mm/huge_memory.c

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