mm/huge_memory.c

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