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