mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.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
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 /*
  39  * By default transparent hugepage support is disabled in order that avoid
  40  * to risk increase the memory footprint of applications without a guaranteed
  41  * benefit. When transparent hugepage support is enabled, is for all mappings,
  42  * and khugepaged scans all mappings.
  43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  44  * for all hugepage allocations.
  45  */
  46 unsigned long transparent_hugepage_flags __read_mostly =
  47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  49 #endif
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  52 #endif
  53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  56
  57 static struct shrinker deferred_split_shrinker;
  58
  59 static atomic_t huge_zero_refcount;
  60 struct page *huge_zero_page __read_mostly;
  61
  62 static struct page *get_huge_zero_page(void)
  63 {
  64         struct page *zero_page;
  65 retry:
  66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  67                 return READ_ONCE(huge_zero_page);
  68
  69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  70                         HPAGE_PMD_ORDER);
  71         if (!zero_page) {
  72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  73                 return NULL;
  74         }
  75         count_vm_event(THP_ZERO_PAGE_ALLOC);
  76         preempt_disable();
  77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  78                 preempt_enable();
  79                 __free_pages(zero_page, compound_order(zero_page));
  80                 goto retry;
  81         }
  82
  83         /* We take additional reference here. It will be put back by shrinker */
  84         atomic_set(&huge_zero_refcount, 2);
  85         preempt_enable();
  86         return READ_ONCE(huge_zero_page);
  87 }
  88
  89 static void put_huge_zero_page(void)
  90 {
  91         /*
  92          * Counter should never go to zero here. Only shrinker can put
  93          * last reference.
  94          */
  95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  96 }
  97
  98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
  99 {
 100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 101                 return READ_ONCE(huge_zero_page);
 102
 103         if (!get_huge_zero_page())
 104                 return NULL;
 105
 106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 107                 put_huge_zero_page();
 108
 109         return READ_ONCE(huge_zero_page);
 110 }
 111
 112 void mm_put_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 put_huge_zero_page();
 116 }
 117
 118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 119                                         struct shrink_control *sc)
 120 {
 121         /* we can free zero page only if last reference remains */
 122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 123 }
 124
 125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 126                                        struct shrink_control *sc)
 127 {
 128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 130                 BUG_ON(zero_page == NULL);
 131                 __free_pages(zero_page, compound_order(zero_page));
 132                 return HPAGE_PMD_NR;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static struct shrinker huge_zero_page_shrinker = {
 139         .count_objects = shrink_huge_zero_page_count,
 140         .scan_objects = shrink_huge_zero_page_scan,
 141         .seeks = DEFAULT_SEEKS,
 142 };
 143
 144 #ifdef CONFIG_SYSFS
 145 static ssize_t enabled_show(struct kobject *kobj,
 146                             struct kobj_attribute *attr, char *buf)
 147 {
 148         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 149                 return sprintf(buf, "[always] madvise never\n");
 150         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 151                 return sprintf(buf, "always [madvise] never\n");
 152         else
 153                 return sprintf(buf, "always madvise [never]\n");
 154 }
 155
 156 static ssize_t enabled_store(struct kobject *kobj,
 157                              struct kobj_attribute *attr,
 158                              const char *buf, size_t count)
 159 {
 160         ssize_t ret = count;
 161
 162         if (!memcmp("always", buf,
 163                     min(sizeof("always")-1, count))) {
 164                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 165                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 166         } else if (!memcmp("madvise", buf,
 167                            min(sizeof("madvise")-1, count))) {
 168                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 170         } else if (!memcmp("never", buf,
 171                            min(sizeof("never")-1, count))) {
 172                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 173                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 174         } else
 175                 ret = -EINVAL;
 176
 177         if (ret > 0) {
 178                 int err = start_stop_khugepaged();
 179                 if (err)
 180                         ret = err;
 181         }
 182         return ret;
 183 }
 184 static struct kobj_attribute enabled_attr =
 185         __ATTR(enabled, 0644, enabled_show, enabled_store);
 186
 187 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 188                                 struct kobj_attribute *attr, char *buf,
 189                                 enum transparent_hugepage_flag flag)
 190 {
 191         return sprintf(buf, "%d\n",
 192                        !!test_bit(flag, &transparent_hugepage_flags));
 193 }
 194
 195 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 196                                  struct kobj_attribute *attr,
 197                                  const char *buf, size_t count,
 198                                  enum transparent_hugepage_flag flag)
 199 {
 200         unsigned long value;
 201         int ret;
 202
 203         ret = kstrtoul(buf, 10, &value);
 204         if (ret < 0)
 205                 return ret;
 206         if (value > 1)
 207                 return -EINVAL;
 208
 209         if (value)
 210                 set_bit(flag, &transparent_hugepage_flags);
 211         else
 212                 clear_bit(flag, &transparent_hugepage_flags);
 213
 214         return count;
 215 }
 216
 217 static ssize_t defrag_show(struct kobject *kobj,
 218                            struct kobj_attribute *attr, char *buf)
 219 {
 220         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 221                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 222         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 223                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 224         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 225                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 226         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 227                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 228         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 229 }
 230
 231 static ssize_t defrag_store(struct kobject *kobj,
 232                             struct kobj_attribute *attr,
 233                             const char *buf, size_t count)
 234 {
 235         if (!memcmp("always", buf,
 236                     min(sizeof("always")-1, count))) {
 237                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 238                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 239                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 240                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 241         } else if (!memcmp("defer", buf,
 242                     min(sizeof("defer")-1, count))) {
 243                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 245                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 246                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 247         } else if (!memcmp("defer+madvise", buf,
 248                     min(sizeof("defer+madvise")-1, count))) {
 249                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 250                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 252                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253         } else if (!memcmp("madvise", buf,
 254                            min(sizeof("madvise")-1, count))) {
 255                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 258                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259         } else if (!memcmp("never", buf,
 260                            min(sizeof("never")-1, count))) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 265         } else
 266                 return -EINVAL;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute defrag_attr =
 271         __ATTR(defrag, 0644, defrag_show, defrag_store);
 272
 273 static ssize_t use_zero_page_show(struct kobject *kobj,
 274                 struct kobj_attribute *attr, char *buf)
 275 {
 276         return single_hugepage_flag_show(kobj, attr, buf,
 277                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 278 }
 279 static ssize_t use_zero_page_store(struct kobject *kobj,
 280                 struct kobj_attribute *attr, const char *buf, size_t count)
 281 {
 282         return single_hugepage_flag_store(kobj, attr, buf, count,
 283                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 284 }
 285 static struct kobj_attribute use_zero_page_attr =
 286         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 287
 288 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 289                 struct kobj_attribute *attr, char *buf)
 290 {
 291         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 292 }
 293 static struct kobj_attribute hpage_pmd_size_attr =
 294         __ATTR_RO(hpage_pmd_size);
 295
 296 #ifdef CONFIG_DEBUG_VM
 297 static ssize_t debug_cow_show(struct kobject *kobj,
 298                                 struct kobj_attribute *attr, char *buf)
 299 {
 300         return single_hugepage_flag_show(kobj, attr, buf,
 301                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 302 }
 303 static ssize_t debug_cow_store(struct kobject *kobj,
 304                                struct kobj_attribute *attr,
 305                                const char *buf, size_t count)
 306 {
 307         return single_hugepage_flag_store(kobj, attr, buf, count,
 308                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 309 }
 310 static struct kobj_attribute debug_cow_attr =
 311         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 312 #endif /* CONFIG_DEBUG_VM */
 313
 314 static struct attribute *hugepage_attr[] = {
 315         &enabled_attr.attr,
 316         &defrag_attr.attr,
 317         &use_zero_page_attr.attr,
 318         &hpage_pmd_size_attr.attr,
 319 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 320         &shmem_enabled_attr.attr,
 321 #endif
 322 #ifdef CONFIG_DEBUG_VM
 323         &debug_cow_attr.attr,
 324 #endif
 325         NULL,
 326 };
 327
 328 static struct attribute_group hugepage_attr_group = {
 329         .attrs = hugepage_attr,
 330 };
 331
 332 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 333 {
 334         int err;
 335
 336         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 337         if (unlikely(!*hugepage_kobj)) {
 338                 pr_err("failed to create transparent hugepage kobject\n");
 339                 return -ENOMEM;
 340         }
 341
 342         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 343         if (err) {
 344                 pr_err("failed to register transparent hugepage group\n");
 345                 goto delete_obj;
 346         }
 347
 348         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 349         if (err) {
 350                 pr_err("failed to register transparent hugepage group\n");
 351                 goto remove_hp_group;
 352         }
 353
 354         return 0;
 355
 356 remove_hp_group:
 357         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 358 delete_obj:
 359         kobject_put(*hugepage_kobj);
 360         return err;
 361 }
 362
 363 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 364 {
 365         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 366         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 367         kobject_put(hugepage_kobj);
 368 }
 369 #else
 370 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 371 {
 372         return 0;
 373 }
 374
 375 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 376 {
 377 }
 378 #endif /* CONFIG_SYSFS */
 379
 380 static int __init hugepage_init(void)
 381 {
 382         int err;
 383         struct kobject *hugepage_kobj;
 384
 385         if (!has_transparent_hugepage()) {
 386                 transparent_hugepage_flags = 0;
 387                 return -EINVAL;
 388         }
 389
 390         /*
 391          * hugepages can't be allocated by the buddy allocator
 392          */
 393         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 394         /*
 395          * we use page->mapping and page->index in second tail page
 396          * as list_head: assuming THP order >= 2
 397          */
 398         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 399
 400         err = hugepage_init_sysfs(&hugepage_kobj);
 401         if (err)
 402                 goto err_sysfs;
 403
 404         err = khugepaged_init();
 405         if (err)
 406                 goto err_slab;
 407
 408         err = register_shrinker(&huge_zero_page_shrinker);
 409         if (err)
 410                 goto err_hzp_shrinker;
 411         err = register_shrinker(&deferred_split_shrinker);
 412         if (err)
 413                 goto err_split_shrinker;
 414
 415         /*
 416          * By default disable transparent hugepages on smaller systems,
 417          * where the extra memory used could hurt more than TLB overhead
 418          * is likely to save.  The admin can still enable it through /sys.
 419          */
 420         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 421                 transparent_hugepage_flags = 0;
 422                 return 0;
 423         }
 424
 425         err = start_stop_khugepaged();
 426         if (err)
 427                 goto err_khugepaged;
 428
 429         return 0;
 430 err_khugepaged:
 431         unregister_shrinker(&deferred_split_shrinker);
 432 err_split_shrinker:
 433         unregister_shrinker(&huge_zero_page_shrinker);
 434 err_hzp_shrinker:
 435         khugepaged_destroy();
 436 err_slab:
 437         hugepage_exit_sysfs(hugepage_kobj);
 438 err_sysfs:
 439         return err;
 440 }
 441 subsys_initcall(hugepage_init);
 442
 443 static int __init setup_transparent_hugepage(char *str)
 444 {
 445         int ret = 0;
 446         if (!str)
 447                 goto out;
 448         if (!strcmp(str, "always")) {
 449                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 450                         &transparent_hugepage_flags);
 451                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 452                           &transparent_hugepage_flags);
 453                 ret = 1;
 454         } else if (!strcmp(str, "madvise")) {
 455                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 456                           &transparent_hugepage_flags);
 457                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 458                         &transparent_hugepage_flags);
 459                 ret = 1;
 460         } else if (!strcmp(str, "never")) {
 461                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 462                           &transparent_hugepage_flags);
 463                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 464                           &transparent_hugepage_flags);
 465                 ret = 1;
 466         }
 467 out:
 468         if (!ret)
 469                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 470         return ret;
 471 }
 472 __setup("transparent_hugepage=", setup_transparent_hugepage);
 473
 474 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 475 {
 476         if (likely(vma->vm_flags & VM_WRITE))
 477                 pmd = pmd_mkwrite(pmd);
 478         return pmd;
 479 }
 480
 481 static inline struct list_head *page_deferred_list(struct page *page)
 482 {
 483         /*
 484          * ->lru in the tail pages is occupied by compound_head.
 485          * Let's use ->mapping + ->index in the second tail page as list_head.
 486          */
 487         return (struct list_head *)&page[2].mapping;
 488 }
 489
 490 void prep_transhuge_page(struct page *page)
 491 {
 492         /*
 493          * we use page->mapping and page->indexlru in second tail page
 494          * as list_head: assuming THP order >= 2
 495          */
 496
 497         INIT_LIST_HEAD(page_deferred_list(page));
 498         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 499 }
 500
 501 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 502                 loff_t off, unsigned long flags, unsigned long size)
 503 {
 504         unsigned long addr;
 505         loff_t off_end = off + len;
 506         loff_t off_align = round_up(off, size);
 507         unsigned long len_pad;
 508
 509         if (off_end <= off_align || (off_end - off_align) < size)
 510                 return 0;
 511
 512         len_pad = len + size;
 513         if (len_pad < len || (off + len_pad) < off)
 514                 return 0;
 515
 516         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 517                                               off >> PAGE_SHIFT, flags);
 518         if (IS_ERR_VALUE(addr))
 519                 return 0;
 520
 521         addr += (off - addr) & (size - 1);
 522         return addr;
 523 }
 524
 525 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 526                 unsigned long len, unsigned long pgoff, unsigned long flags)
 527 {
 528         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 529
 530         if (addr)
 531                 goto out;
 532         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 533                 goto out;
 534
 535         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 536         if (addr)
 537                 return addr;
 538
 539  out:
 540         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 541 }
 542 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 543
 544 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 545                 gfp_t gfp)
 546 {
 547         struct vm_area_struct *vma = vmf->vma;
 548         struct mem_cgroup *memcg;
 549         pgtable_t pgtable;
 550         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 551
 552         VM_BUG_ON_PAGE(!PageCompound(page), page);
 553
 554         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 555                 put_page(page);
 556                 count_vm_event(THP_FAULT_FALLBACK);
 557                 return VM_FAULT_FALLBACK;
 558         }
 559
 560         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 561         if (unlikely(!pgtable)) {
 562                 mem_cgroup_cancel_charge(page, memcg, true);
 563                 put_page(page);
 564                 return VM_FAULT_OOM;
 565         }
 566
 567         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 568         /*
 569          * The memory barrier inside __SetPageUptodate makes sure that
 570          * clear_huge_page writes become visible before the set_pmd_at()
 571          * write.
 572          */
 573         __SetPageUptodate(page);
 574
 575         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 576         if (unlikely(!pmd_none(*vmf->pmd))) {
 577                 spin_unlock(vmf->ptl);
 578                 mem_cgroup_cancel_charge(page, memcg, true);
 579                 put_page(page);
 580                 pte_free(vma->vm_mm, pgtable);
 581         } else {
 582                 pmd_t entry;
 583
 584                 /* Deliver the page fault to userland */
 585                 if (userfaultfd_missing(vma)) {
 586                         int ret;
 587
 588                         spin_unlock(vmf->ptl);
 589                         mem_cgroup_cancel_charge(page, memcg, true);
 590                         put_page(page);
 591                         pte_free(vma->vm_mm, pgtable);
 592                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 593                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 594                         return ret;
 595                 }
 596
 597                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 598                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 599                 page_add_new_anon_rmap(page, vma, haddr, true);
 600                 mem_cgroup_commit_charge(page, memcg, false, true);
 601                 lru_cache_add_active_or_unevictable(page, vma);
 602                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 603                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 604                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 605                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 606                 spin_unlock(vmf->ptl);
 607                 count_vm_event(THP_FAULT_ALLOC);
 608         }
 609
 610         return 0;
 611 }
 612
 613 /*
 614  * always: directly stall for all thp allocations
 615  * defer: wake kswapd and fail if not immediately available
 616  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 617  *                fail if not immediately available
 618  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 619  *          available
 620  * never: never stall for any thp allocation
 621  */
 622 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 623 {
 624         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 625
 626         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 627                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 628         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 629                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 630         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 631                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 632                                                              __GFP_KSWAPD_RECLAIM);
 633         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 634                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 635                                                              0);
 636         return GFP_TRANSHUGE_LIGHT;
 637 }
 638
 639 /* Caller must hold page table lock. */
 640 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 641                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 642                 struct page *zero_page)
 643 {
 644         pmd_t entry;
 645         if (!pmd_none(*pmd))
 646                 return false;
 647         entry = mk_pmd(zero_page, vma->vm_page_prot);
 648         entry = pmd_mkhuge(entry);
 649         if (pgtable)
 650                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 651         set_pmd_at(mm, haddr, pmd, entry);
 652         atomic_long_inc(&mm->nr_ptes);
 653         return true;
 654 }
 655
 656 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 657 {
 658         struct vm_area_struct *vma = vmf->vma;
 659         gfp_t gfp;
 660         struct page *page;
 661         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 662
 663         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 664                 return VM_FAULT_FALLBACK;
 665         if (unlikely(anon_vma_prepare(vma)))
 666                 return VM_FAULT_OOM;
 667         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 668                 return VM_FAULT_OOM;
 669         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 670                         !mm_forbids_zeropage(vma->vm_mm) &&
 671                         transparent_hugepage_use_zero_page()) {
 672                 pgtable_t pgtable;
 673                 struct page *zero_page;
 674                 bool set;
 675                 int ret;
 676                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 677                 if (unlikely(!pgtable))
 678                         return VM_FAULT_OOM;
 679                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 680                 if (unlikely(!zero_page)) {
 681                         pte_free(vma->vm_mm, pgtable);
 682                         count_vm_event(THP_FAULT_FALLBACK);
 683                         return VM_FAULT_FALLBACK;
 684                 }
 685                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 686                 ret = 0;
 687                 set = false;
 688                 if (pmd_none(*vmf->pmd)) {
 689                         if (userfaultfd_missing(vma)) {
 690                                 spin_unlock(vmf->ptl);
 691                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 692                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 693                         } else {
 694                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 695                                                    haddr, vmf->pmd, zero_page);
 696                                 spin_unlock(vmf->ptl);
 697                                 set = true;
 698                         }
 699                 } else
 700                         spin_unlock(vmf->ptl);
 701                 if (!set)
 702                         pte_free(vma->vm_mm, pgtable);
 703                 return ret;
 704         }
 705         gfp = alloc_hugepage_direct_gfpmask(vma);
 706         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 707         if (unlikely(!page)) {
 708                 count_vm_event(THP_FAULT_FALLBACK);
 709                 return VM_FAULT_FALLBACK;
 710         }
 711         prep_transhuge_page(page);
 712         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 713 }
 714
 715 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 716                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 717 {
 718         struct mm_struct *mm = vma->vm_mm;
 719         pmd_t entry;
 720         spinlock_t *ptl;
 721
 722         ptl = pmd_lock(mm, pmd);
 723         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 724         if (pfn_t_devmap(pfn))
 725                 entry = pmd_mkdevmap(entry);
 726         if (write) {
 727                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 728                 entry = maybe_pmd_mkwrite(entry, vma);
 729         }
 730         set_pmd_at(mm, addr, pmd, entry);
 731         update_mmu_cache_pmd(vma, addr, pmd);
 732         spin_unlock(ptl);
 733 }
 734
 735 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 736                         pmd_t *pmd, pfn_t pfn, bool write)
 737 {
 738         pgprot_t pgprot = vma->vm_page_prot;
 739         /*
 740          * If we had pmd_special, we could avoid all these restrictions,
 741          * but we need to be consistent with PTEs and architectures that
 742          * can't support a 'special' bit.
 743          */
 744         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 745         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 746                                                 (VM_PFNMAP|VM_MIXEDMAP));
 747         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 748         BUG_ON(!pfn_t_devmap(pfn));
 749
 750         if (addr < vma->vm_start || addr >= vma->vm_end)
 751                 return VM_FAULT_SIGBUS;
 752
 753         track_pfn_insert(vma, &pgprot, pfn);
 754
 755         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 756         return VM_FAULT_NOPAGE;
 757 }
 758 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 759
 760 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 761 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 762 {
 763         if (likely(vma->vm_flags & VM_WRITE))
 764                 pud = pud_mkwrite(pud);
 765         return pud;
 766 }
 767
 768 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 769                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 770 {
 771         struct mm_struct *mm = vma->vm_mm;
 772         pud_t entry;
 773         spinlock_t *ptl;
 774
 775         ptl = pud_lock(mm, pud);
 776         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 777         if (pfn_t_devmap(pfn))
 778                 entry = pud_mkdevmap(entry);
 779         if (write) {
 780                 entry = pud_mkyoung(pud_mkdirty(entry));
 781                 entry = maybe_pud_mkwrite(entry, vma);
 782         }
 783         set_pud_at(mm, addr, pud, entry);
 784         update_mmu_cache_pud(vma, addr, pud);
 785         spin_unlock(ptl);
 786 }
 787
 788 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 789                         pud_t *pud, pfn_t pfn, bool write)
 790 {
 791         pgprot_t pgprot = vma->vm_page_prot;
 792         /*
 793          * If we had pud_special, we could avoid all these restrictions,
 794          * but we need to be consistent with PTEs and architectures that
 795          * can't support a 'special' bit.
 796          */
 797         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 798         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 799                                                 (VM_PFNMAP|VM_MIXEDMAP));
 800         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 801         BUG_ON(!pfn_t_devmap(pfn));
 802
 803         if (addr < vma->vm_start || addr >= vma->vm_end)
 804                 return VM_FAULT_SIGBUS;
 805
 806         track_pfn_insert(vma, &pgprot, pfn);
 807
 808         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 809         return VM_FAULT_NOPAGE;
 810 }
 811 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 812 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 813
 814 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 815                 pmd_t *pmd)
 816 {
 817         pmd_t _pmd;
 818
 819         /*
 820          * We should set the dirty bit only for FOLL_WRITE but for now
 821          * the dirty bit in the pmd is meaningless.  And if the dirty
 822          * bit will become meaningful and we'll only set it with
 823          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 824          * set the young bit, instead of the current set_pmd_at.
 825          */
 826         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 827         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 828                                 pmd, _pmd,  1))
 829                 update_mmu_cache_pmd(vma, addr, pmd);
 830 }
 831
 832 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 833                 pmd_t *pmd, int flags)
 834 {
 835         unsigned long pfn = pmd_pfn(*pmd);
 836         struct mm_struct *mm = vma->vm_mm;
 837         struct dev_pagemap *pgmap;
 838         struct page *page;
 839
 840         assert_spin_locked(pmd_lockptr(mm, pmd));
 841
 842         /*
 843          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 844          * not be in this function with `flags & FOLL_COW` set.
 845          */
 846         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 847
 848         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 849                 return NULL;
 850
 851         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 852                 /* pass */;
 853         else
 854                 return NULL;
 855
 856         if (flags & FOLL_TOUCH)
 857                 touch_pmd(vma, addr, pmd);
 858
 859         /*
 860          * device mapped pages can only be returned if the
 861          * caller will manage the page reference count.
 862          */
 863         if (!(flags & FOLL_GET))
 864                 return ERR_PTR(-EEXIST);
 865
 866         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 867         pgmap = get_dev_pagemap(pfn, NULL);
 868         if (!pgmap)
 869                 return ERR_PTR(-EFAULT);
 870         page = pfn_to_page(pfn);
 871         get_page(page);
 872         put_dev_pagemap(pgmap);
 873
 874         return page;
 875 }
 876
 877 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 878                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 879                   struct vm_area_struct *vma)
 880 {
 881         spinlock_t *dst_ptl, *src_ptl;
 882         struct page *src_page;
 883         pmd_t pmd;
 884         pgtable_t pgtable = NULL;
 885         int ret = -ENOMEM;
 886
 887         /* Skip if can be re-fill on fault */
 888         if (!vma_is_anonymous(vma))
 889                 return 0;
 890
 891         pgtable = pte_alloc_one(dst_mm, addr);
 892         if (unlikely(!pgtable))
 893                 goto out;
 894
 895         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 896         src_ptl = pmd_lockptr(src_mm, src_pmd);
 897         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 898
 899         ret = -EAGAIN;
 900         pmd = *src_pmd;
 901         if (unlikely(!pmd_trans_huge(pmd))) {
 902                 pte_free(dst_mm, pgtable);
 903                 goto out_unlock;
 904         }
 905         /*
 906          * When page table lock is held, the huge zero pmd should not be
 907          * under splitting since we don't split the page itself, only pmd to
 908          * a page table.
 909          */
 910         if (is_huge_zero_pmd(pmd)) {
 911                 struct page *zero_page;
 912                 /*
 913                  * get_huge_zero_page() will never allocate a new page here,
 914                  * since we already have a zero page to copy. It just takes a
 915                  * reference.
 916                  */
 917                 zero_page = mm_get_huge_zero_page(dst_mm);
 918                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 919                                 zero_page);
 920                 ret = 0;
 921                 goto out_unlock;
 922         }
 923
 924         src_page = pmd_page(pmd);
 925         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 926         get_page(src_page);
 927         page_dup_rmap(src_page, true);
 928         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 929         atomic_long_inc(&dst_mm->nr_ptes);
 930         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 931
 932         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 933         pmd = pmd_mkold(pmd_wrprotect(pmd));
 934         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 935
 936         ret = 0;
 937 out_unlock:
 938         spin_unlock(src_ptl);
 939         spin_unlock(dst_ptl);
 940 out:
 941         return ret;
 942 }
 943
 944 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 945 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 946                 pud_t *pud)
 947 {
 948         pud_t _pud;
 949
 950         /*
 951          * We should set the dirty bit only for FOLL_WRITE but for now
 952          * the dirty bit in the pud is meaningless.  And if the dirty
 953          * bit will become meaningful and we'll only set it with
 954          * FOLL_WRITE, an atomic set_bit will be required on the pud to
 955          * set the young bit, instead of the current set_pud_at.
 956          */
 957         _pud = pud_mkyoung(pud_mkdirty(*pud));
 958         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
 959                                 pud, _pud,  1))
 960                 update_mmu_cache_pud(vma, addr, pud);
 961 }
 962
 963 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 964                 pud_t *pud, int flags)
 965 {
 966         unsigned long pfn = pud_pfn(*pud);
 967         struct mm_struct *mm = vma->vm_mm;
 968         struct dev_pagemap *pgmap;
 969         struct page *page;
 970
 971         assert_spin_locked(pud_lockptr(mm, pud));
 972
 973         if (flags & FOLL_WRITE && !pud_write(*pud))
 974                 return NULL;
 975
 976         if (pud_present(*pud) && pud_devmap(*pud))
 977                 /* pass */;
 978         else
 979                 return NULL;
 980
 981         if (flags & FOLL_TOUCH)
 982                 touch_pud(vma, addr, pud);
 983
 984         /*
 985          * device mapped pages can only be returned if the
 986          * caller will manage the page reference count.
 987          */
 988         if (!(flags & FOLL_GET))
 989                 return ERR_PTR(-EEXIST);
 990
 991         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
 992         pgmap = get_dev_pagemap(pfn, NULL);
 993         if (!pgmap)
 994                 return ERR_PTR(-EFAULT);
 995         page = pfn_to_page(pfn);
 996         get_page(page);
 997         put_dev_pagemap(pgmap);
 998
 999         return page;
1000 }
1001
1002 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1003                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1004                   struct vm_area_struct *vma)
1005 {
1006         spinlock_t *dst_ptl, *src_ptl;
1007         pud_t pud;
1008         int ret;
1009
1010         dst_ptl = pud_lock(dst_mm, dst_pud);
1011         src_ptl = pud_lockptr(src_mm, src_pud);
1012         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1013
1014         ret = -EAGAIN;
1015         pud = *src_pud;
1016         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1017                 goto out_unlock;
1018
1019         /*
1020          * When page table lock is held, the huge zero pud should not be
1021          * under splitting since we don't split the page itself, only pud to
1022          * a page table.
1023          */
1024         if (is_huge_zero_pud(pud)) {
1025                 /* No huge zero pud yet */
1026         }
1027
1028         pudp_set_wrprotect(src_mm, addr, src_pud);
1029         pud = pud_mkold(pud_wrprotect(pud));
1030         set_pud_at(dst_mm, addr, dst_pud, pud);
1031
1032         ret = 0;
1033 out_unlock:
1034         spin_unlock(src_ptl);
1035         spin_unlock(dst_ptl);
1036         return ret;
1037 }
1038
1039 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1040 {
1041         pud_t entry;
1042         unsigned long haddr;
1043         bool write = vmf->flags & FAULT_FLAG_WRITE;
1044
1045         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1046         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1047                 goto unlock;
1048
1049         entry = pud_mkyoung(orig_pud);
1050         if (write)
1051                 entry = pud_mkdirty(entry);
1052         haddr = vmf->address & HPAGE_PUD_MASK;
1053         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1054                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1055
1056 unlock:
1057         spin_unlock(vmf->ptl);
1058 }
1059 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1060
1061 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1062 {
1063         pmd_t entry;
1064         unsigned long haddr;
1065         bool write = vmf->flags & FAULT_FLAG_WRITE;
1066
1067         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1068         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1069                 goto unlock;
1070
1071         entry = pmd_mkyoung(orig_pmd);
1072         if (write)
1073                 entry = pmd_mkdirty(entry);
1074         haddr = vmf->address & HPAGE_PMD_MASK;
1075         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1076                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1077
1078 unlock:
1079         spin_unlock(vmf->ptl);
1080 }
1081
1082 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1083                 struct page *page)
1084 {
1085         struct vm_area_struct *vma = vmf->vma;
1086         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1087         struct mem_cgroup *memcg;
1088         pgtable_t pgtable;
1089         pmd_t _pmd;
1090         int ret = 0, i;
1091         struct page **pages;
1092         unsigned long mmun_start;       /* For mmu_notifiers */
1093         unsigned long mmun_end;         /* For mmu_notifiers */
1094
1095         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1096                         GFP_KERNEL);
1097         if (unlikely(!pages)) {
1098                 ret |= VM_FAULT_OOM;
1099                 goto out;
1100         }
1101
1102         for (i = 0; i < HPAGE_PMD_NR; i++) {
1103                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1104                                                vmf->address, page_to_nid(page));
1105                 if (unlikely(!pages[i] ||
1106                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1107                                      GFP_KERNEL, &memcg, false))) {
1108                         if (pages[i])
1109                                 put_page(pages[i]);
1110                         while (--i >= 0) {
1111                                 memcg = (void *)page_private(pages[i]);
1112                                 set_page_private(pages[i], 0);
1113                                 mem_cgroup_cancel_charge(pages[i], memcg,
1114                                                 false);
1115                                 put_page(pages[i]);
1116                         }
1117                         kfree(pages);
1118                         ret |= VM_FAULT_OOM;
1119                         goto out;
1120                 }
1121                 set_page_private(pages[i], (unsigned long)memcg);
1122         }
1123
1124         for (i = 0; i < HPAGE_PMD_NR; i++) {
1125                 copy_user_highpage(pages[i], page + i,
1126                                    haddr + PAGE_SIZE * i, vma);
1127                 __SetPageUptodate(pages[i]);
1128                 cond_resched();
1129         }
1130
1131         mmun_start = haddr;
1132         mmun_end   = haddr + HPAGE_PMD_SIZE;
1133         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1134
1135         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1136         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1137                 goto out_free_pages;
1138         VM_BUG_ON_PAGE(!PageHead(page), page);
1139
1140         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1141         /* leave pmd empty until pte is filled */
1142
1143         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1144         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1145
1146         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1147                 pte_t entry;
1148                 entry = mk_pte(pages[i], vma->vm_page_prot);
1149                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1150                 memcg = (void *)page_private(pages[i]);
1151                 set_page_private(pages[i], 0);
1152                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1153                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1154                 lru_cache_add_active_or_unevictable(pages[i], vma);
1155                 vmf->pte = pte_offset_map(&_pmd, haddr);
1156                 VM_BUG_ON(!pte_none(*vmf->pte));
1157                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1158                 pte_unmap(vmf->pte);
1159         }
1160         kfree(pages);
1161
1162         smp_wmb(); /* make pte visible before pmd */
1163         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1164         page_remove_rmap(page, true);
1165         spin_unlock(vmf->ptl);
1166
1167         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1168
1169         ret |= VM_FAULT_WRITE;
1170         put_page(page);
1171
1172 out:
1173         return ret;
1174
1175 out_free_pages:
1176         spin_unlock(vmf->ptl);
1177         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1178         for (i = 0; i < HPAGE_PMD_NR; i++) {
1179                 memcg = (void *)page_private(pages[i]);
1180                 set_page_private(pages[i], 0);
1181                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1182                 put_page(pages[i]);
1183         }
1184         kfree(pages);
1185         goto out;
1186 }
1187
1188 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1189 {
1190         struct vm_area_struct *vma = vmf->vma;
1191         struct page *page = NULL, *new_page;
1192         struct mem_cgroup *memcg;
1193         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1194         unsigned long mmun_start;       /* For mmu_notifiers */
1195         unsigned long mmun_end;         /* For mmu_notifiers */
1196         gfp_t huge_gfp;                 /* for allocation and charge */
1197         int ret = 0;
1198
1199         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1200         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1201         if (is_huge_zero_pmd(orig_pmd))
1202                 goto alloc;
1203         spin_lock(vmf->ptl);
1204         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1205                 goto out_unlock;
1206
1207         page = pmd_page(orig_pmd);
1208         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1209         /*
1210          * We can only reuse the page if nobody else maps the huge page or it's
1211          * part.
1212          */
1213         if (page_trans_huge_mapcount(page, NULL) == 1) {
1214                 pmd_t entry;
1215                 entry = pmd_mkyoung(orig_pmd);
1216                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1217                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1218                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1219                 ret |= VM_FAULT_WRITE;
1220                 goto out_unlock;
1221         }
1222         get_page(page);
1223         spin_unlock(vmf->ptl);
1224 alloc:
1225         if (transparent_hugepage_enabled(vma) &&
1226             !transparent_hugepage_debug_cow()) {
1227                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1228                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1229         } else
1230                 new_page = NULL;
1231
1232         if (likely(new_page)) {
1233                 prep_transhuge_page(new_page);
1234         } else {
1235                 if (!page) {
1236                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1237                         ret |= VM_FAULT_FALLBACK;
1238                 } else {
1239                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1240                         if (ret & VM_FAULT_OOM) {
1241                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1242                                 ret |= VM_FAULT_FALLBACK;
1243                         }
1244                         put_page(page);
1245                 }
1246                 count_vm_event(THP_FAULT_FALLBACK);
1247                 goto out;
1248         }
1249
1250         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1251                                         huge_gfp, &memcg, true))) {
1252                 put_page(new_page);
1253                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1254                 if (page)
1255                         put_page(page);
1256                 ret |= VM_FAULT_FALLBACK;
1257                 count_vm_event(THP_FAULT_FALLBACK);
1258                 goto out;
1259         }
1260
1261         count_vm_event(THP_FAULT_ALLOC);
1262
1263         if (!page)
1264                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1265         else
1266                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1267         __SetPageUptodate(new_page);
1268
1269         mmun_start = haddr;
1270         mmun_end   = haddr + HPAGE_PMD_SIZE;
1271         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1272
1273         spin_lock(vmf->ptl);
1274         if (page)
1275                 put_page(page);
1276         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1277                 spin_unlock(vmf->ptl);
1278                 mem_cgroup_cancel_charge(new_page, memcg, true);
1279                 put_page(new_page);
1280                 goto out_mn;
1281         } else {
1282                 pmd_t entry;
1283                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1284                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1285                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1286                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1287                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1288                 lru_cache_add_active_or_unevictable(new_page, vma);
1289                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1290                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1291                 if (!page) {
1292                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1293                 } else {
1294                         VM_BUG_ON_PAGE(!PageHead(page), page);
1295                         page_remove_rmap(page, true);
1296                         put_page(page);
1297                 }
1298                 ret |= VM_FAULT_WRITE;
1299         }
1300         spin_unlock(vmf->ptl);
1301 out_mn:
1302         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1303 out:
1304         return ret;
1305 out_unlock:
1306         spin_unlock(vmf->ptl);
1307         return ret;
1308 }
1309
1310 /*
1311  * FOLL_FORCE can write to even unwritable pmd's, but only
1312  * after we've gone through a COW cycle and they are dirty.
1313  */
1314 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1315 {
1316         return pmd_write(pmd) ||
1317                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1318 }
1319
1320 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1321                                    unsigned long addr,
1322                                    pmd_t *pmd,
1323                                    unsigned int flags)
1324 {
1325         struct mm_struct *mm = vma->vm_mm;
1326         struct page *page = NULL;
1327
1328         assert_spin_locked(pmd_lockptr(mm, pmd));
1329
1330         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1331                 goto out;
1332
1333         /* Avoid dumping huge zero page */
1334         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1335                 return ERR_PTR(-EFAULT);
1336
1337         /* Full NUMA hinting faults to serialise migration in fault paths */
1338         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1339                 goto out;
1340
1341         page = pmd_page(*pmd);
1342         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1343         if (flags & FOLL_TOUCH)
1344                 touch_pmd(vma, addr, pmd);
1345         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1346                 /*
1347                  * We don't mlock() pte-mapped THPs. This way we can avoid
1348                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1349                  *
1350                  * For anon THP:
1351                  *
1352                  * In most cases the pmd is the only mapping of the page as we
1353                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1354                  * writable private mappings in populate_vma_page_range().
1355                  *
1356                  * The only scenario when we have the page shared here is if we
1357                  * mlocking read-only mapping shared over fork(). We skip
1358                  * mlocking such pages.
1359                  *
1360                  * For file THP:
1361                  *
1362                  * We can expect PageDoubleMap() to be stable under page lock:
1363                  * for file pages we set it in page_add_file_rmap(), which
1364                  * requires page to be locked.
1365                  */
1366
1367                 if (PageAnon(page) && compound_mapcount(page) != 1)
1368                         goto skip_mlock;
1369                 if (PageDoubleMap(page) || !page->mapping)
1370                         goto skip_mlock;
1371                 if (!trylock_page(page))
1372                         goto skip_mlock;
1373                 lru_add_drain();
1374                 if (page->mapping && !PageDoubleMap(page))
1375                         mlock_vma_page(page);
1376                 unlock_page(page);
1377         }
1378 skip_mlock:
1379         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1380         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1381         if (flags & FOLL_GET)
1382                 get_page(page);
1383
1384 out:
1385         return page;
1386 }
1387
1388 /* NUMA hinting page fault entry point for trans huge pmds */
1389 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1390 {
1391         struct vm_area_struct *vma = vmf->vma;
1392         struct anon_vma *anon_vma = NULL;
1393         struct page *page;
1394         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1395         int page_nid = -1, this_nid = numa_node_id();
1396         int target_nid, last_cpupid = -1;
1397         bool page_locked;
1398         bool migrated = false;
1399         bool was_writable;
1400         int flags = 0;
1401
1402         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1403         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1404                 goto out_unlock;
1405
1406         /*
1407          * If there are potential migrations, wait for completion and retry
1408          * without disrupting NUMA hinting information. Do not relock and
1409          * check_same as the page may no longer be mapped.
1410          */
1411         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1412                 page = pmd_page(*vmf->pmd);
1413                 spin_unlock(vmf->ptl);
1414                 wait_on_page_locked(page);
1415                 goto out;
1416         }
1417
1418         page = pmd_page(pmd);
1419         BUG_ON(is_huge_zero_page(page));
1420         page_nid = page_to_nid(page);
1421         last_cpupid = page_cpupid_last(page);
1422         count_vm_numa_event(NUMA_HINT_FAULTS);
1423         if (page_nid == this_nid) {
1424                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1425                 flags |= TNF_FAULT_LOCAL;
1426         }
1427
1428         /* See similar comment in do_numa_page for explanation */
1429         if (!pmd_savedwrite(pmd))
1430                 flags |= TNF_NO_GROUP;
1431
1432         /*
1433          * Acquire the page lock to serialise THP migrations but avoid dropping
1434          * page_table_lock if at all possible
1435          */
1436         page_locked = trylock_page(page);
1437         target_nid = mpol_misplaced(page, vma, haddr);
1438         if (target_nid == -1) {
1439                 /* If the page was locked, there are no parallel migrations */
1440                 if (page_locked)
1441                         goto clear_pmdnuma;
1442         }
1443
1444         /* Migration could have started since the pmd_trans_migrating check */
1445         if (!page_locked) {
1446                 spin_unlock(vmf->ptl);
1447                 wait_on_page_locked(page);
1448                 page_nid = -1;
1449                 goto out;
1450         }
1451
1452         /*
1453          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1454          * to serialises splits
1455          */
1456         get_page(page);
1457         spin_unlock(vmf->ptl);
1458         anon_vma = page_lock_anon_vma_read(page);
1459
1460         /* Confirm the PMD did not change while page_table_lock was released */
1461         spin_lock(vmf->ptl);
1462         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1463                 unlock_page(page);
1464                 put_page(page);
1465                 page_nid = -1;
1466                 goto out_unlock;
1467         }
1468
1469         /* Bail if we fail to protect against THP splits for any reason */
1470         if (unlikely(!anon_vma)) {
1471                 put_page(page);
1472                 page_nid = -1;
1473                 goto clear_pmdnuma;
1474         }
1475
1476         /*
1477          * Migrate the THP to the requested node, returns with page unlocked
1478          * and access rights restored.
1479          */
1480         spin_unlock(vmf->ptl);
1481         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1482                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1483         if (migrated) {
1484                 flags |= TNF_MIGRATED;
1485                 page_nid = target_nid;
1486         } else
1487                 flags |= TNF_MIGRATE_FAIL;
1488
1489         goto out;
1490 clear_pmdnuma:
1491         BUG_ON(!PageLocked(page));
1492         was_writable = pmd_savedwrite(pmd);
1493         pmd = pmd_modify(pmd, vma->vm_page_prot);
1494         pmd = pmd_mkyoung(pmd);
1495         if (was_writable)
1496                 pmd = pmd_mkwrite(pmd);
1497         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1498         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1499         unlock_page(page);
1500 out_unlock:
1501         spin_unlock(vmf->ptl);
1502
1503 out:
1504         if (anon_vma)
1505                 page_unlock_anon_vma_read(anon_vma);
1506
1507         if (page_nid != -1)
1508                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1509                                 flags);
1510
1511         return 0;
1512 }
1513
1514 /*
1515  * Return true if we do MADV_FREE successfully on entire pmd page.
1516  * Otherwise, return false.
1517  */
1518 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1519                 pmd_t *pmd, unsigned long addr, unsigned long next)
1520 {
1521         spinlock_t *ptl;
1522         pmd_t orig_pmd;
1523         struct page *page;
1524         struct mm_struct *mm = tlb->mm;
1525         bool ret = false;
1526
1527         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1528
1529         ptl = pmd_trans_huge_lock(pmd, vma);
1530         if (!ptl)
1531                 goto out_unlocked;
1532
1533         orig_pmd = *pmd;
1534         if (is_huge_zero_pmd(orig_pmd))
1535                 goto out;
1536
1537         page = pmd_page(orig_pmd);
1538         /*
1539          * If other processes are mapping this page, we couldn't discard
1540          * the page unless they all do MADV_FREE so let's skip the page.
1541          */
1542         if (page_mapcount(page) != 1)
1543                 goto out;
1544
1545         if (!trylock_page(page))
1546                 goto out;
1547
1548         /*
1549          * If user want to discard part-pages of THP, split it so MADV_FREE
1550          * will deactivate only them.
1551          */
1552         if (next - addr != HPAGE_PMD_SIZE) {
1553                 get_page(page);
1554                 spin_unlock(ptl);
1555                 split_huge_page(page);
1556                 put_page(page);
1557                 unlock_page(page);
1558                 goto out_unlocked;
1559         }
1560
1561         if (PageDirty(page))
1562                 ClearPageDirty(page);
1563         unlock_page(page);
1564
1565         if (PageActive(page))
1566                 deactivate_page(page);
1567
1568         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1569                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1570                         tlb->fullmm);
1571                 orig_pmd = pmd_mkold(orig_pmd);
1572                 orig_pmd = pmd_mkclean(orig_pmd);
1573
1574                 set_pmd_at(mm, addr, pmd, orig_pmd);
1575                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1576         }
1577         ret = true;
1578 out:
1579         spin_unlock(ptl);
1580 out_unlocked:
1581         return ret;
1582 }
1583
1584 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1585 {
1586         pgtable_t pgtable;
1587
1588         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1589         pte_free(mm, pgtable);
1590         atomic_long_dec(&mm->nr_ptes);
1591 }
1592
1593 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1594                  pmd_t *pmd, unsigned long addr)
1595 {
1596         pmd_t orig_pmd;
1597         spinlock_t *ptl;
1598
1599         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1600
1601         ptl = __pmd_trans_huge_lock(pmd, vma);
1602         if (!ptl)
1603                 return 0;
1604         /*
1605          * For architectures like ppc64 we look at deposited pgtable
1606          * when calling pmdp_huge_get_and_clear. So do the
1607          * pgtable_trans_huge_withdraw after finishing pmdp related
1608          * operations.
1609          */
1610         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1611                         tlb->fullmm);
1612         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1613         if (vma_is_dax(vma)) {
1614                 spin_unlock(ptl);
1615                 if (is_huge_zero_pmd(orig_pmd))
1616                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1617         } else if (is_huge_zero_pmd(orig_pmd)) {
1618                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1619                 atomic_long_dec(&tlb->mm->nr_ptes);
1620                 spin_unlock(ptl);
1621                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1622         } else {
1623                 struct page *page = pmd_page(orig_pmd);
1624                 page_remove_rmap(page, true);
1625                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1626                 VM_BUG_ON_PAGE(!PageHead(page), page);
1627                 if (PageAnon(page)) {
1628                         pgtable_t pgtable;
1629                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1630                         pte_free(tlb->mm, pgtable);
1631                         atomic_long_dec(&tlb->mm->nr_ptes);
1632                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1633                 } else {
1634                         if (arch_needs_pgtable_deposit())
1635                                 zap_deposited_table(tlb->mm, pmd);
1636                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1637                 }
1638                 spin_unlock(ptl);
1639                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1640         }
1641         return 1;
1642 }
1643
1644 #ifndef pmd_move_must_withdraw
1645 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1646                                          spinlock_t *old_pmd_ptl,
1647                                          struct vm_area_struct *vma)
1648 {
1649         /*
1650          * With split pmd lock we also need to move preallocated
1651          * PTE page table if new_pmd is on different PMD page table.
1652          *
1653          * We also don't deposit and withdraw tables for file pages.
1654          */
1655         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1656 }
1657 #endif
1658
1659 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1660                   unsigned long new_addr, unsigned long old_end,
1661                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1662 {
1663         spinlock_t *old_ptl, *new_ptl;
1664         pmd_t pmd;
1665         struct mm_struct *mm = vma->vm_mm;
1666         bool force_flush = false;
1667
1668         if ((old_addr & ~HPAGE_PMD_MASK) ||
1669             (new_addr & ~HPAGE_PMD_MASK) ||
1670             old_end - old_addr < HPAGE_PMD_SIZE)
1671                 return false;
1672
1673         /*
1674          * The destination pmd shouldn't be established, free_pgtables()
1675          * should have release it.
1676          */
1677         if (WARN_ON(!pmd_none(*new_pmd))) {
1678                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1679                 return false;
1680         }
1681
1682         /*
1683          * We don't have to worry about the ordering of src and dst
1684          * ptlocks because exclusive mmap_sem prevents deadlock.
1685          */
1686         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1687         if (old_ptl) {
1688                 new_ptl = pmd_lockptr(mm, new_pmd);
1689                 if (new_ptl != old_ptl)
1690                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1691                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1692                 if (pmd_present(pmd) && pmd_dirty(pmd))
1693                         force_flush = true;
1694                 VM_BUG_ON(!pmd_none(*new_pmd));
1695
1696                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1697                         pgtable_t pgtable;
1698                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1699                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1700                 }
1701                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1702                 if (new_ptl != old_ptl)
1703                         spin_unlock(new_ptl);
1704                 if (force_flush)
1705                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1706                 else
1707                         *need_flush = true;
1708                 spin_unlock(old_ptl);
1709                 return true;
1710         }
1711         return false;
1712 }
1713
1714 /*
1715  * Returns
1716  *  - 0 if PMD could not be locked
1717  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1718  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1719  */
1720 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1721                 unsigned long addr, pgprot_t newprot, int prot_numa)
1722 {
1723         struct mm_struct *mm = vma->vm_mm;
1724         spinlock_t *ptl;
1725         int ret = 0;
1726
1727         ptl = __pmd_trans_huge_lock(pmd, vma);
1728         if (ptl) {
1729                 pmd_t entry;
1730                 bool preserve_write = prot_numa && pmd_write(*pmd);
1731                 ret = 1;
1732
1733                 /*
1734                  * Avoid trapping faults against the zero page. The read-only
1735                  * data is likely to be read-cached on the local CPU and
1736                  * local/remote hits to the zero page are not interesting.
1737                  */
1738                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1739                         spin_unlock(ptl);
1740                         return ret;
1741                 }
1742
1743                 if (!prot_numa || !pmd_protnone(*pmd)) {
1744                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1745                         entry = pmd_modify(entry, newprot);
1746                         if (preserve_write)
1747                                 entry = pmd_mk_savedwrite(entry);
1748                         ret = HPAGE_PMD_NR;
1749                         set_pmd_at(mm, addr, pmd, entry);
1750                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1751                                         pmd_write(entry));
1752                 }
1753                 spin_unlock(ptl);
1754         }
1755
1756         return ret;
1757 }
1758
1759 /*
1760  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1761  *
1762  * Note that if it returns page table lock pointer, this routine returns without
1763  * unlocking page table lock. So callers must unlock it.
1764  */
1765 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1766 {
1767         spinlock_t *ptl;
1768         ptl = pmd_lock(vma->vm_mm, pmd);
1769         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1770                 return ptl;
1771         spin_unlock(ptl);
1772         return NULL;
1773 }
1774
1775 /*
1776  * Returns true if a given pud maps a thp, false otherwise.
1777  *
1778  * Note that if it returns true, this routine returns without unlocking page
1779  * table lock. So callers must unlock it.
1780  */
1781 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1782 {
1783         spinlock_t *ptl;
1784
1785         ptl = pud_lock(vma->vm_mm, pud);
1786         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1787                 return ptl;
1788         spin_unlock(ptl);
1789         return NULL;
1790 }
1791
1792 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1793 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1794                  pud_t *pud, unsigned long addr)
1795 {
1796         pud_t orig_pud;
1797         spinlock_t *ptl;
1798
1799         ptl = __pud_trans_huge_lock(pud, vma);
1800         if (!ptl)
1801                 return 0;
1802         /*
1803          * For architectures like ppc64 we look at deposited pgtable
1804          * when calling pudp_huge_get_and_clear. So do the
1805          * pgtable_trans_huge_withdraw after finishing pudp related
1806          * operations.
1807          */
1808         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1809                         tlb->fullmm);
1810         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1811         if (vma_is_dax(vma)) {
1812                 spin_unlock(ptl);
1813                 /* No zero page support yet */
1814         } else {
1815                 /* No support for anonymous PUD pages yet */
1816                 BUG();
1817         }
1818         return 1;
1819 }
1820
1821 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1822                 unsigned long haddr)
1823 {
1824         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1825         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1826         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1827         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1828
1829         count_vm_event(THP_SPLIT_PMD);
1830
1831         pudp_huge_clear_flush_notify(vma, haddr, pud);
1832 }
1833
1834 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1835                 unsigned long address)
1836 {
1837         spinlock_t *ptl;
1838         struct mm_struct *mm = vma->vm_mm;
1839         unsigned long haddr = address & HPAGE_PUD_MASK;
1840
1841         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1842         ptl = pud_lock(mm, pud);
1843         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1844                 goto out;
1845         __split_huge_pud_locked(vma, pud, haddr);
1846
1847 out:
1848         spin_unlock(ptl);
1849         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1850 }
1851 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1852
1853 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1854                 unsigned long haddr, pmd_t *pmd)
1855 {
1856         struct mm_struct *mm = vma->vm_mm;
1857         pgtable_t pgtable;
1858         pmd_t _pmd;
1859         int i;
1860
1861         /* leave pmd empty until pte is filled */
1862         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1863
1864         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1865         pmd_populate(mm, &_pmd, pgtable);
1866
1867         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1868                 pte_t *pte, entry;
1869                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1870                 entry = pte_mkspecial(entry);
1871                 pte = pte_offset_map(&_pmd, haddr);
1872                 VM_BUG_ON(!pte_none(*pte));
1873                 set_pte_at(mm, haddr, pte, entry);
1874                 pte_unmap(pte);
1875         }
1876         smp_wmb(); /* make pte visible before pmd */
1877         pmd_populate(mm, pmd, pgtable);
1878 }
1879
1880 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1881                 unsigned long haddr, bool freeze)
1882 {
1883         struct mm_struct *mm = vma->vm_mm;
1884         struct page *page;
1885         pgtable_t pgtable;
1886         pmd_t _pmd;
1887         bool young, write, dirty, soft_dirty;
1888         unsigned long addr;
1889         int i;
1890
1891         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1892         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1893         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1894         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1895
1896         count_vm_event(THP_SPLIT_PMD);
1897
1898         if (!vma_is_anonymous(vma)) {
1899                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1900                 /*
1901                  * We are going to unmap this huge page. So
1902                  * just go ahead and zap it
1903                  */
1904                 if (arch_needs_pgtable_deposit())
1905                         zap_deposited_table(mm, pmd);
1906                 if (vma_is_dax(vma))
1907                         return;
1908                 page = pmd_page(_pmd);
1909                 if (!PageReferenced(page) && pmd_young(_pmd))
1910                         SetPageReferenced(page);
1911                 page_remove_rmap(page, true);
1912                 put_page(page);
1913                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1914                 return;
1915         } else if (is_huge_zero_pmd(*pmd)) {
1916                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1917         }
1918
1919         page = pmd_page(*pmd);
1920         VM_BUG_ON_PAGE(!page_count(page), page);
1921         page_ref_add(page, HPAGE_PMD_NR - 1);
1922         write = pmd_write(*pmd);
1923         young = pmd_young(*pmd);
1924         dirty = pmd_dirty(*pmd);
1925         soft_dirty = pmd_soft_dirty(*pmd);
1926
1927         pmdp_huge_split_prepare(vma, haddr, pmd);
1928         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1929         pmd_populate(mm, &_pmd, pgtable);
1930
1931         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1932                 pte_t entry, *pte;
1933                 /*
1934                  * Note that NUMA hinting access restrictions are not
1935                  * transferred to avoid any possibility of altering
1936                  * permissions across VMAs.
1937                  */
1938                 if (freeze) {
1939                         swp_entry_t swp_entry;
1940                         swp_entry = make_migration_entry(page + i, write);
1941                         entry = swp_entry_to_pte(swp_entry);
1942                         if (soft_dirty)
1943                                 entry = pte_swp_mksoft_dirty(entry);
1944                 } else {
1945                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1946                         entry = maybe_mkwrite(entry, vma);
1947                         if (!write)
1948                                 entry = pte_wrprotect(entry);
1949                         if (!young)
1950                                 entry = pte_mkold(entry);
1951                         if (soft_dirty)
1952                                 entry = pte_mksoft_dirty(entry);
1953                 }
1954                 if (dirty)
1955                         SetPageDirty(page + i);
1956                 pte = pte_offset_map(&_pmd, addr);
1957                 BUG_ON(!pte_none(*pte));
1958                 set_pte_at(mm, addr, pte, entry);
1959                 atomic_inc(&page[i]._mapcount);
1960                 pte_unmap(pte);
1961         }
1962
1963         /*
1964          * Set PG_double_map before dropping compound_mapcount to avoid
1965          * false-negative page_mapped().
1966          */
1967         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1968                 for (i = 0; i < HPAGE_PMD_NR; i++)
1969                         atomic_inc(&page[i]._mapcount);
1970         }
1971
1972         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1973                 /* Last compound_mapcount is gone. */
1974                 __dec_node_page_state(page, NR_ANON_THPS);
1975                 if (TestClearPageDoubleMap(page)) {
1976                         /* No need in mapcount reference anymore */
1977                         for (i = 0; i < HPAGE_PMD_NR; i++)
1978                                 atomic_dec(&page[i]._mapcount);
1979                 }
1980         }
1981
1982         smp_wmb(); /* make pte visible before pmd */
1983         /*
1984          * Up to this point the pmd is present and huge and userland has the
1985          * whole access to the hugepage during the split (which happens in
1986          * place). If we overwrite the pmd with the not-huge version pointing
1987          * to the pte here (which of course we could if all CPUs were bug
1988          * free), userland could trigger a small page size TLB miss on the
1989          * small sized TLB while the hugepage TLB entry is still established in
1990          * the huge TLB. Some CPU doesn't like that.
1991          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1992          * 383 on page 93. Intel should be safe but is also warns that it's
1993          * only safe if the permission and cache attributes of the two entries
1994          * loaded in the two TLB is identical (which should be the case here).
1995          * But it is generally safer to never allow small and huge TLB entries
1996          * for the same virtual address to be loaded simultaneously. So instead
1997          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1998          * current pmd notpresent (atomically because here the pmd_trans_huge
1999          * and pmd_trans_splitting must remain set at all times on the pmd
2000          * until the split is complete for this pmd), then we flush the SMP TLB
2001          * and finally we write the non-huge version of the pmd entry with
2002          * pmd_populate.
2003          */
2004         pmdp_invalidate(vma, haddr, pmd);
2005         pmd_populate(mm, pmd, pgtable);
2006
2007         if (freeze) {
2008                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2009                         page_remove_rmap(page + i, false);
2010                         put_page(page + i);
2011                 }
2012         }
2013 }
2014
2015 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2016                 unsigned long address, bool freeze, struct page *page)
2017 {
2018         spinlock_t *ptl;
2019         struct mm_struct *mm = vma->vm_mm;
2020         unsigned long haddr = address & HPAGE_PMD_MASK;
2021
2022         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2023         ptl = pmd_lock(mm, pmd);
2024
2025         /*
2026          * If caller asks to setup a migration entries, we need a page to check
2027          * pmd against. Otherwise we can end up replacing wrong page.
2028          */
2029         VM_BUG_ON(freeze && !page);
2030         if (page && page != pmd_page(*pmd))
2031                 goto out;
2032
2033         if (pmd_trans_huge(*pmd)) {
2034                 page = pmd_page(*pmd);
2035                 if (PageMlocked(page))
2036                         clear_page_mlock(page);
2037         } else if (!pmd_devmap(*pmd))
2038                 goto out;
2039         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2040 out:
2041         spin_unlock(ptl);
2042         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2043 }
2044
2045 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2046                 bool freeze, struct page *page)
2047 {
2048         pgd_t *pgd;
2049         pud_t *pud;
2050         pmd_t *pmd;
2051
2052         pgd = pgd_offset(vma->vm_mm, address);
2053         if (!pgd_present(*pgd))
2054                 return;
2055
2056         pud = pud_offset(pgd, address);
2057         if (!pud_present(*pud))
2058                 return;
2059
2060         pmd = pmd_offset(pud, address);
2061
2062         __split_huge_pmd(vma, pmd, address, freeze, page);
2063 }
2064
2065 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2066                              unsigned long start,
2067                              unsigned long end,
2068                              long adjust_next)
2069 {
2070         /*
2071          * If the new start address isn't hpage aligned and it could
2072          * previously contain an hugepage: check if we need to split
2073          * an huge pmd.
2074          */
2075         if (start & ~HPAGE_PMD_MASK &&
2076             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2077             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2078                 split_huge_pmd_address(vma, start, false, NULL);
2079
2080         /*
2081          * If the new end address isn't hpage aligned and it could
2082          * previously contain an hugepage: check if we need to split
2083          * an huge pmd.
2084          */
2085         if (end & ~HPAGE_PMD_MASK &&
2086             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2087             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2088                 split_huge_pmd_address(vma, end, false, NULL);
2089
2090         /*
2091          * If we're also updating the vma->vm_next->vm_start, if the new
2092          * vm_next->vm_start isn't page aligned and it could previously
2093          * contain an hugepage: check if we need to split an huge pmd.
2094          */
2095         if (adjust_next > 0) {
2096                 struct vm_area_struct *next = vma->vm_next;
2097                 unsigned long nstart = next->vm_start;
2098                 nstart += adjust_next << PAGE_SHIFT;
2099                 if (nstart & ~HPAGE_PMD_MASK &&
2100                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2101                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2102                         split_huge_pmd_address(next, nstart, false, NULL);
2103         }
2104 }
2105
2106 static void freeze_page(struct page *page)
2107 {
2108         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2109                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2110         int ret;
2111
2112         VM_BUG_ON_PAGE(!PageHead(page), page);
2113
2114         if (PageAnon(page))
2115                 ttu_flags |= TTU_MIGRATION;
2116
2117         ret = try_to_unmap(page, ttu_flags);
2118         VM_BUG_ON_PAGE(ret, page);
2119 }
2120
2121 static void unfreeze_page(struct page *page)
2122 {
2123         int i;
2124         if (PageTransHuge(page)) {
2125                 remove_migration_ptes(page, page, true);
2126         } else {
2127                 for (i = 0; i < HPAGE_PMD_NR; i++)
2128                         remove_migration_ptes(page + i, page + i, true);
2129         }
2130 }
2131
2132 static void __split_huge_page_tail(struct page *head, int tail,
2133                 struct lruvec *lruvec, struct list_head *list)
2134 {
2135         struct page *page_tail = head + tail;
2136
2137         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2138         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2139
2140         /*
2141          * tail_page->_refcount is zero and not changing from under us. But
2142          * get_page_unless_zero() may be running from under us on the
2143          * tail_page. If we used atomic_set() below instead of atomic_inc() or
2144          * atomic_add(), we would then run atomic_set() concurrently with
2145          * get_page_unless_zero(), and atomic_set() is implemented in C not
2146          * using locked ops. spin_unlock on x86 sometime uses locked ops
2147          * because of PPro errata 66, 92, so unless somebody can guarantee
2148          * atomic_set() here would be safe on all archs (and not only on x86),
2149          * it's safer to use atomic_inc()/atomic_add().
2150          */
2151         if (PageAnon(head)) {
2152                 page_ref_inc(page_tail);
2153         } else {
2154                 /* Additional pin to radix tree */
2155                 page_ref_add(page_tail, 2);
2156         }
2157
2158         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2159         page_tail->flags |= (head->flags &
2160                         ((1L << PG_referenced) |
2161                          (1L << PG_swapbacked) |
2162                          (1L << PG_mlocked) |
2163                          (1L << PG_uptodate) |
2164                          (1L << PG_active) |
2165                          (1L << PG_locked) |
2166                          (1L << PG_unevictable) |
2167                          (1L << PG_dirty)));
2168
2169         /*
2170          * After clearing PageTail the gup refcount can be released.
2171          * Page flags also must be visible before we make the page non-compound.
2172          */
2173         smp_wmb();
2174
2175         clear_compound_head(page_tail);
2176
2177         if (page_is_young(head))
2178                 set_page_young(page_tail);
2179         if (page_is_idle(head))
2180                 set_page_idle(page_tail);
2181
2182         /* ->mapping in first tail page is compound_mapcount */
2183         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2184                         page_tail);
2185         page_tail->mapping = head->mapping;
2186
2187         page_tail->index = head->index + tail;
2188         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2189         lru_add_page_tail(head, page_tail, lruvec, list);
2190 }
2191
2192 static void __split_huge_page(struct page *page, struct list_head *list,
2193                 unsigned long flags)
2194 {
2195         struct page *head = compound_head(page);
2196         struct zone *zone = page_zone(head);
2197         struct lruvec *lruvec;
2198         pgoff_t end = -1;
2199         int i;
2200
2201         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2202
2203         /* complete memcg works before add pages to LRU */
2204         mem_cgroup_split_huge_fixup(head);
2205
2206         if (!PageAnon(page))
2207                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2208
2209         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2210                 __split_huge_page_tail(head, i, lruvec, list);
2211                 /* Some pages can be beyond i_size: drop them from page cache */
2212                 if (head[i].index >= end) {
2213                         __ClearPageDirty(head + i);
2214                         __delete_from_page_cache(head + i, NULL);
2215                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2216                                 shmem_uncharge(head->mapping->host, 1);
2217                         put_page(head + i);
2218                 }
2219         }
2220
2221         ClearPageCompound(head);
2222         /* See comment in __split_huge_page_tail() */
2223         if (PageAnon(head)) {
2224                 page_ref_inc(head);
2225         } else {
2226                 /* Additional pin to radix tree */
2227                 page_ref_add(head, 2);
2228                 spin_unlock(&head->mapping->tree_lock);
2229         }
2230
2231         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2232
2233         unfreeze_page(head);
2234
2235         for (i = 0; i < HPAGE_PMD_NR; i++) {
2236                 struct page *subpage = head + i;
2237                 if (subpage == page)
2238                         continue;
2239                 unlock_page(subpage);
2240
2241                 /*
2242                  * Subpages may be freed if there wasn't any mapping
2243                  * like if add_to_swap() is running on a lru page that
2244                  * had its mapping zapped. And freeing these pages
2245                  * requires taking the lru_lock so we do the put_page
2246                  * of the tail pages after the split is complete.
2247                  */
2248                 put_page(subpage);
2249         }
2250 }
2251
2252 int total_mapcount(struct page *page)
2253 {
2254         int i, compound, ret;
2255
2256         VM_BUG_ON_PAGE(PageTail(page), page);
2257
2258         if (likely(!PageCompound(page)))
2259                 return atomic_read(&page->_mapcount) + 1;
2260
2261         compound = compound_mapcount(page);
2262         if (PageHuge(page))
2263                 return compound;
2264         ret = compound;
2265         for (i = 0; i < HPAGE_PMD_NR; i++)
2266                 ret += atomic_read(&page[i]._mapcount) + 1;
2267         /* File pages has compound_mapcount included in _mapcount */
2268         if (!PageAnon(page))
2269                 return ret - compound * HPAGE_PMD_NR;
2270         if (PageDoubleMap(page))
2271                 ret -= HPAGE_PMD_NR;
2272         return ret;
2273 }
2274
2275 /*
2276  * This calculates accurately how many mappings a transparent hugepage
2277  * has (unlike page_mapcount() which isn't fully accurate). This full
2278  * accuracy is primarily needed to know if copy-on-write faults can
2279  * reuse the page and change the mapping to read-write instead of
2280  * copying them. At the same time this returns the total_mapcount too.
2281  *
2282  * The function returns the highest mapcount any one of the subpages
2283  * has. If the return value is one, even if different processes are
2284  * mapping different subpages of the transparent hugepage, they can
2285  * all reuse it, because each process is reusing a different subpage.
2286  *
2287  * The total_mapcount is instead counting all virtual mappings of the
2288  * subpages. If the total_mapcount is equal to "one", it tells the
2289  * caller all mappings belong to the same "mm" and in turn the
2290  * anon_vma of the transparent hugepage can become the vma->anon_vma
2291  * local one as no other process may be mapping any of the subpages.
2292  *
2293  * It would be more accurate to replace page_mapcount() with
2294  * page_trans_huge_mapcount(), however we only use
2295  * page_trans_huge_mapcount() in the copy-on-write faults where we
2296  * need full accuracy to avoid breaking page pinning, because
2297  * page_trans_huge_mapcount() is slower than page_mapcount().
2298  */
2299 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2300 {
2301         int i, ret, _total_mapcount, mapcount;
2302
2303         /* hugetlbfs shouldn't call it */
2304         VM_BUG_ON_PAGE(PageHuge(page), page);
2305
2306         if (likely(!PageTransCompound(page))) {
2307                 mapcount = atomic_read(&page->_mapcount) + 1;
2308                 if (total_mapcount)
2309                         *total_mapcount = mapcount;
2310                 return mapcount;
2311         }
2312
2313         page = compound_head(page);
2314
2315         _total_mapcount = ret = 0;
2316         for (i = 0; i < HPAGE_PMD_NR; i++) {
2317                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2318                 ret = max(ret, mapcount);
2319                 _total_mapcount += mapcount;
2320         }
2321         if (PageDoubleMap(page)) {
2322                 ret -= 1;
2323                 _total_mapcount -= HPAGE_PMD_NR;
2324         }
2325         mapcount = compound_mapcount(page);
2326         ret += mapcount;
2327         _total_mapcount += mapcount;
2328         if (total_mapcount)
2329                 *total_mapcount = _total_mapcount;
2330         return ret;
2331 }
2332
2333 /*
2334  * This function splits huge page into normal pages. @page can point to any
2335  * subpage of huge page to split. Split doesn't change the position of @page.
2336  *
2337  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2338  * The huge page must be locked.
2339  *
2340  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2341  *
2342  * Both head page and tail pages will inherit mapping, flags, and so on from
2343  * the hugepage.
2344  *
2345  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2346  * they are not mapped.
2347  *
2348  * Returns 0 if the hugepage is split successfully.
2349  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2350  * us.
2351  */
2352 int split_huge_page_to_list(struct page *page, struct list_head *list)
2353 {
2354         struct page *head = compound_head(page);
2355         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2356         struct anon_vma *anon_vma = NULL;
2357         struct address_space *mapping = NULL;
2358         int count, mapcount, extra_pins, ret;
2359         bool mlocked;
2360         unsigned long flags;
2361
2362         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2363         VM_BUG_ON_PAGE(!PageLocked(page), page);
2364         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2365         VM_BUG_ON_PAGE(!PageCompound(page), page);
2366
2367         if (PageAnon(head)) {
2368                 /*
2369                  * The caller does not necessarily hold an mmap_sem that would
2370                  * prevent the anon_vma disappearing so we first we take a
2371                  * reference to it and then lock the anon_vma for write. This
2372                  * is similar to page_lock_anon_vma_read except the write lock
2373                  * is taken to serialise against parallel split or collapse
2374                  * operations.
2375                  */
2376                 anon_vma = page_get_anon_vma(head);
2377                 if (!anon_vma) {
2378                         ret = -EBUSY;
2379                         goto out;
2380                 }
2381                 extra_pins = 0;
2382                 mapping = NULL;
2383                 anon_vma_lock_write(anon_vma);
2384         } else {
2385                 mapping = head->mapping;
2386
2387                 /* Truncated ? */
2388                 if (!mapping) {
2389                         ret = -EBUSY;
2390                         goto out;
2391                 }
2392
2393                 /* Addidional pins from radix tree */
2394                 extra_pins = HPAGE_PMD_NR;
2395                 anon_vma = NULL;
2396                 i_mmap_lock_read(mapping);
2397         }
2398
2399         /*
2400          * Racy check if we can split the page, before freeze_page() will
2401          * split PMDs
2402          */
2403         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2404                 ret = -EBUSY;
2405                 goto out_unlock;
2406         }
2407
2408         mlocked = PageMlocked(page);
2409         freeze_page(head);
2410         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2411
2412         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2413         if (mlocked)
2414                 lru_add_drain();
2415
2416         /* prevent PageLRU to go away from under us, and freeze lru stats */
2417         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2418
2419         if (mapping) {
2420                 void **pslot;
2421
2422                 spin_lock(&mapping->tree_lock);
2423                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2424                                 page_index(head));
2425                 /*
2426                  * Check if the head page is present in radix tree.
2427                  * We assume all tail are present too, if head is there.
2428                  */
2429                 if (radix_tree_deref_slot_protected(pslot,
2430                                         &mapping->tree_lock) != head)
2431                         goto fail;
2432         }
2433
2434         /* Prevent deferred_split_scan() touching ->_refcount */
2435         spin_lock(&pgdata->split_queue_lock);
2436         count = page_count(head);
2437         mapcount = total_mapcount(head);
2438         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2439                 if (!list_empty(page_deferred_list(head))) {
2440                         pgdata->split_queue_len--;
2441                         list_del(page_deferred_list(head));
2442                 }
2443                 if (mapping)
2444                         __dec_node_page_state(page, NR_SHMEM_THPS);
2445                 spin_unlock(&pgdata->split_queue_lock);
2446                 __split_huge_page(page, list, flags);
2447                 ret = 0;
2448         } else {
2449                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2450                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2451                                         mapcount, count);
2452                         if (PageTail(page))
2453                                 dump_page(head, NULL);
2454                         dump_page(page, "total_mapcount(head) > 0");
2455                         BUG();
2456                 }
2457                 spin_unlock(&pgdata->split_queue_lock);
2458 fail:           if (mapping)
2459                         spin_unlock(&mapping->tree_lock);
2460                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2461                 unfreeze_page(head);
2462                 ret = -EBUSY;
2463         }
2464
2465 out_unlock:
2466         if (anon_vma) {
2467                 anon_vma_unlock_write(anon_vma);
2468                 put_anon_vma(anon_vma);
2469         }
2470         if (mapping)
2471                 i_mmap_unlock_read(mapping);
2472 out:
2473         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2474         return ret;
2475 }
2476
2477 void free_transhuge_page(struct page *page)
2478 {
2479         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2480         unsigned long flags;
2481
2482         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2483         if (!list_empty(page_deferred_list(page))) {
2484                 pgdata->split_queue_len--;
2485                 list_del(page_deferred_list(page));
2486         }
2487         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2488         free_compound_page(page);
2489 }
2490
2491 void deferred_split_huge_page(struct page *page)
2492 {
2493         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2494         unsigned long flags;
2495
2496         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2497
2498         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2499         if (list_empty(page_deferred_list(page))) {
2500                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2501                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2502                 pgdata->split_queue_len++;
2503         }
2504         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2505 }
2506
2507 static unsigned long deferred_split_count(struct shrinker *shrink,
2508                 struct shrink_control *sc)
2509 {
2510         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2511         return ACCESS_ONCE(pgdata->split_queue_len);
2512 }
2513
2514 static unsigned long deferred_split_scan(struct shrinker *shrink,
2515                 struct shrink_control *sc)
2516 {
2517         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2518         unsigned long flags;
2519         LIST_HEAD(list), *pos, *next;
2520         struct page *page;
2521         int split = 0;
2522
2523         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2524         /* Take pin on all head pages to avoid freeing them under us */
2525         list_for_each_safe(pos, next, &pgdata->split_queue) {
2526                 page = list_entry((void *)pos, struct page, mapping);
2527                 page = compound_head(page);
2528                 if (get_page_unless_zero(page)) {
2529                         list_move(page_deferred_list(page), &list);
2530                 } else {
2531                         /* We lost race with put_compound_page() */
2532                         list_del_init(page_deferred_list(page));
2533                         pgdata->split_queue_len--;
2534                 }
2535                 if (!--sc->nr_to_scan)
2536                         break;
2537         }
2538         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2539
2540         list_for_each_safe(pos, next, &list) {
2541                 page = list_entry((void *)pos, struct page, mapping);
2542                 lock_page(page);
2543                 /* split_huge_page() removes page from list on success */
2544                 if (!split_huge_page(page))
2545                         split++;
2546                 unlock_page(page);
2547                 put_page(page);
2548         }
2549
2550         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2551         list_splice_tail(&list, &pgdata->split_queue);
2552         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2553
2554         /*
2555          * Stop shrinker if we didn't split any page, but the queue is empty.
2556          * This can happen if pages were freed under us.
2557          */
2558         if (!split && list_empty(&pgdata->split_queue))
2559                 return SHRINK_STOP;
2560         return split;
2561 }
2562
2563 static struct shrinker deferred_split_shrinker = {
2564         .count_objects = deferred_split_count,
2565         .scan_objects = deferred_split_scan,
2566         .seeks = DEFAULT_SEEKS,
2567         .flags = SHRINKER_NUMA_AWARE,
2568 };
2569
2570 #ifdef CONFIG_DEBUG_FS
2571 static int split_huge_pages_set(void *data, u64 val)
2572 {
2573         struct zone *zone;
2574         struct page *page;
2575         unsigned long pfn, max_zone_pfn;
2576         unsigned long total = 0, split = 0;
2577
2578         if (val != 1)
2579                 return -EINVAL;
2580
2581         for_each_populated_zone(zone) {
2582                 max_zone_pfn = zone_end_pfn(zone);
2583                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2584                         if (!pfn_valid(pfn))
2585                                 continue;
2586
2587                         page = pfn_to_page(pfn);
2588                         if (!get_page_unless_zero(page))
2589                                 continue;
2590
2591                         if (zone != page_zone(page))
2592                                 goto next;
2593
2594                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2595                                 goto next;
2596
2597                         total++;
2598                         lock_page(page);
2599                         if (!split_huge_page(page))
2600                                 split++;
2601                         unlock_page(page);
2602 next:
2603                         put_page(page);
2604                 }
2605         }
2606
2607         pr_info("%lu of %lu THP split\n", split, total);
2608
2609         return 0;
2610 }
2611 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2612                 "%llu\n");
2613
2614 static int __init split_huge_pages_debugfs(void)
2615 {
2616         void *ret;
2617
2618         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2619                         &split_huge_pages_fops);
2620         if (!ret)
2621                 pr_warn("Failed to create split_huge_pages in debugfs");
2622         return 0;
2623 }
2624 late_initcall(split_huge_pages_debugfs);
2625 #endif