mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/dax.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 #include <linux/shmem_fs.h>
  34 #include <linux/oom.h>
  35 #include <linux/numa.h>
  36 #include <linux/page_owner.h>
  37
  38 #include <asm/tlb.h>
  39 #include <asm/pgalloc.h>
  40 #include "internal.h"
  41
  42 /*
  43  * By default, transparent hugepage support is disabled in order to avoid
  44  * risking an increased memory footprint for applications that are not
  45  * guaranteed to benefit from it. When transparent hugepage support is
  46  * enabled, it is for all mappings, and khugepaged scans all mappings.
  47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  48  * for all hugepage allocations.
  49  */
  50 unsigned long transparent_hugepage_flags __read_mostly =
  51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  52         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  53 #endif
  54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  55         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  56 #endif
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  59         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  60
  61 static struct shrinker deferred_split_shrinker;
  62
  63 static atomic_t huge_zero_refcount;
  64 struct page *huge_zero_page __read_mostly;
  65 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  66
  67 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  68 {
  69         /* The addr is used to check if the vma size fits */
  70         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  71
  72         if (!transhuge_vma_suitable(vma, addr))
  73                 return false;
  74         if (vma_is_anonymous(vma))
  75                 return __transparent_hugepage_enabled(vma);
  76         if (vma_is_shmem(vma))
  77                 return shmem_huge_enabled(vma);
  78
  79         return false;
  80 }
  81
  82 static bool get_huge_zero_page(void)
  83 {
  84         struct page *zero_page;
  85 retry:
  86         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  87                 return true;
  88
  89         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  90                         HPAGE_PMD_ORDER);
  91         if (!zero_page) {
  92                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  93                 return false;
  94         }
  95         count_vm_event(THP_ZERO_PAGE_ALLOC);
  96         preempt_disable();
  97         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  98                 preempt_enable();
  99                 __free_pages(zero_page, compound_order(zero_page));
 100                 goto retry;
 101         }
 102         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 103
 104         /* We take additional reference here. It will be put back by shrinker */
 105         atomic_set(&huge_zero_refcount, 2);
 106         preempt_enable();
 107         return true;
 108 }
 109
 110 static void put_huge_zero_page(void)
 111 {
 112         /*
 113          * Counter should never go to zero here. Only shrinker can put
 114          * last reference.
 115          */
 116         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 117 }
 118
 119 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 120 {
 121         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 122                 return READ_ONCE(huge_zero_page);
 123
 124         if (!get_huge_zero_page())
 125                 return NULL;
 126
 127         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 128                 put_huge_zero_page();
 129
 130         return READ_ONCE(huge_zero_page);
 131 }
 132
 133 void mm_put_huge_zero_page(struct mm_struct *mm)
 134 {
 135         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 136                 put_huge_zero_page();
 137 }
 138
 139 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 140                                         struct shrink_control *sc)
 141 {
 142         /* we can free zero page only if last reference remains */
 143         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 144 }
 145
 146 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 147                                        struct shrink_control *sc)
 148 {
 149         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 150                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 151                 BUG_ON(zero_page == NULL);
 152                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 153                 __free_pages(zero_page, compound_order(zero_page));
 154                 return HPAGE_PMD_NR;
 155         }
 156
 157         return 0;
 158 }
 159
 160 static struct shrinker huge_zero_page_shrinker = {
 161         .count_objects = shrink_huge_zero_page_count,
 162         .scan_objects = shrink_huge_zero_page_scan,
 163         .seeks = DEFAULT_SEEKS,
 164 };
 165
 166 #ifdef CONFIG_SYSFS
 167 static ssize_t enabled_show(struct kobject *kobj,
 168                             struct kobj_attribute *attr, char *buf)
 169 {
 170         const char *output;
 171
 172         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 173                 output = "[always] madvise never";
 174         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 175                           &transparent_hugepage_flags))
 176                 output = "always [madvise] never";
 177         else
 178                 output = "always madvise [never]";
 179
 180         return sysfs_emit(buf, "%s\n", output);
 181 }
 182
 183 static ssize_t enabled_store(struct kobject *kobj,
 184                              struct kobj_attribute *attr,
 185                              const char *buf, size_t count)
 186 {
 187         ssize_t ret = count;
 188
 189         if (sysfs_streq(buf, "always")) {
 190                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 191                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 192         } else if (sysfs_streq(buf, "madvise")) {
 193                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 194                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 195         } else if (sysfs_streq(buf, "never")) {
 196                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 197                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 198         } else
 199                 ret = -EINVAL;
 200
 201         if (ret > 0) {
 202                 int err = start_stop_khugepaged();
 203                 if (err)
 204                         ret = err;
 205         }
 206         return ret;
 207 }
 208 static struct kobj_attribute enabled_attr =
 209         __ATTR(enabled, 0644, enabled_show, enabled_store);
 210
 211 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 212                                   struct kobj_attribute *attr, char *buf,
 213                                   enum transparent_hugepage_flag flag)
 214 {
 215         return sysfs_emit(buf, "%d\n",
 216                           !!test_bit(flag, &transparent_hugepage_flags));
 217 }
 218
 219 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 220                                  struct kobj_attribute *attr,
 221                                  const char *buf, size_t count,
 222                                  enum transparent_hugepage_flag flag)
 223 {
 224         unsigned long value;
 225         int ret;
 226
 227         ret = kstrtoul(buf, 10, &value);
 228         if (ret < 0)
 229                 return ret;
 230         if (value > 1)
 231                 return -EINVAL;
 232
 233         if (value)
 234                 set_bit(flag, &transparent_hugepage_flags);
 235         else
 236                 clear_bit(flag, &transparent_hugepage_flags);
 237
 238         return count;
 239 }
 240
 241 static ssize_t defrag_show(struct kobject *kobj,
 242                            struct kobj_attribute *attr, char *buf)
 243 {
 244         const char *output;
 245
 246         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 247                      &transparent_hugepage_flags))
 248                 output = "[always] defer defer+madvise madvise never";
 249         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 250                           &transparent_hugepage_flags))
 251                 output = "always [defer] defer+madvise madvise never";
 252         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 253                           &transparent_hugepage_flags))
 254                 output = "always defer [defer+madvise] madvise never";
 255         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 256                           &transparent_hugepage_flags))
 257                 output = "always defer defer+madvise [madvise] never";
 258         else
 259                 output = "always defer defer+madvise madvise [never]";
 260
 261         return sysfs_emit(buf, "%s\n", output);
 262 }
 263
 264 static ssize_t defrag_store(struct kobject *kobj,
 265                             struct kobj_attribute *attr,
 266                             const char *buf, size_t count)
 267 {
 268         if (sysfs_streq(buf, "always")) {
 269                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 272                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 273         } else if (sysfs_streq(buf, "defer+madvise")) {
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 277                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 278         } else if (sysfs_streq(buf, "defer")) {
 279                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 282                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 283         } else if (sysfs_streq(buf, "madvise")) {
 284                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 285                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 287                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 288         } else if (sysfs_streq(buf, "never")) {
 289                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 290                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 291                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 292                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 293         } else
 294                 return -EINVAL;
 295
 296         return count;
 297 }
 298 static struct kobj_attribute defrag_attr =
 299         __ATTR(defrag, 0644, defrag_show, defrag_store);
 300
 301 static ssize_t use_zero_page_show(struct kobject *kobj,
 302                                   struct kobj_attribute *attr, char *buf)
 303 {
 304         return single_hugepage_flag_show(kobj, attr, buf,
 305                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 306 }
 307 static ssize_t use_zero_page_store(struct kobject *kobj,
 308                 struct kobj_attribute *attr, const char *buf, size_t count)
 309 {
 310         return single_hugepage_flag_store(kobj, attr, buf, count,
 311                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 312 }
 313 static struct kobj_attribute use_zero_page_attr =
 314         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 315
 316 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 317                                    struct kobj_attribute *attr, char *buf)
 318 {
 319         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 320 }
 321 static struct kobj_attribute hpage_pmd_size_attr =
 322         __ATTR_RO(hpage_pmd_size);
 323
 324 static struct attribute *hugepage_attr[] = {
 325         &enabled_attr.attr,
 326         &defrag_attr.attr,
 327         &use_zero_page_attr.attr,
 328         &hpage_pmd_size_attr.attr,
 329 #ifdef CONFIG_SHMEM
 330         &shmem_enabled_attr.attr,
 331 #endif
 332         NULL,
 333 };
 334
 335 static const struct attribute_group hugepage_attr_group = {
 336         .attrs = hugepage_attr,
 337 };
 338
 339 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 340 {
 341         int err;
 342
 343         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 344         if (unlikely(!*hugepage_kobj)) {
 345                 pr_err("failed to create transparent hugepage kobject\n");
 346                 return -ENOMEM;
 347         }
 348
 349         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 350         if (err) {
 351                 pr_err("failed to register transparent hugepage group\n");
 352                 goto delete_obj;
 353         }
 354
 355         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 356         if (err) {
 357                 pr_err("failed to register transparent hugepage group\n");
 358                 goto remove_hp_group;
 359         }
 360
 361         return 0;
 362
 363 remove_hp_group:
 364         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 365 delete_obj:
 366         kobject_put(*hugepage_kobj);
 367         return err;
 368 }
 369
 370 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 371 {
 372         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 373         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 374         kobject_put(hugepage_kobj);
 375 }
 376 #else
 377 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 378 {
 379         return 0;
 380 }
 381
 382 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 383 {
 384 }
 385 #endif /* CONFIG_SYSFS */
 386
 387 static int __init hugepage_init(void)
 388 {
 389         int err;
 390         struct kobject *hugepage_kobj;
 391
 392         if (!has_transparent_hugepage()) {
 393                 /*
 394                  * Hardware doesn't support hugepages, hence disable
 395                  * DAX PMD support.
 396                  */
 397                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 398                 return -EINVAL;
 399         }
 400
 401         /*
 402          * hugepages can't be allocated by the buddy allocator
 403          */
 404         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 405         /*
 406          * we use page->mapping and page->index in second tail page
 407          * as list_head: assuming THP order >= 2
 408          */
 409         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 410
 411         err = hugepage_init_sysfs(&hugepage_kobj);
 412         if (err)
 413                 goto err_sysfs;
 414
 415         err = khugepaged_init();
 416         if (err)
 417                 goto err_slab;
 418
 419         err = register_shrinker(&huge_zero_page_shrinker);
 420         if (err)
 421                 goto err_hzp_shrinker;
 422         err = register_shrinker(&deferred_split_shrinker);
 423         if (err)
 424                 goto err_split_shrinker;
 425
 426         /*
 427          * By default disable transparent hugepages on smaller systems,
 428          * where the extra memory used could hurt more than TLB overhead
 429          * is likely to save.  The admin can still enable it through /sys.
 430          */
 431         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 432                 transparent_hugepage_flags = 0;
 433                 return 0;
 434         }
 435
 436         err = start_stop_khugepaged();
 437         if (err)
 438                 goto err_khugepaged;
 439
 440         return 0;
 441 err_khugepaged:
 442         unregister_shrinker(&deferred_split_shrinker);
 443 err_split_shrinker:
 444         unregister_shrinker(&huge_zero_page_shrinker);
 445 err_hzp_shrinker:
 446         khugepaged_destroy();
 447 err_slab:
 448         hugepage_exit_sysfs(hugepage_kobj);
 449 err_sysfs:
 450         return err;
 451 }
 452 subsys_initcall(hugepage_init);
 453
 454 static int __init setup_transparent_hugepage(char *str)
 455 {
 456         int ret = 0;
 457         if (!str)
 458                 goto out;
 459         if (!strcmp(str, "always")) {
 460                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 461                         &transparent_hugepage_flags);
 462                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 463                           &transparent_hugepage_flags);
 464                 ret = 1;
 465         } else if (!strcmp(str, "madvise")) {
 466                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 467                           &transparent_hugepage_flags);
 468                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 469                         &transparent_hugepage_flags);
 470                 ret = 1;
 471         } else if (!strcmp(str, "never")) {
 472                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 473                           &transparent_hugepage_flags);
 474                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 475                           &transparent_hugepage_flags);
 476                 ret = 1;
 477         }
 478 out:
 479         if (!ret)
 480                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 481         return ret;
 482 }
 483 __setup("transparent_hugepage=", setup_transparent_hugepage);
 484
 485 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 486 {
 487         if (likely(vma->vm_flags & VM_WRITE))
 488                 pmd = pmd_mkwrite(pmd);
 489         return pmd;
 490 }
 491
 492 #ifdef CONFIG_MEMCG
 493 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 494 {
 495         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 496         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 497
 498         if (memcg)
 499                 return &memcg->deferred_split_queue;
 500         else
 501                 return &pgdat->deferred_split_queue;
 502 }
 503 #else
 504 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 505 {
 506         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 507
 508         return &pgdat->deferred_split_queue;
 509 }
 510 #endif
 511
 512 void prep_transhuge_page(struct page *page)
 513 {
 514         /*
 515          * we use page->mapping and page->indexlru in second tail page
 516          * as list_head: assuming THP order >= 2
 517          */
 518
 519         INIT_LIST_HEAD(page_deferred_list(page));
 520         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 521 }
 522
 523 bool is_transparent_hugepage(struct page *page)
 524 {
 525         if (!PageCompound(page))
 526                 return false;
 527
 528         page = compound_head(page);
 529         return is_huge_zero_page(page) ||
 530                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 531 }
 532 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 533
 534 static unsigned long __thp_get_unmapped_area(struct file *filp,
 535                 unsigned long addr, unsigned long len,
 536                 loff_t off, unsigned long flags, unsigned long size)
 537 {
 538         loff_t off_end = off + len;
 539         loff_t off_align = round_up(off, size);
 540         unsigned long len_pad, ret;
 541
 542         if (off_end <= off_align || (off_end - off_align) < size)
 543                 return 0;
 544
 545         len_pad = len + size;
 546         if (len_pad < len || (off + len_pad) < off)
 547                 return 0;
 548
 549         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 550                                               off >> PAGE_SHIFT, flags);
 551
 552         /*
 553          * The failure might be due to length padding. The caller will retry
 554          * without the padding.
 555          */
 556         if (IS_ERR_VALUE(ret))
 557                 return 0;
 558
 559         /*
 560          * Do not try to align to THP boundary if allocation at the address
 561          * hint succeeds.
 562          */
 563         if (ret == addr)
 564                 return addr;
 565
 566         ret += (off - ret) & (size - 1);
 567         return ret;
 568 }
 569
 570 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 571                 unsigned long len, unsigned long pgoff, unsigned long flags)
 572 {
 573         unsigned long ret;
 574         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 575
 576         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 577                 goto out;
 578
 579         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 580         if (ret)
 581                 return ret;
 582 out:
 583         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 584 }
 585 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 586
 587 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 588                         struct page *page, gfp_t gfp)
 589 {
 590         struct vm_area_struct *vma = vmf->vma;
 591         pgtable_t pgtable;
 592         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 593         vm_fault_t ret = 0;
 594
 595         VM_BUG_ON_PAGE(!PageCompound(page), page);
 596
 597         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 598                 put_page(page);
 599                 count_vm_event(THP_FAULT_FALLBACK);
 600                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 601                 return VM_FAULT_FALLBACK;
 602         }
 603         cgroup_throttle_swaprate(page, gfp);
 604
 605         pgtable = pte_alloc_one(vma->vm_mm);
 606         if (unlikely(!pgtable)) {
 607                 ret = VM_FAULT_OOM;
 608                 goto release;
 609         }
 610
 611         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 612         /*
 613          * The memory barrier inside __SetPageUptodate makes sure that
 614          * clear_huge_page writes become visible before the set_pmd_at()
 615          * write.
 616          */
 617         __SetPageUptodate(page);
 618
 619         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 620         if (unlikely(!pmd_none(*vmf->pmd))) {
 621                 goto unlock_release;
 622         } else {
 623                 pmd_t entry;
 624
 625                 ret = check_stable_address_space(vma->vm_mm);
 626                 if (ret)
 627                         goto unlock_release;
 628
 629                 /* Deliver the page fault to userland */
 630                 if (userfaultfd_missing(vma)) {
 631                         spin_unlock(vmf->ptl);
 632                         put_page(page);
 633                         pte_free(vma->vm_mm, pgtable);
 634                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 635                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 636                         return ret;
 637                 }
 638
 639                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 640                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 641                 page_add_new_anon_rmap(page, vma, haddr, true);
 642                 lru_cache_add_inactive_or_unevictable(page, vma);
 643                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 644                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 645                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 646                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 647                 mm_inc_nr_ptes(vma->vm_mm);
 648                 spin_unlock(vmf->ptl);
 649                 count_vm_event(THP_FAULT_ALLOC);
 650                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 651         }
 652
 653         return 0;
 654 unlock_release:
 655         spin_unlock(vmf->ptl);
 656 release:
 657         if (pgtable)
 658                 pte_free(vma->vm_mm, pgtable);
 659         put_page(page);
 660         return ret;
 661
 662 }
 663
 664 /*
 665  * always: directly stall for all thp allocations
 666  * defer: wake kswapd and fail if not immediately available
 667  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 668  *                fail if not immediately available
 669  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 670  *          available
 671  * never: never stall for any thp allocation
 672  */
 673 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 674 {
 675         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 676
 677         /* Always do synchronous compaction */
 678         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 679                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 680
 681         /* Kick kcompactd and fail quickly */
 682         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 683                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 684
 685         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 686         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 687                 return GFP_TRANSHUGE_LIGHT |
 688                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 689                                         __GFP_KSWAPD_RECLAIM);
 690
 691         /* Only do synchronous compaction if madvised */
 692         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 693                 return GFP_TRANSHUGE_LIGHT |
 694                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 695
 696         return GFP_TRANSHUGE_LIGHT;
 697 }
 698
 699 /* Caller must hold page table lock. */
 700 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 701                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 702                 struct page *zero_page)
 703 {
 704         pmd_t entry;
 705         if (!pmd_none(*pmd))
 706                 return;
 707         entry = mk_pmd(zero_page, vma->vm_page_prot);
 708         entry = pmd_mkhuge(entry);
 709         if (pgtable)
 710                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 711         set_pmd_at(mm, haddr, pmd, entry);
 712         mm_inc_nr_ptes(mm);
 713 }
 714
 715 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 716 {
 717         struct vm_area_struct *vma = vmf->vma;
 718         gfp_t gfp;
 719         struct page *page;
 720         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 721
 722         if (!transhuge_vma_suitable(vma, haddr))
 723                 return VM_FAULT_FALLBACK;
 724         if (unlikely(anon_vma_prepare(vma)))
 725                 return VM_FAULT_OOM;
 726         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 727                 return VM_FAULT_OOM;
 728         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 729                         !mm_forbids_zeropage(vma->vm_mm) &&
 730                         transparent_hugepage_use_zero_page()) {
 731                 pgtable_t pgtable;
 732                 struct page *zero_page;
 733                 vm_fault_t ret;
 734                 pgtable = pte_alloc_one(vma->vm_mm);
 735                 if (unlikely(!pgtable))
 736                         return VM_FAULT_OOM;
 737                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 738                 if (unlikely(!zero_page)) {
 739                         pte_free(vma->vm_mm, pgtable);
 740                         count_vm_event(THP_FAULT_FALLBACK);
 741                         return VM_FAULT_FALLBACK;
 742                 }
 743                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 744                 ret = 0;
 745                 if (pmd_none(*vmf->pmd)) {
 746                         ret = check_stable_address_space(vma->vm_mm);
 747                         if (ret) {
 748                                 spin_unlock(vmf->ptl);
 749                                 pte_free(vma->vm_mm, pgtable);
 750                         } else if (userfaultfd_missing(vma)) {
 751                                 spin_unlock(vmf->ptl);
 752                                 pte_free(vma->vm_mm, pgtable);
 753                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 754                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 755                         } else {
 756                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 757                                                    haddr, vmf->pmd, zero_page);
 758                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 759                                 spin_unlock(vmf->ptl);
 760                         }
 761                 } else {
 762                         spin_unlock(vmf->ptl);
 763                         pte_free(vma->vm_mm, pgtable);
 764                 }
 765                 return ret;
 766         }
 767         gfp = vma_thp_gfp_mask(vma);
 768         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 769         if (unlikely(!page)) {
 770                 count_vm_event(THP_FAULT_FALLBACK);
 771                 return VM_FAULT_FALLBACK;
 772         }
 773         prep_transhuge_page(page);
 774         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 775 }
 776
 777 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 778                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 779                 pgtable_t pgtable)
 780 {
 781         struct mm_struct *mm = vma->vm_mm;
 782         pmd_t entry;
 783         spinlock_t *ptl;
 784
 785         ptl = pmd_lock(mm, pmd);
 786         if (!pmd_none(*pmd)) {
 787                 if (write) {
 788                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 789                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 790                                 goto out_unlock;
 791                         }
 792                         entry = pmd_mkyoung(*pmd);
 793                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 794                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 795                                 update_mmu_cache_pmd(vma, addr, pmd);
 796                 }
 797
 798                 goto out_unlock;
 799         }
 800
 801         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 802         if (pfn_t_devmap(pfn))
 803                 entry = pmd_mkdevmap(entry);
 804         if (write) {
 805                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 806                 entry = maybe_pmd_mkwrite(entry, vma);
 807         }
 808
 809         if (pgtable) {
 810                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 811                 mm_inc_nr_ptes(mm);
 812                 pgtable = NULL;
 813         }
 814
 815         set_pmd_at(mm, addr, pmd, entry);
 816         update_mmu_cache_pmd(vma, addr, pmd);
 817
 818 out_unlock:
 819         spin_unlock(ptl);
 820         if (pgtable)
 821                 pte_free(mm, pgtable);
 822 }
 823
 824 /**
 825  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 826  * @vmf: Structure describing the fault
 827  * @pfn: pfn to insert
 828  * @pgprot: page protection to use
 829  * @write: whether it's a write fault
 830  *
 831  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 832  * also consult the vmf_insert_mixed_prot() documentation when
 833  * @pgprot != @vmf->vma->vm_page_prot.
 834  *
 835  * Return: vm_fault_t value.
 836  */
 837 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 838                                    pgprot_t pgprot, bool write)
 839 {
 840         unsigned long addr = vmf->address & PMD_MASK;
 841         struct vm_area_struct *vma = vmf->vma;
 842         pgtable_t pgtable = NULL;
 843
 844         /*
 845          * If we had pmd_special, we could avoid all these restrictions,
 846          * but we need to be consistent with PTEs and architectures that
 847          * can't support a 'special' bit.
 848          */
 849         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 850                         !pfn_t_devmap(pfn));
 851         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 852                                                 (VM_PFNMAP|VM_MIXEDMAP));
 853         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 854
 855         if (addr < vma->vm_start || addr >= vma->vm_end)
 856                 return VM_FAULT_SIGBUS;
 857
 858         if (arch_needs_pgtable_deposit()) {
 859                 pgtable = pte_alloc_one(vma->vm_mm);
 860                 if (!pgtable)
 861                         return VM_FAULT_OOM;
 862         }
 863
 864         track_pfn_insert(vma, &pgprot, pfn);
 865
 866         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 867         return VM_FAULT_NOPAGE;
 868 }
 869 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 870
 871 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 872 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 873 {
 874         if (likely(vma->vm_flags & VM_WRITE))
 875                 pud = pud_mkwrite(pud);
 876         return pud;
 877 }
 878
 879 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 880                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 881 {
 882         struct mm_struct *mm = vma->vm_mm;
 883         pud_t entry;
 884         spinlock_t *ptl;
 885
 886         ptl = pud_lock(mm, pud);
 887         if (!pud_none(*pud)) {
 888                 if (write) {
 889                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 890                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 891                                 goto out_unlock;
 892                         }
 893                         entry = pud_mkyoung(*pud);
 894                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 895                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 896                                 update_mmu_cache_pud(vma, addr, pud);
 897                 }
 898                 goto out_unlock;
 899         }
 900
 901         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 902         if (pfn_t_devmap(pfn))
 903                 entry = pud_mkdevmap(entry);
 904         if (write) {
 905                 entry = pud_mkyoung(pud_mkdirty(entry));
 906                 entry = maybe_pud_mkwrite(entry, vma);
 907         }
 908         set_pud_at(mm, addr, pud, entry);
 909         update_mmu_cache_pud(vma, addr, pud);
 910
 911 out_unlock:
 912         spin_unlock(ptl);
 913 }
 914
 915 /**
 916  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 917  * @vmf: Structure describing the fault
 918  * @pfn: pfn to insert
 919  * @pgprot: page protection to use
 920  * @write: whether it's a write fault
 921  *
 922  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 923  * also consult the vmf_insert_mixed_prot() documentation when
 924  * @pgprot != @vmf->vma->vm_page_prot.
 925  *
 926  * Return: vm_fault_t value.
 927  */
 928 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 929                                    pgprot_t pgprot, bool write)
 930 {
 931         unsigned long addr = vmf->address & PUD_MASK;
 932         struct vm_area_struct *vma = vmf->vma;
 933
 934         /*
 935          * If we had pud_special, we could avoid all these restrictions,
 936          * but we need to be consistent with PTEs and architectures that
 937          * can't support a 'special' bit.
 938          */
 939         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 940                         !pfn_t_devmap(pfn));
 941         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 942                                                 (VM_PFNMAP|VM_MIXEDMAP));
 943         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 944
 945         if (addr < vma->vm_start || addr >= vma->vm_end)
 946                 return VM_FAULT_SIGBUS;
 947
 948         track_pfn_insert(vma, &pgprot, pfn);
 949
 950         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 951         return VM_FAULT_NOPAGE;
 952 }
 953 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 954 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 955
 956 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 957                 pmd_t *pmd, int flags)
 958 {
 959         pmd_t _pmd;
 960
 961         _pmd = pmd_mkyoung(*pmd);
 962         if (flags & FOLL_WRITE)
 963                 _pmd = pmd_mkdirty(_pmd);
 964         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 965                                 pmd, _pmd, flags & FOLL_WRITE))
 966                 update_mmu_cache_pmd(vma, addr, pmd);
 967 }
 968
 969 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 970                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 971 {
 972         unsigned long pfn = pmd_pfn(*pmd);
 973         struct mm_struct *mm = vma->vm_mm;
 974         struct page *page;
 975
 976         assert_spin_locked(pmd_lockptr(mm, pmd));
 977
 978         /*
 979          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 980          * not be in this function with `flags & FOLL_COW` set.
 981          */
 982         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 983
 984         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 985         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 986                          (FOLL_PIN | FOLL_GET)))
 987                 return NULL;
 988
 989         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 990                 return NULL;
 991
 992         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 993                 /* pass */;
 994         else
 995                 return NULL;
 996
 997         if (flags & FOLL_TOUCH)
 998                 touch_pmd(vma, addr, pmd, flags);
 999
1000         /*
1001          * device mapped pages can only be returned if the
1002          * caller will manage the page reference count.
1003          */
1004         if (!(flags & (FOLL_GET | FOLL_PIN)))
1005                 return ERR_PTR(-EEXIST);
1006
1007         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1008         *pgmap = get_dev_pagemap(pfn, *pgmap);
1009         if (!*pgmap)
1010                 return ERR_PTR(-EFAULT);
1011         page = pfn_to_page(pfn);
1012         if (!try_grab_page(page, flags))
1013                 page = ERR_PTR(-ENOMEM);
1014
1015         return page;
1016 }
1017
1018 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1019                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1020                   struct vm_area_struct *vma)
1021 {
1022         spinlock_t *dst_ptl, *src_ptl;
1023         struct page *src_page;
1024         pmd_t pmd;
1025         pgtable_t pgtable = NULL;
1026         int ret = -ENOMEM;
1027
1028         /* Skip if can be re-fill on fault */
1029         if (!vma_is_anonymous(vma))
1030                 return 0;
1031
1032         pgtable = pte_alloc_one(dst_mm);
1033         if (unlikely(!pgtable))
1034                 goto out;
1035
1036         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1037         src_ptl = pmd_lockptr(src_mm, src_pmd);
1038         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1039
1040         ret = -EAGAIN;
1041         pmd = *src_pmd;
1042
1043         /*
1044          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1045          * does not have the VM_UFFD_WP, which means that the uffd
1046          * fork event is not enabled.
1047          */
1048         if (!(vma->vm_flags & VM_UFFD_WP))
1049                 pmd = pmd_clear_uffd_wp(pmd);
1050
1051 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1052         if (unlikely(is_swap_pmd(pmd))) {
1053                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1054
1055                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1056                 if (is_write_migration_entry(entry)) {
1057                         make_migration_entry_read(&entry);
1058                         pmd = swp_entry_to_pmd(entry);
1059                         if (pmd_swp_soft_dirty(*src_pmd))
1060                                 pmd = pmd_swp_mksoft_dirty(pmd);
1061                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1062                 }
1063                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1064                 mm_inc_nr_ptes(dst_mm);
1065                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1066                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1067                 ret = 0;
1068                 goto out_unlock;
1069         }
1070 #endif
1071
1072         if (unlikely(!pmd_trans_huge(pmd))) {
1073                 pte_free(dst_mm, pgtable);
1074                 goto out_unlock;
1075         }
1076         /*
1077          * When page table lock is held, the huge zero pmd should not be
1078          * under splitting since we don't split the page itself, only pmd to
1079          * a page table.
1080          */
1081         if (is_huge_zero_pmd(pmd)) {
1082                 struct page *zero_page;
1083                 /*
1084                  * get_huge_zero_page() will never allocate a new page here,
1085                  * since we already have a zero page to copy. It just takes a
1086                  * reference.
1087                  */
1088                 zero_page = mm_get_huge_zero_page(dst_mm);
1089                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1090                                 zero_page);
1091                 ret = 0;
1092                 goto out_unlock;
1093         }
1094
1095         src_page = pmd_page(pmd);
1096         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1097
1098         /*
1099          * If this page is a potentially pinned page, split and retry the fault
1100          * with smaller page size.  Normally this should not happen because the
1101          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1102          * best effort that the pinned pages won't be replaced by another
1103          * random page during the coming copy-on-write.
1104          */
1105         if (unlikely(page_needs_cow_for_dma(vma, src_page))) {
1106                 pte_free(dst_mm, pgtable);
1107                 spin_unlock(src_ptl);
1108                 spin_unlock(dst_ptl);
1109                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1110                 return -EAGAIN;
1111         }
1112
1113         get_page(src_page);
1114         page_dup_rmap(src_page, true);
1115         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1116         mm_inc_nr_ptes(dst_mm);
1117         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1118
1119         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1120         pmd = pmd_mkold(pmd_wrprotect(pmd));
1121         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1122
1123         ret = 0;
1124 out_unlock:
1125         spin_unlock(src_ptl);
1126         spin_unlock(dst_ptl);
1127 out:
1128         return ret;
1129 }
1130
1131 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1132 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1133                 pud_t *pud, int flags)
1134 {
1135         pud_t _pud;
1136
1137         _pud = pud_mkyoung(*pud);
1138         if (flags & FOLL_WRITE)
1139                 _pud = pud_mkdirty(_pud);
1140         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1141                                 pud, _pud, flags & FOLL_WRITE))
1142                 update_mmu_cache_pud(vma, addr, pud);
1143 }
1144
1145 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1146                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1147 {
1148         unsigned long pfn = pud_pfn(*pud);
1149         struct mm_struct *mm = vma->vm_mm;
1150         struct page *page;
1151
1152         assert_spin_locked(pud_lockptr(mm, pud));
1153
1154         if (flags & FOLL_WRITE && !pud_write(*pud))
1155                 return NULL;
1156
1157         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1158         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1159                          (FOLL_PIN | FOLL_GET)))
1160                 return NULL;
1161
1162         if (pud_present(*pud) && pud_devmap(*pud))
1163                 /* pass */;
1164         else
1165                 return NULL;
1166
1167         if (flags & FOLL_TOUCH)
1168                 touch_pud(vma, addr, pud, flags);
1169
1170         /*
1171          * device mapped pages can only be returned if the
1172          * caller will manage the page reference count.
1173          *
1174          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1175          */
1176         if (!(flags & (FOLL_GET | FOLL_PIN)))
1177                 return ERR_PTR(-EEXIST);
1178
1179         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1180         *pgmap = get_dev_pagemap(pfn, *pgmap);
1181         if (!*pgmap)
1182                 return ERR_PTR(-EFAULT);
1183         page = pfn_to_page(pfn);
1184         if (!try_grab_page(page, flags))
1185                 page = ERR_PTR(-ENOMEM);
1186
1187         return page;
1188 }
1189
1190 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1191                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1192                   struct vm_area_struct *vma)
1193 {
1194         spinlock_t *dst_ptl, *src_ptl;
1195         pud_t pud;
1196         int ret;
1197
1198         dst_ptl = pud_lock(dst_mm, dst_pud);
1199         src_ptl = pud_lockptr(src_mm, src_pud);
1200         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1201
1202         ret = -EAGAIN;
1203         pud = *src_pud;
1204         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1205                 goto out_unlock;
1206
1207         /*
1208          * When page table lock is held, the huge zero pud should not be
1209          * under splitting since we don't split the page itself, only pud to
1210          * a page table.
1211          */
1212         if (is_huge_zero_pud(pud)) {
1213                 /* No huge zero pud yet */
1214         }
1215
1216         /* Please refer to comments in copy_huge_pmd() */
1217         if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1218                 spin_unlock(src_ptl);
1219                 spin_unlock(dst_ptl);
1220                 __split_huge_pud(vma, src_pud, addr);
1221                 return -EAGAIN;
1222         }
1223
1224         pudp_set_wrprotect(src_mm, addr, src_pud);
1225         pud = pud_mkold(pud_wrprotect(pud));
1226         set_pud_at(dst_mm, addr, dst_pud, pud);
1227
1228         ret = 0;
1229 out_unlock:
1230         spin_unlock(src_ptl);
1231         spin_unlock(dst_ptl);
1232         return ret;
1233 }
1234
1235 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1236 {
1237         pud_t entry;
1238         unsigned long haddr;
1239         bool write = vmf->flags & FAULT_FLAG_WRITE;
1240
1241         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1242         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1243                 goto unlock;
1244
1245         entry = pud_mkyoung(orig_pud);
1246         if (write)
1247                 entry = pud_mkdirty(entry);
1248         haddr = vmf->address & HPAGE_PUD_MASK;
1249         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1250                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1251
1252 unlock:
1253         spin_unlock(vmf->ptl);
1254 }
1255 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1256
1257 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1258 {
1259         pmd_t entry;
1260         unsigned long haddr;
1261         bool write = vmf->flags & FAULT_FLAG_WRITE;
1262
1263         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1264         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1265                 goto unlock;
1266
1267         entry = pmd_mkyoung(orig_pmd);
1268         if (write)
1269                 entry = pmd_mkdirty(entry);
1270         haddr = vmf->address & HPAGE_PMD_MASK;
1271         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1272                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1273
1274 unlock:
1275         spin_unlock(vmf->ptl);
1276 }
1277
1278 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1279 {
1280         struct vm_area_struct *vma = vmf->vma;
1281         struct page *page;
1282         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1283
1284         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1285         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1286
1287         if (is_huge_zero_pmd(orig_pmd))
1288                 goto fallback;
1289
1290         spin_lock(vmf->ptl);
1291
1292         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1293                 spin_unlock(vmf->ptl);
1294                 return 0;
1295         }
1296
1297         page = pmd_page(orig_pmd);
1298         VM_BUG_ON_PAGE(!PageHead(page), page);
1299
1300         /* Lock page for reuse_swap_page() */
1301         if (!trylock_page(page)) {
1302                 get_page(page);
1303                 spin_unlock(vmf->ptl);
1304                 lock_page(page);
1305                 spin_lock(vmf->ptl);
1306                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1307                         spin_unlock(vmf->ptl);
1308                         unlock_page(page);
1309                         put_page(page);
1310                         return 0;
1311                 }
1312                 put_page(page);
1313         }
1314
1315         /*
1316          * We can only reuse the page if nobody else maps the huge page or it's
1317          * part.
1318          */
1319         if (reuse_swap_page(page, NULL)) {
1320                 pmd_t entry;
1321                 entry = pmd_mkyoung(orig_pmd);
1322                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1323                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1324                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1325                 unlock_page(page);
1326                 spin_unlock(vmf->ptl);
1327                 return VM_FAULT_WRITE;
1328         }
1329
1330         unlock_page(page);
1331         spin_unlock(vmf->ptl);
1332 fallback:
1333         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1334         return VM_FAULT_FALLBACK;
1335 }
1336
1337 /*
1338  * FOLL_FORCE can write to even unwritable pmd's, but only
1339  * after we've gone through a COW cycle and they are dirty.
1340  */
1341 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1342 {
1343         return pmd_write(pmd) ||
1344                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1345 }
1346
1347 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1348                                    unsigned long addr,
1349                                    pmd_t *pmd,
1350                                    unsigned int flags)
1351 {
1352         struct mm_struct *mm = vma->vm_mm;
1353         struct page *page = NULL;
1354
1355         assert_spin_locked(pmd_lockptr(mm, pmd));
1356
1357         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1358                 goto out;
1359
1360         /* Avoid dumping huge zero page */
1361         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1362                 return ERR_PTR(-EFAULT);
1363
1364         /* Full NUMA hinting faults to serialise migration in fault paths */
1365         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1366                 goto out;
1367
1368         page = pmd_page(*pmd);
1369         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1370
1371         if (!try_grab_page(page, flags))
1372                 return ERR_PTR(-ENOMEM);
1373
1374         if (flags & FOLL_TOUCH)
1375                 touch_pmd(vma, addr, pmd, flags);
1376
1377         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1378                 /*
1379                  * We don't mlock() pte-mapped THPs. This way we can avoid
1380                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1381                  *
1382                  * For anon THP:
1383                  *
1384                  * In most cases the pmd is the only mapping of the page as we
1385                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1386                  * writable private mappings in populate_vma_page_range().
1387                  *
1388                  * The only scenario when we have the page shared here is if we
1389                  * mlocking read-only mapping shared over fork(). We skip
1390                  * mlocking such pages.
1391                  *
1392                  * For file THP:
1393                  *
1394                  * We can expect PageDoubleMap() to be stable under page lock:
1395                  * for file pages we set it in page_add_file_rmap(), which
1396                  * requires page to be locked.
1397                  */
1398
1399                 if (PageAnon(page) && compound_mapcount(page) != 1)
1400                         goto skip_mlock;
1401                 if (PageDoubleMap(page) || !page->mapping)
1402                         goto skip_mlock;
1403                 if (!trylock_page(page))
1404                         goto skip_mlock;
1405                 if (page->mapping && !PageDoubleMap(page))
1406                         mlock_vma_page(page);
1407                 unlock_page(page);
1408         }
1409 skip_mlock:
1410         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1411         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1412
1413 out:
1414         return page;
1415 }
1416
1417 /* NUMA hinting page fault entry point for trans huge pmds */
1418 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1419 {
1420         struct vm_area_struct *vma = vmf->vma;
1421         struct anon_vma *anon_vma = NULL;
1422         struct page *page;
1423         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1424         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1425         int target_nid, last_cpupid = -1;
1426         bool page_locked;
1427         bool migrated = false;
1428         bool was_writable;
1429         int flags = 0;
1430
1431         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1432         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1433                 goto out_unlock;
1434
1435         /*
1436          * If there are potential migrations, wait for completion and retry
1437          * without disrupting NUMA hinting information. Do not relock and
1438          * check_same as the page may no longer be mapped.
1439          */
1440         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1441                 page = pmd_page(*vmf->pmd);
1442                 if (!get_page_unless_zero(page))
1443                         goto out_unlock;
1444                 spin_unlock(vmf->ptl);
1445                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1446                 goto out;
1447         }
1448
1449         page = pmd_page(pmd);
1450         BUG_ON(is_huge_zero_page(page));
1451         page_nid = page_to_nid(page);
1452         last_cpupid = page_cpupid_last(page);
1453         count_vm_numa_event(NUMA_HINT_FAULTS);
1454         if (page_nid == this_nid) {
1455                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1456                 flags |= TNF_FAULT_LOCAL;
1457         }
1458
1459         /* See similar comment in do_numa_page for explanation */
1460         if (!pmd_savedwrite(pmd))
1461                 flags |= TNF_NO_GROUP;
1462
1463         /*
1464          * Acquire the page lock to serialise THP migrations but avoid dropping
1465          * page_table_lock if at all possible
1466          */
1467         page_locked = trylock_page(page);
1468         target_nid = mpol_misplaced(page, vma, haddr);
1469         /* Migration could have started since the pmd_trans_migrating check */
1470         if (!page_locked) {
1471                 page_nid = NUMA_NO_NODE;
1472                 if (!get_page_unless_zero(page))
1473                         goto out_unlock;
1474                 spin_unlock(vmf->ptl);
1475                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1476                 goto out;
1477         } else if (target_nid == NUMA_NO_NODE) {
1478                 /* There are no parallel migrations and page is in the right
1479                  * node. Clear the numa hinting info in this pmd.
1480                  */
1481                 goto clear_pmdnuma;
1482         }
1483
1484         /*
1485          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1486          * to serialises splits
1487          */
1488         get_page(page);
1489         spin_unlock(vmf->ptl);
1490         anon_vma = page_lock_anon_vma_read(page);
1491
1492         /* Confirm the PMD did not change while page_table_lock was released */
1493         spin_lock(vmf->ptl);
1494         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1495                 unlock_page(page);
1496                 put_page(page);
1497                 page_nid = NUMA_NO_NODE;
1498                 goto out_unlock;
1499         }
1500
1501         /* Bail if we fail to protect against THP splits for any reason */
1502         if (unlikely(!anon_vma)) {
1503                 put_page(page);
1504                 page_nid = NUMA_NO_NODE;
1505                 goto clear_pmdnuma;
1506         }
1507
1508         /*
1509          * Since we took the NUMA fault, we must have observed the !accessible
1510          * bit. Make sure all other CPUs agree with that, to avoid them
1511          * modifying the page we're about to migrate.
1512          *
1513          * Must be done under PTL such that we'll observe the relevant
1514          * inc_tlb_flush_pending().
1515          *
1516          * We are not sure a pending tlb flush here is for a huge page
1517          * mapping or not. Hence use the tlb range variant
1518          */
1519         if (mm_tlb_flush_pending(vma->vm_mm)) {
1520                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1521                 /*
1522                  * change_huge_pmd() released the pmd lock before
1523                  * invalidating the secondary MMUs sharing the primary
1524                  * MMU pagetables (with ->invalidate_range()). The
1525                  * mmu_notifier_invalidate_range_end() (which
1526                  * internally calls ->invalidate_range()) in
1527                  * change_pmd_range() will run after us, so we can't
1528                  * rely on it here and we need an explicit invalidate.
1529                  */
1530                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1531                                               haddr + HPAGE_PMD_SIZE);
1532         }
1533
1534         /*
1535          * Migrate the THP to the requested node, returns with page unlocked
1536          * and access rights restored.
1537          */
1538         spin_unlock(vmf->ptl);
1539
1540         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1541                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1542         if (migrated) {
1543                 flags |= TNF_MIGRATED;
1544                 page_nid = target_nid;
1545         } else
1546                 flags |= TNF_MIGRATE_FAIL;
1547
1548         goto out;
1549 clear_pmdnuma:
1550         BUG_ON(!PageLocked(page));
1551         was_writable = pmd_savedwrite(pmd);
1552         pmd = pmd_modify(pmd, vma->vm_page_prot);
1553         pmd = pmd_mkyoung(pmd);
1554         if (was_writable)
1555                 pmd = pmd_mkwrite(pmd);
1556         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1557         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1558         unlock_page(page);
1559 out_unlock:
1560         spin_unlock(vmf->ptl);
1561
1562 out:
1563         if (anon_vma)
1564                 page_unlock_anon_vma_read(anon_vma);
1565
1566         if (page_nid != NUMA_NO_NODE)
1567                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1568                                 flags);
1569
1570         return 0;
1571 }
1572
1573 /*
1574  * Return true if we do MADV_FREE successfully on entire pmd page.
1575  * Otherwise, return false.
1576  */
1577 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1578                 pmd_t *pmd, unsigned long addr, unsigned long next)
1579 {
1580         spinlock_t *ptl;
1581         pmd_t orig_pmd;
1582         struct page *page;
1583         struct mm_struct *mm = tlb->mm;
1584         bool ret = false;
1585
1586         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1587
1588         ptl = pmd_trans_huge_lock(pmd, vma);
1589         if (!ptl)
1590                 goto out_unlocked;
1591
1592         orig_pmd = *pmd;
1593         if (is_huge_zero_pmd(orig_pmd))
1594                 goto out;
1595
1596         if (unlikely(!pmd_present(orig_pmd))) {
1597                 VM_BUG_ON(thp_migration_supported() &&
1598                                   !is_pmd_migration_entry(orig_pmd));
1599                 goto out;
1600         }
1601
1602         page = pmd_page(orig_pmd);
1603         /*
1604          * If other processes are mapping this page, we couldn't discard
1605          * the page unless they all do MADV_FREE so let's skip the page.
1606          */
1607         if (page_mapcount(page) != 1)
1608                 goto out;
1609
1610         if (!trylock_page(page))
1611                 goto out;
1612
1613         /*
1614          * If user want to discard part-pages of THP, split it so MADV_FREE
1615          * will deactivate only them.
1616          */
1617         if (next - addr != HPAGE_PMD_SIZE) {
1618                 get_page(page);
1619                 spin_unlock(ptl);
1620                 split_huge_page(page);
1621                 unlock_page(page);
1622                 put_page(page);
1623                 goto out_unlocked;
1624         }
1625
1626         if (PageDirty(page))
1627                 ClearPageDirty(page);
1628         unlock_page(page);
1629
1630         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1631                 pmdp_invalidate(vma, addr, pmd);
1632                 orig_pmd = pmd_mkold(orig_pmd);
1633                 orig_pmd = pmd_mkclean(orig_pmd);
1634
1635                 set_pmd_at(mm, addr, pmd, orig_pmd);
1636                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1637         }
1638
1639         mark_page_lazyfree(page);
1640         ret = true;
1641 out:
1642         spin_unlock(ptl);
1643 out_unlocked:
1644         return ret;
1645 }
1646
1647 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1648 {
1649         pgtable_t pgtable;
1650
1651         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1652         pte_free(mm, pgtable);
1653         mm_dec_nr_ptes(mm);
1654 }
1655
1656 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1657                  pmd_t *pmd, unsigned long addr)
1658 {
1659         pmd_t orig_pmd;
1660         spinlock_t *ptl;
1661
1662         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1663
1664         ptl = __pmd_trans_huge_lock(pmd, vma);
1665         if (!ptl)
1666                 return 0;
1667         /*
1668          * For architectures like ppc64 we look at deposited pgtable
1669          * when calling pmdp_huge_get_and_clear. So do the
1670          * pgtable_trans_huge_withdraw after finishing pmdp related
1671          * operations.
1672          */
1673         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1674                                                 tlb->fullmm);
1675         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1676         if (vma_is_special_huge(vma)) {
1677                 if (arch_needs_pgtable_deposit())
1678                         zap_deposited_table(tlb->mm, pmd);
1679                 spin_unlock(ptl);
1680                 if (is_huge_zero_pmd(orig_pmd))
1681                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1682         } else if (is_huge_zero_pmd(orig_pmd)) {
1683                 zap_deposited_table(tlb->mm, pmd);
1684                 spin_unlock(ptl);
1685                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1686         } else {
1687                 struct page *page = NULL;
1688                 int flush_needed = 1;
1689
1690                 if (pmd_present(orig_pmd)) {
1691                         page = pmd_page(orig_pmd);
1692                         page_remove_rmap(page, true);
1693                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1694                         VM_BUG_ON_PAGE(!PageHead(page), page);
1695                 } else if (thp_migration_supported()) {
1696                         swp_entry_t entry;
1697
1698                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1699                         entry = pmd_to_swp_entry(orig_pmd);
1700                         page = migration_entry_to_page(entry);
1701                         flush_needed = 0;
1702                 } else
1703                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1704
1705                 if (PageAnon(page)) {
1706                         zap_deposited_table(tlb->mm, pmd);
1707                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1708                 } else {
1709                         if (arch_needs_pgtable_deposit())
1710                                 zap_deposited_table(tlb->mm, pmd);
1711                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1712                 }
1713
1714                 spin_unlock(ptl);
1715                 if (flush_needed)
1716                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1717         }
1718         return 1;
1719 }
1720
1721 #ifndef pmd_move_must_withdraw
1722 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1723                                          spinlock_t *old_pmd_ptl,
1724                                          struct vm_area_struct *vma)
1725 {
1726         /*
1727          * With split pmd lock we also need to move preallocated
1728          * PTE page table if new_pmd is on different PMD page table.
1729          *
1730          * We also don't deposit and withdraw tables for file pages.
1731          */
1732         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1733 }
1734 #endif
1735
1736 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1737 {
1738 #ifdef CONFIG_MEM_SOFT_DIRTY
1739         if (unlikely(is_pmd_migration_entry(pmd)))
1740                 pmd = pmd_swp_mksoft_dirty(pmd);
1741         else if (pmd_present(pmd))
1742                 pmd = pmd_mksoft_dirty(pmd);
1743 #endif
1744         return pmd;
1745 }
1746
1747 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1748                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1749 {
1750         spinlock_t *old_ptl, *new_ptl;
1751         pmd_t pmd;
1752         struct mm_struct *mm = vma->vm_mm;
1753         bool force_flush = false;
1754
1755         /*
1756          * The destination pmd shouldn't be established, free_pgtables()
1757          * should have release it.
1758          */
1759         if (WARN_ON(!pmd_none(*new_pmd))) {
1760                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1761                 return false;
1762         }
1763
1764         /*
1765          * We don't have to worry about the ordering of src and dst
1766          * ptlocks because exclusive mmap_lock prevents deadlock.
1767          */
1768         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1769         if (old_ptl) {
1770                 new_ptl = pmd_lockptr(mm, new_pmd);
1771                 if (new_ptl != old_ptl)
1772                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1773                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1774                 if (pmd_present(pmd))
1775                         force_flush = true;
1776                 VM_BUG_ON(!pmd_none(*new_pmd));
1777
1778                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1779                         pgtable_t pgtable;
1780                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1781                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1782                 }
1783                 pmd = move_soft_dirty_pmd(pmd);
1784                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1785                 if (force_flush)
1786                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1787                 if (new_ptl != old_ptl)
1788                         spin_unlock(new_ptl);
1789                 spin_unlock(old_ptl);
1790                 return true;
1791         }
1792         return false;
1793 }
1794
1795 /*
1796  * Returns
1797  *  - 0 if PMD could not be locked
1798  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1799  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1800  */
1801 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1802                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1803 {
1804         struct mm_struct *mm = vma->vm_mm;
1805         spinlock_t *ptl;
1806         pmd_t entry;
1807         bool preserve_write;
1808         int ret;
1809         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1810         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1811         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1812
1813         ptl = __pmd_trans_huge_lock(pmd, vma);
1814         if (!ptl)
1815                 return 0;
1816
1817         preserve_write = prot_numa && pmd_write(*pmd);
1818         ret = 1;
1819
1820 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1821         if (is_swap_pmd(*pmd)) {
1822                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1823
1824                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1825                 if (is_write_migration_entry(entry)) {
1826                         pmd_t newpmd;
1827                         /*
1828                          * A protection check is difficult so
1829                          * just be safe and disable write
1830                          */
1831                         make_migration_entry_read(&entry);
1832                         newpmd = swp_entry_to_pmd(entry);
1833                         if (pmd_swp_soft_dirty(*pmd))
1834                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1835                         set_pmd_at(mm, addr, pmd, newpmd);
1836                 }
1837                 goto unlock;
1838         }
1839 #endif
1840
1841         /*
1842          * Avoid trapping faults against the zero page. The read-only
1843          * data is likely to be read-cached on the local CPU and
1844          * local/remote hits to the zero page are not interesting.
1845          */
1846         if (prot_numa && is_huge_zero_pmd(*pmd))
1847                 goto unlock;
1848
1849         if (prot_numa && pmd_protnone(*pmd))
1850                 goto unlock;
1851
1852         /*
1853          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1854          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1855          * which is also under mmap_read_lock(mm):
1856          *
1857          *      CPU0:                           CPU1:
1858          *                              change_huge_pmd(prot_numa=1)
1859          *                               pmdp_huge_get_and_clear_notify()
1860          * madvise_dontneed()
1861          *  zap_pmd_range()
1862          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1863          *   // skip the pmd
1864          *                               set_pmd_at();
1865          *                               // pmd is re-established
1866          *
1867          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1868          * which may break userspace.
1869          *
1870          * pmdp_invalidate() is required to make sure we don't miss
1871          * dirty/young flags set by hardware.
1872          */
1873         entry = pmdp_invalidate(vma, addr, pmd);
1874
1875         entry = pmd_modify(entry, newprot);
1876         if (preserve_write)
1877                 entry = pmd_mk_savedwrite(entry);
1878         if (uffd_wp) {
1879                 entry = pmd_wrprotect(entry);
1880                 entry = pmd_mkuffd_wp(entry);
1881         } else if (uffd_wp_resolve) {
1882                 /*
1883                  * Leave the write bit to be handled by PF interrupt
1884                  * handler, then things like COW could be properly
1885                  * handled.
1886                  */
1887                 entry = pmd_clear_uffd_wp(entry);
1888         }
1889         ret = HPAGE_PMD_NR;
1890         set_pmd_at(mm, addr, pmd, entry);
1891         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1892 unlock:
1893         spin_unlock(ptl);
1894         return ret;
1895 }
1896
1897 /*
1898  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1899  *
1900  * Note that if it returns page table lock pointer, this routine returns without
1901  * unlocking page table lock. So callers must unlock it.
1902  */
1903 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1904 {
1905         spinlock_t *ptl;
1906         ptl = pmd_lock(vma->vm_mm, pmd);
1907         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1908                         pmd_devmap(*pmd)))
1909                 return ptl;
1910         spin_unlock(ptl);
1911         return NULL;
1912 }
1913
1914 /*
1915  * Returns true if a given pud maps a thp, false otherwise.
1916  *
1917  * Note that if it returns true, this routine returns without unlocking page
1918  * table lock. So callers must unlock it.
1919  */
1920 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1921 {
1922         spinlock_t *ptl;
1923
1924         ptl = pud_lock(vma->vm_mm, pud);
1925         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1926                 return ptl;
1927         spin_unlock(ptl);
1928         return NULL;
1929 }
1930
1931 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1932 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1933                  pud_t *pud, unsigned long addr)
1934 {
1935         spinlock_t *ptl;
1936
1937         ptl = __pud_trans_huge_lock(pud, vma);
1938         if (!ptl)
1939                 return 0;
1940         /*
1941          * For architectures like ppc64 we look at deposited pgtable
1942          * when calling pudp_huge_get_and_clear. So do the
1943          * pgtable_trans_huge_withdraw after finishing pudp related
1944          * operations.
1945          */
1946         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1947         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1948         if (vma_is_special_huge(vma)) {
1949                 spin_unlock(ptl);
1950                 /* No zero page support yet */
1951         } else {
1952                 /* No support for anonymous PUD pages yet */
1953                 BUG();
1954         }
1955         return 1;
1956 }
1957
1958 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1959                 unsigned long haddr)
1960 {
1961         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1962         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1963         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1964         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1965
1966         count_vm_event(THP_SPLIT_PUD);
1967
1968         pudp_huge_clear_flush_notify(vma, haddr, pud);
1969 }
1970
1971 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1972                 unsigned long address)
1973 {
1974         spinlock_t *ptl;
1975         struct mmu_notifier_range range;
1976
1977         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1978                                 address & HPAGE_PUD_MASK,
1979                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1980         mmu_notifier_invalidate_range_start(&range);
1981         ptl = pud_lock(vma->vm_mm, pud);
1982         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1983                 goto out;
1984         __split_huge_pud_locked(vma, pud, range.start);
1985
1986 out:
1987         spin_unlock(ptl);
1988         /*
1989          * No need to double call mmu_notifier->invalidate_range() callback as
1990          * the above pudp_huge_clear_flush_notify() did already call it.
1991          */
1992         mmu_notifier_invalidate_range_only_end(&range);
1993 }
1994 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1995
1996 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1997                 unsigned long haddr, pmd_t *pmd)
1998 {
1999         struct mm_struct *mm = vma->vm_mm;
2000         pgtable_t pgtable;
2001         pmd_t _pmd;
2002         int i;
2003
2004         /*
2005          * Leave pmd empty until pte is filled note that it is fine to delay
2006          * notification until mmu_notifier_invalidate_range_end() as we are
2007          * replacing a zero pmd write protected page with a zero pte write
2008          * protected page.
2009          *
2010          * See Documentation/vm/mmu_notifier.rst
2011          */
2012         pmdp_huge_clear_flush(vma, haddr, pmd);
2013
2014         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2015         pmd_populate(mm, &_pmd, pgtable);
2016
2017         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2018                 pte_t *pte, entry;
2019                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2020                 entry = pte_mkspecial(entry);
2021                 pte = pte_offset_map(&_pmd, haddr);
2022                 VM_BUG_ON(!pte_none(*pte));
2023                 set_pte_at(mm, haddr, pte, entry);
2024                 pte_unmap(pte);
2025         }
2026         smp_wmb(); /* make pte visible before pmd */
2027         pmd_populate(mm, pmd, pgtable);
2028 }
2029
2030 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2031                 unsigned long haddr, bool freeze)
2032 {
2033         struct mm_struct *mm = vma->vm_mm;
2034         struct page *page;
2035         pgtable_t pgtable;
2036         pmd_t old_pmd, _pmd;
2037         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2038         unsigned long addr;
2039         int i;
2040
2041         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2042         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2043         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2044         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2045                                 && !pmd_devmap(*pmd));
2046
2047         count_vm_event(THP_SPLIT_PMD);
2048
2049         if (!vma_is_anonymous(vma)) {
2050                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2051                 /*
2052                  * We are going to unmap this huge page. So
2053                  * just go ahead and zap it
2054                  */
2055                 if (arch_needs_pgtable_deposit())
2056                         zap_deposited_table(mm, pmd);
2057                 if (vma_is_special_huge(vma))
2058                         return;
2059                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2060                         swp_entry_t entry;
2061
2062                         entry = pmd_to_swp_entry(old_pmd);
2063                         page = migration_entry_to_page(entry);
2064                 } else {
2065                         page = pmd_page(old_pmd);
2066                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2067                                 set_page_dirty(page);
2068                         if (!PageReferenced(page) && pmd_young(old_pmd))
2069                                 SetPageReferenced(page);
2070                         page_remove_rmap(page, true);
2071                         put_page(page);
2072                 }
2073                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2074                 return;
2075         }
2076
2077         if (is_huge_zero_pmd(*pmd)) {
2078                 /*
2079                  * FIXME: Do we want to invalidate secondary mmu by calling
2080                  * mmu_notifier_invalidate_range() see comments below inside
2081                  * __split_huge_pmd() ?
2082                  *
2083                  * We are going from a zero huge page write protected to zero
2084                  * small page also write protected so it does not seems useful
2085                  * to invalidate secondary mmu at this time.
2086                  */
2087                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2088         }
2089
2090         /*
2091          * Up to this point the pmd is present and huge and userland has the
2092          * whole access to the hugepage during the split (which happens in
2093          * place). If we overwrite the pmd with the not-huge version pointing
2094          * to the pte here (which of course we could if all CPUs were bug
2095          * free), userland could trigger a small page size TLB miss on the
2096          * small sized TLB while the hugepage TLB entry is still established in
2097          * the huge TLB. Some CPU doesn't like that.
2098          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2099          * 383 on page 105. Intel should be safe but is also warns that it's
2100          * only safe if the permission and cache attributes of the two entries
2101          * loaded in the two TLB is identical (which should be the case here).
2102          * But it is generally safer to never allow small and huge TLB entries
2103          * for the same virtual address to be loaded simultaneously. So instead
2104          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2105          * current pmd notpresent (atomically because here the pmd_trans_huge
2106          * must remain set at all times on the pmd until the split is complete
2107          * for this pmd), then we flush the SMP TLB and finally we write the
2108          * non-huge version of the pmd entry with pmd_populate.
2109          */
2110         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2111
2112         pmd_migration = is_pmd_migration_entry(old_pmd);
2113         if (unlikely(pmd_migration)) {
2114                 swp_entry_t entry;
2115
2116                 entry = pmd_to_swp_entry(old_pmd);
2117                 page = migration_entry_to_page(entry);
2118                 write = is_write_migration_entry(entry);
2119                 young = false;
2120                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2121                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2122         } else {
2123                 page = pmd_page(old_pmd);
2124                 if (pmd_dirty(old_pmd))
2125                         SetPageDirty(page);
2126                 write = pmd_write(old_pmd);
2127                 young = pmd_young(old_pmd);
2128                 soft_dirty = pmd_soft_dirty(old_pmd);
2129                 uffd_wp = pmd_uffd_wp(old_pmd);
2130         }
2131         VM_BUG_ON_PAGE(!page_count(page), page);
2132         page_ref_add(page, HPAGE_PMD_NR - 1);
2133
2134         /*
2135          * Withdraw the table only after we mark the pmd entry invalid.
2136          * This's critical for some architectures (Power).
2137          */
2138         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2139         pmd_populate(mm, &_pmd, pgtable);
2140
2141         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2142                 pte_t entry, *pte;
2143                 /*
2144                  * Note that NUMA hinting access restrictions are not
2145                  * transferred to avoid any possibility of altering
2146                  * permissions across VMAs.
2147                  */
2148                 if (freeze || pmd_migration) {
2149                         swp_entry_t swp_entry;
2150                         swp_entry = make_migration_entry(page + i, write);
2151                         entry = swp_entry_to_pte(swp_entry);
2152                         if (soft_dirty)
2153                                 entry = pte_swp_mksoft_dirty(entry);
2154                         if (uffd_wp)
2155                                 entry = pte_swp_mkuffd_wp(entry);
2156                 } else {
2157                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2158                         entry = maybe_mkwrite(entry, vma);
2159                         if (!write)
2160                                 entry = pte_wrprotect(entry);
2161                         if (!young)
2162                                 entry = pte_mkold(entry);
2163                         if (soft_dirty)
2164                                 entry = pte_mksoft_dirty(entry);
2165                         if (uffd_wp)
2166                                 entry = pte_mkuffd_wp(entry);
2167                 }
2168                 pte = pte_offset_map(&_pmd, addr);
2169                 BUG_ON(!pte_none(*pte));
2170                 set_pte_at(mm, addr, pte, entry);
2171                 if (!pmd_migration)
2172                         atomic_inc(&page[i]._mapcount);
2173                 pte_unmap(pte);
2174         }
2175
2176         if (!pmd_migration) {
2177                 /*
2178                  * Set PG_double_map before dropping compound_mapcount to avoid
2179                  * false-negative page_mapped().
2180                  */
2181                 if (compound_mapcount(page) > 1 &&
2182                     !TestSetPageDoubleMap(page)) {
2183                         for (i = 0; i < HPAGE_PMD_NR; i++)
2184                                 atomic_inc(&page[i]._mapcount);
2185                 }
2186
2187                 lock_page_memcg(page);
2188                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2189                         /* Last compound_mapcount is gone. */
2190                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2191                                                 -HPAGE_PMD_NR);
2192                         if (TestClearPageDoubleMap(page)) {
2193                                 /* No need in mapcount reference anymore */
2194                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2195                                         atomic_dec(&page[i]._mapcount);
2196                         }
2197                 }
2198                 unlock_page_memcg(page);
2199         }
2200
2201         smp_wmb(); /* make pte visible before pmd */
2202         pmd_populate(mm, pmd, pgtable);
2203
2204         if (freeze) {
2205                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2206                         page_remove_rmap(page + i, false);
2207                         put_page(page + i);
2208                 }
2209         }
2210 }
2211
2212 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2213                 unsigned long address, bool freeze, struct page *page)
2214 {
2215         spinlock_t *ptl;
2216         struct mmu_notifier_range range;
2217         bool do_unlock_page = false;
2218         pmd_t _pmd;
2219
2220         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2221                                 address & HPAGE_PMD_MASK,
2222                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2223         mmu_notifier_invalidate_range_start(&range);
2224         ptl = pmd_lock(vma->vm_mm, pmd);
2225
2226         /*
2227          * If caller asks to setup a migration entries, we need a page to check
2228          * pmd against. Otherwise we can end up replacing wrong page.
2229          */
2230         VM_BUG_ON(freeze && !page);
2231         if (page) {
2232                 VM_WARN_ON_ONCE(!PageLocked(page));
2233                 if (page != pmd_page(*pmd))
2234                         goto out;
2235         }
2236
2237 repeat:
2238         if (pmd_trans_huge(*pmd)) {
2239                 if (!page) {
2240                         page = pmd_page(*pmd);
2241                         /*
2242                          * An anonymous page must be locked, to ensure that a
2243                          * concurrent reuse_swap_page() sees stable mapcount;
2244                          * but reuse_swap_page() is not used on shmem or file,
2245                          * and page lock must not be taken when zap_pmd_range()
2246                          * calls __split_huge_pmd() while i_mmap_lock is held.
2247                          */
2248                         if (PageAnon(page)) {
2249                                 if (unlikely(!trylock_page(page))) {
2250                                         get_page(page);
2251                                         _pmd = *pmd;
2252                                         spin_unlock(ptl);
2253                                         lock_page(page);
2254                                         spin_lock(ptl);
2255                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2256                                                 unlock_page(page);
2257                                                 put_page(page);
2258                                                 page = NULL;
2259                                                 goto repeat;
2260                                         }
2261                                         put_page(page);
2262                                 }
2263                                 do_unlock_page = true;
2264                         }
2265                 }
2266                 if (PageMlocked(page))
2267                         clear_page_mlock(page);
2268         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2269                 goto out;
2270         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2271 out:
2272         spin_unlock(ptl);
2273         if (do_unlock_page)
2274                 unlock_page(page);
2275         /*
2276          * No need to double call mmu_notifier->invalidate_range() callback.
2277          * They are 3 cases to consider inside __split_huge_pmd_locked():
2278          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2279          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2280          *    fault will trigger a flush_notify before pointing to a new page
2281          *    (it is fine if the secondary mmu keeps pointing to the old zero
2282          *    page in the meantime)
2283          *  3) Split a huge pmd into pte pointing to the same page. No need
2284          *     to invalidate secondary tlb entry they are all still valid.
2285          *     any further changes to individual pte will notify. So no need
2286          *     to call mmu_notifier->invalidate_range()
2287          */
2288         mmu_notifier_invalidate_range_only_end(&range);
2289 }
2290
2291 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2292                 bool freeze, struct page *page)
2293 {
2294         pgd_t *pgd;
2295         p4d_t *p4d;
2296         pud_t *pud;
2297         pmd_t *pmd;
2298
2299         pgd = pgd_offset(vma->vm_mm, address);
2300         if (!pgd_present(*pgd))
2301                 return;
2302
2303         p4d = p4d_offset(pgd, address);
2304         if (!p4d_present(*p4d))
2305                 return;
2306
2307         pud = pud_offset(p4d, address);
2308         if (!pud_present(*pud))
2309                 return;
2310
2311         pmd = pmd_offset(pud, address);
2312
2313         __split_huge_pmd(vma, pmd, address, freeze, page);
2314 }
2315
2316 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2317 {
2318         /*
2319          * If the new address isn't hpage aligned and it could previously
2320          * contain an hugepage: check if we need to split an huge pmd.
2321          */
2322         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2323             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2324                          ALIGN(address, HPAGE_PMD_SIZE)))
2325                 split_huge_pmd_address(vma, address, false, NULL);
2326 }
2327
2328 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2329                              unsigned long start,
2330                              unsigned long end,
2331                              long adjust_next)
2332 {
2333         /* Check if we need to split start first. */
2334         split_huge_pmd_if_needed(vma, start);
2335
2336         /* Check if we need to split end next. */
2337         split_huge_pmd_if_needed(vma, end);
2338
2339         /*
2340          * If we're also updating the vma->vm_next->vm_start,
2341          * check if we need to split it.
2342          */
2343         if (adjust_next > 0) {
2344                 struct vm_area_struct *next = vma->vm_next;
2345                 unsigned long nstart = next->vm_start;
2346                 nstart += adjust_next;
2347                 split_huge_pmd_if_needed(next, nstart);
2348         }
2349 }
2350
2351 static void unmap_page(struct page *page)
2352 {
2353         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_SYNC |
2354                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2355         bool unmap_success;
2356
2357         VM_BUG_ON_PAGE(!PageHead(page), page);
2358
2359         if (PageAnon(page))
2360                 ttu_flags |= TTU_SPLIT_FREEZE;
2361
2362         unmap_success = try_to_unmap(page, ttu_flags);
2363         VM_BUG_ON_PAGE(!unmap_success, page);
2364 }
2365
2366 static void remap_page(struct page *page, unsigned int nr)
2367 {
2368         int i;
2369         if (PageTransHuge(page)) {
2370                 remove_migration_ptes(page, page, true);
2371         } else {
2372                 for (i = 0; i < nr; i++)
2373                         remove_migration_ptes(page + i, page + i, true);
2374         }
2375 }
2376
2377 static void lru_add_page_tail(struct page *head, struct page *tail,
2378                 struct lruvec *lruvec, struct list_head *list)
2379 {
2380         VM_BUG_ON_PAGE(!PageHead(head), head);
2381         VM_BUG_ON_PAGE(PageCompound(tail), head);
2382         VM_BUG_ON_PAGE(PageLRU(tail), head);
2383         lockdep_assert_held(&lruvec->lru_lock);
2384
2385         if (list) {
2386                 /* page reclaim is reclaiming a huge page */
2387                 VM_WARN_ON(PageLRU(head));
2388                 get_page(tail);
2389                 list_add_tail(&tail->lru, list);
2390         } else {
2391                 /* head is still on lru (and we have it frozen) */
2392                 VM_WARN_ON(!PageLRU(head));
2393                 SetPageLRU(tail);
2394                 list_add_tail(&tail->lru, &head->lru);
2395         }
2396 }
2397
2398 static void __split_huge_page_tail(struct page *head, int tail,
2399                 struct lruvec *lruvec, struct list_head *list)
2400 {
2401         struct page *page_tail = head + tail;
2402
2403         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2404
2405         /*
2406          * Clone page flags before unfreezing refcount.
2407          *
2408          * After successful get_page_unless_zero() might follow flags change,
2409          * for example lock_page() which set PG_waiters.
2410          */
2411         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2412         page_tail->flags |= (head->flags &
2413                         ((1L << PG_referenced) |
2414                          (1L << PG_swapbacked) |
2415                          (1L << PG_swapcache) |
2416                          (1L << PG_mlocked) |
2417                          (1L << PG_uptodate) |
2418                          (1L << PG_active) |
2419                          (1L << PG_workingset) |
2420                          (1L << PG_locked) |
2421                          (1L << PG_unevictable) |
2422 #ifdef CONFIG_64BIT
2423                          (1L << PG_arch_2) |
2424 #endif
2425                          (1L << PG_dirty)));
2426
2427         /* ->mapping in first tail page is compound_mapcount */
2428         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2429                         page_tail);
2430         page_tail->mapping = head->mapping;
2431         page_tail->index = head->index + tail;
2432
2433         /* Page flags must be visible before we make the page non-compound. */
2434         smp_wmb();
2435
2436         /*
2437          * Clear PageTail before unfreezing page refcount.
2438          *
2439          * After successful get_page_unless_zero() might follow put_page()
2440          * which needs correct compound_head().
2441          */
2442         clear_compound_head(page_tail);
2443
2444         /* Finally unfreeze refcount. Additional reference from page cache. */
2445         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2446                                           PageSwapCache(head)));
2447
2448         if (page_is_young(head))
2449                 set_page_young(page_tail);
2450         if (page_is_idle(head))
2451                 set_page_idle(page_tail);
2452
2453         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2454
2455         /*
2456          * always add to the tail because some iterators expect new
2457          * pages to show after the currently processed elements - e.g.
2458          * migrate_pages
2459          */
2460         lru_add_page_tail(head, page_tail, lruvec, list);
2461 }
2462
2463 static void __split_huge_page(struct page *page, struct list_head *list,
2464                 pgoff_t end)
2465 {
2466         struct page *head = compound_head(page);
2467         struct lruvec *lruvec;
2468         struct address_space *swap_cache = NULL;
2469         unsigned long offset = 0;
2470         unsigned int nr = thp_nr_pages(head);
2471         int i;
2472
2473         /* complete memcg works before add pages to LRU */
2474         split_page_memcg(head, nr);
2475
2476         if (PageAnon(head) && PageSwapCache(head)) {
2477                 swp_entry_t entry = { .val = page_private(head) };
2478
2479                 offset = swp_offset(entry);
2480                 swap_cache = swap_address_space(entry);
2481                 xa_lock(&swap_cache->i_pages);
2482         }
2483
2484         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2485         lruvec = lock_page_lruvec(head);
2486
2487         for (i = nr - 1; i >= 1; i--) {
2488                 __split_huge_page_tail(head, i, lruvec, list);
2489                 /* Some pages can be beyond i_size: drop them from page cache */
2490                 if (head[i].index >= end) {
2491                         ClearPageDirty(head + i);
2492                         __delete_from_page_cache(head + i, NULL);
2493                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2494                                 shmem_uncharge(head->mapping->host, 1);
2495                         put_page(head + i);
2496                 } else if (!PageAnon(page)) {
2497                         __xa_store(&head->mapping->i_pages, head[i].index,
2498                                         head + i, 0);
2499                 } else if (swap_cache) {
2500                         __xa_store(&swap_cache->i_pages, offset + i,
2501                                         head + i, 0);
2502                 }
2503         }
2504
2505         ClearPageCompound(head);
2506         unlock_page_lruvec(lruvec);
2507         /* Caller disabled irqs, so they are still disabled here */
2508
2509         split_page_owner(head, nr);
2510
2511         /* See comment in __split_huge_page_tail() */
2512         if (PageAnon(head)) {
2513                 /* Additional pin to swap cache */
2514                 if (PageSwapCache(head)) {
2515                         page_ref_add(head, 2);
2516                         xa_unlock(&swap_cache->i_pages);
2517                 } else {
2518                         page_ref_inc(head);
2519                 }
2520         } else {
2521                 /* Additional pin to page cache */
2522                 page_ref_add(head, 2);
2523                 xa_unlock(&head->mapping->i_pages);
2524         }
2525         local_irq_enable();
2526
2527         remap_page(head, nr);
2528
2529         if (PageSwapCache(head)) {
2530                 swp_entry_t entry = { .val = page_private(head) };
2531
2532                 split_swap_cluster(entry);
2533         }
2534
2535         for (i = 0; i < nr; i++) {
2536                 struct page *subpage = head + i;
2537                 if (subpage == page)
2538                         continue;
2539                 unlock_page(subpage);
2540
2541                 /*
2542                  * Subpages may be freed if there wasn't any mapping
2543                  * like if add_to_swap() is running on a lru page that
2544                  * had its mapping zapped. And freeing these pages
2545                  * requires taking the lru_lock so we do the put_page
2546                  * of the tail pages after the split is complete.
2547                  */
2548                 put_page(subpage);
2549         }
2550 }
2551
2552 int total_mapcount(struct page *page)
2553 {
2554         int i, compound, nr, ret;
2555
2556         VM_BUG_ON_PAGE(PageTail(page), page);
2557
2558         if (likely(!PageCompound(page)))
2559                 return atomic_read(&page->_mapcount) + 1;
2560
2561         compound = compound_mapcount(page);
2562         nr = compound_nr(page);
2563         if (PageHuge(page))
2564                 return compound;
2565         ret = compound;
2566         for (i = 0; i < nr; i++)
2567                 ret += atomic_read(&page[i]._mapcount) + 1;
2568         /* File pages has compound_mapcount included in _mapcount */
2569         if (!PageAnon(page))
2570                 return ret - compound * nr;
2571         if (PageDoubleMap(page))
2572                 ret -= nr;
2573         return ret;
2574 }
2575
2576 /*
2577  * This calculates accurately how many mappings a transparent hugepage
2578  * has (unlike page_mapcount() which isn't fully accurate). This full
2579  * accuracy is primarily needed to know if copy-on-write faults can
2580  * reuse the page and change the mapping to read-write instead of
2581  * copying them. At the same time this returns the total_mapcount too.
2582  *
2583  * The function returns the highest mapcount any one of the subpages
2584  * has. If the return value is one, even if different processes are
2585  * mapping different subpages of the transparent hugepage, they can
2586  * all reuse it, because each process is reusing a different subpage.
2587  *
2588  * The total_mapcount is instead counting all virtual mappings of the
2589  * subpages. If the total_mapcount is equal to "one", it tells the
2590  * caller all mappings belong to the same "mm" and in turn the
2591  * anon_vma of the transparent hugepage can become the vma->anon_vma
2592  * local one as no other process may be mapping any of the subpages.
2593  *
2594  * It would be more accurate to replace page_mapcount() with
2595  * page_trans_huge_mapcount(), however we only use
2596  * page_trans_huge_mapcount() in the copy-on-write faults where we
2597  * need full accuracy to avoid breaking page pinning, because
2598  * page_trans_huge_mapcount() is slower than page_mapcount().
2599  */
2600 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2601 {
2602         int i, ret, _total_mapcount, mapcount;
2603
2604         /* hugetlbfs shouldn't call it */
2605         VM_BUG_ON_PAGE(PageHuge(page), page);
2606
2607         if (likely(!PageTransCompound(page))) {
2608                 mapcount = atomic_read(&page->_mapcount) + 1;
2609                 if (total_mapcount)
2610                         *total_mapcount = mapcount;
2611                 return mapcount;
2612         }
2613
2614         page = compound_head(page);
2615
2616         _total_mapcount = ret = 0;
2617         for (i = 0; i < thp_nr_pages(page); i++) {
2618                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2619                 ret = max(ret, mapcount);
2620                 _total_mapcount += mapcount;
2621         }
2622         if (PageDoubleMap(page)) {
2623                 ret -= 1;
2624                 _total_mapcount -= thp_nr_pages(page);
2625         }
2626         mapcount = compound_mapcount(page);
2627         ret += mapcount;
2628         _total_mapcount += mapcount;
2629         if (total_mapcount)
2630                 *total_mapcount = _total_mapcount;
2631         return ret;
2632 }
2633
2634 /* Racy check whether the huge page can be split */
2635 bool can_split_huge_page(struct page *page, int *pextra_pins)
2636 {
2637         int extra_pins;
2638
2639         /* Additional pins from page cache */
2640         if (PageAnon(page))
2641                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2642         else
2643                 extra_pins = thp_nr_pages(page);
2644         if (pextra_pins)
2645                 *pextra_pins = extra_pins;
2646         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2647 }
2648
2649 /*
2650  * This function splits huge page into normal pages. @page can point to any
2651  * subpage of huge page to split. Split doesn't change the position of @page.
2652  *
2653  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2654  * The huge page must be locked.
2655  *
2656  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2657  *
2658  * Both head page and tail pages will inherit mapping, flags, and so on from
2659  * the hugepage.
2660  *
2661  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2662  * they are not mapped.
2663  *
2664  * Returns 0 if the hugepage is split successfully.
2665  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2666  * us.
2667  */
2668 int split_huge_page_to_list(struct page *page, struct list_head *list)
2669 {
2670         struct page *head = compound_head(page);
2671         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2672         struct anon_vma *anon_vma = NULL;
2673         struct address_space *mapping = NULL;
2674         int count, mapcount, extra_pins, ret;
2675         pgoff_t end;
2676
2677         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2678         VM_BUG_ON_PAGE(!PageLocked(head), head);
2679         VM_BUG_ON_PAGE(!PageCompound(head), head);
2680
2681         if (PageWriteback(head))
2682                 return -EBUSY;
2683
2684         if (PageAnon(head)) {
2685                 /*
2686                  * The caller does not necessarily hold an mmap_lock that would
2687                  * prevent the anon_vma disappearing so we first we take a
2688                  * reference to it and then lock the anon_vma for write. This
2689                  * is similar to page_lock_anon_vma_read except the write lock
2690                  * is taken to serialise against parallel split or collapse
2691                  * operations.
2692                  */
2693                 anon_vma = page_get_anon_vma(head);
2694                 if (!anon_vma) {
2695                         ret = -EBUSY;
2696                         goto out;
2697                 }
2698                 end = -1;
2699                 mapping = NULL;
2700                 anon_vma_lock_write(anon_vma);
2701         } else {
2702                 mapping = head->mapping;
2703
2704                 /* Truncated ? */
2705                 if (!mapping) {
2706                         ret = -EBUSY;
2707                         goto out;
2708                 }
2709
2710                 anon_vma = NULL;
2711                 i_mmap_lock_read(mapping);
2712
2713                 /*
2714                  *__split_huge_page() may need to trim off pages beyond EOF:
2715                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2716                  * which cannot be nested inside the page tree lock. So note
2717                  * end now: i_size itself may be changed at any moment, but
2718                  * head page lock is good enough to serialize the trimming.
2719                  */
2720                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2721         }
2722
2723         /*
2724          * Racy check if we can split the page, before unmap_page() will
2725          * split PMDs
2726          */
2727         if (!can_split_huge_page(head, &extra_pins)) {
2728                 ret = -EBUSY;
2729                 goto out_unlock;
2730         }
2731
2732         unmap_page(head);
2733         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2734
2735         /* block interrupt reentry in xa_lock and spinlock */
2736         local_irq_disable();
2737         if (mapping) {
2738                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2739
2740                 /*
2741                  * Check if the head page is present in page cache.
2742                  * We assume all tail are present too, if head is there.
2743                  */
2744                 xa_lock(&mapping->i_pages);
2745                 if (xas_load(&xas) != head)
2746                         goto fail;
2747         }
2748
2749         /* Prevent deferred_split_scan() touching ->_refcount */
2750         spin_lock(&ds_queue->split_queue_lock);
2751         count = page_count(head);
2752         mapcount = total_mapcount(head);
2753         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2754                 if (!list_empty(page_deferred_list(head))) {
2755                         ds_queue->split_queue_len--;
2756                         list_del(page_deferred_list(head));
2757                 }
2758                 spin_unlock(&ds_queue->split_queue_lock);
2759                 if (mapping) {
2760                         int nr = thp_nr_pages(head);
2761
2762                         if (PageSwapBacked(head))
2763                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2764                                                         -nr);
2765                         else
2766                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2767                                                         -nr);
2768                 }
2769
2770                 __split_huge_page(page, list, end);
2771                 ret = 0;
2772         } else {
2773                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2774                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2775                                         mapcount, count);
2776                         if (PageTail(page))
2777                                 dump_page(head, NULL);
2778                         dump_page(page, "total_mapcount(head) > 0");
2779                         BUG();
2780                 }
2781                 spin_unlock(&ds_queue->split_queue_lock);
2782 fail:           if (mapping)
2783                         xa_unlock(&mapping->i_pages);
2784                 local_irq_enable();
2785                 remap_page(head, thp_nr_pages(head));
2786                 ret = -EBUSY;
2787         }
2788
2789 out_unlock:
2790         if (anon_vma) {
2791                 anon_vma_unlock_write(anon_vma);
2792                 put_anon_vma(anon_vma);
2793         }
2794         if (mapping)
2795                 i_mmap_unlock_read(mapping);
2796 out:
2797         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2798         return ret;
2799 }
2800
2801 void free_transhuge_page(struct page *page)
2802 {
2803         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2804         unsigned long flags;
2805
2806         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2807         if (!list_empty(page_deferred_list(page))) {
2808                 ds_queue->split_queue_len--;
2809                 list_del(page_deferred_list(page));
2810         }
2811         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2812         free_compound_page(page);
2813 }
2814
2815 void deferred_split_huge_page(struct page *page)
2816 {
2817         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2818 #ifdef CONFIG_MEMCG
2819         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2820 #endif
2821         unsigned long flags;
2822
2823         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2824
2825         /*
2826          * The try_to_unmap() in page reclaim path might reach here too,
2827          * this may cause a race condition to corrupt deferred split queue.
2828          * And, if page reclaim is already handling the same page, it is
2829          * unnecessary to handle it again in shrinker.
2830          *
2831          * Check PageSwapCache to determine if the page is being
2832          * handled by page reclaim since THP swap would add the page into
2833          * swap cache before calling try_to_unmap().
2834          */
2835         if (PageSwapCache(page))
2836                 return;
2837
2838         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2839         if (list_empty(page_deferred_list(page))) {
2840                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2841                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2842                 ds_queue->split_queue_len++;
2843 #ifdef CONFIG_MEMCG
2844                 if (memcg)
2845                         set_shrinker_bit(memcg, page_to_nid(page),
2846                                          deferred_split_shrinker.id);
2847 #endif
2848         }
2849         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2850 }
2851
2852 static unsigned long deferred_split_count(struct shrinker *shrink,
2853                 struct shrink_control *sc)
2854 {
2855         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2856         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2857
2858 #ifdef CONFIG_MEMCG
2859         if (sc->memcg)
2860                 ds_queue = &sc->memcg->deferred_split_queue;
2861 #endif
2862         return READ_ONCE(ds_queue->split_queue_len);
2863 }
2864
2865 static unsigned long deferred_split_scan(struct shrinker *shrink,
2866                 struct shrink_control *sc)
2867 {
2868         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2869         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2870         unsigned long flags;
2871         LIST_HEAD(list), *pos, *next;
2872         struct page *page;
2873         int split = 0;
2874
2875 #ifdef CONFIG_MEMCG
2876         if (sc->memcg)
2877                 ds_queue = &sc->memcg->deferred_split_queue;
2878 #endif
2879
2880         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2881         /* Take pin on all head pages to avoid freeing them under us */
2882         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2883                 page = list_entry((void *)pos, struct page, mapping);
2884                 page = compound_head(page);
2885                 if (get_page_unless_zero(page)) {
2886                         list_move(page_deferred_list(page), &list);
2887                 } else {
2888                         /* We lost race with put_compound_page() */
2889                         list_del_init(page_deferred_list(page));
2890                         ds_queue->split_queue_len--;
2891                 }
2892                 if (!--sc->nr_to_scan)
2893                         break;
2894         }
2895         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2896
2897         list_for_each_safe(pos, next, &list) {
2898                 page = list_entry((void *)pos, struct page, mapping);
2899                 if (!trylock_page(page))
2900                         goto next;
2901                 /* split_huge_page() removes page from list on success */
2902                 if (!split_huge_page(page))
2903                         split++;
2904                 unlock_page(page);
2905 next:
2906                 put_page(page);
2907         }
2908
2909         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2910         list_splice_tail(&list, &ds_queue->split_queue);
2911         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2912
2913         /*
2914          * Stop shrinker if we didn't split any page, but the queue is empty.
2915          * This can happen if pages were freed under us.
2916          */
2917         if (!split && list_empty(&ds_queue->split_queue))
2918                 return SHRINK_STOP;
2919         return split;
2920 }
2921
2922 static struct shrinker deferred_split_shrinker = {
2923         .count_objects = deferred_split_count,
2924         .scan_objects = deferred_split_scan,
2925         .seeks = DEFAULT_SEEKS,
2926         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2927                  SHRINKER_NONSLAB,
2928 };
2929
2930 #ifdef CONFIG_DEBUG_FS
2931 static void split_huge_pages_all(void)
2932 {
2933         struct zone *zone;
2934         struct page *page;
2935         unsigned long pfn, max_zone_pfn;
2936         unsigned long total = 0, split = 0;
2937
2938         pr_debug("Split all THPs\n");
2939         for_each_populated_zone(zone) {
2940                 max_zone_pfn = zone_end_pfn(zone);
2941                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2942                         if (!pfn_valid(pfn))
2943                                 continue;
2944
2945                         page = pfn_to_page(pfn);
2946                         if (!get_page_unless_zero(page))
2947                                 continue;
2948
2949                         if (zone != page_zone(page))
2950                                 goto next;
2951
2952                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2953                                 goto next;
2954
2955                         total++;
2956                         lock_page(page);
2957                         if (!split_huge_page(page))
2958                                 split++;
2959                         unlock_page(page);
2960 next:
2961                         put_page(page);
2962                         cond_resched();
2963                 }
2964         }
2965
2966         pr_debug("%lu of %lu THP split\n", split, total);
2967 }
2968
2969 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2970 {
2971         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2972                     is_vm_hugetlb_page(vma);
2973 }
2974
2975 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2976                                 unsigned long vaddr_end)
2977 {
2978         int ret = 0;
2979         struct task_struct *task;
2980         struct mm_struct *mm;
2981         unsigned long total = 0, split = 0;
2982         unsigned long addr;
2983
2984         vaddr_start &= PAGE_MASK;
2985         vaddr_end &= PAGE_MASK;
2986
2987         /* Find the task_struct from pid */
2988         rcu_read_lock();
2989         task = find_task_by_vpid(pid);
2990         if (!task) {
2991                 rcu_read_unlock();
2992                 ret = -ESRCH;
2993                 goto out;
2994         }
2995         get_task_struct(task);
2996         rcu_read_unlock();
2997
2998         /* Find the mm_struct */
2999         mm = get_task_mm(task);
3000         put_task_struct(task);
3001
3002         if (!mm) {
3003                 ret = -EINVAL;
3004                 goto out;
3005         }
3006
3007         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3008                  pid, vaddr_start, vaddr_end);
3009
3010         mmap_read_lock(mm);
3011         /*
3012          * always increase addr by PAGE_SIZE, since we could have a PTE page
3013          * table filled with PTE-mapped THPs, each of which is distinct.
3014          */
3015         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3016                 struct vm_area_struct *vma = find_vma(mm, addr);
3017                 unsigned int follflags;
3018                 struct page *page;
3019
3020                 if (!vma || addr < vma->vm_start)
3021                         break;
3022
3023                 /* skip special VMA and hugetlb VMA */
3024                 if (vma_not_suitable_for_thp_split(vma)) {
3025                         addr = vma->vm_end;
3026                         continue;
3027                 }
3028
3029                 /* FOLL_DUMP to ignore special (like zero) pages */
3030                 follflags = FOLL_GET | FOLL_DUMP;
3031                 page = follow_page(vma, addr, follflags);
3032
3033                 if (IS_ERR(page))
3034                         continue;
3035                 if (!page)
3036                         continue;
3037
3038                 if (!is_transparent_hugepage(page))
3039                         goto next;
3040
3041                 total++;
3042                 if (!can_split_huge_page(compound_head(page), NULL))
3043                         goto next;
3044
3045                 if (!trylock_page(page))
3046                         goto next;
3047
3048                 if (!split_huge_page(page))
3049                         split++;
3050
3051                 unlock_page(page);
3052 next:
3053                 put_page(page);
3054                 cond_resched();
3055         }
3056         mmap_read_unlock(mm);
3057         mmput(mm);
3058
3059         pr_debug("%lu of %lu THP split\n", split, total);
3060
3061 out:
3062         return ret;
3063 }
3064
3065 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3066                                 pgoff_t off_end)
3067 {
3068         struct filename *file;
3069         struct file *candidate;
3070         struct address_space *mapping;
3071         int ret = -EINVAL;
3072         pgoff_t index;
3073         int nr_pages = 1;
3074         unsigned long total = 0, split = 0;
3075
3076         file = getname_kernel(file_path);
3077         if (IS_ERR(file))
3078                 return ret;
3079
3080         candidate = file_open_name(file, O_RDONLY, 0);
3081         if (IS_ERR(candidate))
3082                 goto out;
3083
3084         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3085                  file_path, off_start, off_end);
3086
3087         mapping = candidate->f_mapping;
3088
3089         for (index = off_start; index < off_end; index += nr_pages) {
3090                 struct page *fpage = pagecache_get_page(mapping, index,
3091                                                 FGP_ENTRY | FGP_HEAD, 0);
3092
3093                 nr_pages = 1;
3094                 if (xa_is_value(fpage) || !fpage)
3095                         continue;
3096
3097                 if (!is_transparent_hugepage(fpage))
3098                         goto next;
3099
3100                 total++;
3101                 nr_pages = thp_nr_pages(fpage);
3102
3103                 if (!trylock_page(fpage))
3104                         goto next;
3105
3106                 if (!split_huge_page(fpage))
3107                         split++;
3108
3109                 unlock_page(fpage);
3110 next:
3111                 put_page(fpage);
3112                 cond_resched();
3113         }
3114
3115         filp_close(candidate, NULL);
3116         ret = 0;
3117
3118         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3119 out:
3120         putname(file);
3121         return ret;
3122 }
3123
3124 #define MAX_INPUT_BUF_SZ 255
3125
3126 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3127                                 size_t count, loff_t *ppops)
3128 {
3129         static DEFINE_MUTEX(split_debug_mutex);
3130         ssize_t ret;
3131         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3132         char input_buf[MAX_INPUT_BUF_SZ];
3133         int pid;
3134         unsigned long vaddr_start, vaddr_end;
3135
3136         ret = mutex_lock_interruptible(&split_debug_mutex);
3137         if (ret)
3138                 return ret;
3139
3140         ret = -EFAULT;
3141
3142         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3143         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3144                 goto out;
3145
3146         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3147
3148         if (input_buf[0] == '/') {
3149                 char *tok;
3150                 char *buf = input_buf;
3151                 char file_path[MAX_INPUT_BUF_SZ];
3152                 pgoff_t off_start = 0, off_end = 0;
3153                 size_t input_len = strlen(input_buf);
3154
3155                 tok = strsep(&buf, ",");
3156                 if (tok) {
3157                         strncpy(file_path, tok, MAX_INPUT_BUF_SZ);
3158                 } else {
3159                         ret = -EINVAL;
3160                         goto out;
3161                 }
3162
3163                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3164                 if (ret != 2) {
3165                         ret = -EINVAL;
3166                         goto out;
3167                 }
3168                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3169                 if (!ret)
3170                         ret = input_len;
3171
3172                 goto out;
3173         }
3174
3175         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3176         if (ret == 1 && pid == 1) {
3177                 split_huge_pages_all();
3178                 ret = strlen(input_buf);
3179                 goto out;
3180         } else if (ret != 3) {
3181                 ret = -EINVAL;
3182                 goto out;
3183         }
3184
3185         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3186         if (!ret)
3187                 ret = strlen(input_buf);
3188 out:
3189         mutex_unlock(&split_debug_mutex);
3190         return ret;
3191
3192 }
3193
3194 static const struct file_operations split_huge_pages_fops = {
3195         .owner   = THIS_MODULE,
3196         .write   = split_huge_pages_write,
3197         .llseek  = no_llseek,
3198 };
3199
3200 static int __init split_huge_pages_debugfs(void)
3201 {
3202         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3203                             &split_huge_pages_fops);
3204         return 0;
3205 }
3206 late_initcall(split_huge_pages_debugfs);
3207 #endif
3208
3209 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3210 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3211                 struct page *page)
3212 {
3213         struct vm_area_struct *vma = pvmw->vma;
3214         struct mm_struct *mm = vma->vm_mm;
3215         unsigned long address = pvmw->address;
3216         pmd_t pmdval;
3217         swp_entry_t entry;
3218         pmd_t pmdswp;
3219
3220         if (!(pvmw->pmd && !pvmw->pte))
3221                 return;
3222
3223         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3224         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3225         if (pmd_dirty(pmdval))
3226                 set_page_dirty(page);
3227         entry = make_migration_entry(page, pmd_write(pmdval));
3228         pmdswp = swp_entry_to_pmd(entry);
3229         if (pmd_soft_dirty(pmdval))
3230                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3231         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3232         page_remove_rmap(page, true);
3233         put_page(page);
3234 }
3235
3236 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3237 {
3238         struct vm_area_struct *vma = pvmw->vma;
3239         struct mm_struct *mm = vma->vm_mm;
3240         unsigned long address = pvmw->address;
3241         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3242         pmd_t pmde;
3243         swp_entry_t entry;
3244
3245         if (!(pvmw->pmd && !pvmw->pte))
3246                 return;
3247
3248         entry = pmd_to_swp_entry(*pvmw->pmd);
3249         get_page(new);
3250         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3251         if (pmd_swp_soft_dirty(*pvmw->pmd))
3252                 pmde = pmd_mksoft_dirty(pmde);
3253         if (is_write_migration_entry(entry))
3254                 pmde = maybe_pmd_mkwrite(pmde, vma);
3255
3256         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3257         if (PageAnon(new))
3258                 page_add_anon_rmap(new, vma, mmun_start, true);
3259         else
3260                 page_add_file_rmap(new, true);
3261         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3262         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3263                 mlock_vma_page(new);
3264         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3265 }
3266 #endif