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 struct task_struct *khugepaged_thread __read_mostly;
  93 static DEFINE_MUTEX(khugepaged_mutex);
  94 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  96 /*
  97  * default collapse hugepages if there is at least one pte mapped like
  98  * it would have happened if the vma was large enough during page
  99  * fault.
 100  */
 101 static unsigned int khugepaged_max_ptes_none __read_mostly;
 102
 103 static int khugepaged(void *none);
 104 static int khugepaged_slab_init(void);
 105 static void khugepaged_slab_exit(void);
 106
 107 #define MM_SLOTS_HASH_BITS 10
 108 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 109
 110 static struct kmem_cache *mm_slot_cache __read_mostly;
 111
 112 /**
 113  * struct mm_slot - hash lookup from mm to mm_slot
 114  * @hash: hash collision list
 115  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 116  * @mm: the mm that this information is valid for
 117  */
 118 struct mm_slot {
 119         struct hlist_node hash;
 120         struct list_head mm_node;
 121         struct mm_struct *mm;
 122 };
 123
 124 /**
 125  * struct khugepaged_scan - cursor for scanning
 126  * @mm_head: the head of the mm list to scan
 127  * @mm_slot: the current mm_slot we are scanning
 128  * @address: the next address inside that to be scanned
 129  *
 130  * There is only the one khugepaged_scan instance of this cursor structure.
 131  */
 132 struct khugepaged_scan {
 133         struct list_head mm_head;
 134         struct mm_slot *mm_slot;
 135         unsigned long address;
 136 };
 137 static struct khugepaged_scan khugepaged_scan = {
 138         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 139 };
 140
 141 static struct shrinker deferred_split_shrinker;
 142
 143 static void set_recommended_min_free_kbytes(void)
 144 {
 145         struct zone *zone;
 146         int nr_zones = 0;
 147         unsigned long recommended_min;
 148
 149         for_each_populated_zone(zone)
 150                 nr_zones++;
 151
 152         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 153         recommended_min = pageblock_nr_pages * nr_zones * 2;
 154
 155         /*
 156          * Make sure that on average at least two pageblocks are almost free
 157          * of another type, one for a migratetype to fall back to and a
 158          * second to avoid subsequent fallbacks of other types There are 3
 159          * MIGRATE_TYPES we care about.
 160          */
 161         recommended_min += pageblock_nr_pages * nr_zones *
 162                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 163
 164         /* don't ever allow to reserve more than 5% of the lowmem */
 165         recommended_min = min(recommended_min,
 166                               (unsigned long) nr_free_buffer_pages() / 20);
 167         recommended_min <<= (PAGE_SHIFT-10);
 168
 169         if (recommended_min > min_free_kbytes) {
 170                 if (user_min_free_kbytes >= 0)
 171                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
 172                                 min_free_kbytes, recommended_min);
 173
 174                 min_free_kbytes = recommended_min;
 175         }
 176         setup_per_zone_wmarks();
 177 }
 178
 179 static int start_stop_khugepaged(void)
 180 {
 181         int err = 0;
 182         if (khugepaged_enabled()) {
 183                 if (!khugepaged_thread)
 184                         khugepaged_thread = kthread_run(khugepaged, NULL,
 185                                                         "khugepaged");
 186                 if (IS_ERR(khugepaged_thread)) {
 187                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 188                         err = PTR_ERR(khugepaged_thread);
 189                         khugepaged_thread = NULL;
 190                         goto fail;
 191                 }
 192
 193                 if (!list_empty(&khugepaged_scan.mm_head))
 194                         wake_up_interruptible(&khugepaged_wait);
 195
 196                 set_recommended_min_free_kbytes();
 197         } else if (khugepaged_thread) {
 198                 kthread_stop(khugepaged_thread);
 199                 khugepaged_thread = NULL;
 200         }
 201 fail:
 202         return err;
 203 }
 204
 205 static atomic_t huge_zero_refcount;
 206 struct page *huge_zero_page __read_mostly;
 207
 208 struct page *get_huge_zero_page(void)
 209 {
 210         struct page *zero_page;
 211 retry:
 212         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 213                 return READ_ONCE(huge_zero_page);
 214
 215         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 216                         HPAGE_PMD_ORDER);
 217         if (!zero_page) {
 218                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 219                 return NULL;
 220         }
 221         count_vm_event(THP_ZERO_PAGE_ALLOC);
 222         preempt_disable();
 223         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 224                 preempt_enable();
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 goto retry;
 227         }
 228
 229         /* We take additional reference here. It will be put back by shrinker */
 230         atomic_set(&huge_zero_refcount, 2);
 231         preempt_enable();
 232         return READ_ONCE(huge_zero_page);
 233 }
 234
 235 static void put_huge_zero_page(void)
 236 {
 237         /*
 238          * Counter should never go to zero here. Only shrinker can put
 239          * last reference.
 240          */
 241         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 242 }
 243
 244 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 245                                         struct shrink_control *sc)
 246 {
 247         /* we can free zero page only if last reference remains */
 248         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 249 }
 250
 251 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 252                                        struct shrink_control *sc)
 253 {
 254         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 255                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 256                 BUG_ON(zero_page == NULL);
 257                 __free_pages(zero_page, compound_order(zero_page));
 258                 return HPAGE_PMD_NR;
 259         }
 260
 261         return 0;
 262 }
 263
 264 static struct shrinker huge_zero_page_shrinker = {
 265         .count_objects = shrink_huge_zero_page_count,
 266         .scan_objects = shrink_huge_zero_page_scan,
 267         .seeks = DEFAULT_SEEKS,
 268 };
 269
 270 #ifdef CONFIG_SYSFS
 271
 272 static ssize_t triple_flag_store(struct kobject *kobj,
 273                                  struct kobj_attribute *attr,
 274                                  const char *buf, size_t count,
 275                                  enum transparent_hugepage_flag enabled,
 276                                  enum transparent_hugepage_flag deferred,
 277                                  enum transparent_hugepage_flag req_madv)
 278 {
 279         if (!memcmp("defer", buf,
 280                     min(sizeof("defer")-1, count))) {
 281                 if (enabled == deferred)
 282                         return -EINVAL;
 283                 clear_bit(enabled, &transparent_hugepage_flags);
 284                 clear_bit(req_madv, &transparent_hugepage_flags);
 285                 set_bit(deferred, &transparent_hugepage_flags);
 286         } else if (!memcmp("always", buf,
 287                     min(sizeof("always")-1, count))) {
 288                 clear_bit(deferred, &transparent_hugepage_flags);
 289                 clear_bit(req_madv, &transparent_hugepage_flags);
 290                 set_bit(enabled, &transparent_hugepage_flags);
 291         } else if (!memcmp("madvise", buf,
 292                            min(sizeof("madvise")-1, count))) {
 293                 clear_bit(enabled, &transparent_hugepage_flags);
 294                 clear_bit(deferred, &transparent_hugepage_flags);
 295                 set_bit(req_madv, &transparent_hugepage_flags);
 296         } else if (!memcmp("never", buf,
 297                            min(sizeof("never")-1, count))) {
 298                 clear_bit(enabled, &transparent_hugepage_flags);
 299                 clear_bit(req_madv, &transparent_hugepage_flags);
 300                 clear_bit(deferred, &transparent_hugepage_flags);
 301         } else
 302                 return -EINVAL;
 303
 304         return count;
 305 }
 306
 307 static ssize_t enabled_show(struct kobject *kobj,
 308                             struct kobj_attribute *attr, char *buf)
 309 {
 310         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 311                 return sprintf(buf, "[always] madvise never\n");
 312         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 313                 return sprintf(buf, "always [madvise] never\n");
 314         else
 315                 return sprintf(buf, "always madvise [never]\n");
 316 }
 317
 318 static ssize_t enabled_store(struct kobject *kobj,
 319                              struct kobj_attribute *attr,
 320                              const char *buf, size_t count)
 321 {
 322         ssize_t ret;
 323
 324         ret = triple_flag_store(kobj, attr, buf, count,
 325                                 TRANSPARENT_HUGEPAGE_FLAG,
 326                                 TRANSPARENT_HUGEPAGE_FLAG,
 327                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 328
 329         if (ret > 0) {
 330                 int err;
 331
 332                 mutex_lock(&khugepaged_mutex);
 333                 err = start_stop_khugepaged();
 334                 mutex_unlock(&khugepaged_mutex);
 335
 336                 if (err)
 337                         ret = err;
 338         }
 339
 340         return ret;
 341 }
 342 static struct kobj_attribute enabled_attr =
 343         __ATTR(enabled, 0644, enabled_show, enabled_store);
 344
 345 static ssize_t single_flag_show(struct kobject *kobj,
 346                                 struct kobj_attribute *attr, char *buf,
 347                                 enum transparent_hugepage_flag flag)
 348 {
 349         return sprintf(buf, "%d\n",
 350                        !!test_bit(flag, &transparent_hugepage_flags));
 351 }
 352
 353 static ssize_t single_flag_store(struct kobject *kobj,
 354                                  struct kobj_attribute *attr,
 355                                  const char *buf, size_t count,
 356                                  enum transparent_hugepage_flag flag)
 357 {
 358         unsigned long value;
 359         int ret;
 360
 361         ret = kstrtoul(buf, 10, &value);
 362         if (ret < 0)
 363                 return ret;
 364         if (value > 1)
 365                 return -EINVAL;
 366
 367         if (value)
 368                 set_bit(flag, &transparent_hugepage_flags);
 369         else
 370                 clear_bit(flag, &transparent_hugepage_flags);
 371
 372         return count;
 373 }
 374
 375 /*
 376  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 377  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 378  * memory just to allocate one more hugepage.
 379  */
 380 static ssize_t defrag_show(struct kobject *kobj,
 381                            struct kobj_attribute *attr, char *buf)
 382 {
 383         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 384                 return sprintf(buf, "[always] defer madvise never\n");
 385         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 386                 return sprintf(buf, "always [defer] madvise never\n");
 387         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 388                 return sprintf(buf, "always defer [madvise] never\n");
 389         else
 390                 return sprintf(buf, "always defer madvise [never]\n");
 391
 392 }
 393 static ssize_t defrag_store(struct kobject *kobj,
 394                             struct kobj_attribute *attr,
 395                             const char *buf, size_t count)
 396 {
 397         return triple_flag_store(kobj, attr, buf, count,
 398                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 399                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 400                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 401 }
 402 static struct kobj_attribute defrag_attr =
 403         __ATTR(defrag, 0644, defrag_show, defrag_store);
 404
 405 static ssize_t use_zero_page_show(struct kobject *kobj,
 406                 struct kobj_attribute *attr, char *buf)
 407 {
 408         return single_flag_show(kobj, attr, buf,
 409                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 410 }
 411 static ssize_t use_zero_page_store(struct kobject *kobj,
 412                 struct kobj_attribute *attr, const char *buf, size_t count)
 413 {
 414         return single_flag_store(kobj, attr, buf, count,
 415                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 416 }
 417 static struct kobj_attribute use_zero_page_attr =
 418         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 419 #ifdef CONFIG_DEBUG_VM
 420 static ssize_t debug_cow_show(struct kobject *kobj,
 421                                 struct kobj_attribute *attr, char *buf)
 422 {
 423         return single_flag_show(kobj, attr, buf,
 424                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 425 }
 426 static ssize_t debug_cow_store(struct kobject *kobj,
 427                                struct kobj_attribute *attr,
 428                                const char *buf, size_t count)
 429 {
 430         return single_flag_store(kobj, attr, buf, count,
 431                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 432 }
 433 static struct kobj_attribute debug_cow_attr =
 434         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 435 #endif /* CONFIG_DEBUG_VM */
 436
 437 static struct attribute *hugepage_attr[] = {
 438         &enabled_attr.attr,
 439         &defrag_attr.attr,
 440         &use_zero_page_attr.attr,
 441 #ifdef CONFIG_DEBUG_VM
 442         &debug_cow_attr.attr,
 443 #endif
 444         NULL,
 445 };
 446
 447 static struct attribute_group hugepage_attr_group = {
 448         .attrs = hugepage_attr,
 449 };
 450
 451 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 452                                          struct kobj_attribute *attr,
 453                                          char *buf)
 454 {
 455         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 456 }
 457
 458 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 459                                           struct kobj_attribute *attr,
 460                                           const char *buf, size_t count)
 461 {
 462         unsigned long msecs;
 463         int err;
 464
 465         err = kstrtoul(buf, 10, &msecs);
 466         if (err || msecs > UINT_MAX)
 467                 return -EINVAL;
 468
 469         khugepaged_scan_sleep_millisecs = msecs;
 470         wake_up_interruptible(&khugepaged_wait);
 471
 472         return count;
 473 }
 474 static struct kobj_attribute scan_sleep_millisecs_attr =
 475         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 476                scan_sleep_millisecs_store);
 477
 478 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 479                                           struct kobj_attribute *attr,
 480                                           char *buf)
 481 {
 482         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 483 }
 484
 485 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 486                                            struct kobj_attribute *attr,
 487                                            const char *buf, size_t count)
 488 {
 489         unsigned long msecs;
 490         int err;
 491
 492         err = kstrtoul(buf, 10, &msecs);
 493         if (err || msecs > UINT_MAX)
 494                 return -EINVAL;
 495
 496         khugepaged_alloc_sleep_millisecs = msecs;
 497         wake_up_interruptible(&khugepaged_wait);
 498
 499         return count;
 500 }
 501 static struct kobj_attribute alloc_sleep_millisecs_attr =
 502         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 503                alloc_sleep_millisecs_store);
 504
 505 static ssize_t pages_to_scan_show(struct kobject *kobj,
 506                                   struct kobj_attribute *attr,
 507                                   char *buf)
 508 {
 509         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 510 }
 511 static ssize_t pages_to_scan_store(struct kobject *kobj,
 512                                    struct kobj_attribute *attr,
 513                                    const char *buf, size_t count)
 514 {
 515         int err;
 516         unsigned long pages;
 517
 518         err = kstrtoul(buf, 10, &pages);
 519         if (err || !pages || pages > UINT_MAX)
 520                 return -EINVAL;
 521
 522         khugepaged_pages_to_scan = pages;
 523
 524         return count;
 525 }
 526 static struct kobj_attribute pages_to_scan_attr =
 527         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 528                pages_to_scan_store);
 529
 530 static ssize_t pages_collapsed_show(struct kobject *kobj,
 531                                     struct kobj_attribute *attr,
 532                                     char *buf)
 533 {
 534         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 535 }
 536 static struct kobj_attribute pages_collapsed_attr =
 537         __ATTR_RO(pages_collapsed);
 538
 539 static ssize_t full_scans_show(struct kobject *kobj,
 540                                struct kobj_attribute *attr,
 541                                char *buf)
 542 {
 543         return sprintf(buf, "%u\n", khugepaged_full_scans);
 544 }
 545 static struct kobj_attribute full_scans_attr =
 546         __ATTR_RO(full_scans);
 547
 548 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 549                                       struct kobj_attribute *attr, char *buf)
 550 {
 551         return single_flag_show(kobj, attr, buf,
 552                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 553 }
 554 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 555                                        struct kobj_attribute *attr,
 556                                        const char *buf, size_t count)
 557 {
 558         return single_flag_store(kobj, attr, buf, count,
 559                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 560 }
 561 static struct kobj_attribute khugepaged_defrag_attr =
 562         __ATTR(defrag, 0644, khugepaged_defrag_show,
 563                khugepaged_defrag_store);
 564
 565 /*
 566  * max_ptes_none controls if khugepaged should collapse hugepages over
 567  * any unmapped ptes in turn potentially increasing the memory
 568  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 569  * reduce the available free memory in the system as it
 570  * runs. Increasing max_ptes_none will instead potentially reduce the
 571  * free memory in the system during the khugepaged scan.
 572  */
 573 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 574                                              struct kobj_attribute *attr,
 575                                              char *buf)
 576 {
 577         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 578 }
 579 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 580                                               struct kobj_attribute *attr,
 581                                               const char *buf, size_t count)
 582 {
 583         int err;
 584         unsigned long max_ptes_none;
 585
 586         err = kstrtoul(buf, 10, &max_ptes_none);
 587         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 588                 return -EINVAL;
 589
 590         khugepaged_max_ptes_none = max_ptes_none;
 591
 592         return count;
 593 }
 594 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 595         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 596                khugepaged_max_ptes_none_store);
 597
 598 static struct attribute *khugepaged_attr[] = {
 599         &khugepaged_defrag_attr.attr,
 600         &khugepaged_max_ptes_none_attr.attr,
 601         &pages_to_scan_attr.attr,
 602         &pages_collapsed_attr.attr,
 603         &full_scans_attr.attr,
 604         &scan_sleep_millisecs_attr.attr,
 605         &alloc_sleep_millisecs_attr.attr,
 606         NULL,
 607 };
 608
 609 static struct attribute_group khugepaged_attr_group = {
 610         .attrs = khugepaged_attr,
 611         .name = "khugepaged",
 612 };
 613
 614 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 615 {
 616         int err;
 617
 618         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 619         if (unlikely(!*hugepage_kobj)) {
 620                 pr_err("failed to create transparent hugepage kobject\n");
 621                 return -ENOMEM;
 622         }
 623
 624         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 625         if (err) {
 626                 pr_err("failed to register transparent hugepage group\n");
 627                 goto delete_obj;
 628         }
 629
 630         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 631         if (err) {
 632                 pr_err("failed to register transparent hugepage group\n");
 633                 goto remove_hp_group;
 634         }
 635
 636         return 0;
 637
 638 remove_hp_group:
 639         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 640 delete_obj:
 641         kobject_put(*hugepage_kobj);
 642         return err;
 643 }
 644
 645 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 646 {
 647         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 648         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 649         kobject_put(hugepage_kobj);
 650 }
 651 #else
 652 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 653 {
 654         return 0;
 655 }
 656
 657 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 658 {
 659 }
 660 #endif /* CONFIG_SYSFS */
 661
 662 static int __init hugepage_init(void)
 663 {
 664         int err;
 665         struct kobject *hugepage_kobj;
 666
 667         if (!has_transparent_hugepage()) {
 668                 transparent_hugepage_flags = 0;
 669                 return -EINVAL;
 670         }
 671
 672         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 673         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 674         /*
 675          * hugepages can't be allocated by the buddy allocator
 676          */
 677         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 678         /*
 679          * we use page->mapping and page->index in second tail page
 680          * as list_head: assuming THP order >= 2
 681          */
 682         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 683
 684         err = hugepage_init_sysfs(&hugepage_kobj);
 685         if (err)
 686                 goto err_sysfs;
 687
 688         err = khugepaged_slab_init();
 689         if (err)
 690                 goto err_slab;
 691
 692         err = register_shrinker(&huge_zero_page_shrinker);
 693         if (err)
 694                 goto err_hzp_shrinker;
 695         err = register_shrinker(&deferred_split_shrinker);
 696         if (err)
 697                 goto err_split_shrinker;
 698
 699         /*
 700          * By default disable transparent hugepages on smaller systems,
 701          * where the extra memory used could hurt more than TLB overhead
 702          * is likely to save.  The admin can still enable it through /sys.
 703          */
 704         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 705                 transparent_hugepage_flags = 0;
 706                 return 0;
 707         }
 708
 709         err = start_stop_khugepaged();
 710         if (err)
 711                 goto err_khugepaged;
 712
 713         return 0;
 714 err_khugepaged:
 715         unregister_shrinker(&deferred_split_shrinker);
 716 err_split_shrinker:
 717         unregister_shrinker(&huge_zero_page_shrinker);
 718 err_hzp_shrinker:
 719         khugepaged_slab_exit();
 720 err_slab:
 721         hugepage_exit_sysfs(hugepage_kobj);
 722 err_sysfs:
 723         return err;
 724 }
 725 subsys_initcall(hugepage_init);
 726
 727 static int __init setup_transparent_hugepage(char *str)
 728 {
 729         int ret = 0;
 730         if (!str)
 731                 goto out;
 732         if (!strcmp(str, "always")) {
 733                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 734                         &transparent_hugepage_flags);
 735                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 736                           &transparent_hugepage_flags);
 737                 ret = 1;
 738         } else if (!strcmp(str, "madvise")) {
 739                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 740                           &transparent_hugepage_flags);
 741                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 742                         &transparent_hugepage_flags);
 743                 ret = 1;
 744         } else if (!strcmp(str, "never")) {
 745                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 746                           &transparent_hugepage_flags);
 747                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 748                           &transparent_hugepage_flags);
 749                 ret = 1;
 750         }
 751 out:
 752         if (!ret)
 753                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 754         return ret;
 755 }
 756 __setup("transparent_hugepage=", setup_transparent_hugepage);
 757
 758 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 759 {
 760         if (likely(vma->vm_flags & VM_WRITE))
 761                 pmd = pmd_mkwrite(pmd);
 762         return pmd;
 763 }
 764
 765 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 766 {
 767         pmd_t entry;
 768         entry = mk_pmd(page, prot);
 769         entry = pmd_mkhuge(entry);
 770         return entry;
 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. We can do it by checking page_mapcount() on each sub-page, but
1303          * it's expensive.
1304          * The cheaper way is to check page_count() to be equal 1: every
1305          * mapcount takes page reference reference, so this way we can
1306          * guarantee, that the PMD is the only mapping.
1307          * This can give false negative if somebody pinned the page, but that's
1308          * fine.
1309          */
1310         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1311                 pmd_t entry;
1312                 entry = pmd_mkyoung(orig_pmd);
1313                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1314                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1315                         update_mmu_cache_pmd(vma, address, pmd);
1316                 ret |= VM_FAULT_WRITE;
1317                 goto out_unlock;
1318         }
1319         get_page(page);
1320         spin_unlock(ptl);
1321 alloc:
1322         if (transparent_hugepage_enabled(vma) &&
1323             !transparent_hugepage_debug_cow()) {
1324                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1325                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1326         } else
1327                 new_page = NULL;
1328
1329         if (likely(new_page)) {
1330                 prep_transhuge_page(new_page);
1331         } else {
1332                 if (!page) {
1333                         split_huge_pmd(vma, pmd, address);
1334                         ret |= VM_FAULT_FALLBACK;
1335                 } else {
1336                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1337                                         pmd, orig_pmd, page, haddr);
1338                         if (ret & VM_FAULT_OOM) {
1339                                 split_huge_pmd(vma, pmd, address);
1340                                 ret |= VM_FAULT_FALLBACK;
1341                         }
1342                         put_page(page);
1343                 }
1344                 count_vm_event(THP_FAULT_FALLBACK);
1345                 goto out;
1346         }
1347
1348         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1349                                            true))) {
1350                 put_page(new_page);
1351                 if (page) {
1352                         split_huge_pmd(vma, pmd, address);
1353                         put_page(page);
1354                 } else
1355                         split_huge_pmd(vma, pmd, address);
1356                 ret |= VM_FAULT_FALLBACK;
1357                 count_vm_event(THP_FAULT_FALLBACK);
1358                 goto out;
1359         }
1360
1361         count_vm_event(THP_FAULT_ALLOC);
1362
1363         if (!page)
1364                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1365         else
1366                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1367         __SetPageUptodate(new_page);
1368
1369         mmun_start = haddr;
1370         mmun_end   = haddr + HPAGE_PMD_SIZE;
1371         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1372
1373         spin_lock(ptl);
1374         if (page)
1375                 put_page(page);
1376         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1377                 spin_unlock(ptl);
1378                 mem_cgroup_cancel_charge(new_page, memcg, true);
1379                 put_page(new_page);
1380                 goto out_mn;
1381         } else {
1382                 pmd_t entry;
1383                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1384                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1385                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1386                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1387                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1388                 lru_cache_add_active_or_unevictable(new_page, vma);
1389                 set_pmd_at(mm, haddr, pmd, entry);
1390                 update_mmu_cache_pmd(vma, address, pmd);
1391                 if (!page) {
1392                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1393                         put_huge_zero_page();
1394                 } else {
1395                         VM_BUG_ON_PAGE(!PageHead(page), page);
1396                         page_remove_rmap(page, true);
1397                         put_page(page);
1398                 }
1399                 ret |= VM_FAULT_WRITE;
1400         }
1401         spin_unlock(ptl);
1402 out_mn:
1403         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1404 out:
1405         return ret;
1406 out_unlock:
1407         spin_unlock(ptl);
1408         return ret;
1409 }
1410
1411 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1412                                    unsigned long addr,
1413                                    pmd_t *pmd,
1414                                    unsigned int flags)
1415 {
1416         struct mm_struct *mm = vma->vm_mm;
1417         struct page *page = NULL;
1418
1419         assert_spin_locked(pmd_lockptr(mm, pmd));
1420
1421         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1422                 goto out;
1423
1424         /* Avoid dumping huge zero page */
1425         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1426                 return ERR_PTR(-EFAULT);
1427
1428         /* Full NUMA hinting faults to serialise migration in fault paths */
1429         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1430                 goto out;
1431
1432         page = pmd_page(*pmd);
1433         VM_BUG_ON_PAGE(!PageHead(page), page);
1434         if (flags & FOLL_TOUCH)
1435                 touch_pmd(vma, addr, pmd);
1436         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1437                 /*
1438                  * We don't mlock() pte-mapped THPs. This way we can avoid
1439                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1440                  *
1441                  * In most cases the pmd is the only mapping of the page as we
1442                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1443                  * writable private mappings in populate_vma_page_range().
1444                  *
1445                  * The only scenario when we have the page shared here is if we
1446                  * mlocking read-only mapping shared over fork(). We skip
1447                  * mlocking such pages.
1448                  */
1449                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1450                                 page->mapping && trylock_page(page)) {
1451                         lru_add_drain();
1452                         if (page->mapping)
1453                                 mlock_vma_page(page);
1454                         unlock_page(page);
1455                 }
1456         }
1457         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1458         VM_BUG_ON_PAGE(!PageCompound(page), page);
1459         if (flags & FOLL_GET)
1460                 get_page(page);
1461
1462 out:
1463         return page;
1464 }
1465
1466 /* NUMA hinting page fault entry point for trans huge pmds */
1467 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1468                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1469 {
1470         spinlock_t *ptl;
1471         struct anon_vma *anon_vma = NULL;
1472         struct page *page;
1473         unsigned long haddr = addr & HPAGE_PMD_MASK;
1474         int page_nid = -1, this_nid = numa_node_id();
1475         int target_nid, last_cpupid = -1;
1476         bool page_locked;
1477         bool migrated = false;
1478         bool was_writable;
1479         int flags = 0;
1480
1481         /* A PROT_NONE fault should not end up here */
1482         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1483
1484         ptl = pmd_lock(mm, pmdp);
1485         if (unlikely(!pmd_same(pmd, *pmdp)))
1486                 goto out_unlock;
1487
1488         /*
1489          * If there are potential migrations, wait for completion and retry
1490          * without disrupting NUMA hinting information. Do not relock and
1491          * check_same as the page may no longer be mapped.
1492          */
1493         if (unlikely(pmd_trans_migrating(*pmdp))) {
1494                 page = pmd_page(*pmdp);
1495                 spin_unlock(ptl);
1496                 wait_on_page_locked(page);
1497                 goto out;
1498         }
1499
1500         page = pmd_page(pmd);
1501         BUG_ON(is_huge_zero_page(page));
1502         page_nid = page_to_nid(page);
1503         last_cpupid = page_cpupid_last(page);
1504         count_vm_numa_event(NUMA_HINT_FAULTS);
1505         if (page_nid == this_nid) {
1506                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1507                 flags |= TNF_FAULT_LOCAL;
1508         }
1509
1510         /* See similar comment in do_numa_page for explanation */
1511         if (!(vma->vm_flags & VM_WRITE))
1512                 flags |= TNF_NO_GROUP;
1513
1514         /*
1515          * Acquire the page lock to serialise THP migrations but avoid dropping
1516          * page_table_lock if at all possible
1517          */
1518         page_locked = trylock_page(page);
1519         target_nid = mpol_misplaced(page, vma, haddr);
1520         if (target_nid == -1) {
1521                 /* If the page was locked, there are no parallel migrations */
1522                 if (page_locked)
1523                         goto clear_pmdnuma;
1524         }
1525
1526         /* Migration could have started since the pmd_trans_migrating check */
1527         if (!page_locked) {
1528                 spin_unlock(ptl);
1529                 wait_on_page_locked(page);
1530                 page_nid = -1;
1531                 goto out;
1532         }
1533
1534         /*
1535          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1536          * to serialises splits
1537          */
1538         get_page(page);
1539         spin_unlock(ptl);
1540         anon_vma = page_lock_anon_vma_read(page);
1541
1542         /* Confirm the PMD did not change while page_table_lock was released */
1543         spin_lock(ptl);
1544         if (unlikely(!pmd_same(pmd, *pmdp))) {
1545                 unlock_page(page);
1546                 put_page(page);
1547                 page_nid = -1;
1548                 goto out_unlock;
1549         }
1550
1551         /* Bail if we fail to protect against THP splits for any reason */
1552         if (unlikely(!anon_vma)) {
1553                 put_page(page);
1554                 page_nid = -1;
1555                 goto clear_pmdnuma;
1556         }
1557
1558         /*
1559          * Migrate the THP to the requested node, returns with page unlocked
1560          * and access rights restored.
1561          */
1562         spin_unlock(ptl);
1563         migrated = migrate_misplaced_transhuge_page(mm, vma,
1564                                 pmdp, pmd, addr, page, target_nid);
1565         if (migrated) {
1566                 flags |= TNF_MIGRATED;
1567                 page_nid = target_nid;
1568         } else
1569                 flags |= TNF_MIGRATE_FAIL;
1570
1571         goto out;
1572 clear_pmdnuma:
1573         BUG_ON(!PageLocked(page));
1574         was_writable = pmd_write(pmd);
1575         pmd = pmd_modify(pmd, vma->vm_page_prot);
1576         pmd = pmd_mkyoung(pmd);
1577         if (was_writable)
1578                 pmd = pmd_mkwrite(pmd);
1579         set_pmd_at(mm, haddr, pmdp, pmd);
1580         update_mmu_cache_pmd(vma, addr, pmdp);
1581         unlock_page(page);
1582 out_unlock:
1583         spin_unlock(ptl);
1584
1585 out:
1586         if (anon_vma)
1587                 page_unlock_anon_vma_read(anon_vma);
1588
1589         if (page_nid != -1)
1590                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1591
1592         return 0;
1593 }
1594
1595 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1596                 pmd_t *pmd, unsigned long addr, unsigned long next)
1597
1598 {
1599         spinlock_t *ptl;
1600         pmd_t orig_pmd;
1601         struct page *page;
1602         struct mm_struct *mm = tlb->mm;
1603         int ret = 0;
1604
1605         ptl = pmd_trans_huge_lock(pmd, vma);
1606         if (!ptl)
1607                 goto out_unlocked;
1608
1609         orig_pmd = *pmd;
1610         if (is_huge_zero_pmd(orig_pmd)) {
1611                 ret = 1;
1612                 goto out;
1613         }
1614
1615         page = pmd_page(orig_pmd);
1616         /*
1617          * If other processes are mapping this page, we couldn't discard
1618          * the page unless they all do MADV_FREE so let's skip the page.
1619          */
1620         if (page_mapcount(page) != 1)
1621                 goto out;
1622
1623         if (!trylock_page(page))
1624                 goto out;
1625
1626         /*
1627          * If user want to discard part-pages of THP, split it so MADV_FREE
1628          * will deactivate only them.
1629          */
1630         if (next - addr != HPAGE_PMD_SIZE) {
1631                 get_page(page);
1632                 spin_unlock(ptl);
1633                 if (split_huge_page(page)) {
1634                         put_page(page);
1635                         unlock_page(page);
1636                         goto out_unlocked;
1637                 }
1638                 put_page(page);
1639                 unlock_page(page);
1640                 ret = 1;
1641                 goto out_unlocked;
1642         }
1643
1644         if (PageDirty(page))
1645                 ClearPageDirty(page);
1646         unlock_page(page);
1647
1648         if (PageActive(page))
1649                 deactivate_page(page);
1650
1651         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1652                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1653                         tlb->fullmm);
1654                 orig_pmd = pmd_mkold(orig_pmd);
1655                 orig_pmd = pmd_mkclean(orig_pmd);
1656
1657                 set_pmd_at(mm, addr, pmd, orig_pmd);
1658                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1659         }
1660         ret = 1;
1661 out:
1662         spin_unlock(ptl);
1663 out_unlocked:
1664         return ret;
1665 }
1666
1667 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1668                  pmd_t *pmd, unsigned long addr)
1669 {
1670         pmd_t orig_pmd;
1671         spinlock_t *ptl;
1672
1673         ptl = __pmd_trans_huge_lock(pmd, vma);
1674         if (!ptl)
1675                 return 0;
1676         /*
1677          * For architectures like ppc64 we look at deposited pgtable
1678          * when calling pmdp_huge_get_and_clear. So do the
1679          * pgtable_trans_huge_withdraw after finishing pmdp related
1680          * operations.
1681          */
1682         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1683                         tlb->fullmm);
1684         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1685         if (vma_is_dax(vma)) {
1686                 spin_unlock(ptl);
1687                 if (is_huge_zero_pmd(orig_pmd))
1688                         put_huge_zero_page();
1689         } else if (is_huge_zero_pmd(orig_pmd)) {
1690                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1691                 atomic_long_dec(&tlb->mm->nr_ptes);
1692                 spin_unlock(ptl);
1693                 put_huge_zero_page();
1694         } else {
1695                 struct page *page = pmd_page(orig_pmd);
1696                 page_remove_rmap(page, true);
1697                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1698                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1699                 VM_BUG_ON_PAGE(!PageHead(page), page);
1700                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1701                 atomic_long_dec(&tlb->mm->nr_ptes);
1702                 spin_unlock(ptl);
1703                 tlb_remove_page(tlb, page);
1704         }
1705         return 1;
1706 }
1707
1708 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1709                   unsigned long old_addr,
1710                   unsigned long new_addr, unsigned long old_end,
1711                   pmd_t *old_pmd, pmd_t *new_pmd)
1712 {
1713         spinlock_t *old_ptl, *new_ptl;
1714         pmd_t pmd;
1715
1716         struct mm_struct *mm = vma->vm_mm;
1717
1718         if ((old_addr & ~HPAGE_PMD_MASK) ||
1719             (new_addr & ~HPAGE_PMD_MASK) ||
1720             old_end - old_addr < HPAGE_PMD_SIZE ||
1721             (new_vma->vm_flags & VM_NOHUGEPAGE))
1722                 return false;
1723
1724         /*
1725          * The destination pmd shouldn't be established, free_pgtables()
1726          * should have release it.
1727          */
1728         if (WARN_ON(!pmd_none(*new_pmd))) {
1729                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1730                 return false;
1731         }
1732
1733         /*
1734          * We don't have to worry about the ordering of src and dst
1735          * ptlocks because exclusive mmap_sem prevents deadlock.
1736          */
1737         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1738         if (old_ptl) {
1739                 new_ptl = pmd_lockptr(mm, new_pmd);
1740                 if (new_ptl != old_ptl)
1741                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1742                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1743                 VM_BUG_ON(!pmd_none(*new_pmd));
1744
1745                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1746                                 vma_is_anonymous(vma)) {
1747                         pgtable_t pgtable;
1748                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1749                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1750                 }
1751                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1752                 if (new_ptl != old_ptl)
1753                         spin_unlock(new_ptl);
1754                 spin_unlock(old_ptl);
1755                 return true;
1756         }
1757         return false;
1758 }
1759
1760 /*
1761  * Returns
1762  *  - 0 if PMD could not be locked
1763  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1764  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1765  */
1766 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1767                 unsigned long addr, pgprot_t newprot, int prot_numa)
1768 {
1769         struct mm_struct *mm = vma->vm_mm;
1770         spinlock_t *ptl;
1771         int ret = 0;
1772
1773         ptl = __pmd_trans_huge_lock(pmd, vma);
1774         if (ptl) {
1775                 pmd_t entry;
1776                 bool preserve_write = prot_numa && pmd_write(*pmd);
1777                 ret = 1;
1778
1779                 /*
1780                  * Avoid trapping faults against the zero page. The read-only
1781                  * data is likely to be read-cached on the local CPU and
1782                  * local/remote hits to the zero page are not interesting.
1783                  */
1784                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1785                         spin_unlock(ptl);
1786                         return ret;
1787                 }
1788
1789                 if (!prot_numa || !pmd_protnone(*pmd)) {
1790                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1791                         entry = pmd_modify(entry, newprot);
1792                         if (preserve_write)
1793                                 entry = pmd_mkwrite(entry);
1794                         ret = HPAGE_PMD_NR;
1795                         set_pmd_at(mm, addr, pmd, entry);
1796                         BUG_ON(!preserve_write && pmd_write(entry));
1797                 }
1798                 spin_unlock(ptl);
1799         }
1800
1801         return ret;
1802 }
1803
1804 /*
1805  * Returns true if a given pmd maps a thp, false otherwise.
1806  *
1807  * Note that if it returns true, this routine returns without unlocking page
1808  * table lock. So callers must unlock it.
1809  */
1810 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1811 {
1812         spinlock_t *ptl;
1813         ptl = pmd_lock(vma->vm_mm, pmd);
1814         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1815                 return ptl;
1816         spin_unlock(ptl);
1817         return NULL;
1818 }
1819
1820 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1821
1822 int hugepage_madvise(struct vm_area_struct *vma,
1823                      unsigned long *vm_flags, int advice)
1824 {
1825         switch (advice) {
1826         case MADV_HUGEPAGE:
1827 #ifdef CONFIG_S390
1828                 /*
1829                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1830                  * can't handle this properly after s390_enable_sie, so we simply
1831                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1832                  */
1833                 if (mm_has_pgste(vma->vm_mm))
1834                         return 0;
1835 #endif
1836                 /*
1837                  * Be somewhat over-protective like KSM for now!
1838                  */
1839                 if (*vm_flags & VM_NO_THP)
1840                         return -EINVAL;
1841                 *vm_flags &= ~VM_NOHUGEPAGE;
1842                 *vm_flags |= VM_HUGEPAGE;
1843                 /*
1844                  * If the vma become good for khugepaged to scan,
1845                  * register it here without waiting a page fault that
1846                  * may not happen any time soon.
1847                  */
1848                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1849                         return -ENOMEM;
1850                 break;
1851         case MADV_NOHUGEPAGE:
1852                 /*
1853                  * Be somewhat over-protective like KSM for now!
1854                  */
1855                 if (*vm_flags & VM_NO_THP)
1856                         return -EINVAL;
1857                 *vm_flags &= ~VM_HUGEPAGE;
1858                 *vm_flags |= VM_NOHUGEPAGE;
1859                 /*
1860                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1861                  * this vma even if we leave the mm registered in khugepaged if
1862                  * it got registered before VM_NOHUGEPAGE was set.
1863                  */
1864                 break;
1865         }
1866
1867         return 0;
1868 }
1869
1870 static int __init khugepaged_slab_init(void)
1871 {
1872         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1873                                           sizeof(struct mm_slot),
1874                                           __alignof__(struct mm_slot), 0, NULL);
1875         if (!mm_slot_cache)
1876                 return -ENOMEM;
1877
1878         return 0;
1879 }
1880
1881 static void __init khugepaged_slab_exit(void)
1882 {
1883         kmem_cache_destroy(mm_slot_cache);
1884 }
1885
1886 static inline struct mm_slot *alloc_mm_slot(void)
1887 {
1888         if (!mm_slot_cache)     /* initialization failed */
1889                 return NULL;
1890         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1891 }
1892
1893 static inline void free_mm_slot(struct mm_slot *mm_slot)
1894 {
1895         kmem_cache_free(mm_slot_cache, mm_slot);
1896 }
1897
1898 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1899 {
1900         struct mm_slot *mm_slot;
1901
1902         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1903                 if (mm == mm_slot->mm)
1904                         return mm_slot;
1905
1906         return NULL;
1907 }
1908
1909 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1910                                     struct mm_slot *mm_slot)
1911 {
1912         mm_slot->mm = mm;
1913         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1914 }
1915
1916 static inline int khugepaged_test_exit(struct mm_struct *mm)
1917 {
1918         return atomic_read(&mm->mm_users) == 0;
1919 }
1920
1921 int __khugepaged_enter(struct mm_struct *mm)
1922 {
1923         struct mm_slot *mm_slot;
1924         int wakeup;
1925
1926         mm_slot = alloc_mm_slot();
1927         if (!mm_slot)
1928                 return -ENOMEM;
1929
1930         /* __khugepaged_exit() must not run from under us */
1931         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1932         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1933                 free_mm_slot(mm_slot);
1934                 return 0;
1935         }
1936
1937         spin_lock(&khugepaged_mm_lock);
1938         insert_to_mm_slots_hash(mm, mm_slot);
1939         /*
1940          * Insert just behind the scanning cursor, to let the area settle
1941          * down a little.
1942          */
1943         wakeup = list_empty(&khugepaged_scan.mm_head);
1944         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1945         spin_unlock(&khugepaged_mm_lock);
1946
1947         atomic_inc(&mm->mm_count);
1948         if (wakeup)
1949                 wake_up_interruptible(&khugepaged_wait);
1950
1951         return 0;
1952 }
1953
1954 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1955                                unsigned long vm_flags)
1956 {
1957         unsigned long hstart, hend;
1958         if (!vma->anon_vma)
1959                 /*
1960                  * Not yet faulted in so we will register later in the
1961                  * page fault if needed.
1962                  */
1963                 return 0;
1964         if (vma->vm_ops)
1965                 /* khugepaged not yet working on file or special mappings */
1966                 return 0;
1967         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1968         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1969         hend = vma->vm_end & HPAGE_PMD_MASK;
1970         if (hstart < hend)
1971                 return khugepaged_enter(vma, vm_flags);
1972         return 0;
1973 }
1974
1975 void __khugepaged_exit(struct mm_struct *mm)
1976 {
1977         struct mm_slot *mm_slot;
1978         int free = 0;
1979
1980         spin_lock(&khugepaged_mm_lock);
1981         mm_slot = get_mm_slot(mm);
1982         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1983                 hash_del(&mm_slot->hash);
1984                 list_del(&mm_slot->mm_node);
1985                 free = 1;
1986         }
1987         spin_unlock(&khugepaged_mm_lock);
1988
1989         if (free) {
1990                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1991                 free_mm_slot(mm_slot);
1992                 mmdrop(mm);
1993         } else if (mm_slot) {
1994                 /*
1995                  * This is required to serialize against
1996                  * khugepaged_test_exit() (which is guaranteed to run
1997                  * under mmap sem read mode). Stop here (after we
1998                  * return all pagetables will be destroyed) until
1999                  * khugepaged has finished working on the pagetables
2000                  * under the mmap_sem.
2001                  */
2002                 down_write(&mm->mmap_sem);
2003                 up_write(&mm->mmap_sem);
2004         }
2005 }
2006
2007 static void release_pte_page(struct page *page)
2008 {
2009         /* 0 stands for page_is_file_cache(page) == false */
2010         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2011         unlock_page(page);
2012         putback_lru_page(page);
2013 }
2014
2015 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2016 {
2017         while (--_pte >= pte) {
2018                 pte_t pteval = *_pte;
2019                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2020                         release_pte_page(pte_page(pteval));
2021         }
2022 }
2023
2024 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2025                                         unsigned long address,
2026                                         pte_t *pte)
2027 {
2028         struct page *page = NULL;
2029         pte_t *_pte;
2030         int none_or_zero = 0, result = 0;
2031         bool referenced = false, writable = false;
2032
2033         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2034              _pte++, address += PAGE_SIZE) {
2035                 pte_t pteval = *_pte;
2036                 if (pte_none(pteval) || (pte_present(pteval) &&
2037                                 is_zero_pfn(pte_pfn(pteval)))) {
2038                         if (!userfaultfd_armed(vma) &&
2039                             ++none_or_zero <= khugepaged_max_ptes_none) {
2040                                 continue;
2041                         } else {
2042                                 result = SCAN_EXCEED_NONE_PTE;
2043                                 goto out;
2044                         }
2045                 }
2046                 if (!pte_present(pteval)) {
2047                         result = SCAN_PTE_NON_PRESENT;
2048                         goto out;
2049                 }
2050                 page = vm_normal_page(vma, address, pteval);
2051                 if (unlikely(!page)) {
2052                         result = SCAN_PAGE_NULL;
2053                         goto out;
2054                 }
2055
2056                 VM_BUG_ON_PAGE(PageCompound(page), page);
2057                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2058                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2059
2060                 /*
2061                  * We can do it before isolate_lru_page because the
2062                  * page can't be freed from under us. NOTE: PG_lock
2063                  * is needed to serialize against split_huge_page
2064                  * when invoked from the VM.
2065                  */
2066                 if (!trylock_page(page)) {
2067                         result = SCAN_PAGE_LOCK;
2068                         goto out;
2069                 }
2070
2071                 /*
2072                  * cannot use mapcount: can't collapse if there's a gup pin.
2073                  * The page must only be referenced by the scanned process
2074                  * and page swap cache.
2075                  */
2076                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2077                         unlock_page(page);
2078                         result = SCAN_PAGE_COUNT;
2079                         goto out;
2080                 }
2081                 if (pte_write(pteval)) {
2082                         writable = true;
2083                 } else {
2084                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2085                                 unlock_page(page);
2086                                 result = SCAN_SWAP_CACHE_PAGE;
2087                                 goto out;
2088                         }
2089                         /*
2090                          * Page is not in the swap cache. It can be collapsed
2091                          * into a THP.
2092                          */
2093                 }
2094
2095                 /*
2096                  * Isolate the page to avoid collapsing an hugepage
2097                  * currently in use by the VM.
2098                  */
2099                 if (isolate_lru_page(page)) {
2100                         unlock_page(page);
2101                         result = SCAN_DEL_PAGE_LRU;
2102                         goto out;
2103                 }
2104                 /* 0 stands for page_is_file_cache(page) == false */
2105                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2106                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2107                 VM_BUG_ON_PAGE(PageLRU(page), page);
2108
2109                 /* If there is no mapped pte young don't collapse the page */
2110                 if (pte_young(pteval) ||
2111                     page_is_young(page) || PageReferenced(page) ||
2112                     mmu_notifier_test_young(vma->vm_mm, address))
2113                         referenced = true;
2114         }
2115         if (likely(writable)) {
2116                 if (likely(referenced)) {
2117                         result = SCAN_SUCCEED;
2118                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2119                                                             referenced, writable, result);
2120                         return 1;
2121                 }
2122         } else {
2123                 result = SCAN_PAGE_RO;
2124         }
2125
2126 out:
2127         release_pte_pages(pte, _pte);
2128         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2129                                             referenced, writable, result);
2130         return 0;
2131 }
2132
2133 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2134                                       struct vm_area_struct *vma,
2135                                       unsigned long address,
2136                                       spinlock_t *ptl)
2137 {
2138         pte_t *_pte;
2139         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2140                 pte_t pteval = *_pte;
2141                 struct page *src_page;
2142
2143                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2144                         clear_user_highpage(page, address);
2145                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2146                         if (is_zero_pfn(pte_pfn(pteval))) {
2147                                 /*
2148                                  * ptl mostly unnecessary.
2149                                  */
2150                                 spin_lock(ptl);
2151                                 /*
2152                                  * paravirt calls inside pte_clear here are
2153                                  * superfluous.
2154                                  */
2155                                 pte_clear(vma->vm_mm, address, _pte);
2156                                 spin_unlock(ptl);
2157                         }
2158                 } else {
2159                         src_page = pte_page(pteval);
2160                         copy_user_highpage(page, src_page, address, vma);
2161                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2162                         release_pte_page(src_page);
2163                         /*
2164                          * ptl mostly unnecessary, but preempt has to
2165                          * be disabled to update the per-cpu stats
2166                          * inside page_remove_rmap().
2167                          */
2168                         spin_lock(ptl);
2169                         /*
2170                          * paravirt calls inside pte_clear here are
2171                          * superfluous.
2172                          */
2173                         pte_clear(vma->vm_mm, address, _pte);
2174                         page_remove_rmap(src_page, false);
2175                         spin_unlock(ptl);
2176                         free_page_and_swap_cache(src_page);
2177                 }
2178
2179                 address += PAGE_SIZE;
2180                 page++;
2181         }
2182 }
2183
2184 static void khugepaged_alloc_sleep(void)
2185 {
2186         DEFINE_WAIT(wait);
2187
2188         add_wait_queue(&khugepaged_wait, &wait);
2189         freezable_schedule_timeout_interruptible(
2190                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2191         remove_wait_queue(&khugepaged_wait, &wait);
2192 }
2193
2194 static int khugepaged_node_load[MAX_NUMNODES];
2195
2196 static bool khugepaged_scan_abort(int nid)
2197 {
2198         int i;
2199
2200         /*
2201          * If zone_reclaim_mode is disabled, then no extra effort is made to
2202          * allocate memory locally.
2203          */
2204         if (!zone_reclaim_mode)
2205                 return false;
2206
2207         /* If there is a count for this node already, it must be acceptable */
2208         if (khugepaged_node_load[nid])
2209                 return false;
2210
2211         for (i = 0; i < MAX_NUMNODES; i++) {
2212                 if (!khugepaged_node_load[i])
2213                         continue;
2214                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2215                         return true;
2216         }
2217         return false;
2218 }
2219
2220 #ifdef CONFIG_NUMA
2221 static int khugepaged_find_target_node(void)
2222 {
2223         static int last_khugepaged_target_node = NUMA_NO_NODE;
2224         int nid, target_node = 0, max_value = 0;
2225
2226         /* find first node with max normal pages hit */
2227         for (nid = 0; nid < MAX_NUMNODES; nid++)
2228                 if (khugepaged_node_load[nid] > max_value) {
2229                         max_value = khugepaged_node_load[nid];
2230                         target_node = nid;
2231                 }
2232
2233         /* do some balance if several nodes have the same hit record */
2234         if (target_node <= last_khugepaged_target_node)
2235                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2236                                 nid++)
2237                         if (max_value == khugepaged_node_load[nid]) {
2238                                 target_node = nid;
2239                                 break;
2240                         }
2241
2242         last_khugepaged_target_node = target_node;
2243         return target_node;
2244 }
2245
2246 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2247 {
2248         if (IS_ERR(*hpage)) {
2249                 if (!*wait)
2250                         return false;
2251
2252                 *wait = false;
2253                 *hpage = NULL;
2254                 khugepaged_alloc_sleep();
2255         } else if (*hpage) {
2256                 put_page(*hpage);
2257                 *hpage = NULL;
2258         }
2259
2260         return true;
2261 }
2262
2263 static struct page *
2264 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2265                        unsigned long address, int node)
2266 {
2267         VM_BUG_ON_PAGE(*hpage, *hpage);
2268
2269         /*
2270          * Before allocating the hugepage, release the mmap_sem read lock.
2271          * The allocation can take potentially a long time if it involves
2272          * sync compaction, and we do not need to hold the mmap_sem during
2273          * that. We will recheck the vma after taking it again in write mode.
2274          */
2275         up_read(&mm->mmap_sem);
2276
2277         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2278         if (unlikely(!*hpage)) {
2279                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2280                 *hpage = ERR_PTR(-ENOMEM);
2281                 return NULL;
2282         }
2283
2284         prep_transhuge_page(*hpage);
2285         count_vm_event(THP_COLLAPSE_ALLOC);
2286         return *hpage;
2287 }
2288 #else
2289 static int khugepaged_find_target_node(void)
2290 {
2291         return 0;
2292 }
2293
2294 static inline struct page *alloc_khugepaged_hugepage(void)
2295 {
2296         struct page *page;
2297
2298         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2299                            HPAGE_PMD_ORDER);
2300         if (page)
2301                 prep_transhuge_page(page);
2302         return page;
2303 }
2304
2305 static struct page *khugepaged_alloc_hugepage(bool *wait)
2306 {
2307         struct page *hpage;
2308
2309         do {
2310                 hpage = alloc_khugepaged_hugepage();
2311                 if (!hpage) {
2312                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2313                         if (!*wait)
2314                                 return NULL;
2315
2316                         *wait = false;
2317                         khugepaged_alloc_sleep();
2318                 } else
2319                         count_vm_event(THP_COLLAPSE_ALLOC);
2320         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2321
2322         return hpage;
2323 }
2324
2325 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2326 {
2327         if (!*hpage)
2328                 *hpage = khugepaged_alloc_hugepage(wait);
2329
2330         if (unlikely(!*hpage))
2331                 return false;
2332
2333         return true;
2334 }
2335
2336 static struct page *
2337 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2338                        unsigned long address, int node)
2339 {
2340         up_read(&mm->mmap_sem);
2341         VM_BUG_ON(!*hpage);
2342
2343         return  *hpage;
2344 }
2345 #endif
2346
2347 static bool hugepage_vma_check(struct vm_area_struct *vma)
2348 {
2349         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2350             (vma->vm_flags & VM_NOHUGEPAGE))
2351                 return false;
2352         if (!vma->anon_vma || vma->vm_ops)
2353                 return false;
2354         if (is_vma_temporary_stack(vma))
2355                 return false;
2356         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2357         return true;
2358 }
2359
2360 static void collapse_huge_page(struct mm_struct *mm,
2361                                    unsigned long address,
2362                                    struct page **hpage,
2363                                    struct vm_area_struct *vma,
2364                                    int node)
2365 {
2366         pmd_t *pmd, _pmd;
2367         pte_t *pte;
2368         pgtable_t pgtable;
2369         struct page *new_page;
2370         spinlock_t *pmd_ptl, *pte_ptl;
2371         int isolated = 0, result = 0;
2372         unsigned long hstart, hend;
2373         struct mem_cgroup *memcg;
2374         unsigned long mmun_start;       /* For mmu_notifiers */
2375         unsigned long mmun_end;         /* For mmu_notifiers */
2376         gfp_t gfp;
2377
2378         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2379
2380         /* Only allocate from the target node */
2381         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2382
2383         /* release the mmap_sem read lock. */
2384         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2385         if (!new_page) {
2386                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2387                 goto out_nolock;
2388         }
2389
2390         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2391                 result = SCAN_CGROUP_CHARGE_FAIL;
2392                 goto out_nolock;
2393         }
2394
2395         /*
2396          * Prevent all access to pagetables with the exception of
2397          * gup_fast later hanlded by the ptep_clear_flush and the VM
2398          * handled by the anon_vma lock + PG_lock.
2399          */
2400         down_write(&mm->mmap_sem);
2401         if (unlikely(khugepaged_test_exit(mm))) {
2402                 result = SCAN_ANY_PROCESS;
2403                 goto out;
2404         }
2405
2406         vma = find_vma(mm, address);
2407         if (!vma) {
2408                 result = SCAN_VMA_NULL;
2409                 goto out;
2410         }
2411         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2412         hend = vma->vm_end & HPAGE_PMD_MASK;
2413         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2414                 result = SCAN_ADDRESS_RANGE;
2415                 goto out;
2416         }
2417         if (!hugepage_vma_check(vma)) {
2418                 result = SCAN_VMA_CHECK;
2419                 goto out;
2420         }
2421         pmd = mm_find_pmd(mm, address);
2422         if (!pmd) {
2423                 result = SCAN_PMD_NULL;
2424                 goto out;
2425         }
2426
2427         anon_vma_lock_write(vma->anon_vma);
2428
2429         pte = pte_offset_map(pmd, address);
2430         pte_ptl = pte_lockptr(mm, pmd);
2431
2432         mmun_start = address;
2433         mmun_end   = address + HPAGE_PMD_SIZE;
2434         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2435         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2436         /*
2437          * After this gup_fast can't run anymore. This also removes
2438          * any huge TLB entry from the CPU so we won't allow
2439          * huge and small TLB entries for the same virtual address
2440          * to avoid the risk of CPU bugs in that area.
2441          */
2442         _pmd = pmdp_collapse_flush(vma, address, pmd);
2443         spin_unlock(pmd_ptl);
2444         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2445
2446         spin_lock(pte_ptl);
2447         isolated = __collapse_huge_page_isolate(vma, address, pte);
2448         spin_unlock(pte_ptl);
2449
2450         if (unlikely(!isolated)) {
2451                 pte_unmap(pte);
2452                 spin_lock(pmd_ptl);
2453                 BUG_ON(!pmd_none(*pmd));
2454                 /*
2455                  * We can only use set_pmd_at when establishing
2456                  * hugepmds and never for establishing regular pmds that
2457                  * points to regular pagetables. Use pmd_populate for that
2458                  */
2459                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2460                 spin_unlock(pmd_ptl);
2461                 anon_vma_unlock_write(vma->anon_vma);
2462                 result = SCAN_FAIL;
2463                 goto out;
2464         }
2465
2466         /*
2467          * All pages are isolated and locked so anon_vma rmap
2468          * can't run anymore.
2469          */
2470         anon_vma_unlock_write(vma->anon_vma);
2471
2472         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2473         pte_unmap(pte);
2474         __SetPageUptodate(new_page);
2475         pgtable = pmd_pgtable(_pmd);
2476
2477         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2478         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2479
2480         /*
2481          * spin_lock() below is not the equivalent of smp_wmb(), so
2482          * this is needed to avoid the copy_huge_page writes to become
2483          * visible after the set_pmd_at() write.
2484          */
2485         smp_wmb();
2486
2487         spin_lock(pmd_ptl);
2488         BUG_ON(!pmd_none(*pmd));
2489         page_add_new_anon_rmap(new_page, vma, address, true);
2490         mem_cgroup_commit_charge(new_page, memcg, false, true);
2491         lru_cache_add_active_or_unevictable(new_page, vma);
2492         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2493         set_pmd_at(mm, address, pmd, _pmd);
2494         update_mmu_cache_pmd(vma, address, pmd);
2495         spin_unlock(pmd_ptl);
2496
2497         *hpage = NULL;
2498
2499         khugepaged_pages_collapsed++;
2500         result = SCAN_SUCCEED;
2501 out_up_write:
2502         up_write(&mm->mmap_sem);
2503         trace_mm_collapse_huge_page(mm, isolated, result);
2504         return;
2505
2506 out_nolock:
2507         trace_mm_collapse_huge_page(mm, isolated, result);
2508         return;
2509 out:
2510         mem_cgroup_cancel_charge(new_page, memcg, true);
2511         goto out_up_write;
2512 }
2513
2514 static int khugepaged_scan_pmd(struct mm_struct *mm,
2515                                struct vm_area_struct *vma,
2516                                unsigned long address,
2517                                struct page **hpage)
2518 {
2519         pmd_t *pmd;
2520         pte_t *pte, *_pte;
2521         int ret = 0, none_or_zero = 0, result = 0;
2522         struct page *page = NULL;
2523         unsigned long _address;
2524         spinlock_t *ptl;
2525         int node = NUMA_NO_NODE;
2526         bool writable = false, referenced = false;
2527
2528         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2529
2530         pmd = mm_find_pmd(mm, address);
2531         if (!pmd) {
2532                 result = SCAN_PMD_NULL;
2533                 goto out;
2534         }
2535
2536         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2537         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2538         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2539              _pte++, _address += PAGE_SIZE) {
2540                 pte_t pteval = *_pte;
2541                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2542                         if (!userfaultfd_armed(vma) &&
2543                             ++none_or_zero <= khugepaged_max_ptes_none) {
2544                                 continue;
2545                         } else {
2546                                 result = SCAN_EXCEED_NONE_PTE;
2547                                 goto out_unmap;
2548                         }
2549                 }
2550                 if (!pte_present(pteval)) {
2551                         result = SCAN_PTE_NON_PRESENT;
2552                         goto out_unmap;
2553                 }
2554                 if (pte_write(pteval))
2555                         writable = true;
2556
2557                 page = vm_normal_page(vma, _address, pteval);
2558                 if (unlikely(!page)) {
2559                         result = SCAN_PAGE_NULL;
2560                         goto out_unmap;
2561                 }
2562
2563                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2564                 if (PageCompound(page)) {
2565                         result = SCAN_PAGE_COMPOUND;
2566                         goto out_unmap;
2567                 }
2568
2569                 /*
2570                  * Record which node the original page is from and save this
2571                  * information to khugepaged_node_load[].
2572                  * Khupaged will allocate hugepage from the node has the max
2573                  * hit record.
2574                  */
2575                 node = page_to_nid(page);
2576                 if (khugepaged_scan_abort(node)) {
2577                         result = SCAN_SCAN_ABORT;
2578                         goto out_unmap;
2579                 }
2580                 khugepaged_node_load[node]++;
2581                 if (!PageLRU(page)) {
2582                         result = SCAN_PAGE_LRU;
2583                         goto out_unmap;
2584                 }
2585                 if (PageLocked(page)) {
2586                         result = SCAN_PAGE_LOCK;
2587                         goto out_unmap;
2588                 }
2589                 if (!PageAnon(page)) {
2590                         result = SCAN_PAGE_ANON;
2591                         goto out_unmap;
2592                 }
2593
2594                 /*
2595                  * cannot use mapcount: can't collapse if there's a gup pin.
2596                  * The page must only be referenced by the scanned process
2597                  * and page swap cache.
2598                  */
2599                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2600                         result = SCAN_PAGE_COUNT;
2601                         goto out_unmap;
2602                 }
2603                 if (pte_young(pteval) ||
2604                     page_is_young(page) || PageReferenced(page) ||
2605                     mmu_notifier_test_young(vma->vm_mm, address))
2606                         referenced = true;
2607         }
2608         if (writable) {
2609                 if (referenced) {
2610                         result = SCAN_SUCCEED;
2611                         ret = 1;
2612                 } else {
2613                         result = SCAN_NO_REFERENCED_PAGE;
2614                 }
2615         } else {
2616                 result = SCAN_PAGE_RO;
2617         }
2618 out_unmap:
2619         pte_unmap_unlock(pte, ptl);
2620         if (ret) {
2621                 node = khugepaged_find_target_node();
2622                 /* collapse_huge_page will return with the mmap_sem released */
2623                 collapse_huge_page(mm, address, hpage, vma, node);
2624         }
2625 out:
2626         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2627                                      none_or_zero, result);
2628         return ret;
2629 }
2630
2631 static void collect_mm_slot(struct mm_slot *mm_slot)
2632 {
2633         struct mm_struct *mm = mm_slot->mm;
2634
2635         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2636
2637         if (khugepaged_test_exit(mm)) {
2638                 /* free mm_slot */
2639                 hash_del(&mm_slot->hash);
2640                 list_del(&mm_slot->mm_node);
2641
2642                 /*
2643                  * Not strictly needed because the mm exited already.
2644                  *
2645                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2646                  */
2647
2648                 /* khugepaged_mm_lock actually not necessary for the below */
2649                 free_mm_slot(mm_slot);
2650                 mmdrop(mm);
2651         }
2652 }
2653
2654 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2655                                             struct page **hpage)
2656         __releases(&khugepaged_mm_lock)
2657         __acquires(&khugepaged_mm_lock)
2658 {
2659         struct mm_slot *mm_slot;
2660         struct mm_struct *mm;
2661         struct vm_area_struct *vma;
2662         int progress = 0;
2663
2664         VM_BUG_ON(!pages);
2665         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2666
2667         if (khugepaged_scan.mm_slot)
2668                 mm_slot = khugepaged_scan.mm_slot;
2669         else {
2670                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2671                                      struct mm_slot, mm_node);
2672                 khugepaged_scan.address = 0;
2673                 khugepaged_scan.mm_slot = mm_slot;
2674         }
2675         spin_unlock(&khugepaged_mm_lock);
2676
2677         mm = mm_slot->mm;
2678         down_read(&mm->mmap_sem);
2679         if (unlikely(khugepaged_test_exit(mm)))
2680                 vma = NULL;
2681         else
2682                 vma = find_vma(mm, khugepaged_scan.address);
2683
2684         progress++;
2685         for (; vma; vma = vma->vm_next) {
2686                 unsigned long hstart, hend;
2687
2688                 cond_resched();
2689                 if (unlikely(khugepaged_test_exit(mm))) {
2690                         progress++;
2691                         break;
2692                 }
2693                 if (!hugepage_vma_check(vma)) {
2694 skip:
2695                         progress++;
2696                         continue;
2697                 }
2698                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2699                 hend = vma->vm_end & HPAGE_PMD_MASK;
2700                 if (hstart >= hend)
2701                         goto skip;
2702                 if (khugepaged_scan.address > hend)
2703                         goto skip;
2704                 if (khugepaged_scan.address < hstart)
2705                         khugepaged_scan.address = hstart;
2706                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2707
2708                 while (khugepaged_scan.address < hend) {
2709                         int ret;
2710                         cond_resched();
2711                         if (unlikely(khugepaged_test_exit(mm)))
2712                                 goto breakouterloop;
2713
2714                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2715                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2716                                   hend);
2717                         ret = khugepaged_scan_pmd(mm, vma,
2718                                                   khugepaged_scan.address,
2719                                                   hpage);
2720                         /* move to next address */
2721                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2722                         progress += HPAGE_PMD_NR;
2723                         if (ret)
2724                                 /* we released mmap_sem so break loop */
2725                                 goto breakouterloop_mmap_sem;
2726                         if (progress >= pages)
2727                                 goto breakouterloop;
2728                 }
2729         }
2730 breakouterloop:
2731         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2732 breakouterloop_mmap_sem:
2733
2734         spin_lock(&khugepaged_mm_lock);
2735         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2736         /*
2737          * Release the current mm_slot if this mm is about to die, or
2738          * if we scanned all vmas of this mm.
2739          */
2740         if (khugepaged_test_exit(mm) || !vma) {
2741                 /*
2742                  * Make sure that if mm_users is reaching zero while
2743                  * khugepaged runs here, khugepaged_exit will find
2744                  * mm_slot not pointing to the exiting mm.
2745                  */
2746                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2747                         khugepaged_scan.mm_slot = list_entry(
2748                                 mm_slot->mm_node.next,
2749                                 struct mm_slot, mm_node);
2750                         khugepaged_scan.address = 0;
2751                 } else {
2752                         khugepaged_scan.mm_slot = NULL;
2753                         khugepaged_full_scans++;
2754                 }
2755
2756                 collect_mm_slot(mm_slot);
2757         }
2758
2759         return progress;
2760 }
2761
2762 static int khugepaged_has_work(void)
2763 {
2764         return !list_empty(&khugepaged_scan.mm_head) &&
2765                 khugepaged_enabled();
2766 }
2767
2768 static int khugepaged_wait_event(void)
2769 {
2770         return !list_empty(&khugepaged_scan.mm_head) ||
2771                 kthread_should_stop();
2772 }
2773
2774 static void khugepaged_do_scan(void)
2775 {
2776         struct page *hpage = NULL;
2777         unsigned int progress = 0, pass_through_head = 0;
2778         unsigned int pages = khugepaged_pages_to_scan;
2779         bool wait = true;
2780
2781         barrier(); /* write khugepaged_pages_to_scan to local stack */
2782
2783         while (progress < pages) {
2784                 if (!khugepaged_prealloc_page(&hpage, &wait))
2785                         break;
2786
2787                 cond_resched();
2788
2789                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2790                         break;
2791
2792                 spin_lock(&khugepaged_mm_lock);
2793                 if (!khugepaged_scan.mm_slot)
2794                         pass_through_head++;
2795                 if (khugepaged_has_work() &&
2796                     pass_through_head < 2)
2797                         progress += khugepaged_scan_mm_slot(pages - progress,
2798                                                             &hpage);
2799                 else
2800                         progress = pages;
2801                 spin_unlock(&khugepaged_mm_lock);
2802         }
2803
2804         if (!IS_ERR_OR_NULL(hpage))
2805                 put_page(hpage);
2806 }
2807
2808 static void khugepaged_wait_work(void)
2809 {
2810         if (khugepaged_has_work()) {
2811                 if (!khugepaged_scan_sleep_millisecs)
2812                         return;
2813
2814                 wait_event_freezable_timeout(khugepaged_wait,
2815                                              kthread_should_stop(),
2816                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
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, so a huge pmd cannot
3041          * materialize from under us.
3042          */
3043         __split_huge_pmd(vma, pmd, address, freeze);
3044 }
3045
3046 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3047                              unsigned long start,
3048                              unsigned long end,
3049                              long adjust_next)
3050 {
3051         /*
3052          * If the new start address isn't hpage aligned and it could
3053          * previously contain an hugepage: check if we need to split
3054          * an huge pmd.
3055          */
3056         if (start & ~HPAGE_PMD_MASK &&
3057             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3058             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3059                 split_huge_pmd_address(vma, start, false, NULL);
3060
3061         /*
3062          * If the new end address isn't hpage aligned and it could
3063          * previously contain an hugepage: check if we need to split
3064          * an huge pmd.
3065          */
3066         if (end & ~HPAGE_PMD_MASK &&
3067             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3068             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3069                 split_huge_pmd_address(vma, end, false, NULL);
3070
3071         /*
3072          * If we're also updating the vma->vm_next->vm_start, if the new
3073          * vm_next->vm_start isn't page aligned and it could previously
3074          * contain an hugepage: check if we need to split an huge pmd.
3075          */
3076         if (adjust_next > 0) {
3077                 struct vm_area_struct *next = vma->vm_next;
3078                 unsigned long nstart = next->vm_start;
3079                 nstart += adjust_next << PAGE_SHIFT;
3080                 if (nstart & ~HPAGE_PMD_MASK &&
3081                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3082                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3083                         split_huge_pmd_address(next, nstart, false, NULL);
3084         }
3085 }
3086
3087 static void freeze_page(struct page *page)
3088 {
3089         enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3090                 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3091         int i, ret;
3092
3093         VM_BUG_ON_PAGE(!PageHead(page), page);
3094
3095         /* We only need TTU_SPLIT_HUGE_PMD once */
3096         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3097         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3098                 /* Cut short if the page is unmapped */
3099                 if (page_count(page) == 1)
3100                         return;
3101
3102                 ret = try_to_unmap(page + i, ttu_flags);
3103         }
3104         VM_BUG_ON(ret);
3105 }
3106
3107 static void unfreeze_page(struct page *page)
3108 {
3109         int i;
3110
3111         for (i = 0; i < HPAGE_PMD_NR; i++)
3112                 remove_migration_ptes(page + i, page + i, true);
3113 }
3114
3115 static void __split_huge_page_tail(struct page *head, int tail,
3116                 struct lruvec *lruvec, struct list_head *list)
3117 {
3118         struct page *page_tail = head + tail;
3119
3120         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3121         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3122
3123         /*
3124          * tail_page->_count is zero and not changing from under us. But
3125          * get_page_unless_zero() may be running from under us on the
3126          * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3127          * would then run atomic_set() concurrently with
3128          * get_page_unless_zero(), and atomic_set() is implemented in C not
3129          * using locked ops. spin_unlock on x86 sometime uses locked ops
3130          * because of PPro errata 66, 92, so unless somebody can guarantee
3131          * atomic_set() here would be safe on all archs (and not only on x86),
3132          * it's safer to use atomic_inc().
3133          */
3134         page_ref_inc(page_tail);
3135
3136         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3137         page_tail->flags |= (head->flags &
3138                         ((1L << PG_referenced) |
3139                          (1L << PG_swapbacked) |
3140                          (1L << PG_mlocked) |
3141                          (1L << PG_uptodate) |
3142                          (1L << PG_active) |
3143                          (1L << PG_locked) |
3144                          (1L << PG_unevictable) |
3145                          (1L << PG_dirty)));
3146
3147         /*
3148          * After clearing PageTail the gup refcount can be released.
3149          * Page flags also must be visible before we make the page non-compound.
3150          */
3151         smp_wmb();
3152
3153         clear_compound_head(page_tail);
3154
3155         if (page_is_young(head))
3156                 set_page_young(page_tail);
3157         if (page_is_idle(head))
3158                 set_page_idle(page_tail);
3159
3160         /* ->mapping in first tail page is compound_mapcount */
3161         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3162                         page_tail);
3163         page_tail->mapping = head->mapping;
3164
3165         page_tail->index = head->index + tail;
3166         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3167         lru_add_page_tail(head, page_tail, lruvec, list);
3168 }
3169
3170 static void __split_huge_page(struct page *page, struct list_head *list)
3171 {
3172         struct page *head = compound_head(page);
3173         struct zone *zone = page_zone(head);
3174         struct lruvec *lruvec;
3175         int i;
3176
3177         /* prevent PageLRU to go away from under us, and freeze lru stats */
3178         spin_lock_irq(&zone->lru_lock);
3179         lruvec = mem_cgroup_page_lruvec(head, zone);
3180
3181         /* complete memcg works before add pages to LRU */
3182         mem_cgroup_split_huge_fixup(head);
3183
3184         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3185                 __split_huge_page_tail(head, i, lruvec, list);
3186
3187         ClearPageCompound(head);
3188         spin_unlock_irq(&zone->lru_lock);
3189
3190         unfreeze_page(head);
3191
3192         for (i = 0; i < HPAGE_PMD_NR; i++) {
3193                 struct page *subpage = head + i;
3194                 if (subpage == page)
3195                         continue;
3196                 unlock_page(subpage);
3197
3198                 /*
3199                  * Subpages may be freed if there wasn't any mapping
3200                  * like if add_to_swap() is running on a lru page that
3201                  * had its mapping zapped. And freeing these pages
3202                  * requires taking the lru_lock so we do the put_page
3203                  * of the tail pages after the split is complete.
3204                  */
3205                 put_page(subpage);
3206         }
3207 }
3208
3209 int total_mapcount(struct page *page)
3210 {
3211         int i, ret;
3212
3213         VM_BUG_ON_PAGE(PageTail(page), page);
3214
3215         if (likely(!PageCompound(page)))
3216                 return atomic_read(&page->_mapcount) + 1;
3217
3218         ret = compound_mapcount(page);
3219         if (PageHuge(page))
3220                 return ret;
3221         for (i = 0; i < HPAGE_PMD_NR; i++)
3222                 ret += atomic_read(&page[i]._mapcount) + 1;
3223         if (PageDoubleMap(page))
3224                 ret -= HPAGE_PMD_NR;
3225         return ret;
3226 }
3227
3228 /*
3229  * This function splits huge page into normal pages. @page can point to any
3230  * subpage of huge page to split. Split doesn't change the position of @page.
3231  *
3232  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3233  * The huge page must be locked.
3234  *
3235  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3236  *
3237  * Both head page and tail pages will inherit mapping, flags, and so on from
3238  * the hugepage.
3239  *
3240  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3241  * they are not mapped.
3242  *
3243  * Returns 0 if the hugepage is split successfully.
3244  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3245  * us.
3246  */
3247 int split_huge_page_to_list(struct page *page, struct list_head *list)
3248 {
3249         struct page *head = compound_head(page);
3250         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3251         struct anon_vma *anon_vma;
3252         int count, mapcount, ret;
3253         bool mlocked;
3254         unsigned long flags;
3255
3256         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3257         VM_BUG_ON_PAGE(!PageAnon(page), page);
3258         VM_BUG_ON_PAGE(!PageLocked(page), page);
3259         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3260         VM_BUG_ON_PAGE(!PageCompound(page), page);
3261
3262         /*
3263          * The caller does not necessarily hold an mmap_sem that would prevent
3264          * the anon_vma disappearing so we first we take a reference to it
3265          * and then lock the anon_vma for write. This is similar to
3266          * page_lock_anon_vma_read except the write lock is taken to serialise
3267          * against parallel split or collapse operations.
3268          */
3269         anon_vma = page_get_anon_vma(head);
3270         if (!anon_vma) {
3271                 ret = -EBUSY;
3272                 goto out;
3273         }
3274         anon_vma_lock_write(anon_vma);
3275
3276         /*
3277          * Racy check if we can split the page, before freeze_page() will
3278          * split PMDs
3279          */
3280         if (total_mapcount(head) != page_count(head) - 1) {
3281                 ret = -EBUSY;
3282                 goto out_unlock;
3283         }
3284
3285         mlocked = PageMlocked(page);
3286         freeze_page(head);
3287         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3288
3289         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3290         if (mlocked)
3291                 lru_add_drain();
3292
3293         /* Prevent deferred_split_scan() touching ->_count */
3294         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3295         count = page_count(head);
3296         mapcount = total_mapcount(head);
3297         if (!mapcount && count == 1) {
3298                 if (!list_empty(page_deferred_list(head))) {
3299                         pgdata->split_queue_len--;
3300                         list_del(page_deferred_list(head));
3301                 }
3302                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3303                 __split_huge_page(page, list);
3304                 ret = 0;
3305         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3306                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3307                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3308                                 mapcount, count);
3309                 if (PageTail(page))
3310                         dump_page(head, NULL);
3311                 dump_page(page, "total_mapcount(head) > 0");
3312                 BUG();
3313         } else {
3314                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3315                 unfreeze_page(head);
3316                 ret = -EBUSY;
3317         }
3318
3319 out_unlock:
3320         anon_vma_unlock_write(anon_vma);
3321         put_anon_vma(anon_vma);
3322 out:
3323         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3324         return ret;
3325 }
3326
3327 void free_transhuge_page(struct page *page)
3328 {
3329         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3330         unsigned long flags;
3331
3332         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3333         if (!list_empty(page_deferred_list(page))) {
3334                 pgdata->split_queue_len--;
3335                 list_del(page_deferred_list(page));
3336         }
3337         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3338         free_compound_page(page);
3339 }
3340
3341 void deferred_split_huge_page(struct page *page)
3342 {
3343         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3344         unsigned long flags;
3345
3346         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3347
3348         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3349         if (list_empty(page_deferred_list(page))) {
3350                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3351                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3352                 pgdata->split_queue_len++;
3353         }
3354         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3355 }
3356
3357 static unsigned long deferred_split_count(struct shrinker *shrink,
3358                 struct shrink_control *sc)
3359 {
3360         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3361         return ACCESS_ONCE(pgdata->split_queue_len);
3362 }
3363
3364 static unsigned long deferred_split_scan(struct shrinker *shrink,
3365                 struct shrink_control *sc)
3366 {
3367         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3368         unsigned long flags;
3369         LIST_HEAD(list), *pos, *next;
3370         struct page *page;
3371         int split = 0;
3372
3373         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3374         /* Take pin on all head pages to avoid freeing them under us */
3375         list_for_each_safe(pos, next, &pgdata->split_queue) {
3376                 page = list_entry((void *)pos, struct page, mapping);
3377                 page = compound_head(page);
3378                 if (get_page_unless_zero(page)) {
3379                         list_move(page_deferred_list(page), &list);
3380                 } else {
3381                         /* We lost race with put_compound_page() */
3382                         list_del_init(page_deferred_list(page));
3383                         pgdata->split_queue_len--;
3384                 }
3385                 if (!--sc->nr_to_scan)
3386                         break;
3387         }
3388         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3389
3390         list_for_each_safe(pos, next, &list) {
3391                 page = list_entry((void *)pos, struct page, mapping);
3392                 lock_page(page);
3393                 /* split_huge_page() removes page from list on success */
3394                 if (!split_huge_page(page))
3395                         split++;
3396                 unlock_page(page);
3397                 put_page(page);
3398         }
3399
3400         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3401         list_splice_tail(&list, &pgdata->split_queue);
3402         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3403
3404         /*
3405          * Stop shrinker if we didn't split any page, but the queue is empty.
3406          * This can happen if pages were freed under us.
3407          */
3408         if (!split && list_empty(&pgdata->split_queue))
3409                 return SHRINK_STOP;
3410         return split;
3411 }
3412
3413 static struct shrinker deferred_split_shrinker = {
3414         .count_objects = deferred_split_count,
3415         .scan_objects = deferred_split_scan,
3416         .seeks = DEFAULT_SEEKS,
3417         .flags = SHRINKER_NUMA_AWARE,
3418 };
3419
3420 #ifdef CONFIG_DEBUG_FS
3421 static int split_huge_pages_set(void *data, u64 val)
3422 {
3423         struct zone *zone;
3424         struct page *page;
3425         unsigned long pfn, max_zone_pfn;
3426         unsigned long total = 0, split = 0;
3427
3428         if (val != 1)
3429                 return -EINVAL;
3430
3431         for_each_populated_zone(zone) {
3432                 max_zone_pfn = zone_end_pfn(zone);
3433                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3434                         if (!pfn_valid(pfn))
3435                                 continue;
3436
3437                         page = pfn_to_page(pfn);
3438                         if (!get_page_unless_zero(page))
3439                                 continue;
3440
3441                         if (zone != page_zone(page))
3442                                 goto next;
3443
3444                         if (!PageHead(page) || !PageAnon(page) ||
3445                                         PageHuge(page))
3446                                 goto next;
3447
3448                         total++;
3449                         lock_page(page);
3450                         if (!split_huge_page(page))
3451                                 split++;
3452                         unlock_page(page);
3453 next:
3454                         put_page(page);
3455                 }
3456         }
3457
3458         pr_info("%lu of %lu THP split", split, total);
3459
3460         return 0;
3461 }
3462 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3463                 "%llu\n");
3464
3465 static int __init split_huge_pages_debugfs(void)
3466 {
3467         void *ret;
3468
3469         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3470                         &split_huge_pages_fops);
3471         if (!ret)
3472                 pr_warn("Failed to create split_huge_pages in debugfs");
3473         return 0;
3474 }
3475 late_initcall(split_huge_pages_debugfs);
3476 #endif