mm/huge_memory.c

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