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