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