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/sched/coredump.h>
  13 #include <linux/sched/numa_balancing.h>
  14 #include <linux/highmem.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <linux/rmap.h>
  18 #include <linux/swap.h>
  19 #include <linux/shrinker.h>
  20 #include <linux/mm_inline.h>
  21 #include <linux/swapops.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default transparent hugepage support is disabled in order that avoid
  43  * to risk increase the memory footprint of applications without a guaranteed
  44  * benefit. When transparent hugepage support is enabled, is for all mappings,
  45  * and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 static struct page *get_huge_zero_page(void)
  66 {
  67         struct page *zero_page;
  68 retry:
  69         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  70                 return READ_ONCE(huge_zero_page);
  71
  72         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  73                         HPAGE_PMD_ORDER);
  74         if (!zero_page) {
  75                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  76                 return NULL;
  77         }
  78         count_vm_event(THP_ZERO_PAGE_ALLOC);
  79         preempt_disable();
  80         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  81                 preempt_enable();
  82                 __free_pages(zero_page, compound_order(zero_page));
  83                 goto retry;
  84         }
  85
  86         /* We take additional reference here. It will be put back by shrinker */
  87         atomic_set(&huge_zero_refcount, 2);
  88         preempt_enable();
  89         return READ_ONCE(huge_zero_page);
  90 }
  91
  92 static void put_huge_zero_page(void)
  93 {
  94         /*
  95          * Counter should never go to zero here. Only shrinker can put
  96          * last reference.
  97          */
  98         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  99 }
 100
 101 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 102 {
 103         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 104                 return READ_ONCE(huge_zero_page);
 105
 106         if (!get_huge_zero_page())
 107                 return NULL;
 108
 109         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 110                 put_huge_zero_page();
 111
 112         return READ_ONCE(huge_zero_page);
 113 }
 114
 115 void mm_put_huge_zero_page(struct mm_struct *mm)
 116 {
 117         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 118                 put_huge_zero_page();
 119 }
 120
 121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 122                                         struct shrink_control *sc)
 123 {
 124         /* we can free zero page only if last reference remains */
 125         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 126 }
 127
 128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 129                                        struct shrink_control *sc)
 130 {
 131         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 132                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 133                 BUG_ON(zero_page == NULL);
 134                 __free_pages(zero_page, compound_order(zero_page));
 135                 return HPAGE_PMD_NR;
 136         }
 137
 138         return 0;
 139 }
 140
 141 static struct shrinker huge_zero_page_shrinker = {
 142         .count_objects = shrink_huge_zero_page_count,
 143         .scan_objects = shrink_huge_zero_page_scan,
 144         .seeks = DEFAULT_SEEKS,
 145 };
 146
 147 #ifdef CONFIG_SYSFS
 148 static ssize_t enabled_show(struct kobject *kobj,
 149                             struct kobj_attribute *attr, char *buf)
 150 {
 151         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 152                 return sprintf(buf, "[always] madvise never\n");
 153         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 154                 return sprintf(buf, "always [madvise] never\n");
 155         else
 156                 return sprintf(buf, "always madvise [never]\n");
 157 }
 158
 159 static ssize_t enabled_store(struct kobject *kobj,
 160                              struct kobj_attribute *attr,
 161                              const char *buf, size_t count)
 162 {
 163         ssize_t ret = count;
 164
 165         if (!memcmp("always", buf,
 166                     min(sizeof("always")-1, count))) {
 167                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 168                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169         } else if (!memcmp("madvise", buf,
 170                            min(sizeof("madvise")-1, count))) {
 171                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 172                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 173         } else if (!memcmp("never", buf,
 174                            min(sizeof("never")-1, count))) {
 175                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 176                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 177         } else
 178                 ret = -EINVAL;
 179
 180         if (ret > 0) {
 181                 int err = start_stop_khugepaged();
 182                 if (err)
 183                         ret = err;
 184         }
 185         return ret;
 186 }
 187 static struct kobj_attribute enabled_attr =
 188         __ATTR(enabled, 0644, enabled_show, enabled_store);
 189
 190 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 191                                 struct kobj_attribute *attr, char *buf,
 192                                 enum transparent_hugepage_flag flag)
 193 {
 194         return sprintf(buf, "%d\n",
 195                        !!test_bit(flag, &transparent_hugepage_flags));
 196 }
 197
 198 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 199                                  struct kobj_attribute *attr,
 200                                  const char *buf, size_t count,
 201                                  enum transparent_hugepage_flag flag)
 202 {
 203         unsigned long value;
 204         int ret;
 205
 206         ret = kstrtoul(buf, 10, &value);
 207         if (ret < 0)
 208                 return ret;
 209         if (value > 1)
 210                 return -EINVAL;
 211
 212         if (value)
 213                 set_bit(flag, &transparent_hugepage_flags);
 214         else
 215                 clear_bit(flag, &transparent_hugepage_flags);
 216
 217         return count;
 218 }
 219
 220 static ssize_t defrag_show(struct kobject *kobj,
 221                            struct kobj_attribute *attr, char *buf)
 222 {
 223         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 224                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 225         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 226                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 227         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 228                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 229         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 230                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 231         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 232 }
 233
 234 static ssize_t defrag_store(struct kobject *kobj,
 235                             struct kobj_attribute *attr,
 236                             const char *buf, size_t count)
 237 {
 238         if (!memcmp("always", buf,
 239                     min(sizeof("always")-1, count))) {
 240                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 241                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 242                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 243                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244         } else if (!memcmp("defer+madvise", buf,
 245                     min(sizeof("defer+madvise")-1, count))) {
 246                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 248                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 249                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 250         } else if (!memcmp("defer", buf,
 251                     min(sizeof("defer")-1, count))) {
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 255                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256         } else if (!memcmp("madvise", buf,
 257                            min(sizeof("madvise")-1, count))) {
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 261                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262         } else if (!memcmp("never", buf,
 263                            min(sizeof("never")-1, count))) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273 static struct kobj_attribute defrag_attr =
 274         __ATTR(defrag, 0644, defrag_show, defrag_store);
 275
 276 static ssize_t use_zero_page_show(struct kobject *kobj,
 277                 struct kobj_attribute *attr, char *buf)
 278 {
 279         return single_hugepage_flag_show(kobj, attr, buf,
 280                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 281 }
 282 static ssize_t use_zero_page_store(struct kobject *kobj,
 283                 struct kobj_attribute *attr, const char *buf, size_t count)
 284 {
 285         return single_hugepage_flag_store(kobj, attr, buf, count,
 286                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 287 }
 288 static struct kobj_attribute use_zero_page_attr =
 289         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 290
 291 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 292                 struct kobj_attribute *attr, char *buf)
 293 {
 294         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 295 }
 296 static struct kobj_attribute hpage_pmd_size_attr =
 297         __ATTR_RO(hpage_pmd_size);
 298
 299 #ifdef CONFIG_DEBUG_VM
 300 static ssize_t debug_cow_show(struct kobject *kobj,
 301                                 struct kobj_attribute *attr, char *buf)
 302 {
 303         return single_hugepage_flag_show(kobj, attr, buf,
 304                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 305 }
 306 static ssize_t debug_cow_store(struct kobject *kobj,
 307                                struct kobj_attribute *attr,
 308                                const char *buf, size_t count)
 309 {
 310         return single_hugepage_flag_store(kobj, attr, buf, count,
 311                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static struct kobj_attribute debug_cow_attr =
 314         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 315 #endif /* CONFIG_DEBUG_VM */
 316
 317 static struct attribute *hugepage_attr[] = {
 318         &enabled_attr.attr,
 319         &defrag_attr.attr,
 320         &use_zero_page_attr.attr,
 321         &hpage_pmd_size_attr.attr,
 322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 323         &shmem_enabled_attr.attr,
 324 #endif
 325 #ifdef CONFIG_DEBUG_VM
 326         &debug_cow_attr.attr,
 327 #endif
 328         NULL,
 329 };
 330
 331 static const struct attribute_group hugepage_attr_group = {
 332         .attrs = hugepage_attr,
 333 };
 334
 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 336 {
 337         int err;
 338
 339         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 340         if (unlikely(!*hugepage_kobj)) {
 341                 pr_err("failed to create transparent hugepage kobject\n");
 342                 return -ENOMEM;
 343         }
 344
 345         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 346         if (err) {
 347                 pr_err("failed to register transparent hugepage group\n");
 348                 goto delete_obj;
 349         }
 350
 351         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 352         if (err) {
 353                 pr_err("failed to register transparent hugepage group\n");
 354                 goto remove_hp_group;
 355         }
 356
 357         return 0;
 358
 359 remove_hp_group:
 360         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 361 delete_obj:
 362         kobject_put(*hugepage_kobj);
 363         return err;
 364 }
 365
 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 369         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 370         kobject_put(hugepage_kobj);
 371 }
 372 #else
 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 374 {
 375         return 0;
 376 }
 377
 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 379 {
 380 }
 381 #endif /* CONFIG_SYSFS */
 382
 383 static int __init hugepage_init(void)
 384 {
 385         int err;
 386         struct kobject *hugepage_kobj;
 387
 388         if (!has_transparent_hugepage()) {
 389                 transparent_hugepage_flags = 0;
 390                 return -EINVAL;
 391         }
 392
 393         /*
 394          * hugepages can't be allocated by the buddy allocator
 395          */
 396         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 397         /*
 398          * we use page->mapping and page->index in second tail page
 399          * as list_head: assuming THP order >= 2
 400          */
 401         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 402
 403         err = hugepage_init_sysfs(&hugepage_kobj);
 404         if (err)
 405                 goto err_sysfs;
 406
 407         err = khugepaged_init();
 408         if (err)
 409                 goto err_slab;
 410
 411         err = register_shrinker(&huge_zero_page_shrinker);
 412         if (err)
 413                 goto err_hzp_shrinker;
 414         err = register_shrinker(&deferred_split_shrinker);
 415         if (err)
 416                 goto err_split_shrinker;
 417
 418         /*
 419          * By default disable transparent hugepages on smaller systems,
 420          * where the extra memory used could hurt more than TLB overhead
 421          * is likely to save.  The admin can still enable it through /sys.
 422          */
 423         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 424                 transparent_hugepage_flags = 0;
 425                 return 0;
 426         }
 427
 428         err = start_stop_khugepaged();
 429         if (err)
 430                 goto err_khugepaged;
 431
 432         return 0;
 433 err_khugepaged:
 434         unregister_shrinker(&deferred_split_shrinker);
 435 err_split_shrinker:
 436         unregister_shrinker(&huge_zero_page_shrinker);
 437 err_hzp_shrinker:
 438         khugepaged_destroy();
 439 err_slab:
 440         hugepage_exit_sysfs(hugepage_kobj);
 441 err_sysfs:
 442         return err;
 443 }
 444 subsys_initcall(hugepage_init);
 445
 446 static int __init setup_transparent_hugepage(char *str)
 447 {
 448         int ret = 0;
 449         if (!str)
 450                 goto out;
 451         if (!strcmp(str, "always")) {
 452                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                         &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         } else if (!strcmp(str, "madvise")) {
 458                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 459                           &transparent_hugepage_flags);
 460                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 461                         &transparent_hugepage_flags);
 462                 ret = 1;
 463         } else if (!strcmp(str, "never")) {
 464                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 465                           &transparent_hugepage_flags);
 466                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 467                           &transparent_hugepage_flags);
 468                 ret = 1;
 469         }
 470 out:
 471         if (!ret)
 472                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 473         return ret;
 474 }
 475 __setup("transparent_hugepage=", setup_transparent_hugepage);
 476
 477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 478 {
 479         if (likely(vma->vm_flags & VM_WRITE))
 480                 pmd = pmd_mkwrite(pmd);
 481         return pmd;
 482 }
 483
 484 static inline struct list_head *page_deferred_list(struct page *page)
 485 {
 486         /*
 487          * ->lru in the tail pages is occupied by compound_head.
 488          * Let's use ->mapping + ->index in the second tail page as list_head.
 489          */
 490         return (struct list_head *)&page[2].mapping;
 491 }
 492
 493 void prep_transhuge_page(struct page *page)
 494 {
 495         /*
 496          * we use page->mapping and page->indexlru in second tail page
 497          * as list_head: assuming THP order >= 2
 498          */
 499
 500         INIT_LIST_HEAD(page_deferred_list(page));
 501         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 502 }
 503
 504 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 505                 loff_t off, unsigned long flags, unsigned long size)
 506 {
 507         unsigned long addr;
 508         loff_t off_end = off + len;
 509         loff_t off_align = round_up(off, size);
 510         unsigned long len_pad;
 511
 512         if (off_end <= off_align || (off_end - off_align) < size)
 513                 return 0;
 514
 515         len_pad = len + size;
 516         if (len_pad < len || (off + len_pad) < off)
 517                 return 0;
 518
 519         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 520                                               off >> PAGE_SHIFT, flags);
 521         if (IS_ERR_VALUE(addr))
 522                 return 0;
 523
 524         addr += (off - addr) & (size - 1);
 525         return addr;
 526 }
 527
 528 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 529                 unsigned long len, unsigned long pgoff, unsigned long flags)
 530 {
 531         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 532
 533         if (addr)
 534                 goto out;
 535         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 536                 goto out;
 537
 538         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 539         if (addr)
 540                 return addr;
 541
 542  out:
 543         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 544 }
 545 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 546
 547 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 548                 gfp_t gfp)
 549 {
 550         struct vm_area_struct *vma = vmf->vma;
 551         struct mem_cgroup *memcg;
 552         pgtable_t pgtable;
 553         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 554         int ret = 0;
 555
 556         VM_BUG_ON_PAGE(!PageCompound(page), page);
 557
 558         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 559                 put_page(page);
 560                 count_vm_event(THP_FAULT_FALLBACK);
 561                 return VM_FAULT_FALLBACK;
 562         }
 563
 564         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 565         if (unlikely(!pgtable)) {
 566                 ret = VM_FAULT_OOM;
 567                 goto release;
 568         }
 569
 570         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 571         /*
 572          * The memory barrier inside __SetPageUptodate makes sure that
 573          * clear_huge_page writes become visible before the set_pmd_at()
 574          * write.
 575          */
 576         __SetPageUptodate(page);
 577
 578         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 579         if (unlikely(!pmd_none(*vmf->pmd))) {
 580                 goto unlock_release;
 581         } else {
 582                 pmd_t entry;
 583
 584                 ret = check_stable_address_space(vma->vm_mm);
 585                 if (ret)
 586                         goto unlock_release;
 587
 588                 /* Deliver the page fault to userland */
 589                 if (userfaultfd_missing(vma)) {
 590                         int ret;
 591
 592                         spin_unlock(vmf->ptl);
 593                         mem_cgroup_cancel_charge(page, memcg, true);
 594                         put_page(page);
 595                         pte_free(vma->vm_mm, pgtable);
 596                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 597                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 598                         return ret;
 599                 }
 600
 601                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 602                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 603                 page_add_new_anon_rmap(page, vma, haddr, true);
 604                 mem_cgroup_commit_charge(page, memcg, false, true);
 605                 lru_cache_add_active_or_unevictable(page, vma);
 606                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 607                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 608                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 609                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 610                 spin_unlock(vmf->ptl);
 611                 count_vm_event(THP_FAULT_ALLOC);
 612         }
 613
 614         return 0;
 615 unlock_release:
 616         spin_unlock(vmf->ptl);
 617 release:
 618         if (pgtable)
 619                 pte_free(vma->vm_mm, pgtable);
 620         mem_cgroup_cancel_charge(page, memcg, true);
 621         put_page(page);
 622         return ret;
 623
 624 }
 625
 626 /*
 627  * always: directly stall for all thp allocations
 628  * defer: wake kswapd and fail if not immediately available
 629  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 630  *                fail if not immediately available
 631  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 632  *          available
 633  * never: never stall for any thp allocation
 634  */
 635 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 636 {
 637         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 638
 639         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 640                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 641         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 642                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 643         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 644                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 645                                                              __GFP_KSWAPD_RECLAIM);
 646         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 647                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 648                                                              0);
 649         return GFP_TRANSHUGE_LIGHT;
 650 }
 651
 652 /* Caller must hold page table lock. */
 653 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 654                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 655                 struct page *zero_page)
 656 {
 657         pmd_t entry;
 658         if (!pmd_none(*pmd))
 659                 return false;
 660         entry = mk_pmd(zero_page, vma->vm_page_prot);
 661         entry = pmd_mkhuge(entry);
 662         if (pgtable)
 663                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 664         set_pmd_at(mm, haddr, pmd, entry);
 665         atomic_long_inc(&mm->nr_ptes);
 666         return true;
 667 }
 668
 669 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 670 {
 671         struct vm_area_struct *vma = vmf->vma;
 672         gfp_t gfp;
 673         struct page *page;
 674         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 675
 676         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 677                 return VM_FAULT_FALLBACK;
 678         if (unlikely(anon_vma_prepare(vma)))
 679                 return VM_FAULT_OOM;
 680         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 681                 return VM_FAULT_OOM;
 682         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 683                         !mm_forbids_zeropage(vma->vm_mm) &&
 684                         transparent_hugepage_use_zero_page()) {
 685                 pgtable_t pgtable;
 686                 struct page *zero_page;
 687                 bool set;
 688                 int ret;
 689                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 690                 if (unlikely(!pgtable))
 691                         return VM_FAULT_OOM;
 692                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 693                 if (unlikely(!zero_page)) {
 694                         pte_free(vma->vm_mm, pgtable);
 695                         count_vm_event(THP_FAULT_FALLBACK);
 696                         return VM_FAULT_FALLBACK;
 697                 }
 698                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 699                 ret = 0;
 700                 set = false;
 701                 if (pmd_none(*vmf->pmd)) {
 702                         ret = check_stable_address_space(vma->vm_mm);
 703                         if (ret) {
 704                                 spin_unlock(vmf->ptl);
 705                         } else if (userfaultfd_missing(vma)) {
 706                                 spin_unlock(vmf->ptl);
 707                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 708                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 709                         } else {
 710                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 711                                                    haddr, vmf->pmd, zero_page);
 712                                 spin_unlock(vmf->ptl);
 713                                 set = true;
 714                         }
 715                 } else
 716                         spin_unlock(vmf->ptl);
 717                 if (!set)
 718                         pte_free(vma->vm_mm, pgtable);
 719                 return ret;
 720         }
 721         gfp = alloc_hugepage_direct_gfpmask(vma);
 722         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 723         if (unlikely(!page)) {
 724                 count_vm_event(THP_FAULT_FALLBACK);
 725                 return VM_FAULT_FALLBACK;
 726         }
 727         prep_transhuge_page(page);
 728         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 729 }
 730
 731 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 732                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 733                 pgtable_t pgtable)
 734 {
 735         struct mm_struct *mm = vma->vm_mm;
 736         pmd_t entry;
 737         spinlock_t *ptl;
 738
 739         ptl = pmd_lock(mm, pmd);
 740         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 741         if (pfn_t_devmap(pfn))
 742                 entry = pmd_mkdevmap(entry);
 743         if (write) {
 744                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 745                 entry = maybe_pmd_mkwrite(entry, vma);
 746         }
 747
 748         if (pgtable) {
 749                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 750                 atomic_long_inc(&mm->nr_ptes);
 751         }
 752
 753         set_pmd_at(mm, addr, pmd, entry);
 754         update_mmu_cache_pmd(vma, addr, pmd);
 755         spin_unlock(ptl);
 756 }
 757
 758 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 759                         pmd_t *pmd, pfn_t pfn, bool write)
 760 {
 761         pgprot_t pgprot = vma->vm_page_prot;
 762         pgtable_t pgtable = NULL;
 763         /*
 764          * If we had pmd_special, we could avoid all these restrictions,
 765          * but we need to be consistent with PTEs and architectures that
 766          * can't support a 'special' bit.
 767          */
 768         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 769         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 770                                                 (VM_PFNMAP|VM_MIXEDMAP));
 771         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 772         BUG_ON(!pfn_t_devmap(pfn));
 773
 774         if (addr < vma->vm_start || addr >= vma->vm_end)
 775                 return VM_FAULT_SIGBUS;
 776
 777         if (arch_needs_pgtable_deposit()) {
 778                 pgtable = pte_alloc_one(vma->vm_mm, addr);
 779                 if (!pgtable)
 780                         return VM_FAULT_OOM;
 781         }
 782
 783         track_pfn_insert(vma, &pgprot, pfn);
 784
 785         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 786         return VM_FAULT_NOPAGE;
 787 }
 788 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 789
 790 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 791 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 792 {
 793         if (likely(vma->vm_flags & VM_WRITE))
 794                 pud = pud_mkwrite(pud);
 795         return pud;
 796 }
 797
 798 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 799                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 800 {
 801         struct mm_struct *mm = vma->vm_mm;
 802         pud_t entry;
 803         spinlock_t *ptl;
 804
 805         ptl = pud_lock(mm, pud);
 806         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 807         if (pfn_t_devmap(pfn))
 808                 entry = pud_mkdevmap(entry);
 809         if (write) {
 810                 entry = pud_mkyoung(pud_mkdirty(entry));
 811                 entry = maybe_pud_mkwrite(entry, vma);
 812         }
 813         set_pud_at(mm, addr, pud, entry);
 814         update_mmu_cache_pud(vma, addr, pud);
 815         spin_unlock(ptl);
 816 }
 817
 818 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 819                         pud_t *pud, pfn_t pfn, bool write)
 820 {
 821         pgprot_t pgprot = vma->vm_page_prot;
 822         /*
 823          * If we had pud_special, we could avoid all these restrictions,
 824          * but we need to be consistent with PTEs and architectures that
 825          * can't support a 'special' bit.
 826          */
 827         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 828         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 829                                                 (VM_PFNMAP|VM_MIXEDMAP));
 830         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 831         BUG_ON(!pfn_t_devmap(pfn));
 832
 833         if (addr < vma->vm_start || addr >= vma->vm_end)
 834                 return VM_FAULT_SIGBUS;
 835
 836         track_pfn_insert(vma, &pgprot, pfn);
 837
 838         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 839         return VM_FAULT_NOPAGE;
 840 }
 841 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 842 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 843
 844 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 845                 pmd_t *pmd)
 846 {
 847         pmd_t _pmd;
 848
 849         /*
 850          * We should set the dirty bit only for FOLL_WRITE but for now
 851          * the dirty bit in the pmd is meaningless.  And if the dirty
 852          * bit will become meaningful and we'll only set it with
 853          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 854          * set the young bit, instead of the current set_pmd_at.
 855          */
 856         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 857         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 858                                 pmd, _pmd,  1))
 859                 update_mmu_cache_pmd(vma, addr, pmd);
 860 }
 861
 862 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 863                 pmd_t *pmd, int flags)
 864 {
 865         unsigned long pfn = pmd_pfn(*pmd);
 866         struct mm_struct *mm = vma->vm_mm;
 867         struct dev_pagemap *pgmap;
 868         struct page *page;
 869
 870         assert_spin_locked(pmd_lockptr(mm, pmd));
 871
 872         /*
 873          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 874          * not be in this function with `flags & FOLL_COW` set.
 875          */
 876         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 877
 878         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 879                 return NULL;
 880
 881         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 882                 /* pass */;
 883         else
 884                 return NULL;
 885
 886         if (flags & FOLL_TOUCH)
 887                 touch_pmd(vma, addr, pmd);
 888
 889         /*
 890          * device mapped pages can only be returned if the
 891          * caller will manage the page reference count.
 892          */
 893         if (!(flags & FOLL_GET))
 894                 return ERR_PTR(-EEXIST);
 895
 896         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 897         pgmap = get_dev_pagemap(pfn, NULL);
 898         if (!pgmap)
 899                 return ERR_PTR(-EFAULT);
 900         page = pfn_to_page(pfn);
 901         get_page(page);
 902         put_dev_pagemap(pgmap);
 903
 904         return page;
 905 }
 906
 907 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 908                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 909                   struct vm_area_struct *vma)
 910 {
 911         spinlock_t *dst_ptl, *src_ptl;
 912         struct page *src_page;
 913         pmd_t pmd;
 914         pgtable_t pgtable = NULL;
 915         int ret = -ENOMEM;
 916
 917         /* Skip if can be re-fill on fault */
 918         if (!vma_is_anonymous(vma))
 919                 return 0;
 920
 921         pgtable = pte_alloc_one(dst_mm, addr);
 922         if (unlikely(!pgtable))
 923                 goto out;
 924
 925         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 926         src_ptl = pmd_lockptr(src_mm, src_pmd);
 927         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 928
 929         ret = -EAGAIN;
 930         pmd = *src_pmd;
 931         if (unlikely(!pmd_trans_huge(pmd))) {
 932                 pte_free(dst_mm, pgtable);
 933                 goto out_unlock;
 934         }
 935         /*
 936          * When page table lock is held, the huge zero pmd should not be
 937          * under splitting since we don't split the page itself, only pmd to
 938          * a page table.
 939          */
 940         if (is_huge_zero_pmd(pmd)) {
 941                 struct page *zero_page;
 942                 /*
 943                  * get_huge_zero_page() will never allocate a new page here,
 944                  * since we already have a zero page to copy. It just takes a
 945                  * reference.
 946                  */
 947                 zero_page = mm_get_huge_zero_page(dst_mm);
 948                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 949                                 zero_page);
 950                 ret = 0;
 951                 goto out_unlock;
 952         }
 953
 954         src_page = pmd_page(pmd);
 955         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 956         get_page(src_page);
 957         page_dup_rmap(src_page, true);
 958         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 959         atomic_long_inc(&dst_mm->nr_ptes);
 960         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 961
 962         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 963         pmd = pmd_mkold(pmd_wrprotect(pmd));
 964         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 965
 966         ret = 0;
 967 out_unlock:
 968         spin_unlock(src_ptl);
 969         spin_unlock(dst_ptl);
 970 out:
 971         return ret;
 972 }
 973
 974 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 975 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 976                 pud_t *pud)
 977 {
 978         pud_t _pud;
 979
 980         /*
 981          * We should set the dirty bit only for FOLL_WRITE but for now
 982          * the dirty bit in the pud is meaningless.  And if the dirty
 983          * bit will become meaningful and we'll only set it with
 984          * FOLL_WRITE, an atomic set_bit will be required on the pud to
 985          * set the young bit, instead of the current set_pud_at.
 986          */
 987         _pud = pud_mkyoung(pud_mkdirty(*pud));
 988         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
 989                                 pud, _pud,  1))
 990                 update_mmu_cache_pud(vma, addr, pud);
 991 }
 992
 993 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 994                 pud_t *pud, int flags)
 995 {
 996         unsigned long pfn = pud_pfn(*pud);
 997         struct mm_struct *mm = vma->vm_mm;
 998         struct dev_pagemap *pgmap;
 999         struct page *page;
1000
1001         assert_spin_locked(pud_lockptr(mm, pud));
1002
1003         if (flags & FOLL_WRITE && !pud_write(*pud))
1004                 return NULL;
1005
1006         if (pud_present(*pud) && pud_devmap(*pud))
1007                 /* pass */;
1008         else
1009                 return NULL;
1010
1011         if (flags & FOLL_TOUCH)
1012                 touch_pud(vma, addr, pud);
1013
1014         /*
1015          * device mapped pages can only be returned if the
1016          * caller will manage the page reference count.
1017          */
1018         if (!(flags & FOLL_GET))
1019                 return ERR_PTR(-EEXIST);
1020
1021         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1022         pgmap = get_dev_pagemap(pfn, NULL);
1023         if (!pgmap)
1024                 return ERR_PTR(-EFAULT);
1025         page = pfn_to_page(pfn);
1026         get_page(page);
1027         put_dev_pagemap(pgmap);
1028
1029         return page;
1030 }
1031
1032 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1034                   struct vm_area_struct *vma)
1035 {
1036         spinlock_t *dst_ptl, *src_ptl;
1037         pud_t pud;
1038         int ret;
1039
1040         dst_ptl = pud_lock(dst_mm, dst_pud);
1041         src_ptl = pud_lockptr(src_mm, src_pud);
1042         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1043
1044         ret = -EAGAIN;
1045         pud = *src_pud;
1046         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1047                 goto out_unlock;
1048
1049         /*
1050          * When page table lock is held, the huge zero pud should not be
1051          * under splitting since we don't split the page itself, only pud to
1052          * a page table.
1053          */
1054         if (is_huge_zero_pud(pud)) {
1055                 /* No huge zero pud yet */
1056         }
1057
1058         pudp_set_wrprotect(src_mm, addr, src_pud);
1059         pud = pud_mkold(pud_wrprotect(pud));
1060         set_pud_at(dst_mm, addr, dst_pud, pud);
1061
1062         ret = 0;
1063 out_unlock:
1064         spin_unlock(src_ptl);
1065         spin_unlock(dst_ptl);
1066         return ret;
1067 }
1068
1069 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1070 {
1071         pud_t entry;
1072         unsigned long haddr;
1073         bool write = vmf->flags & FAULT_FLAG_WRITE;
1074
1075         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1076         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1077                 goto unlock;
1078
1079         entry = pud_mkyoung(orig_pud);
1080         if (write)
1081                 entry = pud_mkdirty(entry);
1082         haddr = vmf->address & HPAGE_PUD_MASK;
1083         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1084                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1085
1086 unlock:
1087         spin_unlock(vmf->ptl);
1088 }
1089 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1090
1091 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1092 {
1093         pmd_t entry;
1094         unsigned long haddr;
1095         bool write = vmf->flags & FAULT_FLAG_WRITE;
1096
1097         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1098         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1099                 goto unlock;
1100
1101         entry = pmd_mkyoung(orig_pmd);
1102         if (write)
1103                 entry = pmd_mkdirty(entry);
1104         haddr = vmf->address & HPAGE_PMD_MASK;
1105         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1106                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1107
1108 unlock:
1109         spin_unlock(vmf->ptl);
1110 }
1111
1112 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1113                 struct page *page)
1114 {
1115         struct vm_area_struct *vma = vmf->vma;
1116         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1117         struct mem_cgroup *memcg;
1118         pgtable_t pgtable;
1119         pmd_t _pmd;
1120         int ret = 0, i;
1121         struct page **pages;
1122         unsigned long mmun_start;       /* For mmu_notifiers */
1123         unsigned long mmun_end;         /* For mmu_notifiers */
1124
1125         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1126                         GFP_KERNEL);
1127         if (unlikely(!pages)) {
1128                 ret |= VM_FAULT_OOM;
1129                 goto out;
1130         }
1131
1132         for (i = 0; i < HPAGE_PMD_NR; i++) {
1133                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1134                                                vmf->address, page_to_nid(page));
1135                 if (unlikely(!pages[i] ||
1136                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1137                                      GFP_KERNEL, &memcg, false))) {
1138                         if (pages[i])
1139                                 put_page(pages[i]);
1140                         while (--i >= 0) {
1141                                 memcg = (void *)page_private(pages[i]);
1142                                 set_page_private(pages[i], 0);
1143                                 mem_cgroup_cancel_charge(pages[i], memcg,
1144                                                 false);
1145                                 put_page(pages[i]);
1146                         }
1147                         kfree(pages);
1148                         ret |= VM_FAULT_OOM;
1149                         goto out;
1150                 }
1151                 set_page_private(pages[i], (unsigned long)memcg);
1152         }
1153
1154         for (i = 0; i < HPAGE_PMD_NR; i++) {
1155                 copy_user_highpage(pages[i], page + i,
1156                                    haddr + PAGE_SIZE * i, vma);
1157                 __SetPageUptodate(pages[i]);
1158                 cond_resched();
1159         }
1160
1161         mmun_start = haddr;
1162         mmun_end   = haddr + HPAGE_PMD_SIZE;
1163         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1164
1165         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1166         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1167                 goto out_free_pages;
1168         VM_BUG_ON_PAGE(!PageHead(page), page);
1169
1170         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1171         /* leave pmd empty until pte is filled */
1172
1173         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1174         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1175
1176         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1177                 pte_t entry;
1178                 entry = mk_pte(pages[i], vma->vm_page_prot);
1179                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1180                 memcg = (void *)page_private(pages[i]);
1181                 set_page_private(pages[i], 0);
1182                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1183                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1184                 lru_cache_add_active_or_unevictable(pages[i], vma);
1185                 vmf->pte = pte_offset_map(&_pmd, haddr);
1186                 VM_BUG_ON(!pte_none(*vmf->pte));
1187                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1188                 pte_unmap(vmf->pte);
1189         }
1190         kfree(pages);
1191
1192         smp_wmb(); /* make pte visible before pmd */
1193         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1194         page_remove_rmap(page, true);
1195         spin_unlock(vmf->ptl);
1196
1197         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1198
1199         ret |= VM_FAULT_WRITE;
1200         put_page(page);
1201
1202 out:
1203         return ret;
1204
1205 out_free_pages:
1206         spin_unlock(vmf->ptl);
1207         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1208         for (i = 0; i < HPAGE_PMD_NR; i++) {
1209                 memcg = (void *)page_private(pages[i]);
1210                 set_page_private(pages[i], 0);
1211                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1212                 put_page(pages[i]);
1213         }
1214         kfree(pages);
1215         goto out;
1216 }
1217
1218 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1219 {
1220         struct vm_area_struct *vma = vmf->vma;
1221         struct page *page = NULL, *new_page;
1222         struct mem_cgroup *memcg;
1223         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1224         unsigned long mmun_start;       /* For mmu_notifiers */
1225         unsigned long mmun_end;         /* For mmu_notifiers */
1226         gfp_t huge_gfp;                 /* for allocation and charge */
1227         int ret = 0;
1228
1229         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1230         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1231         if (is_huge_zero_pmd(orig_pmd))
1232                 goto alloc;
1233         spin_lock(vmf->ptl);
1234         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1235                 goto out_unlock;
1236
1237         page = pmd_page(orig_pmd);
1238         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1239         /*
1240          * We can only reuse the page if nobody else maps the huge page or it's
1241          * part.
1242          */
1243         if (!trylock_page(page)) {
1244                 get_page(page);
1245                 spin_unlock(vmf->ptl);
1246                 lock_page(page);
1247                 spin_lock(vmf->ptl);
1248                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1249                         unlock_page(page);
1250                         put_page(page);
1251                         goto out_unlock;
1252                 }
1253                 put_page(page);
1254         }
1255         if (reuse_swap_page(page, NULL)) {
1256                 pmd_t entry;
1257                 entry = pmd_mkyoung(orig_pmd);
1258                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1259                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1260                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1261                 ret |= VM_FAULT_WRITE;
1262                 unlock_page(page);
1263                 goto out_unlock;
1264         }
1265         unlock_page(page);
1266         get_page(page);
1267         spin_unlock(vmf->ptl);
1268 alloc:
1269         if (transparent_hugepage_enabled(vma) &&
1270             !transparent_hugepage_debug_cow()) {
1271                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1272                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1273         } else
1274                 new_page = NULL;
1275
1276         if (likely(new_page)) {
1277                 prep_transhuge_page(new_page);
1278         } else {
1279                 if (!page) {
1280                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1281                         ret |= VM_FAULT_FALLBACK;
1282                 } else {
1283                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1284                         if (ret & VM_FAULT_OOM) {
1285                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1286                                 ret |= VM_FAULT_FALLBACK;
1287                         }
1288                         put_page(page);
1289                 }
1290                 count_vm_event(THP_FAULT_FALLBACK);
1291                 goto out;
1292         }
1293
1294         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1295                                         huge_gfp, &memcg, true))) {
1296                 put_page(new_page);
1297                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1298                 if (page)
1299                         put_page(page);
1300                 ret |= VM_FAULT_FALLBACK;
1301                 count_vm_event(THP_FAULT_FALLBACK);
1302                 goto out;
1303         }
1304
1305         count_vm_event(THP_FAULT_ALLOC);
1306
1307         if (!page)
1308                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1309         else
1310                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1311         __SetPageUptodate(new_page);
1312
1313         mmun_start = haddr;
1314         mmun_end   = haddr + HPAGE_PMD_SIZE;
1315         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1316
1317         spin_lock(vmf->ptl);
1318         if (page)
1319                 put_page(page);
1320         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1321                 spin_unlock(vmf->ptl);
1322                 mem_cgroup_cancel_charge(new_page, memcg, true);
1323                 put_page(new_page);
1324                 goto out_mn;
1325         } else {
1326                 pmd_t entry;
1327                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1328                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1329                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1330                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1331                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1332                 lru_cache_add_active_or_unevictable(new_page, vma);
1333                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1334                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1335                 if (!page) {
1336                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1337                 } else {
1338                         VM_BUG_ON_PAGE(!PageHead(page), page);
1339                         page_remove_rmap(page, true);
1340                         put_page(page);
1341                 }
1342                 ret |= VM_FAULT_WRITE;
1343         }
1344         spin_unlock(vmf->ptl);
1345 out_mn:
1346         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1347 out:
1348         return ret;
1349 out_unlock:
1350         spin_unlock(vmf->ptl);
1351         return ret;
1352 }
1353
1354 /*
1355  * FOLL_FORCE can write to even unwritable pmd's, but only
1356  * after we've gone through a COW cycle and they are dirty.
1357  */
1358 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1359 {
1360         return pmd_write(pmd) ||
1361                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1362 }
1363
1364 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1365                                    unsigned long addr,
1366                                    pmd_t *pmd,
1367                                    unsigned int flags)
1368 {
1369         struct mm_struct *mm = vma->vm_mm;
1370         struct page *page = NULL;
1371
1372         assert_spin_locked(pmd_lockptr(mm, pmd));
1373
1374         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1375                 goto out;
1376
1377         /* Avoid dumping huge zero page */
1378         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1379                 return ERR_PTR(-EFAULT);
1380
1381         /* Full NUMA hinting faults to serialise migration in fault paths */
1382         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1383                 goto out;
1384
1385         page = pmd_page(*pmd);
1386         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1387         if (flags & FOLL_TOUCH)
1388                 touch_pmd(vma, addr, pmd);
1389         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1390                 /*
1391                  * We don't mlock() pte-mapped THPs. This way we can avoid
1392                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1393                  *
1394                  * For anon THP:
1395                  *
1396                  * In most cases the pmd is the only mapping of the page as we
1397                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1398                  * writable private mappings in populate_vma_page_range().
1399                  *
1400                  * The only scenario when we have the page shared here is if we
1401                  * mlocking read-only mapping shared over fork(). We skip
1402                  * mlocking such pages.
1403                  *
1404                  * For file THP:
1405                  *
1406                  * We can expect PageDoubleMap() to be stable under page lock:
1407                  * for file pages we set it in page_add_file_rmap(), which
1408                  * requires page to be locked.
1409                  */
1410
1411                 if (PageAnon(page) && compound_mapcount(page) != 1)
1412                         goto skip_mlock;
1413                 if (PageDoubleMap(page) || !page->mapping)
1414                         goto skip_mlock;
1415                 if (!trylock_page(page))
1416                         goto skip_mlock;
1417                 lru_add_drain();
1418                 if (page->mapping && !PageDoubleMap(page))
1419                         mlock_vma_page(page);
1420                 unlock_page(page);
1421         }
1422 skip_mlock:
1423         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1424         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1425         if (flags & FOLL_GET)
1426                 get_page(page);
1427
1428 out:
1429         return page;
1430 }
1431
1432 /* NUMA hinting page fault entry point for trans huge pmds */
1433 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1434 {
1435         struct vm_area_struct *vma = vmf->vma;
1436         struct anon_vma *anon_vma = NULL;
1437         struct page *page;
1438         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1439         int page_nid = -1, this_nid = numa_node_id();
1440         int target_nid, last_cpupid = -1;
1441         bool page_locked;
1442         bool migrated = false;
1443         bool was_writable;
1444         int flags = 0;
1445
1446         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1447         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1448                 goto out_unlock;
1449
1450         /*
1451          * If there are potential migrations, wait for completion and retry
1452          * without disrupting NUMA hinting information. Do not relock and
1453          * check_same as the page may no longer be mapped.
1454          */
1455         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1456                 page = pmd_page(*vmf->pmd);
1457                 if (!get_page_unless_zero(page))
1458                         goto out_unlock;
1459                 spin_unlock(vmf->ptl);
1460                 wait_on_page_locked(page);
1461                 put_page(page);
1462                 goto out;
1463         }
1464
1465         page = pmd_page(pmd);
1466         BUG_ON(is_huge_zero_page(page));
1467         page_nid = page_to_nid(page);
1468         last_cpupid = page_cpupid_last(page);
1469         count_vm_numa_event(NUMA_HINT_FAULTS);
1470         if (page_nid == this_nid) {
1471                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1472                 flags |= TNF_FAULT_LOCAL;
1473         }
1474
1475         /* See similar comment in do_numa_page for explanation */
1476         if (!pmd_savedwrite(pmd))
1477                 flags |= TNF_NO_GROUP;
1478
1479         /*
1480          * Acquire the page lock to serialise THP migrations but avoid dropping
1481          * page_table_lock if at all possible
1482          */
1483         page_locked = trylock_page(page);
1484         target_nid = mpol_misplaced(page, vma, haddr);
1485         if (target_nid == -1) {
1486                 /* If the page was locked, there are no parallel migrations */
1487                 if (page_locked)
1488                         goto clear_pmdnuma;
1489         }
1490
1491         /* Migration could have started since the pmd_trans_migrating check */
1492         if (!page_locked) {
1493                 page_nid = -1;
1494                 if (!get_page_unless_zero(page))
1495                         goto out_unlock;
1496                 spin_unlock(vmf->ptl);
1497                 wait_on_page_locked(page);
1498                 put_page(page);
1499                 goto out;
1500         }
1501
1502         /*
1503          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1504          * to serialises splits
1505          */
1506         get_page(page);
1507         spin_unlock(vmf->ptl);
1508         anon_vma = page_lock_anon_vma_read(page);
1509
1510         /* Confirm the PMD did not change while page_table_lock was released */
1511         spin_lock(vmf->ptl);
1512         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1513                 unlock_page(page);
1514                 put_page(page);
1515                 page_nid = -1;
1516                 goto out_unlock;
1517         }
1518
1519         /* Bail if we fail to protect against THP splits for any reason */
1520         if (unlikely(!anon_vma)) {
1521                 put_page(page);
1522                 page_nid = -1;
1523                 goto clear_pmdnuma;
1524         }
1525
1526         /*
1527          * Since we took the NUMA fault, we must have observed the !accessible
1528          * bit. Make sure all other CPUs agree with that, to avoid them
1529          * modifying the page we're about to migrate.
1530          *
1531          * Must be done under PTL such that we'll observe the relevant
1532          * inc_tlb_flush_pending().
1533          *
1534          * We are not sure a pending tlb flush here is for a huge page
1535          * mapping or not. Hence use the tlb range variant
1536          */
1537         if (mm_tlb_flush_pending(vma->vm_mm))
1538                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1539
1540         /*
1541          * Migrate the THP to the requested node, returns with page unlocked
1542          * and access rights restored.
1543          */
1544         spin_unlock(vmf->ptl);
1545
1546         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1547                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1548         if (migrated) {
1549                 flags |= TNF_MIGRATED;
1550                 page_nid = target_nid;
1551         } else
1552                 flags |= TNF_MIGRATE_FAIL;
1553
1554         goto out;
1555 clear_pmdnuma:
1556         BUG_ON(!PageLocked(page));
1557         was_writable = pmd_savedwrite(pmd);
1558         pmd = pmd_modify(pmd, vma->vm_page_prot);
1559         pmd = pmd_mkyoung(pmd);
1560         if (was_writable)
1561                 pmd = pmd_mkwrite(pmd);
1562         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1563         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1564         unlock_page(page);
1565 out_unlock:
1566         spin_unlock(vmf->ptl);
1567
1568 out:
1569         if (anon_vma)
1570                 page_unlock_anon_vma_read(anon_vma);
1571
1572         if (page_nid != -1)
1573                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1574                                 flags);
1575
1576         return 0;
1577 }
1578
1579 /*
1580  * Return true if we do MADV_FREE successfully on entire pmd page.
1581  * Otherwise, return false.
1582  */
1583 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1584                 pmd_t *pmd, unsigned long addr, unsigned long next)
1585 {
1586         spinlock_t *ptl;
1587         pmd_t orig_pmd;
1588         struct page *page;
1589         struct mm_struct *mm = tlb->mm;
1590         bool ret = false;
1591
1592         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1593
1594         ptl = pmd_trans_huge_lock(pmd, vma);
1595         if (!ptl)
1596                 goto out_unlocked;
1597
1598         orig_pmd = *pmd;
1599         if (is_huge_zero_pmd(orig_pmd))
1600                 goto out;
1601
1602         page = pmd_page(orig_pmd);
1603         /*
1604          * If other processes are mapping this page, we couldn't discard
1605          * the page unless they all do MADV_FREE so let's skip the page.
1606          */
1607         if (page_mapcount(page) != 1)
1608                 goto out;
1609
1610         if (!trylock_page(page))
1611                 goto out;
1612
1613         /*
1614          * If user want to discard part-pages of THP, split it so MADV_FREE
1615          * will deactivate only them.
1616          */
1617         if (next - addr != HPAGE_PMD_SIZE) {
1618                 get_page(page);
1619                 spin_unlock(ptl);
1620                 split_huge_page(page);
1621                 unlock_page(page);
1622                 put_page(page);
1623                 goto out_unlocked;
1624         }
1625
1626         if (PageDirty(page))
1627                 ClearPageDirty(page);
1628         unlock_page(page);
1629
1630         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1631                 pmdp_invalidate(vma, addr, pmd);
1632                 orig_pmd = pmd_mkold(orig_pmd);
1633                 orig_pmd = pmd_mkclean(orig_pmd);
1634
1635                 set_pmd_at(mm, addr, pmd, orig_pmd);
1636                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1637         }
1638
1639         mark_page_lazyfree(page);
1640         ret = true;
1641 out:
1642         spin_unlock(ptl);
1643 out_unlocked:
1644         return ret;
1645 }
1646
1647 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1648 {
1649         pgtable_t pgtable;
1650
1651         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1652         pte_free(mm, pgtable);
1653         atomic_long_dec(&mm->nr_ptes);
1654 }
1655
1656 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1657                  pmd_t *pmd, unsigned long addr)
1658 {
1659         pmd_t orig_pmd;
1660         spinlock_t *ptl;
1661
1662         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1663
1664         ptl = __pmd_trans_huge_lock(pmd, vma);
1665         if (!ptl)
1666                 return 0;
1667         /*
1668          * For architectures like ppc64 we look at deposited pgtable
1669          * when calling pmdp_huge_get_and_clear. So do the
1670          * pgtable_trans_huge_withdraw after finishing pmdp related
1671          * operations.
1672          */
1673         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1674                         tlb->fullmm);
1675         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1676         if (vma_is_dax(vma)) {
1677                 if (arch_needs_pgtable_deposit())
1678                         zap_deposited_table(tlb->mm, pmd);
1679                 spin_unlock(ptl);
1680                 if (is_huge_zero_pmd(orig_pmd))
1681                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1682         } else if (is_huge_zero_pmd(orig_pmd)) {
1683                 zap_deposited_table(tlb->mm, pmd);
1684                 spin_unlock(ptl);
1685                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1686         } else {
1687                 struct page *page = pmd_page(orig_pmd);
1688                 page_remove_rmap(page, true);
1689                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1690                 VM_BUG_ON_PAGE(!PageHead(page), page);
1691                 if (PageAnon(page)) {
1692                         zap_deposited_table(tlb->mm, pmd);
1693                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1694                 } else {
1695                         if (arch_needs_pgtable_deposit())
1696                                 zap_deposited_table(tlb->mm, pmd);
1697                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1698                 }
1699                 spin_unlock(ptl);
1700                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1701         }
1702         return 1;
1703 }
1704
1705 #ifndef pmd_move_must_withdraw
1706 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1707                                          spinlock_t *old_pmd_ptl,
1708                                          struct vm_area_struct *vma)
1709 {
1710         /*
1711          * With split pmd lock we also need to move preallocated
1712          * PTE page table if new_pmd is on different PMD page table.
1713          *
1714          * We also don't deposit and withdraw tables for file pages.
1715          */
1716         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1717 }
1718 #endif
1719
1720 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1721                   unsigned long new_addr, unsigned long old_end,
1722                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1723 {
1724         spinlock_t *old_ptl, *new_ptl;
1725         pmd_t pmd;
1726         struct mm_struct *mm = vma->vm_mm;
1727         bool force_flush = false;
1728
1729         if ((old_addr & ~HPAGE_PMD_MASK) ||
1730             (new_addr & ~HPAGE_PMD_MASK) ||
1731             old_end - old_addr < HPAGE_PMD_SIZE)
1732                 return false;
1733
1734         /*
1735          * The destination pmd shouldn't be established, free_pgtables()
1736          * should have release it.
1737          */
1738         if (WARN_ON(!pmd_none(*new_pmd))) {
1739                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1740                 return false;
1741         }
1742
1743         /*
1744          * We don't have to worry about the ordering of src and dst
1745          * ptlocks because exclusive mmap_sem prevents deadlock.
1746          */
1747         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1748         if (old_ptl) {
1749                 new_ptl = pmd_lockptr(mm, new_pmd);
1750                 if (new_ptl != old_ptl)
1751                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1752                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1753                 if (pmd_present(pmd) && pmd_dirty(pmd))
1754                         force_flush = true;
1755                 VM_BUG_ON(!pmd_none(*new_pmd));
1756
1757                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1758                         pgtable_t pgtable;
1759                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1760                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1761                 }
1762                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1763                 if (new_ptl != old_ptl)
1764                         spin_unlock(new_ptl);
1765                 if (force_flush)
1766                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1767                 else
1768                         *need_flush = true;
1769                 spin_unlock(old_ptl);
1770                 return true;
1771         }
1772         return false;
1773 }
1774
1775 /*
1776  * Returns
1777  *  - 0 if PMD could not be locked
1778  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1779  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1780  */
1781 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1782                 unsigned long addr, pgprot_t newprot, int prot_numa)
1783 {
1784         struct mm_struct *mm = vma->vm_mm;
1785         spinlock_t *ptl;
1786         pmd_t entry;
1787         bool preserve_write;
1788         int ret;
1789
1790         ptl = __pmd_trans_huge_lock(pmd, vma);
1791         if (!ptl)
1792                 return 0;
1793
1794         preserve_write = prot_numa && pmd_write(*pmd);
1795         ret = 1;
1796
1797         /*
1798          * Avoid trapping faults against the zero page. The read-only
1799          * data is likely to be read-cached on the local CPU and
1800          * local/remote hits to the zero page are not interesting.
1801          */
1802         if (prot_numa && is_huge_zero_pmd(*pmd))
1803                 goto unlock;
1804
1805         if (prot_numa && pmd_protnone(*pmd))
1806                 goto unlock;
1807
1808         /*
1809          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1810          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1811          * which is also under down_read(mmap_sem):
1812          *
1813          *      CPU0:                           CPU1:
1814          *                              change_huge_pmd(prot_numa=1)
1815          *                               pmdp_huge_get_and_clear_notify()
1816          * madvise_dontneed()
1817          *  zap_pmd_range()
1818          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1819          *   // skip the pmd
1820          *                               set_pmd_at();
1821          *                               // pmd is re-established
1822          *
1823          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1824          * which may break userspace.
1825          *
1826          * pmdp_invalidate() is required to make sure we don't miss
1827          * dirty/young flags set by hardware.
1828          */
1829         entry = *pmd;
1830         pmdp_invalidate(vma, addr, pmd);
1831
1832         /*
1833          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1834          * corrupt them.
1835          */
1836         if (pmd_dirty(*pmd))
1837                 entry = pmd_mkdirty(entry);
1838         if (pmd_young(*pmd))
1839                 entry = pmd_mkyoung(entry);
1840
1841         entry = pmd_modify(entry, newprot);
1842         if (preserve_write)
1843                 entry = pmd_mk_savedwrite(entry);
1844         ret = HPAGE_PMD_NR;
1845         set_pmd_at(mm, addr, pmd, entry);
1846         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1847 unlock:
1848         spin_unlock(ptl);
1849         return ret;
1850 }
1851
1852 /*
1853  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1854  *
1855  * Note that if it returns page table lock pointer, this routine returns without
1856  * unlocking page table lock. So callers must unlock it.
1857  */
1858 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1859 {
1860         spinlock_t *ptl;
1861         ptl = pmd_lock(vma->vm_mm, pmd);
1862         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1863                 return ptl;
1864         spin_unlock(ptl);
1865         return NULL;
1866 }
1867
1868 /*
1869  * Returns true if a given pud maps a thp, false otherwise.
1870  *
1871  * Note that if it returns true, this routine returns without unlocking page
1872  * table lock. So callers must unlock it.
1873  */
1874 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1875 {
1876         spinlock_t *ptl;
1877
1878         ptl = pud_lock(vma->vm_mm, pud);
1879         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1880                 return ptl;
1881         spin_unlock(ptl);
1882         return NULL;
1883 }
1884
1885 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1886 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1887                  pud_t *pud, unsigned long addr)
1888 {
1889         pud_t orig_pud;
1890         spinlock_t *ptl;
1891
1892         ptl = __pud_trans_huge_lock(pud, vma);
1893         if (!ptl)
1894                 return 0;
1895         /*
1896          * For architectures like ppc64 we look at deposited pgtable
1897          * when calling pudp_huge_get_and_clear. So do the
1898          * pgtable_trans_huge_withdraw after finishing pudp related
1899          * operations.
1900          */
1901         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1902                         tlb->fullmm);
1903         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1904         if (vma_is_dax(vma)) {
1905                 spin_unlock(ptl);
1906                 /* No zero page support yet */
1907         } else {
1908                 /* No support for anonymous PUD pages yet */
1909                 BUG();
1910         }
1911         return 1;
1912 }
1913
1914 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1915                 unsigned long haddr)
1916 {
1917         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1918         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1919         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1920         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1921
1922         count_vm_event(THP_SPLIT_PUD);
1923
1924         pudp_huge_clear_flush_notify(vma, haddr, pud);
1925 }
1926
1927 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1928                 unsigned long address)
1929 {
1930         spinlock_t *ptl;
1931         struct mm_struct *mm = vma->vm_mm;
1932         unsigned long haddr = address & HPAGE_PUD_MASK;
1933
1934         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1935         ptl = pud_lock(mm, pud);
1936         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1937                 goto out;
1938         __split_huge_pud_locked(vma, pud, haddr);
1939
1940 out:
1941         spin_unlock(ptl);
1942         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1943 }
1944 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1945
1946 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1947                 unsigned long haddr, pmd_t *pmd)
1948 {
1949         struct mm_struct *mm = vma->vm_mm;
1950         pgtable_t pgtable;
1951         pmd_t _pmd;
1952         int i;
1953
1954         /* leave pmd empty until pte is filled */
1955         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1956
1957         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1958         pmd_populate(mm, &_pmd, pgtable);
1959
1960         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1961                 pte_t *pte, entry;
1962                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1963                 entry = pte_mkspecial(entry);
1964                 pte = pte_offset_map(&_pmd, haddr);
1965                 VM_BUG_ON(!pte_none(*pte));
1966                 set_pte_at(mm, haddr, pte, entry);
1967                 pte_unmap(pte);
1968         }
1969         smp_wmb(); /* make pte visible before pmd */
1970         pmd_populate(mm, pmd, pgtable);
1971 }
1972
1973 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1974                 unsigned long haddr, bool freeze)
1975 {
1976         struct mm_struct *mm = vma->vm_mm;
1977         struct page *page;
1978         pgtable_t pgtable;
1979         pmd_t _pmd;
1980         bool young, write, dirty, soft_dirty;
1981         unsigned long addr;
1982         int i;
1983
1984         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1985         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1986         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1987         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1988
1989         count_vm_event(THP_SPLIT_PMD);
1990
1991         if (!vma_is_anonymous(vma)) {
1992                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1993                 /*
1994                  * We are going to unmap this huge page. So
1995                  * just go ahead and zap it
1996                  */
1997                 if (arch_needs_pgtable_deposit())
1998                         zap_deposited_table(mm, pmd);
1999                 if (vma_is_dax(vma))
2000                         return;
2001                 page = pmd_page(_pmd);
2002                 if (!PageReferenced(page) && pmd_young(_pmd))
2003                         SetPageReferenced(page);
2004                 page_remove_rmap(page, true);
2005                 put_page(page);
2006                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2007                 return;
2008         } else if (is_huge_zero_pmd(*pmd)) {
2009                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2010         }
2011
2012         page = pmd_page(*pmd);
2013         VM_BUG_ON_PAGE(!page_count(page), page);
2014         page_ref_add(page, HPAGE_PMD_NR - 1);
2015         write = pmd_write(*pmd);
2016         young = pmd_young(*pmd);
2017         dirty = pmd_dirty(*pmd);
2018         soft_dirty = pmd_soft_dirty(*pmd);
2019
2020         pmdp_huge_split_prepare(vma, haddr, pmd);
2021         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2022         pmd_populate(mm, &_pmd, pgtable);
2023
2024         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2025                 pte_t entry, *pte;
2026                 /*
2027                  * Note that NUMA hinting access restrictions are not
2028                  * transferred to avoid any possibility of altering
2029                  * permissions across VMAs.
2030                  */
2031                 if (freeze) {
2032                         swp_entry_t swp_entry;
2033                         swp_entry = make_migration_entry(page + i, write);
2034                         entry = swp_entry_to_pte(swp_entry);
2035                         if (soft_dirty)
2036                                 entry = pte_swp_mksoft_dirty(entry);
2037                 } else {
2038                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2039                         entry = maybe_mkwrite(entry, vma);
2040                         if (!write)
2041                                 entry = pte_wrprotect(entry);
2042                         if (!young)
2043                                 entry = pte_mkold(entry);
2044                         if (soft_dirty)
2045                                 entry = pte_mksoft_dirty(entry);
2046                 }
2047                 if (dirty)
2048                         SetPageDirty(page + i);
2049                 pte = pte_offset_map(&_pmd, addr);
2050                 BUG_ON(!pte_none(*pte));
2051                 set_pte_at(mm, addr, pte, entry);
2052                 atomic_inc(&page[i]._mapcount);
2053                 pte_unmap(pte);
2054         }
2055
2056         /*
2057          * Set PG_double_map before dropping compound_mapcount to avoid
2058          * false-negative page_mapped().
2059          */
2060         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2061                 for (i = 0; i < HPAGE_PMD_NR; i++)
2062                         atomic_inc(&page[i]._mapcount);
2063         }
2064
2065         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2066                 /* Last compound_mapcount is gone. */
2067                 __dec_node_page_state(page, NR_ANON_THPS);
2068                 if (TestClearPageDoubleMap(page)) {
2069                         /* No need in mapcount reference anymore */
2070                         for (i = 0; i < HPAGE_PMD_NR; i++)
2071                                 atomic_dec(&page[i]._mapcount);
2072                 }
2073         }
2074
2075         smp_wmb(); /* make pte visible before pmd */
2076         /*
2077          * Up to this point the pmd is present and huge and userland has the
2078          * whole access to the hugepage during the split (which happens in
2079          * place). If we overwrite the pmd with the not-huge version pointing
2080          * to the pte here (which of course we could if all CPUs were bug
2081          * free), userland could trigger a small page size TLB miss on the
2082          * small sized TLB while the hugepage TLB entry is still established in
2083          * the huge TLB. Some CPU doesn't like that.
2084          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2085          * 383 on page 93. Intel should be safe but is also warns that it's
2086          * only safe if the permission and cache attributes of the two entries
2087          * loaded in the two TLB is identical (which should be the case here).
2088          * But it is generally safer to never allow small and huge TLB entries
2089          * for the same virtual address to be loaded simultaneously. So instead
2090          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2091          * current pmd notpresent (atomically because here the pmd_trans_huge
2092          * and pmd_trans_splitting must remain set at all times on the pmd
2093          * until the split is complete for this pmd), then we flush the SMP TLB
2094          * and finally we write the non-huge version of the pmd entry with
2095          * pmd_populate.
2096          */
2097         pmdp_invalidate(vma, haddr, pmd);
2098         pmd_populate(mm, pmd, pgtable);
2099
2100         if (freeze) {
2101                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2102                         page_remove_rmap(page + i, false);
2103                         put_page(page + i);
2104                 }
2105         }
2106 }
2107
2108 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2109                 unsigned long address, bool freeze, struct page *page)
2110 {
2111         spinlock_t *ptl;
2112         struct mm_struct *mm = vma->vm_mm;
2113         unsigned long haddr = address & HPAGE_PMD_MASK;
2114
2115         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2116         ptl = pmd_lock(mm, pmd);
2117
2118         /*
2119          * If caller asks to setup a migration entries, we need a page to check
2120          * pmd against. Otherwise we can end up replacing wrong page.
2121          */
2122         VM_BUG_ON(freeze && !page);
2123         if (page && page != pmd_page(*pmd))
2124                 goto out;
2125
2126         if (pmd_trans_huge(*pmd)) {
2127                 page = pmd_page(*pmd);
2128                 if (PageMlocked(page))
2129                         clear_page_mlock(page);
2130         } else if (!pmd_devmap(*pmd))
2131                 goto out;
2132         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2133 out:
2134         spin_unlock(ptl);
2135         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2136 }
2137
2138 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2139                 bool freeze, struct page *page)
2140 {
2141         pgd_t *pgd;
2142         p4d_t *p4d;
2143         pud_t *pud;
2144         pmd_t *pmd;
2145
2146         pgd = pgd_offset(vma->vm_mm, address);
2147         if (!pgd_present(*pgd))
2148                 return;
2149
2150         p4d = p4d_offset(pgd, address);
2151         if (!p4d_present(*p4d))
2152                 return;
2153
2154         pud = pud_offset(p4d, address);
2155         if (!pud_present(*pud))
2156                 return;
2157
2158         pmd = pmd_offset(pud, address);
2159
2160         __split_huge_pmd(vma, pmd, address, freeze, page);
2161 }
2162
2163 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2164                              unsigned long start,
2165                              unsigned long end,
2166                              long adjust_next)
2167 {
2168         /*
2169          * If the new start address isn't hpage aligned and it could
2170          * previously contain an hugepage: check if we need to split
2171          * an huge pmd.
2172          */
2173         if (start & ~HPAGE_PMD_MASK &&
2174             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2175             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2176                 split_huge_pmd_address(vma, start, false, NULL);
2177
2178         /*
2179          * If the new end address isn't hpage aligned and it could
2180          * previously contain an hugepage: check if we need to split
2181          * an huge pmd.
2182          */
2183         if (end & ~HPAGE_PMD_MASK &&
2184             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2185             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2186                 split_huge_pmd_address(vma, end, false, NULL);
2187
2188         /*
2189          * If we're also updating the vma->vm_next->vm_start, if the new
2190          * vm_next->vm_start isn't page aligned and it could previously
2191          * contain an hugepage: check if we need to split an huge pmd.
2192          */
2193         if (adjust_next > 0) {
2194                 struct vm_area_struct *next = vma->vm_next;
2195                 unsigned long nstart = next->vm_start;
2196                 nstart += adjust_next << PAGE_SHIFT;
2197                 if (nstart & ~HPAGE_PMD_MASK &&
2198                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2199                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2200                         split_huge_pmd_address(next, nstart, false, NULL);
2201         }
2202 }
2203
2204 static void freeze_page(struct page *page)
2205 {
2206         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2207                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2208         bool unmap_success;
2209
2210         VM_BUG_ON_PAGE(!PageHead(page), page);
2211
2212         if (PageAnon(page))
2213                 ttu_flags |= TTU_MIGRATION;
2214
2215         unmap_success = try_to_unmap(page, ttu_flags);
2216         VM_BUG_ON_PAGE(!unmap_success, page);
2217 }
2218
2219 static void unfreeze_page(struct page *page)
2220 {
2221         int i;
2222         if (PageTransHuge(page)) {
2223                 remove_migration_ptes(page, page, true);
2224         } else {
2225                 for (i = 0; i < HPAGE_PMD_NR; i++)
2226                         remove_migration_ptes(page + i, page + i, true);
2227         }
2228 }
2229
2230 static void __split_huge_page_tail(struct page *head, int tail,
2231                 struct lruvec *lruvec, struct list_head *list)
2232 {
2233         struct page *page_tail = head + tail;
2234
2235         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2236         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2237
2238         /*
2239          * tail_page->_refcount is zero and not changing from under us. But
2240          * get_page_unless_zero() may be running from under us on the
2241          * tail_page. If we used atomic_set() below instead of atomic_inc() or
2242          * atomic_add(), we would then run atomic_set() concurrently with
2243          * get_page_unless_zero(), and atomic_set() is implemented in C not
2244          * using locked ops. spin_unlock on x86 sometime uses locked ops
2245          * because of PPro errata 66, 92, so unless somebody can guarantee
2246          * atomic_set() here would be safe on all archs (and not only on x86),
2247          * it's safer to use atomic_inc()/atomic_add().
2248          */
2249         if (PageAnon(head) && !PageSwapCache(head)) {
2250                 page_ref_inc(page_tail);
2251         } else {
2252                 /* Additional pin to radix tree */
2253                 page_ref_add(page_tail, 2);
2254         }
2255
2256         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2257         page_tail->flags |= (head->flags &
2258                         ((1L << PG_referenced) |
2259                          (1L << PG_swapbacked) |
2260                          (1L << PG_swapcache) |
2261                          (1L << PG_mlocked) |
2262                          (1L << PG_uptodate) |
2263                          (1L << PG_active) |
2264                          (1L << PG_locked) |
2265                          (1L << PG_unevictable) |
2266                          (1L << PG_dirty)));
2267
2268         /*
2269          * After clearing PageTail the gup refcount can be released.
2270          * Page flags also must be visible before we make the page non-compound.
2271          */
2272         smp_wmb();
2273
2274         clear_compound_head(page_tail);
2275
2276         if (page_is_young(head))
2277                 set_page_young(page_tail);
2278         if (page_is_idle(head))
2279                 set_page_idle(page_tail);
2280
2281         /* ->mapping in first tail page is compound_mapcount */
2282         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2283                         page_tail);
2284         page_tail->mapping = head->mapping;
2285
2286         page_tail->index = head->index + tail;
2287         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2288         lru_add_page_tail(head, page_tail, lruvec, list);
2289 }
2290
2291 static void __split_huge_page(struct page *page, struct list_head *list,
2292                 unsigned long flags)
2293 {
2294         struct page *head = compound_head(page);
2295         struct zone *zone = page_zone(head);
2296         struct lruvec *lruvec;
2297         pgoff_t end = -1;
2298         int i;
2299
2300         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2301
2302         /* complete memcg works before add pages to LRU */
2303         mem_cgroup_split_huge_fixup(head);
2304
2305         if (!PageAnon(page))
2306                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2307
2308         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2309                 __split_huge_page_tail(head, i, lruvec, list);
2310                 /* Some pages can be beyond i_size: drop them from page cache */
2311                 if (head[i].index >= end) {
2312                         __ClearPageDirty(head + i);
2313                         __delete_from_page_cache(head + i, NULL);
2314                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2315                                 shmem_uncharge(head->mapping->host, 1);
2316                         put_page(head + i);
2317                 }
2318         }
2319
2320         ClearPageCompound(head);
2321         /* See comment in __split_huge_page_tail() */
2322         if (PageAnon(head)) {
2323                 /* Additional pin to radix tree of swap cache */
2324                 if (PageSwapCache(head))
2325                         page_ref_add(head, 2);
2326                 else
2327                         page_ref_inc(head);
2328         } else {
2329                 /* Additional pin to radix tree */
2330                 page_ref_add(head, 2);
2331                 spin_unlock(&head->mapping->tree_lock);
2332         }
2333
2334         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2335
2336         unfreeze_page(head);
2337
2338         for (i = 0; i < HPAGE_PMD_NR; i++) {
2339                 struct page *subpage = head + i;
2340                 if (subpage == page)
2341                         continue;
2342                 unlock_page(subpage);
2343
2344                 /*
2345                  * Subpages may be freed if there wasn't any mapping
2346                  * like if add_to_swap() is running on a lru page that
2347                  * had its mapping zapped. And freeing these pages
2348                  * requires taking the lru_lock so we do the put_page
2349                  * of the tail pages after the split is complete.
2350                  */
2351                 put_page(subpage);
2352         }
2353 }
2354
2355 int total_mapcount(struct page *page)
2356 {
2357         int i, compound, ret;
2358
2359         VM_BUG_ON_PAGE(PageTail(page), page);
2360
2361         if (likely(!PageCompound(page)))
2362                 return atomic_read(&page->_mapcount) + 1;
2363
2364         compound = compound_mapcount(page);
2365         if (PageHuge(page))
2366                 return compound;
2367         ret = compound;
2368         for (i = 0; i < HPAGE_PMD_NR; i++)
2369                 ret += atomic_read(&page[i]._mapcount) + 1;
2370         /* File pages has compound_mapcount included in _mapcount */
2371         if (!PageAnon(page))
2372                 return ret - compound * HPAGE_PMD_NR;
2373         if (PageDoubleMap(page))
2374                 ret -= HPAGE_PMD_NR;
2375         return ret;
2376 }
2377
2378 /*
2379  * This calculates accurately how many mappings a transparent hugepage
2380  * has (unlike page_mapcount() which isn't fully accurate). This full
2381  * accuracy is primarily needed to know if copy-on-write faults can
2382  * reuse the page and change the mapping to read-write instead of
2383  * copying them. At the same time this returns the total_mapcount too.
2384  *
2385  * The function returns the highest mapcount any one of the subpages
2386  * has. If the return value is one, even if different processes are
2387  * mapping different subpages of the transparent hugepage, they can
2388  * all reuse it, because each process is reusing a different subpage.
2389  *
2390  * The total_mapcount is instead counting all virtual mappings of the
2391  * subpages. If the total_mapcount is equal to "one", it tells the
2392  * caller all mappings belong to the same "mm" and in turn the
2393  * anon_vma of the transparent hugepage can become the vma->anon_vma
2394  * local one as no other process may be mapping any of the subpages.
2395  *
2396  * It would be more accurate to replace page_mapcount() with
2397  * page_trans_huge_mapcount(), however we only use
2398  * page_trans_huge_mapcount() in the copy-on-write faults where we
2399  * need full accuracy to avoid breaking page pinning, because
2400  * page_trans_huge_mapcount() is slower than page_mapcount().
2401  */
2402 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2403 {
2404         int i, ret, _total_mapcount, mapcount;
2405
2406         /* hugetlbfs shouldn't call it */
2407         VM_BUG_ON_PAGE(PageHuge(page), page);
2408
2409         if (likely(!PageTransCompound(page))) {
2410                 mapcount = atomic_read(&page->_mapcount) + 1;
2411                 if (total_mapcount)
2412                         *total_mapcount = mapcount;
2413                 return mapcount;
2414         }
2415
2416         page = compound_head(page);
2417
2418         _total_mapcount = ret = 0;
2419         for (i = 0; i < HPAGE_PMD_NR; i++) {
2420                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2421                 ret = max(ret, mapcount);
2422                 _total_mapcount += mapcount;
2423         }
2424         if (PageDoubleMap(page)) {
2425                 ret -= 1;
2426                 _total_mapcount -= HPAGE_PMD_NR;
2427         }
2428         mapcount = compound_mapcount(page);
2429         ret += mapcount;
2430         _total_mapcount += mapcount;
2431         if (total_mapcount)
2432                 *total_mapcount = _total_mapcount;
2433         return ret;
2434 }
2435
2436 /* Racy check whether the huge page can be split */
2437 bool can_split_huge_page(struct page *page, int *pextra_pins)
2438 {
2439         int extra_pins;
2440
2441         /* Additional pins from radix tree */
2442         if (PageAnon(page))
2443                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2444         else
2445                 extra_pins = HPAGE_PMD_NR;
2446         if (pextra_pins)
2447                 *pextra_pins = extra_pins;
2448         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2449 }
2450
2451 /*
2452  * This function splits huge page into normal pages. @page can point to any
2453  * subpage of huge page to split. Split doesn't change the position of @page.
2454  *
2455  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2456  * The huge page must be locked.
2457  *
2458  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2459  *
2460  * Both head page and tail pages will inherit mapping, flags, and so on from
2461  * the hugepage.
2462  *
2463  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2464  * they are not mapped.
2465  *
2466  * Returns 0 if the hugepage is split successfully.
2467  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2468  * us.
2469  */
2470 int split_huge_page_to_list(struct page *page, struct list_head *list)
2471 {
2472         struct page *head = compound_head(page);
2473         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2474         struct anon_vma *anon_vma = NULL;
2475         struct address_space *mapping = NULL;
2476         int count, mapcount, extra_pins, ret;
2477         bool mlocked;
2478         unsigned long flags;
2479
2480         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2481         VM_BUG_ON_PAGE(!PageLocked(page), page);
2482         VM_BUG_ON_PAGE(!PageCompound(page), page);
2483
2484         if (PageWriteback(page))
2485                 return -EBUSY;
2486
2487         if (PageAnon(head)) {
2488                 /*
2489                  * The caller does not necessarily hold an mmap_sem that would
2490                  * prevent the anon_vma disappearing so we first we take a
2491                  * reference to it and then lock the anon_vma for write. This
2492                  * is similar to page_lock_anon_vma_read except the write lock
2493                  * is taken to serialise against parallel split or collapse
2494                  * operations.
2495                  */
2496                 anon_vma = page_get_anon_vma(head);
2497                 if (!anon_vma) {
2498                         ret = -EBUSY;
2499                         goto out;
2500                 }
2501                 mapping = NULL;
2502                 anon_vma_lock_write(anon_vma);
2503         } else {
2504                 mapping = head->mapping;
2505
2506                 /* Truncated ? */
2507                 if (!mapping) {
2508                         ret = -EBUSY;
2509                         goto out;
2510                 }
2511
2512                 anon_vma = NULL;
2513                 i_mmap_lock_read(mapping);
2514         }
2515
2516         /*
2517          * Racy check if we can split the page, before freeze_page() will
2518          * split PMDs
2519          */
2520         if (!can_split_huge_page(head, &extra_pins)) {
2521                 ret = -EBUSY;
2522                 goto out_unlock;
2523         }
2524
2525         mlocked = PageMlocked(page);
2526         freeze_page(head);
2527         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2528
2529         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2530         if (mlocked)
2531                 lru_add_drain();
2532
2533         /* prevent PageLRU to go away from under us, and freeze lru stats */
2534         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2535
2536         if (mapping) {
2537                 void **pslot;
2538
2539                 spin_lock(&mapping->tree_lock);
2540                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2541                                 page_index(head));
2542                 /*
2543                  * Check if the head page is present in radix tree.
2544                  * We assume all tail are present too, if head is there.
2545                  */
2546                 if (radix_tree_deref_slot_protected(pslot,
2547                                         &mapping->tree_lock) != head)
2548                         goto fail;
2549         }
2550
2551         /* Prevent deferred_split_scan() touching ->_refcount */
2552         spin_lock(&pgdata->split_queue_lock);
2553         count = page_count(head);
2554         mapcount = total_mapcount(head);
2555         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2556                 if (!list_empty(page_deferred_list(head))) {
2557                         pgdata->split_queue_len--;
2558                         list_del(page_deferred_list(head));
2559                 }
2560                 if (mapping)
2561                         __dec_node_page_state(page, NR_SHMEM_THPS);
2562                 spin_unlock(&pgdata->split_queue_lock);
2563                 __split_huge_page(page, list, flags);
2564                 if (PageSwapCache(head)) {
2565                         swp_entry_t entry = { .val = page_private(head) };
2566
2567                         ret = split_swap_cluster(entry);
2568                 } else
2569                         ret = 0;
2570         } else {
2571                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2572                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2573                                         mapcount, count);
2574                         if (PageTail(page))
2575                                 dump_page(head, NULL);
2576                         dump_page(page, "total_mapcount(head) > 0");
2577                         BUG();
2578                 }
2579                 spin_unlock(&pgdata->split_queue_lock);
2580 fail:           if (mapping)
2581                         spin_unlock(&mapping->tree_lock);
2582                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2583                 unfreeze_page(head);
2584                 ret = -EBUSY;
2585         }
2586
2587 out_unlock:
2588         if (anon_vma) {
2589                 anon_vma_unlock_write(anon_vma);
2590                 put_anon_vma(anon_vma);
2591         }
2592         if (mapping)
2593                 i_mmap_unlock_read(mapping);
2594 out:
2595         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2596         return ret;
2597 }
2598
2599 void free_transhuge_page(struct page *page)
2600 {
2601         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2602         unsigned long flags;
2603
2604         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2605         if (!list_empty(page_deferred_list(page))) {
2606                 pgdata->split_queue_len--;
2607                 list_del(page_deferred_list(page));
2608         }
2609         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2610         free_compound_page(page);
2611 }
2612
2613 void deferred_split_huge_page(struct page *page)
2614 {
2615         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2616         unsigned long flags;
2617
2618         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2619
2620         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2621         if (list_empty(page_deferred_list(page))) {
2622                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2623                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2624                 pgdata->split_queue_len++;
2625         }
2626         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2627 }
2628
2629 static unsigned long deferred_split_count(struct shrinker *shrink,
2630                 struct shrink_control *sc)
2631 {
2632         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2633         return ACCESS_ONCE(pgdata->split_queue_len);
2634 }
2635
2636 static unsigned long deferred_split_scan(struct shrinker *shrink,
2637                 struct shrink_control *sc)
2638 {
2639         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2640         unsigned long flags;
2641         LIST_HEAD(list), *pos, *next;
2642         struct page *page;
2643         int split = 0;
2644
2645         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2646         /* Take pin on all head pages to avoid freeing them under us */
2647         list_for_each_safe(pos, next, &pgdata->split_queue) {
2648                 page = list_entry((void *)pos, struct page, mapping);
2649                 page = compound_head(page);
2650                 if (get_page_unless_zero(page)) {
2651                         list_move(page_deferred_list(page), &list);
2652                 } else {
2653                         /* We lost race with put_compound_page() */
2654                         list_del_init(page_deferred_list(page));
2655                         pgdata->split_queue_len--;
2656                 }
2657                 if (!--sc->nr_to_scan)
2658                         break;
2659         }
2660         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2661
2662         list_for_each_safe(pos, next, &list) {
2663                 page = list_entry((void *)pos, struct page, mapping);
2664                 lock_page(page);
2665                 /* split_huge_page() removes page from list on success */
2666                 if (!split_huge_page(page))
2667                         split++;
2668                 unlock_page(page);
2669                 put_page(page);
2670         }
2671
2672         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2673         list_splice_tail(&list, &pgdata->split_queue);
2674         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2675
2676         /*
2677          * Stop shrinker if we didn't split any page, but the queue is empty.
2678          * This can happen if pages were freed under us.
2679          */
2680         if (!split && list_empty(&pgdata->split_queue))
2681                 return SHRINK_STOP;
2682         return split;
2683 }
2684
2685 static struct shrinker deferred_split_shrinker = {
2686         .count_objects = deferred_split_count,
2687         .scan_objects = deferred_split_scan,
2688         .seeks = DEFAULT_SEEKS,
2689         .flags = SHRINKER_NUMA_AWARE,
2690 };
2691
2692 #ifdef CONFIG_DEBUG_FS
2693 static int split_huge_pages_set(void *data, u64 val)
2694 {
2695         struct zone *zone;
2696         struct page *page;
2697         unsigned long pfn, max_zone_pfn;
2698         unsigned long total = 0, split = 0;
2699
2700         if (val != 1)
2701                 return -EINVAL;
2702
2703         for_each_populated_zone(zone) {
2704                 max_zone_pfn = zone_end_pfn(zone);
2705                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2706                         if (!pfn_valid(pfn))
2707                                 continue;
2708
2709                         page = pfn_to_page(pfn);
2710                         if (!get_page_unless_zero(page))
2711                                 continue;
2712
2713                         if (zone != page_zone(page))
2714                                 goto next;
2715
2716                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2717                                 goto next;
2718
2719                         total++;
2720                         lock_page(page);
2721                         if (!split_huge_page(page))
2722                                 split++;
2723                         unlock_page(page);
2724 next:
2725                         put_page(page);
2726                 }
2727         }
2728
2729         pr_info("%lu of %lu THP split\n", split, total);
2730
2731         return 0;
2732 }
2733 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2734                 "%llu\n");
2735
2736 static int __init split_huge_pages_debugfs(void)
2737 {
2738         void *ret;
2739
2740         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2741                         &split_huge_pages_fops);
2742         if (!ret)
2743                 pr_warn("Failed to create split_huge_pages in debugfs");
2744         return 0;
2745 }
2746 late_initcall(split_huge_pages_debugfs);
2747 #endif