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