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_writable_migration_entry(entry)) {
1058                         entry = make_readable_migration_entry(
1059                                                         swp_offset(entry));
1060                         pmd = swp_entry_to_pmd(entry);
1061                         if (pmd_swp_soft_dirty(*src_pmd))
1062                                 pmd = pmd_swp_mksoft_dirty(pmd);
1063                         if (pmd_swp_uffd_wp(*src_pmd))
1064                                 pmd = pmd_swp_mkuffd_wp(pmd);
1065                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1066                 }
1067                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068                 mm_inc_nr_ptes(dst_mm);
1069                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070                 if (!userfaultfd_wp(dst_vma))
1071                         pmd = pmd_swp_clear_uffd_wp(pmd);
1072                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1073                 ret = 0;
1074                 goto out_unlock;
1075         }
1076 #endif
1077
1078         if (unlikely(!pmd_trans_huge(pmd))) {
1079                 pte_free(dst_mm, pgtable);
1080                 goto out_unlock;
1081         }
1082         /*
1083          * When page table lock is held, the huge zero pmd should not be
1084          * under splitting since we don't split the page itself, only pmd to
1085          * a page table.
1086          */
1087         if (is_huge_zero_pmd(pmd)) {
1088                 /*
1089                  * get_huge_zero_page() will never allocate a new page here,
1090                  * since we already have a zero page to copy. It just takes a
1091                  * reference.
1092                  */
1093                 mm_get_huge_zero_page(dst_mm);
1094                 goto out_zero_page;
1095         }
1096
1097         src_page = pmd_page(pmd);
1098         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1099
1100         /*
1101          * If this page is a potentially pinned page, split and retry the fault
1102          * with smaller page size.  Normally this should not happen because the
1103          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1104          * best effort that the pinned pages won't be replaced by another
1105          * random page during the coming copy-on-write.
1106          */
1107         if (unlikely(page_needs_cow_for_dma(src_vma, src_page))) {
1108                 pte_free(dst_mm, pgtable);
1109                 spin_unlock(src_ptl);
1110                 spin_unlock(dst_ptl);
1111                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1112                 return -EAGAIN;
1113         }
1114
1115         get_page(src_page);
1116         page_dup_rmap(src_page, true);
1117         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118 out_zero_page:
1119         mm_inc_nr_ptes(dst_mm);
1120         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1121         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1122         if (!userfaultfd_wp(dst_vma))
1123                 pmd = pmd_clear_uffd_wp(pmd);
1124         pmd = pmd_mkold(pmd_wrprotect(pmd));
1125         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1126
1127         ret = 0;
1128 out_unlock:
1129         spin_unlock(src_ptl);
1130         spin_unlock(dst_ptl);
1131 out:
1132         return ret;
1133 }
1134
1135 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1136 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1137                 pud_t *pud, int flags)
1138 {
1139         pud_t _pud;
1140
1141         _pud = pud_mkyoung(*pud);
1142         if (flags & FOLL_WRITE)
1143                 _pud = pud_mkdirty(_pud);
1144         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1145                                 pud, _pud, flags & FOLL_WRITE))
1146                 update_mmu_cache_pud(vma, addr, pud);
1147 }
1148
1149 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1150                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1151 {
1152         unsigned long pfn = pud_pfn(*pud);
1153         struct mm_struct *mm = vma->vm_mm;
1154         struct page *page;
1155
1156         assert_spin_locked(pud_lockptr(mm, pud));
1157
1158         if (flags & FOLL_WRITE && !pud_write(*pud))
1159                 return NULL;
1160
1161         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1162         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1163                          (FOLL_PIN | FOLL_GET)))
1164                 return NULL;
1165
1166         if (pud_present(*pud) && pud_devmap(*pud))
1167                 /* pass */;
1168         else
1169                 return NULL;
1170
1171         if (flags & FOLL_TOUCH)
1172                 touch_pud(vma, addr, pud, flags);
1173
1174         /*
1175          * device mapped pages can only be returned if the
1176          * caller will manage the page reference count.
1177          *
1178          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1179          */
1180         if (!(flags & (FOLL_GET | FOLL_PIN)))
1181                 return ERR_PTR(-EEXIST);
1182
1183         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1184         *pgmap = get_dev_pagemap(pfn, *pgmap);
1185         if (!*pgmap)
1186                 return ERR_PTR(-EFAULT);
1187         page = pfn_to_page(pfn);
1188         if (!try_grab_page(page, flags))
1189                 page = ERR_PTR(-ENOMEM);
1190
1191         return page;
1192 }
1193
1194 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1195                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1196                   struct vm_area_struct *vma)
1197 {
1198         spinlock_t *dst_ptl, *src_ptl;
1199         pud_t pud;
1200         int ret;
1201
1202         dst_ptl = pud_lock(dst_mm, dst_pud);
1203         src_ptl = pud_lockptr(src_mm, src_pud);
1204         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1205
1206         ret = -EAGAIN;
1207         pud = *src_pud;
1208         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1209                 goto out_unlock;
1210
1211         /*
1212          * When page table lock is held, the huge zero pud should not be
1213          * under splitting since we don't split the page itself, only pud to
1214          * a page table.
1215          */
1216         if (is_huge_zero_pud(pud)) {
1217                 /* No huge zero pud yet */
1218         }
1219
1220         /* Please refer to comments in copy_huge_pmd() */
1221         if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1222                 spin_unlock(src_ptl);
1223                 spin_unlock(dst_ptl);
1224                 __split_huge_pud(vma, src_pud, addr);
1225                 return -EAGAIN;
1226         }
1227
1228         pudp_set_wrprotect(src_mm, addr, src_pud);
1229         pud = pud_mkold(pud_wrprotect(pud));
1230         set_pud_at(dst_mm, addr, dst_pud, pud);
1231
1232         ret = 0;
1233 out_unlock:
1234         spin_unlock(src_ptl);
1235         spin_unlock(dst_ptl);
1236         return ret;
1237 }
1238
1239 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1240 {
1241         pud_t entry;
1242         unsigned long haddr;
1243         bool write = vmf->flags & FAULT_FLAG_WRITE;
1244
1245         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1246         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1247                 goto unlock;
1248
1249         entry = pud_mkyoung(orig_pud);
1250         if (write)
1251                 entry = pud_mkdirty(entry);
1252         haddr = vmf->address & HPAGE_PUD_MASK;
1253         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1254                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1255
1256 unlock:
1257         spin_unlock(vmf->ptl);
1258 }
1259 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1260
1261 void huge_pmd_set_accessed(struct vm_fault *vmf)
1262 {
1263         pmd_t entry;
1264         unsigned long haddr;
1265         bool write = vmf->flags & FAULT_FLAG_WRITE;
1266         pmd_t orig_pmd = vmf->orig_pmd;
1267
1268         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1269         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1270                 goto unlock;
1271
1272         entry = pmd_mkyoung(orig_pmd);
1273         if (write)
1274                 entry = pmd_mkdirty(entry);
1275         haddr = vmf->address & HPAGE_PMD_MASK;
1276         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1277                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1278
1279 unlock:
1280         spin_unlock(vmf->ptl);
1281 }
1282
1283 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1284 {
1285         struct vm_area_struct *vma = vmf->vma;
1286         struct page *page;
1287         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1288         pmd_t orig_pmd = vmf->orig_pmd;
1289
1290         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1291         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1292
1293         if (is_huge_zero_pmd(orig_pmd))
1294                 goto fallback;
1295
1296         spin_lock(vmf->ptl);
1297
1298         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1299                 spin_unlock(vmf->ptl);
1300                 return 0;
1301         }
1302
1303         page = pmd_page(orig_pmd);
1304         VM_BUG_ON_PAGE(!PageHead(page), page);
1305
1306         /* Lock page for reuse_swap_page() */
1307         if (!trylock_page(page)) {
1308                 get_page(page);
1309                 spin_unlock(vmf->ptl);
1310                 lock_page(page);
1311                 spin_lock(vmf->ptl);
1312                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1313                         spin_unlock(vmf->ptl);
1314                         unlock_page(page);
1315                         put_page(page);
1316                         return 0;
1317                 }
1318                 put_page(page);
1319         }
1320
1321         /*
1322          * We can only reuse the page if nobody else maps the huge page or it's
1323          * part.
1324          */
1325         if (reuse_swap_page(page, NULL)) {
1326                 pmd_t entry;
1327                 entry = pmd_mkyoung(orig_pmd);
1328                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1329                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1330                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1331                 unlock_page(page);
1332                 spin_unlock(vmf->ptl);
1333                 return VM_FAULT_WRITE;
1334         }
1335
1336         unlock_page(page);
1337         spin_unlock(vmf->ptl);
1338 fallback:
1339         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1340         return VM_FAULT_FALLBACK;
1341 }
1342
1343 /*
1344  * FOLL_FORCE can write to even unwritable pmd's, but only
1345  * after we've gone through a COW cycle and they are dirty.
1346  */
1347 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1348 {
1349         return pmd_write(pmd) ||
1350                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1351 }
1352
1353 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1354                                    unsigned long addr,
1355                                    pmd_t *pmd,
1356                                    unsigned int flags)
1357 {
1358         struct mm_struct *mm = vma->vm_mm;
1359         struct page *page = NULL;
1360
1361         assert_spin_locked(pmd_lockptr(mm, pmd));
1362
1363         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1364                 goto out;
1365
1366         /* Avoid dumping huge zero page */
1367         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1368                 return ERR_PTR(-EFAULT);
1369
1370         /* Full NUMA hinting faults to serialise migration in fault paths */
1371         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1372                 goto out;
1373
1374         page = pmd_page(*pmd);
1375         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1376
1377         if (!try_grab_page(page, flags))
1378                 return ERR_PTR(-ENOMEM);
1379
1380         if (flags & FOLL_TOUCH)
1381                 touch_pmd(vma, addr, pmd, flags);
1382
1383         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1384                 /*
1385                  * We don't mlock() pte-mapped THPs. This way we can avoid
1386                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1387                  *
1388                  * For anon THP:
1389                  *
1390                  * In most cases the pmd is the only mapping of the page as we
1391                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1392                  * writable private mappings in populate_vma_page_range().
1393                  *
1394                  * The only scenario when we have the page shared here is if we
1395                  * mlocking read-only mapping shared over fork(). We skip
1396                  * mlocking such pages.
1397                  *
1398                  * For file THP:
1399                  *
1400                  * We can expect PageDoubleMap() to be stable under page lock:
1401                  * for file pages we set it in page_add_file_rmap(), which
1402                  * requires page to be locked.
1403                  */
1404
1405                 if (PageAnon(page) && compound_mapcount(page) != 1)
1406                         goto skip_mlock;
1407                 if (PageDoubleMap(page) || !page->mapping)
1408                         goto skip_mlock;
1409                 if (!trylock_page(page))
1410                         goto skip_mlock;
1411                 if (page->mapping && !PageDoubleMap(page))
1412                         mlock_vma_page(page);
1413                 unlock_page(page);
1414         }
1415 skip_mlock:
1416         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1417         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1418
1419 out:
1420         return page;
1421 }
1422
1423 /* NUMA hinting page fault entry point for trans huge pmds */
1424 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1425 {
1426         struct vm_area_struct *vma = vmf->vma;
1427         pmd_t oldpmd = vmf->orig_pmd;
1428         pmd_t pmd;
1429         struct page *page;
1430         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1431         int page_nid = NUMA_NO_NODE;
1432         int target_nid, last_cpupid = -1;
1433         bool migrated = false;
1434         bool was_writable = pmd_savedwrite(oldpmd);
1435         int flags = 0;
1436
1437         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1438         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1439                 spin_unlock(vmf->ptl);
1440                 goto out;
1441         }
1442
1443         /*
1444          * Since we took the NUMA fault, we must have observed the !accessible
1445          * bit. Make sure all other CPUs agree with that, to avoid them
1446          * modifying the page we're about to migrate.
1447          *
1448          * Must be done under PTL such that we'll observe the relevant
1449          * inc_tlb_flush_pending().
1450          *
1451          * We are not sure a pending tlb flush here is for a huge page
1452          * mapping or not. Hence use the tlb range variant
1453          */
1454         if (mm_tlb_flush_pending(vma->vm_mm)) {
1455                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1456                 /*
1457                  * change_huge_pmd() released the pmd lock before
1458                  * invalidating the secondary MMUs sharing the primary
1459                  * MMU pagetables (with ->invalidate_range()). The
1460                  * mmu_notifier_invalidate_range_end() (which
1461                  * internally calls ->invalidate_range()) in
1462                  * change_pmd_range() will run after us, so we can't
1463                  * rely on it here and we need an explicit invalidate.
1464                  */
1465                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1466                                               haddr + HPAGE_PMD_SIZE);
1467         }
1468
1469         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1470         page = vm_normal_page_pmd(vma, haddr, pmd);
1471         if (!page)
1472                 goto out_map;
1473
1474         /* See similar comment in do_numa_page for explanation */
1475         if (!was_writable)
1476                 flags |= TNF_NO_GROUP;
1477
1478         page_nid = page_to_nid(page);
1479         last_cpupid = page_cpupid_last(page);
1480         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1481                                        &flags);
1482
1483         if (target_nid == NUMA_NO_NODE) {
1484                 put_page(page);
1485                 goto out_map;
1486         }
1487
1488         spin_unlock(vmf->ptl);
1489
1490         migrated = migrate_misplaced_page(page, vma, target_nid);
1491         if (migrated) {
1492                 flags |= TNF_MIGRATED;
1493                 page_nid = target_nid;
1494         } else {
1495                 flags |= TNF_MIGRATE_FAIL;
1496                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1497                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1498                         spin_unlock(vmf->ptl);
1499                         goto out;
1500                 }
1501                 goto out_map;
1502         }
1503
1504 out:
1505         if (page_nid != NUMA_NO_NODE)
1506                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1507                                 flags);
1508
1509         return 0;
1510
1511 out_map:
1512         /* Restore the PMD */
1513         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1514         pmd = pmd_mkyoung(pmd);
1515         if (was_writable)
1516                 pmd = pmd_mkwrite(pmd);
1517         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1518         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1519         spin_unlock(vmf->ptl);
1520         goto out;
1521 }
1522
1523 /*
1524  * Return true if we do MADV_FREE successfully on entire pmd page.
1525  * Otherwise, return false.
1526  */
1527 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1528                 pmd_t *pmd, unsigned long addr, unsigned long next)
1529 {
1530         spinlock_t *ptl;
1531         pmd_t orig_pmd;
1532         struct page *page;
1533         struct mm_struct *mm = tlb->mm;
1534         bool ret = false;
1535
1536         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1537
1538         ptl = pmd_trans_huge_lock(pmd, vma);
1539         if (!ptl)
1540                 goto out_unlocked;
1541
1542         orig_pmd = *pmd;
1543         if (is_huge_zero_pmd(orig_pmd))
1544                 goto out;
1545
1546         if (unlikely(!pmd_present(orig_pmd))) {
1547                 VM_BUG_ON(thp_migration_supported() &&
1548                                   !is_pmd_migration_entry(orig_pmd));
1549                 goto out;
1550         }
1551
1552         page = pmd_page(orig_pmd);
1553         /*
1554          * If other processes are mapping this page, we couldn't discard
1555          * the page unless they all do MADV_FREE so let's skip the page.
1556          */
1557         if (total_mapcount(page) != 1)
1558                 goto out;
1559
1560         if (!trylock_page(page))
1561                 goto out;
1562
1563         /*
1564          * If user want to discard part-pages of THP, split it so MADV_FREE
1565          * will deactivate only them.
1566          */
1567         if (next - addr != HPAGE_PMD_SIZE) {
1568                 get_page(page);
1569                 spin_unlock(ptl);
1570                 split_huge_page(page);
1571                 unlock_page(page);
1572                 put_page(page);
1573                 goto out_unlocked;
1574         }
1575
1576         if (PageDirty(page))
1577                 ClearPageDirty(page);
1578         unlock_page(page);
1579
1580         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1581                 pmdp_invalidate(vma, addr, pmd);
1582                 orig_pmd = pmd_mkold(orig_pmd);
1583                 orig_pmd = pmd_mkclean(orig_pmd);
1584
1585                 set_pmd_at(mm, addr, pmd, orig_pmd);
1586                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1587         }
1588
1589         mark_page_lazyfree(page);
1590         ret = true;
1591 out:
1592         spin_unlock(ptl);
1593 out_unlocked:
1594         return ret;
1595 }
1596
1597 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1598 {
1599         pgtable_t pgtable;
1600
1601         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1602         pte_free(mm, pgtable);
1603         mm_dec_nr_ptes(mm);
1604 }
1605
1606 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1607                  pmd_t *pmd, unsigned long addr)
1608 {
1609         pmd_t orig_pmd;
1610         spinlock_t *ptl;
1611
1612         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1613
1614         ptl = __pmd_trans_huge_lock(pmd, vma);
1615         if (!ptl)
1616                 return 0;
1617         /*
1618          * For architectures like ppc64 we look at deposited pgtable
1619          * when calling pmdp_huge_get_and_clear. So do the
1620          * pgtable_trans_huge_withdraw after finishing pmdp related
1621          * operations.
1622          */
1623         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1624                                                 tlb->fullmm);
1625         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1626         if (vma_is_special_huge(vma)) {
1627                 if (arch_needs_pgtable_deposit())
1628                         zap_deposited_table(tlb->mm, pmd);
1629                 spin_unlock(ptl);
1630         } else if (is_huge_zero_pmd(orig_pmd)) {
1631                 zap_deposited_table(tlb->mm, pmd);
1632                 spin_unlock(ptl);
1633         } else {
1634                 struct page *page = NULL;
1635                 int flush_needed = 1;
1636
1637                 if (pmd_present(orig_pmd)) {
1638                         page = pmd_page(orig_pmd);
1639                         page_remove_rmap(page, true);
1640                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1641                         VM_BUG_ON_PAGE(!PageHead(page), page);
1642                 } else if (thp_migration_supported()) {
1643                         swp_entry_t entry;
1644
1645                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1646                         entry = pmd_to_swp_entry(orig_pmd);
1647                         page = pfn_swap_entry_to_page(entry);
1648                         flush_needed = 0;
1649                 } else
1650                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1651
1652                 if (PageAnon(page)) {
1653                         zap_deposited_table(tlb->mm, pmd);
1654                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1655                 } else {
1656                         if (arch_needs_pgtable_deposit())
1657                                 zap_deposited_table(tlb->mm, pmd);
1658                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1659                 }
1660
1661                 spin_unlock(ptl);
1662                 if (flush_needed)
1663                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1664         }
1665         return 1;
1666 }
1667
1668 #ifndef pmd_move_must_withdraw
1669 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1670                                          spinlock_t *old_pmd_ptl,
1671                                          struct vm_area_struct *vma)
1672 {
1673         /*
1674          * With split pmd lock we also need to move preallocated
1675          * PTE page table if new_pmd is on different PMD page table.
1676          *
1677          * We also don't deposit and withdraw tables for file pages.
1678          */
1679         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1680 }
1681 #endif
1682
1683 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1684 {
1685 #ifdef CONFIG_MEM_SOFT_DIRTY
1686         if (unlikely(is_pmd_migration_entry(pmd)))
1687                 pmd = pmd_swp_mksoft_dirty(pmd);
1688         else if (pmd_present(pmd))
1689                 pmd = pmd_mksoft_dirty(pmd);
1690 #endif
1691         return pmd;
1692 }
1693
1694 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1695                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1696 {
1697         spinlock_t *old_ptl, *new_ptl;
1698         pmd_t pmd;
1699         struct mm_struct *mm = vma->vm_mm;
1700         bool force_flush = false;
1701
1702         /*
1703          * The destination pmd shouldn't be established, free_pgtables()
1704          * should have release it.
1705          */
1706         if (WARN_ON(!pmd_none(*new_pmd))) {
1707                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1708                 return false;
1709         }
1710
1711         /*
1712          * We don't have to worry about the ordering of src and dst
1713          * ptlocks because exclusive mmap_lock prevents deadlock.
1714          */
1715         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1716         if (old_ptl) {
1717                 new_ptl = pmd_lockptr(mm, new_pmd);
1718                 if (new_ptl != old_ptl)
1719                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1720                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1721                 if (pmd_present(pmd))
1722                         force_flush = true;
1723                 VM_BUG_ON(!pmd_none(*new_pmd));
1724
1725                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1726                         pgtable_t pgtable;
1727                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1728                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1729                 }
1730                 pmd = move_soft_dirty_pmd(pmd);
1731                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1732                 if (force_flush)
1733                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1734                 if (new_ptl != old_ptl)
1735                         spin_unlock(new_ptl);
1736                 spin_unlock(old_ptl);
1737                 return true;
1738         }
1739         return false;
1740 }
1741
1742 /*
1743  * Returns
1744  *  - 0 if PMD could not be locked
1745  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1746  *      or if prot_numa but THP migration is not supported
1747  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1748  */
1749 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1750                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1751 {
1752         struct mm_struct *mm = vma->vm_mm;
1753         spinlock_t *ptl;
1754         pmd_t entry;
1755         bool preserve_write;
1756         int ret;
1757         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1758         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1759         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1760
1761         if (prot_numa && !thp_migration_supported())
1762                 return 1;
1763
1764         ptl = __pmd_trans_huge_lock(pmd, vma);
1765         if (!ptl)
1766                 return 0;
1767
1768         preserve_write = prot_numa && pmd_write(*pmd);
1769         ret = 1;
1770
1771 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1772         if (is_swap_pmd(*pmd)) {
1773                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1774
1775                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1776                 if (is_writable_migration_entry(entry)) {
1777                         pmd_t newpmd;
1778                         /*
1779                          * A protection check is difficult so
1780                          * just be safe and disable write
1781                          */
1782                         entry = make_readable_migration_entry(
1783                                                         swp_offset(entry));
1784                         newpmd = swp_entry_to_pmd(entry);
1785                         if (pmd_swp_soft_dirty(*pmd))
1786                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1787                         if (pmd_swp_uffd_wp(*pmd))
1788                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1789                         set_pmd_at(mm, addr, pmd, newpmd);
1790                 }
1791                 goto unlock;
1792         }
1793 #endif
1794
1795         /*
1796          * Avoid trapping faults against the zero page. The read-only
1797          * data is likely to be read-cached on the local CPU and
1798          * local/remote hits to the zero page are not interesting.
1799          */
1800         if (prot_numa && is_huge_zero_pmd(*pmd))
1801                 goto unlock;
1802
1803         if (prot_numa && pmd_protnone(*pmd))
1804                 goto unlock;
1805
1806         /*
1807          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1808          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1809          * which is also under mmap_read_lock(mm):
1810          *
1811          *      CPU0:                           CPU1:
1812          *                              change_huge_pmd(prot_numa=1)
1813          *                               pmdp_huge_get_and_clear_notify()
1814          * madvise_dontneed()
1815          *  zap_pmd_range()
1816          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1817          *   // skip the pmd
1818          *                               set_pmd_at();
1819          *                               // pmd is re-established
1820          *
1821          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1822          * which may break userspace.
1823          *
1824          * pmdp_invalidate() is required to make sure we don't miss
1825          * dirty/young flags set by hardware.
1826          */
1827         entry = pmdp_invalidate(vma, addr, pmd);
1828
1829         entry = pmd_modify(entry, newprot);
1830         if (preserve_write)
1831                 entry = pmd_mk_savedwrite(entry);
1832         if (uffd_wp) {
1833                 entry = pmd_wrprotect(entry);
1834                 entry = pmd_mkuffd_wp(entry);
1835         } else if (uffd_wp_resolve) {
1836                 /*
1837                  * Leave the write bit to be handled by PF interrupt
1838                  * handler, then things like COW could be properly
1839                  * handled.
1840                  */
1841                 entry = pmd_clear_uffd_wp(entry);
1842         }
1843         ret = HPAGE_PMD_NR;
1844         set_pmd_at(mm, addr, pmd, entry);
1845         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1846 unlock:
1847         spin_unlock(ptl);
1848         return ret;
1849 }
1850
1851 /*
1852  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1853  *
1854  * Note that if it returns page table lock pointer, this routine returns without
1855  * unlocking page table lock. So callers must unlock it.
1856  */
1857 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1858 {
1859         spinlock_t *ptl;
1860         ptl = pmd_lock(vma->vm_mm, pmd);
1861         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1862                         pmd_devmap(*pmd)))
1863                 return ptl;
1864         spin_unlock(ptl);
1865         return NULL;
1866 }
1867
1868 /*
1869  * Returns true if a given pud maps a thp, false otherwise.
1870  *
1871  * Note that if it returns true, this routine returns without unlocking page
1872  * table lock. So callers must unlock it.
1873  */
1874 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1875 {
1876         spinlock_t *ptl;
1877
1878         ptl = pud_lock(vma->vm_mm, pud);
1879         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1880                 return ptl;
1881         spin_unlock(ptl);
1882         return NULL;
1883 }
1884
1885 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1886 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1887                  pud_t *pud, unsigned long addr)
1888 {
1889         spinlock_t *ptl;
1890
1891         ptl = __pud_trans_huge_lock(pud, vma);
1892         if (!ptl)
1893                 return 0;
1894         /*
1895          * For architectures like ppc64 we look at deposited pgtable
1896          * when calling pudp_huge_get_and_clear. So do the
1897          * pgtable_trans_huge_withdraw after finishing pudp related
1898          * operations.
1899          */
1900         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1901         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1902         if (vma_is_special_huge(vma)) {
1903                 spin_unlock(ptl);
1904                 /* No zero page support yet */
1905         } else {
1906                 /* No support for anonymous PUD pages yet */
1907                 BUG();
1908         }
1909         return 1;
1910 }
1911
1912 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1913                 unsigned long haddr)
1914 {
1915         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1916         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1917         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1918         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1919
1920         count_vm_event(THP_SPLIT_PUD);
1921
1922         pudp_huge_clear_flush_notify(vma, haddr, pud);
1923 }
1924
1925 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1926                 unsigned long address)
1927 {
1928         spinlock_t *ptl;
1929         struct mmu_notifier_range range;
1930
1931         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1932                                 address & HPAGE_PUD_MASK,
1933                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1934         mmu_notifier_invalidate_range_start(&range);
1935         ptl = pud_lock(vma->vm_mm, pud);
1936         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1937                 goto out;
1938         __split_huge_pud_locked(vma, pud, range.start);
1939
1940 out:
1941         spin_unlock(ptl);
1942         /*
1943          * No need to double call mmu_notifier->invalidate_range() callback as
1944          * the above pudp_huge_clear_flush_notify() did already call it.
1945          */
1946         mmu_notifier_invalidate_range_only_end(&range);
1947 }
1948 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1949
1950 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1951                 unsigned long haddr, pmd_t *pmd)
1952 {
1953         struct mm_struct *mm = vma->vm_mm;
1954         pgtable_t pgtable;
1955         pmd_t _pmd;
1956         int i;
1957
1958         /*
1959          * Leave pmd empty until pte is filled note that it is fine to delay
1960          * notification until mmu_notifier_invalidate_range_end() as we are
1961          * replacing a zero pmd write protected page with a zero pte write
1962          * protected page.
1963          *
1964          * See Documentation/vm/mmu_notifier.rst
1965          */
1966         pmdp_huge_clear_flush(vma, haddr, pmd);
1967
1968         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1969         pmd_populate(mm, &_pmd, pgtable);
1970
1971         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1972                 pte_t *pte, entry;
1973                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1974                 entry = pte_mkspecial(entry);
1975                 pte = pte_offset_map(&_pmd, haddr);
1976                 VM_BUG_ON(!pte_none(*pte));
1977                 set_pte_at(mm, haddr, pte, entry);
1978                 pte_unmap(pte);
1979         }
1980         smp_wmb(); /* make pte visible before pmd */
1981         pmd_populate(mm, pmd, pgtable);
1982 }
1983
1984 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1985                 unsigned long haddr, bool freeze)
1986 {
1987         struct mm_struct *mm = vma->vm_mm;
1988         struct page *page;
1989         pgtable_t pgtable;
1990         pmd_t old_pmd, _pmd;
1991         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1992         unsigned long addr;
1993         int i;
1994
1995         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1996         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1997         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1998         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
1999                                 && !pmd_devmap(*pmd));
2000
2001         count_vm_event(THP_SPLIT_PMD);
2002
2003         if (!vma_is_anonymous(vma)) {
2004                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2005                 /*
2006                  * We are going to unmap this huge page. So
2007                  * just go ahead and zap it
2008                  */
2009                 if (arch_needs_pgtable_deposit())
2010                         zap_deposited_table(mm, pmd);
2011                 if (vma_is_special_huge(vma))
2012                         return;
2013                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2014                         swp_entry_t entry;
2015
2016                         entry = pmd_to_swp_entry(old_pmd);
2017                         page = pfn_swap_entry_to_page(entry);
2018                 } else {
2019                         page = pmd_page(old_pmd);
2020                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2021                                 set_page_dirty(page);
2022                         if (!PageReferenced(page) && pmd_young(old_pmd))
2023                                 SetPageReferenced(page);
2024                         page_remove_rmap(page, true);
2025                         put_page(page);
2026                 }
2027                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2028                 return;
2029         }
2030
2031         if (is_huge_zero_pmd(*pmd)) {
2032                 /*
2033                  * FIXME: Do we want to invalidate secondary mmu by calling
2034                  * mmu_notifier_invalidate_range() see comments below inside
2035                  * __split_huge_pmd() ?
2036                  *
2037                  * We are going from a zero huge page write protected to zero
2038                  * small page also write protected so it does not seems useful
2039                  * to invalidate secondary mmu at this time.
2040                  */
2041                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2042         }
2043
2044         /*
2045          * Up to this point the pmd is present and huge and userland has the
2046          * whole access to the hugepage during the split (which happens in
2047          * place). If we overwrite the pmd with the not-huge version pointing
2048          * to the pte here (which of course we could if all CPUs were bug
2049          * free), userland could trigger a small page size TLB miss on the
2050          * small sized TLB while the hugepage TLB entry is still established in
2051          * the huge TLB. Some CPU doesn't like that.
2052          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2053          * 383 on page 105. Intel should be safe but is also warns that it's
2054          * only safe if the permission and cache attributes of the two entries
2055          * loaded in the two TLB is identical (which should be the case here).
2056          * But it is generally safer to never allow small and huge TLB entries
2057          * for the same virtual address to be loaded simultaneously. So instead
2058          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2059          * current pmd notpresent (atomically because here the pmd_trans_huge
2060          * must remain set at all times on the pmd until the split is complete
2061          * for this pmd), then we flush the SMP TLB and finally we write the
2062          * non-huge version of the pmd entry with pmd_populate.
2063          */
2064         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2065
2066         pmd_migration = is_pmd_migration_entry(old_pmd);
2067         if (unlikely(pmd_migration)) {
2068                 swp_entry_t entry;
2069
2070                 entry = pmd_to_swp_entry(old_pmd);
2071                 page = pfn_swap_entry_to_page(entry);
2072                 write = is_writable_migration_entry(entry);
2073                 young = false;
2074                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2075                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2076         } else {
2077                 page = pmd_page(old_pmd);
2078                 if (pmd_dirty(old_pmd))
2079                         SetPageDirty(page);
2080                 write = pmd_write(old_pmd);
2081                 young = pmd_young(old_pmd);
2082                 soft_dirty = pmd_soft_dirty(old_pmd);
2083                 uffd_wp = pmd_uffd_wp(old_pmd);
2084         }
2085         VM_BUG_ON_PAGE(!page_count(page), page);
2086         page_ref_add(page, HPAGE_PMD_NR - 1);
2087
2088         /*
2089          * Withdraw the table only after we mark the pmd entry invalid.
2090          * This's critical for some architectures (Power).
2091          */
2092         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2093         pmd_populate(mm, &_pmd, pgtable);
2094
2095         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2096                 pte_t entry, *pte;
2097                 /*
2098                  * Note that NUMA hinting access restrictions are not
2099                  * transferred to avoid any possibility of altering
2100                  * permissions across VMAs.
2101                  */
2102                 if (freeze || pmd_migration) {
2103                         swp_entry_t swp_entry;
2104                         if (write)
2105                                 swp_entry = make_writable_migration_entry(
2106                                                         page_to_pfn(page + i));
2107                         else
2108                                 swp_entry = make_readable_migration_entry(
2109                                                         page_to_pfn(page + i));
2110                         entry = swp_entry_to_pte(swp_entry);
2111                         if (soft_dirty)
2112                                 entry = pte_swp_mksoft_dirty(entry);
2113                         if (uffd_wp)
2114                                 entry = pte_swp_mkuffd_wp(entry);
2115                 } else {
2116                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2117                         entry = maybe_mkwrite(entry, vma);
2118                         if (!write)
2119                                 entry = pte_wrprotect(entry);
2120                         if (!young)
2121                                 entry = pte_mkold(entry);
2122                         if (soft_dirty)
2123                                 entry = pte_mksoft_dirty(entry);
2124                         if (uffd_wp)
2125                                 entry = pte_mkuffd_wp(entry);
2126                 }
2127                 pte = pte_offset_map(&_pmd, addr);
2128                 BUG_ON(!pte_none(*pte));
2129                 set_pte_at(mm, addr, pte, entry);
2130                 if (!pmd_migration)
2131                         atomic_inc(&page[i]._mapcount);
2132                 pte_unmap(pte);
2133         }
2134
2135         if (!pmd_migration) {
2136                 /*
2137                  * Set PG_double_map before dropping compound_mapcount to avoid
2138                  * false-negative page_mapped().
2139                  */
2140                 if (compound_mapcount(page) > 1 &&
2141                     !TestSetPageDoubleMap(page)) {
2142                         for (i = 0; i < HPAGE_PMD_NR; i++)
2143                                 atomic_inc(&page[i]._mapcount);
2144                 }
2145
2146                 lock_page_memcg(page);
2147                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2148                         /* Last compound_mapcount is gone. */
2149                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2150                                                 -HPAGE_PMD_NR);
2151                         if (TestClearPageDoubleMap(page)) {
2152                                 /* No need in mapcount reference anymore */
2153                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2154                                         atomic_dec(&page[i]._mapcount);
2155                         }
2156                 }
2157                 unlock_page_memcg(page);
2158         }
2159
2160         smp_wmb(); /* make pte visible before pmd */
2161         pmd_populate(mm, pmd, pgtable);
2162
2163         if (freeze) {
2164                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2165                         page_remove_rmap(page + i, false);
2166                         put_page(page + i);
2167                 }
2168         }
2169 }
2170
2171 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2172                 unsigned long address, bool freeze, struct page *page)
2173 {
2174         spinlock_t *ptl;
2175         struct mmu_notifier_range range;
2176         bool do_unlock_page = false;
2177         pmd_t _pmd;
2178
2179         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2180                                 address & HPAGE_PMD_MASK,
2181                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2182         mmu_notifier_invalidate_range_start(&range);
2183         ptl = pmd_lock(vma->vm_mm, pmd);
2184
2185         /*
2186          * If caller asks to setup a migration entries, we need a page to check
2187          * pmd against. Otherwise we can end up replacing wrong page.
2188          */
2189         VM_BUG_ON(freeze && !page);
2190         if (page) {
2191                 VM_WARN_ON_ONCE(!PageLocked(page));
2192                 if (page != pmd_page(*pmd))
2193                         goto out;
2194         }
2195
2196 repeat:
2197         if (pmd_trans_huge(*pmd)) {
2198                 if (!page) {
2199                         page = pmd_page(*pmd);
2200                         /*
2201                          * An anonymous page must be locked, to ensure that a
2202                          * concurrent reuse_swap_page() sees stable mapcount;
2203                          * but reuse_swap_page() is not used on shmem or file,
2204                          * and page lock must not be taken when zap_pmd_range()
2205                          * calls __split_huge_pmd() while i_mmap_lock is held.
2206                          */
2207                         if (PageAnon(page)) {
2208                                 if (unlikely(!trylock_page(page))) {
2209                                         get_page(page);
2210                                         _pmd = *pmd;
2211                                         spin_unlock(ptl);
2212                                         lock_page(page);
2213                                         spin_lock(ptl);
2214                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2215                                                 unlock_page(page);
2216                                                 put_page(page);
2217                                                 page = NULL;
2218                                                 goto repeat;
2219                                         }
2220                                         put_page(page);
2221                                 }
2222                                 do_unlock_page = true;
2223                         }
2224                 }
2225                 if (PageMlocked(page))
2226                         clear_page_mlock(page);
2227         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2228                 goto out;
2229         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2230 out:
2231         spin_unlock(ptl);
2232         if (do_unlock_page)
2233                 unlock_page(page);
2234         /*
2235          * No need to double call mmu_notifier->invalidate_range() callback.
2236          * They are 3 cases to consider inside __split_huge_pmd_locked():
2237          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2238          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2239          *    fault will trigger a flush_notify before pointing to a new page
2240          *    (it is fine if the secondary mmu keeps pointing to the old zero
2241          *    page in the meantime)
2242          *  3) Split a huge pmd into pte pointing to the same page. No need
2243          *     to invalidate secondary tlb entry they are all still valid.
2244          *     any further changes to individual pte will notify. So no need
2245          *     to call mmu_notifier->invalidate_range()
2246          */
2247         mmu_notifier_invalidate_range_only_end(&range);
2248 }
2249
2250 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2251                 bool freeze, struct page *page)
2252 {
2253         pgd_t *pgd;
2254         p4d_t *p4d;
2255         pud_t *pud;
2256         pmd_t *pmd;
2257
2258         pgd = pgd_offset(vma->vm_mm, address);
2259         if (!pgd_present(*pgd))
2260                 return;
2261
2262         p4d = p4d_offset(pgd, address);
2263         if (!p4d_present(*p4d))
2264                 return;
2265
2266         pud = pud_offset(p4d, address);
2267         if (!pud_present(*pud))
2268                 return;
2269
2270         pmd = pmd_offset(pud, address);
2271
2272         __split_huge_pmd(vma, pmd, address, freeze, page);
2273 }
2274
2275 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2276 {
2277         /*
2278          * If the new address isn't hpage aligned and it could previously
2279          * contain an hugepage: check if we need to split an huge pmd.
2280          */
2281         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2282             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2283                          ALIGN(address, HPAGE_PMD_SIZE)))
2284                 split_huge_pmd_address(vma, address, false, NULL);
2285 }
2286
2287 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2288                              unsigned long start,
2289                              unsigned long end,
2290                              long adjust_next)
2291 {
2292         /* Check if we need to split start first. */
2293         split_huge_pmd_if_needed(vma, start);
2294
2295         /* Check if we need to split end next. */
2296         split_huge_pmd_if_needed(vma, end);
2297
2298         /*
2299          * If we're also updating the vma->vm_next->vm_start,
2300          * check if we need to split it.
2301          */
2302         if (adjust_next > 0) {
2303                 struct vm_area_struct *next = vma->vm_next;
2304                 unsigned long nstart = next->vm_start;
2305                 nstart += adjust_next;
2306                 split_huge_pmd_if_needed(next, nstart);
2307         }
2308 }
2309
2310 static void unmap_page(struct page *page)
2311 {
2312         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2313                 TTU_SYNC;
2314
2315         VM_BUG_ON_PAGE(!PageHead(page), page);
2316
2317         /*
2318          * Anon pages need migration entries to preserve them, but file
2319          * pages can simply be left unmapped, then faulted back on demand.
2320          * If that is ever changed (perhaps for mlock), update remap_page().
2321          */
2322         if (PageAnon(page))
2323                 try_to_migrate(page, ttu_flags);
2324         else
2325                 try_to_unmap(page, ttu_flags | TTU_IGNORE_MLOCK);
2326
2327         VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2328 }
2329
2330 static void remap_page(struct page *page, unsigned int nr)
2331 {
2332         int i;
2333
2334         /* If TTU_SPLIT_FREEZE is ever extended to file, remove this check */
2335         if (!PageAnon(page))
2336                 return;
2337         if (PageTransHuge(page)) {
2338                 remove_migration_ptes(page, page, true);
2339         } else {
2340                 for (i = 0; i < nr; i++)
2341                         remove_migration_ptes(page + i, page + i, true);
2342         }
2343 }
2344
2345 static void lru_add_page_tail(struct page *head, struct page *tail,
2346                 struct lruvec *lruvec, struct list_head *list)
2347 {
2348         VM_BUG_ON_PAGE(!PageHead(head), head);
2349         VM_BUG_ON_PAGE(PageCompound(tail), head);
2350         VM_BUG_ON_PAGE(PageLRU(tail), head);
2351         lockdep_assert_held(&lruvec->lru_lock);
2352
2353         if (list) {
2354                 /* page reclaim is reclaiming a huge page */
2355                 VM_WARN_ON(PageLRU(head));
2356                 get_page(tail);
2357                 list_add_tail(&tail->lru, list);
2358         } else {
2359                 /* head is still on lru (and we have it frozen) */
2360                 VM_WARN_ON(!PageLRU(head));
2361                 SetPageLRU(tail);
2362                 list_add_tail(&tail->lru, &head->lru);
2363         }
2364 }
2365
2366 static void __split_huge_page_tail(struct page *head, int tail,
2367                 struct lruvec *lruvec, struct list_head *list)
2368 {
2369         struct page *page_tail = head + tail;
2370
2371         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2372
2373         /*
2374          * Clone page flags before unfreezing refcount.
2375          *
2376          * After successful get_page_unless_zero() might follow flags change,
2377          * for example lock_page() which set PG_waiters.
2378          */
2379         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2380         page_tail->flags |= (head->flags &
2381                         ((1L << PG_referenced) |
2382                          (1L << PG_swapbacked) |
2383                          (1L << PG_swapcache) |
2384                          (1L << PG_mlocked) |
2385                          (1L << PG_uptodate) |
2386                          (1L << PG_active) |
2387                          (1L << PG_workingset) |
2388                          (1L << PG_locked) |
2389                          (1L << PG_unevictable) |
2390 #ifdef CONFIG_64BIT
2391                          (1L << PG_arch_2) |
2392 #endif
2393                          (1L << PG_dirty)));
2394
2395         /* ->mapping in first tail page is compound_mapcount */
2396         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2397                         page_tail);
2398         page_tail->mapping = head->mapping;
2399         page_tail->index = head->index + tail;
2400
2401         /* Page flags must be visible before we make the page non-compound. */
2402         smp_wmb();
2403
2404         /*
2405          * Clear PageTail before unfreezing page refcount.
2406          *
2407          * After successful get_page_unless_zero() might follow put_page()
2408          * which needs correct compound_head().
2409          */
2410         clear_compound_head(page_tail);
2411
2412         /* Finally unfreeze refcount. Additional reference from page cache. */
2413         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2414                                           PageSwapCache(head)));
2415
2416         if (page_is_young(head))
2417                 set_page_young(page_tail);
2418         if (page_is_idle(head))
2419                 set_page_idle(page_tail);
2420
2421         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2422
2423         /*
2424          * always add to the tail because some iterators expect new
2425          * pages to show after the currently processed elements - e.g.
2426          * migrate_pages
2427          */
2428         lru_add_page_tail(head, page_tail, lruvec, list);
2429 }
2430
2431 static void __split_huge_page(struct page *page, struct list_head *list,
2432                 pgoff_t end)
2433 {
2434         struct page *head = compound_head(page);
2435         struct lruvec *lruvec;
2436         struct address_space *swap_cache = NULL;
2437         unsigned long offset = 0;
2438         unsigned int nr = thp_nr_pages(head);
2439         int i;
2440
2441         /* complete memcg works before add pages to LRU */
2442         split_page_memcg(head, nr);
2443
2444         if (PageAnon(head) && PageSwapCache(head)) {
2445                 swp_entry_t entry = { .val = page_private(head) };
2446
2447                 offset = swp_offset(entry);
2448                 swap_cache = swap_address_space(entry);
2449                 xa_lock(&swap_cache->i_pages);
2450         }
2451
2452         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2453         lruvec = lock_page_lruvec(head);
2454
2455         for (i = nr - 1; i >= 1; i--) {
2456                 __split_huge_page_tail(head, i, lruvec, list);
2457                 /* Some pages can be beyond i_size: drop them from page cache */
2458                 if (head[i].index >= end) {
2459                         ClearPageDirty(head + i);
2460                         __delete_from_page_cache(head + i, NULL);
2461                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2462                                 shmem_uncharge(head->mapping->host, 1);
2463                         put_page(head + i);
2464                 } else if (!PageAnon(page)) {
2465                         __xa_store(&head->mapping->i_pages, head[i].index,
2466                                         head + i, 0);
2467                 } else if (swap_cache) {
2468                         __xa_store(&swap_cache->i_pages, offset + i,
2469                                         head + i, 0);
2470                 }
2471         }
2472
2473         ClearPageCompound(head);
2474         unlock_page_lruvec(lruvec);
2475         /* Caller disabled irqs, so they are still disabled here */
2476
2477         split_page_owner(head, nr);
2478
2479         /* See comment in __split_huge_page_tail() */
2480         if (PageAnon(head)) {
2481                 /* Additional pin to swap cache */
2482                 if (PageSwapCache(head)) {
2483                         page_ref_add(head, 2);
2484                         xa_unlock(&swap_cache->i_pages);
2485                 } else {
2486                         page_ref_inc(head);
2487                 }
2488         } else {
2489                 /* Additional pin to page cache */
2490                 page_ref_add(head, 2);
2491                 xa_unlock(&head->mapping->i_pages);
2492         }
2493         local_irq_enable();
2494
2495         remap_page(head, nr);
2496
2497         if (PageSwapCache(head)) {
2498                 swp_entry_t entry = { .val = page_private(head) };
2499
2500                 split_swap_cluster(entry);
2501         }
2502
2503         for (i = 0; i < nr; i++) {
2504                 struct page *subpage = head + i;
2505                 if (subpage == page)
2506                         continue;
2507                 unlock_page(subpage);
2508
2509                 /*
2510                  * Subpages may be freed if there wasn't any mapping
2511                  * like if add_to_swap() is running on a lru page that
2512                  * had its mapping zapped. And freeing these pages
2513                  * requires taking the lru_lock so we do the put_page
2514                  * of the tail pages after the split is complete.
2515                  */
2516                 put_page(subpage);
2517         }
2518 }
2519
2520 int total_mapcount(struct page *page)
2521 {
2522         int i, compound, nr, ret;
2523
2524         VM_BUG_ON_PAGE(PageTail(page), page);
2525
2526         if (likely(!PageCompound(page)))
2527                 return atomic_read(&page->_mapcount) + 1;
2528
2529         compound = compound_mapcount(page);
2530         nr = compound_nr(page);
2531         if (PageHuge(page))
2532                 return compound;
2533         ret = compound;
2534         for (i = 0; i < nr; i++)
2535                 ret += atomic_read(&page[i]._mapcount) + 1;
2536         /* File pages has compound_mapcount included in _mapcount */
2537         if (!PageAnon(page))
2538                 return ret - compound * nr;
2539         if (PageDoubleMap(page))
2540                 ret -= nr;
2541         return ret;
2542 }
2543
2544 /*
2545  * This calculates accurately how many mappings a transparent hugepage
2546  * has (unlike page_mapcount() which isn't fully accurate). This full
2547  * accuracy is primarily needed to know if copy-on-write faults can
2548  * reuse the page and change the mapping to read-write instead of
2549  * copying them. At the same time this returns the total_mapcount too.
2550  *
2551  * The function returns the highest mapcount any one of the subpages
2552  * has. If the return value is one, even if different processes are
2553  * mapping different subpages of the transparent hugepage, they can
2554  * all reuse it, because each process is reusing a different subpage.
2555  *
2556  * The total_mapcount is instead counting all virtual mappings of the
2557  * subpages. If the total_mapcount is equal to "one", it tells the
2558  * caller all mappings belong to the same "mm" and in turn the
2559  * anon_vma of the transparent hugepage can become the vma->anon_vma
2560  * local one as no other process may be mapping any of the subpages.
2561  *
2562  * It would be more accurate to replace page_mapcount() with
2563  * page_trans_huge_mapcount(), however we only use
2564  * page_trans_huge_mapcount() in the copy-on-write faults where we
2565  * need full accuracy to avoid breaking page pinning, because
2566  * page_trans_huge_mapcount() is slower than page_mapcount().
2567  */
2568 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2569 {
2570         int i, ret, _total_mapcount, mapcount;
2571
2572         /* hugetlbfs shouldn't call it */
2573         VM_BUG_ON_PAGE(PageHuge(page), page);
2574
2575         if (likely(!PageTransCompound(page))) {
2576                 mapcount = atomic_read(&page->_mapcount) + 1;
2577                 if (total_mapcount)
2578                         *total_mapcount = mapcount;
2579                 return mapcount;
2580         }
2581
2582         page = compound_head(page);
2583
2584         _total_mapcount = ret = 0;
2585         for (i = 0; i < thp_nr_pages(page); i++) {
2586                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2587                 ret = max(ret, mapcount);
2588                 _total_mapcount += mapcount;
2589         }
2590         if (PageDoubleMap(page)) {
2591                 ret -= 1;
2592                 _total_mapcount -= thp_nr_pages(page);
2593         }
2594         mapcount = compound_mapcount(page);
2595         ret += mapcount;
2596         _total_mapcount += mapcount;
2597         if (total_mapcount)
2598                 *total_mapcount = _total_mapcount;
2599         return ret;
2600 }
2601
2602 /* Racy check whether the huge page can be split */
2603 bool can_split_huge_page(struct page *page, int *pextra_pins)
2604 {
2605         int extra_pins;
2606
2607         /* Additional pins from page cache */
2608         if (PageAnon(page))
2609                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2610         else
2611                 extra_pins = thp_nr_pages(page);
2612         if (pextra_pins)
2613                 *pextra_pins = extra_pins;
2614         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2615 }
2616
2617 /*
2618  * This function splits huge page into normal pages. @page can point to any
2619  * subpage of huge page to split. Split doesn't change the position of @page.
2620  *
2621  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2622  * The huge page must be locked.
2623  *
2624  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2625  *
2626  * Both head page and tail pages will inherit mapping, flags, and so on from
2627  * the hugepage.
2628  *
2629  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2630  * they are not mapped.
2631  *
2632  * Returns 0 if the hugepage is split successfully.
2633  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2634  * us.
2635  */
2636 int split_huge_page_to_list(struct page *page, struct list_head *list)
2637 {
2638         struct page *head = compound_head(page);
2639         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2640         struct anon_vma *anon_vma = NULL;
2641         struct address_space *mapping = NULL;
2642         int extra_pins, ret;
2643         pgoff_t end;
2644
2645         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2646         VM_BUG_ON_PAGE(!PageLocked(head), head);
2647         VM_BUG_ON_PAGE(!PageCompound(head), head);
2648
2649         if (PageWriteback(head))
2650                 return -EBUSY;
2651
2652         if (PageAnon(head)) {
2653                 /*
2654                  * The caller does not necessarily hold an mmap_lock that would
2655                  * prevent the anon_vma disappearing so we first we take a
2656                  * reference to it and then lock the anon_vma for write. This
2657                  * is similar to page_lock_anon_vma_read except the write lock
2658                  * is taken to serialise against parallel split or collapse
2659                  * operations.
2660                  */
2661                 anon_vma = page_get_anon_vma(head);
2662                 if (!anon_vma) {
2663                         ret = -EBUSY;
2664                         goto out;
2665                 }
2666                 end = -1;
2667                 mapping = NULL;
2668                 anon_vma_lock_write(anon_vma);
2669         } else {
2670                 mapping = head->mapping;
2671
2672                 /* Truncated ? */
2673                 if (!mapping) {
2674                         ret = -EBUSY;
2675                         goto out;
2676                 }
2677
2678                 anon_vma = NULL;
2679                 i_mmap_lock_read(mapping);
2680
2681                 /*
2682                  *__split_huge_page() may need to trim off pages beyond EOF:
2683                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2684                  * which cannot be nested inside the page tree lock. So note
2685                  * end now: i_size itself may be changed at any moment, but
2686                  * head page lock is good enough to serialize the trimming.
2687                  */
2688                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2689         }
2690
2691         /*
2692          * Racy check if we can split the page, before unmap_page() will
2693          * split PMDs
2694          */
2695         if (!can_split_huge_page(head, &extra_pins)) {
2696                 ret = -EBUSY;
2697                 goto out_unlock;
2698         }
2699
2700         unmap_page(head);
2701
2702         /* block interrupt reentry in xa_lock and spinlock */
2703         local_irq_disable();
2704         if (mapping) {
2705                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2706
2707                 /*
2708                  * Check if the head page is present in page cache.
2709                  * We assume all tail are present too, if head is there.
2710                  */
2711                 xa_lock(&mapping->i_pages);
2712                 if (xas_load(&xas) != head)
2713                         goto fail;
2714         }
2715
2716         /* Prevent deferred_split_scan() touching ->_refcount */
2717         spin_lock(&ds_queue->split_queue_lock);
2718         if (page_ref_freeze(head, 1 + extra_pins)) {
2719                 if (!list_empty(page_deferred_list(head))) {
2720                         ds_queue->split_queue_len--;
2721                         list_del(page_deferred_list(head));
2722                 }
2723                 spin_unlock(&ds_queue->split_queue_lock);
2724                 if (mapping) {
2725                         int nr = thp_nr_pages(head);
2726
2727                         if (PageSwapBacked(head))
2728                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2729                                                         -nr);
2730                         else
2731                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2732                                                         -nr);
2733                 }
2734
2735                 __split_huge_page(page, list, end);
2736                 ret = 0;
2737         } else {
2738                 spin_unlock(&ds_queue->split_queue_lock);
2739 fail:
2740                 if (mapping)
2741                         xa_unlock(&mapping->i_pages);
2742                 local_irq_enable();
2743                 remap_page(head, thp_nr_pages(head));
2744                 ret = -EBUSY;
2745         }
2746
2747 out_unlock:
2748         if (anon_vma) {
2749                 anon_vma_unlock_write(anon_vma);
2750                 put_anon_vma(anon_vma);
2751         }
2752         if (mapping)
2753                 i_mmap_unlock_read(mapping);
2754 out:
2755         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2756         return ret;
2757 }
2758
2759 void free_transhuge_page(struct page *page)
2760 {
2761         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2762         unsigned long flags;
2763
2764         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2765         if (!list_empty(page_deferred_list(page))) {
2766                 ds_queue->split_queue_len--;
2767                 list_del(page_deferred_list(page));
2768         }
2769         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2770         free_compound_page(page);
2771 }
2772
2773 void deferred_split_huge_page(struct page *page)
2774 {
2775         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2776 #ifdef CONFIG_MEMCG
2777         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2778 #endif
2779         unsigned long flags;
2780
2781         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2782
2783         /*
2784          * The try_to_unmap() in page reclaim path might reach here too,
2785          * this may cause a race condition to corrupt deferred split queue.
2786          * And, if page reclaim is already handling the same page, it is
2787          * unnecessary to handle it again in shrinker.
2788          *
2789          * Check PageSwapCache to determine if the page is being
2790          * handled by page reclaim since THP swap would add the page into
2791          * swap cache before calling try_to_unmap().
2792          */
2793         if (PageSwapCache(page))
2794                 return;
2795
2796         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2797         if (list_empty(page_deferred_list(page))) {
2798                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2799                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2800                 ds_queue->split_queue_len++;
2801 #ifdef CONFIG_MEMCG
2802                 if (memcg)
2803                         set_shrinker_bit(memcg, page_to_nid(page),
2804                                          deferred_split_shrinker.id);
2805 #endif
2806         }
2807         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2808 }
2809
2810 static unsigned long deferred_split_count(struct shrinker *shrink,
2811                 struct shrink_control *sc)
2812 {
2813         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2814         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2815
2816 #ifdef CONFIG_MEMCG
2817         if (sc->memcg)
2818                 ds_queue = &sc->memcg->deferred_split_queue;
2819 #endif
2820         return READ_ONCE(ds_queue->split_queue_len);
2821 }
2822
2823 static unsigned long deferred_split_scan(struct shrinker *shrink,
2824                 struct shrink_control *sc)
2825 {
2826         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2827         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2828         unsigned long flags;
2829         LIST_HEAD(list), *pos, *next;
2830         struct page *page;
2831         int split = 0;
2832
2833 #ifdef CONFIG_MEMCG
2834         if (sc->memcg)
2835                 ds_queue = &sc->memcg->deferred_split_queue;
2836 #endif
2837
2838         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2839         /* Take pin on all head pages to avoid freeing them under us */
2840         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2841                 page = list_entry((void *)pos, struct page, deferred_list);
2842                 page = compound_head(page);
2843                 if (get_page_unless_zero(page)) {
2844                         list_move(page_deferred_list(page), &list);
2845                 } else {
2846                         /* We lost race with put_compound_page() */
2847                         list_del_init(page_deferred_list(page));
2848                         ds_queue->split_queue_len--;
2849                 }
2850                 if (!--sc->nr_to_scan)
2851                         break;
2852         }
2853         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2854
2855         list_for_each_safe(pos, next, &list) {
2856                 page = list_entry((void *)pos, struct page, deferred_list);
2857                 if (!trylock_page(page))
2858                         goto next;
2859                 /* split_huge_page() removes page from list on success */
2860                 if (!split_huge_page(page))
2861                         split++;
2862                 unlock_page(page);
2863 next:
2864                 put_page(page);
2865         }
2866
2867         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2868         list_splice_tail(&list, &ds_queue->split_queue);
2869         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2870
2871         /*
2872          * Stop shrinker if we didn't split any page, but the queue is empty.
2873          * This can happen if pages were freed under us.
2874          */
2875         if (!split && list_empty(&ds_queue->split_queue))
2876                 return SHRINK_STOP;
2877         return split;
2878 }
2879
2880 static struct shrinker deferred_split_shrinker = {
2881         .count_objects = deferred_split_count,
2882         .scan_objects = deferred_split_scan,
2883         .seeks = DEFAULT_SEEKS,
2884         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2885                  SHRINKER_NONSLAB,
2886 };
2887
2888 #ifdef CONFIG_DEBUG_FS
2889 static void split_huge_pages_all(void)
2890 {
2891         struct zone *zone;
2892         struct page *page;
2893         unsigned long pfn, max_zone_pfn;
2894         unsigned long total = 0, split = 0;
2895
2896         pr_debug("Split all THPs\n");
2897         for_each_populated_zone(zone) {
2898                 max_zone_pfn = zone_end_pfn(zone);
2899                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2900                         if (!pfn_valid(pfn))
2901                                 continue;
2902
2903                         page = pfn_to_page(pfn);
2904                         if (!get_page_unless_zero(page))
2905                                 continue;
2906
2907                         if (zone != page_zone(page))
2908                                 goto next;
2909
2910                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2911                                 goto next;
2912
2913                         total++;
2914                         lock_page(page);
2915                         if (!split_huge_page(page))
2916                                 split++;
2917                         unlock_page(page);
2918 next:
2919                         put_page(page);
2920                         cond_resched();
2921                 }
2922         }
2923
2924         pr_debug("%lu of %lu THP split\n", split, total);
2925 }
2926
2927 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2928 {
2929         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2930                     is_vm_hugetlb_page(vma);
2931 }
2932
2933 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2934                                 unsigned long vaddr_end)
2935 {
2936         int ret = 0;
2937         struct task_struct *task;
2938         struct mm_struct *mm;
2939         unsigned long total = 0, split = 0;
2940         unsigned long addr;
2941
2942         vaddr_start &= PAGE_MASK;
2943         vaddr_end &= PAGE_MASK;
2944
2945         /* Find the task_struct from pid */
2946         rcu_read_lock();
2947         task = find_task_by_vpid(pid);
2948         if (!task) {
2949                 rcu_read_unlock();
2950                 ret = -ESRCH;
2951                 goto out;
2952         }
2953         get_task_struct(task);
2954         rcu_read_unlock();
2955
2956         /* Find the mm_struct */
2957         mm = get_task_mm(task);
2958         put_task_struct(task);
2959
2960         if (!mm) {
2961                 ret = -EINVAL;
2962                 goto out;
2963         }
2964
2965         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2966                  pid, vaddr_start, vaddr_end);
2967
2968         mmap_read_lock(mm);
2969         /*
2970          * always increase addr by PAGE_SIZE, since we could have a PTE page
2971          * table filled with PTE-mapped THPs, each of which is distinct.
2972          */
2973         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2974                 struct vm_area_struct *vma = find_vma(mm, addr);
2975                 unsigned int follflags;
2976                 struct page *page;
2977
2978                 if (!vma || addr < vma->vm_start)
2979                         break;
2980
2981                 /* skip special VMA and hugetlb VMA */
2982                 if (vma_not_suitable_for_thp_split(vma)) {
2983                         addr = vma->vm_end;
2984                         continue;
2985                 }
2986
2987                 /* FOLL_DUMP to ignore special (like zero) pages */
2988                 follflags = FOLL_GET | FOLL_DUMP;
2989                 page = follow_page(vma, addr, follflags);
2990
2991                 if (IS_ERR(page))
2992                         continue;
2993                 if (!page)
2994                         continue;
2995
2996                 if (!is_transparent_hugepage(page))
2997                         goto next;
2998
2999                 total++;
3000                 if (!can_split_huge_page(compound_head(page), NULL))
3001                         goto next;
3002
3003                 if (!trylock_page(page))
3004                         goto next;
3005
3006                 if (!split_huge_page(page))
3007                         split++;
3008
3009                 unlock_page(page);
3010 next:
3011                 put_page(page);
3012                 cond_resched();
3013         }
3014         mmap_read_unlock(mm);
3015         mmput(mm);
3016
3017         pr_debug("%lu of %lu THP split\n", split, total);
3018
3019 out:
3020         return ret;
3021 }
3022
3023 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3024                                 pgoff_t off_end)
3025 {
3026         struct filename *file;
3027         struct file *candidate;
3028         struct address_space *mapping;
3029         int ret = -EINVAL;
3030         pgoff_t index;
3031         int nr_pages = 1;
3032         unsigned long total = 0, split = 0;
3033
3034         file = getname_kernel(file_path);
3035         if (IS_ERR(file))
3036                 return ret;
3037
3038         candidate = file_open_name(file, O_RDONLY, 0);
3039         if (IS_ERR(candidate))
3040                 goto out;
3041
3042         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3043                  file_path, off_start, off_end);
3044
3045         mapping = candidate->f_mapping;
3046
3047         for (index = off_start; index < off_end; index += nr_pages) {
3048                 struct page *fpage = pagecache_get_page(mapping, index,
3049                                                 FGP_ENTRY | FGP_HEAD, 0);
3050
3051                 nr_pages = 1;
3052                 if (xa_is_value(fpage) || !fpage)
3053                         continue;
3054
3055                 if (!is_transparent_hugepage(fpage))
3056                         goto next;
3057
3058                 total++;
3059                 nr_pages = thp_nr_pages(fpage);
3060
3061                 if (!trylock_page(fpage))
3062                         goto next;
3063
3064                 if (!split_huge_page(fpage))
3065                         split++;
3066
3067                 unlock_page(fpage);
3068 next:
3069                 put_page(fpage);
3070                 cond_resched();
3071         }
3072
3073         filp_close(candidate, NULL);
3074         ret = 0;
3075
3076         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3077 out:
3078         putname(file);
3079         return ret;
3080 }
3081
3082 #define MAX_INPUT_BUF_SZ 255
3083
3084 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3085                                 size_t count, loff_t *ppops)
3086 {
3087         static DEFINE_MUTEX(split_debug_mutex);
3088         ssize_t ret;
3089         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3090         char input_buf[MAX_INPUT_BUF_SZ];
3091         int pid;
3092         unsigned long vaddr_start, vaddr_end;
3093
3094         ret = mutex_lock_interruptible(&split_debug_mutex);
3095         if (ret)
3096                 return ret;
3097
3098         ret = -EFAULT;
3099
3100         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3101         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3102                 goto out;
3103
3104         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3105
3106         if (input_buf[0] == '/') {
3107                 char *tok;
3108                 char *buf = input_buf;
3109                 char file_path[MAX_INPUT_BUF_SZ];
3110                 pgoff_t off_start = 0, off_end = 0;
3111                 size_t input_len = strlen(input_buf);
3112
3113                 tok = strsep(&buf, ",");
3114                 if (tok) {
3115                         strcpy(file_path, tok);
3116                 } else {
3117                         ret = -EINVAL;
3118                         goto out;
3119                 }
3120
3121                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3122                 if (ret != 2) {
3123                         ret = -EINVAL;
3124                         goto out;
3125                 }
3126                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3127                 if (!ret)
3128                         ret = input_len;
3129
3130                 goto out;
3131         }
3132
3133         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3134         if (ret == 1 && pid == 1) {
3135                 split_huge_pages_all();
3136                 ret = strlen(input_buf);
3137                 goto out;
3138         } else if (ret != 3) {
3139                 ret = -EINVAL;
3140                 goto out;
3141         }
3142
3143         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3144         if (!ret)
3145                 ret = strlen(input_buf);
3146 out:
3147         mutex_unlock(&split_debug_mutex);
3148         return ret;
3149
3150 }
3151
3152 static const struct file_operations split_huge_pages_fops = {
3153         .owner   = THIS_MODULE,
3154         .write   = split_huge_pages_write,
3155         .llseek  = no_llseek,
3156 };
3157
3158 static int __init split_huge_pages_debugfs(void)
3159 {
3160         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3161                             &split_huge_pages_fops);
3162         return 0;
3163 }
3164 late_initcall(split_huge_pages_debugfs);
3165 #endif
3166
3167 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3168 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3169                 struct page *page)
3170 {
3171         struct vm_area_struct *vma = pvmw->vma;
3172         struct mm_struct *mm = vma->vm_mm;
3173         unsigned long address = pvmw->address;
3174         pmd_t pmdval;
3175         swp_entry_t entry;
3176         pmd_t pmdswp;
3177
3178         if (!(pvmw->pmd && !pvmw->pte))
3179                 return;
3180
3181         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3182         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3183         if (pmd_dirty(pmdval))
3184                 set_page_dirty(page);
3185         if (pmd_write(pmdval))
3186                 entry = make_writable_migration_entry(page_to_pfn(page));
3187         else
3188                 entry = make_readable_migration_entry(page_to_pfn(page));
3189         pmdswp = swp_entry_to_pmd(entry);
3190         if (pmd_soft_dirty(pmdval))
3191                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3192         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3193         page_remove_rmap(page, true);
3194         put_page(page);
3195 }
3196
3197 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3198 {
3199         struct vm_area_struct *vma = pvmw->vma;
3200         struct mm_struct *mm = vma->vm_mm;
3201         unsigned long address = pvmw->address;
3202         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3203         pmd_t pmde;
3204         swp_entry_t entry;
3205
3206         if (!(pvmw->pmd && !pvmw->pte))
3207                 return;
3208
3209         entry = pmd_to_swp_entry(*pvmw->pmd);
3210         get_page(new);
3211         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3212         if (pmd_swp_soft_dirty(*pvmw->pmd))
3213                 pmde = pmd_mksoft_dirty(pmde);
3214         if (is_writable_migration_entry(entry))
3215                 pmde = maybe_pmd_mkwrite(pmde, vma);
3216         if (pmd_swp_uffd_wp(*pvmw->pmd))
3217                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3218
3219         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3220         if (PageAnon(new))
3221                 page_add_anon_rmap(new, vma, mmun_start, true);
3222         else
3223                 page_add_file_rmap(new, true);
3224         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3225         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3226                 mlock_vma_page(new);
3227         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3228 }
3229 #endif