mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/sched/coredump.h>
  13 #include <linux/sched/numa_balancing.h>
  14 #include <linux/highmem.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <linux/rmap.h>
  18 #include <linux/swap.h>
  19 #include <linux/shrinker.h>
  20 #include <linux/mm_inline.h>
  21 #include <linux/swapops.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35
  36 #include <asm/tlb.h>
  37 #include <asm/pgalloc.h>
  38 #include "internal.h"
  39
  40 /*
  41  * By default transparent hugepage support is disabled in order that avoid
  42  * to risk increase the memory footprint of applications without a guaranteed
  43  * benefit. When transparent hugepage support is enabled, is for all mappings,
  44  * and khugepaged scans all mappings.
  45  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  46  * for all hugepage allocations.
  47  */
  48 unsigned long transparent_hugepage_flags __read_mostly =
  49 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  50         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  51 #endif
  52 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  53         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  54 #endif
  55         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  58
  59 static struct shrinker deferred_split_shrinker;
  60
  61 static atomic_t huge_zero_refcount;
  62 struct page *huge_zero_page __read_mostly;
  63
  64 static struct page *get_huge_zero_page(void)
  65 {
  66         struct page *zero_page;
  67 retry:
  68         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  69                 return READ_ONCE(huge_zero_page);
  70
  71         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  72                         HPAGE_PMD_ORDER);
  73         if (!zero_page) {
  74                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  75                 return NULL;
  76         }
  77         count_vm_event(THP_ZERO_PAGE_ALLOC);
  78         preempt_disable();
  79         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  80                 preempt_enable();
  81                 __free_pages(zero_page, compound_order(zero_page));
  82                 goto retry;
  83         }
  84
  85         /* We take additional reference here. It will be put back by shrinker */
  86         atomic_set(&huge_zero_refcount, 2);
  87         preempt_enable();
  88         return READ_ONCE(huge_zero_page);
  89 }
  90
  91 static void put_huge_zero_page(void)
  92 {
  93         /*
  94          * Counter should never go to zero here. Only shrinker can put
  95          * last reference.
  96          */
  97         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  98 }
  99
 100 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 101 {
 102         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 103                 return READ_ONCE(huge_zero_page);
 104
 105         if (!get_huge_zero_page())
 106                 return NULL;
 107
 108         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 109                 put_huge_zero_page();
 110
 111         return READ_ONCE(huge_zero_page);
 112 }
 113
 114 void mm_put_huge_zero_page(struct mm_struct *mm)
 115 {
 116         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 117                 put_huge_zero_page();
 118 }
 119
 120 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 121                                         struct shrink_control *sc)
 122 {
 123         /* we can free zero page only if last reference remains */
 124         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 125 }
 126
 127 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 128                                        struct shrink_control *sc)
 129 {
 130         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 131                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 132                 BUG_ON(zero_page == NULL);
 133                 __free_pages(zero_page, compound_order(zero_page));
 134                 return HPAGE_PMD_NR;
 135         }
 136
 137         return 0;
 138 }
 139
 140 static struct shrinker huge_zero_page_shrinker = {
 141         .count_objects = shrink_huge_zero_page_count,
 142         .scan_objects = shrink_huge_zero_page_scan,
 143         .seeks = DEFAULT_SEEKS,
 144 };
 145
 146 #ifdef CONFIG_SYSFS
 147 static ssize_t enabled_show(struct kobject *kobj,
 148                             struct kobj_attribute *attr, char *buf)
 149 {
 150         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 151                 return sprintf(buf, "[always] madvise never\n");
 152         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 153                 return sprintf(buf, "always [madvise] never\n");
 154         else
 155                 return sprintf(buf, "always madvise [never]\n");
 156 }
 157
 158 static ssize_t enabled_store(struct kobject *kobj,
 159                              struct kobj_attribute *attr,
 160                              const char *buf, size_t count)
 161 {
 162         ssize_t ret = count;
 163
 164         if (!memcmp("always", buf,
 165                     min(sizeof("always")-1, count))) {
 166                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 167                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 168         } else if (!memcmp("madvise", buf,
 169                            min(sizeof("madvise")-1, count))) {
 170                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 171                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 172         } else if (!memcmp("never", buf,
 173                            min(sizeof("never")-1, count))) {
 174                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 175                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 176         } else
 177                 ret = -EINVAL;
 178
 179         if (ret > 0) {
 180                 int err = start_stop_khugepaged();
 181                 if (err)
 182                         ret = err;
 183         }
 184         return ret;
 185 }
 186 static struct kobj_attribute enabled_attr =
 187         __ATTR(enabled, 0644, enabled_show, enabled_store);
 188
 189 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 190                                 struct kobj_attribute *attr, char *buf,
 191                                 enum transparent_hugepage_flag flag)
 192 {
 193         return sprintf(buf, "%d\n",
 194                        !!test_bit(flag, &transparent_hugepage_flags));
 195 }
 196
 197 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 198                                  struct kobj_attribute *attr,
 199                                  const char *buf, size_t count,
 200                                  enum transparent_hugepage_flag flag)
 201 {
 202         unsigned long value;
 203         int ret;
 204
 205         ret = kstrtoul(buf, 10, &value);
 206         if (ret < 0)
 207                 return ret;
 208         if (value > 1)
 209                 return -EINVAL;
 210
 211         if (value)
 212                 set_bit(flag, &transparent_hugepage_flags);
 213         else
 214                 clear_bit(flag, &transparent_hugepage_flags);
 215
 216         return count;
 217 }
 218
 219 static ssize_t defrag_show(struct kobject *kobj,
 220                            struct kobj_attribute *attr, char *buf)
 221 {
 222         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 223                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 224         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 225                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 226         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 227                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 228         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 229                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 230         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 231 }
 232
 233 static ssize_t defrag_store(struct kobject *kobj,
 234                             struct kobj_attribute *attr,
 235                             const char *buf, size_t count)
 236 {
 237         if (!memcmp("always", buf,
 238                     min(sizeof("always")-1, count))) {
 239                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 240                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 241                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 242                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 243         } else if (!memcmp("defer", buf,
 244                     min(sizeof("defer")-1, count))) {
 245                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 246                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 248                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 249         } else if (!memcmp("defer+madvise", buf,
 250                     min(sizeof("defer+madvise")-1, count))) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 255         } else if (!memcmp("madvise", buf,
 256                            min(sizeof("madvise")-1, count))) {
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 261         } else if (!memcmp("never", buf,
 262                            min(sizeof("never")-1, count))) {
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 267         } else
 268                 return -EINVAL;
 269
 270         return count;
 271 }
 272 static struct kobj_attribute defrag_attr =
 273         __ATTR(defrag, 0644, defrag_show, defrag_store);
 274
 275 static ssize_t use_zero_page_show(struct kobject *kobj,
 276                 struct kobj_attribute *attr, char *buf)
 277 {
 278         return single_hugepage_flag_show(kobj, attr, buf,
 279                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 280 }
 281 static ssize_t use_zero_page_store(struct kobject *kobj,
 282                 struct kobj_attribute *attr, const char *buf, size_t count)
 283 {
 284         return single_hugepage_flag_store(kobj, attr, buf, count,
 285                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 286 }
 287 static struct kobj_attribute use_zero_page_attr =
 288         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 289
 290 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 291                 struct kobj_attribute *attr, char *buf)
 292 {
 293         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 294 }
 295 static struct kobj_attribute hpage_pmd_size_attr =
 296         __ATTR_RO(hpage_pmd_size);
 297
 298 #ifdef CONFIG_DEBUG_VM
 299 static ssize_t debug_cow_show(struct kobject *kobj,
 300                                 struct kobj_attribute *attr, char *buf)
 301 {
 302         return single_hugepage_flag_show(kobj, attr, buf,
 303                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 304 }
 305 static ssize_t debug_cow_store(struct kobject *kobj,
 306                                struct kobj_attribute *attr,
 307                                const char *buf, size_t count)
 308 {
 309         return single_hugepage_flag_store(kobj, attr, buf, count,
 310                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 311 }
 312 static struct kobj_attribute debug_cow_attr =
 313         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 314 #endif /* CONFIG_DEBUG_VM */
 315
 316 static struct attribute *hugepage_attr[] = {
 317         &enabled_attr.attr,
 318         &defrag_attr.attr,
 319         &use_zero_page_attr.attr,
 320         &hpage_pmd_size_attr.attr,
 321 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 322         &shmem_enabled_attr.attr,
 323 #endif
 324 #ifdef CONFIG_DEBUG_VM
 325         &debug_cow_attr.attr,
 326 #endif
 327         NULL,
 328 };
 329
 330 static struct attribute_group hugepage_attr_group = {
 331         .attrs = hugepage_attr,
 332 };
 333
 334 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 335 {
 336         int err;
 337
 338         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 339         if (unlikely(!*hugepage_kobj)) {
 340                 pr_err("failed to create transparent hugepage kobject\n");
 341                 return -ENOMEM;
 342         }
 343
 344         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 345         if (err) {
 346                 pr_err("failed to register transparent hugepage group\n");
 347                 goto delete_obj;
 348         }
 349
 350         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 351         if (err) {
 352                 pr_err("failed to register transparent hugepage group\n");
 353                 goto remove_hp_group;
 354         }
 355
 356         return 0;
 357
 358 remove_hp_group:
 359         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 360 delete_obj:
 361         kobject_put(*hugepage_kobj);
 362         return err;
 363 }
 364
 365 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 366 {
 367         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 368         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 369         kobject_put(hugepage_kobj);
 370 }
 371 #else
 372 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 373 {
 374         return 0;
 375 }
 376
 377 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 378 {
 379 }
 380 #endif /* CONFIG_SYSFS */
 381
 382 static int __init hugepage_init(void)
 383 {
 384         int err;
 385         struct kobject *hugepage_kobj;
 386
 387         if (!has_transparent_hugepage()) {
 388                 transparent_hugepage_flags = 0;
 389                 return -EINVAL;
 390         }
 391
 392         /*
 393          * hugepages can't be allocated by the buddy allocator
 394          */
 395         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 396         /*
 397          * we use page->mapping and page->index in second tail page
 398          * as list_head: assuming THP order >= 2
 399          */
 400         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 401
 402         err = hugepage_init_sysfs(&hugepage_kobj);
 403         if (err)
 404                 goto err_sysfs;
 405
 406         err = khugepaged_init();
 407         if (err)
 408                 goto err_slab;
 409
 410         err = register_shrinker(&huge_zero_page_shrinker);
 411         if (err)
 412                 goto err_hzp_shrinker;
 413         err = register_shrinker(&deferred_split_shrinker);
 414         if (err)
 415                 goto err_split_shrinker;
 416
 417         /*
 418          * By default disable transparent hugepages on smaller systems,
 419          * where the extra memory used could hurt more than TLB overhead
 420          * is likely to save.  The admin can still enable it through /sys.
 421          */
 422         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 423                 transparent_hugepage_flags = 0;
 424                 return 0;
 425         }
 426
 427         err = start_stop_khugepaged();
 428         if (err)
 429                 goto err_khugepaged;
 430
 431         return 0;
 432 err_khugepaged:
 433         unregister_shrinker(&deferred_split_shrinker);
 434 err_split_shrinker:
 435         unregister_shrinker(&huge_zero_page_shrinker);
 436 err_hzp_shrinker:
 437         khugepaged_destroy();
 438 err_slab:
 439         hugepage_exit_sysfs(hugepage_kobj);
 440 err_sysfs:
 441         return err;
 442 }
 443 subsys_initcall(hugepage_init);
 444
 445 static int __init setup_transparent_hugepage(char *str)
 446 {
 447         int ret = 0;
 448         if (!str)
 449                 goto out;
 450         if (!strcmp(str, "always")) {
 451                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 452                         &transparent_hugepage_flags);
 453                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 454                           &transparent_hugepage_flags);
 455                 ret = 1;
 456         } else if (!strcmp(str, "madvise")) {
 457                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 458                           &transparent_hugepage_flags);
 459                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 460                         &transparent_hugepage_flags);
 461                 ret = 1;
 462         } else if (!strcmp(str, "never")) {
 463                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 464                           &transparent_hugepage_flags);
 465                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 466                           &transparent_hugepage_flags);
 467                 ret = 1;
 468         }
 469 out:
 470         if (!ret)
 471                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 472         return ret;
 473 }
 474 __setup("transparent_hugepage=", setup_transparent_hugepage);
 475
 476 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 477 {
 478         if (likely(vma->vm_flags & VM_WRITE))
 479                 pmd = pmd_mkwrite(pmd);
 480         return pmd;
 481 }
 482
 483 static inline struct list_head *page_deferred_list(struct page *page)
 484 {
 485         /*
 486          * ->lru in the tail pages is occupied by compound_head.
 487          * Let's use ->mapping + ->index in the second tail page as list_head.
 488          */
 489         return (struct list_head *)&page[2].mapping;
 490 }
 491
 492 void prep_transhuge_page(struct page *page)
 493 {
 494         /*
 495          * we use page->mapping and page->indexlru in second tail page
 496          * as list_head: assuming THP order >= 2
 497          */
 498
 499         INIT_LIST_HEAD(page_deferred_list(page));
 500         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 501 }
 502
 503 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 504                 loff_t off, unsigned long flags, unsigned long size)
 505 {
 506         unsigned long addr;
 507         loff_t off_end = off + len;
 508         loff_t off_align = round_up(off, size);
 509         unsigned long len_pad;
 510
 511         if (off_end <= off_align || (off_end - off_align) < size)
 512                 return 0;
 513
 514         len_pad = len + size;
 515         if (len_pad < len || (off + len_pad) < off)
 516                 return 0;
 517
 518         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 519                                               off >> PAGE_SHIFT, flags);
 520         if (IS_ERR_VALUE(addr))
 521                 return 0;
 522
 523         addr += (off - addr) & (size - 1);
 524         return addr;
 525 }
 526
 527 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 528                 unsigned long len, unsigned long pgoff, unsigned long flags)
 529 {
 530         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 531
 532         if (addr)
 533                 goto out;
 534         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 535                 goto out;
 536
 537         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 538         if (addr)
 539                 return addr;
 540
 541  out:
 542         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 543 }
 544 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 545
 546 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 547                 gfp_t gfp)
 548 {
 549         struct vm_area_struct *vma = vmf->vma;
 550         struct mem_cgroup *memcg;
 551         pgtable_t pgtable;
 552         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 553
 554         VM_BUG_ON_PAGE(!PageCompound(page), page);
 555
 556         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 557                 put_page(page);
 558                 count_vm_event(THP_FAULT_FALLBACK);
 559                 return VM_FAULT_FALLBACK;
 560         }
 561
 562         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 563         if (unlikely(!pgtable)) {
 564                 mem_cgroup_cancel_charge(page, memcg, true);
 565                 put_page(page);
 566                 return VM_FAULT_OOM;
 567         }
 568
 569         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 570         /*
 571          * The memory barrier inside __SetPageUptodate makes sure that
 572          * clear_huge_page writes become visible before the set_pmd_at()
 573          * write.
 574          */
 575         __SetPageUptodate(page);
 576
 577         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 578         if (unlikely(!pmd_none(*vmf->pmd))) {
 579                 spin_unlock(vmf->ptl);
 580                 mem_cgroup_cancel_charge(page, memcg, true);
 581                 put_page(page);
 582                 pte_free(vma->vm_mm, pgtable);
 583         } else {
 584                 pmd_t entry;
 585
 586                 /* Deliver the page fault to userland */
 587                 if (userfaultfd_missing(vma)) {
 588                         int ret;
 589
 590                         spin_unlock(vmf->ptl);
 591                         mem_cgroup_cancel_charge(page, memcg, true);
 592                         put_page(page);
 593                         pte_free(vma->vm_mm, pgtable);
 594                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 595                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 596                         return ret;
 597                 }
 598
 599                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 600                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 601                 page_add_new_anon_rmap(page, vma, haddr, true);
 602                 mem_cgroup_commit_charge(page, memcg, false, true);
 603                 lru_cache_add_active_or_unevictable(page, vma);
 604                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 605                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 606                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 607                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 608                 spin_unlock(vmf->ptl);
 609                 count_vm_event(THP_FAULT_ALLOC);
 610         }
 611
 612         return 0;
 613 }
 614
 615 /*
 616  * always: directly stall for all thp allocations
 617  * defer: wake kswapd and fail if not immediately available
 618  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 619  *                fail if not immediately available
 620  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 621  *          available
 622  * never: never stall for any thp allocation
 623  */
 624 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 625 {
 626         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 627
 628         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 629                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 630         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 631                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 632         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 633                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 634                                                              __GFP_KSWAPD_RECLAIM);
 635         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 636                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 637                                                              0);
 638         return GFP_TRANSHUGE_LIGHT;
 639 }
 640
 641 /* Caller must hold page table lock. */
 642 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 643                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 644                 struct page *zero_page)
 645 {
 646         pmd_t entry;
 647         if (!pmd_none(*pmd))
 648                 return false;
 649         entry = mk_pmd(zero_page, vma->vm_page_prot);
 650         entry = pmd_mkhuge(entry);
 651         if (pgtable)
 652                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 653         set_pmd_at(mm, haddr, pmd, entry);
 654         atomic_long_inc(&mm->nr_ptes);
 655         return true;
 656 }
 657
 658 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 659 {
 660         struct vm_area_struct *vma = vmf->vma;
 661         gfp_t gfp;
 662         struct page *page;
 663         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 664
 665         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 666                 return VM_FAULT_FALLBACK;
 667         if (unlikely(anon_vma_prepare(vma)))
 668                 return VM_FAULT_OOM;
 669         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 670                 return VM_FAULT_OOM;
 671         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 672                         !mm_forbids_zeropage(vma->vm_mm) &&
 673                         transparent_hugepage_use_zero_page()) {
 674                 pgtable_t pgtable;
 675                 struct page *zero_page;
 676                 bool set;
 677                 int ret;
 678                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 679                 if (unlikely(!pgtable))
 680                         return VM_FAULT_OOM;
 681                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 682                 if (unlikely(!zero_page)) {
 683                         pte_free(vma->vm_mm, pgtable);
 684                         count_vm_event(THP_FAULT_FALLBACK);
 685                         return VM_FAULT_FALLBACK;
 686                 }
 687                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 688                 ret = 0;
 689                 set = false;
 690                 if (pmd_none(*vmf->pmd)) {
 691                         if (userfaultfd_missing(vma)) {
 692                                 spin_unlock(vmf->ptl);
 693                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 694                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 695                         } else {
 696                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 697                                                    haddr, vmf->pmd, zero_page);
 698                                 spin_unlock(vmf->ptl);
 699                                 set = true;
 700                         }
 701                 } else
 702                         spin_unlock(vmf->ptl);
 703                 if (!set)
 704                         pte_free(vma->vm_mm, pgtable);
 705                 return ret;
 706         }
 707         gfp = alloc_hugepage_direct_gfpmask(vma);
 708         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 709         if (unlikely(!page)) {
 710                 count_vm_event(THP_FAULT_FALLBACK);
 711                 return VM_FAULT_FALLBACK;
 712         }
 713         prep_transhuge_page(page);
 714         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 715 }
 716
 717 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 718                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 719 {
 720         struct mm_struct *mm = vma->vm_mm;
 721         pmd_t entry;
 722         spinlock_t *ptl;
 723
 724         ptl = pmd_lock(mm, pmd);
 725         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 726         if (pfn_t_devmap(pfn))
 727                 entry = pmd_mkdevmap(entry);
 728         if (write) {
 729                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 730                 entry = maybe_pmd_mkwrite(entry, vma);
 731         }
 732         set_pmd_at(mm, addr, pmd, entry);
 733         update_mmu_cache_pmd(vma, addr, pmd);
 734         spin_unlock(ptl);
 735 }
 736
 737 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 738                         pmd_t *pmd, pfn_t pfn, bool write)
 739 {
 740         pgprot_t pgprot = vma->vm_page_prot;
 741         /*
 742          * If we had pmd_special, we could avoid all these restrictions,
 743          * but we need to be consistent with PTEs and architectures that
 744          * can't support a 'special' bit.
 745          */
 746         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 747         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 748                                                 (VM_PFNMAP|VM_MIXEDMAP));
 749         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 750         BUG_ON(!pfn_t_devmap(pfn));
 751
 752         if (addr < vma->vm_start || addr >= vma->vm_end)
 753                 return VM_FAULT_SIGBUS;
 754
 755         track_pfn_insert(vma, &pgprot, pfn);
 756
 757         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 758         return VM_FAULT_NOPAGE;
 759 }
 760 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 761
 762 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 763 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 764 {
 765         if (likely(vma->vm_flags & VM_WRITE))
 766                 pud = pud_mkwrite(pud);
 767         return pud;
 768 }
 769
 770 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 771                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 772 {
 773         struct mm_struct *mm = vma->vm_mm;
 774         pud_t entry;
 775         spinlock_t *ptl;
 776
 777         ptl = pud_lock(mm, pud);
 778         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 779         if (pfn_t_devmap(pfn))
 780                 entry = pud_mkdevmap(entry);
 781         if (write) {
 782                 entry = pud_mkyoung(pud_mkdirty(entry));
 783                 entry = maybe_pud_mkwrite(entry, vma);
 784         }
 785         set_pud_at(mm, addr, pud, entry);
 786         update_mmu_cache_pud(vma, addr, pud);
 787         spin_unlock(ptl);
 788 }
 789
 790 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 791                         pud_t *pud, pfn_t pfn, bool write)
 792 {
 793         pgprot_t pgprot = vma->vm_page_prot;
 794         /*
 795          * If we had pud_special, we could avoid all these restrictions,
 796          * but we need to be consistent with PTEs and architectures that
 797          * can't support a 'special' bit.
 798          */
 799         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 800         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 801                                                 (VM_PFNMAP|VM_MIXEDMAP));
 802         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 803         BUG_ON(!pfn_t_devmap(pfn));
 804
 805         if (addr < vma->vm_start || addr >= vma->vm_end)
 806                 return VM_FAULT_SIGBUS;
 807
 808         track_pfn_insert(vma, &pgprot, pfn);
 809
 810         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 811         return VM_FAULT_NOPAGE;
 812 }
 813 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 814 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 815
 816 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 817                 pmd_t *pmd)
 818 {
 819         pmd_t _pmd;
 820
 821         /*
 822          * We should set the dirty bit only for FOLL_WRITE but for now
 823          * the dirty bit in the pmd is meaningless.  And if the dirty
 824          * bit will become meaningful and we'll only set it with
 825          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 826          * set the young bit, instead of the current set_pmd_at.
 827          */
 828         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 829         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 830                                 pmd, _pmd,  1))
 831                 update_mmu_cache_pmd(vma, addr, pmd);
 832 }
 833
 834 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 835                 pmd_t *pmd, int flags)
 836 {
 837         unsigned long pfn = pmd_pfn(*pmd);
 838         struct mm_struct *mm = vma->vm_mm;
 839         struct dev_pagemap *pgmap;
 840         struct page *page;
 841
 842         assert_spin_locked(pmd_lockptr(mm, pmd));
 843
 844         /*
 845          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 846          * not be in this function with `flags & FOLL_COW` set.
 847          */
 848         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 849
 850         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 851                 return NULL;
 852
 853         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 854                 /* pass */;
 855         else
 856                 return NULL;
 857
 858         if (flags & FOLL_TOUCH)
 859                 touch_pmd(vma, addr, pmd);
 860
 861         /*
 862          * device mapped pages can only be returned if the
 863          * caller will manage the page reference count.
 864          */
 865         if (!(flags & FOLL_GET))
 866                 return ERR_PTR(-EEXIST);
 867
 868         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 869         pgmap = get_dev_pagemap(pfn, NULL);
 870         if (!pgmap)
 871                 return ERR_PTR(-EFAULT);
 872         page = pfn_to_page(pfn);
 873         get_page(page);
 874         put_dev_pagemap(pgmap);
 875
 876         return page;
 877 }
 878
 879 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 880                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 881                   struct vm_area_struct *vma)
 882 {
 883         spinlock_t *dst_ptl, *src_ptl;
 884         struct page *src_page;
 885         pmd_t pmd;
 886         pgtable_t pgtable = NULL;
 887         int ret = -ENOMEM;
 888
 889         /* Skip if can be re-fill on fault */
 890         if (!vma_is_anonymous(vma))
 891                 return 0;
 892
 893         pgtable = pte_alloc_one(dst_mm, addr);
 894         if (unlikely(!pgtable))
 895                 goto out;
 896
 897         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 898         src_ptl = pmd_lockptr(src_mm, src_pmd);
 899         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 900
 901         ret = -EAGAIN;
 902         pmd = *src_pmd;
 903         if (unlikely(!pmd_trans_huge(pmd))) {
 904                 pte_free(dst_mm, pgtable);
 905                 goto out_unlock;
 906         }
 907         /*
 908          * When page table lock is held, the huge zero pmd should not be
 909          * under splitting since we don't split the page itself, only pmd to
 910          * a page table.
 911          */
 912         if (is_huge_zero_pmd(pmd)) {
 913                 struct page *zero_page;
 914                 /*
 915                  * get_huge_zero_page() will never allocate a new page here,
 916                  * since we already have a zero page to copy. It just takes a
 917                  * reference.
 918                  */
 919                 zero_page = mm_get_huge_zero_page(dst_mm);
 920                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 921                                 zero_page);
 922                 ret = 0;
 923                 goto out_unlock;
 924         }
 925
 926         src_page = pmd_page(pmd);
 927         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 928         get_page(src_page);
 929         page_dup_rmap(src_page, true);
 930         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 931         atomic_long_inc(&dst_mm->nr_ptes);
 932         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 933
 934         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 935         pmd = pmd_mkold(pmd_wrprotect(pmd));
 936         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 937
 938         ret = 0;
 939 out_unlock:
 940         spin_unlock(src_ptl);
 941         spin_unlock(dst_ptl);
 942 out:
 943         return ret;
 944 }
 945
 946 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 947 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 948                 pud_t *pud)
 949 {
 950         pud_t _pud;
 951
 952         /*
 953          * We should set the dirty bit only for FOLL_WRITE but for now
 954          * the dirty bit in the pud is meaningless.  And if the dirty
 955          * bit will become meaningful and we'll only set it with
 956          * FOLL_WRITE, an atomic set_bit will be required on the pud to
 957          * set the young bit, instead of the current set_pud_at.
 958          */
 959         _pud = pud_mkyoung(pud_mkdirty(*pud));
 960         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
 961                                 pud, _pud,  1))
 962                 update_mmu_cache_pud(vma, addr, pud);
 963 }
 964
 965 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 966                 pud_t *pud, int flags)
 967 {
 968         unsigned long pfn = pud_pfn(*pud);
 969         struct mm_struct *mm = vma->vm_mm;
 970         struct dev_pagemap *pgmap;
 971         struct page *page;
 972
 973         assert_spin_locked(pud_lockptr(mm, pud));
 974
 975         if (flags & FOLL_WRITE && !pud_write(*pud))
 976                 return NULL;
 977
 978         if (pud_present(*pud) && pud_devmap(*pud))
 979                 /* pass */;
 980         else
 981                 return NULL;
 982
 983         if (flags & FOLL_TOUCH)
 984                 touch_pud(vma, addr, pud);
 985
 986         /*
 987          * device mapped pages can only be returned if the
 988          * caller will manage the page reference count.
 989          */
 990         if (!(flags & FOLL_GET))
 991                 return ERR_PTR(-EEXIST);
 992
 993         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
 994         pgmap = get_dev_pagemap(pfn, NULL);
 995         if (!pgmap)
 996                 return ERR_PTR(-EFAULT);
 997         page = pfn_to_page(pfn);
 998         get_page(page);
 999         put_dev_pagemap(pgmap);
1000
1001         return page;
1002 }
1003
1004 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1005                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1006                   struct vm_area_struct *vma)
1007 {
1008         spinlock_t *dst_ptl, *src_ptl;
1009         pud_t pud;
1010         int ret;
1011
1012         dst_ptl = pud_lock(dst_mm, dst_pud);
1013         src_ptl = pud_lockptr(src_mm, src_pud);
1014         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1015
1016         ret = -EAGAIN;
1017         pud = *src_pud;
1018         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1019                 goto out_unlock;
1020
1021         /*
1022          * When page table lock is held, the huge zero pud should not be
1023          * under splitting since we don't split the page itself, only pud to
1024          * a page table.
1025          */
1026         if (is_huge_zero_pud(pud)) {
1027                 /* No huge zero pud yet */
1028         }
1029
1030         pudp_set_wrprotect(src_mm, addr, src_pud);
1031         pud = pud_mkold(pud_wrprotect(pud));
1032         set_pud_at(dst_mm, addr, dst_pud, pud);
1033
1034         ret = 0;
1035 out_unlock:
1036         spin_unlock(src_ptl);
1037         spin_unlock(dst_ptl);
1038         return ret;
1039 }
1040
1041 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1042 {
1043         pud_t entry;
1044         unsigned long haddr;
1045         bool write = vmf->flags & FAULT_FLAG_WRITE;
1046
1047         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1048         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1049                 goto unlock;
1050
1051         entry = pud_mkyoung(orig_pud);
1052         if (write)
1053                 entry = pud_mkdirty(entry);
1054         haddr = vmf->address & HPAGE_PUD_MASK;
1055         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1056                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1057
1058 unlock:
1059         spin_unlock(vmf->ptl);
1060 }
1061 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1062
1063 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1064 {
1065         pmd_t entry;
1066         unsigned long haddr;
1067         bool write = vmf->flags & FAULT_FLAG_WRITE;
1068
1069         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1070         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1071                 goto unlock;
1072
1073         entry = pmd_mkyoung(orig_pmd);
1074         if (write)
1075                 entry = pmd_mkdirty(entry);
1076         haddr = vmf->address & HPAGE_PMD_MASK;
1077         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1078                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1079
1080 unlock:
1081         spin_unlock(vmf->ptl);
1082 }
1083
1084 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1085                 struct page *page)
1086 {
1087         struct vm_area_struct *vma = vmf->vma;
1088         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1089         struct mem_cgroup *memcg;
1090         pgtable_t pgtable;
1091         pmd_t _pmd;
1092         int ret = 0, i;
1093         struct page **pages;
1094         unsigned long mmun_start;       /* For mmu_notifiers */
1095         unsigned long mmun_end;         /* For mmu_notifiers */
1096
1097         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1098                         GFP_KERNEL);
1099         if (unlikely(!pages)) {
1100                 ret |= VM_FAULT_OOM;
1101                 goto out;
1102         }
1103
1104         for (i = 0; i < HPAGE_PMD_NR; i++) {
1105                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1106                                                vmf->address, page_to_nid(page));
1107                 if (unlikely(!pages[i] ||
1108                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1109                                      GFP_KERNEL, &memcg, false))) {
1110                         if (pages[i])
1111                                 put_page(pages[i]);
1112                         while (--i >= 0) {
1113                                 memcg = (void *)page_private(pages[i]);
1114                                 set_page_private(pages[i], 0);
1115                                 mem_cgroup_cancel_charge(pages[i], memcg,
1116                                                 false);
1117                                 put_page(pages[i]);
1118                         }
1119                         kfree(pages);
1120                         ret |= VM_FAULT_OOM;
1121                         goto out;
1122                 }
1123                 set_page_private(pages[i], (unsigned long)memcg);
1124         }
1125
1126         for (i = 0; i < HPAGE_PMD_NR; i++) {
1127                 copy_user_highpage(pages[i], page + i,
1128                                    haddr + PAGE_SIZE * i, vma);
1129                 __SetPageUptodate(pages[i]);
1130                 cond_resched();
1131         }
1132
1133         mmun_start = haddr;
1134         mmun_end   = haddr + HPAGE_PMD_SIZE;
1135         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1136
1137         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1138         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1139                 goto out_free_pages;
1140         VM_BUG_ON_PAGE(!PageHead(page), page);
1141
1142         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1143         /* leave pmd empty until pte is filled */
1144
1145         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1146         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1147
1148         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1149                 pte_t entry;
1150                 entry = mk_pte(pages[i], vma->vm_page_prot);
1151                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1152                 memcg = (void *)page_private(pages[i]);
1153                 set_page_private(pages[i], 0);
1154                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1155                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1156                 lru_cache_add_active_or_unevictable(pages[i], vma);
1157                 vmf->pte = pte_offset_map(&_pmd, haddr);
1158                 VM_BUG_ON(!pte_none(*vmf->pte));
1159                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1160                 pte_unmap(vmf->pte);
1161         }
1162         kfree(pages);
1163
1164         smp_wmb(); /* make pte visible before pmd */
1165         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1166         page_remove_rmap(page, true);
1167         spin_unlock(vmf->ptl);
1168
1169         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1170
1171         ret |= VM_FAULT_WRITE;
1172         put_page(page);
1173
1174 out:
1175         return ret;
1176
1177 out_free_pages:
1178         spin_unlock(vmf->ptl);
1179         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1180         for (i = 0; i < HPAGE_PMD_NR; i++) {
1181                 memcg = (void *)page_private(pages[i]);
1182                 set_page_private(pages[i], 0);
1183                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1184                 put_page(pages[i]);
1185         }
1186         kfree(pages);
1187         goto out;
1188 }
1189
1190 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1191 {
1192         struct vm_area_struct *vma = vmf->vma;
1193         struct page *page = NULL, *new_page;
1194         struct mem_cgroup *memcg;
1195         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1196         unsigned long mmun_start;       /* For mmu_notifiers */
1197         unsigned long mmun_end;         /* For mmu_notifiers */
1198         gfp_t huge_gfp;                 /* for allocation and charge */
1199         int ret = 0;
1200
1201         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1202         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1203         if (is_huge_zero_pmd(orig_pmd))
1204                 goto alloc;
1205         spin_lock(vmf->ptl);
1206         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1207                 goto out_unlock;
1208
1209         page = pmd_page(orig_pmd);
1210         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1211         /*
1212          * We can only reuse the page if nobody else maps the huge page or it's
1213          * part.
1214          */
1215         if (page_trans_huge_mapcount(page, NULL) == 1) {
1216                 pmd_t entry;
1217                 entry = pmd_mkyoung(orig_pmd);
1218                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1219                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1220                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1221                 ret |= VM_FAULT_WRITE;
1222                 goto out_unlock;
1223         }
1224         get_page(page);
1225         spin_unlock(vmf->ptl);
1226 alloc:
1227         if (transparent_hugepage_enabled(vma) &&
1228             !transparent_hugepage_debug_cow()) {
1229                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1230                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1231         } else
1232                 new_page = NULL;
1233
1234         if (likely(new_page)) {
1235                 prep_transhuge_page(new_page);
1236         } else {
1237                 if (!page) {
1238                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1239                         ret |= VM_FAULT_FALLBACK;
1240                 } else {
1241                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1242                         if (ret & VM_FAULT_OOM) {
1243                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1244                                 ret |= VM_FAULT_FALLBACK;
1245                         }
1246                         put_page(page);
1247                 }
1248                 count_vm_event(THP_FAULT_FALLBACK);
1249                 goto out;
1250         }
1251
1252         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1253                                         huge_gfp, &memcg, true))) {
1254                 put_page(new_page);
1255                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1256                 if (page)
1257                         put_page(page);
1258                 ret |= VM_FAULT_FALLBACK;
1259                 count_vm_event(THP_FAULT_FALLBACK);
1260                 goto out;
1261         }
1262
1263         count_vm_event(THP_FAULT_ALLOC);
1264
1265         if (!page)
1266                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1267         else
1268                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1269         __SetPageUptodate(new_page);
1270
1271         mmun_start = haddr;
1272         mmun_end   = haddr + HPAGE_PMD_SIZE;
1273         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1274
1275         spin_lock(vmf->ptl);
1276         if (page)
1277                 put_page(page);
1278         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1279                 spin_unlock(vmf->ptl);
1280                 mem_cgroup_cancel_charge(new_page, memcg, true);
1281                 put_page(new_page);
1282                 goto out_mn;
1283         } else {
1284                 pmd_t entry;
1285                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1286                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1287                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1288                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1289                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1290                 lru_cache_add_active_or_unevictable(new_page, vma);
1291                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1292                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1293                 if (!page) {
1294                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1295                 } else {
1296                         VM_BUG_ON_PAGE(!PageHead(page), page);
1297                         page_remove_rmap(page, true);
1298                         put_page(page);
1299                 }
1300                 ret |= VM_FAULT_WRITE;
1301         }
1302         spin_unlock(vmf->ptl);
1303 out_mn:
1304         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1305 out:
1306         return ret;
1307 out_unlock:
1308         spin_unlock(vmf->ptl);
1309         return ret;
1310 }
1311
1312 /*
1313  * FOLL_FORCE can write to even unwritable pmd's, but only
1314  * after we've gone through a COW cycle and they are dirty.
1315  */
1316 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1317 {
1318         return pmd_write(pmd) ||
1319                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1320 }
1321
1322 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1323                                    unsigned long addr,
1324                                    pmd_t *pmd,
1325                                    unsigned int flags)
1326 {
1327         struct mm_struct *mm = vma->vm_mm;
1328         struct page *page = NULL;
1329
1330         assert_spin_locked(pmd_lockptr(mm, pmd));
1331
1332         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1333                 goto out;
1334
1335         /* Avoid dumping huge zero page */
1336         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1337                 return ERR_PTR(-EFAULT);
1338
1339         /* Full NUMA hinting faults to serialise migration in fault paths */
1340         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1341                 goto out;
1342
1343         page = pmd_page(*pmd);
1344         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1345         if (flags & FOLL_TOUCH)
1346                 touch_pmd(vma, addr, pmd);
1347         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1348                 /*
1349                  * We don't mlock() pte-mapped THPs. This way we can avoid
1350                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1351                  *
1352                  * For anon THP:
1353                  *
1354                  * In most cases the pmd is the only mapping of the page as we
1355                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1356                  * writable private mappings in populate_vma_page_range().
1357                  *
1358                  * The only scenario when we have the page shared here is if we
1359                  * mlocking read-only mapping shared over fork(). We skip
1360                  * mlocking such pages.
1361                  *
1362                  * For file THP:
1363                  *
1364                  * We can expect PageDoubleMap() to be stable under page lock:
1365                  * for file pages we set it in page_add_file_rmap(), which
1366                  * requires page to be locked.
1367                  */
1368
1369                 if (PageAnon(page) && compound_mapcount(page) != 1)
1370                         goto skip_mlock;
1371                 if (PageDoubleMap(page) || !page->mapping)
1372                         goto skip_mlock;
1373                 if (!trylock_page(page))
1374                         goto skip_mlock;
1375                 lru_add_drain();
1376                 if (page->mapping && !PageDoubleMap(page))
1377                         mlock_vma_page(page);
1378                 unlock_page(page);
1379         }
1380 skip_mlock:
1381         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1382         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1383         if (flags & FOLL_GET)
1384                 get_page(page);
1385
1386 out:
1387         return page;
1388 }
1389
1390 /* NUMA hinting page fault entry point for trans huge pmds */
1391 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1392 {
1393         struct vm_area_struct *vma = vmf->vma;
1394         struct anon_vma *anon_vma = NULL;
1395         struct page *page;
1396         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1397         int page_nid = -1, this_nid = numa_node_id();
1398         int target_nid, last_cpupid = -1;
1399         bool page_locked;
1400         bool migrated = false;
1401         bool was_writable;
1402         int flags = 0;
1403
1404         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1405         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1406                 goto out_unlock;
1407
1408         /*
1409          * If there are potential migrations, wait for completion and retry
1410          * without disrupting NUMA hinting information. Do not relock and
1411          * check_same as the page may no longer be mapped.
1412          */
1413         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1414                 page = pmd_page(*vmf->pmd);
1415                 spin_unlock(vmf->ptl);
1416                 wait_on_page_locked(page);
1417                 goto out;
1418         }
1419
1420         page = pmd_page(pmd);
1421         BUG_ON(is_huge_zero_page(page));
1422         page_nid = page_to_nid(page);
1423         last_cpupid = page_cpupid_last(page);
1424         count_vm_numa_event(NUMA_HINT_FAULTS);
1425         if (page_nid == this_nid) {
1426                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1427                 flags |= TNF_FAULT_LOCAL;
1428         }
1429
1430         /* See similar comment in do_numa_page for explanation */
1431         if (!pmd_savedwrite(pmd))
1432                 flags |= TNF_NO_GROUP;
1433
1434         /*
1435          * Acquire the page lock to serialise THP migrations but avoid dropping
1436          * page_table_lock if at all possible
1437          */
1438         page_locked = trylock_page(page);
1439         target_nid = mpol_misplaced(page, vma, haddr);
1440         if (target_nid == -1) {
1441                 /* If the page was locked, there are no parallel migrations */
1442                 if (page_locked)
1443                         goto clear_pmdnuma;
1444         }
1445
1446         /* Migration could have started since the pmd_trans_migrating check */
1447         if (!page_locked) {
1448                 spin_unlock(vmf->ptl);
1449                 wait_on_page_locked(page);
1450                 page_nid = -1;
1451                 goto out;
1452         }
1453
1454         /*
1455          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1456          * to serialises splits
1457          */
1458         get_page(page);
1459         spin_unlock(vmf->ptl);
1460         anon_vma = page_lock_anon_vma_read(page);
1461
1462         /* Confirm the PMD did not change while page_table_lock was released */
1463         spin_lock(vmf->ptl);
1464         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1465                 unlock_page(page);
1466                 put_page(page);
1467                 page_nid = -1;
1468                 goto out_unlock;
1469         }
1470
1471         /* Bail if we fail to protect against THP splits for any reason */
1472         if (unlikely(!anon_vma)) {
1473                 put_page(page);
1474                 page_nid = -1;
1475                 goto clear_pmdnuma;
1476         }
1477
1478         /*
1479          * Migrate the THP to the requested node, returns with page unlocked
1480          * and access rights restored.
1481          */
1482         spin_unlock(vmf->ptl);
1483         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1484                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1485         if (migrated) {
1486                 flags |= TNF_MIGRATED;
1487                 page_nid = target_nid;
1488         } else
1489                 flags |= TNF_MIGRATE_FAIL;
1490
1491         goto out;
1492 clear_pmdnuma:
1493         BUG_ON(!PageLocked(page));
1494         was_writable = pmd_savedwrite(pmd);
1495         pmd = pmd_modify(pmd, vma->vm_page_prot);
1496         pmd = pmd_mkyoung(pmd);
1497         if (was_writable)
1498                 pmd = pmd_mkwrite(pmd);
1499         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1500         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1501         unlock_page(page);
1502 out_unlock:
1503         spin_unlock(vmf->ptl);
1504
1505 out:
1506         if (anon_vma)
1507                 page_unlock_anon_vma_read(anon_vma);
1508
1509         if (page_nid != -1)
1510                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1511                                 flags);
1512
1513         return 0;
1514 }
1515
1516 /*
1517  * Return true if we do MADV_FREE successfully on entire pmd page.
1518  * Otherwise, return false.
1519  */
1520 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1521                 pmd_t *pmd, unsigned long addr, unsigned long next)
1522 {
1523         spinlock_t *ptl;
1524         pmd_t orig_pmd;
1525         struct page *page;
1526         struct mm_struct *mm = tlb->mm;
1527         bool ret = false;
1528
1529         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1530
1531         ptl = pmd_trans_huge_lock(pmd, vma);
1532         if (!ptl)
1533                 goto out_unlocked;
1534
1535         orig_pmd = *pmd;
1536         if (is_huge_zero_pmd(orig_pmd))
1537                 goto out;
1538
1539         page = pmd_page(orig_pmd);
1540         /*
1541          * If other processes are mapping this page, we couldn't discard
1542          * the page unless they all do MADV_FREE so let's skip the page.
1543          */
1544         if (page_mapcount(page) != 1)
1545                 goto out;
1546
1547         if (!trylock_page(page))
1548                 goto out;
1549
1550         /*
1551          * If user want to discard part-pages of THP, split it so MADV_FREE
1552          * will deactivate only them.
1553          */
1554         if (next - addr != HPAGE_PMD_SIZE) {
1555                 get_page(page);
1556                 spin_unlock(ptl);
1557                 split_huge_page(page);
1558                 put_page(page);
1559                 unlock_page(page);
1560                 goto out_unlocked;
1561         }
1562
1563         if (PageDirty(page))
1564                 ClearPageDirty(page);
1565         unlock_page(page);
1566
1567         if (PageActive(page))
1568                 deactivate_page(page);
1569
1570         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1571                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1572                         tlb->fullmm);
1573                 orig_pmd = pmd_mkold(orig_pmd);
1574                 orig_pmd = pmd_mkclean(orig_pmd);
1575
1576                 set_pmd_at(mm, addr, pmd, orig_pmd);
1577                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1578         }
1579         ret = true;
1580 out:
1581         spin_unlock(ptl);
1582 out_unlocked:
1583         return ret;
1584 }
1585
1586 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1587 {
1588         pgtable_t pgtable;
1589
1590         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1591         pte_free(mm, pgtable);
1592         atomic_long_dec(&mm->nr_ptes);
1593 }
1594
1595 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1596                  pmd_t *pmd, unsigned long addr)
1597 {
1598         pmd_t orig_pmd;
1599         spinlock_t *ptl;
1600
1601         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1602
1603         ptl = __pmd_trans_huge_lock(pmd, vma);
1604         if (!ptl)
1605                 return 0;
1606         /*
1607          * For architectures like ppc64 we look at deposited pgtable
1608          * when calling pmdp_huge_get_and_clear. So do the
1609          * pgtable_trans_huge_withdraw after finishing pmdp related
1610          * operations.
1611          */
1612         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1613                         tlb->fullmm);
1614         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1615         if (vma_is_dax(vma)) {
1616                 spin_unlock(ptl);
1617                 if (is_huge_zero_pmd(orig_pmd))
1618                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1619         } else if (is_huge_zero_pmd(orig_pmd)) {
1620                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1621                 atomic_long_dec(&tlb->mm->nr_ptes);
1622                 spin_unlock(ptl);
1623                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1624         } else {
1625                 struct page *page = pmd_page(orig_pmd);
1626                 page_remove_rmap(page, true);
1627                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1628                 VM_BUG_ON_PAGE(!PageHead(page), page);
1629                 if (PageAnon(page)) {
1630                         pgtable_t pgtable;
1631                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1632                         pte_free(tlb->mm, pgtable);
1633                         atomic_long_dec(&tlb->mm->nr_ptes);
1634                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1635                 } else {
1636                         if (arch_needs_pgtable_deposit())
1637                                 zap_deposited_table(tlb->mm, pmd);
1638                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1639                 }
1640                 spin_unlock(ptl);
1641                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1642         }
1643         return 1;
1644 }
1645
1646 #ifndef pmd_move_must_withdraw
1647 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1648                                          spinlock_t *old_pmd_ptl,
1649                                          struct vm_area_struct *vma)
1650 {
1651         /*
1652          * With split pmd lock we also need to move preallocated
1653          * PTE page table if new_pmd is on different PMD page table.
1654          *
1655          * We also don't deposit and withdraw tables for file pages.
1656          */
1657         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1658 }
1659 #endif
1660
1661 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1662                   unsigned long new_addr, unsigned long old_end,
1663                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1664 {
1665         spinlock_t *old_ptl, *new_ptl;
1666         pmd_t pmd;
1667         struct mm_struct *mm = vma->vm_mm;
1668         bool force_flush = false;
1669
1670         if ((old_addr & ~HPAGE_PMD_MASK) ||
1671             (new_addr & ~HPAGE_PMD_MASK) ||
1672             old_end - old_addr < HPAGE_PMD_SIZE)
1673                 return false;
1674
1675         /*
1676          * The destination pmd shouldn't be established, free_pgtables()
1677          * should have release it.
1678          */
1679         if (WARN_ON(!pmd_none(*new_pmd))) {
1680                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1681                 return false;
1682         }
1683
1684         /*
1685          * We don't have to worry about the ordering of src and dst
1686          * ptlocks because exclusive mmap_sem prevents deadlock.
1687          */
1688         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1689         if (old_ptl) {
1690                 new_ptl = pmd_lockptr(mm, new_pmd);
1691                 if (new_ptl != old_ptl)
1692                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1693                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1694                 if (pmd_present(pmd) && pmd_dirty(pmd))
1695                         force_flush = true;
1696                 VM_BUG_ON(!pmd_none(*new_pmd));
1697
1698                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1699                         pgtable_t pgtable;
1700                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1701                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1702                 }
1703                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1704                 if (new_ptl != old_ptl)
1705                         spin_unlock(new_ptl);
1706                 if (force_flush)
1707                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1708                 else
1709                         *need_flush = true;
1710                 spin_unlock(old_ptl);
1711                 return true;
1712         }
1713         return false;
1714 }
1715
1716 /*
1717  * Returns
1718  *  - 0 if PMD could not be locked
1719  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1720  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1721  */
1722 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1723                 unsigned long addr, pgprot_t newprot, int prot_numa)
1724 {
1725         struct mm_struct *mm = vma->vm_mm;
1726         spinlock_t *ptl;
1727         int ret = 0;
1728
1729         ptl = __pmd_trans_huge_lock(pmd, vma);
1730         if (ptl) {
1731                 pmd_t entry;
1732                 bool preserve_write = prot_numa && pmd_write(*pmd);
1733                 ret = 1;
1734
1735                 /*
1736                  * Avoid trapping faults against the zero page. The read-only
1737                  * data is likely to be read-cached on the local CPU and
1738                  * local/remote hits to the zero page are not interesting.
1739                  */
1740                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1741                         spin_unlock(ptl);
1742                         return ret;
1743                 }
1744
1745                 if (!prot_numa || !pmd_protnone(*pmd)) {
1746                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1747                         entry = pmd_modify(entry, newprot);
1748                         if (preserve_write)
1749                                 entry = pmd_mk_savedwrite(entry);
1750                         ret = HPAGE_PMD_NR;
1751                         set_pmd_at(mm, addr, pmd, entry);
1752                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1753                                         pmd_write(entry));
1754                 }
1755                 spin_unlock(ptl);
1756         }
1757
1758         return ret;
1759 }
1760
1761 /*
1762  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1763  *
1764  * Note that if it returns page table lock pointer, this routine returns without
1765  * unlocking page table lock. So callers must unlock it.
1766  */
1767 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1768 {
1769         spinlock_t *ptl;
1770         ptl = pmd_lock(vma->vm_mm, pmd);
1771         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1772                 return ptl;
1773         spin_unlock(ptl);
1774         return NULL;
1775 }
1776
1777 /*
1778  * Returns true if a given pud maps a thp, false otherwise.
1779  *
1780  * Note that if it returns true, this routine returns without unlocking page
1781  * table lock. So callers must unlock it.
1782  */
1783 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1784 {
1785         spinlock_t *ptl;
1786
1787         ptl = pud_lock(vma->vm_mm, pud);
1788         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1789                 return ptl;
1790         spin_unlock(ptl);
1791         return NULL;
1792 }
1793
1794 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1795 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1796                  pud_t *pud, unsigned long addr)
1797 {
1798         pud_t orig_pud;
1799         spinlock_t *ptl;
1800
1801         ptl = __pud_trans_huge_lock(pud, vma);
1802         if (!ptl)
1803                 return 0;
1804         /*
1805          * For architectures like ppc64 we look at deposited pgtable
1806          * when calling pudp_huge_get_and_clear. So do the
1807          * pgtable_trans_huge_withdraw after finishing pudp related
1808          * operations.
1809          */
1810         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1811                         tlb->fullmm);
1812         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1813         if (vma_is_dax(vma)) {
1814                 spin_unlock(ptl);
1815                 /* No zero page support yet */
1816         } else {
1817                 /* No support for anonymous PUD pages yet */
1818                 BUG();
1819         }
1820         return 1;
1821 }
1822
1823 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1824                 unsigned long haddr)
1825 {
1826         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1827         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1828         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1829         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1830
1831         count_vm_event(THP_SPLIT_PMD);
1832
1833         pudp_huge_clear_flush_notify(vma, haddr, pud);
1834 }
1835
1836 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1837                 unsigned long address)
1838 {
1839         spinlock_t *ptl;
1840         struct mm_struct *mm = vma->vm_mm;
1841         unsigned long haddr = address & HPAGE_PUD_MASK;
1842
1843         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1844         ptl = pud_lock(mm, pud);
1845         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1846                 goto out;
1847         __split_huge_pud_locked(vma, pud, haddr);
1848
1849 out:
1850         spin_unlock(ptl);
1851         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1852 }
1853 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1854
1855 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1856                 unsigned long haddr, pmd_t *pmd)
1857 {
1858         struct mm_struct *mm = vma->vm_mm;
1859         pgtable_t pgtable;
1860         pmd_t _pmd;
1861         int i;
1862
1863         /* leave pmd empty until pte is filled */
1864         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1865
1866         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1867         pmd_populate(mm, &_pmd, pgtable);
1868
1869         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1870                 pte_t *pte, entry;
1871                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1872                 entry = pte_mkspecial(entry);
1873                 pte = pte_offset_map(&_pmd, haddr);
1874                 VM_BUG_ON(!pte_none(*pte));
1875                 set_pte_at(mm, haddr, pte, entry);
1876                 pte_unmap(pte);
1877         }
1878         smp_wmb(); /* make pte visible before pmd */
1879         pmd_populate(mm, pmd, pgtable);
1880 }
1881
1882 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1883                 unsigned long haddr, bool freeze)
1884 {
1885         struct mm_struct *mm = vma->vm_mm;
1886         struct page *page;
1887         pgtable_t pgtable;
1888         pmd_t _pmd;
1889         bool young, write, dirty, soft_dirty;
1890         unsigned long addr;
1891         int i;
1892
1893         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1894         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1895         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1896         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1897
1898         count_vm_event(THP_SPLIT_PMD);
1899
1900         if (!vma_is_anonymous(vma)) {
1901                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1902                 /*
1903                  * We are going to unmap this huge page. So
1904                  * just go ahead and zap it
1905                  */
1906                 if (arch_needs_pgtable_deposit())
1907                         zap_deposited_table(mm, pmd);
1908                 if (vma_is_dax(vma))
1909                         return;
1910                 page = pmd_page(_pmd);
1911                 if (!PageReferenced(page) && pmd_young(_pmd))
1912                         SetPageReferenced(page);
1913                 page_remove_rmap(page, true);
1914                 put_page(page);
1915                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1916                 return;
1917         } else if (is_huge_zero_pmd(*pmd)) {
1918                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1919         }
1920
1921         page = pmd_page(*pmd);
1922         VM_BUG_ON_PAGE(!page_count(page), page);
1923         page_ref_add(page, HPAGE_PMD_NR - 1);
1924         write = pmd_write(*pmd);
1925         young = pmd_young(*pmd);
1926         dirty = pmd_dirty(*pmd);
1927         soft_dirty = pmd_soft_dirty(*pmd);
1928
1929         pmdp_huge_split_prepare(vma, haddr, pmd);
1930         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1931         pmd_populate(mm, &_pmd, pgtable);
1932
1933         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1934                 pte_t entry, *pte;
1935                 /*
1936                  * Note that NUMA hinting access restrictions are not
1937                  * transferred to avoid any possibility of altering
1938                  * permissions across VMAs.
1939                  */
1940                 if (freeze) {
1941                         swp_entry_t swp_entry;
1942                         swp_entry = make_migration_entry(page + i, write);
1943                         entry = swp_entry_to_pte(swp_entry);
1944                         if (soft_dirty)
1945                                 entry = pte_swp_mksoft_dirty(entry);
1946                 } else {
1947                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1948                         entry = maybe_mkwrite(entry, vma);
1949                         if (!write)
1950                                 entry = pte_wrprotect(entry);
1951                         if (!young)
1952                                 entry = pte_mkold(entry);
1953                         if (soft_dirty)
1954                                 entry = pte_mksoft_dirty(entry);
1955                 }
1956                 if (dirty)
1957                         SetPageDirty(page + i);
1958                 pte = pte_offset_map(&_pmd, addr);
1959                 BUG_ON(!pte_none(*pte));
1960                 set_pte_at(mm, addr, pte, entry);
1961                 atomic_inc(&page[i]._mapcount);
1962                 pte_unmap(pte);
1963         }
1964
1965         /*
1966          * Set PG_double_map before dropping compound_mapcount to avoid
1967          * false-negative page_mapped().
1968          */
1969         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1970                 for (i = 0; i < HPAGE_PMD_NR; i++)
1971                         atomic_inc(&page[i]._mapcount);
1972         }
1973
1974         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1975                 /* Last compound_mapcount is gone. */
1976                 __dec_node_page_state(page, NR_ANON_THPS);
1977                 if (TestClearPageDoubleMap(page)) {
1978                         /* No need in mapcount reference anymore */
1979                         for (i = 0; i < HPAGE_PMD_NR; i++)
1980                                 atomic_dec(&page[i]._mapcount);
1981                 }
1982         }
1983
1984         smp_wmb(); /* make pte visible before pmd */
1985         /*
1986          * Up to this point the pmd is present and huge and userland has the
1987          * whole access to the hugepage during the split (which happens in
1988          * place). If we overwrite the pmd with the not-huge version pointing
1989          * to the pte here (which of course we could if all CPUs were bug
1990          * free), userland could trigger a small page size TLB miss on the
1991          * small sized TLB while the hugepage TLB entry is still established in
1992          * the huge TLB. Some CPU doesn't like that.
1993          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1994          * 383 on page 93. Intel should be safe but is also warns that it's
1995          * only safe if the permission and cache attributes of the two entries
1996          * loaded in the two TLB is identical (which should be the case here).
1997          * But it is generally safer to never allow small and huge TLB entries
1998          * for the same virtual address to be loaded simultaneously. So instead
1999          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2000          * current pmd notpresent (atomically because here the pmd_trans_huge
2001          * and pmd_trans_splitting must remain set at all times on the pmd
2002          * until the split is complete for this pmd), then we flush the SMP TLB
2003          * and finally we write the non-huge version of the pmd entry with
2004          * pmd_populate.
2005          */
2006         pmdp_invalidate(vma, haddr, pmd);
2007         pmd_populate(mm, pmd, pgtable);
2008
2009         if (freeze) {
2010                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2011                         page_remove_rmap(page + i, false);
2012                         put_page(page + i);
2013                 }
2014         }
2015 }
2016
2017 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2018                 unsigned long address, bool freeze, struct page *page)
2019 {
2020         spinlock_t *ptl;
2021         struct mm_struct *mm = vma->vm_mm;
2022         unsigned long haddr = address & HPAGE_PMD_MASK;
2023
2024         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2025         ptl = pmd_lock(mm, pmd);
2026
2027         /*
2028          * If caller asks to setup a migration entries, we need a page to check
2029          * pmd against. Otherwise we can end up replacing wrong page.
2030          */
2031         VM_BUG_ON(freeze && !page);
2032         if (page && page != pmd_page(*pmd))
2033                 goto out;
2034
2035         if (pmd_trans_huge(*pmd)) {
2036                 page = pmd_page(*pmd);
2037                 if (PageMlocked(page))
2038                         clear_page_mlock(page);
2039         } else if (!pmd_devmap(*pmd))
2040                 goto out;
2041         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2042 out:
2043         spin_unlock(ptl);
2044         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2045 }
2046
2047 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2048                 bool freeze, struct page *page)
2049 {
2050         pgd_t *pgd;
2051         pud_t *pud;
2052         pmd_t *pmd;
2053
2054         pgd = pgd_offset(vma->vm_mm, address);
2055         if (!pgd_present(*pgd))
2056                 return;
2057
2058         pud = pud_offset(pgd, address);
2059         if (!pud_present(*pud))
2060                 return;
2061
2062         pmd = pmd_offset(pud, address);
2063
2064         __split_huge_pmd(vma, pmd, address, freeze, page);
2065 }
2066
2067 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2068                              unsigned long start,
2069                              unsigned long end,
2070                              long adjust_next)
2071 {
2072         /*
2073          * If the new start address isn't hpage aligned and it could
2074          * previously contain an hugepage: check if we need to split
2075          * an huge pmd.
2076          */
2077         if (start & ~HPAGE_PMD_MASK &&
2078             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2079             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2080                 split_huge_pmd_address(vma, start, false, NULL);
2081
2082         /*
2083          * If the new end address isn't hpage aligned and it could
2084          * previously contain an hugepage: check if we need to split
2085          * an huge pmd.
2086          */
2087         if (end & ~HPAGE_PMD_MASK &&
2088             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2089             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2090                 split_huge_pmd_address(vma, end, false, NULL);
2091
2092         /*
2093          * If we're also updating the vma->vm_next->vm_start, if the new
2094          * vm_next->vm_start isn't page aligned and it could previously
2095          * contain an hugepage: check if we need to split an huge pmd.
2096          */
2097         if (adjust_next > 0) {
2098                 struct vm_area_struct *next = vma->vm_next;
2099                 unsigned long nstart = next->vm_start;
2100                 nstart += adjust_next << PAGE_SHIFT;
2101                 if (nstart & ~HPAGE_PMD_MASK &&
2102                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2103                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2104                         split_huge_pmd_address(next, nstart, false, NULL);
2105         }
2106 }
2107
2108 static void freeze_page(struct page *page)
2109 {
2110         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2111                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2112         int ret;
2113
2114         VM_BUG_ON_PAGE(!PageHead(page), page);
2115
2116         if (PageAnon(page))
2117                 ttu_flags |= TTU_MIGRATION;
2118
2119         ret = try_to_unmap(page, ttu_flags);
2120         VM_BUG_ON_PAGE(ret, page);
2121 }
2122
2123 static void unfreeze_page(struct page *page)
2124 {
2125         int i;
2126         if (PageTransHuge(page)) {
2127                 remove_migration_ptes(page, page, true);
2128         } else {
2129                 for (i = 0; i < HPAGE_PMD_NR; i++)
2130                         remove_migration_ptes(page + i, page + i, true);
2131         }
2132 }
2133
2134 static void __split_huge_page_tail(struct page *head, int tail,
2135                 struct lruvec *lruvec, struct list_head *list)
2136 {
2137         struct page *page_tail = head + tail;
2138
2139         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2140         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2141
2142         /*
2143          * tail_page->_refcount is zero and not changing from under us. But
2144          * get_page_unless_zero() may be running from under us on the
2145          * tail_page. If we used atomic_set() below instead of atomic_inc() or
2146          * atomic_add(), we would then run atomic_set() concurrently with
2147          * get_page_unless_zero(), and atomic_set() is implemented in C not
2148          * using locked ops. spin_unlock on x86 sometime uses locked ops
2149          * because of PPro errata 66, 92, so unless somebody can guarantee
2150          * atomic_set() here would be safe on all archs (and not only on x86),
2151          * it's safer to use atomic_inc()/atomic_add().
2152          */
2153         if (PageAnon(head)) {
2154                 page_ref_inc(page_tail);
2155         } else {
2156                 /* Additional pin to radix tree */
2157                 page_ref_add(page_tail, 2);
2158         }
2159
2160         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2161         page_tail->flags |= (head->flags &
2162                         ((1L << PG_referenced) |
2163                          (1L << PG_swapbacked) |
2164                          (1L << PG_mlocked) |
2165                          (1L << PG_uptodate) |
2166                          (1L << PG_active) |
2167                          (1L << PG_locked) |
2168                          (1L << PG_unevictable) |
2169                          (1L << PG_dirty)));
2170
2171         /*
2172          * After clearing PageTail the gup refcount can be released.
2173          * Page flags also must be visible before we make the page non-compound.
2174          */
2175         smp_wmb();
2176
2177         clear_compound_head(page_tail);
2178
2179         if (page_is_young(head))
2180                 set_page_young(page_tail);
2181         if (page_is_idle(head))
2182                 set_page_idle(page_tail);
2183
2184         /* ->mapping in first tail page is compound_mapcount */
2185         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2186                         page_tail);
2187         page_tail->mapping = head->mapping;
2188
2189         page_tail->index = head->index + tail;
2190         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2191         lru_add_page_tail(head, page_tail, lruvec, list);
2192 }
2193
2194 static void __split_huge_page(struct page *page, struct list_head *list,
2195                 unsigned long flags)
2196 {
2197         struct page *head = compound_head(page);
2198         struct zone *zone = page_zone(head);
2199         struct lruvec *lruvec;
2200         pgoff_t end = -1;
2201         int i;
2202
2203         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2204
2205         /* complete memcg works before add pages to LRU */
2206         mem_cgroup_split_huge_fixup(head);
2207
2208         if (!PageAnon(page))
2209                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2210
2211         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2212                 __split_huge_page_tail(head, i, lruvec, list);
2213                 /* Some pages can be beyond i_size: drop them from page cache */
2214                 if (head[i].index >= end) {
2215                         __ClearPageDirty(head + i);
2216                         __delete_from_page_cache(head + i, NULL);
2217                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2218                                 shmem_uncharge(head->mapping->host, 1);
2219                         put_page(head + i);
2220                 }
2221         }
2222
2223         ClearPageCompound(head);
2224         /* See comment in __split_huge_page_tail() */
2225         if (PageAnon(head)) {
2226                 page_ref_inc(head);
2227         } else {
2228                 /* Additional pin to radix tree */
2229                 page_ref_add(head, 2);
2230                 spin_unlock(&head->mapping->tree_lock);
2231         }
2232
2233         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2234
2235         unfreeze_page(head);
2236
2237         for (i = 0; i < HPAGE_PMD_NR; i++) {
2238                 struct page *subpage = head + i;
2239                 if (subpage == page)
2240                         continue;
2241                 unlock_page(subpage);
2242
2243                 /*
2244                  * Subpages may be freed if there wasn't any mapping
2245                  * like if add_to_swap() is running on a lru page that
2246                  * had its mapping zapped. And freeing these pages
2247                  * requires taking the lru_lock so we do the put_page
2248                  * of the tail pages after the split is complete.
2249                  */
2250                 put_page(subpage);
2251         }
2252 }
2253
2254 int total_mapcount(struct page *page)
2255 {
2256         int i, compound, ret;
2257
2258         VM_BUG_ON_PAGE(PageTail(page), page);
2259
2260         if (likely(!PageCompound(page)))
2261                 return atomic_read(&page->_mapcount) + 1;
2262
2263         compound = compound_mapcount(page);
2264         if (PageHuge(page))
2265                 return compound;
2266         ret = compound;
2267         for (i = 0; i < HPAGE_PMD_NR; i++)
2268                 ret += atomic_read(&page[i]._mapcount) + 1;
2269         /* File pages has compound_mapcount included in _mapcount */
2270         if (!PageAnon(page))
2271                 return ret - compound * HPAGE_PMD_NR;
2272         if (PageDoubleMap(page))
2273                 ret -= HPAGE_PMD_NR;
2274         return ret;
2275 }
2276
2277 /*
2278  * This calculates accurately how many mappings a transparent hugepage
2279  * has (unlike page_mapcount() which isn't fully accurate). This full
2280  * accuracy is primarily needed to know if copy-on-write faults can
2281  * reuse the page and change the mapping to read-write instead of
2282  * copying them. At the same time this returns the total_mapcount too.
2283  *
2284  * The function returns the highest mapcount any one of the subpages
2285  * has. If the return value is one, even if different processes are
2286  * mapping different subpages of the transparent hugepage, they can
2287  * all reuse it, because each process is reusing a different subpage.
2288  *
2289  * The total_mapcount is instead counting all virtual mappings of the
2290  * subpages. If the total_mapcount is equal to "one", it tells the
2291  * caller all mappings belong to the same "mm" and in turn the
2292  * anon_vma of the transparent hugepage can become the vma->anon_vma
2293  * local one as no other process may be mapping any of the subpages.
2294  *
2295  * It would be more accurate to replace page_mapcount() with
2296  * page_trans_huge_mapcount(), however we only use
2297  * page_trans_huge_mapcount() in the copy-on-write faults where we
2298  * need full accuracy to avoid breaking page pinning, because
2299  * page_trans_huge_mapcount() is slower than page_mapcount().
2300  */
2301 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2302 {
2303         int i, ret, _total_mapcount, mapcount;
2304
2305         /* hugetlbfs shouldn't call it */
2306         VM_BUG_ON_PAGE(PageHuge(page), page);
2307
2308         if (likely(!PageTransCompound(page))) {
2309                 mapcount = atomic_read(&page->_mapcount) + 1;
2310                 if (total_mapcount)
2311                         *total_mapcount = mapcount;
2312                 return mapcount;
2313         }
2314
2315         page = compound_head(page);
2316
2317         _total_mapcount = ret = 0;
2318         for (i = 0; i < HPAGE_PMD_NR; i++) {
2319                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2320                 ret = max(ret, mapcount);
2321                 _total_mapcount += mapcount;
2322         }
2323         if (PageDoubleMap(page)) {
2324                 ret -= 1;
2325                 _total_mapcount -= HPAGE_PMD_NR;
2326         }
2327         mapcount = compound_mapcount(page);
2328         ret += mapcount;
2329         _total_mapcount += mapcount;
2330         if (total_mapcount)
2331                 *total_mapcount = _total_mapcount;
2332         return ret;
2333 }
2334
2335 /*
2336  * This function splits huge page into normal pages. @page can point to any
2337  * subpage of huge page to split. Split doesn't change the position of @page.
2338  *
2339  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2340  * The huge page must be locked.
2341  *
2342  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2343  *
2344  * Both head page and tail pages will inherit mapping, flags, and so on from
2345  * the hugepage.
2346  *
2347  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2348  * they are not mapped.
2349  *
2350  * Returns 0 if the hugepage is split successfully.
2351  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2352  * us.
2353  */
2354 int split_huge_page_to_list(struct page *page, struct list_head *list)
2355 {
2356         struct page *head = compound_head(page);
2357         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2358         struct anon_vma *anon_vma = NULL;
2359         struct address_space *mapping = NULL;
2360         int count, mapcount, extra_pins, ret;
2361         bool mlocked;
2362         unsigned long flags;
2363
2364         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2365         VM_BUG_ON_PAGE(!PageLocked(page), page);
2366         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2367         VM_BUG_ON_PAGE(!PageCompound(page), page);
2368
2369         if (PageAnon(head)) {
2370                 /*
2371                  * The caller does not necessarily hold an mmap_sem that would
2372                  * prevent the anon_vma disappearing so we first we take a
2373                  * reference to it and then lock the anon_vma for write. This
2374                  * is similar to page_lock_anon_vma_read except the write lock
2375                  * is taken to serialise against parallel split or collapse
2376                  * operations.
2377                  */
2378                 anon_vma = page_get_anon_vma(head);
2379                 if (!anon_vma) {
2380                         ret = -EBUSY;
2381                         goto out;
2382                 }
2383                 extra_pins = 0;
2384                 mapping = NULL;
2385                 anon_vma_lock_write(anon_vma);
2386         } else {
2387                 mapping = head->mapping;
2388
2389                 /* Truncated ? */
2390                 if (!mapping) {
2391                         ret = -EBUSY;
2392                         goto out;
2393                 }
2394
2395                 /* Addidional pins from radix tree */
2396                 extra_pins = HPAGE_PMD_NR;
2397                 anon_vma = NULL;
2398                 i_mmap_lock_read(mapping);
2399         }
2400
2401         /*
2402          * Racy check if we can split the page, before freeze_page() will
2403          * split PMDs
2404          */
2405         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2406                 ret = -EBUSY;
2407                 goto out_unlock;
2408         }
2409
2410         mlocked = PageMlocked(page);
2411         freeze_page(head);
2412         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2413
2414         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2415         if (mlocked)
2416                 lru_add_drain();
2417
2418         /* prevent PageLRU to go away from under us, and freeze lru stats */
2419         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2420
2421         if (mapping) {
2422                 void **pslot;
2423
2424                 spin_lock(&mapping->tree_lock);
2425                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2426                                 page_index(head));
2427                 /*
2428                  * Check if the head page is present in radix tree.
2429                  * We assume all tail are present too, if head is there.
2430                  */
2431                 if (radix_tree_deref_slot_protected(pslot,
2432                                         &mapping->tree_lock) != head)
2433                         goto fail;
2434         }
2435
2436         /* Prevent deferred_split_scan() touching ->_refcount */
2437         spin_lock(&pgdata->split_queue_lock);
2438         count = page_count(head);
2439         mapcount = total_mapcount(head);
2440         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2441                 if (!list_empty(page_deferred_list(head))) {
2442                         pgdata->split_queue_len--;
2443                         list_del(page_deferred_list(head));
2444                 }
2445                 if (mapping)
2446                         __dec_node_page_state(page, NR_SHMEM_THPS);
2447                 spin_unlock(&pgdata->split_queue_lock);
2448                 __split_huge_page(page, list, flags);
2449                 ret = 0;
2450         } else {
2451                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2452                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2453                                         mapcount, count);
2454                         if (PageTail(page))
2455                                 dump_page(head, NULL);
2456                         dump_page(page, "total_mapcount(head) > 0");
2457                         BUG();
2458                 }
2459                 spin_unlock(&pgdata->split_queue_lock);
2460 fail:           if (mapping)
2461                         spin_unlock(&mapping->tree_lock);
2462                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2463                 unfreeze_page(head);
2464                 ret = -EBUSY;
2465         }
2466
2467 out_unlock:
2468         if (anon_vma) {
2469                 anon_vma_unlock_write(anon_vma);
2470                 put_anon_vma(anon_vma);
2471         }
2472         if (mapping)
2473                 i_mmap_unlock_read(mapping);
2474 out:
2475         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2476         return ret;
2477 }
2478
2479 void free_transhuge_page(struct page *page)
2480 {
2481         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2482         unsigned long flags;
2483
2484         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2485         if (!list_empty(page_deferred_list(page))) {
2486                 pgdata->split_queue_len--;
2487                 list_del(page_deferred_list(page));
2488         }
2489         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2490         free_compound_page(page);
2491 }
2492
2493 void deferred_split_huge_page(struct page *page)
2494 {
2495         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2496         unsigned long flags;
2497
2498         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2499
2500         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2501         if (list_empty(page_deferred_list(page))) {
2502                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2503                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2504                 pgdata->split_queue_len++;
2505         }
2506         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2507 }
2508
2509 static unsigned long deferred_split_count(struct shrinker *shrink,
2510                 struct shrink_control *sc)
2511 {
2512         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2513         return ACCESS_ONCE(pgdata->split_queue_len);
2514 }
2515
2516 static unsigned long deferred_split_scan(struct shrinker *shrink,
2517                 struct shrink_control *sc)
2518 {
2519         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2520         unsigned long flags;
2521         LIST_HEAD(list), *pos, *next;
2522         struct page *page;
2523         int split = 0;
2524
2525         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2526         /* Take pin on all head pages to avoid freeing them under us */
2527         list_for_each_safe(pos, next, &pgdata->split_queue) {
2528                 page = list_entry((void *)pos, struct page, mapping);
2529                 page = compound_head(page);
2530                 if (get_page_unless_zero(page)) {
2531                         list_move(page_deferred_list(page), &list);
2532                 } else {
2533                         /* We lost race with put_compound_page() */
2534                         list_del_init(page_deferred_list(page));
2535                         pgdata->split_queue_len--;
2536                 }
2537                 if (!--sc->nr_to_scan)
2538                         break;
2539         }
2540         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2541
2542         list_for_each_safe(pos, next, &list) {
2543                 page = list_entry((void *)pos, struct page, mapping);
2544                 lock_page(page);
2545                 /* split_huge_page() removes page from list on success */
2546                 if (!split_huge_page(page))
2547                         split++;
2548                 unlock_page(page);
2549                 put_page(page);
2550         }
2551
2552         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2553         list_splice_tail(&list, &pgdata->split_queue);
2554         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2555
2556         /*
2557          * Stop shrinker if we didn't split any page, but the queue is empty.
2558          * This can happen if pages were freed under us.
2559          */
2560         if (!split && list_empty(&pgdata->split_queue))
2561                 return SHRINK_STOP;
2562         return split;
2563 }
2564
2565 static struct shrinker deferred_split_shrinker = {
2566         .count_objects = deferred_split_count,
2567         .scan_objects = deferred_split_scan,
2568         .seeks = DEFAULT_SEEKS,
2569         .flags = SHRINKER_NUMA_AWARE,
2570 };
2571
2572 #ifdef CONFIG_DEBUG_FS
2573 static int split_huge_pages_set(void *data, u64 val)
2574 {
2575         struct zone *zone;
2576         struct page *page;
2577         unsigned long pfn, max_zone_pfn;
2578         unsigned long total = 0, split = 0;
2579
2580         if (val != 1)
2581                 return -EINVAL;
2582
2583         for_each_populated_zone(zone) {
2584                 max_zone_pfn = zone_end_pfn(zone);
2585                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2586                         if (!pfn_valid(pfn))
2587                                 continue;
2588
2589                         page = pfn_to_page(pfn);
2590                         if (!get_page_unless_zero(page))
2591                                 continue;
2592
2593                         if (zone != page_zone(page))
2594                                 goto next;
2595
2596                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2597                                 goto next;
2598
2599                         total++;
2600                         lock_page(page);
2601                         if (!split_huge_page(page))
2602                                 split++;
2603                         unlock_page(page);
2604 next:
2605                         put_page(page);
2606                 }
2607         }
2608
2609         pr_info("%lu of %lu THP split\n", split, total);
2610
2611         return 0;
2612 }
2613 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2614                 "%llu\n");
2615
2616 static int __init split_huge_pages_debugfs(void)
2617 {
2618         void *ret;
2619
2620         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2621                         &split_huge_pages_fops);
2622         if (!ret)
2623                 pr_warn("Failed to create split_huge_pages in debugfs");
2624         return 0;
2625 }
2626 late_initcall(split_huge_pages_debugfs);
2627 #endif