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