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