mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 /*
  39  * By default transparent hugepage support is disabled in order that avoid
  40  * to risk increase the memory footprint of applications without a guaranteed
  41  * benefit. When transparent hugepage support is enabled, is for all mappings,
  42  * and khugepaged scans all mappings.
  43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  44  * for all hugepage allocations.
  45  */
  46 unsigned long transparent_hugepage_flags __read_mostly =
  47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  49 #endif
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  52 #endif
  53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  56
  57 static struct shrinker deferred_split_shrinker;
  58
  59 static atomic_t huge_zero_refcount;
  60 struct page *huge_zero_page __read_mostly;
  61
  62 static struct page *get_huge_zero_page(void)
  63 {
  64         struct page *zero_page;
  65 retry:
  66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  67                 return READ_ONCE(huge_zero_page);
  68
  69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  70                         HPAGE_PMD_ORDER);
  71         if (!zero_page) {
  72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  73                 return NULL;
  74         }
  75         count_vm_event(THP_ZERO_PAGE_ALLOC);
  76         preempt_disable();
  77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  78                 preempt_enable();
  79                 __free_pages(zero_page, compound_order(zero_page));
  80                 goto retry;
  81         }
  82
  83         /* We take additional reference here. It will be put back by shrinker */
  84         atomic_set(&huge_zero_refcount, 2);
  85         preempt_enable();
  86         return READ_ONCE(huge_zero_page);
  87 }
  88
  89 static void put_huge_zero_page(void)
  90 {
  91         /*
  92          * Counter should never go to zero here. Only shrinker can put
  93          * last reference.
  94          */
  95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  96 }
  97
  98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
  99 {
 100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 101                 return READ_ONCE(huge_zero_page);
 102
 103         if (!get_huge_zero_page())
 104                 return NULL;
 105
 106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 107                 put_huge_zero_page();
 108
 109         return READ_ONCE(huge_zero_page);
 110 }
 111
 112 void mm_put_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 put_huge_zero_page();
 116 }
 117
 118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 119                                         struct shrink_control *sc)
 120 {
 121         /* we can free zero page only if last reference remains */
 122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 123 }
 124
 125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 126                                        struct shrink_control *sc)
 127 {
 128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 130                 BUG_ON(zero_page == NULL);
 131                 __free_pages(zero_page, compound_order(zero_page));
 132                 return HPAGE_PMD_NR;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static struct shrinker huge_zero_page_shrinker = {
 139         .count_objects = shrink_huge_zero_page_count,
 140         .scan_objects = shrink_huge_zero_page_scan,
 141         .seeks = DEFAULT_SEEKS,
 142 };
 143
 144 #ifdef CONFIG_SYSFS
 145
 146 static ssize_t triple_flag_store(struct kobject *kobj,
 147                                  struct kobj_attribute *attr,
 148                                  const char *buf, size_t count,
 149                                  enum transparent_hugepage_flag enabled,
 150                                  enum transparent_hugepage_flag deferred,
 151                                  enum transparent_hugepage_flag req_madv)
 152 {
 153         if (!memcmp("defer", buf,
 154                     min(sizeof("defer")-1, count))) {
 155                 if (enabled == deferred)
 156                         return -EINVAL;
 157                 clear_bit(enabled, &transparent_hugepage_flags);
 158                 clear_bit(req_madv, &transparent_hugepage_flags);
 159                 set_bit(deferred, &transparent_hugepage_flags);
 160         } else if (!memcmp("always", buf,
 161                     min(sizeof("always")-1, count))) {
 162                 clear_bit(deferred, &transparent_hugepage_flags);
 163                 clear_bit(req_madv, &transparent_hugepage_flags);
 164                 set_bit(enabled, &transparent_hugepage_flags);
 165         } else if (!memcmp("madvise", buf,
 166                            min(sizeof("madvise")-1, count))) {
 167                 clear_bit(enabled, &transparent_hugepage_flags);
 168                 clear_bit(deferred, &transparent_hugepage_flags);
 169                 set_bit(req_madv, &transparent_hugepage_flags);
 170         } else if (!memcmp("never", buf,
 171                            min(sizeof("never")-1, count))) {
 172                 clear_bit(enabled, &transparent_hugepage_flags);
 173                 clear_bit(req_madv, &transparent_hugepage_flags);
 174                 clear_bit(deferred, &transparent_hugepage_flags);
 175         } else
 176                 return -EINVAL;
 177
 178         return count;
 179 }
 180
 181 static ssize_t enabled_show(struct kobject *kobj,
 182                             struct kobj_attribute *attr, char *buf)
 183 {
 184         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 185                 return sprintf(buf, "[always] madvise never\n");
 186         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 187                 return sprintf(buf, "always [madvise] never\n");
 188         else
 189                 return sprintf(buf, "always madvise [never]\n");
 190 }
 191
 192 static ssize_t enabled_store(struct kobject *kobj,
 193                              struct kobj_attribute *attr,
 194                              const char *buf, size_t count)
 195 {
 196         ssize_t ret;
 197
 198         ret = triple_flag_store(kobj, attr, buf, count,
 199                                 TRANSPARENT_HUGEPAGE_FLAG,
 200                                 TRANSPARENT_HUGEPAGE_FLAG,
 201                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 202
 203         if (ret > 0) {
 204                 int err = start_stop_khugepaged();
 205                 if (err)
 206                         ret = err;
 207         }
 208
 209         return ret;
 210 }
 211 static struct kobj_attribute enabled_attr =
 212         __ATTR(enabled, 0644, enabled_show, enabled_store);
 213
 214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 215                                 struct kobj_attribute *attr, char *buf,
 216                                 enum transparent_hugepage_flag flag)
 217 {
 218         return sprintf(buf, "%d\n",
 219                        !!test_bit(flag, &transparent_hugepage_flags));
 220 }
 221
 222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 223                                  struct kobj_attribute *attr,
 224                                  const char *buf, size_t count,
 225                                  enum transparent_hugepage_flag flag)
 226 {
 227         unsigned long value;
 228         int ret;
 229
 230         ret = kstrtoul(buf, 10, &value);
 231         if (ret < 0)
 232                 return ret;
 233         if (value > 1)
 234                 return -EINVAL;
 235
 236         if (value)
 237                 set_bit(flag, &transparent_hugepage_flags);
 238         else
 239                 clear_bit(flag, &transparent_hugepage_flags);
 240
 241         return count;
 242 }
 243
 244 /*
 245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 247  * memory just to allocate one more hugepage.
 248  */
 249 static ssize_t defrag_show(struct kobject *kobj,
 250                            struct kobj_attribute *attr, char *buf)
 251 {
 252         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 253                 return sprintf(buf, "[always] defer madvise never\n");
 254         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 255                 return sprintf(buf, "always [defer] madvise never\n");
 256         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 257                 return sprintf(buf, "always defer [madvise] never\n");
 258         else
 259                 return sprintf(buf, "always defer madvise [never]\n");
 260
 261 }
 262 static ssize_t defrag_store(struct kobject *kobj,
 263                             struct kobj_attribute *attr,
 264                             const char *buf, size_t count)
 265 {
 266         return triple_flag_store(kobj, attr, buf, count,
 267                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 268                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 269                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 270 }
 271 static struct kobj_attribute defrag_attr =
 272         __ATTR(defrag, 0644, defrag_show, defrag_store);
 273
 274 static ssize_t use_zero_page_show(struct kobject *kobj,
 275                 struct kobj_attribute *attr, char *buf)
 276 {
 277         return single_hugepage_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 279 }
 280 static ssize_t use_zero_page_store(struct kobject *kobj,
 281                 struct kobj_attribute *attr, const char *buf, size_t count)
 282 {
 283         return single_hugepage_flag_store(kobj, attr, buf, count,
 284                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 285 }
 286 static struct kobj_attribute use_zero_page_attr =
 287         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 288 #ifdef CONFIG_DEBUG_VM
 289 static ssize_t debug_cow_show(struct kobject *kobj,
 290                                 struct kobj_attribute *attr, char *buf)
 291 {
 292         return single_hugepage_flag_show(kobj, attr, buf,
 293                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 294 }
 295 static ssize_t debug_cow_store(struct kobject *kobj,
 296                                struct kobj_attribute *attr,
 297                                const char *buf, size_t count)
 298 {
 299         return single_hugepage_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 301 }
 302 static struct kobj_attribute debug_cow_attr =
 303         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 304 #endif /* CONFIG_DEBUG_VM */
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309         &use_zero_page_attr.attr,
 310 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 311         &shmem_enabled_attr.attr,
 312 #endif
 313 #ifdef CONFIG_DEBUG_VM
 314         &debug_cow_attr.attr,
 315 #endif
 316         NULL,
 317 };
 318
 319 static struct attribute_group hugepage_attr_group = {
 320         .attrs = hugepage_attr,
 321 };
 322
 323 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 324 {
 325         int err;
 326
 327         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 328         if (unlikely(!*hugepage_kobj)) {
 329                 pr_err("failed to create transparent hugepage kobject\n");
 330                 return -ENOMEM;
 331         }
 332
 333         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 334         if (err) {
 335                 pr_err("failed to register transparent hugepage group\n");
 336                 goto delete_obj;
 337         }
 338
 339         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 340         if (err) {
 341                 pr_err("failed to register transparent hugepage group\n");
 342                 goto remove_hp_group;
 343         }
 344
 345         return 0;
 346
 347 remove_hp_group:
 348         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 349 delete_obj:
 350         kobject_put(*hugepage_kobj);
 351         return err;
 352 }
 353
 354 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 355 {
 356         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 357         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 358         kobject_put(hugepage_kobj);
 359 }
 360 #else
 361 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 362 {
 363         return 0;
 364 }
 365
 366 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368 }
 369 #endif /* CONFIG_SYSFS */
 370
 371 static int __init hugepage_init(void)
 372 {
 373         int err;
 374         struct kobject *hugepage_kobj;
 375
 376         if (!has_transparent_hugepage()) {
 377                 transparent_hugepage_flags = 0;
 378                 return -EINVAL;
 379         }
 380
 381         /*
 382          * hugepages can't be allocated by the buddy allocator
 383          */
 384         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 385         /*
 386          * we use page->mapping and page->index in second tail page
 387          * as list_head: assuming THP order >= 2
 388          */
 389         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 390
 391         err = hugepage_init_sysfs(&hugepage_kobj);
 392         if (err)
 393                 goto err_sysfs;
 394
 395         err = khugepaged_init();
 396         if (err)
 397                 goto err_slab;
 398
 399         err = register_shrinker(&huge_zero_page_shrinker);
 400         if (err)
 401                 goto err_hzp_shrinker;
 402         err = register_shrinker(&deferred_split_shrinker);
 403         if (err)
 404                 goto err_split_shrinker;
 405
 406         /*
 407          * By default disable transparent hugepages on smaller systems,
 408          * where the extra memory used could hurt more than TLB overhead
 409          * is likely to save.  The admin can still enable it through /sys.
 410          */
 411         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 412                 transparent_hugepage_flags = 0;
 413                 return 0;
 414         }
 415
 416         err = start_stop_khugepaged();
 417         if (err)
 418                 goto err_khugepaged;
 419
 420         return 0;
 421 err_khugepaged:
 422         unregister_shrinker(&deferred_split_shrinker);
 423 err_split_shrinker:
 424         unregister_shrinker(&huge_zero_page_shrinker);
 425 err_hzp_shrinker:
 426         khugepaged_destroy();
 427 err_slab:
 428         hugepage_exit_sysfs(hugepage_kobj);
 429 err_sysfs:
 430         return err;
 431 }
 432 subsys_initcall(hugepage_init);
 433
 434 static int __init setup_transparent_hugepage(char *str)
 435 {
 436         int ret = 0;
 437         if (!str)
 438                 goto out;
 439         if (!strcmp(str, "always")) {
 440                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 441                         &transparent_hugepage_flags);
 442                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 443                           &transparent_hugepage_flags);
 444                 ret = 1;
 445         } else if (!strcmp(str, "madvise")) {
 446                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 447                           &transparent_hugepage_flags);
 448                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 449                         &transparent_hugepage_flags);
 450                 ret = 1;
 451         } else if (!strcmp(str, "never")) {
 452                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                           &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         }
 458 out:
 459         if (!ret)
 460                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 461         return ret;
 462 }
 463 __setup("transparent_hugepage=", setup_transparent_hugepage);
 464
 465 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 466 {
 467         if (likely(vma->vm_flags & VM_WRITE))
 468                 pmd = pmd_mkwrite(pmd);
 469         return pmd;
 470 }
 471
 472 static inline struct list_head *page_deferred_list(struct page *page)
 473 {
 474         /*
 475          * ->lru in the tail pages is occupied by compound_head.
 476          * Let's use ->mapping + ->index in the second tail page as list_head.
 477          */
 478         return (struct list_head *)&page[2].mapping;
 479 }
 480
 481 void prep_transhuge_page(struct page *page)
 482 {
 483         /*
 484          * we use page->mapping and page->indexlru in second tail page
 485          * as list_head: assuming THP order >= 2
 486          */
 487
 488         INIT_LIST_HEAD(page_deferred_list(page));
 489         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 490 }
 491
 492 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 493                 loff_t off, unsigned long flags, unsigned long size)
 494 {
 495         unsigned long addr;
 496         loff_t off_end = off + len;
 497         loff_t off_align = round_up(off, size);
 498         unsigned long len_pad;
 499
 500         if (off_end <= off_align || (off_end - off_align) < size)
 501                 return 0;
 502
 503         len_pad = len + size;
 504         if (len_pad < len || (off + len_pad) < off)
 505                 return 0;
 506
 507         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 508                                               off >> PAGE_SHIFT, flags);
 509         if (IS_ERR_VALUE(addr))
 510                 return 0;
 511
 512         addr += (off - addr) & (size - 1);
 513         return addr;
 514 }
 515
 516 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 517                 unsigned long len, unsigned long pgoff, unsigned long flags)
 518 {
 519         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 520
 521         if (addr)
 522                 goto out;
 523         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 524                 goto out;
 525
 526         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 527         if (addr)
 528                 return addr;
 529
 530  out:
 531         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 532 }
 533 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 534
 535 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 536                 gfp_t gfp)
 537 {
 538         struct vm_area_struct *vma = fe->vma;
 539         struct mem_cgroup *memcg;
 540         pgtable_t pgtable;
 541         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 542
 543         VM_BUG_ON_PAGE(!PageCompound(page), page);
 544
 545         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 546                 put_page(page);
 547                 count_vm_event(THP_FAULT_FALLBACK);
 548                 return VM_FAULT_FALLBACK;
 549         }
 550
 551         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 552         if (unlikely(!pgtable)) {
 553                 mem_cgroup_cancel_charge(page, memcg, true);
 554                 put_page(page);
 555                 return VM_FAULT_OOM;
 556         }
 557
 558         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 559         /*
 560          * The memory barrier inside __SetPageUptodate makes sure that
 561          * clear_huge_page writes become visible before the set_pmd_at()
 562          * write.
 563          */
 564         __SetPageUptodate(page);
 565
 566         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 567         if (unlikely(!pmd_none(*fe->pmd))) {
 568                 spin_unlock(fe->ptl);
 569                 mem_cgroup_cancel_charge(page, memcg, true);
 570                 put_page(page);
 571                 pte_free(vma->vm_mm, pgtable);
 572         } else {
 573                 pmd_t entry;
 574
 575                 /* Deliver the page fault to userland */
 576                 if (userfaultfd_missing(vma)) {
 577                         int ret;
 578
 579                         spin_unlock(fe->ptl);
 580                         mem_cgroup_cancel_charge(page, memcg, true);
 581                         put_page(page);
 582                         pte_free(vma->vm_mm, pgtable);
 583                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 584                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 585                         return ret;
 586                 }
 587
 588                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 589                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 590                 page_add_new_anon_rmap(page, vma, haddr, true);
 591                 mem_cgroup_commit_charge(page, memcg, false, true);
 592                 lru_cache_add_active_or_unevictable(page, vma);
 593                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 594                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 595                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 596                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 597                 spin_unlock(fe->ptl);
 598                 count_vm_event(THP_FAULT_ALLOC);
 599         }
 600
 601         return 0;
 602 }
 603
 604 /*
 605  * If THP defrag is set to always then directly reclaim/compact as necessary
 606  * If set to defer then do only background reclaim/compact and defer to khugepaged
 607  * If set to madvise and the VMA is flagged then directly reclaim/compact
 608  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
 609  */
 610 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 611 {
 612         bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 613
 614         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 615                                 &transparent_hugepage_flags) && vma_madvised)
 616                 return GFP_TRANSHUGE;
 617         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 618                                                 &transparent_hugepage_flags))
 619                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 620         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 621                                                 &transparent_hugepage_flags))
 622                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 623
 624         return GFP_TRANSHUGE_LIGHT;
 625 }
 626
 627 /* Caller must hold page table lock. */
 628 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 629                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 630                 struct page *zero_page)
 631 {
 632         pmd_t entry;
 633         if (!pmd_none(*pmd))
 634                 return false;
 635         entry = mk_pmd(zero_page, vma->vm_page_prot);
 636         entry = pmd_mkhuge(entry);
 637         if (pgtable)
 638                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 639         set_pmd_at(mm, haddr, pmd, entry);
 640         atomic_long_inc(&mm->nr_ptes);
 641         return true;
 642 }
 643
 644 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 645 {
 646         struct vm_area_struct *vma = fe->vma;
 647         gfp_t gfp;
 648         struct page *page;
 649         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 650
 651         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 652                 return VM_FAULT_FALLBACK;
 653         if (unlikely(anon_vma_prepare(vma)))
 654                 return VM_FAULT_OOM;
 655         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 656                 return VM_FAULT_OOM;
 657         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 658                         !mm_forbids_zeropage(vma->vm_mm) &&
 659                         transparent_hugepage_use_zero_page()) {
 660                 pgtable_t pgtable;
 661                 struct page *zero_page;
 662                 bool set;
 663                 int ret;
 664                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 665                 if (unlikely(!pgtable))
 666                         return VM_FAULT_OOM;
 667                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 668                 if (unlikely(!zero_page)) {
 669                         pte_free(vma->vm_mm, pgtable);
 670                         count_vm_event(THP_FAULT_FALLBACK);
 671                         return VM_FAULT_FALLBACK;
 672                 }
 673                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 674                 ret = 0;
 675                 set = false;
 676                 if (pmd_none(*fe->pmd)) {
 677                         if (userfaultfd_missing(vma)) {
 678                                 spin_unlock(fe->ptl);
 679                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 680                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 681                         } else {
 682                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 683                                                    haddr, fe->pmd, zero_page);
 684                                 spin_unlock(fe->ptl);
 685                                 set = true;
 686                         }
 687                 } else
 688                         spin_unlock(fe->ptl);
 689                 if (!set)
 690                         pte_free(vma->vm_mm, pgtable);
 691                 return ret;
 692         }
 693         gfp = alloc_hugepage_direct_gfpmask(vma);
 694         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 695         if (unlikely(!page)) {
 696                 count_vm_event(THP_FAULT_FALLBACK);
 697                 return VM_FAULT_FALLBACK;
 698         }
 699         prep_transhuge_page(page);
 700         return __do_huge_pmd_anonymous_page(fe, page, gfp);
 701 }
 702
 703 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 704                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 705 {
 706         struct mm_struct *mm = vma->vm_mm;
 707         pmd_t entry;
 708         spinlock_t *ptl;
 709
 710         ptl = pmd_lock(mm, pmd);
 711         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 712         if (pfn_t_devmap(pfn))
 713                 entry = pmd_mkdevmap(entry);
 714         if (write) {
 715                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 716                 entry = maybe_pmd_mkwrite(entry, vma);
 717         }
 718         set_pmd_at(mm, addr, pmd, entry);
 719         update_mmu_cache_pmd(vma, addr, pmd);
 720         spin_unlock(ptl);
 721 }
 722
 723 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 724                         pmd_t *pmd, pfn_t pfn, bool write)
 725 {
 726         pgprot_t pgprot = vma->vm_page_prot;
 727         /*
 728          * If we had pmd_special, we could avoid all these restrictions,
 729          * but we need to be consistent with PTEs and architectures that
 730          * can't support a 'special' bit.
 731          */
 732         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 733         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 734                                                 (VM_PFNMAP|VM_MIXEDMAP));
 735         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 736         BUG_ON(!pfn_t_devmap(pfn));
 737
 738         if (addr < vma->vm_start || addr >= vma->vm_end)
 739                 return VM_FAULT_SIGBUS;
 740         if (track_pfn_insert(vma, &pgprot, pfn))
 741                 return VM_FAULT_SIGBUS;
 742         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 743         return VM_FAULT_NOPAGE;
 744 }
 745 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 746
 747 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 748                 pmd_t *pmd)
 749 {
 750         pmd_t _pmd;
 751
 752         /*
 753          * We should set the dirty bit only for FOLL_WRITE but for now
 754          * the dirty bit in the pmd is meaningless.  And if the dirty
 755          * bit will become meaningful and we'll only set it with
 756          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 757          * set the young bit, instead of the current set_pmd_at.
 758          */
 759         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 760         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 761                                 pmd, _pmd,  1))
 762                 update_mmu_cache_pmd(vma, addr, pmd);
 763 }
 764
 765 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 766                 pmd_t *pmd, int flags)
 767 {
 768         unsigned long pfn = pmd_pfn(*pmd);
 769         struct mm_struct *mm = vma->vm_mm;
 770         struct dev_pagemap *pgmap;
 771         struct page *page;
 772
 773         assert_spin_locked(pmd_lockptr(mm, pmd));
 774
 775         /*
 776          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 777          * not be in this function with `flags & FOLL_COW` set.
 778          */
 779         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 780
 781         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 782                 return NULL;
 783
 784         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 785                 /* pass */;
 786         else
 787                 return NULL;
 788
 789         if (flags & FOLL_TOUCH)
 790                 touch_pmd(vma, addr, pmd);
 791
 792         /*
 793          * device mapped pages can only be returned if the
 794          * caller will manage the page reference count.
 795          */
 796         if (!(flags & FOLL_GET))
 797                 return ERR_PTR(-EEXIST);
 798
 799         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 800         pgmap = get_dev_pagemap(pfn, NULL);
 801         if (!pgmap)
 802                 return ERR_PTR(-EFAULT);
 803         page = pfn_to_page(pfn);
 804         get_page(page);
 805         put_dev_pagemap(pgmap);
 806
 807         return page;
 808 }
 809
 810 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 811                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 812                   struct vm_area_struct *vma)
 813 {
 814         spinlock_t *dst_ptl, *src_ptl;
 815         struct page *src_page;
 816         pmd_t pmd;
 817         pgtable_t pgtable = NULL;
 818         int ret = -ENOMEM;
 819
 820         /* Skip if can be re-fill on fault */
 821         if (!vma_is_anonymous(vma))
 822                 return 0;
 823
 824         pgtable = pte_alloc_one(dst_mm, addr);
 825         if (unlikely(!pgtable))
 826                 goto out;
 827
 828         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 829         src_ptl = pmd_lockptr(src_mm, src_pmd);
 830         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 831
 832         ret = -EAGAIN;
 833         pmd = *src_pmd;
 834         if (unlikely(!pmd_trans_huge(pmd))) {
 835                 pte_free(dst_mm, pgtable);
 836                 goto out_unlock;
 837         }
 838         /*
 839          * When page table lock is held, the huge zero pmd should not be
 840          * under splitting since we don't split the page itself, only pmd to
 841          * a page table.
 842          */
 843         if (is_huge_zero_pmd(pmd)) {
 844                 struct page *zero_page;
 845                 /*
 846                  * get_huge_zero_page() will never allocate a new page here,
 847                  * since we already have a zero page to copy. It just takes a
 848                  * reference.
 849                  */
 850                 zero_page = mm_get_huge_zero_page(dst_mm);
 851                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 852                                 zero_page);
 853                 ret = 0;
 854                 goto out_unlock;
 855         }
 856
 857         src_page = pmd_page(pmd);
 858         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 859         get_page(src_page);
 860         page_dup_rmap(src_page, true);
 861         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 862         atomic_long_inc(&dst_mm->nr_ptes);
 863         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 864
 865         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 866         pmd = pmd_mkold(pmd_wrprotect(pmd));
 867         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 868
 869         ret = 0;
 870 out_unlock:
 871         spin_unlock(src_ptl);
 872         spin_unlock(dst_ptl);
 873 out:
 874         return ret;
 875 }
 876
 877 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
 878 {
 879         pmd_t entry;
 880         unsigned long haddr;
 881         bool write = fe->flags & FAULT_FLAG_WRITE;
 882
 883         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
 884         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 885                 goto unlock;
 886
 887         entry = pmd_mkyoung(orig_pmd);
 888         if (write)
 889                 entry = pmd_mkdirty(entry);
 890         haddr = fe->address & HPAGE_PMD_MASK;
 891         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
 892                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
 893
 894 unlock:
 895         spin_unlock(fe->ptl);
 896 }
 897
 898 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
 899                 struct page *page)
 900 {
 901         struct vm_area_struct *vma = fe->vma;
 902         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 903         struct mem_cgroup *memcg;
 904         pgtable_t pgtable;
 905         pmd_t _pmd;
 906         int ret = 0, i;
 907         struct page **pages;
 908         unsigned long mmun_start;       /* For mmu_notifiers */
 909         unsigned long mmun_end;         /* For mmu_notifiers */
 910
 911         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 912                         GFP_KERNEL);
 913         if (unlikely(!pages)) {
 914                 ret |= VM_FAULT_OOM;
 915                 goto out;
 916         }
 917
 918         for (i = 0; i < HPAGE_PMD_NR; i++) {
 919                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 920                                                __GFP_OTHER_NODE, vma,
 921                                                fe->address, page_to_nid(page));
 922                 if (unlikely(!pages[i] ||
 923                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 924                                      GFP_KERNEL, &memcg, false))) {
 925                         if (pages[i])
 926                                 put_page(pages[i]);
 927                         while (--i >= 0) {
 928                                 memcg = (void *)page_private(pages[i]);
 929                                 set_page_private(pages[i], 0);
 930                                 mem_cgroup_cancel_charge(pages[i], memcg,
 931                                                 false);
 932                                 put_page(pages[i]);
 933                         }
 934                         kfree(pages);
 935                         ret |= VM_FAULT_OOM;
 936                         goto out;
 937                 }
 938                 set_page_private(pages[i], (unsigned long)memcg);
 939         }
 940
 941         for (i = 0; i < HPAGE_PMD_NR; i++) {
 942                 copy_user_highpage(pages[i], page + i,
 943                                    haddr + PAGE_SIZE * i, vma);
 944                 __SetPageUptodate(pages[i]);
 945                 cond_resched();
 946         }
 947
 948         mmun_start = haddr;
 949         mmun_end   = haddr + HPAGE_PMD_SIZE;
 950         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 951
 952         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 953         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 954                 goto out_free_pages;
 955         VM_BUG_ON_PAGE(!PageHead(page), page);
 956
 957         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
 958         /* leave pmd empty until pte is filled */
 959
 960         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
 961         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 962
 963         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 964                 pte_t entry;
 965                 entry = mk_pte(pages[i], vma->vm_page_prot);
 966                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 967                 memcg = (void *)page_private(pages[i]);
 968                 set_page_private(pages[i], 0);
 969                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
 970                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 971                 lru_cache_add_active_or_unevictable(pages[i], vma);
 972                 fe->pte = pte_offset_map(&_pmd, haddr);
 973                 VM_BUG_ON(!pte_none(*fe->pte));
 974                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
 975                 pte_unmap(fe->pte);
 976         }
 977         kfree(pages);
 978
 979         smp_wmb(); /* make pte visible before pmd */
 980         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
 981         page_remove_rmap(page, true);
 982         spin_unlock(fe->ptl);
 983
 984         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 985
 986         ret |= VM_FAULT_WRITE;
 987         put_page(page);
 988
 989 out:
 990         return ret;
 991
 992 out_free_pages:
 993         spin_unlock(fe->ptl);
 994         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 995         for (i = 0; i < HPAGE_PMD_NR; i++) {
 996                 memcg = (void *)page_private(pages[i]);
 997                 set_page_private(pages[i], 0);
 998                 mem_cgroup_cancel_charge(pages[i], memcg, false);
 999                 put_page(pages[i]);
1000         }
1001         kfree(pages);
1002         goto out;
1003 }
1004
1005 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1006 {
1007         struct vm_area_struct *vma = fe->vma;
1008         struct page *page = NULL, *new_page;
1009         struct mem_cgroup *memcg;
1010         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1011         unsigned long mmun_start;       /* For mmu_notifiers */
1012         unsigned long mmun_end;         /* For mmu_notifiers */
1013         gfp_t huge_gfp;                 /* for allocation and charge */
1014         int ret = 0;
1015
1016         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1017         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1018         if (is_huge_zero_pmd(orig_pmd))
1019                 goto alloc;
1020         spin_lock(fe->ptl);
1021         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1022                 goto out_unlock;
1023
1024         page = pmd_page(orig_pmd);
1025         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1026         /*
1027          * We can only reuse the page if nobody else maps the huge page or it's
1028          * part.
1029          */
1030         if (page_trans_huge_mapcount(page, NULL) == 1) {
1031                 pmd_t entry;
1032                 entry = pmd_mkyoung(orig_pmd);
1033                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1034                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1035                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1036                 ret |= VM_FAULT_WRITE;
1037                 goto out_unlock;
1038         }
1039         get_page(page);
1040         spin_unlock(fe->ptl);
1041 alloc:
1042         if (transparent_hugepage_enabled(vma) &&
1043             !transparent_hugepage_debug_cow()) {
1044                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1045                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1046         } else
1047                 new_page = NULL;
1048
1049         if (likely(new_page)) {
1050                 prep_transhuge_page(new_page);
1051         } else {
1052                 if (!page) {
1053                         split_huge_pmd(vma, fe->pmd, fe->address);
1054                         ret |= VM_FAULT_FALLBACK;
1055                 } else {
1056                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1057                         if (ret & VM_FAULT_OOM) {
1058                                 split_huge_pmd(vma, fe->pmd, fe->address);
1059                                 ret |= VM_FAULT_FALLBACK;
1060                         }
1061                         put_page(page);
1062                 }
1063                 count_vm_event(THP_FAULT_FALLBACK);
1064                 goto out;
1065         }
1066
1067         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1068                                         huge_gfp, &memcg, true))) {
1069                 put_page(new_page);
1070                 split_huge_pmd(vma, fe->pmd, fe->address);
1071                 if (page)
1072                         put_page(page);
1073                 ret |= VM_FAULT_FALLBACK;
1074                 count_vm_event(THP_FAULT_FALLBACK);
1075                 goto out;
1076         }
1077
1078         count_vm_event(THP_FAULT_ALLOC);
1079
1080         if (!page)
1081                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1082         else
1083                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1084         __SetPageUptodate(new_page);
1085
1086         mmun_start = haddr;
1087         mmun_end   = haddr + HPAGE_PMD_SIZE;
1088         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1089
1090         spin_lock(fe->ptl);
1091         if (page)
1092                 put_page(page);
1093         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1094                 spin_unlock(fe->ptl);
1095                 mem_cgroup_cancel_charge(new_page, memcg, true);
1096                 put_page(new_page);
1097                 goto out_mn;
1098         } else {
1099                 pmd_t entry;
1100                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1101                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1102                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1103                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1104                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1105                 lru_cache_add_active_or_unevictable(new_page, vma);
1106                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1107                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1108                 if (!page) {
1109                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1110                 } else {
1111                         VM_BUG_ON_PAGE(!PageHead(page), page);
1112                         page_remove_rmap(page, true);
1113                         put_page(page);
1114                 }
1115                 ret |= VM_FAULT_WRITE;
1116         }
1117         spin_unlock(fe->ptl);
1118 out_mn:
1119         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1120 out:
1121         return ret;
1122 out_unlock:
1123         spin_unlock(fe->ptl);
1124         return ret;
1125 }
1126
1127 /*
1128  * FOLL_FORCE can write to even unwritable pmd's, but only
1129  * after we've gone through a COW cycle and they are dirty.
1130  */
1131 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1132 {
1133         return pmd_write(pmd) ||
1134                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1135 }
1136
1137 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1138                                    unsigned long addr,
1139                                    pmd_t *pmd,
1140                                    unsigned int flags)
1141 {
1142         struct mm_struct *mm = vma->vm_mm;
1143         struct page *page = NULL;
1144
1145         assert_spin_locked(pmd_lockptr(mm, pmd));
1146
1147         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1148                 goto out;
1149
1150         /* Avoid dumping huge zero page */
1151         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1152                 return ERR_PTR(-EFAULT);
1153
1154         /* Full NUMA hinting faults to serialise migration in fault paths */
1155         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1156                 goto out;
1157
1158         page = pmd_page(*pmd);
1159         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1160         if (flags & FOLL_TOUCH)
1161                 touch_pmd(vma, addr, pmd);
1162         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1163                 /*
1164                  * We don't mlock() pte-mapped THPs. This way we can avoid
1165                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1166                  *
1167                  * For anon THP:
1168                  *
1169                  * In most cases the pmd is the only mapping of the page as we
1170                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1171                  * writable private mappings in populate_vma_page_range().
1172                  *
1173                  * The only scenario when we have the page shared here is if we
1174                  * mlocking read-only mapping shared over fork(). We skip
1175                  * mlocking such pages.
1176                  *
1177                  * For file THP:
1178                  *
1179                  * We can expect PageDoubleMap() to be stable under page lock:
1180                  * for file pages we set it in page_add_file_rmap(), which
1181                  * requires page to be locked.
1182                  */
1183
1184                 if (PageAnon(page) && compound_mapcount(page) != 1)
1185                         goto skip_mlock;
1186                 if (PageDoubleMap(page) || !page->mapping)
1187                         goto skip_mlock;
1188                 if (!trylock_page(page))
1189                         goto skip_mlock;
1190                 lru_add_drain();
1191                 if (page->mapping && !PageDoubleMap(page))
1192                         mlock_vma_page(page);
1193                 unlock_page(page);
1194         }
1195 skip_mlock:
1196         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1197         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1198         if (flags & FOLL_GET)
1199                 get_page(page);
1200
1201 out:
1202         return page;
1203 }
1204
1205 /* NUMA hinting page fault entry point for trans huge pmds */
1206 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1207 {
1208         struct vm_area_struct *vma = fe->vma;
1209         struct anon_vma *anon_vma = NULL;
1210         struct page *page;
1211         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1212         int page_nid = -1, this_nid = numa_node_id();
1213         int target_nid, last_cpupid = -1;
1214         bool page_locked;
1215         bool migrated = false;
1216         bool was_writable;
1217         int flags = 0;
1218
1219         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1220         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1221                 goto out_unlock;
1222
1223         /*
1224          * If there are potential migrations, wait for completion and retry
1225          * without disrupting NUMA hinting information. Do not relock and
1226          * check_same as the page may no longer be mapped.
1227          */
1228         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1229                 page = pmd_page(*fe->pmd);
1230                 spin_unlock(fe->ptl);
1231                 wait_on_page_locked(page);
1232                 goto out;
1233         }
1234
1235         page = pmd_page(pmd);
1236         BUG_ON(is_huge_zero_page(page));
1237         page_nid = page_to_nid(page);
1238         last_cpupid = page_cpupid_last(page);
1239         count_vm_numa_event(NUMA_HINT_FAULTS);
1240         if (page_nid == this_nid) {
1241                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1242                 flags |= TNF_FAULT_LOCAL;
1243         }
1244
1245         /* See similar comment in do_numa_page for explanation */
1246         if (!pmd_write(pmd))
1247                 flags |= TNF_NO_GROUP;
1248
1249         /*
1250          * Acquire the page lock to serialise THP migrations but avoid dropping
1251          * page_table_lock if at all possible
1252          */
1253         page_locked = trylock_page(page);
1254         target_nid = mpol_misplaced(page, vma, haddr);
1255         if (target_nid == -1) {
1256                 /* If the page was locked, there are no parallel migrations */
1257                 if (page_locked)
1258                         goto clear_pmdnuma;
1259         }
1260
1261         /* Migration could have started since the pmd_trans_migrating check */
1262         if (!page_locked) {
1263                 spin_unlock(fe->ptl);
1264                 wait_on_page_locked(page);
1265                 page_nid = -1;
1266                 goto out;
1267         }
1268
1269         /*
1270          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1271          * to serialises splits
1272          */
1273         get_page(page);
1274         spin_unlock(fe->ptl);
1275         anon_vma = page_lock_anon_vma_read(page);
1276
1277         /* Confirm the PMD did not change while page_table_lock was released */
1278         spin_lock(fe->ptl);
1279         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1280                 unlock_page(page);
1281                 put_page(page);
1282                 page_nid = -1;
1283                 goto out_unlock;
1284         }
1285
1286         /* Bail if we fail to protect against THP splits for any reason */
1287         if (unlikely(!anon_vma)) {
1288                 put_page(page);
1289                 page_nid = -1;
1290                 goto clear_pmdnuma;
1291         }
1292
1293         /*
1294          * Migrate the THP to the requested node, returns with page unlocked
1295          * and access rights restored.
1296          */
1297         spin_unlock(fe->ptl);
1298         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1299                                 fe->pmd, pmd, fe->address, page, target_nid);
1300         if (migrated) {
1301                 flags |= TNF_MIGRATED;
1302                 page_nid = target_nid;
1303         } else
1304                 flags |= TNF_MIGRATE_FAIL;
1305
1306         goto out;
1307 clear_pmdnuma:
1308         BUG_ON(!PageLocked(page));
1309         was_writable = pmd_write(pmd);
1310         pmd = pmd_modify(pmd, vma->vm_page_prot);
1311         pmd = pmd_mkyoung(pmd);
1312         if (was_writable)
1313                 pmd = pmd_mkwrite(pmd);
1314         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1315         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1316         unlock_page(page);
1317 out_unlock:
1318         spin_unlock(fe->ptl);
1319
1320 out:
1321         if (anon_vma)
1322                 page_unlock_anon_vma_read(anon_vma);
1323
1324         if (page_nid != -1)
1325                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1326
1327         return 0;
1328 }
1329
1330 /*
1331  * Return true if we do MADV_FREE successfully on entire pmd page.
1332  * Otherwise, return false.
1333  */
1334 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1335                 pmd_t *pmd, unsigned long addr, unsigned long next)
1336 {
1337         spinlock_t *ptl;
1338         pmd_t orig_pmd;
1339         struct page *page;
1340         struct mm_struct *mm = tlb->mm;
1341         bool ret = false;
1342
1343         ptl = pmd_trans_huge_lock(pmd, vma);
1344         if (!ptl)
1345                 goto out_unlocked;
1346
1347         orig_pmd = *pmd;
1348         if (is_huge_zero_pmd(orig_pmd))
1349                 goto out;
1350
1351         page = pmd_page(orig_pmd);
1352         /*
1353          * If other processes are mapping this page, we couldn't discard
1354          * the page unless they all do MADV_FREE so let's skip the page.
1355          */
1356         if (page_mapcount(page) != 1)
1357                 goto out;
1358
1359         if (!trylock_page(page))
1360                 goto out;
1361
1362         /*
1363          * If user want to discard part-pages of THP, split it so MADV_FREE
1364          * will deactivate only them.
1365          */
1366         if (next - addr != HPAGE_PMD_SIZE) {
1367                 get_page(page);
1368                 spin_unlock(ptl);
1369                 split_huge_page(page);
1370                 put_page(page);
1371                 unlock_page(page);
1372                 goto out_unlocked;
1373         }
1374
1375         if (PageDirty(page))
1376                 ClearPageDirty(page);
1377         unlock_page(page);
1378
1379         if (PageActive(page))
1380                 deactivate_page(page);
1381
1382         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1383                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1384                         tlb->fullmm);
1385                 orig_pmd = pmd_mkold(orig_pmd);
1386                 orig_pmd = pmd_mkclean(orig_pmd);
1387
1388                 set_pmd_at(mm, addr, pmd, orig_pmd);
1389                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1390         }
1391         ret = true;
1392 out:
1393         spin_unlock(ptl);
1394 out_unlocked:
1395         return ret;
1396 }
1397
1398 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1399                  pmd_t *pmd, unsigned long addr)
1400 {
1401         pmd_t orig_pmd;
1402         spinlock_t *ptl;
1403
1404         ptl = __pmd_trans_huge_lock(pmd, vma);
1405         if (!ptl)
1406                 return 0;
1407         /*
1408          * For architectures like ppc64 we look at deposited pgtable
1409          * when calling pmdp_huge_get_and_clear. So do the
1410          * pgtable_trans_huge_withdraw after finishing pmdp related
1411          * operations.
1412          */
1413         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1414                         tlb->fullmm);
1415         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1416         if (vma_is_dax(vma)) {
1417                 spin_unlock(ptl);
1418                 if (is_huge_zero_pmd(orig_pmd))
1419                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1420         } else if (is_huge_zero_pmd(orig_pmd)) {
1421                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1422                 atomic_long_dec(&tlb->mm->nr_ptes);
1423                 spin_unlock(ptl);
1424                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1425         } else {
1426                 struct page *page = pmd_page(orig_pmd);
1427                 page_remove_rmap(page, true);
1428                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1429                 VM_BUG_ON_PAGE(!PageHead(page), page);
1430                 if (PageAnon(page)) {
1431                         pgtable_t pgtable;
1432                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1433                         pte_free(tlb->mm, pgtable);
1434                         atomic_long_dec(&tlb->mm->nr_ptes);
1435                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1436                 } else {
1437                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1438                 }
1439                 spin_unlock(ptl);
1440                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1441         }
1442         return 1;
1443 }
1444
1445 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1446                   unsigned long new_addr, unsigned long old_end,
1447                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1448 {
1449         spinlock_t *old_ptl, *new_ptl;
1450         pmd_t pmd;
1451         struct mm_struct *mm = vma->vm_mm;
1452         bool force_flush = false;
1453
1454         if ((old_addr & ~HPAGE_PMD_MASK) ||
1455             (new_addr & ~HPAGE_PMD_MASK) ||
1456             old_end - old_addr < HPAGE_PMD_SIZE)
1457                 return false;
1458
1459         /*
1460          * The destination pmd shouldn't be established, free_pgtables()
1461          * should have release it.
1462          */
1463         if (WARN_ON(!pmd_none(*new_pmd))) {
1464                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1465                 return false;
1466         }
1467
1468         /*
1469          * We don't have to worry about the ordering of src and dst
1470          * ptlocks because exclusive mmap_sem prevents deadlock.
1471          */
1472         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1473         if (old_ptl) {
1474                 new_ptl = pmd_lockptr(mm, new_pmd);
1475                 if (new_ptl != old_ptl)
1476                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1477                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1478                 if (pmd_present(pmd) && pmd_dirty(pmd))
1479                         force_flush = true;
1480                 VM_BUG_ON(!pmd_none(*new_pmd));
1481
1482                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1483                                 vma_is_anonymous(vma)) {
1484                         pgtable_t pgtable;
1485                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1486                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1487                 }
1488                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1489                 if (new_ptl != old_ptl)
1490                         spin_unlock(new_ptl);
1491                 if (force_flush)
1492                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1493                 else
1494                         *need_flush = true;
1495                 spin_unlock(old_ptl);
1496                 return true;
1497         }
1498         return false;
1499 }
1500
1501 /*
1502  * Returns
1503  *  - 0 if PMD could not be locked
1504  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1505  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1506  */
1507 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1508                 unsigned long addr, pgprot_t newprot, int prot_numa)
1509 {
1510         struct mm_struct *mm = vma->vm_mm;
1511         spinlock_t *ptl;
1512         int ret = 0;
1513
1514         ptl = __pmd_trans_huge_lock(pmd, vma);
1515         if (ptl) {
1516                 pmd_t entry;
1517                 bool preserve_write = prot_numa && pmd_write(*pmd);
1518                 ret = 1;
1519
1520                 /*
1521                  * Avoid trapping faults against the zero page. The read-only
1522                  * data is likely to be read-cached on the local CPU and
1523                  * local/remote hits to the zero page are not interesting.
1524                  */
1525                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1526                         spin_unlock(ptl);
1527                         return ret;
1528                 }
1529
1530                 if (!prot_numa || !pmd_protnone(*pmd)) {
1531                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1532                         entry = pmd_modify(entry, newprot);
1533                         if (preserve_write)
1534                                 entry = pmd_mkwrite(entry);
1535                         ret = HPAGE_PMD_NR;
1536                         set_pmd_at(mm, addr, pmd, entry);
1537                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1538                                         pmd_write(entry));
1539                 }
1540                 spin_unlock(ptl);
1541         }
1542
1543         return ret;
1544 }
1545
1546 /*
1547  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1548  *
1549  * Note that if it returns page table lock pointer, this routine returns without
1550  * unlocking page table lock. So callers must unlock it.
1551  */
1552 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1553 {
1554         spinlock_t *ptl;
1555         ptl = pmd_lock(vma->vm_mm, pmd);
1556         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1557                 return ptl;
1558         spin_unlock(ptl);
1559         return NULL;
1560 }
1561
1562 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1563                 unsigned long haddr, pmd_t *pmd)
1564 {
1565         struct mm_struct *mm = vma->vm_mm;
1566         pgtable_t pgtable;
1567         pmd_t _pmd;
1568         int i;
1569
1570         /* leave pmd empty until pte is filled */
1571         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1572
1573         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1574         pmd_populate(mm, &_pmd, pgtable);
1575
1576         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1577                 pte_t *pte, entry;
1578                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1579                 entry = pte_mkspecial(entry);
1580                 pte = pte_offset_map(&_pmd, haddr);
1581                 VM_BUG_ON(!pte_none(*pte));
1582                 set_pte_at(mm, haddr, pte, entry);
1583                 pte_unmap(pte);
1584         }
1585         smp_wmb(); /* make pte visible before pmd */
1586         pmd_populate(mm, pmd, pgtable);
1587 }
1588
1589 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1590                 unsigned long haddr, bool freeze)
1591 {
1592         struct mm_struct *mm = vma->vm_mm;
1593         struct page *page;
1594         pgtable_t pgtable;
1595         pmd_t _pmd;
1596         bool young, write, dirty, soft_dirty;
1597         unsigned long addr;
1598         int i;
1599
1600         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1601         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1602         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1603         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1604
1605         count_vm_event(THP_SPLIT_PMD);
1606
1607         if (!vma_is_anonymous(vma)) {
1608                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1609                 if (vma_is_dax(vma))
1610                         return;
1611                 page = pmd_page(_pmd);
1612                 if (!PageReferenced(page) && pmd_young(_pmd))
1613                         SetPageReferenced(page);
1614                 page_remove_rmap(page, true);
1615                 put_page(page);
1616                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1617                 return;
1618         } else if (is_huge_zero_pmd(*pmd)) {
1619                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1620         }
1621
1622         page = pmd_page(*pmd);
1623         VM_BUG_ON_PAGE(!page_count(page), page);
1624         page_ref_add(page, HPAGE_PMD_NR - 1);
1625         write = pmd_write(*pmd);
1626         young = pmd_young(*pmd);
1627         dirty = pmd_dirty(*pmd);
1628         soft_dirty = pmd_soft_dirty(*pmd);
1629
1630         pmdp_huge_split_prepare(vma, haddr, pmd);
1631         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1632         pmd_populate(mm, &_pmd, pgtable);
1633
1634         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1635                 pte_t entry, *pte;
1636                 /*
1637                  * Note that NUMA hinting access restrictions are not
1638                  * transferred to avoid any possibility of altering
1639                  * permissions across VMAs.
1640                  */
1641                 if (freeze) {
1642                         swp_entry_t swp_entry;
1643                         swp_entry = make_migration_entry(page + i, write);
1644                         entry = swp_entry_to_pte(swp_entry);
1645                         if (soft_dirty)
1646                                 entry = pte_swp_mksoft_dirty(entry);
1647                 } else {
1648                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1649                         entry = maybe_mkwrite(entry, vma);
1650                         if (!write)
1651                                 entry = pte_wrprotect(entry);
1652                         if (!young)
1653                                 entry = pte_mkold(entry);
1654                         if (soft_dirty)
1655                                 entry = pte_mksoft_dirty(entry);
1656                 }
1657                 if (dirty)
1658                         SetPageDirty(page + i);
1659                 pte = pte_offset_map(&_pmd, addr);
1660                 BUG_ON(!pte_none(*pte));
1661                 set_pte_at(mm, addr, pte, entry);
1662                 atomic_inc(&page[i]._mapcount);
1663                 pte_unmap(pte);
1664         }
1665
1666         /*
1667          * Set PG_double_map before dropping compound_mapcount to avoid
1668          * false-negative page_mapped().
1669          */
1670         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1671                 for (i = 0; i < HPAGE_PMD_NR; i++)
1672                         atomic_inc(&page[i]._mapcount);
1673         }
1674
1675         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1676                 /* Last compound_mapcount is gone. */
1677                 __dec_node_page_state(page, NR_ANON_THPS);
1678                 if (TestClearPageDoubleMap(page)) {
1679                         /* No need in mapcount reference anymore */
1680                         for (i = 0; i < HPAGE_PMD_NR; i++)
1681                                 atomic_dec(&page[i]._mapcount);
1682                 }
1683         }
1684
1685         smp_wmb(); /* make pte visible before pmd */
1686         /*
1687          * Up to this point the pmd is present and huge and userland has the
1688          * whole access to the hugepage during the split (which happens in
1689          * place). If we overwrite the pmd with the not-huge version pointing
1690          * to the pte here (which of course we could if all CPUs were bug
1691          * free), userland could trigger a small page size TLB miss on the
1692          * small sized TLB while the hugepage TLB entry is still established in
1693          * the huge TLB. Some CPU doesn't like that.
1694          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1695          * 383 on page 93. Intel should be safe but is also warns that it's
1696          * only safe if the permission and cache attributes of the two entries
1697          * loaded in the two TLB is identical (which should be the case here).
1698          * But it is generally safer to never allow small and huge TLB entries
1699          * for the same virtual address to be loaded simultaneously. So instead
1700          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1701          * current pmd notpresent (atomically because here the pmd_trans_huge
1702          * and pmd_trans_splitting must remain set at all times on the pmd
1703          * until the split is complete for this pmd), then we flush the SMP TLB
1704          * and finally we write the non-huge version of the pmd entry with
1705          * pmd_populate.
1706          */
1707         pmdp_invalidate(vma, haddr, pmd);
1708         pmd_populate(mm, pmd, pgtable);
1709
1710         if (freeze) {
1711                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1712                         page_remove_rmap(page + i, false);
1713                         put_page(page + i);
1714                 }
1715         }
1716 }
1717
1718 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1719                 unsigned long address, bool freeze, struct page *page)
1720 {
1721         spinlock_t *ptl;
1722         struct mm_struct *mm = vma->vm_mm;
1723         unsigned long haddr = address & HPAGE_PMD_MASK;
1724
1725         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1726         ptl = pmd_lock(mm, pmd);
1727
1728         /*
1729          * If caller asks to setup a migration entries, we need a page to check
1730          * pmd against. Otherwise we can end up replacing wrong page.
1731          */
1732         VM_BUG_ON(freeze && !page);
1733         if (page && page != pmd_page(*pmd))
1734                 goto out;
1735
1736         if (pmd_trans_huge(*pmd)) {
1737                 page = pmd_page(*pmd);
1738                 if (PageMlocked(page))
1739                         clear_page_mlock(page);
1740         } else if (!pmd_devmap(*pmd))
1741                 goto out;
1742         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1743 out:
1744         spin_unlock(ptl);
1745         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1746 }
1747
1748 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1749                 bool freeze, struct page *page)
1750 {
1751         pgd_t *pgd;
1752         pud_t *pud;
1753         pmd_t *pmd;
1754
1755         pgd = pgd_offset(vma->vm_mm, address);
1756         if (!pgd_present(*pgd))
1757                 return;
1758
1759         pud = pud_offset(pgd, address);
1760         if (!pud_present(*pud))
1761                 return;
1762
1763         pmd = pmd_offset(pud, address);
1764
1765         __split_huge_pmd(vma, pmd, address, freeze, page);
1766 }
1767
1768 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1769                              unsigned long start,
1770                              unsigned long end,
1771                              long adjust_next)
1772 {
1773         /*
1774          * If the new start address isn't hpage aligned and it could
1775          * previously contain an hugepage: check if we need to split
1776          * an huge pmd.
1777          */
1778         if (start & ~HPAGE_PMD_MASK &&
1779             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1780             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1781                 split_huge_pmd_address(vma, start, false, NULL);
1782
1783         /*
1784          * If the new end address isn't hpage aligned and it could
1785          * previously contain an hugepage: check if we need to split
1786          * an huge pmd.
1787          */
1788         if (end & ~HPAGE_PMD_MASK &&
1789             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1790             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1791                 split_huge_pmd_address(vma, end, false, NULL);
1792
1793         /*
1794          * If we're also updating the vma->vm_next->vm_start, if the new
1795          * vm_next->vm_start isn't page aligned and it could previously
1796          * contain an hugepage: check if we need to split an huge pmd.
1797          */
1798         if (adjust_next > 0) {
1799                 struct vm_area_struct *next = vma->vm_next;
1800                 unsigned long nstart = next->vm_start;
1801                 nstart += adjust_next << PAGE_SHIFT;
1802                 if (nstart & ~HPAGE_PMD_MASK &&
1803                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1804                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1805                         split_huge_pmd_address(next, nstart, false, NULL);
1806         }
1807 }
1808
1809 static void freeze_page(struct page *page)
1810 {
1811         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1812                 TTU_RMAP_LOCKED;
1813         int i, ret;
1814
1815         VM_BUG_ON_PAGE(!PageHead(page), page);
1816
1817         if (PageAnon(page))
1818                 ttu_flags |= TTU_MIGRATION;
1819
1820         /* We only need TTU_SPLIT_HUGE_PMD once */
1821         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1822         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1823                 /* Cut short if the page is unmapped */
1824                 if (page_count(page) == 1)
1825                         return;
1826
1827                 ret = try_to_unmap(page + i, ttu_flags);
1828         }
1829         VM_BUG_ON_PAGE(ret, page + i - 1);
1830 }
1831
1832 static void unfreeze_page(struct page *page)
1833 {
1834         int i;
1835
1836         for (i = 0; i < HPAGE_PMD_NR; i++)
1837                 remove_migration_ptes(page + i, page + i, true);
1838 }
1839
1840 static void __split_huge_page_tail(struct page *head, int tail,
1841                 struct lruvec *lruvec, struct list_head *list)
1842 {
1843         struct page *page_tail = head + tail;
1844
1845         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1846         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1847
1848         /*
1849          * tail_page->_refcount is zero and not changing from under us. But
1850          * get_page_unless_zero() may be running from under us on the
1851          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1852          * atomic_add(), we would then run atomic_set() concurrently with
1853          * get_page_unless_zero(), and atomic_set() is implemented in C not
1854          * using locked ops. spin_unlock on x86 sometime uses locked ops
1855          * because of PPro errata 66, 92, so unless somebody can guarantee
1856          * atomic_set() here would be safe on all archs (and not only on x86),
1857          * it's safer to use atomic_inc()/atomic_add().
1858          */
1859         if (PageAnon(head)) {
1860                 page_ref_inc(page_tail);
1861         } else {
1862                 /* Additional pin to radix tree */
1863                 page_ref_add(page_tail, 2);
1864         }
1865
1866         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1867         page_tail->flags |= (head->flags &
1868                         ((1L << PG_referenced) |
1869                          (1L << PG_swapbacked) |
1870                          (1L << PG_mlocked) |
1871                          (1L << PG_uptodate) |
1872                          (1L << PG_active) |
1873                          (1L << PG_locked) |
1874                          (1L << PG_unevictable) |
1875                          (1L << PG_dirty)));
1876
1877         /*
1878          * After clearing PageTail the gup refcount can be released.
1879          * Page flags also must be visible before we make the page non-compound.
1880          */
1881         smp_wmb();
1882
1883         clear_compound_head(page_tail);
1884
1885         if (page_is_young(head))
1886                 set_page_young(page_tail);
1887         if (page_is_idle(head))
1888                 set_page_idle(page_tail);
1889
1890         /* ->mapping in first tail page is compound_mapcount */
1891         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1892                         page_tail);
1893         page_tail->mapping = head->mapping;
1894
1895         page_tail->index = head->index + tail;
1896         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1897         lru_add_page_tail(head, page_tail, lruvec, list);
1898 }
1899
1900 static void __split_huge_page(struct page *page, struct list_head *list,
1901                 unsigned long flags)
1902 {
1903         struct page *head = compound_head(page);
1904         struct zone *zone = page_zone(head);
1905         struct lruvec *lruvec;
1906         pgoff_t end = -1;
1907         int i;
1908
1909         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1910
1911         /* complete memcg works before add pages to LRU */
1912         mem_cgroup_split_huge_fixup(head);
1913
1914         if (!PageAnon(page))
1915                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1916
1917         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1918                 __split_huge_page_tail(head, i, lruvec, list);
1919                 /* Some pages can be beyond i_size: drop them from page cache */
1920                 if (head[i].index >= end) {
1921                         __ClearPageDirty(head + i);
1922                         __delete_from_page_cache(head + i, NULL);
1923                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1924                                 shmem_uncharge(head->mapping->host, 1);
1925                         put_page(head + i);
1926                 }
1927         }
1928
1929         ClearPageCompound(head);
1930         /* See comment in __split_huge_page_tail() */
1931         if (PageAnon(head)) {
1932                 page_ref_inc(head);
1933         } else {
1934                 /* Additional pin to radix tree */
1935                 page_ref_add(head, 2);
1936                 spin_unlock(&head->mapping->tree_lock);
1937         }
1938
1939         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1940
1941         unfreeze_page(head);
1942
1943         for (i = 0; i < HPAGE_PMD_NR; i++) {
1944                 struct page *subpage = head + i;
1945                 if (subpage == page)
1946                         continue;
1947                 unlock_page(subpage);
1948
1949                 /*
1950                  * Subpages may be freed if there wasn't any mapping
1951                  * like if add_to_swap() is running on a lru page that
1952                  * had its mapping zapped. And freeing these pages
1953                  * requires taking the lru_lock so we do the put_page
1954                  * of the tail pages after the split is complete.
1955                  */
1956                 put_page(subpage);
1957         }
1958 }
1959
1960 int total_mapcount(struct page *page)
1961 {
1962         int i, compound, ret;
1963
1964         VM_BUG_ON_PAGE(PageTail(page), page);
1965
1966         if (likely(!PageCompound(page)))
1967                 return atomic_read(&page->_mapcount) + 1;
1968
1969         compound = compound_mapcount(page);
1970         if (PageHuge(page))
1971                 return compound;
1972         ret = compound;
1973         for (i = 0; i < HPAGE_PMD_NR; i++)
1974                 ret += atomic_read(&page[i]._mapcount) + 1;
1975         /* File pages has compound_mapcount included in _mapcount */
1976         if (!PageAnon(page))
1977                 return ret - compound * HPAGE_PMD_NR;
1978         if (PageDoubleMap(page))
1979                 ret -= HPAGE_PMD_NR;
1980         return ret;
1981 }
1982
1983 /*
1984  * This calculates accurately how many mappings a transparent hugepage
1985  * has (unlike page_mapcount() which isn't fully accurate). This full
1986  * accuracy is primarily needed to know if copy-on-write faults can
1987  * reuse the page and change the mapping to read-write instead of
1988  * copying them. At the same time this returns the total_mapcount too.
1989  *
1990  * The function returns the highest mapcount any one of the subpages
1991  * has. If the return value is one, even if different processes are
1992  * mapping different subpages of the transparent hugepage, they can
1993  * all reuse it, because each process is reusing a different subpage.
1994  *
1995  * The total_mapcount is instead counting all virtual mappings of the
1996  * subpages. If the total_mapcount is equal to "one", it tells the
1997  * caller all mappings belong to the same "mm" and in turn the
1998  * anon_vma of the transparent hugepage can become the vma->anon_vma
1999  * local one as no other process may be mapping any of the subpages.
2000  *
2001  * It would be more accurate to replace page_mapcount() with
2002  * page_trans_huge_mapcount(), however we only use
2003  * page_trans_huge_mapcount() in the copy-on-write faults where we
2004  * need full accuracy to avoid breaking page pinning, because
2005  * page_trans_huge_mapcount() is slower than page_mapcount().
2006  */
2007 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2008 {
2009         int i, ret, _total_mapcount, mapcount;
2010
2011         /* hugetlbfs shouldn't call it */
2012         VM_BUG_ON_PAGE(PageHuge(page), page);
2013
2014         if (likely(!PageTransCompound(page))) {
2015                 mapcount = atomic_read(&page->_mapcount) + 1;
2016                 if (total_mapcount)
2017                         *total_mapcount = mapcount;
2018                 return mapcount;
2019         }
2020
2021         page = compound_head(page);
2022
2023         _total_mapcount = ret = 0;
2024         for (i = 0; i < HPAGE_PMD_NR; i++) {
2025                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2026                 ret = max(ret, mapcount);
2027                 _total_mapcount += mapcount;
2028         }
2029         if (PageDoubleMap(page)) {
2030                 ret -= 1;
2031                 _total_mapcount -= HPAGE_PMD_NR;
2032         }
2033         mapcount = compound_mapcount(page);
2034         ret += mapcount;
2035         _total_mapcount += mapcount;
2036         if (total_mapcount)
2037                 *total_mapcount = _total_mapcount;
2038         return ret;
2039 }
2040
2041 /*
2042  * This function splits huge page into normal pages. @page can point to any
2043  * subpage of huge page to split. Split doesn't change the position of @page.
2044  *
2045  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2046  * The huge page must be locked.
2047  *
2048  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2049  *
2050  * Both head page and tail pages will inherit mapping, flags, and so on from
2051  * the hugepage.
2052  *
2053  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2054  * they are not mapped.
2055  *
2056  * Returns 0 if the hugepage is split successfully.
2057  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2058  * us.
2059  */
2060 int split_huge_page_to_list(struct page *page, struct list_head *list)
2061 {
2062         struct page *head = compound_head(page);
2063         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2064         struct anon_vma *anon_vma = NULL;
2065         struct address_space *mapping = NULL;
2066         int count, mapcount, extra_pins, ret;
2067         bool mlocked;
2068         unsigned long flags;
2069
2070         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2071         VM_BUG_ON_PAGE(!PageLocked(page), page);
2072         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2073         VM_BUG_ON_PAGE(!PageCompound(page), page);
2074
2075         if (PageAnon(head)) {
2076                 /*
2077                  * The caller does not necessarily hold an mmap_sem that would
2078                  * prevent the anon_vma disappearing so we first we take a
2079                  * reference to it and then lock the anon_vma for write. This
2080                  * is similar to page_lock_anon_vma_read except the write lock
2081                  * is taken to serialise against parallel split or collapse
2082                  * operations.
2083                  */
2084                 anon_vma = page_get_anon_vma(head);
2085                 if (!anon_vma) {
2086                         ret = -EBUSY;
2087                         goto out;
2088                 }
2089                 extra_pins = 0;
2090                 mapping = NULL;
2091                 anon_vma_lock_write(anon_vma);
2092         } else {
2093                 mapping = head->mapping;
2094
2095                 /* Truncated ? */
2096                 if (!mapping) {
2097                         ret = -EBUSY;
2098                         goto out;
2099                 }
2100
2101                 /* Addidional pins from radix tree */
2102                 extra_pins = HPAGE_PMD_NR;
2103                 anon_vma = NULL;
2104                 i_mmap_lock_read(mapping);
2105         }
2106
2107         /*
2108          * Racy check if we can split the page, before freeze_page() will
2109          * split PMDs
2110          */
2111         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2112                 ret = -EBUSY;
2113                 goto out_unlock;
2114         }
2115
2116         mlocked = PageMlocked(page);
2117         freeze_page(head);
2118         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2119
2120         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2121         if (mlocked)
2122                 lru_add_drain();
2123
2124         /* prevent PageLRU to go away from under us, and freeze lru stats */
2125         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2126
2127         if (mapping) {
2128                 void **pslot;
2129
2130                 spin_lock(&mapping->tree_lock);
2131                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2132                                 page_index(head));
2133                 /*
2134                  * Check if the head page is present in radix tree.
2135                  * We assume all tail are present too, if head is there.
2136                  */
2137                 if (radix_tree_deref_slot_protected(pslot,
2138                                         &mapping->tree_lock) != head)
2139                         goto fail;
2140         }
2141
2142         /* Prevent deferred_split_scan() touching ->_refcount */
2143         spin_lock(&pgdata->split_queue_lock);
2144         count = page_count(head);
2145         mapcount = total_mapcount(head);
2146         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2147                 if (!list_empty(page_deferred_list(head))) {
2148                         pgdata->split_queue_len--;
2149                         list_del(page_deferred_list(head));
2150                 }
2151                 if (mapping)
2152                         __dec_node_page_state(page, NR_SHMEM_THPS);
2153                 spin_unlock(&pgdata->split_queue_lock);
2154                 __split_huge_page(page, list, flags);
2155                 ret = 0;
2156         } else {
2157                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2158                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2159                                         mapcount, count);
2160                         if (PageTail(page))
2161                                 dump_page(head, NULL);
2162                         dump_page(page, "total_mapcount(head) > 0");
2163                         BUG();
2164                 }
2165                 spin_unlock(&pgdata->split_queue_lock);
2166 fail:           if (mapping)
2167                         spin_unlock(&mapping->tree_lock);
2168                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2169                 unfreeze_page(head);
2170                 ret = -EBUSY;
2171         }
2172
2173 out_unlock:
2174         if (anon_vma) {
2175                 anon_vma_unlock_write(anon_vma);
2176                 put_anon_vma(anon_vma);
2177         }
2178         if (mapping)
2179                 i_mmap_unlock_read(mapping);
2180 out:
2181         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2182         return ret;
2183 }
2184
2185 void free_transhuge_page(struct page *page)
2186 {
2187         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2188         unsigned long flags;
2189
2190         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2191         if (!list_empty(page_deferred_list(page))) {
2192                 pgdata->split_queue_len--;
2193                 list_del(page_deferred_list(page));
2194         }
2195         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2196         free_compound_page(page);
2197 }
2198
2199 void deferred_split_huge_page(struct page *page)
2200 {
2201         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2202         unsigned long flags;
2203
2204         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2205
2206         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2207         if (list_empty(page_deferred_list(page))) {
2208                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2209                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2210                 pgdata->split_queue_len++;
2211         }
2212         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2213 }
2214
2215 static unsigned long deferred_split_count(struct shrinker *shrink,
2216                 struct shrink_control *sc)
2217 {
2218         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2219         return ACCESS_ONCE(pgdata->split_queue_len);
2220 }
2221
2222 static unsigned long deferred_split_scan(struct shrinker *shrink,
2223                 struct shrink_control *sc)
2224 {
2225         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2226         unsigned long flags;
2227         LIST_HEAD(list), *pos, *next;
2228         struct page *page;
2229         int split = 0;
2230
2231         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2232         /* Take pin on all head pages to avoid freeing them under us */
2233         list_for_each_safe(pos, next, &pgdata->split_queue) {
2234                 page = list_entry((void *)pos, struct page, mapping);
2235                 page = compound_head(page);
2236                 if (get_page_unless_zero(page)) {
2237                         list_move(page_deferred_list(page), &list);
2238                 } else {
2239                         /* We lost race with put_compound_page() */
2240                         list_del_init(page_deferred_list(page));
2241                         pgdata->split_queue_len--;
2242                 }
2243                 if (!--sc->nr_to_scan)
2244                         break;
2245         }
2246         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2247
2248         list_for_each_safe(pos, next, &list) {
2249                 page = list_entry((void *)pos, struct page, mapping);
2250                 lock_page(page);
2251                 /* split_huge_page() removes page from list on success */
2252                 if (!split_huge_page(page))
2253                         split++;
2254                 unlock_page(page);
2255                 put_page(page);
2256         }
2257
2258         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2259         list_splice_tail(&list, &pgdata->split_queue);
2260         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2261
2262         /*
2263          * Stop shrinker if we didn't split any page, but the queue is empty.
2264          * This can happen if pages were freed under us.
2265          */
2266         if (!split && list_empty(&pgdata->split_queue))
2267                 return SHRINK_STOP;
2268         return split;
2269 }
2270
2271 static struct shrinker deferred_split_shrinker = {
2272         .count_objects = deferred_split_count,
2273         .scan_objects = deferred_split_scan,
2274         .seeks = DEFAULT_SEEKS,
2275         .flags = SHRINKER_NUMA_AWARE,
2276 };
2277
2278 #ifdef CONFIG_DEBUG_FS
2279 static int split_huge_pages_set(void *data, u64 val)
2280 {
2281         struct zone *zone;
2282         struct page *page;
2283         unsigned long pfn, max_zone_pfn;
2284         unsigned long total = 0, split = 0;
2285
2286         if (val != 1)
2287                 return -EINVAL;
2288
2289         for_each_populated_zone(zone) {
2290                 max_zone_pfn = zone_end_pfn(zone);
2291                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2292                         if (!pfn_valid(pfn))
2293                                 continue;
2294
2295                         page = pfn_to_page(pfn);
2296                         if (!get_page_unless_zero(page))
2297                                 continue;
2298
2299                         if (zone != page_zone(page))
2300                                 goto next;
2301
2302                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2303                                 goto next;
2304
2305                         total++;
2306                         lock_page(page);
2307                         if (!split_huge_page(page))
2308                                 split++;
2309                         unlock_page(page);
2310 next:
2311                         put_page(page);
2312                 }
2313         }
2314
2315         pr_info("%lu of %lu THP split\n", split, total);
2316
2317         return 0;
2318 }
2319 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2320                 "%llu\n");
2321
2322 static int __init split_huge_pages_debugfs(void)
2323 {
2324         void *ret;
2325
2326         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2327                         &split_huge_pages_fops);
2328         if (!ret)
2329                 pr_warn("Failed to create split_huge_pages in debugfs");
2330         return 0;
2331 }
2332 late_initcall(split_huge_pages_debugfs);
2333 #endif