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
1017 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1018                 pmd_t *pmd)
1019 {
1020         pmd_t _pmd;
1021
1022         /*
1023          * We should set the dirty bit only for FOLL_WRITE but for now
1024          * the dirty bit in the pmd is meaningless.  And if the dirty
1025          * bit will become meaningful and we'll only set it with
1026          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1027          * set the young bit, instead of the current set_pmd_at.
1028          */
1029         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1030         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031                                 pmd, _pmd,  1))
1032                 update_mmu_cache_pmd(vma, addr, pmd);
1033 }
1034
1035 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036                 pmd_t *pmd, int flags)
1037 {
1038         unsigned long pfn = pmd_pfn(*pmd);
1039         struct mm_struct *mm = vma->vm_mm;
1040         struct dev_pagemap *pgmap;
1041         struct page *page;
1042
1043         assert_spin_locked(pmd_lockptr(mm, pmd));
1044
1045         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1046                 return NULL;
1047
1048         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1049                 /* pass */;
1050         else
1051                 return NULL;
1052
1053         if (flags & FOLL_TOUCH)
1054                 touch_pmd(vma, addr, pmd);
1055
1056         /*
1057          * device mapped pages can only be returned if the
1058          * caller will manage the page reference count.
1059          */
1060         if (!(flags & FOLL_GET))
1061                 return ERR_PTR(-EEXIST);
1062
1063         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1064         pgmap = get_dev_pagemap(pfn, NULL);
1065         if (!pgmap)
1066                 return ERR_PTR(-EFAULT);
1067         page = pfn_to_page(pfn);
1068         get_page(page);
1069         put_dev_pagemap(pgmap);
1070
1071         return page;
1072 }
1073
1074 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1075                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1076                   struct vm_area_struct *vma)
1077 {
1078         spinlock_t *dst_ptl, *src_ptl;
1079         struct page *src_page;
1080         pmd_t pmd;
1081         pgtable_t pgtable = NULL;
1082         int ret;
1083
1084         if (!vma_is_dax(vma)) {
1085                 ret = -ENOMEM;
1086                 pgtable = pte_alloc_one(dst_mm, addr);
1087                 if (unlikely(!pgtable))
1088                         goto out;
1089         }
1090
1091         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1092         src_ptl = pmd_lockptr(src_mm, src_pmd);
1093         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1094
1095         ret = -EAGAIN;
1096         pmd = *src_pmd;
1097         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1098                 pte_free(dst_mm, pgtable);
1099                 goto out_unlock;
1100         }
1101         /*
1102          * When page table lock is held, the huge zero pmd should not be
1103          * under splitting since we don't split the page itself, only pmd to
1104          * a page table.
1105          */
1106         if (is_huge_zero_pmd(pmd)) {
1107                 struct page *zero_page;
1108                 /*
1109                  * get_huge_zero_page() will never allocate a new page here,
1110                  * since we already have a zero page to copy. It just takes a
1111                  * reference.
1112                  */
1113                 zero_page = get_huge_zero_page();
1114                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1115                                 zero_page);
1116                 ret = 0;
1117                 goto out_unlock;
1118         }
1119
1120         if (!vma_is_dax(vma)) {
1121                 /* thp accounting separate from pmd_devmap accounting */
1122                 src_page = pmd_page(pmd);
1123                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1124                 get_page(src_page);
1125                 page_dup_rmap(src_page, true);
1126                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1127                 atomic_long_inc(&dst_mm->nr_ptes);
1128                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1129         }
1130
1131         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1132         pmd = pmd_mkold(pmd_wrprotect(pmd));
1133         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1134
1135         ret = 0;
1136 out_unlock:
1137         spin_unlock(src_ptl);
1138         spin_unlock(dst_ptl);
1139 out:
1140         return ret;
1141 }
1142
1143 void huge_pmd_set_accessed(struct mm_struct *mm,
1144                            struct vm_area_struct *vma,
1145                            unsigned long address,
1146                            pmd_t *pmd, pmd_t orig_pmd,
1147                            int dirty)
1148 {
1149         spinlock_t *ptl;
1150         pmd_t entry;
1151         unsigned long haddr;
1152
1153         ptl = pmd_lock(mm, pmd);
1154         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1155                 goto unlock;
1156
1157         entry = pmd_mkyoung(orig_pmd);
1158         haddr = address & HPAGE_PMD_MASK;
1159         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1160                 update_mmu_cache_pmd(vma, address, pmd);
1161
1162 unlock:
1163         spin_unlock(ptl);
1164 }
1165
1166 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1167                                         struct vm_area_struct *vma,
1168                                         unsigned long address,
1169                                         pmd_t *pmd, pmd_t orig_pmd,
1170                                         struct page *page,
1171                                         unsigned long haddr)
1172 {
1173         struct mem_cgroup *memcg;
1174         spinlock_t *ptl;
1175         pgtable_t pgtable;
1176         pmd_t _pmd;
1177         int ret = 0, i;
1178         struct page **pages;
1179         unsigned long mmun_start;       /* For mmu_notifiers */
1180         unsigned long mmun_end;         /* For mmu_notifiers */
1181
1182         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1183                         GFP_KERNEL);
1184         if (unlikely(!pages)) {
1185                 ret |= VM_FAULT_OOM;
1186                 goto out;
1187         }
1188
1189         for (i = 0; i < HPAGE_PMD_NR; i++) {
1190                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1191                                                __GFP_OTHER_NODE,
1192                                                vma, address, page_to_nid(page));
1193                 if (unlikely(!pages[i] ||
1194                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1195                                                    &memcg, false))) {
1196                         if (pages[i])
1197                                 put_page(pages[i]);
1198                         while (--i >= 0) {
1199                                 memcg = (void *)page_private(pages[i]);
1200                                 set_page_private(pages[i], 0);
1201                                 mem_cgroup_cancel_charge(pages[i], memcg,
1202                                                 false);
1203                                 put_page(pages[i]);
1204                         }
1205                         kfree(pages);
1206                         ret |= VM_FAULT_OOM;
1207                         goto out;
1208                 }
1209                 set_page_private(pages[i], (unsigned long)memcg);
1210         }
1211
1212         for (i = 0; i < HPAGE_PMD_NR; i++) {
1213                 copy_user_highpage(pages[i], page + i,
1214                                    haddr + PAGE_SIZE * i, vma);
1215                 __SetPageUptodate(pages[i]);
1216                 cond_resched();
1217         }
1218
1219         mmun_start = haddr;
1220         mmun_end   = haddr + HPAGE_PMD_SIZE;
1221         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1222
1223         ptl = pmd_lock(mm, pmd);
1224         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1225                 goto out_free_pages;
1226         VM_BUG_ON_PAGE(!PageHead(page), page);
1227
1228         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1229         /* leave pmd empty until pte is filled */
1230
1231         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1232         pmd_populate(mm, &_pmd, pgtable);
1233
1234         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1235                 pte_t *pte, entry;
1236                 entry = mk_pte(pages[i], vma->vm_page_prot);
1237                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1238                 memcg = (void *)page_private(pages[i]);
1239                 set_page_private(pages[i], 0);
1240                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1241                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1242                 lru_cache_add_active_or_unevictable(pages[i], vma);
1243                 pte = pte_offset_map(&_pmd, haddr);
1244                 VM_BUG_ON(!pte_none(*pte));
1245                 set_pte_at(mm, haddr, pte, entry);
1246                 pte_unmap(pte);
1247         }
1248         kfree(pages);
1249
1250         smp_wmb(); /* make pte visible before pmd */
1251         pmd_populate(mm, pmd, pgtable);
1252         page_remove_rmap(page, true);
1253         spin_unlock(ptl);
1254
1255         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1256
1257         ret |= VM_FAULT_WRITE;
1258         put_page(page);
1259
1260 out:
1261         return ret;
1262
1263 out_free_pages:
1264         spin_unlock(ptl);
1265         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1266         for (i = 0; i < HPAGE_PMD_NR; i++) {
1267                 memcg = (void *)page_private(pages[i]);
1268                 set_page_private(pages[i], 0);
1269                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1270                 put_page(pages[i]);
1271         }
1272         kfree(pages);
1273         goto out;
1274 }
1275
1276 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1277                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1278 {
1279         spinlock_t *ptl;
1280         int ret = 0;
1281         struct page *page = NULL, *new_page;
1282         struct mem_cgroup *memcg;
1283         unsigned long haddr;
1284         unsigned long mmun_start;       /* For mmu_notifiers */
1285         unsigned long mmun_end;         /* For mmu_notifiers */
1286         gfp_t huge_gfp;                 /* for allocation and charge */
1287
1288         ptl = pmd_lockptr(mm, pmd);
1289         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1290         haddr = address & HPAGE_PMD_MASK;
1291         if (is_huge_zero_pmd(orig_pmd))
1292                 goto alloc;
1293         spin_lock(ptl);
1294         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1295                 goto out_unlock;
1296
1297         page = pmd_page(orig_pmd);
1298         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1299         /*
1300          * We can only reuse the page if nobody else maps the huge page or it's
1301          * part. We can do it by checking page_mapcount() on each sub-page, but
1302          * it's expensive.
1303          * The cheaper way is to check page_count() to be equal 1: every
1304          * mapcount takes page reference reference, so this way we can
1305          * guarantee, that the PMD is the only mapping.
1306          * This can give false negative if somebody pinned the page, but that's
1307          * fine.
1308          */
1309         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1310                 pmd_t entry;
1311                 entry = pmd_mkyoung(orig_pmd);
1312                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1313                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1314                         update_mmu_cache_pmd(vma, address, pmd);
1315                 ret |= VM_FAULT_WRITE;
1316                 goto out_unlock;
1317         }
1318         get_page(page);
1319         spin_unlock(ptl);
1320 alloc:
1321         if (transparent_hugepage_enabled(vma) &&
1322             !transparent_hugepage_debug_cow()) {
1323                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1324                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1325         } else
1326                 new_page = NULL;
1327
1328         if (likely(new_page)) {
1329                 prep_transhuge_page(new_page);
1330         } else {
1331                 if (!page) {
1332                         split_huge_pmd(vma, pmd, address);
1333                         ret |= VM_FAULT_FALLBACK;
1334                 } else {
1335                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1336                                         pmd, orig_pmd, page, haddr);
1337                         if (ret & VM_FAULT_OOM) {
1338                                 split_huge_pmd(vma, pmd, address);
1339                                 ret |= VM_FAULT_FALLBACK;
1340                         }
1341                         put_page(page);
1342                 }
1343                 count_vm_event(THP_FAULT_FALLBACK);
1344                 goto out;
1345         }
1346
1347         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1348                                            true))) {
1349                 put_page(new_page);
1350                 if (page) {
1351                         split_huge_pmd(vma, pmd, address);
1352                         put_page(page);
1353                 } else
1354                         split_huge_pmd(vma, pmd, address);
1355                 ret |= VM_FAULT_FALLBACK;
1356                 count_vm_event(THP_FAULT_FALLBACK);
1357                 goto out;
1358         }
1359
1360         count_vm_event(THP_FAULT_ALLOC);
1361
1362         if (!page)
1363                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1364         else
1365                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1366         __SetPageUptodate(new_page);
1367
1368         mmun_start = haddr;
1369         mmun_end   = haddr + HPAGE_PMD_SIZE;
1370         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1371
1372         spin_lock(ptl);
1373         if (page)
1374                 put_page(page);
1375         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1376                 spin_unlock(ptl);
1377                 mem_cgroup_cancel_charge(new_page, memcg, true);
1378                 put_page(new_page);
1379                 goto out_mn;
1380         } else {
1381                 pmd_t entry;
1382                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1383                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1384                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1385                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1386                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1387                 lru_cache_add_active_or_unevictable(new_page, vma);
1388                 set_pmd_at(mm, haddr, pmd, entry);
1389                 update_mmu_cache_pmd(vma, address, pmd);
1390                 if (!page) {
1391                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1392                         put_huge_zero_page();
1393                 } else {
1394                         VM_BUG_ON_PAGE(!PageHead(page), page);
1395                         page_remove_rmap(page, true);
1396                         put_page(page);
1397                 }
1398                 ret |= VM_FAULT_WRITE;
1399         }
1400         spin_unlock(ptl);
1401 out_mn:
1402         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1403 out:
1404         return ret;
1405 out_unlock:
1406         spin_unlock(ptl);
1407         return ret;
1408 }
1409
1410 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1411                                    unsigned long addr,
1412                                    pmd_t *pmd,
1413                                    unsigned int flags)
1414 {
1415         struct mm_struct *mm = vma->vm_mm;
1416         struct page *page = NULL;
1417
1418         assert_spin_locked(pmd_lockptr(mm, pmd));
1419
1420         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1421                 goto out;
1422
1423         /* Avoid dumping huge zero page */
1424         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1425                 return ERR_PTR(-EFAULT);
1426
1427         /* Full NUMA hinting faults to serialise migration in fault paths */
1428         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1429                 goto out;
1430
1431         page = pmd_page(*pmd);
1432         VM_BUG_ON_PAGE(!PageHead(page), page);
1433         if (flags & FOLL_TOUCH)
1434                 touch_pmd(vma, addr, pmd);
1435         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1436                 /*
1437                  * We don't mlock() pte-mapped THPs. This way we can avoid
1438                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1439                  *
1440                  * In most cases the pmd is the only mapping of the page as we
1441                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1442                  * writable private mappings in populate_vma_page_range().
1443                  *
1444                  * The only scenario when we have the page shared here is if we
1445                  * mlocking read-only mapping shared over fork(). We skip
1446                  * mlocking such pages.
1447                  */
1448                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1449                                 page->mapping && trylock_page(page)) {
1450                         lru_add_drain();
1451                         if (page->mapping)
1452                                 mlock_vma_page(page);
1453                         unlock_page(page);
1454                 }
1455         }
1456         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1457         VM_BUG_ON_PAGE(!PageCompound(page), page);
1458         if (flags & FOLL_GET)
1459                 get_page(page);
1460
1461 out:
1462         return page;
1463 }
1464
1465 /* NUMA hinting page fault entry point for trans huge pmds */
1466 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1467                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1468 {
1469         spinlock_t *ptl;
1470         struct anon_vma *anon_vma = NULL;
1471         struct page *page;
1472         unsigned long haddr = addr & HPAGE_PMD_MASK;
1473         int page_nid = -1, this_nid = numa_node_id();
1474         int target_nid, last_cpupid = -1;
1475         bool page_locked;
1476         bool migrated = false;
1477         bool was_writable;
1478         int flags = 0;
1479
1480         /* A PROT_NONE fault should not end up here */
1481         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1482
1483         ptl = pmd_lock(mm, pmdp);
1484         if (unlikely(!pmd_same(pmd, *pmdp)))
1485                 goto out_unlock;
1486
1487         /*
1488          * If there are potential migrations, wait for completion and retry
1489          * without disrupting NUMA hinting information. Do not relock and
1490          * check_same as the page may no longer be mapped.
1491          */
1492         if (unlikely(pmd_trans_migrating(*pmdp))) {
1493                 page = pmd_page(*pmdp);
1494                 spin_unlock(ptl);
1495                 wait_on_page_locked(page);
1496                 goto out;
1497         }
1498
1499         page = pmd_page(pmd);
1500         BUG_ON(is_huge_zero_page(page));
1501         page_nid = page_to_nid(page);
1502         last_cpupid = page_cpupid_last(page);
1503         count_vm_numa_event(NUMA_HINT_FAULTS);
1504         if (page_nid == this_nid) {
1505                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1506                 flags |= TNF_FAULT_LOCAL;
1507         }
1508
1509         /* See similar comment in do_numa_page for explanation */
1510         if (!(vma->vm_flags & VM_WRITE))
1511                 flags |= TNF_NO_GROUP;
1512
1513         /*
1514          * Acquire the page lock to serialise THP migrations but avoid dropping
1515          * page_table_lock if at all possible
1516          */
1517         page_locked = trylock_page(page);
1518         target_nid = mpol_misplaced(page, vma, haddr);
1519         if (target_nid == -1) {
1520                 /* If the page was locked, there are no parallel migrations */
1521                 if (page_locked)
1522                         goto clear_pmdnuma;
1523         }
1524
1525         /* Migration could have started since the pmd_trans_migrating check */
1526         if (!page_locked) {
1527                 spin_unlock(ptl);
1528                 wait_on_page_locked(page);
1529                 page_nid = -1;
1530                 goto out;
1531         }
1532
1533         /*
1534          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1535          * to serialises splits
1536          */
1537         get_page(page);
1538         spin_unlock(ptl);
1539         anon_vma = page_lock_anon_vma_read(page);
1540
1541         /* Confirm the PMD did not change while page_table_lock was released */
1542         spin_lock(ptl);
1543         if (unlikely(!pmd_same(pmd, *pmdp))) {
1544                 unlock_page(page);
1545                 put_page(page);
1546                 page_nid = -1;
1547                 goto out_unlock;
1548         }
1549
1550         /* Bail if we fail to protect against THP splits for any reason */
1551         if (unlikely(!anon_vma)) {
1552                 put_page(page);
1553                 page_nid = -1;
1554                 goto clear_pmdnuma;
1555         }
1556
1557         /*
1558          * Migrate the THP to the requested node, returns with page unlocked
1559          * and access rights restored.
1560          */
1561         spin_unlock(ptl);
1562         migrated = migrate_misplaced_transhuge_page(mm, vma,
1563                                 pmdp, pmd, addr, page, target_nid);
1564         if (migrated) {
1565                 flags |= TNF_MIGRATED;
1566                 page_nid = target_nid;
1567         } else
1568                 flags |= TNF_MIGRATE_FAIL;
1569
1570         goto out;
1571 clear_pmdnuma:
1572         BUG_ON(!PageLocked(page));
1573         was_writable = pmd_write(pmd);
1574         pmd = pmd_modify(pmd, vma->vm_page_prot);
1575         pmd = pmd_mkyoung(pmd);
1576         if (was_writable)
1577                 pmd = pmd_mkwrite(pmd);
1578         set_pmd_at(mm, haddr, pmdp, pmd);
1579         update_mmu_cache_pmd(vma, addr, pmdp);
1580         unlock_page(page);
1581 out_unlock:
1582         spin_unlock(ptl);
1583
1584 out:
1585         if (anon_vma)
1586                 page_unlock_anon_vma_read(anon_vma);
1587
1588         if (page_nid != -1)
1589                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1590
1591         return 0;
1592 }
1593
1594 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1595                 pmd_t *pmd, unsigned long addr, unsigned long next)
1596
1597 {
1598         spinlock_t *ptl;
1599         pmd_t orig_pmd;
1600         struct page *page;
1601         struct mm_struct *mm = tlb->mm;
1602         int ret = 0;
1603
1604         ptl = pmd_trans_huge_lock(pmd, vma);
1605         if (!ptl)
1606                 goto out_unlocked;
1607
1608         orig_pmd = *pmd;
1609         if (is_huge_zero_pmd(orig_pmd)) {
1610                 ret = 1;
1611                 goto out;
1612         }
1613
1614         page = pmd_page(orig_pmd);
1615         /*
1616          * If other processes are mapping this page, we couldn't discard
1617          * the page unless they all do MADV_FREE so let's skip the page.
1618          */
1619         if (page_mapcount(page) != 1)
1620                 goto out;
1621
1622         if (!trylock_page(page))
1623                 goto out;
1624
1625         /*
1626          * If user want to discard part-pages of THP, split it so MADV_FREE
1627          * will deactivate only them.
1628          */
1629         if (next - addr != HPAGE_PMD_SIZE) {
1630                 get_page(page);
1631                 spin_unlock(ptl);
1632                 if (split_huge_page(page)) {
1633                         put_page(page);
1634                         unlock_page(page);
1635                         goto out_unlocked;
1636                 }
1637                 put_page(page);
1638                 unlock_page(page);
1639                 ret = 1;
1640                 goto out_unlocked;
1641         }
1642
1643         if (PageDirty(page))
1644                 ClearPageDirty(page);
1645         unlock_page(page);
1646
1647         if (PageActive(page))
1648                 deactivate_page(page);
1649
1650         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1651                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1652                         tlb->fullmm);
1653                 orig_pmd = pmd_mkold(orig_pmd);
1654                 orig_pmd = pmd_mkclean(orig_pmd);
1655
1656                 set_pmd_at(mm, addr, pmd, orig_pmd);
1657                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1658         }
1659         ret = 1;
1660 out:
1661         spin_unlock(ptl);
1662 out_unlocked:
1663         return ret;
1664 }
1665
1666 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1667                  pmd_t *pmd, unsigned long addr)
1668 {
1669         pmd_t orig_pmd;
1670         spinlock_t *ptl;
1671
1672         ptl = __pmd_trans_huge_lock(pmd, vma);
1673         if (!ptl)
1674                 return 0;
1675         /*
1676          * For architectures like ppc64 we look at deposited pgtable
1677          * when calling pmdp_huge_get_and_clear. So do the
1678          * pgtable_trans_huge_withdraw after finishing pmdp related
1679          * operations.
1680          */
1681         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1682                         tlb->fullmm);
1683         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1684         if (vma_is_dax(vma)) {
1685                 spin_unlock(ptl);
1686                 if (is_huge_zero_pmd(orig_pmd))
1687                         put_huge_zero_page();
1688         } else if (is_huge_zero_pmd(orig_pmd)) {
1689                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1690                 atomic_long_dec(&tlb->mm->nr_ptes);
1691                 spin_unlock(ptl);
1692                 put_huge_zero_page();
1693         } else {
1694                 struct page *page = pmd_page(orig_pmd);
1695                 page_remove_rmap(page, true);
1696                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1697                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1698                 VM_BUG_ON_PAGE(!PageHead(page), page);
1699                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1700                 atomic_long_dec(&tlb->mm->nr_ptes);
1701                 spin_unlock(ptl);
1702                 tlb_remove_page(tlb, page);
1703         }
1704         return 1;
1705 }
1706
1707 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1708                   unsigned long old_addr,
1709                   unsigned long new_addr, unsigned long old_end,
1710                   pmd_t *old_pmd, pmd_t *new_pmd)
1711 {
1712         spinlock_t *old_ptl, *new_ptl;
1713         pmd_t pmd;
1714
1715         struct mm_struct *mm = vma->vm_mm;
1716
1717         if ((old_addr & ~HPAGE_PMD_MASK) ||
1718             (new_addr & ~HPAGE_PMD_MASK) ||
1719             old_end - old_addr < HPAGE_PMD_SIZE ||
1720             (new_vma->vm_flags & VM_NOHUGEPAGE))
1721                 return false;
1722
1723         /*
1724          * The destination pmd shouldn't be established, free_pgtables()
1725          * should have release it.
1726          */
1727         if (WARN_ON(!pmd_none(*new_pmd))) {
1728                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1729                 return false;
1730         }
1731
1732         /*
1733          * We don't have to worry about the ordering of src and dst
1734          * ptlocks because exclusive mmap_sem prevents deadlock.
1735          */
1736         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1737         if (old_ptl) {
1738                 new_ptl = pmd_lockptr(mm, new_pmd);
1739                 if (new_ptl != old_ptl)
1740                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1741                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1742                 VM_BUG_ON(!pmd_none(*new_pmd));
1743
1744                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1745                                 vma_is_anonymous(vma)) {
1746                         pgtable_t pgtable;
1747                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1748                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1749                 }
1750                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1751                 if (new_ptl != old_ptl)
1752                         spin_unlock(new_ptl);
1753                 spin_unlock(old_ptl);
1754                 return true;
1755         }
1756         return false;
1757 }
1758
1759 /*
1760  * Returns
1761  *  - 0 if PMD could not be locked
1762  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1763  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1764  */
1765 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1766                 unsigned long addr, pgprot_t newprot, int prot_numa)
1767 {
1768         struct mm_struct *mm = vma->vm_mm;
1769         spinlock_t *ptl;
1770         int ret = 0;
1771
1772         ptl = __pmd_trans_huge_lock(pmd, vma);
1773         if (ptl) {
1774                 pmd_t entry;
1775                 bool preserve_write = prot_numa && pmd_write(*pmd);
1776                 ret = 1;
1777
1778                 /*
1779                  * Avoid trapping faults against the zero page. The read-only
1780                  * data is likely to be read-cached on the local CPU and
1781                  * local/remote hits to the zero page are not interesting.
1782                  */
1783                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1784                         spin_unlock(ptl);
1785                         return ret;
1786                 }
1787
1788                 if (!prot_numa || !pmd_protnone(*pmd)) {
1789                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1790                         entry = pmd_modify(entry, newprot);
1791                         if (preserve_write)
1792                                 entry = pmd_mkwrite(entry);
1793                         ret = HPAGE_PMD_NR;
1794                         set_pmd_at(mm, addr, pmd, entry);
1795                         BUG_ON(!preserve_write && pmd_write(entry));
1796                 }
1797                 spin_unlock(ptl);
1798         }
1799
1800         return ret;
1801 }
1802
1803 /*
1804  * Returns true if a given pmd maps a thp, false otherwise.
1805  *
1806  * Note that if it returns true, this routine returns without unlocking page
1807  * table lock. So callers must unlock it.
1808  */
1809 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1810 {
1811         spinlock_t *ptl;
1812         ptl = pmd_lock(vma->vm_mm, pmd);
1813         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1814                 return ptl;
1815         spin_unlock(ptl);
1816         return NULL;
1817 }
1818
1819 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1820
1821 int hugepage_madvise(struct vm_area_struct *vma,
1822                      unsigned long *vm_flags, int advice)
1823 {
1824         switch (advice) {
1825         case MADV_HUGEPAGE:
1826 #ifdef CONFIG_S390
1827                 /*
1828                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1829                  * can't handle this properly after s390_enable_sie, so we simply
1830                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1831                  */
1832                 if (mm_has_pgste(vma->vm_mm))
1833                         return 0;
1834 #endif
1835                 /*
1836                  * Be somewhat over-protective like KSM for now!
1837                  */
1838                 if (*vm_flags & VM_NO_THP)
1839                         return -EINVAL;
1840                 *vm_flags &= ~VM_NOHUGEPAGE;
1841                 *vm_flags |= VM_HUGEPAGE;
1842                 /*
1843                  * If the vma become good for khugepaged to scan,
1844                  * register it here without waiting a page fault that
1845                  * may not happen any time soon.
1846                  */
1847                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1848                         return -ENOMEM;
1849                 break;
1850         case MADV_NOHUGEPAGE:
1851                 /*
1852                  * Be somewhat over-protective like KSM for now!
1853                  */
1854                 if (*vm_flags & VM_NO_THP)
1855                         return -EINVAL;
1856                 *vm_flags &= ~VM_HUGEPAGE;
1857                 *vm_flags |= VM_NOHUGEPAGE;
1858                 /*
1859                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1860                  * this vma even if we leave the mm registered in khugepaged if
1861                  * it got registered before VM_NOHUGEPAGE was set.
1862                  */
1863                 break;
1864         }
1865
1866         return 0;
1867 }
1868
1869 static int __init khugepaged_slab_init(void)
1870 {
1871         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1872                                           sizeof(struct mm_slot),
1873                                           __alignof__(struct mm_slot), 0, NULL);
1874         if (!mm_slot_cache)
1875                 return -ENOMEM;
1876
1877         return 0;
1878 }
1879
1880 static void __init khugepaged_slab_exit(void)
1881 {
1882         kmem_cache_destroy(mm_slot_cache);
1883 }
1884
1885 static inline struct mm_slot *alloc_mm_slot(void)
1886 {
1887         if (!mm_slot_cache)     /* initialization failed */
1888                 return NULL;
1889         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1890 }
1891
1892 static inline void free_mm_slot(struct mm_slot *mm_slot)
1893 {
1894         kmem_cache_free(mm_slot_cache, mm_slot);
1895 }
1896
1897 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1898 {
1899         struct mm_slot *mm_slot;
1900
1901         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1902                 if (mm == mm_slot->mm)
1903                         return mm_slot;
1904
1905         return NULL;
1906 }
1907
1908 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1909                                     struct mm_slot *mm_slot)
1910 {
1911         mm_slot->mm = mm;
1912         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1913 }
1914
1915 static inline int khugepaged_test_exit(struct mm_struct *mm)
1916 {
1917         return atomic_read(&mm->mm_users) == 0;
1918 }
1919
1920 int __khugepaged_enter(struct mm_struct *mm)
1921 {
1922         struct mm_slot *mm_slot;
1923         int wakeup;
1924
1925         mm_slot = alloc_mm_slot();
1926         if (!mm_slot)
1927                 return -ENOMEM;
1928
1929         /* __khugepaged_exit() must not run from under us */
1930         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1931         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1932                 free_mm_slot(mm_slot);
1933                 return 0;
1934         }
1935
1936         spin_lock(&khugepaged_mm_lock);
1937         insert_to_mm_slots_hash(mm, mm_slot);
1938         /*
1939          * Insert just behind the scanning cursor, to let the area settle
1940          * down a little.
1941          */
1942         wakeup = list_empty(&khugepaged_scan.mm_head);
1943         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1944         spin_unlock(&khugepaged_mm_lock);
1945
1946         atomic_inc(&mm->mm_count);
1947         if (wakeup)
1948                 wake_up_interruptible(&khugepaged_wait);
1949
1950         return 0;
1951 }
1952
1953 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1954                                unsigned long vm_flags)
1955 {
1956         unsigned long hstart, hend;
1957         if (!vma->anon_vma)
1958                 /*
1959                  * Not yet faulted in so we will register later in the
1960                  * page fault if needed.
1961                  */
1962                 return 0;
1963         if (vma->vm_ops)
1964                 /* khugepaged not yet working on file or special mappings */
1965                 return 0;
1966         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1967         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1968         hend = vma->vm_end & HPAGE_PMD_MASK;
1969         if (hstart < hend)
1970                 return khugepaged_enter(vma, vm_flags);
1971         return 0;
1972 }
1973
1974 void __khugepaged_exit(struct mm_struct *mm)
1975 {
1976         struct mm_slot *mm_slot;
1977         int free = 0;
1978
1979         spin_lock(&khugepaged_mm_lock);
1980         mm_slot = get_mm_slot(mm);
1981         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1982                 hash_del(&mm_slot->hash);
1983                 list_del(&mm_slot->mm_node);
1984                 free = 1;
1985         }
1986         spin_unlock(&khugepaged_mm_lock);
1987
1988         if (free) {
1989                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1990                 free_mm_slot(mm_slot);
1991                 mmdrop(mm);
1992         } else if (mm_slot) {
1993                 /*
1994                  * This is required to serialize against
1995                  * khugepaged_test_exit() (which is guaranteed to run
1996                  * under mmap sem read mode). Stop here (after we
1997                  * return all pagetables will be destroyed) until
1998                  * khugepaged has finished working on the pagetables
1999                  * under the mmap_sem.
2000                  */
2001                 down_write(&mm->mmap_sem);
2002                 up_write(&mm->mmap_sem);
2003         }
2004 }
2005
2006 static void release_pte_page(struct page *page)
2007 {
2008         /* 0 stands for page_is_file_cache(page) == false */
2009         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2010         unlock_page(page);
2011         putback_lru_page(page);
2012 }
2013
2014 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2015 {
2016         while (--_pte >= pte) {
2017                 pte_t pteval = *_pte;
2018                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2019                         release_pte_page(pte_page(pteval));
2020         }
2021 }
2022
2023 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2024                                         unsigned long address,
2025                                         pte_t *pte)
2026 {
2027         struct page *page = NULL;
2028         pte_t *_pte;
2029         int none_or_zero = 0, result = 0;
2030         bool referenced = false, writable = false;
2031
2032         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2033              _pte++, address += PAGE_SIZE) {
2034                 pte_t pteval = *_pte;
2035                 if (pte_none(pteval) || (pte_present(pteval) &&
2036                                 is_zero_pfn(pte_pfn(pteval)))) {
2037                         if (!userfaultfd_armed(vma) &&
2038                             ++none_or_zero <= khugepaged_max_ptes_none) {
2039                                 continue;
2040                         } else {
2041                                 result = SCAN_EXCEED_NONE_PTE;
2042                                 goto out;
2043                         }
2044                 }
2045                 if (!pte_present(pteval)) {
2046                         result = SCAN_PTE_NON_PRESENT;
2047                         goto out;
2048                 }
2049                 page = vm_normal_page(vma, address, pteval);
2050                 if (unlikely(!page)) {
2051                         result = SCAN_PAGE_NULL;
2052                         goto out;
2053                 }
2054
2055                 VM_BUG_ON_PAGE(PageCompound(page), page);
2056                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2057                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2058
2059                 /*
2060                  * We can do it before isolate_lru_page because the
2061                  * page can't be freed from under us. NOTE: PG_lock
2062                  * is needed to serialize against split_huge_page
2063                  * when invoked from the VM.
2064                  */
2065                 if (!trylock_page(page)) {
2066                         result = SCAN_PAGE_LOCK;
2067                         goto out;
2068                 }
2069
2070                 /*
2071                  * cannot use mapcount: can't collapse if there's a gup pin.
2072                  * The page must only be referenced by the scanned process
2073                  * and page swap cache.
2074                  */
2075                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2076                         unlock_page(page);
2077                         result = SCAN_PAGE_COUNT;
2078                         goto out;
2079                 }
2080                 if (pte_write(pteval)) {
2081                         writable = true;
2082                 } else {
2083                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2084                                 unlock_page(page);
2085                                 result = SCAN_SWAP_CACHE_PAGE;
2086                                 goto out;
2087                         }
2088                         /*
2089                          * Page is not in the swap cache. It can be collapsed
2090                          * into a THP.
2091                          */
2092                 }
2093
2094                 /*
2095                  * Isolate the page to avoid collapsing an hugepage
2096                  * currently in use by the VM.
2097                  */
2098                 if (isolate_lru_page(page)) {
2099                         unlock_page(page);
2100                         result = SCAN_DEL_PAGE_LRU;
2101                         goto out;
2102                 }
2103                 /* 0 stands for page_is_file_cache(page) == false */
2104                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2105                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2106                 VM_BUG_ON_PAGE(PageLRU(page), page);
2107
2108                 /* If there is no mapped pte young don't collapse the page */
2109                 if (pte_young(pteval) ||
2110                     page_is_young(page) || PageReferenced(page) ||
2111                     mmu_notifier_test_young(vma->vm_mm, address))
2112                         referenced = true;
2113         }
2114         if (likely(writable)) {
2115                 if (likely(referenced)) {
2116                         result = SCAN_SUCCEED;
2117                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2118                                                             referenced, writable, result);
2119                         return 1;
2120                 }
2121         } else {
2122                 result = SCAN_PAGE_RO;
2123         }
2124
2125 out:
2126         release_pte_pages(pte, _pte);
2127         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2128                                             referenced, writable, result);
2129         return 0;
2130 }
2131
2132 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2133                                       struct vm_area_struct *vma,
2134                                       unsigned long address,
2135                                       spinlock_t *ptl)
2136 {
2137         pte_t *_pte;
2138         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2139                 pte_t pteval = *_pte;
2140                 struct page *src_page;
2141
2142                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2143                         clear_user_highpage(page, address);
2144                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2145                         if (is_zero_pfn(pte_pfn(pteval))) {
2146                                 /*
2147                                  * ptl mostly unnecessary.
2148                                  */
2149                                 spin_lock(ptl);
2150                                 /*
2151                                  * paravirt calls inside pte_clear here are
2152                                  * superfluous.
2153                                  */
2154                                 pte_clear(vma->vm_mm, address, _pte);
2155                                 spin_unlock(ptl);
2156                         }
2157                 } else {
2158                         src_page = pte_page(pteval);
2159                         copy_user_highpage(page, src_page, address, vma);
2160                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2161                         release_pte_page(src_page);
2162                         /*
2163                          * ptl mostly unnecessary, but preempt has to
2164                          * be disabled to update the per-cpu stats
2165                          * inside page_remove_rmap().
2166                          */
2167                         spin_lock(ptl);
2168                         /*
2169                          * paravirt calls inside pte_clear here are
2170                          * superfluous.
2171                          */
2172                         pte_clear(vma->vm_mm, address, _pte);
2173                         page_remove_rmap(src_page, false);
2174                         spin_unlock(ptl);
2175                         free_page_and_swap_cache(src_page);
2176                 }
2177
2178                 address += PAGE_SIZE;
2179                 page++;
2180         }
2181 }
2182
2183 static void khugepaged_alloc_sleep(void)
2184 {
2185         DEFINE_WAIT(wait);
2186
2187         add_wait_queue(&khugepaged_wait, &wait);
2188         freezable_schedule_timeout_interruptible(
2189                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2190         remove_wait_queue(&khugepaged_wait, &wait);
2191 }
2192
2193 static int khugepaged_node_load[MAX_NUMNODES];
2194
2195 static bool khugepaged_scan_abort(int nid)
2196 {
2197         int i;
2198
2199         /*
2200          * If zone_reclaim_mode is disabled, then no extra effort is made to
2201          * allocate memory locally.
2202          */
2203         if (!zone_reclaim_mode)
2204                 return false;
2205
2206         /* If there is a count for this node already, it must be acceptable */
2207         if (khugepaged_node_load[nid])
2208                 return false;
2209
2210         for (i = 0; i < MAX_NUMNODES; i++) {
2211                 if (!khugepaged_node_load[i])
2212                         continue;
2213                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2214                         return true;
2215         }
2216         return false;
2217 }
2218
2219 #ifdef CONFIG_NUMA
2220 static int khugepaged_find_target_node(void)
2221 {
2222         static int last_khugepaged_target_node = NUMA_NO_NODE;
2223         int nid, target_node = 0, max_value = 0;
2224
2225         /* find first node with max normal pages hit */
2226         for (nid = 0; nid < MAX_NUMNODES; nid++)
2227                 if (khugepaged_node_load[nid] > max_value) {
2228                         max_value = khugepaged_node_load[nid];
2229                         target_node = nid;
2230                 }
2231
2232         /* do some balance if several nodes have the same hit record */
2233         if (target_node <= last_khugepaged_target_node)
2234                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2235                                 nid++)
2236                         if (max_value == khugepaged_node_load[nid]) {
2237                                 target_node = nid;
2238                                 break;
2239                         }
2240
2241         last_khugepaged_target_node = target_node;
2242         return target_node;
2243 }
2244
2245 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2246 {
2247         if (IS_ERR(*hpage)) {
2248                 if (!*wait)
2249                         return false;
2250
2251                 *wait = false;
2252                 *hpage = NULL;
2253                 khugepaged_alloc_sleep();
2254         } else if (*hpage) {
2255                 put_page(*hpage);
2256                 *hpage = NULL;
2257         }
2258
2259         return true;
2260 }
2261
2262 static struct page *
2263 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2264                        unsigned long address, int node)
2265 {
2266         VM_BUG_ON_PAGE(*hpage, *hpage);
2267
2268         /*
2269          * Before allocating the hugepage, release the mmap_sem read lock.
2270          * The allocation can take potentially a long time if it involves
2271          * sync compaction, and we do not need to hold the mmap_sem during
2272          * that. We will recheck the vma after taking it again in write mode.
2273          */
2274         up_read(&mm->mmap_sem);
2275
2276         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2277         if (unlikely(!*hpage)) {
2278                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2279                 *hpage = ERR_PTR(-ENOMEM);
2280                 return NULL;
2281         }
2282
2283         prep_transhuge_page(*hpage);
2284         count_vm_event(THP_COLLAPSE_ALLOC);
2285         return *hpage;
2286 }
2287 #else
2288 static int khugepaged_find_target_node(void)
2289 {
2290         return 0;
2291 }
2292
2293 static inline struct page *alloc_khugepaged_hugepage(void)
2294 {
2295         struct page *page;
2296
2297         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2298                            HPAGE_PMD_ORDER);
2299         if (page)
2300                 prep_transhuge_page(page);
2301         return page;
2302 }
2303
2304 static struct page *khugepaged_alloc_hugepage(bool *wait)
2305 {
2306         struct page *hpage;
2307
2308         do {
2309                 hpage = alloc_khugepaged_hugepage();
2310                 if (!hpage) {
2311                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2312                         if (!*wait)
2313                                 return NULL;
2314
2315                         *wait = false;
2316                         khugepaged_alloc_sleep();
2317                 } else
2318                         count_vm_event(THP_COLLAPSE_ALLOC);
2319         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2320
2321         return hpage;
2322 }
2323
2324 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2325 {
2326         if (!*hpage)
2327                 *hpage = khugepaged_alloc_hugepage(wait);
2328
2329         if (unlikely(!*hpage))
2330                 return false;
2331
2332         return true;
2333 }
2334
2335 static struct page *
2336 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2337                        unsigned long address, int node)
2338 {
2339         up_read(&mm->mmap_sem);
2340         VM_BUG_ON(!*hpage);
2341
2342         return  *hpage;
2343 }
2344 #endif
2345
2346 static bool hugepage_vma_check(struct vm_area_struct *vma)
2347 {
2348         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2349             (vma->vm_flags & VM_NOHUGEPAGE))
2350                 return false;
2351         if (!vma->anon_vma || vma->vm_ops)
2352                 return false;
2353         if (is_vma_temporary_stack(vma))
2354                 return false;
2355         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2356         return true;
2357 }
2358
2359 static void collapse_huge_page(struct mm_struct *mm,
2360                                    unsigned long address,
2361                                    struct page **hpage,
2362                                    struct vm_area_struct *vma,
2363                                    int node)
2364 {
2365         pmd_t *pmd, _pmd;
2366         pte_t *pte;
2367         pgtable_t pgtable;
2368         struct page *new_page;
2369         spinlock_t *pmd_ptl, *pte_ptl;
2370         int isolated = 0, result = 0;
2371         unsigned long hstart, hend;
2372         struct mem_cgroup *memcg;
2373         unsigned long mmun_start;       /* For mmu_notifiers */
2374         unsigned long mmun_end;         /* For mmu_notifiers */
2375         gfp_t gfp;
2376
2377         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2378
2379         /* Only allocate from the target node */
2380         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2381
2382         /* release the mmap_sem read lock. */
2383         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2384         if (!new_page) {
2385                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2386                 goto out_nolock;
2387         }
2388
2389         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2390                 result = SCAN_CGROUP_CHARGE_FAIL;
2391                 goto out_nolock;
2392         }
2393
2394         /*
2395          * Prevent all access to pagetables with the exception of
2396          * gup_fast later hanlded by the ptep_clear_flush and the VM
2397          * handled by the anon_vma lock + PG_lock.
2398          */
2399         down_write(&mm->mmap_sem);
2400         if (unlikely(khugepaged_test_exit(mm))) {
2401                 result = SCAN_ANY_PROCESS;
2402                 goto out;
2403         }
2404
2405         vma = find_vma(mm, address);
2406         if (!vma) {
2407                 result = SCAN_VMA_NULL;
2408                 goto out;
2409         }
2410         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2411         hend = vma->vm_end & HPAGE_PMD_MASK;
2412         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2413                 result = SCAN_ADDRESS_RANGE;
2414                 goto out;
2415         }
2416         if (!hugepage_vma_check(vma)) {
2417                 result = SCAN_VMA_CHECK;
2418                 goto out;
2419         }
2420         pmd = mm_find_pmd(mm, address);
2421         if (!pmd) {
2422                 result = SCAN_PMD_NULL;
2423                 goto out;
2424         }
2425
2426         anon_vma_lock_write(vma->anon_vma);
2427
2428         pte = pte_offset_map(pmd, address);
2429         pte_ptl = pte_lockptr(mm, pmd);
2430
2431         mmun_start = address;
2432         mmun_end   = address + HPAGE_PMD_SIZE;
2433         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2434         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2435         /*
2436          * After this gup_fast can't run anymore. This also removes
2437          * any huge TLB entry from the CPU so we won't allow
2438          * huge and small TLB entries for the same virtual address
2439          * to avoid the risk of CPU bugs in that area.
2440          */
2441         _pmd = pmdp_collapse_flush(vma, address, pmd);
2442         spin_unlock(pmd_ptl);
2443         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2444
2445         spin_lock(pte_ptl);
2446         isolated = __collapse_huge_page_isolate(vma, address, pte);
2447         spin_unlock(pte_ptl);
2448
2449         if (unlikely(!isolated)) {
2450                 pte_unmap(pte);
2451                 spin_lock(pmd_ptl);
2452                 BUG_ON(!pmd_none(*pmd));
2453                 /*
2454                  * We can only use set_pmd_at when establishing
2455                  * hugepmds and never for establishing regular pmds that
2456                  * points to regular pagetables. Use pmd_populate for that
2457                  */
2458                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2459                 spin_unlock(pmd_ptl);
2460                 anon_vma_unlock_write(vma->anon_vma);
2461                 result = SCAN_FAIL;
2462                 goto out;
2463         }
2464
2465         /*
2466          * All pages are isolated and locked so anon_vma rmap
2467          * can't run anymore.
2468          */
2469         anon_vma_unlock_write(vma->anon_vma);
2470
2471         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2472         pte_unmap(pte);
2473         __SetPageUptodate(new_page);
2474         pgtable = pmd_pgtable(_pmd);
2475
2476         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2477         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2478
2479         /*
2480          * spin_lock() below is not the equivalent of smp_wmb(), so
2481          * this is needed to avoid the copy_huge_page writes to become
2482          * visible after the set_pmd_at() write.
2483          */
2484         smp_wmb();
2485
2486         spin_lock(pmd_ptl);
2487         BUG_ON(!pmd_none(*pmd));
2488         page_add_new_anon_rmap(new_page, vma, address, true);
2489         mem_cgroup_commit_charge(new_page, memcg, false, true);
2490         lru_cache_add_active_or_unevictable(new_page, vma);
2491         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2492         set_pmd_at(mm, address, pmd, _pmd);
2493         update_mmu_cache_pmd(vma, address, pmd);
2494         spin_unlock(pmd_ptl);
2495
2496         *hpage = NULL;
2497
2498         khugepaged_pages_collapsed++;
2499         result = SCAN_SUCCEED;
2500 out_up_write:
2501         up_write(&mm->mmap_sem);
2502         trace_mm_collapse_huge_page(mm, isolated, result);
2503         return;
2504
2505 out_nolock:
2506         trace_mm_collapse_huge_page(mm, isolated, result);
2507         return;
2508 out:
2509         mem_cgroup_cancel_charge(new_page, memcg, true);
2510         goto out_up_write;
2511 }
2512
2513 static int khugepaged_scan_pmd(struct mm_struct *mm,
2514                                struct vm_area_struct *vma,
2515                                unsigned long address,
2516                                struct page **hpage)
2517 {
2518         pmd_t *pmd;
2519         pte_t *pte, *_pte;
2520         int ret = 0, none_or_zero = 0, result = 0;
2521         struct page *page = NULL;
2522         unsigned long _address;
2523         spinlock_t *ptl;
2524         int node = NUMA_NO_NODE;
2525         bool writable = false, referenced = false;
2526
2527         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2528
2529         pmd = mm_find_pmd(mm, address);
2530         if (!pmd) {
2531                 result = SCAN_PMD_NULL;
2532                 goto out;
2533         }
2534
2535         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2536         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2537         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2538              _pte++, _address += PAGE_SIZE) {
2539                 pte_t pteval = *_pte;
2540                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2541                         if (!userfaultfd_armed(vma) &&
2542                             ++none_or_zero <= khugepaged_max_ptes_none) {
2543                                 continue;
2544                         } else {
2545                                 result = SCAN_EXCEED_NONE_PTE;
2546                                 goto out_unmap;
2547                         }
2548                 }
2549                 if (!pte_present(pteval)) {
2550                         result = SCAN_PTE_NON_PRESENT;
2551                         goto out_unmap;
2552                 }
2553                 if (pte_write(pteval))
2554                         writable = true;
2555
2556                 page = vm_normal_page(vma, _address, pteval);
2557                 if (unlikely(!page)) {
2558                         result = SCAN_PAGE_NULL;
2559                         goto out_unmap;
2560                 }
2561
2562                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2563                 if (PageCompound(page)) {
2564                         result = SCAN_PAGE_COMPOUND;
2565                         goto out_unmap;
2566                 }
2567
2568                 /*
2569                  * Record which node the original page is from and save this
2570                  * information to khugepaged_node_load[].
2571                  * Khupaged will allocate hugepage from the node has the max
2572                  * hit record.
2573                  */
2574                 node = page_to_nid(page);
2575                 if (khugepaged_scan_abort(node)) {
2576                         result = SCAN_SCAN_ABORT;
2577                         goto out_unmap;
2578                 }
2579                 khugepaged_node_load[node]++;
2580                 if (!PageLRU(page)) {
2581                         result = SCAN_PAGE_LRU;
2582                         goto out_unmap;
2583                 }
2584                 if (PageLocked(page)) {
2585                         result = SCAN_PAGE_LOCK;
2586                         goto out_unmap;
2587                 }
2588                 if (!PageAnon(page)) {
2589                         result = SCAN_PAGE_ANON;
2590                         goto out_unmap;
2591                 }
2592
2593                 /*
2594                  * cannot use mapcount: can't collapse if there's a gup pin.
2595                  * The page must only be referenced by the scanned process
2596                  * and page swap cache.
2597                  */
2598                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2599                         result = SCAN_PAGE_COUNT;
2600                         goto out_unmap;
2601                 }
2602                 if (pte_young(pteval) ||
2603                     page_is_young(page) || PageReferenced(page) ||
2604                     mmu_notifier_test_young(vma->vm_mm, address))
2605                         referenced = true;
2606         }
2607         if (writable) {
2608                 if (referenced) {
2609                         result = SCAN_SUCCEED;
2610                         ret = 1;
2611                 } else {
2612                         result = SCAN_NO_REFERENCED_PAGE;
2613                 }
2614         } else {
2615                 result = SCAN_PAGE_RO;
2616         }
2617 out_unmap:
2618         pte_unmap_unlock(pte, ptl);
2619         if (ret) {
2620                 node = khugepaged_find_target_node();
2621                 /* collapse_huge_page will return with the mmap_sem released */
2622                 collapse_huge_page(mm, address, hpage, vma, node);
2623         }
2624 out:
2625         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2626                                      none_or_zero, result);
2627         return ret;
2628 }
2629
2630 static void collect_mm_slot(struct mm_slot *mm_slot)
2631 {
2632         struct mm_struct *mm = mm_slot->mm;
2633
2634         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2635
2636         if (khugepaged_test_exit(mm)) {
2637                 /* free mm_slot */
2638                 hash_del(&mm_slot->hash);
2639                 list_del(&mm_slot->mm_node);
2640
2641                 /*
2642                  * Not strictly needed because the mm exited already.
2643                  *
2644                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2645                  */
2646
2647                 /* khugepaged_mm_lock actually not necessary for the below */
2648                 free_mm_slot(mm_slot);
2649                 mmdrop(mm);
2650         }
2651 }
2652
2653 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2654                                             struct page **hpage)
2655         __releases(&khugepaged_mm_lock)
2656         __acquires(&khugepaged_mm_lock)
2657 {
2658         struct mm_slot *mm_slot;
2659         struct mm_struct *mm;
2660         struct vm_area_struct *vma;
2661         int progress = 0;
2662
2663         VM_BUG_ON(!pages);
2664         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2665
2666         if (khugepaged_scan.mm_slot)
2667                 mm_slot = khugepaged_scan.mm_slot;
2668         else {
2669                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2670                                      struct mm_slot, mm_node);
2671                 khugepaged_scan.address = 0;
2672                 khugepaged_scan.mm_slot = mm_slot;
2673         }
2674         spin_unlock(&khugepaged_mm_lock);
2675
2676         mm = mm_slot->mm;
2677         down_read(&mm->mmap_sem);
2678         if (unlikely(khugepaged_test_exit(mm)))
2679                 vma = NULL;
2680         else
2681                 vma = find_vma(mm, khugepaged_scan.address);
2682
2683         progress++;
2684         for (; vma; vma = vma->vm_next) {
2685                 unsigned long hstart, hend;
2686
2687                 cond_resched();
2688                 if (unlikely(khugepaged_test_exit(mm))) {
2689                         progress++;
2690                         break;
2691                 }
2692                 if (!hugepage_vma_check(vma)) {
2693 skip:
2694                         progress++;
2695                         continue;
2696                 }
2697                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2698                 hend = vma->vm_end & HPAGE_PMD_MASK;
2699                 if (hstart >= hend)
2700                         goto skip;
2701                 if (khugepaged_scan.address > hend)
2702                         goto skip;
2703                 if (khugepaged_scan.address < hstart)
2704                         khugepaged_scan.address = hstart;
2705                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2706
2707                 while (khugepaged_scan.address < hend) {
2708                         int ret;
2709                         cond_resched();
2710                         if (unlikely(khugepaged_test_exit(mm)))
2711                                 goto breakouterloop;
2712
2713                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2714                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2715                                   hend);
2716                         ret = khugepaged_scan_pmd(mm, vma,
2717                                                   khugepaged_scan.address,
2718                                                   hpage);
2719                         /* move to next address */
2720                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2721                         progress += HPAGE_PMD_NR;
2722                         if (ret)
2723                                 /* we released mmap_sem so break loop */
2724                                 goto breakouterloop_mmap_sem;
2725                         if (progress >= pages)
2726                                 goto breakouterloop;
2727                 }
2728         }
2729 breakouterloop:
2730         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2731 breakouterloop_mmap_sem:
2732
2733         spin_lock(&khugepaged_mm_lock);
2734         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2735         /*
2736          * Release the current mm_slot if this mm is about to die, or
2737          * if we scanned all vmas of this mm.
2738          */
2739         if (khugepaged_test_exit(mm) || !vma) {
2740                 /*
2741                  * Make sure that if mm_users is reaching zero while
2742                  * khugepaged runs here, khugepaged_exit will find
2743                  * mm_slot not pointing to the exiting mm.
2744                  */
2745                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2746                         khugepaged_scan.mm_slot = list_entry(
2747                                 mm_slot->mm_node.next,
2748                                 struct mm_slot, mm_node);
2749                         khugepaged_scan.address = 0;
2750                 } else {
2751                         khugepaged_scan.mm_slot = NULL;
2752                         khugepaged_full_scans++;
2753                 }
2754
2755                 collect_mm_slot(mm_slot);
2756         }
2757
2758         return progress;
2759 }
2760
2761 static int khugepaged_has_work(void)
2762 {
2763         return !list_empty(&khugepaged_scan.mm_head) &&
2764                 khugepaged_enabled();
2765 }
2766
2767 static int khugepaged_wait_event(void)
2768 {
2769         return !list_empty(&khugepaged_scan.mm_head) ||
2770                 kthread_should_stop();
2771 }
2772
2773 static void khugepaged_do_scan(void)
2774 {
2775         struct page *hpage = NULL;
2776         unsigned int progress = 0, pass_through_head = 0;
2777         unsigned int pages = khugepaged_pages_to_scan;
2778         bool wait = true;
2779
2780         barrier(); /* write khugepaged_pages_to_scan to local stack */
2781
2782         while (progress < pages) {
2783                 if (!khugepaged_prealloc_page(&hpage, &wait))
2784                         break;
2785
2786                 cond_resched();
2787
2788                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2789                         break;
2790
2791                 spin_lock(&khugepaged_mm_lock);
2792                 if (!khugepaged_scan.mm_slot)
2793                         pass_through_head++;
2794                 if (khugepaged_has_work() &&
2795                     pass_through_head < 2)
2796                         progress += khugepaged_scan_mm_slot(pages - progress,
2797                                                             &hpage);
2798                 else
2799                         progress = pages;
2800                 spin_unlock(&khugepaged_mm_lock);
2801         }
2802
2803         if (!IS_ERR_OR_NULL(hpage))
2804                 put_page(hpage);
2805 }
2806
2807 static void khugepaged_wait_work(void)
2808 {
2809         if (khugepaged_has_work()) {
2810                 if (!khugepaged_scan_sleep_millisecs)
2811                         return;
2812
2813                 wait_event_freezable_timeout(khugepaged_wait,
2814                                              kthread_should_stop(),
2815                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2816                 return;
2817         }
2818
2819         if (khugepaged_enabled())
2820                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2821 }
2822
2823 static int khugepaged(void *none)
2824 {
2825         struct mm_slot *mm_slot;
2826
2827         set_freezable();
2828         set_user_nice(current, MAX_NICE);
2829
2830         while (!kthread_should_stop()) {
2831                 khugepaged_do_scan();
2832                 khugepaged_wait_work();
2833         }
2834
2835         spin_lock(&khugepaged_mm_lock);
2836         mm_slot = khugepaged_scan.mm_slot;
2837         khugepaged_scan.mm_slot = NULL;
2838         if (mm_slot)
2839                 collect_mm_slot(mm_slot);
2840         spin_unlock(&khugepaged_mm_lock);
2841         return 0;
2842 }
2843
2844 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2845                 unsigned long haddr, pmd_t *pmd)
2846 {
2847         struct mm_struct *mm = vma->vm_mm;
2848         pgtable_t pgtable;
2849         pmd_t _pmd;
2850         int i;
2851
2852         /* leave pmd empty until pte is filled */
2853         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2854
2855         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2856         pmd_populate(mm, &_pmd, pgtable);
2857
2858         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2859                 pte_t *pte, entry;
2860                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2861                 entry = pte_mkspecial(entry);
2862                 pte = pte_offset_map(&_pmd, haddr);
2863                 VM_BUG_ON(!pte_none(*pte));
2864                 set_pte_at(mm, haddr, pte, entry);
2865                 pte_unmap(pte);
2866         }
2867         smp_wmb(); /* make pte visible before pmd */
2868         pmd_populate(mm, pmd, pgtable);
2869         put_huge_zero_page();
2870 }
2871
2872 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2873                 unsigned long haddr, bool freeze)
2874 {
2875         struct mm_struct *mm = vma->vm_mm;
2876         struct page *page;
2877         pgtable_t pgtable;
2878         pmd_t _pmd;
2879         bool young, write, dirty;
2880         unsigned long addr;
2881         int i;
2882
2883         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2884         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2885         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2886         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2887
2888         count_vm_event(THP_SPLIT_PMD);
2889
2890         if (vma_is_dax(vma)) {
2891                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2892                 if (is_huge_zero_pmd(_pmd))
2893                         put_huge_zero_page();
2894                 return;
2895         } else if (is_huge_zero_pmd(*pmd)) {
2896                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2897         }
2898
2899         page = pmd_page(*pmd);
2900         VM_BUG_ON_PAGE(!page_count(page), page);
2901         page_ref_add(page, HPAGE_PMD_NR - 1);
2902         write = pmd_write(*pmd);
2903         young = pmd_young(*pmd);
2904         dirty = pmd_dirty(*pmd);
2905
2906         pmdp_huge_split_prepare(vma, haddr, pmd);
2907         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2908         pmd_populate(mm, &_pmd, pgtable);
2909
2910         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2911                 pte_t entry, *pte;
2912                 /*
2913                  * Note that NUMA hinting access restrictions are not
2914                  * transferred to avoid any possibility of altering
2915                  * permissions across VMAs.
2916                  */
2917                 if (freeze) {
2918                         swp_entry_t swp_entry;
2919                         swp_entry = make_migration_entry(page + i, write);
2920                         entry = swp_entry_to_pte(swp_entry);
2921                 } else {
2922                         entry = mk_pte(page + i, vma->vm_page_prot);
2923                         entry = maybe_mkwrite(entry, vma);
2924                         if (!write)
2925                                 entry = pte_wrprotect(entry);
2926                         if (!young)
2927                                 entry = pte_mkold(entry);
2928                 }
2929                 if (dirty)
2930                         SetPageDirty(page + i);
2931                 pte = pte_offset_map(&_pmd, addr);
2932                 BUG_ON(!pte_none(*pte));
2933                 set_pte_at(mm, addr, pte, entry);
2934                 atomic_inc(&page[i]._mapcount);
2935                 pte_unmap(pte);
2936         }
2937
2938         /*
2939          * Set PG_double_map before dropping compound_mapcount to avoid
2940          * false-negative page_mapped().
2941          */
2942         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2943                 for (i = 0; i < HPAGE_PMD_NR; i++)
2944                         atomic_inc(&page[i]._mapcount);
2945         }
2946
2947         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2948                 /* Last compound_mapcount is gone. */
2949                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2950                 if (TestClearPageDoubleMap(page)) {
2951                         /* No need in mapcount reference anymore */
2952                         for (i = 0; i < HPAGE_PMD_NR; i++)
2953                                 atomic_dec(&page[i]._mapcount);
2954                 }
2955         }
2956
2957         smp_wmb(); /* make pte visible before pmd */
2958         /*
2959          * Up to this point the pmd is present and huge and userland has the
2960          * whole access to the hugepage during the split (which happens in
2961          * place). If we overwrite the pmd with the not-huge version pointing
2962          * to the pte here (which of course we could if all CPUs were bug
2963          * free), userland could trigger a small page size TLB miss on the
2964          * small sized TLB while the hugepage TLB entry is still established in
2965          * the huge TLB. Some CPU doesn't like that.
2966          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2967          * 383 on page 93. Intel should be safe but is also warns that it's
2968          * only safe if the permission and cache attributes of the two entries
2969          * loaded in the two TLB is identical (which should be the case here).
2970          * But it is generally safer to never allow small and huge TLB entries
2971          * for the same virtual address to be loaded simultaneously. So instead
2972          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2973          * current pmd notpresent (atomically because here the pmd_trans_huge
2974          * and pmd_trans_splitting must remain set at all times on the pmd
2975          * until the split is complete for this pmd), then we flush the SMP TLB
2976          * and finally we write the non-huge version of the pmd entry with
2977          * pmd_populate.
2978          */
2979         pmdp_invalidate(vma, haddr, pmd);
2980         pmd_populate(mm, pmd, pgtable);
2981
2982         if (freeze) {
2983                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2984                         page_remove_rmap(page + i, false);
2985                         put_page(page + i);
2986                 }
2987         }
2988 }
2989
2990 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2991                 unsigned long address, bool freeze)
2992 {
2993         spinlock_t *ptl;
2994         struct mm_struct *mm = vma->vm_mm;
2995         unsigned long haddr = address & HPAGE_PMD_MASK;
2996
2997         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2998         ptl = pmd_lock(mm, pmd);
2999         if (pmd_trans_huge(*pmd)) {
3000                 struct page *page = pmd_page(*pmd);
3001                 if (PageMlocked(page))
3002                         clear_page_mlock(page);
3003         } else if (!pmd_devmap(*pmd))
3004                 goto out;
3005         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
3006 out:
3007         spin_unlock(ptl);
3008         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3009 }
3010
3011 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3012                 bool freeze, struct page *page)
3013 {
3014         pgd_t *pgd;
3015         pud_t *pud;
3016         pmd_t *pmd;
3017
3018         pgd = pgd_offset(vma->vm_mm, address);
3019         if (!pgd_present(*pgd))
3020                 return;
3021
3022         pud = pud_offset(pgd, address);
3023         if (!pud_present(*pud))
3024                 return;
3025
3026         pmd = pmd_offset(pud, address);
3027         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
3028                 return;
3029
3030         /*
3031          * If caller asks to setup a migration entries, we need a page to check
3032          * pmd against. Otherwise we can end up replacing wrong page.
3033          */
3034         VM_BUG_ON(freeze && !page);
3035         if (page && page != pmd_page(*pmd))
3036                 return;
3037
3038         /*
3039          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3040          * materialize from under us.
3041          */
3042         __split_huge_pmd(vma, pmd, address, freeze);
3043 }
3044
3045 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3046                              unsigned long start,
3047                              unsigned long end,
3048                              long adjust_next)
3049 {
3050         /*
3051          * If the new start address isn't hpage aligned and it could
3052          * previously contain an hugepage: check if we need to split
3053          * an huge pmd.
3054          */
3055         if (start & ~HPAGE_PMD_MASK &&
3056             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3057             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3058                 split_huge_pmd_address(vma, start, false, NULL);
3059
3060         /*
3061          * If the new end address isn't hpage aligned and it could
3062          * previously contain an hugepage: check if we need to split
3063          * an huge pmd.
3064          */
3065         if (end & ~HPAGE_PMD_MASK &&
3066             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3067             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3068                 split_huge_pmd_address(vma, end, false, NULL);
3069
3070         /*
3071          * If we're also updating the vma->vm_next->vm_start, if the new
3072          * vm_next->vm_start isn't page aligned and it could previously
3073          * contain an hugepage: check if we need to split an huge pmd.
3074          */
3075         if (adjust_next > 0) {
3076                 struct vm_area_struct *next = vma->vm_next;
3077                 unsigned long nstart = next->vm_start;
3078                 nstart += adjust_next << PAGE_SHIFT;
3079                 if (nstart & ~HPAGE_PMD_MASK &&
3080                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3081                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3082                         split_huge_pmd_address(next, nstart, false, NULL);
3083         }
3084 }
3085
3086 static void freeze_page(struct page *page)
3087 {
3088         enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3089                 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3090         int i, ret;
3091
3092         VM_BUG_ON_PAGE(!PageHead(page), page);
3093
3094         /* We only need TTU_SPLIT_HUGE_PMD once */
3095         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3096         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3097                 /* Cut short if the page is unmapped */
3098                 if (page_count(page) == 1)
3099                         return;
3100
3101                 ret = try_to_unmap(page + i, ttu_flags);
3102         }
3103         VM_BUG_ON(ret);
3104 }
3105
3106 static void unfreeze_page(struct page *page)
3107 {
3108         int i;
3109
3110         for (i = 0; i < HPAGE_PMD_NR; i++)
3111                 remove_migration_ptes(page + i, page + i, true);
3112 }
3113
3114 static void __split_huge_page_tail(struct page *head, int tail,
3115                 struct lruvec *lruvec, struct list_head *list)
3116 {
3117         struct page *page_tail = head + tail;
3118
3119         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3120         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3121
3122         /*
3123          * tail_page->_count is zero and not changing from under us. But
3124          * get_page_unless_zero() may be running from under us on the
3125          * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3126          * would then run atomic_set() concurrently with
3127          * get_page_unless_zero(), and atomic_set() is implemented in C not
3128          * using locked ops. spin_unlock on x86 sometime uses locked ops
3129          * because of PPro errata 66, 92, so unless somebody can guarantee
3130          * atomic_set() here would be safe on all archs (and not only on x86),
3131          * it's safer to use atomic_inc().
3132          */
3133         page_ref_inc(page_tail);
3134
3135         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3136         page_tail->flags |= (head->flags &
3137                         ((1L << PG_referenced) |
3138                          (1L << PG_swapbacked) |
3139                          (1L << PG_mlocked) |
3140                          (1L << PG_uptodate) |
3141                          (1L << PG_active) |
3142                          (1L << PG_locked) |
3143                          (1L << PG_unevictable) |
3144                          (1L << PG_dirty)));
3145
3146         /*
3147          * After clearing PageTail the gup refcount can be released.
3148          * Page flags also must be visible before we make the page non-compound.
3149          */
3150         smp_wmb();
3151
3152         clear_compound_head(page_tail);
3153
3154         if (page_is_young(head))
3155                 set_page_young(page_tail);
3156         if (page_is_idle(head))
3157                 set_page_idle(page_tail);
3158
3159         /* ->mapping in first tail page is compound_mapcount */
3160         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3161                         page_tail);
3162         page_tail->mapping = head->mapping;
3163
3164         page_tail->index = head->index + tail;
3165         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3166         lru_add_page_tail(head, page_tail, lruvec, list);
3167 }
3168
3169 static void __split_huge_page(struct page *page, struct list_head *list)
3170 {
3171         struct page *head = compound_head(page);
3172         struct zone *zone = page_zone(head);
3173         struct lruvec *lruvec;
3174         int i;
3175
3176         /* prevent PageLRU to go away from under us, and freeze lru stats */
3177         spin_lock_irq(&zone->lru_lock);
3178         lruvec = mem_cgroup_page_lruvec(head, zone);
3179
3180         /* complete memcg works before add pages to LRU */
3181         mem_cgroup_split_huge_fixup(head);
3182
3183         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3184                 __split_huge_page_tail(head, i, lruvec, list);
3185
3186         ClearPageCompound(head);
3187         spin_unlock_irq(&zone->lru_lock);
3188
3189         unfreeze_page(head);
3190
3191         for (i = 0; i < HPAGE_PMD_NR; i++) {
3192                 struct page *subpage = head + i;
3193                 if (subpage == page)
3194                         continue;
3195                 unlock_page(subpage);
3196
3197                 /*
3198                  * Subpages may be freed if there wasn't any mapping
3199                  * like if add_to_swap() is running on a lru page that
3200                  * had its mapping zapped. And freeing these pages
3201                  * requires taking the lru_lock so we do the put_page
3202                  * of the tail pages after the split is complete.
3203                  */
3204                 put_page(subpage);
3205         }
3206 }
3207
3208 int total_mapcount(struct page *page)
3209 {
3210         int i, ret;
3211
3212         VM_BUG_ON_PAGE(PageTail(page), page);
3213
3214         if (likely(!PageCompound(page)))
3215                 return atomic_read(&page->_mapcount) + 1;
3216
3217         ret = compound_mapcount(page);
3218         if (PageHuge(page))
3219                 return ret;
3220         for (i = 0; i < HPAGE_PMD_NR; i++)
3221                 ret += atomic_read(&page[i]._mapcount) + 1;
3222         if (PageDoubleMap(page))
3223                 ret -= HPAGE_PMD_NR;
3224         return ret;
3225 }
3226
3227 /*
3228  * This function splits huge page into normal pages. @page can point to any
3229  * subpage of huge page to split. Split doesn't change the position of @page.
3230  *
3231  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3232  * The huge page must be locked.
3233  *
3234  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3235  *
3236  * Both head page and tail pages will inherit mapping, flags, and so on from
3237  * the hugepage.
3238  *
3239  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3240  * they are not mapped.
3241  *
3242  * Returns 0 if the hugepage is split successfully.
3243  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3244  * us.
3245  */
3246 int split_huge_page_to_list(struct page *page, struct list_head *list)
3247 {
3248         struct page *head = compound_head(page);
3249         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3250         struct anon_vma *anon_vma;
3251         int count, mapcount, ret;
3252         bool mlocked;
3253         unsigned long flags;
3254
3255         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3256         VM_BUG_ON_PAGE(!PageAnon(page), page);
3257         VM_BUG_ON_PAGE(!PageLocked(page), page);
3258         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3259         VM_BUG_ON_PAGE(!PageCompound(page), page);
3260
3261         /*
3262          * The caller does not necessarily hold an mmap_sem that would prevent
3263          * the anon_vma disappearing so we first we take a reference to it
3264          * and then lock the anon_vma for write. This is similar to
3265          * page_lock_anon_vma_read except the write lock is taken to serialise
3266          * against parallel split or collapse operations.
3267          */
3268         anon_vma = page_get_anon_vma(head);
3269         if (!anon_vma) {
3270                 ret = -EBUSY;
3271                 goto out;
3272         }
3273         anon_vma_lock_write(anon_vma);
3274
3275         /*
3276          * Racy check if we can split the page, before freeze_page() will
3277          * split PMDs
3278          */
3279         if (total_mapcount(head) != page_count(head) - 1) {
3280                 ret = -EBUSY;
3281                 goto out_unlock;
3282         }
3283
3284         mlocked = PageMlocked(page);
3285         freeze_page(head);
3286         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3287
3288         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3289         if (mlocked)
3290                 lru_add_drain();
3291
3292         /* Prevent deferred_split_scan() touching ->_count */
3293         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3294         count = page_count(head);
3295         mapcount = total_mapcount(head);
3296         if (!mapcount && count == 1) {
3297                 if (!list_empty(page_deferred_list(head))) {
3298                         pgdata->split_queue_len--;
3299                         list_del(page_deferred_list(head));
3300                 }
3301                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3302                 __split_huge_page(page, list);
3303                 ret = 0;
3304         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3305                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3306                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3307                                 mapcount, count);
3308                 if (PageTail(page))
3309                         dump_page(head, NULL);
3310                 dump_page(page, "total_mapcount(head) > 0");
3311                 BUG();
3312         } else {
3313                 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3314                 unfreeze_page(head);
3315                 ret = -EBUSY;
3316         }
3317
3318 out_unlock:
3319         anon_vma_unlock_write(anon_vma);
3320         put_anon_vma(anon_vma);
3321 out:
3322         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3323         return ret;
3324 }
3325
3326 void free_transhuge_page(struct page *page)
3327 {
3328         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3329         unsigned long flags;
3330
3331         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3332         if (!list_empty(page_deferred_list(page))) {
3333                 pgdata->split_queue_len--;
3334                 list_del(page_deferred_list(page));
3335         }
3336         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3337         free_compound_page(page);
3338 }
3339
3340 void deferred_split_huge_page(struct page *page)
3341 {
3342         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3343         unsigned long flags;
3344
3345         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3346
3347         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3348         if (list_empty(page_deferred_list(page))) {
3349                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3350                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3351                 pgdata->split_queue_len++;
3352         }
3353         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3354 }
3355
3356 static unsigned long deferred_split_count(struct shrinker *shrink,
3357                 struct shrink_control *sc)
3358 {
3359         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3360         return ACCESS_ONCE(pgdata->split_queue_len);
3361 }
3362
3363 static unsigned long deferred_split_scan(struct shrinker *shrink,
3364                 struct shrink_control *sc)
3365 {
3366         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3367         unsigned long flags;
3368         LIST_HEAD(list), *pos, *next;
3369         struct page *page;
3370         int split = 0;
3371
3372         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3373         /* Take pin on all head pages to avoid freeing them under us */
3374         list_for_each_safe(pos, next, &pgdata->split_queue) {
3375                 page = list_entry((void *)pos, struct page, mapping);
3376                 page = compound_head(page);
3377                 if (get_page_unless_zero(page)) {
3378                         list_move(page_deferred_list(page), &list);
3379                 } else {
3380                         /* We lost race with put_compound_page() */
3381                         list_del_init(page_deferred_list(page));
3382                         pgdata->split_queue_len--;
3383                 }
3384                 if (!--sc->nr_to_scan)
3385                         break;
3386         }
3387         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3388
3389         list_for_each_safe(pos, next, &list) {
3390                 page = list_entry((void *)pos, struct page, mapping);
3391                 lock_page(page);
3392                 /* split_huge_page() removes page from list on success */
3393                 if (!split_huge_page(page))
3394                         split++;
3395                 unlock_page(page);
3396                 put_page(page);
3397         }
3398
3399         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3400         list_splice_tail(&list, &pgdata->split_queue);
3401         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3402
3403         /*
3404          * Stop shrinker if we didn't split any page, but the queue is empty.
3405          * This can happen if pages were freed under us.
3406          */
3407         if (!split && list_empty(&pgdata->split_queue))
3408                 return SHRINK_STOP;
3409         return split;
3410 }
3411
3412 static struct shrinker deferred_split_shrinker = {
3413         .count_objects = deferred_split_count,
3414         .scan_objects = deferred_split_scan,
3415         .seeks = DEFAULT_SEEKS,
3416         .flags = SHRINKER_NUMA_AWARE,
3417 };
3418
3419 #ifdef CONFIG_DEBUG_FS
3420 static int split_huge_pages_set(void *data, u64 val)
3421 {
3422         struct zone *zone;
3423         struct page *page;
3424         unsigned long pfn, max_zone_pfn;
3425         unsigned long total = 0, split = 0;
3426
3427         if (val != 1)
3428                 return -EINVAL;
3429
3430         for_each_populated_zone(zone) {
3431                 max_zone_pfn = zone_end_pfn(zone);
3432                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3433                         if (!pfn_valid(pfn))
3434                                 continue;
3435
3436                         page = pfn_to_page(pfn);
3437                         if (!get_page_unless_zero(page))
3438                                 continue;
3439
3440                         if (zone != page_zone(page))
3441                                 goto next;
3442
3443                         if (!PageHead(page) || !PageAnon(page) ||
3444                                         PageHuge(page))
3445                                 goto next;
3446
3447                         total++;
3448                         lock_page(page);
3449                         if (!split_huge_page(page))
3450                                 split++;
3451                         unlock_page(page);
3452 next:
3453                         put_page(page);
3454                 }
3455         }
3456
3457         pr_info("%lu of %lu THP split", split, total);
3458
3459         return 0;
3460 }
3461 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3462                 "%llu\n");
3463
3464 static int __init split_huge_pages_debugfs(void)
3465 {
3466         void *ret;
3467
3468         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3469                         &split_huge_pages_fops);
3470         if (!ret)
3471                 pr_warn("Failed to create split_huge_pages in debugfs");
3472         return 0;
3473 }
3474 late_initcall(split_huge_pages_debugfs);
3475 #endif