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