1978         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1979                 hash_del(&mm_slot->hash);
1980                 list_del(&mm_slot->mm_node);
1981                 free = 1;
1982         }
1983         spin_unlock(&khugepaged_mm_lock);
1984
1985         if (free) {
1986                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1987                 free_mm_slot(mm_slot);
1988                 mmdrop(mm);
1989         } else if (mm_slot) {
1990                 /*
1991                  * This is required to serialize against
1992                  * khugepaged_test_exit() (which is guaranteed to run
1993                  * under mmap sem read mode). Stop here (after we
1994                  * return all pagetables will be destroyed) until
1995                  * khugepaged has finished working on the pagetables
1996                  * under the mmap_sem.
1997                  */
1998                 down_write(&mm->mmap_sem);
1999                 up_write(&mm->mmap_sem);
2000         }
2001 }
2002
2003 static void release_pte_page(struct page *page)
2004 {
2005         /* 0 stands for page_is_file_cache(page) == false */
2006         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2007         unlock_page(page);
2008         putback_lru_page(page);
2009 }
2010
2011 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2012 {
2013         while (--_pte >= pte) {
2014                 pte_t pteval = *_pte;
2015                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2016                         release_pte_page(pte_page(pteval));
2017         }
2018 }
2019
2020 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2021                                         unsigned long address,
2022                                         pte_t *pte)
2023 {
2024         struct page *page = NULL;
2025         pte_t *_pte;
2026         int none_or_zero = 0, result = 0;
2027         bool referenced = false, writable = false;
2028
2029         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2030              _pte++, address += PAGE_SIZE) {
2031                 pte_t pteval = *_pte;
2032                 if (pte_none(pteval) || (pte_present(pteval) &&
2033                                 is_zero_pfn(pte_pfn(pteval)))) {
2034                         if (!userfaultfd_armed(vma) &&
2035                             ++none_or_zero <= khugepaged_max_ptes_none) {
2036                                 continue;
2037                         } else {
2038                                 result = SCAN_EXCEED_NONE_PTE;
2039                                 goto out;
2040                         }
2041                 }
2042                 if (!pte_present(pteval)) {
2043                         result = SCAN_PTE_NON_PRESENT;
2044                         goto out;
2045                 }
2046                 page = vm_normal_page(vma, address, pteval);
2047                 if (unlikely(!page)) {
2048                         result = SCAN_PAGE_NULL;
2049                         goto out;
2050                 }
2051
2052                 VM_BUG_ON_PAGE(PageCompound(page), page);
2053                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2054                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2055
2056                 /*
2057                  * We can do it before isolate_lru_page because the
2058                  * page can't be freed from under us. NOTE: PG_lock
2059                  * is needed to serialize against split_huge_page
2060                  * when invoked from the VM.
2061                  */
2062                 if (!trylock_page(page)) {
2063                         result = SCAN_PAGE_LOCK;
2064                         goto out;
2065                 }
2066
2067                 /*
2068                  * cannot use mapcount: can't collapse if there's a gup pin.
2069                  * The page must only be referenced by the scanned process
2070                  * and page swap cache.
2071                  */
2072                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2073                         unlock_page(page);
2074                         result = SCAN_PAGE_COUNT;
2075                         goto out;
2076                 }
2077                 if (pte_write(pteval)) {
2078                         writable = true;
2079                 } else {
2080                         if (PageSwapCache(page) &&
2081                             !reuse_swap_page(page, NULL)) {
2082                                 unlock_page(page);
2083                                 result = SCAN_SWAP_CACHE_PAGE;
2084                                 goto out;
2085                         }
2086                         /*
2087                          * Page is not in the swap cache. It can be collapsed
2088                          * into a THP.
2089                          */
2090                 }
2091
2092                 /*
2093                  * Isolate the page to avoid collapsing an hugepage
2094                  * currently in use by the VM.
2095                  */
2096                 if (isolate_lru_page(page)) {
2097                         unlock_page(page);
2098                         result = SCAN_DEL_PAGE_LRU;
2099                         goto out;
2100                 }
2101                 /* 0 stands for page_is_file_cache(page) == false */
2102                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2103                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2104                 VM_BUG_ON_PAGE(PageLRU(page), page);
2105
2106                 /* If there is no mapped pte young don't collapse the page */
2107                 if (pte_young(pteval) ||
2108                     page_is_young(page) || PageReferenced(page) ||
2109                     mmu_notifier_test_young(vma->vm_mm, address))
2110                         referenced = true;
2111         }
2112         if (likely(writable)) {
2113                 if (likely(referenced)) {
2114                         result = SCAN_SUCCEED;
2115                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2116                                                             referenced, writable, result);
2117                         return 1;
2118                 }
2119         } else {
2120                 result = SCAN_PAGE_RO;
2121         }
2122
2123 out:
2124         release_pte_pages(pte, _pte);
2125         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2126                                             referenced, writable, result);
2127         return 0;
2128 }
2129
2130 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2131                                       struct vm_area_struct *vma,
2132                                       unsigned long address,
2133                                       spinlock_t *ptl)
2134 {
2135         pte_t *_pte;
2136         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2137                 pte_t pteval = *_pte;
2138                 struct page *src_page;
2139
2140                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2141                         clear_user_highpage(page, address);
2142                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2143                         if (is_zero_pfn(pte_pfn(pteval))) {
2144                                 /*
2145                                  * ptl mostly unnecessary.
2146                                  */
2147                                 spin_lock(ptl);
2148                                 /*
2149                                  * paravirt calls inside pte_clear here are
2150                                  * superfluous.
2151                                  */
2152                                 pte_clear(vma->vm_mm, address, _pte);
2153                                 spin_unlock(ptl);
2154                         }
2155                 } else {
2156                         src_page = pte_page(pteval);
2157                         copy_user_highpage(page, src_page, address, vma);
2158                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2159                         release_pte_page(src_page);
2160                         /*
2161                          * ptl mostly unnecessary, but preempt has to
2162                          * be disabled to update the per-cpu stats
2163                          * inside page_remove_rmap().
2164                          */
2165                         spin_lock(ptl);
2166                         /*
2167                          * paravirt calls inside pte_clear here are
2168                          * superfluous.
2169                          */
2170                         pte_clear(vma->vm_mm, address, _pte);
2171                         page_remove_rmap(src_page, false);
2172                         spin_unlock(ptl);
2173                         free_page_and_swap_cache(src_page);
2174                 }
2175
2176                 address += PAGE_SIZE;
2177                 page++;
2178         }
2179 }
2180
2181 static void khugepaged_alloc_sleep(void)
2182 {
2183         DEFINE_WAIT(wait);
2184
2185         add_wait_queue(&khugepaged_wait, &wait);
2186         freezable_schedule_timeout_interruptible(
2187                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2188         remove_wait_queue(&khugepaged_wait, &wait);
2189 }
2190
2191 static int khugepaged_node_load[MAX_NUMNODES];
2192
2193 static bool khugepaged_scan_abort(int nid)
2194 {
2195         int i;
2196
2197         /*
2198          * If zone_reclaim_mode is disabled, then no extra effort is made to
2199          * allocate memory locally.
2200          */
2201         if (!zone_reclaim_mode)
2202                 return false;
2203
2204         /* If there is a count for this node already, it must be acceptable */
2205         if (khugepaged_node_load[nid])
2206                 return false;
2207
2208         for (i = 0; i < MAX_NUMNODES; i++) {
2209                 if (!khugepaged_node_load[i])
2210                         continue;
2211                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2212                         return true;
2213         }
2214         return false;
2215 }
2216
2217 #ifdef CONFIG_NUMA
2218 static int khugepaged_find_target_node(void)
2219 {
2220         static int last_khugepaged_target_node = NUMA_NO_NODE;
2221         int nid, target_node = 0, max_value = 0;
2222
2223         /* find first node with max normal pages hit */
2224         for (nid = 0; nid < MAX_NUMNODES; nid++)
2225                 if (khugepaged_node_load[nid] > max_value) {
2226                         max_value = khugepaged_node_load[nid];
2227                         target_node = nid;
2228                 }
2229
2230         /* do some balance if several nodes have the same hit record */
2231         if (target_node <= last_khugepaged_target_node)
2232                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2233                                 nid++)
2234                         if (max_value == khugepaged_node_load[nid]) {
2235                                 target_node = nid;
2236                                 break;
2237                         }
2238
2239         last_khugepaged_target_node = target_node;
2240         return target_node;
2241 }
2242
2243 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2244 {
2245         if (IS_ERR(*hpage)) {
2246                 if (!*wait)
2247                         return false;
2248
2249                 *wait = false;
2250                 *hpage = NULL;
2251                 khugepaged_alloc_sleep();
2252         } else if (*hpage) {
2253                 put_page(*hpage);
2254                 *hpage = NULL;
2255         }
2256
2257         return true;
2258 }
2259
2260 static struct page *
2261 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2262                        unsigned long address, int node)
2263 {
2264         VM_BUG_ON_PAGE(*hpage, *hpage);
2265
2266         /*
2267          * Before allocating the hugepage, release the mmap_sem read lock.
2268          * The allocation can take potentially a long time if it involves
2269          * sync compaction, and we do not need to hold the mmap_sem during
2270          * that. We will recheck the vma after taking it again in write mode.
2271          */
2272         up_read(&mm->mmap_sem);
2273
2274         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2275         if (unlikely(!*hpage)) {
2276                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2277                 *hpage = ERR_PTR(-ENOMEM);
2278                 return NULL;
2279         }
2280
2281         prep_transhuge_page(*hpage);
2282         count_vm_event(THP_COLLAPSE_ALLOC);
2283         return *hpage;
2284 }
2285 #else
2286 static int khugepaged_find_target_node(void)
2287 {
2288         return 0;
2289 }
2290
2291 static inline struct page *alloc_khugepaged_hugepage(void)
2292 {
2293         struct page *page;
2294
2295         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2296                            HPAGE_PMD_ORDER);
2297         if (page)
2298                 prep_transhuge_page(page);
2299         return page;
2300 }
2301
2302 static struct page *khugepaged_alloc_hugepage(bool *wait)
2303 {
2304         struct page *hpage;
2305
2306         do {
2307                 hpage = alloc_khugepaged_hugepage();
2308                 if (!hpage) {
2309                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2310                         if (!*wait)
2311                                 return NULL;
2312
2313                         *wait = false;
2314                         khugepaged_alloc_sleep();
2315                 } else
2316                         count_vm_event(THP_COLLAPSE_ALLOC);
2317         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2318
2319         return hpage;
2320 }
2321
2322 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2323 {
2324         if (!*hpage)
2325                 *hpage = khugepaged_alloc_hugepage(wait);
2326
2327         if (unlikely(!*hpage))
2328                 return false;
2329
2330         return true;
2331 }
2332
2333 static struct page *
2334 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2335                        unsigned long address, int node)
2336 {
2337         up_read(&mm->mmap_sem);
2338         VM_BUG_ON(!*hpage);
2339
2340         return  *hpage;
2341 }
2342 #endif
2343
2344 static bool hugepage_vma_check(struct vm_area_struct *vma)
2345 {
2346         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2347             (vma->vm_flags & VM_NOHUGEPAGE))
2348                 return false;
2349         if (!vma->anon_vma || vma->vm_ops)
2350                 return false;
2351         if (is_vma_temporary_stack(vma))
2352                 return false;
2353         return !(vma->vm_flags & VM_NO_THP);
2354 }
2355
2356 /*
2357  * If mmap_sem temporarily dropped, revalidate vma
2358  * before taking mmap_sem.
2359  * Return 0 if succeeds, otherwise return none-zero
2360  * value (scan code).
2361  */
2362
2363 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
2364 {
2365         struct vm_area_struct *vma;
2366         unsigned long hstart, hend;
2367
2368         if (unlikely(khugepaged_test_exit(mm)))
2369                 return SCAN_ANY_PROCESS;
2370
2371         vma = find_vma(mm, address);
2372         if (!vma)
2373                 return SCAN_VMA_NULL;
2374
2375         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2376         hend = vma->vm_end & HPAGE_PMD_MASK;
2377         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2378                 return SCAN_ADDRESS_RANGE;
2379         if (!hugepage_vma_check(vma))
2380                 return SCAN_VMA_CHECK;
2381         return 0;
2382 }
2383
2384 /*
2385  * Bring missing pages in from swap, to complete THP collapse.
2386  * Only done if khugepaged_scan_pmd believes it is worthwhile.
2387  *
2388  * Called and returns without pte mapped or spinlocks held,
2389  * but with mmap_sem held to protect against vma changes.
2390  */
2391
2392 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
2393                                         struct vm_area_struct *vma,
2394                                         unsigned long address, pmd_t *pmd)
2395 {
2396         pte_t pteval;
2397         int swapped_in = 0, ret = 0;
2398         struct fault_env fe = {
2399                 .vma = vma,
2400                 .address = address,
2401                 .flags = FAULT_FLAG_ALLOW_RETRY,
2402                 .pmd = pmd,
2403         };
2404
2405         fe.pte = pte_offset_map(pmd, address);
2406         for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
2407                         fe.pte++, fe.address += PAGE_SIZE) {
2408                 pteval = *fe.pte;
2409                 if (!is_swap_pte(pteval))
2410                         continue;
2411                 swapped_in++;
2412                 ret = do_swap_page(&fe, pteval);
2413                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
2414                 if (ret & VM_FAULT_RETRY) {
2415                         down_read(&mm->mmap_sem);
2416                         /* vma is no longer available, don't continue to swapin */
2417                         if (hugepage_vma_revalidate(mm, address))
2418                                 return false;
2419                         /* check if the pmd is still valid */
2420                         if (mm_find_pmd(mm, address) != pmd)
2421                                 return false;
2422                 }
2423                 if (ret & VM_FAULT_ERROR) {
2424                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0);
2425                         return false;
2426                 }
2427                 /* pte is unmapped now, we need to map it */
2428                 fe.pte = pte_offset_map(pmd, fe.address);
2429         }
2430         fe.pte--;
2431         pte_unmap(fe.pte);
2432         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1);
2433         return true;
2434 }
2435
2436 static void collapse_huge_page(struct mm_struct *mm,
2437                                    unsigned long address,
2438                                    struct page **hpage,
2439                                    struct vm_area_struct *vma,
2440                                    int node)
2441 {
2442         pmd_t *pmd, _pmd;
2443         pte_t *pte;
2444         pgtable_t pgtable;
2445         struct page *new_page;
2446         spinlock_t *pmd_ptl, *pte_ptl;
2447         int isolated = 0, result = 0;
2448         struct mem_cgroup *memcg;
2449         unsigned long mmun_start;       /* For mmu_notifiers */
2450         unsigned long mmun_end;         /* For mmu_notifiers */
2451         gfp_t gfp;
2452
2453         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2454
2455         /* Only allocate from the target node */
2456         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2457
2458         /* release the mmap_sem read lock. */
2459         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2460         if (!new_page) {
2461                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2462                 goto out_nolock;
2463         }
2464
2465         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2466                 result = SCAN_CGROUP_CHARGE_FAIL;
2467                 goto out_nolock;
2468         }
2469
2470         down_read(&mm->mmap_sem);
2471         result = hugepage_vma_revalidate(mm, address);
2472         if (result) {
2473                 mem_cgroup_cancel_charge(new_page, memcg, true);
2474                 up_read(&mm->mmap_sem);
2475                 goto out_nolock;
2476         }
2477
2478         pmd = mm_find_pmd(mm, address);
2479         if (!pmd) {
2480                 result = SCAN_PMD_NULL;
2481                 mem_cgroup_cancel_charge(new_page, memcg, true);
2482                 up_read(&mm->mmap_sem);
2483                 goto out_nolock;
2484         }
2485
2486         /*
2487          * __collapse_huge_page_swapin always returns with mmap_sem locked.
2488          * If it fails, release mmap_sem and jump directly out.
2489          * Continuing to collapse causes inconsistency.
2490          */
2491         if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) {
2492                 mem_cgroup_cancel_charge(new_page, memcg, true);
2493                 up_read(&mm->mmap_sem);
2494                 goto out_nolock;
2495         }
2496
2497         up_read(&mm->mmap_sem);
2498         /*
2499          * Prevent all access to pagetables with the exception of
2500          * gup_fast later handled by the ptep_clear_flush and the VM
2501          * handled by the anon_vma lock + PG_lock.
2502          */
2503         down_write(&mm->mmap_sem);
2504         result = hugepage_vma_revalidate(mm, address);
2505         if (result)
2506                 goto out;
2507         /* check if the pmd is still valid */
2508         if (mm_find_pmd(mm, address) != pmd)
2509                 goto out;
2510
2511         anon_vma_lock_write(vma->anon_vma);
2512
2513         pte = pte_offset_map(pmd, address);
2514         pte_ptl = pte_lockptr(mm, pmd);
2515
2516         mmun_start = address;
2517         mmun_end   = address + HPAGE_PMD_SIZE;
2518         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2519         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2520         /*
2521          * After this gup_fast can't run anymore. This also removes
2522          * any huge TLB entry from the CPU so we won't allow
2523          * huge and small TLB entries for the same virtual address
2524          * to avoid the risk of CPU bugs in that area.
2525          */
2526         _pmd = pmdp_collapse_flush(vma, address, pmd);
2527         spin_unlock(pmd_ptl);
2528         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2529
2530         spin_lock(pte_ptl);
2531         isolated = __collapse_huge_page_isolate(vma, address, pte);
2532         spin_unlock(pte_ptl);
2533
2534         if (unlikely(!isolated)) {
2535                 pte_unmap(pte);
2536                 spin_lock(pmd_ptl);
2537                 BUG_ON(!pmd_none(*pmd));
2538                 /*
2539                  * We can only use set_pmd_at when establishing
2540                  * hugepmds and never for establishing regular pmds that
2541                  * points to regular pagetables. Use pmd_populate for that
2542                  */
2543                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2544                 spin_unlock(pmd_ptl);
2545                 anon_vma_unlock_write(vma->anon_vma);
2546                 result = SCAN_FAIL;
2547                 goto out;
2548         }
2549
2550         /*
2551          * All pages are isolated and locked so anon_vma rmap
2552          * can't run anymore.
2553          */
2554         anon_vma_unlock_write(vma->anon_vma);
2555
2556         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2557         pte_unmap(pte);
2558         __SetPageUptodate(new_page);
2559         pgtable = pmd_pgtable(_pmd);
2560
2561         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2562         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2563
2564         /*
2565          * spin_lock() below is not the equivalent of smp_wmb(), so
2566          * this is needed to avoid the copy_huge_page writes to become
2567          * visible after the set_pmd_at() write.
2568          */
2569         smp_wmb();
2570
2571         spin_lock(pmd_ptl);
2572         BUG_ON(!pmd_none(*pmd));
2573         page_add_new_anon_rmap(new_page, vma, address, true);
2574         mem_cgroup_commit_charge(new_page, memcg, false, true);
2575         lru_cache_add_active_or_unevictable(new_page, vma);
2576         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2577         set_pmd_at(mm, address, pmd, _pmd);
2578         update_mmu_cache_pmd(vma, address, pmd);
2579         spin_unlock(pmd_ptl);
2580
2581         *hpage = NULL;
2582
2583         khugepaged_pages_collapsed++;
2584         result = SCAN_SUCCEED;
2585 out_up_write:
2586         up_write(&mm->mmap_sem);
2587 out_nolock:
2588         trace_mm_collapse_huge_page(mm, isolated, result);
2589         return;
2590 out:
2591         mem_cgroup_cancel_charge(new_page, memcg, true);
2592         goto out_up_write;
2593 }
2594
2595 static int khugepaged_scan_pmd(struct mm_struct *mm,
2596                                struct vm_area_struct *vma,
2597                                unsigned long address,
2598                                struct page **hpage)
2599 {
2600         pmd_t *pmd;
2601         pte_t *pte, *_pte;
2602         int ret = 0, none_or_zero = 0, result = 0;
2603         struct page *page = NULL;
2604         unsigned long _address;
2605         spinlock_t *ptl;
2606         int node = NUMA_NO_NODE, unmapped = 0;
2607         bool writable = false, referenced = false;
2608
2609         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2610
2611         pmd = mm_find_pmd(mm, address);
2612         if (!pmd) {
2613                 result = SCAN_PMD_NULL;
2614                 goto out;
2615         }
2616
2617         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2618         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2619         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2620              _pte++, _address += PAGE_SIZE) {
2621                 pte_t pteval = *_pte;
2622                 if (is_swap_pte(pteval)) {
2623                         if (++unmapped <= khugepaged_max_ptes_swap) {
2624                                 continue;
2625                         } else {
2626                                 result = SCAN_EXCEED_SWAP_PTE;
2627                                 goto out_unmap;
2628                         }
2629                 }
2630                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2631                         if (!userfaultfd_armed(vma) &&
2632                             ++none_or_zero <= khugepaged_max_ptes_none) {
2633                                 continue;
2634                         } else {
2635                                 result = SCAN_EXCEED_NONE_PTE;
2636                                 goto out_unmap;
2637                         }
2638                 }
2639                 if (!pte_present(pteval)) {
2640                         result = SCAN_PTE_NON_PRESENT;
2641                         goto out_unmap;
2642                 }
2643                 if (pte_write(pteval))
2644                         writable = true;
2645
2646                 page = vm_normal_page(vma, _address, pteval);
2647                 if (unlikely(!page)) {
2648                         result = SCAN_PAGE_NULL;
2649                         goto out_unmap;
2650                 }
2651
2652                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2653                 if (PageCompound(page)) {
2654                         result = SCAN_PAGE_COMPOUND;
2655                         goto out_unmap;
2656                 }
2657
2658                 /*
2659                  * Record which node the original page is from and save this
2660                  * information to khugepaged_node_load[].
2661                  * Khupaged will allocate hugepage from the node has the max
2662                  * hit record.
2663                  */
2664                 node = page_to_nid(page);
2665                 if (khugepaged_scan_abort(node)) {
2666                         result = SCAN_SCAN_ABORT;
2667                         goto out_unmap;
2668                 }
2669                 khugepaged_node_load[node]++;
2670                 if (!PageLRU(page)) {
2671                         result = SCAN_PAGE_LRU;
2672                         goto out_unmap;
2673                 }
2674                 if (PageLocked(page)) {
2675                         result = SCAN_PAGE_LOCK;
2676                         goto out_unmap;
2677                 }
2678                 if (!PageAnon(page)) {
2679                         result = SCAN_PAGE_ANON;
2680                         goto out_unmap;
2681                 }
2682
2683                 /*
2684                  * cannot use mapcount: can't collapse if there's a gup pin.
2685                  * The page must only be referenced by the scanned process
2686                  * and page swap cache.
2687                  */
2688                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2689                         result = SCAN_PAGE_COUNT;
2690                         goto out_unmap;
2691                 }
2692                 if (pte_young(pteval) ||
2693                     page_is_young(page) || PageReferenced(page) ||
2694                     mmu_notifier_test_young(vma->vm_mm, address))
2695                         referenced = true;
2696         }
2697         if (writable) {
2698                 if (referenced) {
2699                         result = SCAN_SUCCEED;
2700                         ret = 1;
2701                 } else {
2702                         result = SCAN_NO_REFERENCED_PAGE;
2703                 }
2704         } else {
2705                 result = SCAN_PAGE_RO;
2706         }
2707 out_unmap:
2708         pte_unmap_unlock(pte, ptl);
2709         if (ret) {
2710                 node = khugepaged_find_target_node();
2711                 /* collapse_huge_page will return with the mmap_sem released */
2712                 collapse_huge_page(mm, address, hpage, vma, node);
2713         }
2714 out:
2715         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2716                                      none_or_zero, result, unmapped);
2717         return ret;
2718 }
2719
2720 static void collect_mm_slot(struct mm_slot *mm_slot)
2721 {
2722         struct mm_struct *mm = mm_slot->mm;
2723
2724         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2725
2726         if (khugepaged_test_exit(mm)) {
2727                 /* free mm_slot */
2728                 hash_del(&mm_slot->hash);
2729                 list_del(&mm_slot->mm_node);
2730
2731                 /*
2732                  * Not strictly needed because the mm exited already.
2733                  *
2734                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2735                  */
2736
2737                 /* khugepaged_mm_lock actually not necessary for the below */
2738                 free_mm_slot(mm_slot);
2739                 mmdrop(mm);
2740         }
2741 }
2742
2743 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2744                                             struct page **hpage)
2745         __releases(&khugepaged_mm_lock)
2746         __acquires(&khugepaged_mm_lock)
2747 {
2748         struct mm_slot *mm_slot;
2749         struct mm_struct *mm;
2750         struct vm_area_struct *vma;
2751         int progress = 0;
2752
2753         VM_BUG_ON(!pages);
2754         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2755
2756         if (khugepaged_scan.mm_slot)
2757                 mm_slot = khugepaged_scan.mm_slot;
2758         else {
2759                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2760                                      struct mm_slot, mm_node);
2761                 khugepaged_scan.address = 0;
2762                 khugepaged_scan.mm_slot = mm_slot;
2763         }
2764         spin_unlock(&khugepaged_mm_lock);
2765
2766         mm = mm_slot->mm;
2767         down_read(&mm->mmap_sem);
2768         if (unlikely(khugepaged_test_exit(mm)))
2769                 vma = NULL;
2770         else
2771                 vma = find_vma(mm, khugepaged_scan.address);
2772
2773         progress++;
2774         for (; vma; vma = vma->vm_next) {
2775                 unsigned long hstart, hend;
2776
2777                 cond_resched();
2778                 if (unlikely(khugepaged_test_exit(mm))) {
2779                         progress++;
2780                         break;
2781                 }
2782                 if (!hugepage_vma_check(vma)) {
2783 skip:
2784                         progress++;
2785                         continue;
2786                 }
2787                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2788                 hend = vma->vm_end & HPAGE_PMD_MASK;
2789                 if (hstart >= hend)
2790                         goto skip;
2791                 if (khugepaged_scan.address > hend)
2792                         goto skip;
2793                 if (khugepaged_scan.address < hstart)
2794                         khugepaged_scan.address = hstart;
2795                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2796
2797                 while (khugepaged_scan.address < hend) {
2798                         int ret;
2799                         cond_resched();
2800                         if (unlikely(khugepaged_test_exit(mm)))
2801                                 goto breakouterloop;
2802
2803                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2804                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2805                                   hend);
2806                         ret = khugepaged_scan_pmd(mm, vma,
2807                                                   khugepaged_scan.address,
2808                                                   hpage);
2809                         /* move to next address */
2810                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2811                         progress += HPAGE_PMD_NR;
2812                         if (ret)
2813                                 /* we released mmap_sem so break loop */
2814                                 goto breakouterloop_mmap_sem;
2815                         if (progress >= pages)
2816                                 goto breakouterloop;
2817                 }
2818         }
2819 breakouterloop:
2820         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2821 breakouterloop_mmap_sem:
2822
2823         spin_lock(&khugepaged_mm_lock);
2824         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2825         /*
2826          * Release the current mm_slot if this mm is about to die, or
2827          * if we scanned all vmas of this mm.
2828          */
2829         if (khugepaged_test_exit(mm) || !vma) {
2830                 /*
2831                  * Make sure that if mm_users is reaching zero while
2832                  * khugepaged runs here, khugepaged_exit will find
2833                  * mm_slot not pointing to the exiting mm.
2834                  */
2835                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2836                         khugepaged_scan.mm_slot = list_entry(
2837                                 mm_slot->mm_node.next,
2838                                 struct mm_slot, mm_node);
2839                         khugepaged_scan.address = 0;
2840                 } else {
2841                         khugepaged_scan.mm_slot = NULL;
2842                         khugepaged_full_scans++;
2843                 }
2844
2845                 collect_mm_slot(mm_slot);
2846         }
2847
2848         return progress;
2849 }
2850
2851 static int khugepaged_has_work(void)
2852 {
2853         return !list_empty(&khugepaged_scan.mm_head) &&
2854                 khugepaged_enabled();
2855 }
2856
2857 static int khugepaged_wait_event(void)
2858 {
2859         return !list_empty(&khugepaged_scan.mm_head) ||
2860                 kthread_should_stop();
2861 }
2862
2863 static void khugepaged_do_scan(void)
2864 {
2865         struct page *hpage = NULL;
2866         unsigned int progress = 0, pass_through_head = 0;
2867         unsigned int pages = khugepaged_pages_to_scan;
2868         bool wait = true;
2869
2870         barrier(); /* write khugepaged_pages_to_scan to local stack */
2871
2872         while (progress < pages) {
2873                 if (!khugepaged_prealloc_page(&hpage, &wait))
2874                         break;
2875
2876                 cond_resched();
2877
2878                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2879                         break;
2880
2881                 spin_lock(&khugepaged_mm_lock);
2882                 if (!khugepaged_scan.mm_slot)
2883                         pass_through_head++;
2884                 if (khugepaged_has_work() &&
2885                     pass_through_head < 2)
2886                         progress += khugepaged_scan_mm_slot(pages - progress,
2887                                                             &hpage);
2888                 else
2889                         progress = pages;
2890                 spin_unlock(&khugepaged_mm_lock);
2891         }
2892
2893         if (!IS_ERR_OR_NULL(hpage))
2894                 put_page(hpage);
2895 }
2896
2897 static bool khugepaged_should_wakeup(void)
2898 {
2899         return kthread_should_stop() ||
2900                time_after_eq(jiffies, khugepaged_sleep_expire);
2901 }
2902
2903 static void khugepaged_wait_work(void)
2904 {
2905         if (khugepaged_has_work()) {
2906                 const unsigned long scan_sleep_jiffies =
2907                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2908
2909                 if (!scan_sleep_jiffies)
2910                         return;
2911
2912                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2913                 wait_event_freezable_timeout(khugepaged_wait,
2914                                              khugepaged_should_wakeup(),
2915                                              scan_sleep_jiffies);
2916                 return;
2917         }
2918
2919         if (khugepaged_enabled())
2920                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2921 }
2922
2923 static int khugepaged(void *none)
2924 {
2925         struct mm_slot *mm_slot;
2926
2927         set_freezable();
2928         set_user_nice(current, MAX_NICE);
2929
2930         while (!kthread_should_stop()) {
2931                 khugepaged_do_scan();
2932                 khugepaged_wait_work();
2933         }
2934
2935         spin_lock(&khugepaged_mm_lock);
2936         mm_slot = khugepaged_scan.mm_slot;
2937         khugepaged_scan.mm_slot = NULL;
2938         if (mm_slot)
2939                 collect_mm_slot(mm_slot);
2940         spin_unlock(&khugepaged_mm_lock);
2941         return 0;
2942 }
2943
2944 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2945                 unsigned long haddr, pmd_t *pmd)
2946 {
2947         struct mm_struct *mm = vma->vm_mm;
2948         pgtable_t pgtable;
2949         pmd_t _pmd;
2950         int i;
2951
2952         /* leave pmd empty until pte is filled */
2953         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2954
2955         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2956         pmd_populate(mm, &_pmd, pgtable);
2957
2958         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2959                 pte_t *pte, entry;
2960                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2961                 entry = pte_mkspecial(entry);
2962                 pte = pte_offset_map(&_pmd, haddr);
2963                 VM_BUG_ON(!pte_none(*pte));
2964                 set_pte_at(mm, haddr, pte, entry);
2965                 pte_unmap(pte);
2966         }
2967         smp_wmb(); /* make pte visible before pmd */
2968         pmd_populate(mm, pmd, pgtable);
2969         put_huge_zero_page();
2970 }
2971
2972 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2973                 unsigned long haddr, bool freeze)
2974 {
2975         struct mm_struct *mm = vma->vm_mm;
2976         struct page *page;
2977         pgtable_t pgtable;
2978         pmd_t _pmd;
2979         bool young, write, dirty;
2980         unsigned long addr;
2981         int i;
2982
2983         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2984         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2985         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2986         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2987
2988         count_vm_event(THP_SPLIT_PMD);
2989
2990         if (!vma_is_anonymous(vma)) {
2991                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2992                 if (is_huge_zero_pmd(_pmd))
2993                         put_huge_zero_page();
2994                 if (vma_is_dax(vma))
2995                         return;
2996                 page = pmd_page(_pmd);
2997                 if (!PageReferenced(page) && pmd_young(_pmd))
2998                         SetPageReferenced(page);
2999                 page_remove_rmap(page, true);
3000                 put_page(page);
3001                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
3002                 return;
3003         } else if (is_huge_zero_pmd(*pmd)) {
3004                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
3005         }
3006
3007         page = pmd_page(*pmd);
3008         VM_BUG_ON_PAGE(!page_count(page), page);
3009         page_ref_add(page, HPAGE_PMD_NR - 1);
3010         write = pmd_write(*pmd);
3011         young = pmd_young(*pmd);
3012         dirty = pmd_dirty(*pmd);
3013
3014         pmdp_huge_split_prepare(vma, haddr, pmd);
3015         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
3016         pmd_populate(mm, &_pmd, pgtable);
3017
3018         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
3019                 pte_t entry, *pte;
3020                 /*
3021                  * Note that NUMA hinting access restrictions are not
3022                  * transferred to avoid any possibility of altering
3023                  * permissions across VMAs.
3024                  */
3025                 if (freeze) {
3026                         swp_entry_t swp_entry;
3027                         swp_entry = make_migration_entry(page + i, write);
3028                         entry = swp_entry_to_pte(swp_entry);
3029                 } else {
3030                         entry = mk_pte(page + i, vma->vm_page_prot);
3031                         entry = maybe_mkwrite(entry, vma);
3032                         if (!write)
3033                                 entry = pte_wrprotect(entry);
3034                         if (!young)
3035                                 entry = pte_mkold(entry);
3036                 }
3037                 if (dirty)
3038                         SetPageDirty(page + i);
3039                 pte = pte_offset_map(&_pmd, addr);
3040                 BUG_ON(!pte_none(*pte));
3041                 set_pte_at(mm, addr, pte, entry);
3042                 atomic_inc(&page[i]._mapcount);
3043                 pte_unmap(pte);
3044         }
3045
3046         /*
3047          * Set PG_double_map before dropping compound_mapcount to avoid
3048          * false-negative page_mapped().
3049          */
3050         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
3051                 for (i = 0; i < HPAGE_PMD_NR; i++)
3052                         atomic_inc(&page[i]._mapcount);
3053         }
3054
3055         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
3056                 /* Last compound_mapcount is gone. */
3057                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
3058                 if (TestClearPageDoubleMap(page)) {
3059                         /* No need in mapcount reference anymore */
3060                         for (i = 0; i < HPAGE_PMD_NR; i++)
3061                                 atomic_dec(&page[i]._mapcount);
3062                 }
3063         }
3064
3065         smp_wmb(); /* make pte visible before pmd */
3066         /*
3067          * Up to this point the pmd is present and huge and userland has the
3068          * whole access to the hugepage during the split (which happens in
3069          * place). If we overwrite the pmd with the not-huge version pointing
3070          * to the pte here (which of course we could if all CPUs were bug
3071          * free), userland could trigger a small page size TLB miss on the
3072          * small sized TLB while the hugepage TLB entry is still established in
3073          * the huge TLB. Some CPU doesn't like that.
3074          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
3075          * 383 on page 93. Intel should be safe but is also warns that it's
3076          * only safe if the permission and cache attributes of the two entries
3077          * loaded in the two TLB is identical (which should be the case here).
3078          * But it is generally safer to never allow small and huge TLB entries
3079          * for the same virtual address to be loaded simultaneously. So instead
3080          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
3081          * current pmd notpresent (atomically because here the pmd_trans_huge
3082          * and pmd_trans_splitting must remain set at all times on the pmd
3083          * until the split is complete for this pmd), then we flush the SMP TLB
3084          * and finally we write the non-huge version of the pmd entry with
3085          * pmd_populate.
3086          */
3087         pmdp_invalidate(vma, haddr, pmd);
3088         pmd_populate(mm, pmd, pgtable);
3089
3090         if (freeze) {
3091                 for (i = 0; i < HPAGE_PMD_NR; i++) {
3092                         page_remove_rmap(page + i, false);
3093                         put_page(page + i);
3094                 }
3095         }
3096 }
3097
3098 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
3099                 unsigned long address, bool freeze, struct page *page)
3100 {
3101         spinlock_t *ptl;
3102         struct mm_struct *mm = vma->vm_mm;
3103         unsigned long haddr = address & HPAGE_PMD_MASK;
3104
3105         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
3106         ptl = pmd_lock(mm, pmd);
3107
3108         /*
3109          * If caller asks to setup a migration entries, we need a page to check
3110          * pmd against. Otherwise we can end up replacing wrong page.
3111          */
3112         VM_BUG_ON(freeze && !page);
3113         if (page && page != pmd_page(*pmd))
3114                 goto out;
3115
3116         if (pmd_trans_huge(*pmd)) {
3117                 page = pmd_page(*pmd);
3118                 if (PageMlocked(page))
3119                         clear_page_mlock(page);
3120         } else if (!pmd_devmap(*pmd))
3121                 goto out;
3122         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
3123 out:
3124         spin_unlock(ptl);
3125         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3126 }
3127
3128 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3129                 bool freeze, struct page *page)
3130 {
3131         pgd_t *pgd;
3132         pud_t *pud;
3133         pmd_t *pmd;
3134
3135         pgd = pgd_offset(vma->vm_mm, address);
3136         if (!pgd_present(*pgd))
3137                 return;
3138
3139         pud = pud_offset(pgd, address);
3140         if (!pud_present(*pud))
3141                 return;
3142
3143         pmd = pmd_offset(pud, address);
3144
3145         __split_huge_pmd(vma, pmd, address, freeze, page);
3146 }
3147
3148 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3149                              unsigned long start,
3150                              unsigned long end,
3151                              long adjust_next)
3152 {
3153         /*
3154          * If the new start address isn't hpage aligned and it could
3155          * previously contain an hugepage: check if we need to split
3156          * an huge pmd.
3157          */
3158         if (start & ~HPAGE_PMD_MASK &&
3159             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3160             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3161                 split_huge_pmd_address(vma, start, false, NULL);
3162
3163         /*
3164          * If the new end address isn't hpage aligned and it could
3165          * previously contain an hugepage: check if we need to split
3166          * an huge pmd.
3167          */
3168         if (end & ~HPAGE_PMD_MASK &&
3169             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3170             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3171                 split_huge_pmd_address(vma, end, false, NULL);
3172
3173         /*
3174          * If we're also updating the vma->vm_next->vm_start, if the new
3175          * vm_next->vm_start isn't page aligned and it could previously
3176          * contain an hugepage: check if we need to split an huge pmd.
3177          */
3178         if (adjust_next > 0) {
3179                 struct vm_area_struct *next = vma->vm_next;
3180                 unsigned long nstart = next->vm_start;
3181                 nstart += adjust_next << PAGE_SHIFT;
3182                 if (nstart & ~HPAGE_PMD_MASK &&
3183                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3184                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3185                         split_huge_pmd_address(next, nstart, false, NULL);
3186         }
3187 }
3188
3189 static void freeze_page(struct page *page)
3190 {
3191         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
3192                 TTU_RMAP_LOCKED;
3193         int i, ret;
3194
3195         VM_BUG_ON_PAGE(!PageHead(page), page);
3196
3197         if (PageAnon(page))
3198                 ttu_flags |= TTU_MIGRATION;
3199
3200         /* We only need TTU_SPLIT_HUGE_PMD once */
3201         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3202         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3203                 /* Cut short if the page is unmapped */
3204                 if (page_count(page) == 1)
3205                         return;
3206
3207                 ret = try_to_unmap(page + i, ttu_flags);
3208         }
3209         VM_BUG_ON_PAGE(ret, page + i - 1);
3210 }
3211
3212 static void unfreeze_page(struct page *page)
3213 {
3214         int i;
3215
3216         for (i = 0; i < HPAGE_PMD_NR; i++)
3217                 remove_migration_ptes(page + i, page + i, true);
3218 }
3219
3220 static void __split_huge_page_tail(struct page *head, int tail,
3221                 struct lruvec *lruvec, struct list_head *list)
3222 {
3223         struct page *page_tail = head + tail;
3224
3225         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3226         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3227
3228         /*
3229          * tail_page->_refcount is zero and not changing from under us. But
3230          * get_page_unless_zero() may be running from under us on the
3231          * tail_page. If we used atomic_set() below instead of atomic_inc() or
3232          * atomic_add(), we would then run atomic_set() concurrently with
3233          * get_page_unless_zero(), and atomic_set() is implemented in C not
3234          * using locked ops. spin_unlock on x86 sometime uses locked ops
3235          * because of PPro errata 66, 92, so unless somebody can guarantee
3236          * atomic_set() here would be safe on all archs (and not only on x86),
3237          * it's safer to use atomic_inc()/atomic_add().
3238          */
3239         if (PageAnon(head)) {
3240                 page_ref_inc(page_tail);
3241         } else {
3242                 /* Additional pin to radix tree */
3243                 page_ref_add(page_tail, 2);
3244         }
3245
3246         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3247         page_tail->flags |= (head->flags &
3248                         ((1L << PG_referenced) |
3249                          (1L << PG_swapbacked) |
3250                          (1L << PG_mlocked) |
3251                          (1L << PG_uptodate) |
3252                          (1L << PG_active) |
3253                          (1L << PG_locked) |
3254                          (1L << PG_unevictable) |
3255                          (1L << PG_dirty)));
3256
3257         /*
3258          * After clearing PageTail the gup refcount can be released.
3259          * Page flags also must be visible before we make the page non-compound.
3260          */
3261         smp_wmb();
3262
3263         clear_compound_head(page_tail);
3264
3265         if (page_is_young(head))
3266                 set_page_young(page_tail);
3267         if (page_is_idle(head))
3268                 set_page_idle(page_tail);
3269
3270         /* ->mapping in first tail page is compound_mapcount */
3271         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3272                         page_tail);
3273         page_tail->mapping = head->mapping;
3274
3275         page_tail->index = head->index + tail;
3276         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3277         lru_add_page_tail(head, page_tail, lruvec, list);
3278 }
3279
3280 static void __split_huge_page(struct page *page, struct list_head *list,
3281                 unsigned long flags)
3282 {
3283         struct page *head = compound_head(page);
3284         struct zone *zone = page_zone(head);
3285         struct lruvec *lruvec;
3286         pgoff_t end = -1;
3287         int i;
3288
3289         lruvec = mem_cgroup_page_lruvec(head, zone);
3290
3291         /* complete memcg works before add pages to LRU */
3292         mem_cgroup_split_huge_fixup(head);
3293
3294         if (!PageAnon(page))
3295                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
3296
3297         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
3298                 __split_huge_page_tail(head, i, lruvec, list);
3299                 /* Some pages can be beyond i_size: drop them from page cache */
3300                 if (head[i].index >= end) {
3301                         __ClearPageDirty(head + i);
3302                         __delete_from_page_cache(head + i, NULL);
3303                         put_page(head + i);
3304                 }
3305         }
3306
3307         ClearPageCompound(head);
3308         /* See comment in __split_huge_page_tail() */
3309         if (PageAnon(head)) {
3310                 page_ref_inc(head);
3311         } else {
3312                 /* Additional pin to radix tree */
3313                 page_ref_add(head, 2);
3314                 spin_unlock(&head->mapping->tree_lock);
3315         }
3316
3317         spin_unlock_irqrestore(&page_zone(head)->lru_lock, flags);
3318
3319         unfreeze_page(head);
3320
3321         for (i = 0; i < HPAGE_PMD_NR; i++) {
3322                 struct page *subpage = head + i;
3323                 if (subpage == page)
3324                         continue;
3325                 unlock_page(subpage);
3326
3327                 /*
3328                  * Subpages may be freed if there wasn't any mapping
3329                  * like if add_to_swap() is running on a lru page that
3330                  * had its mapping zapped. And freeing these pages
3331                  * requires taking the lru_lock so we do the put_page
3332                  * of the tail pages after the split is complete.
3333                  */
3334                 put_page(subpage);
3335         }
3336 }
3337
3338 int total_mapcount(struct page *page)
3339 {
3340         int i, compound, ret;
3341
3342         VM_BUG_ON_PAGE(PageTail(page), page);
3343
3344         if (likely(!PageCompound(page)))
3345                 return atomic_read(&page->_mapcount) + 1;
3346
3347         compound = compound_mapcount(page);
3348         if (PageHuge(page))
3349                 return compound;
3350         ret = compound;
3351         for (i = 0; i < HPAGE_PMD_NR; i++)
3352                 ret += atomic_read(&page[i]._mapcount) + 1;
3353         /* File pages has compound_mapcount included in _mapcount */
3354         if (!PageAnon(page))
3355                 return ret - compound * HPAGE_PMD_NR;
3356         if (PageDoubleMap(page))
3357                 ret -= HPAGE_PMD_NR;
3358         return ret;
3359 }
3360
3361 /*
3362  * This calculates accurately how many mappings a transparent hugepage
3363  * has (unlike page_mapcount() which isn't fully accurate). This full
3364  * accuracy is primarily needed to know if copy-on-write faults can
3365  * reuse the page and change the mapping to read-write instead of
3366  * copying them. At the same time this returns the total_mapcount too.
3367  *
3368  * The function returns the highest mapcount any one of the subpages
3369  * has. If the return value is one, even if different processes are
3370  * mapping different subpages of the transparent hugepage, they can
3371  * all reuse it, because each process is reusing a different subpage.
3372  *
3373  * The total_mapcount is instead counting all virtual mappings of the
3374  * subpages. If the total_mapcount is equal to "one", it tells the
3375  * caller all mappings belong to the same "mm" and in turn the
3376  * anon_vma of the transparent hugepage can become the vma->anon_vma
3377  * local one as no other process may be mapping any of the subpages.
3378  *
3379  * It would be more accurate to replace page_mapcount() with
3380  * page_trans_huge_mapcount(), however we only use
3381  * page_trans_huge_mapcount() in the copy-on-write faults where we
3382  * need full accuracy to avoid breaking page pinning, because
3383  * page_trans_huge_mapcount() is slower than page_mapcount().
3384  */
3385 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
3386 {
3387         int i, ret, _total_mapcount, mapcount;
3388
3389         /* hugetlbfs shouldn't call it */
3390         VM_BUG_ON_PAGE(PageHuge(page), page);
3391
3392         if (likely(!PageTransCompound(page))) {
3393                 mapcount = atomic_read(&page->_mapcount) + 1;
3394                 if (total_mapcount)
3395                         *total_mapcount = mapcount;
3396                 return mapcount;
3397         }
3398
3399         page = compound_head(page);
3400
3401         _total_mapcount = ret = 0;
3402         for (i = 0; i < HPAGE_PMD_NR; i++) {
3403                 mapcount = atomic_read(&page[i]._mapcount) + 1;
3404                 ret = max(ret, mapcount);
3405                 _total_mapcount += mapcount;
3406         }
3407         if (PageDoubleMap(page)) {
3408                 ret -= 1;
3409                 _total_mapcount -= HPAGE_PMD_NR;
3410         }
3411         mapcount = compound_mapcount(page);
3412         ret += mapcount;
3413         _total_mapcount += mapcount;
3414         if (total_mapcount)
3415                 *total_mapcount = _total_mapcount;
3416         return ret;
3417 }
3418
3419 /*
3420  * This function splits huge page into normal pages. @page can point to any
3421  * subpage of huge page to split. Split doesn't change the position of @page.
3422  *
3423  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3424  * The huge page must be locked.
3425  *
3426  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3427  *
3428  * Both head page and tail pages will inherit mapping, flags, and so on from
3429  * the hugepage.
3430  *
3431  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3432  * they are not mapped.
3433  *
3434  * Returns 0 if the hugepage is split successfully.
3435  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3436  * us.
3437  */
3438 int split_huge_page_to_list(struct page *page, struct list_head *list)
3439 {
3440         struct page *head = compound_head(page);
3441         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3442         struct anon_vma *anon_vma = NULL;
3443         struct address_space *mapping = NULL;
3444         int count, mapcount, extra_pins, ret;
3445         bool mlocked;
3446         unsigned long flags;
3447
3448         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3449         VM_BUG_ON_PAGE(!PageLocked(page), page);
3450         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3451         VM_BUG_ON_PAGE(!PageCompound(page), page);
3452
3453         if (PageAnon(head)) {
3454                 /*
3455                  * The caller does not necessarily hold an mmap_sem that would
3456                  * prevent the anon_vma disappearing so we first we take a
3457                  * reference to it and then lock the anon_vma for write. This
3458                  * is similar to page_lock_anon_vma_read except the write lock
3459                  * is taken to serialise against parallel split or collapse
3460                  * operations.
3461                  */
3462                 anon_vma = page_get_anon_vma(head);
3463                 if (!anon_vma) {
3464                         ret = -EBUSY;
3465                         goto out;
3466                 }
3467                 extra_pins = 0;
3468                 mapping = NULL;
3469                 anon_vma_lock_write(anon_vma);
3470         } else {
3471                 mapping = head->mapping;
3472
3473                 /* Truncated ? */
3474                 if (!mapping) {
3475                         ret = -EBUSY;
3476                         goto out;
3477                 }
3478
3479                 /* Addidional pins from radix tree */
3480                 extra_pins = HPAGE_PMD_NR;
3481                 anon_vma = NULL;
3482                 i_mmap_lock_read(mapping);
3483         }
3484
3485         /*
3486          * Racy check if we can split the page, before freeze_page() will
3487          * split PMDs
3488          */
3489         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
3490                 ret = -EBUSY;
3491                 goto out_unlock;
3492         }
3493
3494         mlocked = PageMlocked(page);
3495         freeze_page(head);
3496         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3497
3498         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3499         if (mlocked)
3500                 lru_add_drain();
3501
3502         /* prevent PageLRU to go away from under us, and freeze lru stats */
3503         spin_lock_irqsave(&page_zone(head)->lru_lock, flags);
3504
3505         if (mapping) {
3506                 void **pslot;
3507
3508                 spin_lock(&mapping->tree_lock);
3509                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
3510                                 page_index(head));
3511                 /*
3512                  * Check if the head page is present in radix tree.
3513                  * We assume all tail are present too, if head is there.
3514                  */
3515                 if (radix_tree_deref_slot_protected(pslot,
3516                                         &mapping->tree_lock) != head)
3517                         goto fail;
3518         }
3519
3520         /* Prevent deferred_split_scan() touching ->_refcount */
3521         spin_lock(&pgdata->split_queue_lock);
3522         count = page_count(head);
3523         mapcount = total_mapcount(head);
3524         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
3525                 if (!list_empty(page_deferred_list(head))) {
3526                         pgdata->split_queue_len--;
3527                         list_del(page_deferred_list(head));
3528                 }
3529                 spin_unlock(&pgdata->split_queue_lock);
3530                 __split_huge_page(page, list, flags);
3531                 ret = 0;
3532         } else {
3533                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3534                         pr_alert("total_mapcount: %u, page_count(): %u\n",
3535                                         mapcount, count);
3536                         if (PageTail(page))
3537                                 dump_page(head, NULL);
3538                         dump_page(page, "total_mapcount(head) > 0");
3539                         BUG();
3540                 }
3541                 spin_unlock(&pgdata->split_queue_lock);
3542 fail:           if (mapping)
3543                         spin_unlock(&mapping->tree_lock);
3544                 spin_unlock_irqrestore(&page_zone(head)->lru_lock, flags);
3545                 unfreeze_page(head);
3546                 ret = -EBUSY;
3547         }
3548
3549 out_unlock:
3550         if (anon_vma) {
3551                 anon_vma_unlock_write(anon_vma);
3552                 put_anon_vma(anon_vma);
3553         }
3554         if (mapping)
3555                 i_mmap_unlock_read(mapping);
3556 out:
3557         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3558         return ret;
3559 }
3560
3561 void free_transhuge_page(struct page *page)
3562 {
3563         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3564         unsigned long flags;
3565
3566         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3567         if (!list_empty(page_deferred_list(page))) {
3568                 pgdata->split_queue_len--;
3569                 list_del(page_deferred_list(page));
3570         }
3571         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3572         free_compound_page(page);
3573 }
3574
3575 void deferred_split_huge_page(struct page *page)
3576 {
3577         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3578         unsigned long flags;
3579
3580         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3581
3582         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3583         if (list_empty(page_deferred_list(page))) {
3584                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3585                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3586                 pgdata->split_queue_len++;
3587         }
3588         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3589 }
3590
3591 static unsigned long deferred_split_count(struct shrinker *shrink,
3592                 struct shrink_control *sc)
3593 {
3594         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3595         return ACCESS_ONCE(pgdata->split_queue_len);
3596 }
3597
3598 static unsigned long deferred_split_scan(struct shrinker *shrink,
3599                 struct shrink_control *sc)
3600 {
3601         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3602         unsigned long flags;
3603         LIST_HEAD(list), *pos, *next;
3604         struct page *page;
3605         int split = 0;
3606
3607         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3608         /* Take pin on all head pages to avoid freeing them under us */
3609         list_for_each_safe(pos, next, &pgdata->split_queue) {
3610                 page = list_entry((void *)pos, struct page, mapping);
3611                 page = compound_head(page);
3612                 if (get_page_unless_zero(page)) {
3613                         list_move(page_deferred_list(page), &list);
3614                 } else {
3615                         /* We lost race with put_compound_page() */
3616                         list_del_init(page_deferred_list(page));
3617                         pgdata->split_queue_len--;
3618                 }
3619                 if (!--sc->nr_to_scan)
3620                         break;
3621         }
3622         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3623
3624         list_for_each_safe(pos, next, &list) {
3625                 page = list_entry((void *)pos, struct page, mapping);
3626                 lock_page(page);
3627                 /* split_huge_page() removes page from list on success */
3628                 if (!split_huge_page(page))
3629                         split++;
3630                 unlock_page(page);
3631                 put_page(page);
3632         }
3633
3634         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3635         list_splice_tail(&list, &pgdata->split_queue);
3636         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3637
3638         /*
3639          * Stop shrinker if we didn't split any page, but the queue is empty.
3640          * This can happen if pages were freed under us.
3641          */
3642         if (!split && list_empty(&pgdata->split_queue))
3643                 return SHRINK_STOP;
3644         return split;
3645 }
3646
3647 static struct shrinker deferred_split_shrinker = {
3648         .count_objects = deferred_split_count,
3649         .scan_objects = deferred_split_scan,
3650         .seeks = DEFAULT_SEEKS,
3651         .flags = SHRINKER_NUMA_AWARE,
3652 };
3653
3654 #ifdef CONFIG_DEBUG_FS
3655 static int split_huge_pages_set(void *data, u64 val)
3656 {
3657         struct zone *zone;
3658         struct page *page;
3659         unsigned long pfn, max_zone_pfn;
3660         unsigned long total = 0, split = 0;
3661
3662         if (val != 1)
3663                 return -EINVAL;
3664
3665         for_each_populated_zone(zone) {
3666                 max_zone_pfn = zone_end_pfn(zone);
3667                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3668                         if (!pfn_valid(pfn))
3669                                 continue;
3670
3671                         page = pfn_to_page(pfn);
3672                         if (!get_page_unless_zero(page))
3673                                 continue;
3674
3675                         if (zone != page_zone(page))
3676                                 goto next;
3677
3678                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3679                                 goto next;
3680
3681                         total++;
3682                         lock_page(page);
3683                         if (!split_huge_page(page))
3684                                 split++;
3685                         unlock_page(page);
3686 next:
3687                         put_page(page);
3688                 }
3689         }
3690
3691         pr_info("%lu of %lu THP split\n", split, total);
3692
3693         return 0;
3694 }
3695 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3696                 "%llu\n");
3697
3698 static int __init split_huge_pages_debugfs(void)
3699 {
3700         void *ret;
3701
3702         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3703                         &split_huge_pages_fops);
3704         if (!ret)
3705                 pr_warn("Failed to create split_huge_pages in debugfs");
3706         return 0;
3707 }
3708 late_initcall(split_huge_pages_debugfs);
3709 #endif