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