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