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