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/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35
  36 #include <asm/tlb.h>
  37 #include <asm/pgalloc.h>
  38 #include "internal.h"
  39
  40 /*
  41  * By default, transparent hugepage support is disabled in order to avoid
  42  * risking an increased memory footprint for applications that are not
  43  * guaranteed to benefit from it. When transparent hugepage support is
  44  * enabled, it is for all mappings, and khugepaged scans all mappings.
  45  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  46  * for all hugepage allocations.
  47  */
  48 unsigned long transparent_hugepage_flags __read_mostly =
  49 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  50         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  51 #endif
  52 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  53         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  54 #endif
  55         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  58
  59 static struct shrinker deferred_split_shrinker;
  60
  61 static atomic_t huge_zero_refcount;
  62 struct page *huge_zero_page __read_mostly;
  63
  64 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  65 {
  66         /* The addr is used to check if the vma size fits */
  67         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  68
  69         if (!transhuge_vma_suitable(vma, addr))
  70                 return false;
  71         if (vma_is_anonymous(vma))
  72                 return __transparent_hugepage_enabled(vma);
  73         if (vma_is_shmem(vma))
  74                 return shmem_huge_enabled(vma);
  75
  76         return false;
  77 }
  78
  79 static struct page *get_huge_zero_page(void)
  80 {
  81         struct page *zero_page;
  82 retry:
  83         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  84                 return READ_ONCE(huge_zero_page);
  85
  86         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  87                         HPAGE_PMD_ORDER);
  88         if (!zero_page) {
  89                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  90                 return NULL;
  91         }
  92         count_vm_event(THP_ZERO_PAGE_ALLOC);
  93         preempt_disable();
  94         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  95                 preempt_enable();
  96                 __free_pages(zero_page, compound_order(zero_page));
  97                 goto retry;
  98         }
  99
 100         /* We take additional reference here. It will be put back by shrinker */
 101         atomic_set(&huge_zero_refcount, 2);
 102         preempt_enable();
 103         return READ_ONCE(huge_zero_page);
 104 }
 105
 106 static void put_huge_zero_page(void)
 107 {
 108         /*
 109          * Counter should never go to zero here. Only shrinker can put
 110          * last reference.
 111          */
 112         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 113 }
 114
 115 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 116 {
 117         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 118                 return READ_ONCE(huge_zero_page);
 119
 120         if (!get_huge_zero_page())
 121                 return NULL;
 122
 123         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 124                 put_huge_zero_page();
 125
 126         return READ_ONCE(huge_zero_page);
 127 }
 128
 129 void mm_put_huge_zero_page(struct mm_struct *mm)
 130 {
 131         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 132                 put_huge_zero_page();
 133 }
 134
 135 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 136                                         struct shrink_control *sc)
 137 {
 138         /* we can free zero page only if last reference remains */
 139         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 140 }
 141
 142 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 143                                        struct shrink_control *sc)
 144 {
 145         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 146                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 147                 BUG_ON(zero_page == NULL);
 148                 __free_pages(zero_page, compound_order(zero_page));
 149                 return HPAGE_PMD_NR;
 150         }
 151
 152         return 0;
 153 }
 154
 155 static struct shrinker huge_zero_page_shrinker = {
 156         .count_objects = shrink_huge_zero_page_count,
 157         .scan_objects = shrink_huge_zero_page_scan,
 158         .seeks = DEFAULT_SEEKS,
 159 };
 160
 161 #ifdef CONFIG_SYSFS
 162 static ssize_t enabled_show(struct kobject *kobj,
 163                             struct kobj_attribute *attr, char *buf)
 164 {
 165         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 166                 return sprintf(buf, "[always] madvise never\n");
 167         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 168                 return sprintf(buf, "always [madvise] never\n");
 169         else
 170                 return sprintf(buf, "always madvise [never]\n");
 171 }
 172
 173 static ssize_t enabled_store(struct kobject *kobj,
 174                              struct kobj_attribute *attr,
 175                              const char *buf, size_t count)
 176 {
 177         ssize_t ret = count;
 178
 179         if (!memcmp("always", buf,
 180                     min(sizeof("always")-1, count))) {
 181                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 182                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183         } else if (!memcmp("madvise", buf,
 184                            min(sizeof("madvise")-1, count))) {
 185                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 186                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 187         } else if (!memcmp("never", buf,
 188                            min(sizeof("never")-1, count))) {
 189                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 190                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 191         } else
 192                 ret = -EINVAL;
 193
 194         if (ret > 0) {
 195                 int err = start_stop_khugepaged();
 196                 if (err)
 197                         ret = err;
 198         }
 199         return ret;
 200 }
 201 static struct kobj_attribute enabled_attr =
 202         __ATTR(enabled, 0644, enabled_show, enabled_store);
 203
 204 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 205                                 struct kobj_attribute *attr, char *buf,
 206                                 enum transparent_hugepage_flag flag)
 207 {
 208         return sprintf(buf, "%d\n",
 209                        !!test_bit(flag, &transparent_hugepage_flags));
 210 }
 211
 212 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 213                                  struct kobj_attribute *attr,
 214                                  const char *buf, size_t count,
 215                                  enum transparent_hugepage_flag flag)
 216 {
 217         unsigned long value;
 218         int ret;
 219
 220         ret = kstrtoul(buf, 10, &value);
 221         if (ret < 0)
 222                 return ret;
 223         if (value > 1)
 224                 return -EINVAL;
 225
 226         if (value)
 227                 set_bit(flag, &transparent_hugepage_flags);
 228         else
 229                 clear_bit(flag, &transparent_hugepage_flags);
 230
 231         return count;
 232 }
 233
 234 static ssize_t defrag_show(struct kobject *kobj,
 235                            struct kobj_attribute *attr, char *buf)
 236 {
 237         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 238                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 239         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 240                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 241         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 242                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 243         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 244                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 245         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 246 }
 247
 248 static ssize_t defrag_store(struct kobject *kobj,
 249                             struct kobj_attribute *attr,
 250                             const char *buf, size_t count)
 251 {
 252         if (!memcmp("always", buf,
 253                     min(sizeof("always")-1, count))) {
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 255                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 257                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 258         } else if (!memcmp("defer+madvise", buf,
 259                     min(sizeof("defer+madvise")-1, count))) {
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 263                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 264         } else if (!memcmp("defer", buf,
 265                     min(sizeof("defer")-1, count))) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 270         } else if (!memcmp("madvise", buf,
 271                            min(sizeof("madvise")-1, count))) {
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 275                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 276         } else if (!memcmp("never", buf,
 277                            min(sizeof("never")-1, count))) {
 278                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 279                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 282         } else
 283                 return -EINVAL;
 284
 285         return count;
 286 }
 287 static struct kobj_attribute defrag_attr =
 288         __ATTR(defrag, 0644, defrag_show, defrag_store);
 289
 290 static ssize_t use_zero_page_show(struct kobject *kobj,
 291                 struct kobj_attribute *attr, char *buf)
 292 {
 293         return single_hugepage_flag_show(kobj, attr, buf,
 294                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 295 }
 296 static ssize_t use_zero_page_store(struct kobject *kobj,
 297                 struct kobj_attribute *attr, const char *buf, size_t count)
 298 {
 299         return single_hugepage_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 301 }
 302 static struct kobj_attribute use_zero_page_attr =
 303         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 304
 305 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 306                 struct kobj_attribute *attr, char *buf)
 307 {
 308         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 309 }
 310 static struct kobj_attribute hpage_pmd_size_attr =
 311         __ATTR_RO(hpage_pmd_size);
 312
 313 #ifdef CONFIG_DEBUG_VM
 314 static ssize_t debug_cow_show(struct kobject *kobj,
 315                                 struct kobj_attribute *attr, char *buf)
 316 {
 317         return single_hugepage_flag_show(kobj, attr, buf,
 318                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 319 }
 320 static ssize_t debug_cow_store(struct kobject *kobj,
 321                                struct kobj_attribute *attr,
 322                                const char *buf, size_t count)
 323 {
 324         return single_hugepage_flag_store(kobj, attr, buf, count,
 325                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 326 }
 327 static struct kobj_attribute debug_cow_attr =
 328         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 329 #endif /* CONFIG_DEBUG_VM */
 330
 331 static struct attribute *hugepage_attr[] = {
 332         &enabled_attr.attr,
 333         &defrag_attr.attr,
 334         &use_zero_page_attr.attr,
 335         &hpage_pmd_size_attr.attr,
 336 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 337         &shmem_enabled_attr.attr,
 338 #endif
 339 #ifdef CONFIG_DEBUG_VM
 340         &debug_cow_attr.attr,
 341 #endif
 342         NULL,
 343 };
 344
 345 static const struct attribute_group hugepage_attr_group = {
 346         .attrs = hugepage_attr,
 347 };
 348
 349 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 350 {
 351         int err;
 352
 353         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 354         if (unlikely(!*hugepage_kobj)) {
 355                 pr_err("failed to create transparent hugepage kobject\n");
 356                 return -ENOMEM;
 357         }
 358
 359         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 360         if (err) {
 361                 pr_err("failed to register transparent hugepage group\n");
 362                 goto delete_obj;
 363         }
 364
 365         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 366         if (err) {
 367                 pr_err("failed to register transparent hugepage group\n");
 368                 goto remove_hp_group;
 369         }
 370
 371         return 0;
 372
 373 remove_hp_group:
 374         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 375 delete_obj:
 376         kobject_put(*hugepage_kobj);
 377         return err;
 378 }
 379
 380 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 381 {
 382         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 383         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 384         kobject_put(hugepage_kobj);
 385 }
 386 #else
 387 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 388 {
 389         return 0;
 390 }
 391
 392 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 393 {
 394 }
 395 #endif /* CONFIG_SYSFS */
 396
 397 static int __init hugepage_init(void)
 398 {
 399         int err;
 400         struct kobject *hugepage_kobj;
 401
 402         if (!has_transparent_hugepage()) {
 403                 transparent_hugepage_flags = 0;
 404                 return -EINVAL;
 405         }
 406
 407         /*
 408          * hugepages can't be allocated by the buddy allocator
 409          */
 410         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 411         /*
 412          * we use page->mapping and page->index in second tail page
 413          * as list_head: assuming THP order >= 2
 414          */
 415         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 416
 417         err = hugepage_init_sysfs(&hugepage_kobj);
 418         if (err)
 419                 goto err_sysfs;
 420
 421         err = khugepaged_init();
 422         if (err)
 423                 goto err_slab;
 424
 425         err = register_shrinker(&huge_zero_page_shrinker);
 426         if (err)
 427                 goto err_hzp_shrinker;
 428         err = register_shrinker(&deferred_split_shrinker);
 429         if (err)
 430                 goto err_split_shrinker;
 431
 432         /*
 433          * By default disable transparent hugepages on smaller systems,
 434          * where the extra memory used could hurt more than TLB overhead
 435          * is likely to save.  The admin can still enable it through /sys.
 436          */
 437         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 438                 transparent_hugepage_flags = 0;
 439                 return 0;
 440         }
 441
 442         err = start_stop_khugepaged();
 443         if (err)
 444                 goto err_khugepaged;
 445
 446         return 0;
 447 err_khugepaged:
 448         unregister_shrinker(&deferred_split_shrinker);
 449 err_split_shrinker:
 450         unregister_shrinker(&huge_zero_page_shrinker);
 451 err_hzp_shrinker:
 452         khugepaged_destroy();
 453 err_slab:
 454         hugepage_exit_sysfs(hugepage_kobj);
 455 err_sysfs:
 456         return err;
 457 }
 458 subsys_initcall(hugepage_init);
 459
 460 static int __init setup_transparent_hugepage(char *str)
 461 {
 462         int ret = 0;
 463         if (!str)
 464                 goto out;
 465         if (!strcmp(str, "always")) {
 466                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 467                         &transparent_hugepage_flags);
 468                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 469                           &transparent_hugepage_flags);
 470                 ret = 1;
 471         } else if (!strcmp(str, "madvise")) {
 472                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 473                           &transparent_hugepage_flags);
 474                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 475                         &transparent_hugepage_flags);
 476                 ret = 1;
 477         } else if (!strcmp(str, "never")) {
 478                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 479                           &transparent_hugepage_flags);
 480                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 481                           &transparent_hugepage_flags);
 482                 ret = 1;
 483         }
 484 out:
 485         if (!ret)
 486                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 487         return ret;
 488 }
 489 __setup("transparent_hugepage=", setup_transparent_hugepage);
 490
 491 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 492 {
 493         if (likely(vma->vm_flags & VM_WRITE))
 494                 pmd = pmd_mkwrite(pmd);
 495         return pmd;
 496 }
 497
 498 static inline struct list_head *page_deferred_list(struct page *page)
 499 {
 500         /* ->lru in the tail pages is occupied by compound_head. */
 501         return &page[2].deferred_list;
 502 }
 503
 504 void prep_transhuge_page(struct page *page)
 505 {
 506         /*
 507          * we use page->mapping and page->indexlru in second tail page
 508          * as list_head: assuming THP order >= 2
 509          */
 510
 511         INIT_LIST_HEAD(page_deferred_list(page));
 512         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 513 }
 514
 515 static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 516                 loff_t off, unsigned long flags, unsigned long size)
 517 {
 518         unsigned long addr;
 519         loff_t off_end = off + len;
 520         loff_t off_align = round_up(off, size);
 521         unsigned long len_pad;
 522
 523         if (off_end <= off_align || (off_end - off_align) < size)
 524                 return 0;
 525
 526         len_pad = len + size;
 527         if (len_pad < len || (off + len_pad) < off)
 528                 return 0;
 529
 530         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 531                                               off >> PAGE_SHIFT, flags);
 532         if (IS_ERR_VALUE(addr))
 533                 return 0;
 534
 535         addr += (off - addr) & (size - 1);
 536         return addr;
 537 }
 538
 539 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 540                 unsigned long len, unsigned long pgoff, unsigned long flags)
 541 {
 542         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 543
 544         if (addr)
 545                 goto out;
 546         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 547                 goto out;
 548
 549         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 550         if (addr)
 551                 return addr;
 552
 553  out:
 554         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 555 }
 556 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 557
 558 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 559                         struct page *page, gfp_t gfp)
 560 {
 561         struct vm_area_struct *vma = vmf->vma;
 562         struct mem_cgroup *memcg;
 563         pgtable_t pgtable;
 564         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 565         vm_fault_t ret = 0;
 566
 567         VM_BUG_ON_PAGE(!PageCompound(page), page);
 568
 569         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 570                 put_page(page);
 571                 count_vm_event(THP_FAULT_FALLBACK);
 572                 return VM_FAULT_FALLBACK;
 573         }
 574
 575         pgtable = pte_alloc_one(vma->vm_mm);
 576         if (unlikely(!pgtable)) {
 577                 ret = VM_FAULT_OOM;
 578                 goto release;
 579         }
 580
 581         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 582         /*
 583          * The memory barrier inside __SetPageUptodate makes sure that
 584          * clear_huge_page writes become visible before the set_pmd_at()
 585          * write.
 586          */
 587         __SetPageUptodate(page);
 588
 589         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 590         if (unlikely(!pmd_none(*vmf->pmd))) {
 591                 goto unlock_release;
 592         } else {
 593                 pmd_t entry;
 594
 595                 ret = check_stable_address_space(vma->vm_mm);
 596                 if (ret)
 597                         goto unlock_release;
 598
 599                 /* Deliver the page fault to userland */
 600                 if (userfaultfd_missing(vma)) {
 601                         vm_fault_t ret2;
 602
 603                         spin_unlock(vmf->ptl);
 604                         mem_cgroup_cancel_charge(page, memcg, true);
 605                         put_page(page);
 606                         pte_free(vma->vm_mm, pgtable);
 607                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 608                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 609                         return ret2;
 610                 }
 611
 612                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 613                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 614                 page_add_new_anon_rmap(page, vma, haddr, true);
 615                 mem_cgroup_commit_charge(page, memcg, false, true);
 616                 lru_cache_add_active_or_unevictable(page, vma);
 617                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 618                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 619                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 620                 mm_inc_nr_ptes(vma->vm_mm);
 621                 spin_unlock(vmf->ptl);
 622                 count_vm_event(THP_FAULT_ALLOC);
 623                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 624         }
 625
 626         return 0;
 627 unlock_release:
 628         spin_unlock(vmf->ptl);
 629 release:
 630         if (pgtable)
 631                 pte_free(vma->vm_mm, pgtable);
 632         mem_cgroup_cancel_charge(page, memcg, true);
 633         put_page(page);
 634         return ret;
 635
 636 }
 637
 638 /*
 639  * always: directly stall for all thp allocations
 640  * defer: wake kswapd and fail if not immediately available
 641  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 642  *                fail if not immediately available
 643  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 644  *          available
 645  * never: never stall for any thp allocation
 646  */
 647 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 648 {
 649         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 650
 651         /* Always do synchronous compaction */
 652         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 653                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 654
 655         /* Kick kcompactd and fail quickly */
 656         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 657                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 658
 659         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 660         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 661                 return GFP_TRANSHUGE_LIGHT |
 662                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 663                                         __GFP_KSWAPD_RECLAIM);
 664
 665         /* Only do synchronous compaction if madvised */
 666         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 667                 return GFP_TRANSHUGE_LIGHT |
 668                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 669
 670         return GFP_TRANSHUGE_LIGHT;
 671 }
 672
 673 /* Caller must hold page table lock. */
 674 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 675                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 676                 struct page *zero_page)
 677 {
 678         pmd_t entry;
 679         if (!pmd_none(*pmd))
 680                 return false;
 681         entry = mk_pmd(zero_page, vma->vm_page_prot);
 682         entry = pmd_mkhuge(entry);
 683         if (pgtable)
 684                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 685         set_pmd_at(mm, haddr, pmd, entry);
 686         mm_inc_nr_ptes(mm);
 687         return true;
 688 }
 689
 690 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 691 {
 692         struct vm_area_struct *vma = vmf->vma;
 693         gfp_t gfp;
 694         struct page *page;
 695         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 696
 697         if (!transhuge_vma_suitable(vma, haddr))
 698                 return VM_FAULT_FALLBACK;
 699         if (unlikely(anon_vma_prepare(vma)))
 700                 return VM_FAULT_OOM;
 701         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 702                 return VM_FAULT_OOM;
 703         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 704                         !mm_forbids_zeropage(vma->vm_mm) &&
 705                         transparent_hugepage_use_zero_page()) {
 706                 pgtable_t pgtable;
 707                 struct page *zero_page;
 708                 bool set;
 709                 vm_fault_t ret;
 710                 pgtable = pte_alloc_one(vma->vm_mm);
 711                 if (unlikely(!pgtable))
 712                         return VM_FAULT_OOM;
 713                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 714                 if (unlikely(!zero_page)) {
 715                         pte_free(vma->vm_mm, pgtable);
 716                         count_vm_event(THP_FAULT_FALLBACK);
 717                         return VM_FAULT_FALLBACK;
 718                 }
 719                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 720                 ret = 0;
 721                 set = false;
 722                 if (pmd_none(*vmf->pmd)) {
 723                         ret = check_stable_address_space(vma->vm_mm);
 724                         if (ret) {
 725                                 spin_unlock(vmf->ptl);
 726                         } else if (userfaultfd_missing(vma)) {
 727                                 spin_unlock(vmf->ptl);
 728                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 729                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 730                         } else {
 731                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 732                                                    haddr, vmf->pmd, zero_page);
 733                                 spin_unlock(vmf->ptl);
 734                                 set = true;
 735                         }
 736                 } else
 737                         spin_unlock(vmf->ptl);
 738                 if (!set)
 739                         pte_free(vma->vm_mm, pgtable);
 740                 return ret;
 741         }
 742         gfp = alloc_hugepage_direct_gfpmask(vma);
 743         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 744         if (unlikely(!page)) {
 745                 count_vm_event(THP_FAULT_FALLBACK);
 746                 return VM_FAULT_FALLBACK;
 747         }
 748         prep_transhuge_page(page);
 749         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 750 }
 751
 752 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 753                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 754                 pgtable_t pgtable)
 755 {
 756         struct mm_struct *mm = vma->vm_mm;
 757         pmd_t entry;
 758         spinlock_t *ptl;
 759
 760         ptl = pmd_lock(mm, pmd);
 761         if (!pmd_none(*pmd)) {
 762                 if (write) {
 763                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 764                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 765                                 goto out_unlock;
 766                         }
 767                         entry = pmd_mkyoung(*pmd);
 768                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 769                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 770                                 update_mmu_cache_pmd(vma, addr, pmd);
 771                 }
 772
 773                 goto out_unlock;
 774         }
 775
 776         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 777         if (pfn_t_devmap(pfn))
 778                 entry = pmd_mkdevmap(entry);
 779         if (write) {
 780                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 781                 entry = maybe_pmd_mkwrite(entry, vma);
 782         }
 783
 784         if (pgtable) {
 785                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 786                 mm_inc_nr_ptes(mm);
 787                 pgtable = NULL;
 788         }
 789
 790         set_pmd_at(mm, addr, pmd, entry);
 791         update_mmu_cache_pmd(vma, addr, pmd);
 792
 793 out_unlock:
 794         spin_unlock(ptl);
 795         if (pgtable)
 796                 pte_free(mm, pgtable);
 797 }
 798
 799 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 800 {
 801         unsigned long addr = vmf->address & PMD_MASK;
 802         struct vm_area_struct *vma = vmf->vma;
 803         pgprot_t pgprot = vma->vm_page_prot;
 804         pgtable_t pgtable = NULL;
 805
 806         /*
 807          * If we had pmd_special, we could avoid all these restrictions,
 808          * but we need to be consistent with PTEs and architectures that
 809          * can't support a 'special' bit.
 810          */
 811         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 812                         !pfn_t_devmap(pfn));
 813         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 814                                                 (VM_PFNMAP|VM_MIXEDMAP));
 815         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 816
 817         if (addr < vma->vm_start || addr >= vma->vm_end)
 818                 return VM_FAULT_SIGBUS;
 819
 820         if (arch_needs_pgtable_deposit()) {
 821                 pgtable = pte_alloc_one(vma->vm_mm);
 822                 if (!pgtable)
 823                         return VM_FAULT_OOM;
 824         }
 825
 826         track_pfn_insert(vma, &pgprot, pfn);
 827
 828         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 829         return VM_FAULT_NOPAGE;
 830 }
 831 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 832
 833 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 834 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 835 {
 836         if (likely(vma->vm_flags & VM_WRITE))
 837                 pud = pud_mkwrite(pud);
 838         return pud;
 839 }
 840
 841 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 842                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 843 {
 844         struct mm_struct *mm = vma->vm_mm;
 845         pud_t entry;
 846         spinlock_t *ptl;
 847
 848         ptl = pud_lock(mm, pud);
 849         if (!pud_none(*pud)) {
 850                 if (write) {
 851                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 852                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 853                                 goto out_unlock;
 854                         }
 855                         entry = pud_mkyoung(*pud);
 856                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 857                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 858                                 update_mmu_cache_pud(vma, addr, pud);
 859                 }
 860                 goto out_unlock;
 861         }
 862
 863         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 864         if (pfn_t_devmap(pfn))
 865                 entry = pud_mkdevmap(entry);
 866         if (write) {
 867                 entry = pud_mkyoung(pud_mkdirty(entry));
 868                 entry = maybe_pud_mkwrite(entry, vma);
 869         }
 870         set_pud_at(mm, addr, pud, entry);
 871         update_mmu_cache_pud(vma, addr, pud);
 872
 873 out_unlock:
 874         spin_unlock(ptl);
 875 }
 876
 877 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 878 {
 879         unsigned long addr = vmf->address & PUD_MASK;
 880         struct vm_area_struct *vma = vmf->vma;
 881         pgprot_t pgprot = vma->vm_page_prot;
 882
 883         /*
 884          * If we had pud_special, we could avoid all these restrictions,
 885          * but we need to be consistent with PTEs and architectures that
 886          * can't support a 'special' bit.
 887          */
 888         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 889                         !pfn_t_devmap(pfn));
 890         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 891                                                 (VM_PFNMAP|VM_MIXEDMAP));
 892         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 893
 894         if (addr < vma->vm_start || addr >= vma->vm_end)
 895                 return VM_FAULT_SIGBUS;
 896
 897         track_pfn_insert(vma, &pgprot, pfn);
 898
 899         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 900         return VM_FAULT_NOPAGE;
 901 }
 902 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 903 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 904
 905 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 906                 pmd_t *pmd, int flags)
 907 {
 908         pmd_t _pmd;
 909
 910         _pmd = pmd_mkyoung(*pmd);
 911         if (flags & FOLL_WRITE)
 912                 _pmd = pmd_mkdirty(_pmd);
 913         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 914                                 pmd, _pmd, flags & FOLL_WRITE))
 915                 update_mmu_cache_pmd(vma, addr, pmd);
 916 }
 917
 918 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 919                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 920 {
 921         unsigned long pfn = pmd_pfn(*pmd);
 922         struct mm_struct *mm = vma->vm_mm;
 923         struct page *page;
 924
 925         assert_spin_locked(pmd_lockptr(mm, pmd));
 926
 927         /*
 928          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 929          * not be in this function with `flags & FOLL_COW` set.
 930          */
 931         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 932
 933         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 934                 return NULL;
 935
 936         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 937                 /* pass */;
 938         else
 939                 return NULL;
 940
 941         if (flags & FOLL_TOUCH)
 942                 touch_pmd(vma, addr, pmd, flags);
 943
 944         /*
 945          * device mapped pages can only be returned if the
 946          * caller will manage the page reference count.
 947          */
 948         if (!(flags & FOLL_GET))
 949                 return ERR_PTR(-EEXIST);
 950
 951         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 952         *pgmap = get_dev_pagemap(pfn, *pgmap);
 953         if (!*pgmap)
 954                 return ERR_PTR(-EFAULT);
 955         page = pfn_to_page(pfn);
 956         get_page(page);
 957
 958         return page;
 959 }
 960
 961 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 962                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 963                   struct vm_area_struct *vma)
 964 {
 965         spinlock_t *dst_ptl, *src_ptl;
 966         struct page *src_page;
 967         pmd_t pmd;
 968         pgtable_t pgtable = NULL;
 969         int ret = -ENOMEM;
 970
 971         /* Skip if can be re-fill on fault */
 972         if (!vma_is_anonymous(vma))
 973                 return 0;
 974
 975         pgtable = pte_alloc_one(dst_mm);
 976         if (unlikely(!pgtable))
 977                 goto out;
 978
 979         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 980         src_ptl = pmd_lockptr(src_mm, src_pmd);
 981         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 982
 983         ret = -EAGAIN;
 984         pmd = *src_pmd;
 985
 986 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 987         if (unlikely(is_swap_pmd(pmd))) {
 988                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 989
 990                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 991                 if (is_write_migration_entry(entry)) {
 992                         make_migration_entry_read(&entry);
 993                         pmd = swp_entry_to_pmd(entry);
 994                         if (pmd_swp_soft_dirty(*src_pmd))
 995                                 pmd = pmd_swp_mksoft_dirty(pmd);
 996                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 997                 }
 998                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 999                 mm_inc_nr_ptes(dst_mm);
1000                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1001                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1002                 ret = 0;
1003                 goto out_unlock;
1004         }
1005 #endif
1006
1007         if (unlikely(!pmd_trans_huge(pmd))) {
1008                 pte_free(dst_mm, pgtable);
1009                 goto out_unlock;
1010         }
1011         /*
1012          * When page table lock is held, the huge zero pmd should not be
1013          * under splitting since we don't split the page itself, only pmd to
1014          * a page table.
1015          */
1016         if (is_huge_zero_pmd(pmd)) {
1017                 struct page *zero_page;
1018                 /*
1019                  * get_huge_zero_page() will never allocate a new page here,
1020                  * since we already have a zero page to copy. It just takes a
1021                  * reference.
1022                  */
1023                 zero_page = mm_get_huge_zero_page(dst_mm);
1024                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1025                                 zero_page);
1026                 ret = 0;
1027                 goto out_unlock;
1028         }
1029
1030         src_page = pmd_page(pmd);
1031         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1032         get_page(src_page);
1033         page_dup_rmap(src_page, true);
1034         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1035         mm_inc_nr_ptes(dst_mm);
1036         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1037
1038         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1039         pmd = pmd_mkold(pmd_wrprotect(pmd));
1040         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1041
1042         ret = 0;
1043 out_unlock:
1044         spin_unlock(src_ptl);
1045         spin_unlock(dst_ptl);
1046 out:
1047         return ret;
1048 }
1049
1050 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1051 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1052                 pud_t *pud, int flags)
1053 {
1054         pud_t _pud;
1055
1056         _pud = pud_mkyoung(*pud);
1057         if (flags & FOLL_WRITE)
1058                 _pud = pud_mkdirty(_pud);
1059         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1060                                 pud, _pud, flags & FOLL_WRITE))
1061                 update_mmu_cache_pud(vma, addr, pud);
1062 }
1063
1064 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1065                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1066 {
1067         unsigned long pfn = pud_pfn(*pud);
1068         struct mm_struct *mm = vma->vm_mm;
1069         struct page *page;
1070
1071         assert_spin_locked(pud_lockptr(mm, pud));
1072
1073         if (flags & FOLL_WRITE && !pud_write(*pud))
1074                 return NULL;
1075
1076         if (pud_present(*pud) && pud_devmap(*pud))
1077                 /* pass */;
1078         else
1079                 return NULL;
1080
1081         if (flags & FOLL_TOUCH)
1082                 touch_pud(vma, addr, pud, flags);
1083
1084         /*
1085          * device mapped pages can only be returned if the
1086          * caller will manage the page reference count.
1087          */
1088         if (!(flags & FOLL_GET))
1089                 return ERR_PTR(-EEXIST);
1090
1091         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1092         *pgmap = get_dev_pagemap(pfn, *pgmap);
1093         if (!*pgmap)
1094                 return ERR_PTR(-EFAULT);
1095         page = pfn_to_page(pfn);
1096         get_page(page);
1097
1098         return page;
1099 }
1100
1101 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1102                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1103                   struct vm_area_struct *vma)
1104 {
1105         spinlock_t *dst_ptl, *src_ptl;
1106         pud_t pud;
1107         int ret;
1108
1109         dst_ptl = pud_lock(dst_mm, dst_pud);
1110         src_ptl = pud_lockptr(src_mm, src_pud);
1111         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1112
1113         ret = -EAGAIN;
1114         pud = *src_pud;
1115         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1116                 goto out_unlock;
1117
1118         /*
1119          * When page table lock is held, the huge zero pud should not be
1120          * under splitting since we don't split the page itself, only pud to
1121          * a page table.
1122          */
1123         if (is_huge_zero_pud(pud)) {
1124                 /* No huge zero pud yet */
1125         }
1126
1127         pudp_set_wrprotect(src_mm, addr, src_pud);
1128         pud = pud_mkold(pud_wrprotect(pud));
1129         set_pud_at(dst_mm, addr, dst_pud, pud);
1130
1131         ret = 0;
1132 out_unlock:
1133         spin_unlock(src_ptl);
1134         spin_unlock(dst_ptl);
1135         return ret;
1136 }
1137
1138 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1139 {
1140         pud_t entry;
1141         unsigned long haddr;
1142         bool write = vmf->flags & FAULT_FLAG_WRITE;
1143
1144         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1145         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1146                 goto unlock;
1147
1148         entry = pud_mkyoung(orig_pud);
1149         if (write)
1150                 entry = pud_mkdirty(entry);
1151         haddr = vmf->address & HPAGE_PUD_MASK;
1152         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1153                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1154
1155 unlock:
1156         spin_unlock(vmf->ptl);
1157 }
1158 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1159
1160 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1161 {
1162         pmd_t entry;
1163         unsigned long haddr;
1164         bool write = vmf->flags & FAULT_FLAG_WRITE;
1165
1166         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1167         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1168                 goto unlock;
1169
1170         entry = pmd_mkyoung(orig_pmd);
1171         if (write)
1172                 entry = pmd_mkdirty(entry);
1173         haddr = vmf->address & HPAGE_PMD_MASK;
1174         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1175                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1176
1177 unlock:
1178         spin_unlock(vmf->ptl);
1179 }
1180
1181 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1182                         pmd_t orig_pmd, struct page *page)
1183 {
1184         struct vm_area_struct *vma = vmf->vma;
1185         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1186         struct mem_cgroup *memcg;
1187         pgtable_t pgtable;
1188         pmd_t _pmd;
1189         int i;
1190         vm_fault_t ret = 0;
1191         struct page **pages;
1192         struct mmu_notifier_range range;
1193
1194         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1195                               GFP_KERNEL);
1196         if (unlikely(!pages)) {
1197                 ret |= VM_FAULT_OOM;
1198                 goto out;
1199         }
1200
1201         for (i = 0; i < HPAGE_PMD_NR; i++) {
1202                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1203                                                vmf->address, page_to_nid(page));
1204                 if (unlikely(!pages[i] ||
1205                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1206                                      GFP_KERNEL, &memcg, false))) {
1207                         if (pages[i])
1208                                 put_page(pages[i]);
1209                         while (--i >= 0) {
1210                                 memcg = (void *)page_private(pages[i]);
1211                                 set_page_private(pages[i], 0);
1212                                 mem_cgroup_cancel_charge(pages[i], memcg,
1213                                                 false);
1214                                 put_page(pages[i]);
1215                         }
1216                         kfree(pages);
1217                         ret |= VM_FAULT_OOM;
1218                         goto out;
1219                 }
1220                 set_page_private(pages[i], (unsigned long)memcg);
1221         }
1222
1223         for (i = 0; i < HPAGE_PMD_NR; i++) {
1224                 copy_user_highpage(pages[i], page + i,
1225                                    haddr + PAGE_SIZE * i, vma);
1226                 __SetPageUptodate(pages[i]);
1227                 cond_resched();
1228         }
1229
1230         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1231                                 haddr, haddr + HPAGE_PMD_SIZE);
1232         mmu_notifier_invalidate_range_start(&range);
1233
1234         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1235         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1236                 goto out_free_pages;
1237         VM_BUG_ON_PAGE(!PageHead(page), page);
1238
1239         /*
1240          * Leave pmd empty until pte is filled note we must notify here as
1241          * concurrent CPU thread might write to new page before the call to
1242          * mmu_notifier_invalidate_range_end() happens which can lead to a
1243          * device seeing memory write in different order than CPU.
1244          *
1245          * See Documentation/vm/mmu_notifier.rst
1246          */
1247         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1248
1249         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1250         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1251
1252         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1253                 pte_t entry;
1254                 entry = mk_pte(pages[i], vma->vm_page_prot);
1255                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1256                 memcg = (void *)page_private(pages[i]);
1257                 set_page_private(pages[i], 0);
1258                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1259                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1260                 lru_cache_add_active_or_unevictable(pages[i], vma);
1261                 vmf->pte = pte_offset_map(&_pmd, haddr);
1262                 VM_BUG_ON(!pte_none(*vmf->pte));
1263                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1264                 pte_unmap(vmf->pte);
1265         }
1266         kfree(pages);
1267
1268         smp_wmb(); /* make pte visible before pmd */
1269         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1270         page_remove_rmap(page, true);
1271         spin_unlock(vmf->ptl);
1272
1273         /*
1274          * No need to double call mmu_notifier->invalidate_range() callback as
1275          * the above pmdp_huge_clear_flush_notify() did already call it.
1276          */
1277         mmu_notifier_invalidate_range_only_end(&range);
1278
1279         ret |= VM_FAULT_WRITE;
1280         put_page(page);
1281
1282 out:
1283         return ret;
1284
1285 out_free_pages:
1286         spin_unlock(vmf->ptl);
1287         mmu_notifier_invalidate_range_end(&range);
1288         for (i = 0; i < HPAGE_PMD_NR; i++) {
1289                 memcg = (void *)page_private(pages[i]);
1290                 set_page_private(pages[i], 0);
1291                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1292                 put_page(pages[i]);
1293         }
1294         kfree(pages);
1295         goto out;
1296 }
1297
1298 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1299 {
1300         struct vm_area_struct *vma = vmf->vma;
1301         struct page *page = NULL, *new_page;
1302         struct mem_cgroup *memcg;
1303         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1304         struct mmu_notifier_range range;
1305         gfp_t huge_gfp;                 /* for allocation and charge */
1306         vm_fault_t ret = 0;
1307
1308         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1309         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1310         if (is_huge_zero_pmd(orig_pmd))
1311                 goto alloc;
1312         spin_lock(vmf->ptl);
1313         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1314                 goto out_unlock;
1315
1316         page = pmd_page(orig_pmd);
1317         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1318         /*
1319          * We can only reuse the page if nobody else maps the huge page or it's
1320          * part.
1321          */
1322         if (!trylock_page(page)) {
1323                 get_page(page);
1324                 spin_unlock(vmf->ptl);
1325                 lock_page(page);
1326                 spin_lock(vmf->ptl);
1327                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1328                         unlock_page(page);
1329                         put_page(page);
1330                         goto out_unlock;
1331                 }
1332                 put_page(page);
1333         }
1334         if (reuse_swap_page(page, NULL)) {
1335                 pmd_t entry;
1336                 entry = pmd_mkyoung(orig_pmd);
1337                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1338                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1339                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1340                 ret |= VM_FAULT_WRITE;
1341                 unlock_page(page);
1342                 goto out_unlock;
1343         }
1344         unlock_page(page);
1345         get_page(page);
1346         spin_unlock(vmf->ptl);
1347 alloc:
1348         if (__transparent_hugepage_enabled(vma) &&
1349             !transparent_hugepage_debug_cow()) {
1350                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1351                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1352         } else
1353                 new_page = NULL;
1354
1355         if (likely(new_page)) {
1356                 prep_transhuge_page(new_page);
1357         } else {
1358                 if (!page) {
1359                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1360                         ret |= VM_FAULT_FALLBACK;
1361                 } else {
1362                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1363                         if (ret & VM_FAULT_OOM) {
1364                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1365                                 ret |= VM_FAULT_FALLBACK;
1366                         }
1367                         put_page(page);
1368                 }
1369                 count_vm_event(THP_FAULT_FALLBACK);
1370                 goto out;
1371         }
1372
1373         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1374                                         huge_gfp, &memcg, true))) {
1375                 put_page(new_page);
1376                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1377                 if (page)
1378                         put_page(page);
1379                 ret |= VM_FAULT_FALLBACK;
1380                 count_vm_event(THP_FAULT_FALLBACK);
1381                 goto out;
1382         }
1383
1384         count_vm_event(THP_FAULT_ALLOC);
1385         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1386
1387         if (!page)
1388                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1389         else
1390                 copy_user_huge_page(new_page, page, vmf->address,
1391                                     vma, HPAGE_PMD_NR);
1392         __SetPageUptodate(new_page);
1393
1394         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1395                                 haddr, haddr + HPAGE_PMD_SIZE);
1396         mmu_notifier_invalidate_range_start(&range);
1397
1398         spin_lock(vmf->ptl);
1399         if (page)
1400                 put_page(page);
1401         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1402                 spin_unlock(vmf->ptl);
1403                 mem_cgroup_cancel_charge(new_page, memcg, true);
1404                 put_page(new_page);
1405                 goto out_mn;
1406         } else {
1407                 pmd_t entry;
1408                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1409                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1410                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1411                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1412                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1413                 lru_cache_add_active_or_unevictable(new_page, vma);
1414                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1415                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1416                 if (!page) {
1417                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1418                 } else {
1419                         VM_BUG_ON_PAGE(!PageHead(page), page);
1420                         page_remove_rmap(page, true);
1421                         put_page(page);
1422                 }
1423                 ret |= VM_FAULT_WRITE;
1424         }
1425         spin_unlock(vmf->ptl);
1426 out_mn:
1427         /*
1428          * No need to double call mmu_notifier->invalidate_range() callback as
1429          * the above pmdp_huge_clear_flush_notify() did already call it.
1430          */
1431         mmu_notifier_invalidate_range_only_end(&range);
1432 out:
1433         return ret;
1434 out_unlock:
1435         spin_unlock(vmf->ptl);
1436         return ret;
1437 }
1438
1439 /*
1440  * FOLL_FORCE can write to even unwritable pmd's, but only
1441  * after we've gone through a COW cycle and they are dirty.
1442  */
1443 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1444 {
1445         return pmd_write(pmd) ||
1446                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1447 }
1448
1449 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1450                                    unsigned long addr,
1451                                    pmd_t *pmd,
1452                                    unsigned int flags)
1453 {
1454         struct mm_struct *mm = vma->vm_mm;
1455         struct page *page = NULL;
1456
1457         assert_spin_locked(pmd_lockptr(mm, pmd));
1458
1459         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1460                 goto out;
1461
1462         /* Avoid dumping huge zero page */
1463         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1464                 return ERR_PTR(-EFAULT);
1465
1466         /* Full NUMA hinting faults to serialise migration in fault paths */
1467         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1468                 goto out;
1469
1470         page = pmd_page(*pmd);
1471         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1472         if (flags & FOLL_TOUCH)
1473                 touch_pmd(vma, addr, pmd, flags);
1474         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1475                 /*
1476                  * We don't mlock() pte-mapped THPs. This way we can avoid
1477                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1478                  *
1479                  * For anon THP:
1480                  *
1481                  * In most cases the pmd is the only mapping of the page as we
1482                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1483                  * writable private mappings in populate_vma_page_range().
1484                  *
1485                  * The only scenario when we have the page shared here is if we
1486                  * mlocking read-only mapping shared over fork(). We skip
1487                  * mlocking such pages.
1488                  *
1489                  * For file THP:
1490                  *
1491                  * We can expect PageDoubleMap() to be stable under page lock:
1492                  * for file pages we set it in page_add_file_rmap(), which
1493                  * requires page to be locked.
1494                  */
1495
1496                 if (PageAnon(page) && compound_mapcount(page) != 1)
1497                         goto skip_mlock;
1498                 if (PageDoubleMap(page) || !page->mapping)
1499                         goto skip_mlock;
1500                 if (!trylock_page(page))
1501                         goto skip_mlock;
1502                 lru_add_drain();
1503                 if (page->mapping && !PageDoubleMap(page))
1504                         mlock_vma_page(page);
1505                 unlock_page(page);
1506         }
1507 skip_mlock:
1508         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1509         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1510         if (flags & FOLL_GET)
1511                 get_page(page);
1512
1513 out:
1514         return page;
1515 }
1516
1517 /* NUMA hinting page fault entry point for trans huge pmds */
1518 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1519 {
1520         struct vm_area_struct *vma = vmf->vma;
1521         struct anon_vma *anon_vma = NULL;
1522         struct page *page;
1523         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1524         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1525         int target_nid, last_cpupid = -1;
1526         bool page_locked;
1527         bool migrated = false;
1528         bool was_writable;
1529         int flags = 0;
1530
1531         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1532         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1533                 goto out_unlock;
1534
1535         /*
1536          * If there are potential migrations, wait for completion and retry
1537          * without disrupting NUMA hinting information. Do not relock and
1538          * check_same as the page may no longer be mapped.
1539          */
1540         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1541                 page = pmd_page(*vmf->pmd);
1542                 if (!get_page_unless_zero(page))
1543                         goto out_unlock;
1544                 spin_unlock(vmf->ptl);
1545                 put_and_wait_on_page_locked(page);
1546                 goto out;
1547         }
1548
1549         page = pmd_page(pmd);
1550         BUG_ON(is_huge_zero_page(page));
1551         page_nid = page_to_nid(page);
1552         last_cpupid = page_cpupid_last(page);
1553         count_vm_numa_event(NUMA_HINT_FAULTS);
1554         if (page_nid == this_nid) {
1555                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1556                 flags |= TNF_FAULT_LOCAL;
1557         }
1558
1559         /* See similar comment in do_numa_page for explanation */
1560         if (!pmd_savedwrite(pmd))
1561                 flags |= TNF_NO_GROUP;
1562
1563         /*
1564          * Acquire the page lock to serialise THP migrations but avoid dropping
1565          * page_table_lock if at all possible
1566          */
1567         page_locked = trylock_page(page);
1568         target_nid = mpol_misplaced(page, vma, haddr);
1569         if (target_nid == NUMA_NO_NODE) {
1570                 /* If the page was locked, there are no parallel migrations */
1571                 if (page_locked)
1572                         goto clear_pmdnuma;
1573         }
1574
1575         /* Migration could have started since the pmd_trans_migrating check */
1576         if (!page_locked) {
1577                 page_nid = NUMA_NO_NODE;
1578                 if (!get_page_unless_zero(page))
1579                         goto out_unlock;
1580                 spin_unlock(vmf->ptl);
1581                 put_and_wait_on_page_locked(page);
1582                 goto out;
1583         }
1584
1585         /*
1586          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1587          * to serialises splits
1588          */
1589         get_page(page);
1590         spin_unlock(vmf->ptl);
1591         anon_vma = page_lock_anon_vma_read(page);
1592
1593         /* Confirm the PMD did not change while page_table_lock was released */
1594         spin_lock(vmf->ptl);
1595         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1596                 unlock_page(page);
1597                 put_page(page);
1598                 page_nid = NUMA_NO_NODE;
1599                 goto out_unlock;
1600         }
1601
1602         /* Bail if we fail to protect against THP splits for any reason */
1603         if (unlikely(!anon_vma)) {
1604                 put_page(page);
1605                 page_nid = NUMA_NO_NODE;
1606                 goto clear_pmdnuma;
1607         }
1608
1609         /*
1610          * Since we took the NUMA fault, we must have observed the !accessible
1611          * bit. Make sure all other CPUs agree with that, to avoid them
1612          * modifying the page we're about to migrate.
1613          *
1614          * Must be done under PTL such that we'll observe the relevant
1615          * inc_tlb_flush_pending().
1616          *
1617          * We are not sure a pending tlb flush here is for a huge page
1618          * mapping or not. Hence use the tlb range variant
1619          */
1620         if (mm_tlb_flush_pending(vma->vm_mm)) {
1621                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1622                 /*
1623                  * change_huge_pmd() released the pmd lock before
1624                  * invalidating the secondary MMUs sharing the primary
1625                  * MMU pagetables (with ->invalidate_range()). The
1626                  * mmu_notifier_invalidate_range_end() (which
1627                  * internally calls ->invalidate_range()) in
1628                  * change_pmd_range() will run after us, so we can't
1629                  * rely on it here and we need an explicit invalidate.
1630                  */
1631                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1632                                               haddr + HPAGE_PMD_SIZE);
1633         }
1634
1635         /*
1636          * Migrate the THP to the requested node, returns with page unlocked
1637          * and access rights restored.
1638          */
1639         spin_unlock(vmf->ptl);
1640
1641         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1642                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1643         if (migrated) {
1644                 flags |= TNF_MIGRATED;
1645                 page_nid = target_nid;
1646         } else
1647                 flags |= TNF_MIGRATE_FAIL;
1648
1649         goto out;
1650 clear_pmdnuma:
1651         BUG_ON(!PageLocked(page));
1652         was_writable = pmd_savedwrite(pmd);
1653         pmd = pmd_modify(pmd, vma->vm_page_prot);
1654         pmd = pmd_mkyoung(pmd);
1655         if (was_writable)
1656                 pmd = pmd_mkwrite(pmd);
1657         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1658         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1659         unlock_page(page);
1660 out_unlock:
1661         spin_unlock(vmf->ptl);
1662
1663 out:
1664         if (anon_vma)
1665                 page_unlock_anon_vma_read(anon_vma);
1666
1667         if (page_nid != NUMA_NO_NODE)
1668                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1669                                 flags);
1670
1671         return 0;
1672 }
1673
1674 /*
1675  * Return true if we do MADV_FREE successfully on entire pmd page.
1676  * Otherwise, return false.
1677  */
1678 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1679                 pmd_t *pmd, unsigned long addr, unsigned long next)
1680 {
1681         spinlock_t *ptl;
1682         pmd_t orig_pmd;
1683         struct page *page;
1684         struct mm_struct *mm = tlb->mm;
1685         bool ret = false;
1686
1687         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1688
1689         ptl = pmd_trans_huge_lock(pmd, vma);
1690         if (!ptl)
1691                 goto out_unlocked;
1692
1693         orig_pmd = *pmd;
1694         if (is_huge_zero_pmd(orig_pmd))
1695                 goto out;
1696
1697         if (unlikely(!pmd_present(orig_pmd))) {
1698                 VM_BUG_ON(thp_migration_supported() &&
1699                                   !is_pmd_migration_entry(orig_pmd));
1700                 goto out;
1701         }
1702
1703         page = pmd_page(orig_pmd);
1704         /*
1705          * If other processes are mapping this page, we couldn't discard
1706          * the page unless they all do MADV_FREE so let's skip the page.
1707          */
1708         if (page_mapcount(page) != 1)
1709                 goto out;
1710
1711         if (!trylock_page(page))
1712                 goto out;
1713
1714         /*
1715          * If user want to discard part-pages of THP, split it so MADV_FREE
1716          * will deactivate only them.
1717          */
1718         if (next - addr != HPAGE_PMD_SIZE) {
1719                 get_page(page);
1720                 spin_unlock(ptl);
1721                 split_huge_page(page);
1722                 unlock_page(page);
1723                 put_page(page);
1724                 goto out_unlocked;
1725         }
1726
1727         if (PageDirty(page))
1728                 ClearPageDirty(page);
1729         unlock_page(page);
1730
1731         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1732                 pmdp_invalidate(vma, addr, pmd);
1733                 orig_pmd = pmd_mkold(orig_pmd);
1734                 orig_pmd = pmd_mkclean(orig_pmd);
1735
1736                 set_pmd_at(mm, addr, pmd, orig_pmd);
1737                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1738         }
1739
1740         mark_page_lazyfree(page);
1741         ret = true;
1742 out:
1743         spin_unlock(ptl);
1744 out_unlocked:
1745         return ret;
1746 }
1747
1748 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1749 {
1750         pgtable_t pgtable;
1751
1752         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1753         pte_free(mm, pgtable);
1754         mm_dec_nr_ptes(mm);
1755 }
1756
1757 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1758                  pmd_t *pmd, unsigned long addr)
1759 {
1760         pmd_t orig_pmd;
1761         spinlock_t *ptl;
1762
1763         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1764
1765         ptl = __pmd_trans_huge_lock(pmd, vma);
1766         if (!ptl)
1767                 return 0;
1768         /*
1769          * For architectures like ppc64 we look at deposited pgtable
1770          * when calling pmdp_huge_get_and_clear. So do the
1771          * pgtable_trans_huge_withdraw after finishing pmdp related
1772          * operations.
1773          */
1774         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1775                         tlb->fullmm);
1776         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1777         if (vma_is_dax(vma)) {
1778                 if (arch_needs_pgtable_deposit())
1779                         zap_deposited_table(tlb->mm, pmd);
1780                 spin_unlock(ptl);
1781                 if (is_huge_zero_pmd(orig_pmd))
1782                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1783         } else if (is_huge_zero_pmd(orig_pmd)) {
1784                 zap_deposited_table(tlb->mm, pmd);
1785                 spin_unlock(ptl);
1786                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1787         } else {
1788                 struct page *page = NULL;
1789                 int flush_needed = 1;
1790
1791                 if (pmd_present(orig_pmd)) {
1792                         page = pmd_page(orig_pmd);
1793                         page_remove_rmap(page, true);
1794                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1795                         VM_BUG_ON_PAGE(!PageHead(page), page);
1796                 } else if (thp_migration_supported()) {
1797                         swp_entry_t entry;
1798
1799                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1800                         entry = pmd_to_swp_entry(orig_pmd);
1801                         page = pfn_to_page(swp_offset(entry));
1802                         flush_needed = 0;
1803                 } else
1804                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1805
1806                 if (PageAnon(page)) {
1807                         zap_deposited_table(tlb->mm, pmd);
1808                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1809                 } else {
1810                         if (arch_needs_pgtable_deposit())
1811                                 zap_deposited_table(tlb->mm, pmd);
1812                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1813                 }
1814
1815                 spin_unlock(ptl);
1816                 if (flush_needed)
1817                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1818         }
1819         return 1;
1820 }
1821
1822 #ifndef pmd_move_must_withdraw
1823 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1824                                          spinlock_t *old_pmd_ptl,
1825                                          struct vm_area_struct *vma)
1826 {
1827         /*
1828          * With split pmd lock we also need to move preallocated
1829          * PTE page table if new_pmd is on different PMD page table.
1830          *
1831          * We also don't deposit and withdraw tables for file pages.
1832          */
1833         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1834 }
1835 #endif
1836
1837 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1838 {
1839 #ifdef CONFIG_MEM_SOFT_DIRTY
1840         if (unlikely(is_pmd_migration_entry(pmd)))
1841                 pmd = pmd_swp_mksoft_dirty(pmd);
1842         else if (pmd_present(pmd))
1843                 pmd = pmd_mksoft_dirty(pmd);
1844 #endif
1845         return pmd;
1846 }
1847
1848 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1849                   unsigned long new_addr, unsigned long old_end,
1850                   pmd_t *old_pmd, pmd_t *new_pmd)
1851 {
1852         spinlock_t *old_ptl, *new_ptl;
1853         pmd_t pmd;
1854         struct mm_struct *mm = vma->vm_mm;
1855         bool force_flush = false;
1856
1857         if ((old_addr & ~HPAGE_PMD_MASK) ||
1858             (new_addr & ~HPAGE_PMD_MASK) ||
1859             old_end - old_addr < HPAGE_PMD_SIZE)
1860                 return false;
1861
1862         /*
1863          * The destination pmd shouldn't be established, free_pgtables()
1864          * should have release it.
1865          */
1866         if (WARN_ON(!pmd_none(*new_pmd))) {
1867                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1868                 return false;
1869         }
1870
1871         /*
1872          * We don't have to worry about the ordering of src and dst
1873          * ptlocks because exclusive mmap_sem prevents deadlock.
1874          */
1875         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1876         if (old_ptl) {
1877                 new_ptl = pmd_lockptr(mm, new_pmd);
1878                 if (new_ptl != old_ptl)
1879                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1880                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1881                 if (pmd_present(pmd))
1882                         force_flush = true;
1883                 VM_BUG_ON(!pmd_none(*new_pmd));
1884
1885                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1886                         pgtable_t pgtable;
1887                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1888                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1889                 }
1890                 pmd = move_soft_dirty_pmd(pmd);
1891                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1892                 if (force_flush)
1893                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1894                 if (new_ptl != old_ptl)
1895                         spin_unlock(new_ptl);
1896                 spin_unlock(old_ptl);
1897                 return true;
1898         }
1899         return false;
1900 }
1901
1902 /*
1903  * Returns
1904  *  - 0 if PMD could not be locked
1905  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1906  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1907  */
1908 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1909                 unsigned long addr, pgprot_t newprot, int prot_numa)
1910 {
1911         struct mm_struct *mm = vma->vm_mm;
1912         spinlock_t *ptl;
1913         pmd_t entry;
1914         bool preserve_write;
1915         int ret;
1916
1917         ptl = __pmd_trans_huge_lock(pmd, vma);
1918         if (!ptl)
1919                 return 0;
1920
1921         preserve_write = prot_numa && pmd_write(*pmd);
1922         ret = 1;
1923
1924 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1925         if (is_swap_pmd(*pmd)) {
1926                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1927
1928                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1929                 if (is_write_migration_entry(entry)) {
1930                         pmd_t newpmd;
1931                         /*
1932                          * A protection check is difficult so
1933                          * just be safe and disable write
1934                          */
1935                         make_migration_entry_read(&entry);
1936                         newpmd = swp_entry_to_pmd(entry);
1937                         if (pmd_swp_soft_dirty(*pmd))
1938                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1939                         set_pmd_at(mm, addr, pmd, newpmd);
1940                 }
1941                 goto unlock;
1942         }
1943 #endif
1944
1945         /*
1946          * Avoid trapping faults against the zero page. The read-only
1947          * data is likely to be read-cached on the local CPU and
1948          * local/remote hits to the zero page are not interesting.
1949          */
1950         if (prot_numa && is_huge_zero_pmd(*pmd))
1951                 goto unlock;
1952
1953         if (prot_numa && pmd_protnone(*pmd))
1954                 goto unlock;
1955
1956         /*
1957          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1958          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1959          * which is also under down_read(mmap_sem):
1960          *
1961          *      CPU0:                           CPU1:
1962          *                              change_huge_pmd(prot_numa=1)
1963          *                               pmdp_huge_get_and_clear_notify()
1964          * madvise_dontneed()
1965          *  zap_pmd_range()
1966          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1967          *   // skip the pmd
1968          *                               set_pmd_at();
1969          *                               // pmd is re-established
1970          *
1971          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1972          * which may break userspace.
1973          *
1974          * pmdp_invalidate() is required to make sure we don't miss
1975          * dirty/young flags set by hardware.
1976          */
1977         entry = pmdp_invalidate(vma, addr, pmd);
1978
1979         entry = pmd_modify(entry, newprot);
1980         if (preserve_write)
1981                 entry = pmd_mk_savedwrite(entry);
1982         ret = HPAGE_PMD_NR;
1983         set_pmd_at(mm, addr, pmd, entry);
1984         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1985 unlock:
1986         spin_unlock(ptl);
1987         return ret;
1988 }
1989
1990 /*
1991  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1992  *
1993  * Note that if it returns page table lock pointer, this routine returns without
1994  * unlocking page table lock. So callers must unlock it.
1995  */
1996 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1997 {
1998         spinlock_t *ptl;
1999         ptl = pmd_lock(vma->vm_mm, pmd);
2000         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2001                         pmd_devmap(*pmd)))
2002                 return ptl;
2003         spin_unlock(ptl);
2004         return NULL;
2005 }
2006
2007 /*
2008  * Returns true if a given pud maps a thp, false otherwise.
2009  *
2010  * Note that if it returns true, this routine returns without unlocking page
2011  * table lock. So callers must unlock it.
2012  */
2013 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2014 {
2015         spinlock_t *ptl;
2016
2017         ptl = pud_lock(vma->vm_mm, pud);
2018         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2019                 return ptl;
2020         spin_unlock(ptl);
2021         return NULL;
2022 }
2023
2024 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2025 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2026                  pud_t *pud, unsigned long addr)
2027 {
2028         spinlock_t *ptl;
2029
2030         ptl = __pud_trans_huge_lock(pud, vma);
2031         if (!ptl)
2032                 return 0;
2033         /*
2034          * For architectures like ppc64 we look at deposited pgtable
2035          * when calling pudp_huge_get_and_clear. So do the
2036          * pgtable_trans_huge_withdraw after finishing pudp related
2037          * operations.
2038          */
2039         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2040         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2041         if (vma_is_dax(vma)) {
2042                 spin_unlock(ptl);
2043                 /* No zero page support yet */
2044         } else {
2045                 /* No support for anonymous PUD pages yet */
2046                 BUG();
2047         }
2048         return 1;
2049 }
2050
2051 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2052                 unsigned long haddr)
2053 {
2054         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2055         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2056         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2057         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2058
2059         count_vm_event(THP_SPLIT_PUD);
2060
2061         pudp_huge_clear_flush_notify(vma, haddr, pud);
2062 }
2063
2064 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2065                 unsigned long address)
2066 {
2067         spinlock_t *ptl;
2068         struct mmu_notifier_range range;
2069
2070         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2071                                 address & HPAGE_PUD_MASK,
2072                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2073         mmu_notifier_invalidate_range_start(&range);
2074         ptl = pud_lock(vma->vm_mm, pud);
2075         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2076                 goto out;
2077         __split_huge_pud_locked(vma, pud, range.start);
2078
2079 out:
2080         spin_unlock(ptl);
2081         /*
2082          * No need to double call mmu_notifier->invalidate_range() callback as
2083          * the above pudp_huge_clear_flush_notify() did already call it.
2084          */
2085         mmu_notifier_invalidate_range_only_end(&range);
2086 }
2087 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2088
2089 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2090                 unsigned long haddr, pmd_t *pmd)
2091 {
2092         struct mm_struct *mm = vma->vm_mm;
2093         pgtable_t pgtable;
2094         pmd_t _pmd;
2095         int i;
2096
2097         /*
2098          * Leave pmd empty until pte is filled note that it is fine to delay
2099          * notification until mmu_notifier_invalidate_range_end() as we are
2100          * replacing a zero pmd write protected page with a zero pte write
2101          * protected page.
2102          *
2103          * See Documentation/vm/mmu_notifier.rst
2104          */
2105         pmdp_huge_clear_flush(vma, haddr, pmd);
2106
2107         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2108         pmd_populate(mm, &_pmd, pgtable);
2109
2110         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2111                 pte_t *pte, entry;
2112                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2113                 entry = pte_mkspecial(entry);
2114                 pte = pte_offset_map(&_pmd, haddr);
2115                 VM_BUG_ON(!pte_none(*pte));
2116                 set_pte_at(mm, haddr, pte, entry);
2117                 pte_unmap(pte);
2118         }
2119         smp_wmb(); /* make pte visible before pmd */
2120         pmd_populate(mm, pmd, pgtable);
2121 }
2122
2123 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2124                 unsigned long haddr, bool freeze)
2125 {
2126         struct mm_struct *mm = vma->vm_mm;
2127         struct page *page;
2128         pgtable_t pgtable;
2129         pmd_t old_pmd, _pmd;
2130         bool young, write, soft_dirty, pmd_migration = false;
2131         unsigned long addr;
2132         int i;
2133
2134         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2135         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2136         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2137         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2138                                 && !pmd_devmap(*pmd));
2139
2140         count_vm_event(THP_SPLIT_PMD);
2141
2142         if (!vma_is_anonymous(vma)) {
2143                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2144                 /*
2145                  * We are going to unmap this huge page. So
2146                  * just go ahead and zap it
2147                  */
2148                 if (arch_needs_pgtable_deposit())
2149                         zap_deposited_table(mm, pmd);
2150                 if (vma_is_dax(vma))
2151                         return;
2152                 page = pmd_page(_pmd);
2153                 if (!PageDirty(page) && pmd_dirty(_pmd))
2154                         set_page_dirty(page);
2155                 if (!PageReferenced(page) && pmd_young(_pmd))
2156                         SetPageReferenced(page);
2157                 page_remove_rmap(page, true);
2158                 put_page(page);
2159                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2160                 return;
2161         } else if (is_huge_zero_pmd(*pmd)) {
2162                 /*
2163                  * FIXME: Do we want to invalidate secondary mmu by calling
2164                  * mmu_notifier_invalidate_range() see comments below inside
2165                  * __split_huge_pmd() ?
2166                  *
2167                  * We are going from a zero huge page write protected to zero
2168                  * small page also write protected so it does not seems useful
2169                  * to invalidate secondary mmu at this time.
2170                  */
2171                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2172         }
2173
2174         /*
2175          * Up to this point the pmd is present and huge and userland has the
2176          * whole access to the hugepage during the split (which happens in
2177          * place). If we overwrite the pmd with the not-huge version pointing
2178          * to the pte here (which of course we could if all CPUs were bug
2179          * free), userland could trigger a small page size TLB miss on the
2180          * small sized TLB while the hugepage TLB entry is still established in
2181          * the huge TLB. Some CPU doesn't like that.
2182          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2183          * 383 on page 93. Intel should be safe but is also warns that it's
2184          * only safe if the permission and cache attributes of the two entries
2185          * loaded in the two TLB is identical (which should be the case here).
2186          * But it is generally safer to never allow small and huge TLB entries
2187          * for the same virtual address to be loaded simultaneously. So instead
2188          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2189          * current pmd notpresent (atomically because here the pmd_trans_huge
2190          * must remain set at all times on the pmd until the split is complete
2191          * for this pmd), then we flush the SMP TLB and finally we write the
2192          * non-huge version of the pmd entry with pmd_populate.
2193          */
2194         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2195
2196         pmd_migration = is_pmd_migration_entry(old_pmd);
2197         if (unlikely(pmd_migration)) {
2198                 swp_entry_t entry;
2199
2200                 entry = pmd_to_swp_entry(old_pmd);
2201                 page = pfn_to_page(swp_offset(entry));
2202                 write = is_write_migration_entry(entry);
2203                 young = false;
2204                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2205         } else {
2206                 page = pmd_page(old_pmd);
2207                 if (pmd_dirty(old_pmd))
2208                         SetPageDirty(page);
2209                 write = pmd_write(old_pmd);
2210                 young = pmd_young(old_pmd);
2211                 soft_dirty = pmd_soft_dirty(old_pmd);
2212         }
2213         VM_BUG_ON_PAGE(!page_count(page), page);
2214         page_ref_add(page, HPAGE_PMD_NR - 1);
2215
2216         /*
2217          * Withdraw the table only after we mark the pmd entry invalid.
2218          * This's critical for some architectures (Power).
2219          */
2220         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2221         pmd_populate(mm, &_pmd, pgtable);
2222
2223         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2224                 pte_t entry, *pte;
2225                 /*
2226                  * Note that NUMA hinting access restrictions are not
2227                  * transferred to avoid any possibility of altering
2228                  * permissions across VMAs.
2229                  */
2230                 if (freeze || pmd_migration) {
2231                         swp_entry_t swp_entry;
2232                         swp_entry = make_migration_entry(page + i, write);
2233                         entry = swp_entry_to_pte(swp_entry);
2234                         if (soft_dirty)
2235                                 entry = pte_swp_mksoft_dirty(entry);
2236                 } else {
2237                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2238                         entry = maybe_mkwrite(entry, vma);
2239                         if (!write)
2240                                 entry = pte_wrprotect(entry);
2241                         if (!young)
2242                                 entry = pte_mkold(entry);
2243                         if (soft_dirty)
2244                                 entry = pte_mksoft_dirty(entry);
2245                 }
2246                 pte = pte_offset_map(&_pmd, addr);
2247                 BUG_ON(!pte_none(*pte));
2248                 set_pte_at(mm, addr, pte, entry);
2249                 atomic_inc(&page[i]._mapcount);
2250                 pte_unmap(pte);
2251         }
2252
2253         /*
2254          * Set PG_double_map before dropping compound_mapcount to avoid
2255          * false-negative page_mapped().
2256          */
2257         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2258                 for (i = 0; i < HPAGE_PMD_NR; i++)
2259                         atomic_inc(&page[i]._mapcount);
2260         }
2261
2262         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2263                 /* Last compound_mapcount is gone. */
2264                 __dec_node_page_state(page, NR_ANON_THPS);
2265                 if (TestClearPageDoubleMap(page)) {
2266                         /* No need in mapcount reference anymore */
2267                         for (i = 0; i < HPAGE_PMD_NR; i++)
2268                                 atomic_dec(&page[i]._mapcount);
2269                 }
2270         }
2271
2272         smp_wmb(); /* make pte visible before pmd */
2273         pmd_populate(mm, pmd, pgtable);
2274
2275         if (freeze) {
2276                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2277                         page_remove_rmap(page + i, false);
2278                         put_page(page + i);
2279                 }
2280         }
2281 }
2282
2283 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2284                 unsigned long address, bool freeze, struct page *page)
2285 {
2286         spinlock_t *ptl;
2287         struct mmu_notifier_range range;
2288
2289         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2290                                 address & HPAGE_PMD_MASK,
2291                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2292         mmu_notifier_invalidate_range_start(&range);
2293         ptl = pmd_lock(vma->vm_mm, pmd);
2294
2295         /*
2296          * If caller asks to setup a migration entries, we need a page to check
2297          * pmd against. Otherwise we can end up replacing wrong page.
2298          */
2299         VM_BUG_ON(freeze && !page);
2300         if (page && page != pmd_page(*pmd))
2301                 goto out;
2302
2303         if (pmd_trans_huge(*pmd)) {
2304                 page = pmd_page(*pmd);
2305                 if (PageMlocked(page))
2306                         clear_page_mlock(page);
2307         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2308                 goto out;
2309         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2310 out:
2311         spin_unlock(ptl);
2312         /*
2313          * No need to double call mmu_notifier->invalidate_range() callback.
2314          * They are 3 cases to consider inside __split_huge_pmd_locked():
2315          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2316          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2317          *    fault will trigger a flush_notify before pointing to a new page
2318          *    (it is fine if the secondary mmu keeps pointing to the old zero
2319          *    page in the meantime)
2320          *  3) Split a huge pmd into pte pointing to the same page. No need
2321          *     to invalidate secondary tlb entry they are all still valid.
2322          *     any further changes to individual pte will notify. So no need
2323          *     to call mmu_notifier->invalidate_range()
2324          */
2325         mmu_notifier_invalidate_range_only_end(&range);
2326 }
2327
2328 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2329                 bool freeze, struct page *page)
2330 {
2331         pgd_t *pgd;
2332         p4d_t *p4d;
2333         pud_t *pud;
2334         pmd_t *pmd;
2335
2336         pgd = pgd_offset(vma->vm_mm, address);
2337         if (!pgd_present(*pgd))
2338                 return;
2339
2340         p4d = p4d_offset(pgd, address);
2341         if (!p4d_present(*p4d))
2342                 return;
2343
2344         pud = pud_offset(p4d, address);
2345         if (!pud_present(*pud))
2346                 return;
2347
2348         pmd = pmd_offset(pud, address);
2349
2350         __split_huge_pmd(vma, pmd, address, freeze, page);
2351 }
2352
2353 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2354                              unsigned long start,
2355                              unsigned long end,
2356                              long adjust_next)
2357 {
2358         /*
2359          * If the new start address isn't hpage aligned and it could
2360          * previously contain an hugepage: check if we need to split
2361          * an huge pmd.
2362          */
2363         if (start & ~HPAGE_PMD_MASK &&
2364             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2365             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2366                 split_huge_pmd_address(vma, start, false, NULL);
2367
2368         /*
2369          * If the new end address isn't hpage aligned and it could
2370          * previously contain an hugepage: check if we need to split
2371          * an huge pmd.
2372          */
2373         if (end & ~HPAGE_PMD_MASK &&
2374             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2375             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2376                 split_huge_pmd_address(vma, end, false, NULL);
2377
2378         /*
2379          * If we're also updating the vma->vm_next->vm_start, if the new
2380          * vm_next->vm_start isn't page aligned and it could previously
2381          * contain an hugepage: check if we need to split an huge pmd.
2382          */
2383         if (adjust_next > 0) {
2384                 struct vm_area_struct *next = vma->vm_next;
2385                 unsigned long nstart = next->vm_start;
2386                 nstart += adjust_next << PAGE_SHIFT;
2387                 if (nstart & ~HPAGE_PMD_MASK &&
2388                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2389                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2390                         split_huge_pmd_address(next, nstart, false, NULL);
2391         }
2392 }
2393
2394 static void unmap_page(struct page *page)
2395 {
2396         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2397                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2398         bool unmap_success;
2399
2400         VM_BUG_ON_PAGE(!PageHead(page), page);
2401
2402         if (PageAnon(page))
2403                 ttu_flags |= TTU_SPLIT_FREEZE;
2404
2405         unmap_success = try_to_unmap(page, ttu_flags);
2406         VM_BUG_ON_PAGE(!unmap_success, page);
2407 }
2408
2409 static void remap_page(struct page *page)
2410 {
2411         int i;
2412         if (PageTransHuge(page)) {
2413                 remove_migration_ptes(page, page, true);
2414         } else {
2415                 for (i = 0; i < HPAGE_PMD_NR; i++)
2416                         remove_migration_ptes(page + i, page + i, true);
2417         }
2418 }
2419
2420 static void __split_huge_page_tail(struct page *head, int tail,
2421                 struct lruvec *lruvec, struct list_head *list)
2422 {
2423         struct page *page_tail = head + tail;
2424
2425         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2426
2427         /*
2428          * Clone page flags before unfreezing refcount.
2429          *
2430          * After successful get_page_unless_zero() might follow flags change,
2431          * for exmaple lock_page() which set PG_waiters.
2432          */
2433         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2434         page_tail->flags |= (head->flags &
2435                         ((1L << PG_referenced) |
2436                          (1L << PG_swapbacked) |
2437                          (1L << PG_swapcache) |
2438                          (1L << PG_mlocked) |
2439                          (1L << PG_uptodate) |
2440                          (1L << PG_active) |
2441                          (1L << PG_workingset) |
2442                          (1L << PG_locked) |
2443                          (1L << PG_unevictable) |
2444                          (1L << PG_dirty)));
2445
2446         /* ->mapping in first tail page is compound_mapcount */
2447         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2448                         page_tail);
2449         page_tail->mapping = head->mapping;
2450         page_tail->index = head->index + tail;
2451
2452         /* Page flags must be visible before we make the page non-compound. */
2453         smp_wmb();
2454
2455         /*
2456          * Clear PageTail before unfreezing page refcount.
2457          *
2458          * After successful get_page_unless_zero() might follow put_page()
2459          * which needs correct compound_head().
2460          */
2461         clear_compound_head(page_tail);
2462
2463         /* Finally unfreeze refcount. Additional reference from page cache. */
2464         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2465                                           PageSwapCache(head)));
2466
2467         if (page_is_young(head))
2468                 set_page_young(page_tail);
2469         if (page_is_idle(head))
2470                 set_page_idle(page_tail);
2471
2472         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2473
2474         /*
2475          * always add to the tail because some iterators expect new
2476          * pages to show after the currently processed elements - e.g.
2477          * migrate_pages
2478          */
2479         lru_add_page_tail(head, page_tail, lruvec, list);
2480 }
2481
2482 static void __split_huge_page(struct page *page, struct list_head *list,
2483                 pgoff_t end, unsigned long flags)
2484 {
2485         struct page *head = compound_head(page);
2486         pg_data_t *pgdat = page_pgdat(head);
2487         struct lruvec *lruvec;
2488         int i;
2489
2490         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2491
2492         /* complete memcg works before add pages to LRU */
2493         mem_cgroup_split_huge_fixup(head);
2494
2495         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2496                 __split_huge_page_tail(head, i, lruvec, list);
2497                 /* Some pages can be beyond i_size: drop them from page cache */
2498                 if (head[i].index >= end) {
2499                         ClearPageDirty(head + i);
2500                         __delete_from_page_cache(head + i, NULL);
2501                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2502                                 shmem_uncharge(head->mapping->host, 1);
2503                         put_page(head + i);
2504                 }
2505         }
2506
2507         ClearPageCompound(head);
2508         /* See comment in __split_huge_page_tail() */
2509         if (PageAnon(head)) {
2510                 /* Additional pin to swap cache */
2511                 if (PageSwapCache(head))
2512                         page_ref_add(head, 2);
2513                 else
2514                         page_ref_inc(head);
2515         } else {
2516                 /* Additional pin to page cache */
2517                 page_ref_add(head, 2);
2518                 xa_unlock(&head->mapping->i_pages);
2519         }
2520
2521         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2522
2523         remap_page(head);
2524
2525         for (i = 0; i < HPAGE_PMD_NR; i++) {
2526                 struct page *subpage = head + i;
2527                 if (subpage == page)
2528                         continue;
2529                 unlock_page(subpage);
2530
2531                 /*
2532                  * Subpages may be freed if there wasn't any mapping
2533                  * like if add_to_swap() is running on a lru page that
2534                  * had its mapping zapped. And freeing these pages
2535                  * requires taking the lru_lock so we do the put_page
2536                  * of the tail pages after the split is complete.
2537                  */
2538                 put_page(subpage);
2539         }
2540 }
2541
2542 int total_mapcount(struct page *page)
2543 {
2544         int i, compound, ret;
2545
2546         VM_BUG_ON_PAGE(PageTail(page), page);
2547
2548         if (likely(!PageCompound(page)))
2549                 return atomic_read(&page->_mapcount) + 1;
2550
2551         compound = compound_mapcount(page);
2552         if (PageHuge(page))
2553                 return compound;
2554         ret = compound;
2555         for (i = 0; i < HPAGE_PMD_NR; i++)
2556                 ret += atomic_read(&page[i]._mapcount) + 1;
2557         /* File pages has compound_mapcount included in _mapcount */
2558         if (!PageAnon(page))
2559                 return ret - compound * HPAGE_PMD_NR;
2560         if (PageDoubleMap(page))
2561                 ret -= HPAGE_PMD_NR;
2562         return ret;
2563 }
2564
2565 /*
2566  * This calculates accurately how many mappings a transparent hugepage
2567  * has (unlike page_mapcount() which isn't fully accurate). This full
2568  * accuracy is primarily needed to know if copy-on-write faults can
2569  * reuse the page and change the mapping to read-write instead of
2570  * copying them. At the same time this returns the total_mapcount too.
2571  *
2572  * The function returns the highest mapcount any one of the subpages
2573  * has. If the return value is one, even if different processes are
2574  * mapping different subpages of the transparent hugepage, they can
2575  * all reuse it, because each process is reusing a different subpage.
2576  *
2577  * The total_mapcount is instead counting all virtual mappings of the
2578  * subpages. If the total_mapcount is equal to "one", it tells the
2579  * caller all mappings belong to the same "mm" and in turn the
2580  * anon_vma of the transparent hugepage can become the vma->anon_vma
2581  * local one as no other process may be mapping any of the subpages.
2582  *
2583  * It would be more accurate to replace page_mapcount() with
2584  * page_trans_huge_mapcount(), however we only use
2585  * page_trans_huge_mapcount() in the copy-on-write faults where we
2586  * need full accuracy to avoid breaking page pinning, because
2587  * page_trans_huge_mapcount() is slower than page_mapcount().
2588  */
2589 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2590 {
2591         int i, ret, _total_mapcount, mapcount;
2592
2593         /* hugetlbfs shouldn't call it */
2594         VM_BUG_ON_PAGE(PageHuge(page), page);
2595
2596         if (likely(!PageTransCompound(page))) {
2597                 mapcount = atomic_read(&page->_mapcount) + 1;
2598                 if (total_mapcount)
2599                         *total_mapcount = mapcount;
2600                 return mapcount;
2601         }
2602
2603         page = compound_head(page);
2604
2605         _total_mapcount = ret = 0;
2606         for (i = 0; i < HPAGE_PMD_NR; i++) {
2607                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2608                 ret = max(ret, mapcount);
2609                 _total_mapcount += mapcount;
2610         }
2611         if (PageDoubleMap(page)) {
2612                 ret -= 1;
2613                 _total_mapcount -= HPAGE_PMD_NR;
2614         }
2615         mapcount = compound_mapcount(page);
2616         ret += mapcount;
2617         _total_mapcount += mapcount;
2618         if (total_mapcount)
2619                 *total_mapcount = _total_mapcount;
2620         return ret;
2621 }
2622
2623 /* Racy check whether the huge page can be split */
2624 bool can_split_huge_page(struct page *page, int *pextra_pins)
2625 {
2626         int extra_pins;
2627
2628         /* Additional pins from page cache */
2629         if (PageAnon(page))
2630                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2631         else
2632                 extra_pins = HPAGE_PMD_NR;
2633         if (pextra_pins)
2634                 *pextra_pins = extra_pins;
2635         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2636 }
2637
2638 /*
2639  * This function splits huge page into normal pages. @page can point to any
2640  * subpage of huge page to split. Split doesn't change the position of @page.
2641  *
2642  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2643  * The huge page must be locked.
2644  *
2645  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2646  *
2647  * Both head page and tail pages will inherit mapping, flags, and so on from
2648  * the hugepage.
2649  *
2650  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2651  * they are not mapped.
2652  *
2653  * Returns 0 if the hugepage is split successfully.
2654  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2655  * us.
2656  */
2657 int split_huge_page_to_list(struct page *page, struct list_head *list)
2658 {
2659         struct page *head = compound_head(page);
2660         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2661         struct anon_vma *anon_vma = NULL;
2662         struct address_space *mapping = NULL;
2663         int count, mapcount, extra_pins, ret;
2664         bool mlocked;
2665         unsigned long flags;
2666         pgoff_t end;
2667
2668         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2669         VM_BUG_ON_PAGE(!PageLocked(page), page);
2670         VM_BUG_ON_PAGE(!PageCompound(page), page);
2671
2672         if (PageWriteback(page))
2673                 return -EBUSY;
2674
2675         if (PageAnon(head)) {
2676                 /*
2677                  * The caller does not necessarily hold an mmap_sem that would
2678                  * prevent the anon_vma disappearing so we first we take a
2679                  * reference to it and then lock the anon_vma for write. This
2680                  * is similar to page_lock_anon_vma_read except the write lock
2681                  * is taken to serialise against parallel split or collapse
2682                  * operations.
2683                  */
2684                 anon_vma = page_get_anon_vma(head);
2685                 if (!anon_vma) {
2686                         ret = -EBUSY;
2687                         goto out;
2688                 }
2689                 end = -1;
2690                 mapping = NULL;
2691                 anon_vma_lock_write(anon_vma);
2692         } else {
2693                 mapping = head->mapping;
2694
2695                 /* Truncated ? */
2696                 if (!mapping) {
2697                         ret = -EBUSY;
2698                         goto out;
2699                 }
2700
2701                 anon_vma = NULL;
2702                 i_mmap_lock_read(mapping);
2703
2704                 /*
2705                  *__split_huge_page() may need to trim off pages beyond EOF:
2706                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2707                  * which cannot be nested inside the page tree lock. So note
2708                  * end now: i_size itself may be changed at any moment, but
2709                  * head page lock is good enough to serialize the trimming.
2710                  */
2711                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2712         }
2713
2714         /*
2715          * Racy check if we can split the page, before unmap_page() will
2716          * split PMDs
2717          */
2718         if (!can_split_huge_page(head, &extra_pins)) {
2719                 ret = -EBUSY;
2720                 goto out_unlock;
2721         }
2722
2723         mlocked = PageMlocked(page);
2724         unmap_page(head);
2725         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2726
2727         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2728         if (mlocked)
2729                 lru_add_drain();
2730
2731         /* prevent PageLRU to go away from under us, and freeze lru stats */
2732         spin_lock_irqsave(&pgdata->lru_lock, flags);
2733
2734         if (mapping) {
2735                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2736
2737                 /*
2738                  * Check if the head page is present in page cache.
2739                  * We assume all tail are present too, if head is there.
2740                  */
2741                 xa_lock(&mapping->i_pages);
2742                 if (xas_load(&xas) != head)
2743                         goto fail;
2744         }
2745
2746         /* Prevent deferred_split_scan() touching ->_refcount */
2747         spin_lock(&pgdata->split_queue_lock);
2748         count = page_count(head);
2749         mapcount = total_mapcount(head);
2750         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2751                 if (!list_empty(page_deferred_list(head))) {
2752                         pgdata->split_queue_len--;
2753                         list_del(page_deferred_list(head));
2754                 }
2755                 if (mapping)
2756                         __dec_node_page_state(page, NR_SHMEM_THPS);
2757                 spin_unlock(&pgdata->split_queue_lock);
2758                 __split_huge_page(page, list, end, flags);
2759                 if (PageSwapCache(head)) {
2760                         swp_entry_t entry = { .val = page_private(head) };
2761
2762                         ret = split_swap_cluster(entry);
2763                 } else
2764                         ret = 0;
2765         } else {
2766                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2767                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2768                                         mapcount, count);
2769                         if (PageTail(page))
2770                                 dump_page(head, NULL);
2771                         dump_page(page, "total_mapcount(head) > 0");
2772                         BUG();
2773                 }
2774                 spin_unlock(&pgdata->split_queue_lock);
2775 fail:           if (mapping)
2776                         xa_unlock(&mapping->i_pages);
2777                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2778                 remap_page(head);
2779                 ret = -EBUSY;
2780         }
2781
2782 out_unlock:
2783         if (anon_vma) {
2784                 anon_vma_unlock_write(anon_vma);
2785                 put_anon_vma(anon_vma);
2786         }
2787         if (mapping)
2788                 i_mmap_unlock_read(mapping);
2789 out:
2790         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2791         return ret;
2792 }
2793
2794 void free_transhuge_page(struct page *page)
2795 {
2796         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2797         unsigned long flags;
2798
2799         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2800         if (!list_empty(page_deferred_list(page))) {
2801                 pgdata->split_queue_len--;
2802                 list_del(page_deferred_list(page));
2803         }
2804         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2805         free_compound_page(page);
2806 }
2807
2808 void deferred_split_huge_page(struct page *page)
2809 {
2810         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2811         unsigned long flags;
2812
2813         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2814
2815         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2816         if (list_empty(page_deferred_list(page))) {
2817                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2818                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2819                 pgdata->split_queue_len++;
2820         }
2821         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2822 }
2823
2824 static unsigned long deferred_split_count(struct shrinker *shrink,
2825                 struct shrink_control *sc)
2826 {
2827         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2828         return READ_ONCE(pgdata->split_queue_len);
2829 }
2830
2831 static unsigned long deferred_split_scan(struct shrinker *shrink,
2832                 struct shrink_control *sc)
2833 {
2834         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2835         unsigned long flags;
2836         LIST_HEAD(list), *pos, *next;
2837         struct page *page;
2838         int split = 0;
2839
2840         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2841         /* Take pin on all head pages to avoid freeing them under us */
2842         list_for_each_safe(pos, next, &pgdata->split_queue) {
2843                 page = list_entry((void *)pos, struct page, mapping);
2844                 page = compound_head(page);
2845                 if (get_page_unless_zero(page)) {
2846                         list_move(page_deferred_list(page), &list);
2847                 } else {
2848                         /* We lost race with put_compound_page() */
2849                         list_del_init(page_deferred_list(page));
2850                         pgdata->split_queue_len--;
2851                 }
2852                 if (!--sc->nr_to_scan)
2853                         break;
2854         }
2855         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2856
2857         list_for_each_safe(pos, next, &list) {
2858                 page = list_entry((void *)pos, struct page, mapping);
2859                 if (!trylock_page(page))
2860                         goto next;
2861                 /* split_huge_page() removes page from list on success */
2862                 if (!split_huge_page(page))
2863                         split++;
2864                 unlock_page(page);
2865 next:
2866                 put_page(page);
2867         }
2868
2869         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2870         list_splice_tail(&list, &pgdata->split_queue);
2871         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2872
2873         /*
2874          * Stop shrinker if we didn't split any page, but the queue is empty.
2875          * This can happen if pages were freed under us.
2876          */
2877         if (!split && list_empty(&pgdata->split_queue))
2878                 return SHRINK_STOP;
2879         return split;
2880 }
2881
2882 static struct shrinker deferred_split_shrinker = {
2883         .count_objects = deferred_split_count,
2884         .scan_objects = deferred_split_scan,
2885         .seeks = DEFAULT_SEEKS,
2886         .flags = SHRINKER_NUMA_AWARE,
2887 };
2888
2889 #ifdef CONFIG_DEBUG_FS
2890 static int split_huge_pages_set(void *data, u64 val)
2891 {
2892         struct zone *zone;
2893         struct page *page;
2894         unsigned long pfn, max_zone_pfn;
2895         unsigned long total = 0, split = 0;
2896
2897         if (val != 1)
2898                 return -EINVAL;
2899
2900         for_each_populated_zone(zone) {
2901                 max_zone_pfn = zone_end_pfn(zone);
2902                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2903                         if (!pfn_valid(pfn))
2904                                 continue;
2905
2906                         page = pfn_to_page(pfn);
2907                         if (!get_page_unless_zero(page))
2908                                 continue;
2909
2910                         if (zone != page_zone(page))
2911                                 goto next;
2912
2913                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2914                                 goto next;
2915
2916                         total++;
2917                         lock_page(page);
2918                         if (!split_huge_page(page))
2919                                 split++;
2920                         unlock_page(page);
2921 next:
2922                         put_page(page);
2923                 }
2924         }
2925
2926         pr_info("%lu of %lu THP split\n", split, total);
2927
2928         return 0;
2929 }
2930 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2931                 "%llu\n");
2932
2933 static int __init split_huge_pages_debugfs(void)
2934 {
2935         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2936                             &split_huge_pages_fops);
2937         return 0;
2938 }
2939 late_initcall(split_huge_pages_debugfs);
2940 #endif
2941
2942 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2943 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2944                 struct page *page)
2945 {
2946         struct vm_area_struct *vma = pvmw->vma;
2947         struct mm_struct *mm = vma->vm_mm;
2948         unsigned long address = pvmw->address;
2949         pmd_t pmdval;
2950         swp_entry_t entry;
2951         pmd_t pmdswp;
2952
2953         if (!(pvmw->pmd && !pvmw->pte))
2954                 return;
2955
2956         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2957         pmdval = *pvmw->pmd;
2958         pmdp_invalidate(vma, address, pvmw->pmd);
2959         if (pmd_dirty(pmdval))
2960                 set_page_dirty(page);
2961         entry = make_migration_entry(page, pmd_write(pmdval));
2962         pmdswp = swp_entry_to_pmd(entry);
2963         if (pmd_soft_dirty(pmdval))
2964                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2965         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2966         page_remove_rmap(page, true);
2967         put_page(page);
2968 }
2969
2970 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2971 {
2972         struct vm_area_struct *vma = pvmw->vma;
2973         struct mm_struct *mm = vma->vm_mm;
2974         unsigned long address = pvmw->address;
2975         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2976         pmd_t pmde;
2977         swp_entry_t entry;
2978
2979         if (!(pvmw->pmd && !pvmw->pte))
2980                 return;
2981
2982         entry = pmd_to_swp_entry(*pvmw->pmd);
2983         get_page(new);
2984         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2985         if (pmd_swp_soft_dirty(*pvmw->pmd))
2986                 pmde = pmd_mksoft_dirty(pmde);
2987         if (is_write_migration_entry(entry))
2988                 pmde = maybe_pmd_mkwrite(pmde, vma);
2989
2990         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2991         if (PageAnon(new))
2992                 page_add_anon_rmap(new, vma, mmun_start, true);
2993         else
2994                 page_add_file_rmap(new, true);
2995         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2996         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2997                 mlock_vma_page(new);
2998         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2999 }
3000 #endif