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