mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/dax.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33 #include <linux/shmem_fs.h>
  34 #include <linux/oom.h>
  35 #include <linux/numa.h>
  36 #include <linux/page_owner.h>
  37
  38 #include <asm/tlb.h>
  39 #include <asm/pgalloc.h>
  40 #include "internal.h"
  41
  42 /*
  43  * By default, transparent hugepage support is disabled in order to avoid
  44  * risking an increased memory footprint for applications that are not
  45  * guaranteed to benefit from it. When transparent hugepage support is
  46  * enabled, it is for all mappings, and khugepaged scans all mappings.
  47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  48  * for all hugepage allocations.
  49  */
  50 unsigned long transparent_hugepage_flags __read_mostly =
  51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  52         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  53 #endif
  54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  55         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  56 #endif
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  59         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  60
  61 static struct shrinker deferred_split_shrinker;
  62
  63 static atomic_t huge_zero_refcount;
  64 struct page *huge_zero_page __read_mostly;
  65
  66 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  67 {
  68         /* The addr is used to check if the vma size fits */
  69         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  70
  71         if (!transhuge_vma_suitable(vma, addr))
  72                 return false;
  73         if (vma_is_anonymous(vma))
  74                 return __transparent_hugepage_enabled(vma);
  75         if (vma_is_shmem(vma))
  76                 return shmem_huge_enabled(vma);
  77
  78         return false;
  79 }
  80
  81 static bool get_huge_zero_page(void)
  82 {
  83         struct page *zero_page;
  84 retry:
  85         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  86                 return true;
  87
  88         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  89                         HPAGE_PMD_ORDER);
  90         if (!zero_page) {
  91                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  92                 return false;
  93         }
  94         count_vm_event(THP_ZERO_PAGE_ALLOC);
  95         preempt_disable();
  96         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  97                 preempt_enable();
  98                 __free_pages(zero_page, compound_order(zero_page));
  99                 goto retry;
 100         }
 101
 102         /* We take additional reference here. It will be put back by shrinker */
 103         atomic_set(&huge_zero_refcount, 2);
 104         preempt_enable();
 105         return true;
 106 }
 107
 108 static void put_huge_zero_page(void)
 109 {
 110         /*
 111          * Counter should never go to zero here. Only shrinker can put
 112          * last reference.
 113          */
 114         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 115 }
 116
 117 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 118 {
 119         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 120                 return READ_ONCE(huge_zero_page);
 121
 122         if (!get_huge_zero_page())
 123                 return NULL;
 124
 125         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 126                 put_huge_zero_page();
 127
 128         return READ_ONCE(huge_zero_page);
 129 }
 130
 131 void mm_put_huge_zero_page(struct mm_struct *mm)
 132 {
 133         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 134                 put_huge_zero_page();
 135 }
 136
 137 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 138                                         struct shrink_control *sc)
 139 {
 140         /* we can free zero page only if last reference remains */
 141         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 142 }
 143
 144 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 145                                        struct shrink_control *sc)
 146 {
 147         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 148                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 149                 BUG_ON(zero_page == NULL);
 150                 __free_pages(zero_page, compound_order(zero_page));
 151                 return HPAGE_PMD_NR;
 152         }
 153
 154         return 0;
 155 }
 156
 157 static struct shrinker huge_zero_page_shrinker = {
 158         .count_objects = shrink_huge_zero_page_count,
 159         .scan_objects = shrink_huge_zero_page_scan,
 160         .seeks = DEFAULT_SEEKS,
 161 };
 162
 163 #ifdef CONFIG_SYSFS
 164 static ssize_t enabled_show(struct kobject *kobj,
 165                             struct kobj_attribute *attr, char *buf)
 166 {
 167         const char *output;
 168
 169         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 170                 output = "[always] madvise never";
 171         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 172                           &transparent_hugepage_flags))
 173                 output = "always [madvise] never";
 174         else
 175                 output = "always madvise [never]";
 176
 177         return sysfs_emit(buf, "%s\n", output);
 178 }
 179
 180 static ssize_t enabled_store(struct kobject *kobj,
 181                              struct kobj_attribute *attr,
 182                              const char *buf, size_t count)
 183 {
 184         ssize_t ret = count;
 185
 186         if (sysfs_streq(buf, "always")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 188                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 189         } else if (sysfs_streq(buf, "madvise")) {
 190                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 191                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 192         } else if (sysfs_streq(buf, "never")) {
 193                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 194                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 195         } else
 196                 ret = -EINVAL;
 197
 198         if (ret > 0) {
 199                 int err = start_stop_khugepaged();
 200                 if (err)
 201                         ret = err;
 202         }
 203         return ret;
 204 }
 205 static struct kobj_attribute enabled_attr =
 206         __ATTR(enabled, 0644, enabled_show, enabled_store);
 207
 208 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 209                                   struct kobj_attribute *attr, char *buf,
 210                                   enum transparent_hugepage_flag flag)
 211 {
 212         return sysfs_emit(buf, "%d\n",
 213                           !!test_bit(flag, &transparent_hugepage_flags));
 214 }
 215
 216 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 217                                  struct kobj_attribute *attr,
 218                                  const char *buf, size_t count,
 219                                  enum transparent_hugepage_flag flag)
 220 {
 221         unsigned long value;
 222         int ret;
 223
 224         ret = kstrtoul(buf, 10, &value);
 225         if (ret < 0)
 226                 return ret;
 227         if (value > 1)
 228                 return -EINVAL;
 229
 230         if (value)
 231                 set_bit(flag, &transparent_hugepage_flags);
 232         else
 233                 clear_bit(flag, &transparent_hugepage_flags);
 234
 235         return count;
 236 }
 237
 238 static ssize_t defrag_show(struct kobject *kobj,
 239                            struct kobj_attribute *attr, char *buf)
 240 {
 241         const char *output;
 242
 243         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 244                      &transparent_hugepage_flags))
 245                 output = "[always] defer defer+madvise madvise never";
 246         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 247                           &transparent_hugepage_flags))
 248                 output = "always [defer] defer+madvise madvise never";
 249         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 250                           &transparent_hugepage_flags))
 251                 output = "always defer [defer+madvise] madvise never";
 252         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 253                           &transparent_hugepage_flags))
 254                 output = "always defer defer+madvise [madvise] never";
 255         else
 256                 output = "always defer defer+madvise madvise [never]";
 257
 258         return sysfs_emit(buf, "%s\n", output);
 259 }
 260
 261 static ssize_t defrag_store(struct kobject *kobj,
 262                             struct kobj_attribute *attr,
 263                             const char *buf, size_t count)
 264 {
 265         if (sysfs_streq(buf, "always")) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "defer+madvise")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 274                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 275         } else if (sysfs_streq(buf, "defer")) {
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 278                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 279                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 280         } else if (sysfs_streq(buf, "madvise")) {
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 282                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 283                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 284                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 285         } else if (sysfs_streq(buf, "never")) {
 286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 287                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 288                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 289                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 290         } else
 291                 return -EINVAL;
 292
 293         return count;
 294 }
 295 static struct kobj_attribute defrag_attr =
 296         __ATTR(defrag, 0644, defrag_show, defrag_store);
 297
 298 static ssize_t use_zero_page_show(struct kobject *kobj,
 299                                   struct kobj_attribute *attr, char *buf)
 300 {
 301         return single_hugepage_flag_show(kobj, attr, buf,
 302                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 303 }
 304 static ssize_t use_zero_page_store(struct kobject *kobj,
 305                 struct kobj_attribute *attr, const char *buf, size_t count)
 306 {
 307         return single_hugepage_flag_store(kobj, attr, buf, count,
 308                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 309 }
 310 static struct kobj_attribute use_zero_page_attr =
 311         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 312
 313 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 314                                    struct kobj_attribute *attr, char *buf)
 315 {
 316         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 317 }
 318 static struct kobj_attribute hpage_pmd_size_attr =
 319         __ATTR_RO(hpage_pmd_size);
 320
 321 static struct attribute *hugepage_attr[] = {
 322         &enabled_attr.attr,
 323         &defrag_attr.attr,
 324         &use_zero_page_attr.attr,
 325         &hpage_pmd_size_attr.attr,
 326 #ifdef CONFIG_SHMEM
 327         &shmem_enabled_attr.attr,
 328 #endif
 329         NULL,
 330 };
 331
 332 static const struct attribute_group hugepage_attr_group = {
 333         .attrs = hugepage_attr,
 334 };
 335
 336 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 337 {
 338         int err;
 339
 340         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 341         if (unlikely(!*hugepage_kobj)) {
 342                 pr_err("failed to create transparent hugepage kobject\n");
 343                 return -ENOMEM;
 344         }
 345
 346         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 347         if (err) {
 348                 pr_err("failed to register transparent hugepage group\n");
 349                 goto delete_obj;
 350         }
 351
 352         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 353         if (err) {
 354                 pr_err("failed to register transparent hugepage group\n");
 355                 goto remove_hp_group;
 356         }
 357
 358         return 0;
 359
 360 remove_hp_group:
 361         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 362 delete_obj:
 363         kobject_put(*hugepage_kobj);
 364         return err;
 365 }
 366
 367 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 368 {
 369         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 370         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 371         kobject_put(hugepage_kobj);
 372 }
 373 #else
 374 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 375 {
 376         return 0;
 377 }
 378
 379 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 380 {
 381 }
 382 #endif /* CONFIG_SYSFS */
 383
 384 static int __init hugepage_init(void)
 385 {
 386         int err;
 387         struct kobject *hugepage_kobj;
 388
 389         if (!has_transparent_hugepage()) {
 390                 /*
 391                  * Hardware doesn't support hugepages, hence disable
 392                  * DAX PMD support.
 393                  */
 394                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 395                 return -EINVAL;
 396         }
 397
 398         /*
 399          * hugepages can't be allocated by the buddy allocator
 400          */
 401         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 402         /*
 403          * we use page->mapping and page->index in second tail page
 404          * as list_head: assuming THP order >= 2
 405          */
 406         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 407
 408         err = hugepage_init_sysfs(&hugepage_kobj);
 409         if (err)
 410                 goto err_sysfs;
 411
 412         err = khugepaged_init();
 413         if (err)
 414                 goto err_slab;
 415
 416         err = register_shrinker(&huge_zero_page_shrinker);
 417         if (err)
 418                 goto err_hzp_shrinker;
 419         err = register_shrinker(&deferred_split_shrinker);
 420         if (err)
 421                 goto err_split_shrinker;
 422
 423         /*
 424          * By default disable transparent hugepages on smaller systems,
 425          * where the extra memory used could hurt more than TLB overhead
 426          * is likely to save.  The admin can still enable it through /sys.
 427          */
 428         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 429                 transparent_hugepage_flags = 0;
 430                 return 0;
 431         }
 432
 433         err = start_stop_khugepaged();
 434         if (err)
 435                 goto err_khugepaged;
 436
 437         return 0;
 438 err_khugepaged:
 439         unregister_shrinker(&deferred_split_shrinker);
 440 err_split_shrinker:
 441         unregister_shrinker(&huge_zero_page_shrinker);
 442 err_hzp_shrinker:
 443         khugepaged_destroy();
 444 err_slab:
 445         hugepage_exit_sysfs(hugepage_kobj);
 446 err_sysfs:
 447         return err;
 448 }
 449 subsys_initcall(hugepage_init);
 450
 451 static int __init setup_transparent_hugepage(char *str)
 452 {
 453         int ret = 0;
 454         if (!str)
 455                 goto out;
 456         if (!strcmp(str, "always")) {
 457                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 458                         &transparent_hugepage_flags);
 459                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 460                           &transparent_hugepage_flags);
 461                 ret = 1;
 462         } else if (!strcmp(str, "madvise")) {
 463                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 464                           &transparent_hugepage_flags);
 465                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 466                         &transparent_hugepage_flags);
 467                 ret = 1;
 468         } else if (!strcmp(str, "never")) {
 469                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 470                           &transparent_hugepage_flags);
 471                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 472                           &transparent_hugepage_flags);
 473                 ret = 1;
 474         }
 475 out:
 476         if (!ret)
 477                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 478         return ret;
 479 }
 480 __setup("transparent_hugepage=", setup_transparent_hugepage);
 481
 482 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 483 {
 484         if (likely(vma->vm_flags & VM_WRITE))
 485                 pmd = pmd_mkwrite(pmd);
 486         return pmd;
 487 }
 488
 489 #ifdef CONFIG_MEMCG
 490 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 491 {
 492         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 493         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 494
 495         if (memcg)
 496                 return &memcg->deferred_split_queue;
 497         else
 498                 return &pgdat->deferred_split_queue;
 499 }
 500 #else
 501 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 502 {
 503         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 504
 505         return &pgdat->deferred_split_queue;
 506 }
 507 #endif
 508
 509 void prep_transhuge_page(struct page *page)
 510 {
 511         /*
 512          * we use page->mapping and page->indexlru in second tail page
 513          * as list_head: assuming THP order >= 2
 514          */
 515
 516         INIT_LIST_HEAD(page_deferred_list(page));
 517         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 518 }
 519
 520 bool is_transparent_hugepage(struct page *page)
 521 {
 522         if (!PageCompound(page))
 523                 return false;
 524
 525         page = compound_head(page);
 526         return is_huge_zero_page(page) ||
 527                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 528 }
 529 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 530
 531 static unsigned long __thp_get_unmapped_area(struct file *filp,
 532                 unsigned long addr, unsigned long len,
 533                 loff_t off, unsigned long flags, unsigned long size)
 534 {
 535         loff_t off_end = off + len;
 536         loff_t off_align = round_up(off, size);
 537         unsigned long len_pad, ret;
 538
 539         if (off_end <= off_align || (off_end - off_align) < size)
 540                 return 0;
 541
 542         len_pad = len + size;
 543         if (len_pad < len || (off + len_pad) < off)
 544                 return 0;
 545
 546         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 547                                               off >> PAGE_SHIFT, flags);
 548
 549         /*
 550          * The failure might be due to length padding. The caller will retry
 551          * without the padding.
 552          */
 553         if (IS_ERR_VALUE(ret))
 554                 return 0;
 555
 556         /*
 557          * Do not try to align to THP boundary if allocation at the address
 558          * hint succeeds.
 559          */
 560         if (ret == addr)
 561                 return addr;
 562
 563         ret += (off - ret) & (size - 1);
 564         return ret;
 565 }
 566
 567 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 568                 unsigned long len, unsigned long pgoff, unsigned long flags)
 569 {
 570         unsigned long ret;
 571         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 572
 573         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 574                 goto out;
 575
 576         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 577         if (ret)
 578                 return ret;
 579 out:
 580         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 581 }
 582 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 583
 584 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 585                         struct page *page, gfp_t gfp)
 586 {
 587         struct vm_area_struct *vma = vmf->vma;
 588         pgtable_t pgtable;
 589         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 590         vm_fault_t ret = 0;
 591
 592         VM_BUG_ON_PAGE(!PageCompound(page), page);
 593
 594         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 595                 put_page(page);
 596                 count_vm_event(THP_FAULT_FALLBACK);
 597                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 598                 return VM_FAULT_FALLBACK;
 599         }
 600         cgroup_throttle_swaprate(page, gfp);
 601
 602         pgtable = pte_alloc_one(vma->vm_mm);
 603         if (unlikely(!pgtable)) {
 604                 ret = VM_FAULT_OOM;
 605                 goto release;
 606         }
 607
 608         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 609         /*
 610          * The memory barrier inside __SetPageUptodate makes sure that
 611          * clear_huge_page writes become visible before the set_pmd_at()
 612          * write.
 613          */
 614         __SetPageUptodate(page);
 615
 616         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 617         if (unlikely(!pmd_none(*vmf->pmd))) {
 618                 goto unlock_release;
 619         } else {
 620                 pmd_t entry;
 621
 622                 ret = check_stable_address_space(vma->vm_mm);
 623                 if (ret)
 624                         goto unlock_release;
 625
 626                 /* Deliver the page fault to userland */
 627                 if (userfaultfd_missing(vma)) {
 628                         spin_unlock(vmf->ptl);
 629                         put_page(page);
 630                         pte_free(vma->vm_mm, pgtable);
 631                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 632                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 633                         return ret;
 634                 }
 635
 636                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 637                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 638                 page_add_new_anon_rmap(page, vma, haddr, true);
 639                 lru_cache_add_inactive_or_unevictable(page, vma);
 640                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 641                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 642                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 643                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 644                 mm_inc_nr_ptes(vma->vm_mm);
 645                 spin_unlock(vmf->ptl);
 646                 count_vm_event(THP_FAULT_ALLOC);
 647                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 648         }
 649
 650         return 0;
 651 unlock_release:
 652         spin_unlock(vmf->ptl);
 653 release:
 654         if (pgtable)
 655                 pte_free(vma->vm_mm, pgtable);
 656         put_page(page);
 657         return ret;
 658
 659 }
 660
 661 /*
 662  * always: directly stall for all thp allocations
 663  * defer: wake kswapd and fail if not immediately available
 664  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 665  *                fail if not immediately available
 666  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 667  *          available
 668  * never: never stall for any thp allocation
 669  */
 670 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 671 {
 672         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 673
 674         /* Always do synchronous compaction */
 675         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 676                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 677
 678         /* Kick kcompactd and fail quickly */
 679         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 680                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 681
 682         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 683         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 684                 return GFP_TRANSHUGE_LIGHT |
 685                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 686                                         __GFP_KSWAPD_RECLAIM);
 687
 688         /* Only do synchronous compaction if madvised */
 689         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 690                 return GFP_TRANSHUGE_LIGHT |
 691                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 692
 693         return GFP_TRANSHUGE_LIGHT;
 694 }
 695
 696 /* Caller must hold page table lock. */
 697 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 698                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 699                 struct page *zero_page)
 700 {
 701         pmd_t entry;
 702         if (!pmd_none(*pmd))
 703                 return;
 704         entry = mk_pmd(zero_page, vma->vm_page_prot);
 705         entry = pmd_mkhuge(entry);
 706         if (pgtable)
 707                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 708         set_pmd_at(mm, haddr, pmd, entry);
 709         mm_inc_nr_ptes(mm);
 710 }
 711
 712 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 713 {
 714         struct vm_area_struct *vma = vmf->vma;
 715         gfp_t gfp;
 716         struct page *page;
 717         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 718
 719         if (!transhuge_vma_suitable(vma, haddr))
 720                 return VM_FAULT_FALLBACK;
 721         if (unlikely(anon_vma_prepare(vma)))
 722                 return VM_FAULT_OOM;
 723         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 724                 return VM_FAULT_OOM;
 725         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 726                         !mm_forbids_zeropage(vma->vm_mm) &&
 727                         transparent_hugepage_use_zero_page()) {
 728                 pgtable_t pgtable;
 729                 struct page *zero_page;
 730                 vm_fault_t ret;
 731                 pgtable = pte_alloc_one(vma->vm_mm);
 732                 if (unlikely(!pgtable))
 733                         return VM_FAULT_OOM;
 734                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 735                 if (unlikely(!zero_page)) {
 736                         pte_free(vma->vm_mm, pgtable);
 737                         count_vm_event(THP_FAULT_FALLBACK);
 738                         return VM_FAULT_FALLBACK;
 739                 }
 740                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 741                 ret = 0;
 742                 if (pmd_none(*vmf->pmd)) {
 743                         ret = check_stable_address_space(vma->vm_mm);
 744                         if (ret) {
 745                                 spin_unlock(vmf->ptl);
 746                                 pte_free(vma->vm_mm, pgtable);
 747                         } else if (userfaultfd_missing(vma)) {
 748                                 spin_unlock(vmf->ptl);
 749                                 pte_free(vma->vm_mm, pgtable);
 750                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 751                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 752                         } else {
 753                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 754                                                    haddr, vmf->pmd, zero_page);
 755                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 756                                 spin_unlock(vmf->ptl);
 757                         }
 758                 } else {
 759                         spin_unlock(vmf->ptl);
 760                         pte_free(vma->vm_mm, pgtable);
 761                 }
 762                 return ret;
 763         }
 764         gfp = vma_thp_gfp_mask(vma);
 765         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 766         if (unlikely(!page)) {
 767                 count_vm_event(THP_FAULT_FALLBACK);
 768                 return VM_FAULT_FALLBACK;
 769         }
 770         prep_transhuge_page(page);
 771         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 772 }
 773
 774 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 775                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 776                 pgtable_t pgtable)
 777 {
 778         struct mm_struct *mm = vma->vm_mm;
 779         pmd_t entry;
 780         spinlock_t *ptl;
 781
 782         ptl = pmd_lock(mm, pmd);
 783         if (!pmd_none(*pmd)) {
 784                 if (write) {
 785                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 786                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 787                                 goto out_unlock;
 788                         }
 789                         entry = pmd_mkyoung(*pmd);
 790                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 791                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 792                                 update_mmu_cache_pmd(vma, addr, pmd);
 793                 }
 794
 795                 goto out_unlock;
 796         }
 797
 798         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 799         if (pfn_t_devmap(pfn))
 800                 entry = pmd_mkdevmap(entry);
 801         if (write) {
 802                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 803                 entry = maybe_pmd_mkwrite(entry, vma);
 804         }
 805
 806         if (pgtable) {
 807                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 808                 mm_inc_nr_ptes(mm);
 809                 pgtable = NULL;
 810         }
 811
 812         set_pmd_at(mm, addr, pmd, entry);
 813         update_mmu_cache_pmd(vma, addr, pmd);
 814
 815 out_unlock:
 816         spin_unlock(ptl);
 817         if (pgtable)
 818                 pte_free(mm, pgtable);
 819 }
 820
 821 /**
 822  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 823  * @vmf: Structure describing the fault
 824  * @pfn: pfn to insert
 825  * @pgprot: page protection to use
 826  * @write: whether it's a write fault
 827  *
 828  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 829  * also consult the vmf_insert_mixed_prot() documentation when
 830  * @pgprot != @vmf->vma->vm_page_prot.
 831  *
 832  * Return: vm_fault_t value.
 833  */
 834 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 835                                    pgprot_t pgprot, bool write)
 836 {
 837         unsigned long addr = vmf->address & PMD_MASK;
 838         struct vm_area_struct *vma = vmf->vma;
 839         pgtable_t pgtable = NULL;
 840
 841         /*
 842          * If we had pmd_special, we could avoid all these restrictions,
 843          * but we need to be consistent with PTEs and architectures that
 844          * can't support a 'special' bit.
 845          */
 846         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 847                         !pfn_t_devmap(pfn));
 848         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 849                                                 (VM_PFNMAP|VM_MIXEDMAP));
 850         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 851
 852         if (addr < vma->vm_start || addr >= vma->vm_end)
 853                 return VM_FAULT_SIGBUS;
 854
 855         if (arch_needs_pgtable_deposit()) {
 856                 pgtable = pte_alloc_one(vma->vm_mm);
 857                 if (!pgtable)
 858                         return VM_FAULT_OOM;
 859         }
 860
 861         track_pfn_insert(vma, &pgprot, pfn);
 862
 863         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 864         return VM_FAULT_NOPAGE;
 865 }
 866 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 867
 868 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 869 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 870 {
 871         if (likely(vma->vm_flags & VM_WRITE))
 872                 pud = pud_mkwrite(pud);
 873         return pud;
 874 }
 875
 876 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 877                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 878 {
 879         struct mm_struct *mm = vma->vm_mm;
 880         pud_t entry;
 881         spinlock_t *ptl;
 882
 883         ptl = pud_lock(mm, pud);
 884         if (!pud_none(*pud)) {
 885                 if (write) {
 886                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 887                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 888                                 goto out_unlock;
 889                         }
 890                         entry = pud_mkyoung(*pud);
 891                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 892                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 893                                 update_mmu_cache_pud(vma, addr, pud);
 894                 }
 895                 goto out_unlock;
 896         }
 897
 898         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 899         if (pfn_t_devmap(pfn))
 900                 entry = pud_mkdevmap(entry);
 901         if (write) {
 902                 entry = pud_mkyoung(pud_mkdirty(entry));
 903                 entry = maybe_pud_mkwrite(entry, vma);
 904         }
 905         set_pud_at(mm, addr, pud, entry);
 906         update_mmu_cache_pud(vma, addr, pud);
 907
 908 out_unlock:
 909         spin_unlock(ptl);
 910 }
 911
 912 /**
 913  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 914  * @vmf: Structure describing the fault
 915  * @pfn: pfn to insert
 916  * @pgprot: page protection to use
 917  * @write: whether it's a write fault
 918  *
 919  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 920  * also consult the vmf_insert_mixed_prot() documentation when
 921  * @pgprot != @vmf->vma->vm_page_prot.
 922  *
 923  * Return: vm_fault_t value.
 924  */
 925 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 926                                    pgprot_t pgprot, bool write)
 927 {
 928         unsigned long addr = vmf->address & PUD_MASK;
 929         struct vm_area_struct *vma = vmf->vma;
 930
 931         /*
 932          * If we had pud_special, we could avoid all these restrictions,
 933          * but we need to be consistent with PTEs and architectures that
 934          * can't support a 'special' bit.
 935          */
 936         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 937                         !pfn_t_devmap(pfn));
 938         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 939                                                 (VM_PFNMAP|VM_MIXEDMAP));
 940         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 941
 942         if (addr < vma->vm_start || addr >= vma->vm_end)
 943                 return VM_FAULT_SIGBUS;
 944
 945         track_pfn_insert(vma, &pgprot, pfn);
 946
 947         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 948         return VM_FAULT_NOPAGE;
 949 }
 950 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 951 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 952
 953 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 954                 pmd_t *pmd, int flags)
 955 {
 956         pmd_t _pmd;
 957
 958         _pmd = pmd_mkyoung(*pmd);
 959         if (flags & FOLL_WRITE)
 960                 _pmd = pmd_mkdirty(_pmd);
 961         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 962                                 pmd, _pmd, flags & FOLL_WRITE))
 963                 update_mmu_cache_pmd(vma, addr, pmd);
 964 }
 965
 966 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 967                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 968 {
 969         unsigned long pfn = pmd_pfn(*pmd);
 970         struct mm_struct *mm = vma->vm_mm;
 971         struct page *page;
 972
 973         assert_spin_locked(pmd_lockptr(mm, pmd));
 974
 975         /*
 976          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 977          * not be in this function with `flags & FOLL_COW` set.
 978          */
 979         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 980
 981         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 982         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 983                          (FOLL_PIN | FOLL_GET)))
 984                 return NULL;
 985
 986         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 987                 return NULL;
 988
 989         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 990                 /* pass */;
 991         else
 992                 return NULL;
 993
 994         if (flags & FOLL_TOUCH)
 995                 touch_pmd(vma, addr, pmd, flags);
 996
 997         /*
 998          * device mapped pages can only be returned if the
 999          * caller will manage the page reference count.
1000          */
1001         if (!(flags & (FOLL_GET | FOLL_PIN)))
1002                 return ERR_PTR(-EEXIST);
1003
1004         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1005         *pgmap = get_dev_pagemap(pfn, *pgmap);
1006         if (!*pgmap)
1007                 return ERR_PTR(-EFAULT);
1008         page = pfn_to_page(pfn);
1009         if (!try_grab_page(page, flags))
1010                 page = ERR_PTR(-ENOMEM);
1011
1012         return page;
1013 }
1014
1015 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1016                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1017                   struct vm_area_struct *vma)
1018 {
1019         spinlock_t *dst_ptl, *src_ptl;
1020         struct page *src_page;
1021         pmd_t pmd;
1022         pgtable_t pgtable = NULL;
1023         int ret = -ENOMEM;
1024
1025         /* Skip if can be re-fill on fault */
1026         if (!vma_is_anonymous(vma))
1027                 return 0;
1028
1029         pgtable = pte_alloc_one(dst_mm);
1030         if (unlikely(!pgtable))
1031                 goto out;
1032
1033         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1034         src_ptl = pmd_lockptr(src_mm, src_pmd);
1035         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1036
1037         ret = -EAGAIN;
1038         pmd = *src_pmd;
1039
1040         /*
1041          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1042          * does not have the VM_UFFD_WP, which means that the uffd
1043          * fork event is not enabled.
1044          */
1045         if (!(vma->vm_flags & VM_UFFD_WP))
1046                 pmd = pmd_clear_uffd_wp(pmd);
1047
1048 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1049         if (unlikely(is_swap_pmd(pmd))) {
1050                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1051
1052                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1053                 if (is_write_migration_entry(entry)) {
1054                         make_migration_entry_read(&entry);
1055                         pmd = swp_entry_to_pmd(entry);
1056                         if (pmd_swp_soft_dirty(*src_pmd))
1057                                 pmd = pmd_swp_mksoft_dirty(pmd);
1058                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1059                 }
1060                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1061                 mm_inc_nr_ptes(dst_mm);
1062                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1063                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1064                 ret = 0;
1065                 goto out_unlock;
1066         }
1067 #endif
1068
1069         if (unlikely(!pmd_trans_huge(pmd))) {
1070                 pte_free(dst_mm, pgtable);
1071                 goto out_unlock;
1072         }
1073         /*
1074          * When page table lock is held, the huge zero pmd should not be
1075          * under splitting since we don't split the page itself, only pmd to
1076          * a page table.
1077          */
1078         if (is_huge_zero_pmd(pmd)) {
1079                 struct page *zero_page;
1080                 /*
1081                  * get_huge_zero_page() will never allocate a new page here,
1082                  * since we already have a zero page to copy. It just takes a
1083                  * reference.
1084                  */
1085                 zero_page = mm_get_huge_zero_page(dst_mm);
1086                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1087                                 zero_page);
1088                 ret = 0;
1089                 goto out_unlock;
1090         }
1091
1092         src_page = pmd_page(pmd);
1093         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1094
1095         /*
1096          * If this page is a potentially pinned page, split and retry the fault
1097          * with smaller page size.  Normally this should not happen because the
1098          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1099          * best effort that the pinned pages won't be replaced by another
1100          * random page during the coming copy-on-write.
1101          */
1102         if (unlikely(page_needs_cow_for_dma(vma, src_page))) {
1103                 pte_free(dst_mm, pgtable);
1104                 spin_unlock(src_ptl);
1105                 spin_unlock(dst_ptl);
1106                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1107                 return -EAGAIN;
1108         }
1109
1110         get_page(src_page);
1111         page_dup_rmap(src_page, true);
1112         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1113         mm_inc_nr_ptes(dst_mm);
1114         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1115
1116         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1117         pmd = pmd_mkold(pmd_wrprotect(pmd));
1118         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1119
1120         ret = 0;
1121 out_unlock:
1122         spin_unlock(src_ptl);
1123         spin_unlock(dst_ptl);
1124 out:
1125         return ret;
1126 }
1127
1128 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1129 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1130                 pud_t *pud, int flags)
1131 {
1132         pud_t _pud;
1133
1134         _pud = pud_mkyoung(*pud);
1135         if (flags & FOLL_WRITE)
1136                 _pud = pud_mkdirty(_pud);
1137         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1138                                 pud, _pud, flags & FOLL_WRITE))
1139                 update_mmu_cache_pud(vma, addr, pud);
1140 }
1141
1142 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1143                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1144 {
1145         unsigned long pfn = pud_pfn(*pud);
1146         struct mm_struct *mm = vma->vm_mm;
1147         struct page *page;
1148
1149         assert_spin_locked(pud_lockptr(mm, pud));
1150
1151         if (flags & FOLL_WRITE && !pud_write(*pud))
1152                 return NULL;
1153
1154         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1155         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1156                          (FOLL_PIN | FOLL_GET)))
1157                 return NULL;
1158
1159         if (pud_present(*pud) && pud_devmap(*pud))
1160                 /* pass */;
1161         else
1162                 return NULL;
1163
1164         if (flags & FOLL_TOUCH)
1165                 touch_pud(vma, addr, pud, flags);
1166
1167         /*
1168          * device mapped pages can only be returned if the
1169          * caller will manage the page reference count.
1170          *
1171          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1172          */
1173         if (!(flags & (FOLL_GET | FOLL_PIN)))
1174                 return ERR_PTR(-EEXIST);
1175
1176         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1177         *pgmap = get_dev_pagemap(pfn, *pgmap);
1178         if (!*pgmap)
1179                 return ERR_PTR(-EFAULT);
1180         page = pfn_to_page(pfn);
1181         if (!try_grab_page(page, flags))
1182                 page = ERR_PTR(-ENOMEM);
1183
1184         return page;
1185 }
1186
1187 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1188                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1189                   struct vm_area_struct *vma)
1190 {
1191         spinlock_t *dst_ptl, *src_ptl;
1192         pud_t pud;
1193         int ret;
1194
1195         dst_ptl = pud_lock(dst_mm, dst_pud);
1196         src_ptl = pud_lockptr(src_mm, src_pud);
1197         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1198
1199         ret = -EAGAIN;
1200         pud = *src_pud;
1201         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1202                 goto out_unlock;
1203
1204         /*
1205          * When page table lock is held, the huge zero pud should not be
1206          * under splitting since we don't split the page itself, only pud to
1207          * a page table.
1208          */
1209         if (is_huge_zero_pud(pud)) {
1210                 /* No huge zero pud yet */
1211         }
1212
1213         /* Please refer to comments in copy_huge_pmd() */
1214         if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1215                 spin_unlock(src_ptl);
1216                 spin_unlock(dst_ptl);
1217                 __split_huge_pud(vma, src_pud, addr);
1218                 return -EAGAIN;
1219         }
1220
1221         pudp_set_wrprotect(src_mm, addr, src_pud);
1222         pud = pud_mkold(pud_wrprotect(pud));
1223         set_pud_at(dst_mm, addr, dst_pud, pud);
1224
1225         ret = 0;
1226 out_unlock:
1227         spin_unlock(src_ptl);
1228         spin_unlock(dst_ptl);
1229         return ret;
1230 }
1231
1232 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1233 {
1234         pud_t entry;
1235         unsigned long haddr;
1236         bool write = vmf->flags & FAULT_FLAG_WRITE;
1237
1238         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1239         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1240                 goto unlock;
1241
1242         entry = pud_mkyoung(orig_pud);
1243         if (write)
1244                 entry = pud_mkdirty(entry);
1245         haddr = vmf->address & HPAGE_PUD_MASK;
1246         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1247                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1248
1249 unlock:
1250         spin_unlock(vmf->ptl);
1251 }
1252 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1253
1254 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1255 {
1256         pmd_t entry;
1257         unsigned long haddr;
1258         bool write = vmf->flags & FAULT_FLAG_WRITE;
1259
1260         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1261         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1262                 goto unlock;
1263
1264         entry = pmd_mkyoung(orig_pmd);
1265         if (write)
1266                 entry = pmd_mkdirty(entry);
1267         haddr = vmf->address & HPAGE_PMD_MASK;
1268         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1269                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1270
1271 unlock:
1272         spin_unlock(vmf->ptl);
1273 }
1274
1275 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1276 {
1277         struct vm_area_struct *vma = vmf->vma;
1278         struct page *page;
1279         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1280
1281         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1282         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1283
1284         if (is_huge_zero_pmd(orig_pmd))
1285                 goto fallback;
1286
1287         spin_lock(vmf->ptl);
1288
1289         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1290                 spin_unlock(vmf->ptl);
1291                 return 0;
1292         }
1293
1294         page = pmd_page(orig_pmd);
1295         VM_BUG_ON_PAGE(!PageHead(page), page);
1296
1297         /* Lock page for reuse_swap_page() */
1298         if (!trylock_page(page)) {
1299                 get_page(page);
1300                 spin_unlock(vmf->ptl);
1301                 lock_page(page);
1302                 spin_lock(vmf->ptl);
1303                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1304                         spin_unlock(vmf->ptl);
1305                         unlock_page(page);
1306                         put_page(page);
1307                         return 0;
1308                 }
1309                 put_page(page);
1310         }
1311
1312         /*
1313          * We can only reuse the page if nobody else maps the huge page or it's
1314          * part.
1315          */
1316         if (reuse_swap_page(page, NULL)) {
1317                 pmd_t entry;
1318                 entry = pmd_mkyoung(orig_pmd);
1319                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1320                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1321                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1322                 unlock_page(page);
1323                 spin_unlock(vmf->ptl);
1324                 return VM_FAULT_WRITE;
1325         }
1326
1327         unlock_page(page);
1328         spin_unlock(vmf->ptl);
1329 fallback:
1330         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1331         return VM_FAULT_FALLBACK;
1332 }
1333
1334 /*
1335  * FOLL_FORCE can write to even unwritable pmd's, but only
1336  * after we've gone through a COW cycle and they are dirty.
1337  */
1338 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1339 {
1340         return pmd_write(pmd) ||
1341                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1342 }
1343
1344 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1345                                    unsigned long addr,
1346                                    pmd_t *pmd,
1347                                    unsigned int flags)
1348 {
1349         struct mm_struct *mm = vma->vm_mm;
1350         struct page *page = NULL;
1351
1352         assert_spin_locked(pmd_lockptr(mm, pmd));
1353
1354         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1355                 goto out;
1356
1357         /* Avoid dumping huge zero page */
1358         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1359                 return ERR_PTR(-EFAULT);
1360
1361         /* Full NUMA hinting faults to serialise migration in fault paths */
1362         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1363                 goto out;
1364
1365         page = pmd_page(*pmd);
1366         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1367
1368         if (!try_grab_page(page, flags))
1369                 return ERR_PTR(-ENOMEM);
1370
1371         if (flags & FOLL_TOUCH)
1372                 touch_pmd(vma, addr, pmd, flags);
1373
1374         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1375                 /*
1376                  * We don't mlock() pte-mapped THPs. This way we can avoid
1377                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1378                  *
1379                  * For anon THP:
1380                  *
1381                  * In most cases the pmd is the only mapping of the page as we
1382                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1383                  * writable private mappings in populate_vma_page_range().
1384                  *
1385                  * The only scenario when we have the page shared here is if we
1386                  * mlocking read-only mapping shared over fork(). We skip
1387                  * mlocking such pages.
1388                  *
1389                  * For file THP:
1390                  *
1391                  * We can expect PageDoubleMap() to be stable under page lock:
1392                  * for file pages we set it in page_add_file_rmap(), which
1393                  * requires page to be locked.
1394                  */
1395
1396                 if (PageAnon(page) && compound_mapcount(page) != 1)
1397                         goto skip_mlock;
1398                 if (PageDoubleMap(page) || !page->mapping)
1399                         goto skip_mlock;
1400                 if (!trylock_page(page))
1401                         goto skip_mlock;
1402                 if (page->mapping && !PageDoubleMap(page))
1403                         mlock_vma_page(page);
1404                 unlock_page(page);
1405         }
1406 skip_mlock:
1407         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1408         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1409
1410 out:
1411         return page;
1412 }
1413
1414 /* NUMA hinting page fault entry point for trans huge pmds */
1415 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1416 {
1417         struct vm_area_struct *vma = vmf->vma;
1418         struct anon_vma *anon_vma = NULL;
1419         struct page *page;
1420         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1421         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1422         int target_nid, last_cpupid = -1;
1423         bool page_locked;
1424         bool migrated = false;
1425         bool was_writable;
1426         int flags = 0;
1427
1428         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1429         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1430                 goto out_unlock;
1431
1432         /*
1433          * If there are potential migrations, wait for completion and retry
1434          * without disrupting NUMA hinting information. Do not relock and
1435          * check_same as the page may no longer be mapped.
1436          */
1437         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1438                 page = pmd_page(*vmf->pmd);
1439                 if (!get_page_unless_zero(page))
1440                         goto out_unlock;
1441                 spin_unlock(vmf->ptl);
1442                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1443                 goto out;
1444         }
1445
1446         page = pmd_page(pmd);
1447         BUG_ON(is_huge_zero_page(page));
1448         page_nid = page_to_nid(page);
1449         last_cpupid = page_cpupid_last(page);
1450         count_vm_numa_event(NUMA_HINT_FAULTS);
1451         if (page_nid == this_nid) {
1452                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1453                 flags |= TNF_FAULT_LOCAL;
1454         }
1455
1456         /* See similar comment in do_numa_page for explanation */
1457         if (!pmd_savedwrite(pmd))
1458                 flags |= TNF_NO_GROUP;
1459
1460         /*
1461          * Acquire the page lock to serialise THP migrations but avoid dropping
1462          * page_table_lock if at all possible
1463          */
1464         page_locked = trylock_page(page);
1465         target_nid = mpol_misplaced(page, vma, haddr);
1466         /* Migration could have started since the pmd_trans_migrating check */
1467         if (!page_locked) {
1468                 page_nid = NUMA_NO_NODE;
1469                 if (!get_page_unless_zero(page))
1470                         goto out_unlock;
1471                 spin_unlock(vmf->ptl);
1472                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1473                 goto out;
1474         } else if (target_nid == NUMA_NO_NODE) {
1475                 /* There are no parallel migrations and page is in the right
1476                  * node. Clear the numa hinting info in this pmd.
1477                  */
1478                 goto clear_pmdnuma;
1479         }
1480
1481         /*
1482          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1483          * to serialises splits
1484          */
1485         get_page(page);
1486         spin_unlock(vmf->ptl);
1487         anon_vma = page_lock_anon_vma_read(page);
1488
1489         /* Confirm the PMD did not change while page_table_lock was released */
1490         spin_lock(vmf->ptl);
1491         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1492                 unlock_page(page);
1493                 put_page(page);
1494                 page_nid = NUMA_NO_NODE;
1495                 goto out_unlock;
1496         }
1497
1498         /* Bail if we fail to protect against THP splits for any reason */
1499         if (unlikely(!anon_vma)) {
1500                 put_page(page);
1501                 page_nid = NUMA_NO_NODE;
1502                 goto clear_pmdnuma;
1503         }
1504
1505         /*
1506          * Since we took the NUMA fault, we must have observed the !accessible
1507          * bit. Make sure all other CPUs agree with that, to avoid them
1508          * modifying the page we're about to migrate.
1509          *
1510          * Must be done under PTL such that we'll observe the relevant
1511          * inc_tlb_flush_pending().
1512          *
1513          * We are not sure a pending tlb flush here is for a huge page
1514          * mapping or not. Hence use the tlb range variant
1515          */
1516         if (mm_tlb_flush_pending(vma->vm_mm)) {
1517                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1518                 /*
1519                  * change_huge_pmd() released the pmd lock before
1520                  * invalidating the secondary MMUs sharing the primary
1521                  * MMU pagetables (with ->invalidate_range()). The
1522                  * mmu_notifier_invalidate_range_end() (which
1523                  * internally calls ->invalidate_range()) in
1524                  * change_pmd_range() will run after us, so we can't
1525                  * rely on it here and we need an explicit invalidate.
1526                  */
1527                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1528                                               haddr + HPAGE_PMD_SIZE);
1529         }
1530
1531         /*
1532          * Migrate the THP to the requested node, returns with page unlocked
1533          * and access rights restored.
1534          */
1535         spin_unlock(vmf->ptl);
1536
1537         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1538                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1539         if (migrated) {
1540                 flags |= TNF_MIGRATED;
1541                 page_nid = target_nid;
1542         } else
1543                 flags |= TNF_MIGRATE_FAIL;
1544
1545         goto out;
1546 clear_pmdnuma:
1547         BUG_ON(!PageLocked(page));
1548         was_writable = pmd_savedwrite(pmd);
1549         pmd = pmd_modify(pmd, vma->vm_page_prot);
1550         pmd = pmd_mkyoung(pmd);
1551         if (was_writable)
1552                 pmd = pmd_mkwrite(pmd);
1553         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1554         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1555         unlock_page(page);
1556 out_unlock:
1557         spin_unlock(vmf->ptl);
1558
1559 out:
1560         if (anon_vma)
1561                 page_unlock_anon_vma_read(anon_vma);
1562
1563         if (page_nid != NUMA_NO_NODE)
1564                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1565                                 flags);
1566
1567         return 0;
1568 }
1569
1570 /*
1571  * Return true if we do MADV_FREE successfully on entire pmd page.
1572  * Otherwise, return false.
1573  */
1574 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1575                 pmd_t *pmd, unsigned long addr, unsigned long next)
1576 {
1577         spinlock_t *ptl;
1578         pmd_t orig_pmd;
1579         struct page *page;
1580         struct mm_struct *mm = tlb->mm;
1581         bool ret = false;
1582
1583         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1584
1585         ptl = pmd_trans_huge_lock(pmd, vma);
1586         if (!ptl)
1587                 goto out_unlocked;
1588
1589         orig_pmd = *pmd;
1590         if (is_huge_zero_pmd(orig_pmd))
1591                 goto out;
1592
1593         if (unlikely(!pmd_present(orig_pmd))) {
1594                 VM_BUG_ON(thp_migration_supported() &&
1595                                   !is_pmd_migration_entry(orig_pmd));
1596                 goto out;
1597         }
1598
1599         page = pmd_page(orig_pmd);
1600         /*
1601          * If other processes are mapping this page, we couldn't discard
1602          * the page unless they all do MADV_FREE so let's skip the page.
1603          */
1604         if (page_mapcount(page) != 1)
1605                 goto out;
1606
1607         if (!trylock_page(page))
1608                 goto out;
1609
1610         /*
1611          * If user want to discard part-pages of THP, split it so MADV_FREE
1612          * will deactivate only them.
1613          */
1614         if (next - addr != HPAGE_PMD_SIZE) {
1615                 get_page(page);
1616                 spin_unlock(ptl);
1617                 split_huge_page(page);
1618                 unlock_page(page);
1619                 put_page(page);
1620                 goto out_unlocked;
1621         }
1622
1623         if (PageDirty(page))
1624                 ClearPageDirty(page);
1625         unlock_page(page);
1626
1627         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1628                 pmdp_invalidate(vma, addr, pmd);
1629                 orig_pmd = pmd_mkold(orig_pmd);
1630                 orig_pmd = pmd_mkclean(orig_pmd);
1631
1632                 set_pmd_at(mm, addr, pmd, orig_pmd);
1633                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1634         }
1635
1636         mark_page_lazyfree(page);
1637         ret = true;
1638 out:
1639         spin_unlock(ptl);
1640 out_unlocked:
1641         return ret;
1642 }
1643
1644 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1645 {
1646         pgtable_t pgtable;
1647
1648         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1649         pte_free(mm, pgtable);
1650         mm_dec_nr_ptes(mm);
1651 }
1652
1653 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1654                  pmd_t *pmd, unsigned long addr)
1655 {
1656         pmd_t orig_pmd;
1657         spinlock_t *ptl;
1658
1659         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1660
1661         ptl = __pmd_trans_huge_lock(pmd, vma);
1662         if (!ptl)
1663                 return 0;
1664         /*
1665          * For architectures like ppc64 we look at deposited pgtable
1666          * when calling pmdp_huge_get_and_clear. So do the
1667          * pgtable_trans_huge_withdraw after finishing pmdp related
1668          * operations.
1669          */
1670         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1671                                                 tlb->fullmm);
1672         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1673         if (vma_is_special_huge(vma)) {
1674                 if (arch_needs_pgtable_deposit())
1675                         zap_deposited_table(tlb->mm, pmd);
1676                 spin_unlock(ptl);
1677                 if (is_huge_zero_pmd(orig_pmd))
1678                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1679         } else if (is_huge_zero_pmd(orig_pmd)) {
1680                 zap_deposited_table(tlb->mm, pmd);
1681                 spin_unlock(ptl);
1682                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1683         } else {
1684                 struct page *page = NULL;
1685                 int flush_needed = 1;
1686
1687                 if (pmd_present(orig_pmd)) {
1688                         page = pmd_page(orig_pmd);
1689                         page_remove_rmap(page, true);
1690                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1691                         VM_BUG_ON_PAGE(!PageHead(page), page);
1692                 } else if (thp_migration_supported()) {
1693                         swp_entry_t entry;
1694
1695                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1696                         entry = pmd_to_swp_entry(orig_pmd);
1697                         page = migration_entry_to_page(entry);
1698                         flush_needed = 0;
1699                 } else
1700                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1701
1702                 if (PageAnon(page)) {
1703                         zap_deposited_table(tlb->mm, pmd);
1704                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1705                 } else {
1706                         if (arch_needs_pgtable_deposit())
1707                                 zap_deposited_table(tlb->mm, pmd);
1708                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1709                 }
1710
1711                 spin_unlock(ptl);
1712                 if (flush_needed)
1713                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1714         }
1715         return 1;
1716 }
1717
1718 #ifndef pmd_move_must_withdraw
1719 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1720                                          spinlock_t *old_pmd_ptl,
1721                                          struct vm_area_struct *vma)
1722 {
1723         /*
1724          * With split pmd lock we also need to move preallocated
1725          * PTE page table if new_pmd is on different PMD page table.
1726          *
1727          * We also don't deposit and withdraw tables for file pages.
1728          */
1729         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1730 }
1731 #endif
1732
1733 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1734 {
1735 #ifdef CONFIG_MEM_SOFT_DIRTY
1736         if (unlikely(is_pmd_migration_entry(pmd)))
1737                 pmd = pmd_swp_mksoft_dirty(pmd);
1738         else if (pmd_present(pmd))
1739                 pmd = pmd_mksoft_dirty(pmd);
1740 #endif
1741         return pmd;
1742 }
1743
1744 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1745                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1746 {
1747         spinlock_t *old_ptl, *new_ptl;
1748         pmd_t pmd;
1749         struct mm_struct *mm = vma->vm_mm;
1750         bool force_flush = false;
1751
1752         /*
1753          * The destination pmd shouldn't be established, free_pgtables()
1754          * should have release it.
1755          */
1756         if (WARN_ON(!pmd_none(*new_pmd))) {
1757                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1758                 return false;
1759         }
1760
1761         /*
1762          * We don't have to worry about the ordering of src and dst
1763          * ptlocks because exclusive mmap_lock prevents deadlock.
1764          */
1765         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1766         if (old_ptl) {
1767                 new_ptl = pmd_lockptr(mm, new_pmd);
1768                 if (new_ptl != old_ptl)
1769                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1770                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1771                 if (pmd_present(pmd))
1772                         force_flush = true;
1773                 VM_BUG_ON(!pmd_none(*new_pmd));
1774
1775                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1776                         pgtable_t pgtable;
1777                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1778                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1779                 }
1780                 pmd = move_soft_dirty_pmd(pmd);
1781                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1782                 if (force_flush)
1783                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1784                 if (new_ptl != old_ptl)
1785                         spin_unlock(new_ptl);
1786                 spin_unlock(old_ptl);
1787                 return true;
1788         }
1789         return false;
1790 }
1791
1792 /*
1793  * Returns
1794  *  - 0 if PMD could not be locked
1795  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1796  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1797  */
1798 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1799                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1800 {
1801         struct mm_struct *mm = vma->vm_mm;
1802         spinlock_t *ptl;
1803         pmd_t entry;
1804         bool preserve_write;
1805         int ret;
1806         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1807         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1808         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1809
1810         ptl = __pmd_trans_huge_lock(pmd, vma);
1811         if (!ptl)
1812                 return 0;
1813
1814         preserve_write = prot_numa && pmd_write(*pmd);
1815         ret = 1;
1816
1817 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1818         if (is_swap_pmd(*pmd)) {
1819                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1820
1821                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1822                 if (is_write_migration_entry(entry)) {
1823                         pmd_t newpmd;
1824                         /*
1825                          * A protection check is difficult so
1826                          * just be safe and disable write
1827                          */
1828                         make_migration_entry_read(&entry);
1829                         newpmd = swp_entry_to_pmd(entry);
1830                         if (pmd_swp_soft_dirty(*pmd))
1831                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1832                         set_pmd_at(mm, addr, pmd, newpmd);
1833                 }
1834                 goto unlock;
1835         }
1836 #endif
1837
1838         /*
1839          * Avoid trapping faults against the zero page. The read-only
1840          * data is likely to be read-cached on the local CPU and
1841          * local/remote hits to the zero page are not interesting.
1842          */
1843         if (prot_numa && is_huge_zero_pmd(*pmd))
1844                 goto unlock;
1845
1846         if (prot_numa && pmd_protnone(*pmd))
1847                 goto unlock;
1848
1849         /*
1850          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1851          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1852          * which is also under mmap_read_lock(mm):
1853          *
1854          *      CPU0:                           CPU1:
1855          *                              change_huge_pmd(prot_numa=1)
1856          *                               pmdp_huge_get_and_clear_notify()
1857          * madvise_dontneed()
1858          *  zap_pmd_range()
1859          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1860          *   // skip the pmd
1861          *                               set_pmd_at();
1862          *                               // pmd is re-established
1863          *
1864          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1865          * which may break userspace.
1866          *
1867          * pmdp_invalidate() is required to make sure we don't miss
1868          * dirty/young flags set by hardware.
1869          */
1870         entry = pmdp_invalidate(vma, addr, pmd);
1871
1872         entry = pmd_modify(entry, newprot);
1873         if (preserve_write)
1874                 entry = pmd_mk_savedwrite(entry);
1875         if (uffd_wp) {
1876                 entry = pmd_wrprotect(entry);
1877                 entry = pmd_mkuffd_wp(entry);
1878         } else if (uffd_wp_resolve) {
1879                 /*
1880                  * Leave the write bit to be handled by PF interrupt
1881                  * handler, then things like COW could be properly
1882                  * handled.
1883                  */
1884                 entry = pmd_clear_uffd_wp(entry);
1885         }
1886         ret = HPAGE_PMD_NR;
1887         set_pmd_at(mm, addr, pmd, entry);
1888         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1889 unlock:
1890         spin_unlock(ptl);
1891         return ret;
1892 }
1893
1894 /*
1895  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1896  *
1897  * Note that if it returns page table lock pointer, this routine returns without
1898  * unlocking page table lock. So callers must unlock it.
1899  */
1900 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1901 {
1902         spinlock_t *ptl;
1903         ptl = pmd_lock(vma->vm_mm, pmd);
1904         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1905                         pmd_devmap(*pmd)))
1906                 return ptl;
1907         spin_unlock(ptl);
1908         return NULL;
1909 }
1910
1911 /*
1912  * Returns true if a given pud maps a thp, false otherwise.
1913  *
1914  * Note that if it returns true, this routine returns without unlocking page
1915  * table lock. So callers must unlock it.
1916  */
1917 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1918 {
1919         spinlock_t *ptl;
1920
1921         ptl = pud_lock(vma->vm_mm, pud);
1922         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1923                 return ptl;
1924         spin_unlock(ptl);
1925         return NULL;
1926 }
1927
1928 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1929 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1930                  pud_t *pud, unsigned long addr)
1931 {
1932         spinlock_t *ptl;
1933
1934         ptl = __pud_trans_huge_lock(pud, vma);
1935         if (!ptl)
1936                 return 0;
1937         /*
1938          * For architectures like ppc64 we look at deposited pgtable
1939          * when calling pudp_huge_get_and_clear. So do the
1940          * pgtable_trans_huge_withdraw after finishing pudp related
1941          * operations.
1942          */
1943         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1944         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1945         if (vma_is_special_huge(vma)) {
1946                 spin_unlock(ptl);
1947                 /* No zero page support yet */
1948         } else {
1949                 /* No support for anonymous PUD pages yet */
1950                 BUG();
1951         }
1952         return 1;
1953 }
1954
1955 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1956                 unsigned long haddr)
1957 {
1958         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1959         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1960         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1961         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1962
1963         count_vm_event(THP_SPLIT_PUD);
1964
1965         pudp_huge_clear_flush_notify(vma, haddr, pud);
1966 }
1967
1968 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1969                 unsigned long address)
1970 {
1971         spinlock_t *ptl;
1972         struct mmu_notifier_range range;
1973
1974         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1975                                 address & HPAGE_PUD_MASK,
1976                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1977         mmu_notifier_invalidate_range_start(&range);
1978         ptl = pud_lock(vma->vm_mm, pud);
1979         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1980                 goto out;
1981         __split_huge_pud_locked(vma, pud, range.start);
1982
1983 out:
1984         spin_unlock(ptl);
1985         /*
1986          * No need to double call mmu_notifier->invalidate_range() callback as
1987          * the above pudp_huge_clear_flush_notify() did already call it.
1988          */
1989         mmu_notifier_invalidate_range_only_end(&range);
1990 }
1991 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1992
1993 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1994                 unsigned long haddr, pmd_t *pmd)
1995 {
1996         struct mm_struct *mm = vma->vm_mm;
1997         pgtable_t pgtable;
1998         pmd_t _pmd;
1999         int i;
2000
2001         /*
2002          * Leave pmd empty until pte is filled note that it is fine to delay
2003          * notification until mmu_notifier_invalidate_range_end() as we are
2004          * replacing a zero pmd write protected page with a zero pte write
2005          * protected page.
2006          *
2007          * See Documentation/vm/mmu_notifier.rst
2008          */
2009         pmdp_huge_clear_flush(vma, haddr, pmd);
2010
2011         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2012         pmd_populate(mm, &_pmd, pgtable);
2013
2014         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2015                 pte_t *pte, entry;
2016                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2017                 entry = pte_mkspecial(entry);
2018                 pte = pte_offset_map(&_pmd, haddr);
2019                 VM_BUG_ON(!pte_none(*pte));
2020                 set_pte_at(mm, haddr, pte, entry);
2021                 pte_unmap(pte);
2022         }
2023         smp_wmb(); /* make pte visible before pmd */
2024         pmd_populate(mm, pmd, pgtable);
2025 }
2026
2027 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2028                 unsigned long haddr, bool freeze)
2029 {
2030         struct mm_struct *mm = vma->vm_mm;
2031         struct page *page;
2032         pgtable_t pgtable;
2033         pmd_t old_pmd, _pmd;
2034         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2035         unsigned long addr;
2036         int i;
2037
2038         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2039         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2040         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2041         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2042                                 && !pmd_devmap(*pmd));
2043
2044         count_vm_event(THP_SPLIT_PMD);
2045
2046         if (!vma_is_anonymous(vma)) {
2047                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2048                 /*
2049                  * We are going to unmap this huge page. So
2050                  * just go ahead and zap it
2051                  */
2052                 if (arch_needs_pgtable_deposit())
2053                         zap_deposited_table(mm, pmd);
2054                 if (vma_is_special_huge(vma))
2055                         return;
2056                 page = pmd_page(_pmd);
2057                 if (!PageDirty(page) && pmd_dirty(_pmd))
2058                         set_page_dirty(page);
2059                 if (!PageReferenced(page) && pmd_young(_pmd))
2060                         SetPageReferenced(page);
2061                 page_remove_rmap(page, true);
2062                 put_page(page);
2063                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2064                 return;
2065         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2066                 /*
2067                  * FIXME: Do we want to invalidate secondary mmu by calling
2068                  * mmu_notifier_invalidate_range() see comments below inside
2069                  * __split_huge_pmd() ?
2070                  *
2071                  * We are going from a zero huge page write protected to zero
2072                  * small page also write protected so it does not seems useful
2073                  * to invalidate secondary mmu at this time.
2074                  */
2075                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2076         }
2077
2078         /*
2079          * Up to this point the pmd is present and huge and userland has the
2080          * whole access to the hugepage during the split (which happens in
2081          * place). If we overwrite the pmd with the not-huge version pointing
2082          * to the pte here (which of course we could if all CPUs were bug
2083          * free), userland could trigger a small page size TLB miss on the
2084          * small sized TLB while the hugepage TLB entry is still established in
2085          * the huge TLB. Some CPU doesn't like that.
2086          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2087          * 383 on page 105. Intel should be safe but is also warns that it's
2088          * only safe if the permission and cache attributes of the two entries
2089          * loaded in the two TLB is identical (which should be the case here).
2090          * But it is generally safer to never allow small and huge TLB entries
2091          * for the same virtual address to be loaded simultaneously. So instead
2092          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2093          * current pmd notpresent (atomically because here the pmd_trans_huge
2094          * must remain set at all times on the pmd until the split is complete
2095          * for this pmd), then we flush the SMP TLB and finally we write the
2096          * non-huge version of the pmd entry with pmd_populate.
2097          */
2098         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2099
2100         pmd_migration = is_pmd_migration_entry(old_pmd);
2101         if (unlikely(pmd_migration)) {
2102                 swp_entry_t entry;
2103
2104                 entry = pmd_to_swp_entry(old_pmd);
2105                 page = migration_entry_to_page(entry);
2106                 write = is_write_migration_entry(entry);
2107                 young = false;
2108                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2109                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2110         } else {
2111                 page = pmd_page(old_pmd);
2112                 if (pmd_dirty(old_pmd))
2113                         SetPageDirty(page);
2114                 write = pmd_write(old_pmd);
2115                 young = pmd_young(old_pmd);
2116                 soft_dirty = pmd_soft_dirty(old_pmd);
2117                 uffd_wp = pmd_uffd_wp(old_pmd);
2118         }
2119         VM_BUG_ON_PAGE(!page_count(page), page);
2120         page_ref_add(page, HPAGE_PMD_NR - 1);
2121
2122         /*
2123          * Withdraw the table only after we mark the pmd entry invalid.
2124          * This's critical for some architectures (Power).
2125          */
2126         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2127         pmd_populate(mm, &_pmd, pgtable);
2128
2129         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2130                 pte_t entry, *pte;
2131                 /*
2132                  * Note that NUMA hinting access restrictions are not
2133                  * transferred to avoid any possibility of altering
2134                  * permissions across VMAs.
2135                  */
2136                 if (freeze || pmd_migration) {
2137                         swp_entry_t swp_entry;
2138                         swp_entry = make_migration_entry(page + i, write);
2139                         entry = swp_entry_to_pte(swp_entry);
2140                         if (soft_dirty)
2141                                 entry = pte_swp_mksoft_dirty(entry);
2142                         if (uffd_wp)
2143                                 entry = pte_swp_mkuffd_wp(entry);
2144                 } else {
2145                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2146                         entry = maybe_mkwrite(entry, vma);
2147                         if (!write)
2148                                 entry = pte_wrprotect(entry);
2149                         if (!young)
2150                                 entry = pte_mkold(entry);
2151                         if (soft_dirty)
2152                                 entry = pte_mksoft_dirty(entry);
2153                         if (uffd_wp)
2154                                 entry = pte_mkuffd_wp(entry);
2155                 }
2156                 pte = pte_offset_map(&_pmd, addr);
2157                 BUG_ON(!pte_none(*pte));
2158                 set_pte_at(mm, addr, pte, entry);
2159                 if (!pmd_migration)
2160                         atomic_inc(&page[i]._mapcount);
2161                 pte_unmap(pte);
2162         }
2163
2164         if (!pmd_migration) {
2165                 /*
2166                  * Set PG_double_map before dropping compound_mapcount to avoid
2167                  * false-negative page_mapped().
2168                  */
2169                 if (compound_mapcount(page) > 1 &&
2170                     !TestSetPageDoubleMap(page)) {
2171                         for (i = 0; i < HPAGE_PMD_NR; i++)
2172                                 atomic_inc(&page[i]._mapcount);
2173                 }
2174
2175                 lock_page_memcg(page);
2176                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2177                         /* Last compound_mapcount is gone. */
2178                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2179                                                 -HPAGE_PMD_NR);
2180                         if (TestClearPageDoubleMap(page)) {
2181                                 /* No need in mapcount reference anymore */
2182                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2183                                         atomic_dec(&page[i]._mapcount);
2184                         }
2185                 }
2186                 unlock_page_memcg(page);
2187         }
2188
2189         smp_wmb(); /* make pte visible before pmd */
2190         pmd_populate(mm, pmd, pgtable);
2191
2192         if (freeze) {
2193                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2194                         page_remove_rmap(page + i, false);
2195                         put_page(page + i);
2196                 }
2197         }
2198 }
2199
2200 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2201                 unsigned long address, bool freeze, struct page *page)
2202 {
2203         spinlock_t *ptl;
2204         struct mmu_notifier_range range;
2205         bool do_unlock_page = false;
2206         pmd_t _pmd;
2207
2208         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2209                                 address & HPAGE_PMD_MASK,
2210                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2211         mmu_notifier_invalidate_range_start(&range);
2212         ptl = pmd_lock(vma->vm_mm, pmd);
2213
2214         /*
2215          * If caller asks to setup a migration entries, we need a page to check
2216          * pmd against. Otherwise we can end up replacing wrong page.
2217          */
2218         VM_BUG_ON(freeze && !page);
2219         if (page) {
2220                 VM_WARN_ON_ONCE(!PageLocked(page));
2221                 if (page != pmd_page(*pmd))
2222                         goto out;
2223         }
2224
2225 repeat:
2226         if (pmd_trans_huge(*pmd)) {
2227                 if (!page) {
2228                         page = pmd_page(*pmd);
2229                         /*
2230                          * An anonymous page must be locked, to ensure that a
2231                          * concurrent reuse_swap_page() sees stable mapcount;
2232                          * but reuse_swap_page() is not used on shmem or file,
2233                          * and page lock must not be taken when zap_pmd_range()
2234                          * calls __split_huge_pmd() while i_mmap_lock is held.
2235                          */
2236                         if (PageAnon(page)) {
2237                                 if (unlikely(!trylock_page(page))) {
2238                                         get_page(page);
2239                                         _pmd = *pmd;
2240                                         spin_unlock(ptl);
2241                                         lock_page(page);
2242                                         spin_lock(ptl);
2243                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2244                                                 unlock_page(page);
2245                                                 put_page(page);
2246                                                 page = NULL;
2247                                                 goto repeat;
2248                                         }
2249                                         put_page(page);
2250                                 }
2251                                 do_unlock_page = true;
2252                         }
2253                 }
2254                 if (PageMlocked(page))
2255                         clear_page_mlock(page);
2256         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2257                 goto out;
2258         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2259 out:
2260         spin_unlock(ptl);
2261         if (do_unlock_page)
2262                 unlock_page(page);
2263         /*
2264          * No need to double call mmu_notifier->invalidate_range() callback.
2265          * They are 3 cases to consider inside __split_huge_pmd_locked():
2266          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2267          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2268          *    fault will trigger a flush_notify before pointing to a new page
2269          *    (it is fine if the secondary mmu keeps pointing to the old zero
2270          *    page in the meantime)
2271          *  3) Split a huge pmd into pte pointing to the same page. No need
2272          *     to invalidate secondary tlb entry they are all still valid.
2273          *     any further changes to individual pte will notify. So no need
2274          *     to call mmu_notifier->invalidate_range()
2275          */
2276         mmu_notifier_invalidate_range_only_end(&range);
2277 }
2278
2279 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2280                 bool freeze, struct page *page)
2281 {
2282         pgd_t *pgd;
2283         p4d_t *p4d;
2284         pud_t *pud;
2285         pmd_t *pmd;
2286
2287         pgd = pgd_offset(vma->vm_mm, address);
2288         if (!pgd_present(*pgd))
2289                 return;
2290
2291         p4d = p4d_offset(pgd, address);
2292         if (!p4d_present(*p4d))
2293                 return;
2294
2295         pud = pud_offset(p4d, address);
2296         if (!pud_present(*pud))
2297                 return;
2298
2299         pmd = pmd_offset(pud, address);
2300
2301         __split_huge_pmd(vma, pmd, address, freeze, page);
2302 }
2303
2304 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2305 {
2306         /*
2307          * If the new address isn't hpage aligned and it could previously
2308          * contain an hugepage: check if we need to split an huge pmd.
2309          */
2310         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2311             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2312                          ALIGN(address, HPAGE_PMD_SIZE)))
2313                 split_huge_pmd_address(vma, address, false, NULL);
2314 }
2315
2316 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2317                              unsigned long start,
2318                              unsigned long end,
2319                              long adjust_next)
2320 {
2321         /* Check if we need to split start first. */
2322         split_huge_pmd_if_needed(vma, start);
2323
2324         /* Check if we need to split end next. */
2325         split_huge_pmd_if_needed(vma, end);
2326
2327         /*
2328          * If we're also updating the vma->vm_next->vm_start,
2329          * check if we need to split it.
2330          */
2331         if (adjust_next > 0) {
2332                 struct vm_area_struct *next = vma->vm_next;
2333                 unsigned long nstart = next->vm_start;
2334                 nstart += adjust_next;
2335                 split_huge_pmd_if_needed(next, nstart);
2336         }
2337 }
2338
2339 static void unmap_page(struct page *page)
2340 {
2341         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2342                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2343         bool unmap_success;
2344
2345         VM_BUG_ON_PAGE(!PageHead(page), page);
2346
2347         if (PageAnon(page))
2348                 ttu_flags |= TTU_SPLIT_FREEZE;
2349
2350         unmap_success = try_to_unmap(page, ttu_flags);
2351         VM_BUG_ON_PAGE(!unmap_success, page);
2352 }
2353
2354 static void remap_page(struct page *page, unsigned int nr)
2355 {
2356         int i;
2357         if (PageTransHuge(page)) {
2358                 remove_migration_ptes(page, page, true);
2359         } else {
2360                 for (i = 0; i < nr; i++)
2361                         remove_migration_ptes(page + i, page + i, true);
2362         }
2363 }
2364
2365 static void lru_add_page_tail(struct page *head, struct page *tail,
2366                 struct lruvec *lruvec, struct list_head *list)
2367 {
2368         VM_BUG_ON_PAGE(!PageHead(head), head);
2369         VM_BUG_ON_PAGE(PageCompound(tail), head);
2370         VM_BUG_ON_PAGE(PageLRU(tail), head);
2371         lockdep_assert_held(&lruvec->lru_lock);
2372
2373         if (list) {
2374                 /* page reclaim is reclaiming a huge page */
2375                 VM_WARN_ON(PageLRU(head));
2376                 get_page(tail);
2377                 list_add_tail(&tail->lru, list);
2378         } else {
2379                 /* head is still on lru (and we have it frozen) */
2380                 VM_WARN_ON(!PageLRU(head));
2381                 SetPageLRU(tail);
2382                 list_add_tail(&tail->lru, &head->lru);
2383         }
2384 }
2385
2386 static void __split_huge_page_tail(struct page *head, int tail,
2387                 struct lruvec *lruvec, struct list_head *list)
2388 {
2389         struct page *page_tail = head + tail;
2390
2391         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2392
2393         /*
2394          * Clone page flags before unfreezing refcount.
2395          *
2396          * After successful get_page_unless_zero() might follow flags change,
2397          * for example lock_page() which set PG_waiters.
2398          */
2399         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2400         page_tail->flags |= (head->flags &
2401                         ((1L << PG_referenced) |
2402                          (1L << PG_swapbacked) |
2403                          (1L << PG_swapcache) |
2404                          (1L << PG_mlocked) |
2405                          (1L << PG_uptodate) |
2406                          (1L << PG_active) |
2407                          (1L << PG_workingset) |
2408                          (1L << PG_locked) |
2409                          (1L << PG_unevictable) |
2410 #ifdef CONFIG_64BIT
2411                          (1L << PG_arch_2) |
2412 #endif
2413                          (1L << PG_dirty)));
2414
2415         /* ->mapping in first tail page is compound_mapcount */
2416         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2417                         page_tail);
2418         page_tail->mapping = head->mapping;
2419         page_tail->index = head->index + tail;
2420
2421         /* Page flags must be visible before we make the page non-compound. */
2422         smp_wmb();
2423
2424         /*
2425          * Clear PageTail before unfreezing page refcount.
2426          *
2427          * After successful get_page_unless_zero() might follow put_page()
2428          * which needs correct compound_head().
2429          */
2430         clear_compound_head(page_tail);
2431
2432         /* Finally unfreeze refcount. Additional reference from page cache. */
2433         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2434                                           PageSwapCache(head)));
2435
2436         if (page_is_young(head))
2437                 set_page_young(page_tail);
2438         if (page_is_idle(head))
2439                 set_page_idle(page_tail);
2440
2441         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2442
2443         /*
2444          * always add to the tail because some iterators expect new
2445          * pages to show after the currently processed elements - e.g.
2446          * migrate_pages
2447          */
2448         lru_add_page_tail(head, page_tail, lruvec, list);
2449 }
2450
2451 static void __split_huge_page(struct page *page, struct list_head *list,
2452                 pgoff_t end)
2453 {
2454         struct page *head = compound_head(page);
2455         struct lruvec *lruvec;
2456         struct address_space *swap_cache = NULL;
2457         unsigned long offset = 0;
2458         unsigned int nr = thp_nr_pages(head);
2459         int i;
2460
2461         /* complete memcg works before add pages to LRU */
2462         split_page_memcg(head, nr);
2463
2464         if (PageAnon(head) && PageSwapCache(head)) {
2465                 swp_entry_t entry = { .val = page_private(head) };
2466
2467                 offset = swp_offset(entry);
2468                 swap_cache = swap_address_space(entry);
2469                 xa_lock(&swap_cache->i_pages);
2470         }
2471
2472         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2473         lruvec = lock_page_lruvec(head);
2474
2475         for (i = nr - 1; i >= 1; i--) {
2476                 __split_huge_page_tail(head, i, lruvec, list);
2477                 /* Some pages can be beyond i_size: drop them from page cache */
2478                 if (head[i].index >= end) {
2479                         ClearPageDirty(head + i);
2480                         __delete_from_page_cache(head + i, NULL);
2481                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2482                                 shmem_uncharge(head->mapping->host, 1);
2483                         put_page(head + i);
2484                 } else if (!PageAnon(page)) {
2485                         __xa_store(&head->mapping->i_pages, head[i].index,
2486                                         head + i, 0);
2487                 } else if (swap_cache) {
2488                         __xa_store(&swap_cache->i_pages, offset + i,
2489                                         head + i, 0);
2490                 }
2491         }
2492
2493         ClearPageCompound(head);
2494         unlock_page_lruvec(lruvec);
2495         /* Caller disabled irqs, so they are still disabled here */
2496
2497         split_page_owner(head, nr);
2498
2499         /* See comment in __split_huge_page_tail() */
2500         if (PageAnon(head)) {
2501                 /* Additional pin to swap cache */
2502                 if (PageSwapCache(head)) {
2503                         page_ref_add(head, 2);
2504                         xa_unlock(&swap_cache->i_pages);
2505                 } else {
2506                         page_ref_inc(head);
2507                 }
2508         } else {
2509                 /* Additional pin to page cache */
2510                 page_ref_add(head, 2);
2511                 xa_unlock(&head->mapping->i_pages);
2512         }
2513         local_irq_enable();
2514
2515         remap_page(head, nr);
2516
2517         if (PageSwapCache(head)) {
2518                 swp_entry_t entry = { .val = page_private(head) };
2519
2520                 split_swap_cluster(entry);
2521         }
2522
2523         for (i = 0; i < nr; i++) {
2524                 struct page *subpage = head + i;
2525                 if (subpage == page)
2526                         continue;
2527                 unlock_page(subpage);
2528
2529                 /*
2530                  * Subpages may be freed if there wasn't any mapping
2531                  * like if add_to_swap() is running on a lru page that
2532                  * had its mapping zapped. And freeing these pages
2533                  * requires taking the lru_lock so we do the put_page
2534                  * of the tail pages after the split is complete.
2535                  */
2536                 put_page(subpage);
2537         }
2538 }
2539
2540 int total_mapcount(struct page *page)
2541 {
2542         int i, compound, nr, ret;
2543
2544         VM_BUG_ON_PAGE(PageTail(page), page);
2545
2546         if (likely(!PageCompound(page)))
2547                 return atomic_read(&page->_mapcount) + 1;
2548
2549         compound = compound_mapcount(page);
2550         nr = compound_nr(page);
2551         if (PageHuge(page))
2552                 return compound;
2553         ret = compound;
2554         for (i = 0; i < nr; i++)
2555                 ret += atomic_read(&page[i]._mapcount) + 1;
2556         /* File pages has compound_mapcount included in _mapcount */
2557         if (!PageAnon(page))
2558                 return ret - compound * nr;
2559         if (PageDoubleMap(page))
2560                 ret -= nr;
2561         return ret;
2562 }
2563
2564 /*
2565  * This calculates accurately how many mappings a transparent hugepage
2566  * has (unlike page_mapcount() which isn't fully accurate). This full
2567  * accuracy is primarily needed to know if copy-on-write faults can
2568  * reuse the page and change the mapping to read-write instead of
2569  * copying them. At the same time this returns the total_mapcount too.
2570  *
2571  * The function returns the highest mapcount any one of the subpages
2572  * has. If the return value is one, even if different processes are
2573  * mapping different subpages of the transparent hugepage, they can
2574  * all reuse it, because each process is reusing a different subpage.
2575  *
2576  * The total_mapcount is instead counting all virtual mappings of the
2577  * subpages. If the total_mapcount is equal to "one", it tells the
2578  * caller all mappings belong to the same "mm" and in turn the
2579  * anon_vma of the transparent hugepage can become the vma->anon_vma
2580  * local one as no other process may be mapping any of the subpages.
2581  *
2582  * It would be more accurate to replace page_mapcount() with
2583  * page_trans_huge_mapcount(), however we only use
2584  * page_trans_huge_mapcount() in the copy-on-write faults where we
2585  * need full accuracy to avoid breaking page pinning, because
2586  * page_trans_huge_mapcount() is slower than page_mapcount().
2587  */
2588 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2589 {
2590         int i, ret, _total_mapcount, mapcount;
2591
2592         /* hugetlbfs shouldn't call it */
2593         VM_BUG_ON_PAGE(PageHuge(page), page);
2594
2595         if (likely(!PageTransCompound(page))) {
2596                 mapcount = atomic_read(&page->_mapcount) + 1;
2597                 if (total_mapcount)
2598                         *total_mapcount = mapcount;
2599                 return mapcount;
2600         }
2601
2602         page = compound_head(page);
2603
2604         _total_mapcount = ret = 0;
2605         for (i = 0; i < thp_nr_pages(page); i++) {
2606                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2607                 ret = max(ret, mapcount);
2608                 _total_mapcount += mapcount;
2609         }
2610         if (PageDoubleMap(page)) {
2611                 ret -= 1;
2612                 _total_mapcount -= thp_nr_pages(page);
2613         }
2614         mapcount = compound_mapcount(page);
2615         ret += mapcount;
2616         _total_mapcount += mapcount;
2617         if (total_mapcount)
2618                 *total_mapcount = _total_mapcount;
2619         return ret;
2620 }
2621
2622 /* Racy check whether the huge page can be split */
2623 bool can_split_huge_page(struct page *page, int *pextra_pins)
2624 {
2625         int extra_pins;
2626
2627         /* Additional pins from page cache */
2628         if (PageAnon(page))
2629                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2630         else
2631                 extra_pins = thp_nr_pages(page);
2632         if (pextra_pins)
2633                 *pextra_pins = extra_pins;
2634         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2635 }
2636
2637 /*
2638  * This function splits huge page into normal pages. @page can point to any
2639  * subpage of huge page to split. Split doesn't change the position of @page.
2640  *
2641  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2642  * The huge page must be locked.
2643  *
2644  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2645  *
2646  * Both head page and tail pages will inherit mapping, flags, and so on from
2647  * the hugepage.
2648  *
2649  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2650  * they are not mapped.
2651  *
2652  * Returns 0 if the hugepage is split successfully.
2653  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2654  * us.
2655  */
2656 int split_huge_page_to_list(struct page *page, struct list_head *list)
2657 {
2658         struct page *head = compound_head(page);
2659         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2660         struct anon_vma *anon_vma = NULL;
2661         struct address_space *mapping = NULL;
2662         int count, mapcount, extra_pins, ret;
2663         pgoff_t end;
2664
2665         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2666         VM_BUG_ON_PAGE(!PageLocked(head), head);
2667         VM_BUG_ON_PAGE(!PageCompound(head), head);
2668
2669         if (PageWriteback(head))
2670                 return -EBUSY;
2671
2672         if (PageAnon(head)) {
2673                 /*
2674                  * The caller does not necessarily hold an mmap_lock that would
2675                  * prevent the anon_vma disappearing so we first we take a
2676                  * reference to it and then lock the anon_vma for write. This
2677                  * is similar to page_lock_anon_vma_read except the write lock
2678                  * is taken to serialise against parallel split or collapse
2679                  * operations.
2680                  */
2681                 anon_vma = page_get_anon_vma(head);
2682                 if (!anon_vma) {
2683                         ret = -EBUSY;
2684                         goto out;
2685                 }
2686                 end = -1;
2687                 mapping = NULL;
2688                 anon_vma_lock_write(anon_vma);
2689         } else {
2690                 mapping = head->mapping;
2691
2692                 /* Truncated ? */
2693                 if (!mapping) {
2694                         ret = -EBUSY;
2695                         goto out;
2696                 }
2697
2698                 anon_vma = NULL;
2699                 i_mmap_lock_read(mapping);
2700
2701                 /*
2702                  *__split_huge_page() may need to trim off pages beyond EOF:
2703                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2704                  * which cannot be nested inside the page tree lock. So note
2705                  * end now: i_size itself may be changed at any moment, but
2706                  * head page lock is good enough to serialize the trimming.
2707                  */
2708                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2709         }
2710
2711         /*
2712          * Racy check if we can split the page, before unmap_page() will
2713          * split PMDs
2714          */
2715         if (!can_split_huge_page(head, &extra_pins)) {
2716                 ret = -EBUSY;
2717                 goto out_unlock;
2718         }
2719
2720         unmap_page(head);
2721         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2722
2723         /* block interrupt reentry in xa_lock and spinlock */
2724         local_irq_disable();
2725         if (mapping) {
2726                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2727
2728                 /*
2729                  * Check if the head page is present in page cache.
2730                  * We assume all tail are present too, if head is there.
2731                  */
2732                 xa_lock(&mapping->i_pages);
2733                 if (xas_load(&xas) != head)
2734                         goto fail;
2735         }
2736
2737         /* Prevent deferred_split_scan() touching ->_refcount */
2738         spin_lock(&ds_queue->split_queue_lock);
2739         count = page_count(head);
2740         mapcount = total_mapcount(head);
2741         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2742                 if (!list_empty(page_deferred_list(head))) {
2743                         ds_queue->split_queue_len--;
2744                         list_del(page_deferred_list(head));
2745                 }
2746                 spin_unlock(&ds_queue->split_queue_lock);
2747                 if (mapping) {
2748                         int nr = thp_nr_pages(head);
2749
2750                         if (PageSwapBacked(head))
2751                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2752                                                         -nr);
2753                         else
2754                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2755                                                         -nr);
2756                 }
2757
2758                 __split_huge_page(page, list, end);
2759                 ret = 0;
2760         } else {
2761                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2762                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2763                                         mapcount, count);
2764                         if (PageTail(page))
2765                                 dump_page(head, NULL);
2766                         dump_page(page, "total_mapcount(head) > 0");
2767                         BUG();
2768                 }
2769                 spin_unlock(&ds_queue->split_queue_lock);
2770 fail:           if (mapping)
2771                         xa_unlock(&mapping->i_pages);
2772                 local_irq_enable();
2773                 remap_page(head, thp_nr_pages(head));
2774                 ret = -EBUSY;
2775         }
2776
2777 out_unlock:
2778         if (anon_vma) {
2779                 anon_vma_unlock_write(anon_vma);
2780                 put_anon_vma(anon_vma);
2781         }
2782         if (mapping)
2783                 i_mmap_unlock_read(mapping);
2784 out:
2785         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2786         return ret;
2787 }
2788
2789 void free_transhuge_page(struct page *page)
2790 {
2791         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2792         unsigned long flags;
2793
2794         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2795         if (!list_empty(page_deferred_list(page))) {
2796                 ds_queue->split_queue_len--;
2797                 list_del(page_deferred_list(page));
2798         }
2799         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2800         free_compound_page(page);
2801 }
2802
2803 void deferred_split_huge_page(struct page *page)
2804 {
2805         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2806 #ifdef CONFIG_MEMCG
2807         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2808 #endif
2809         unsigned long flags;
2810
2811         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2812
2813         /*
2814          * The try_to_unmap() in page reclaim path might reach here too,
2815          * this may cause a race condition to corrupt deferred split queue.
2816          * And, if page reclaim is already handling the same page, it is
2817          * unnecessary to handle it again in shrinker.
2818          *
2819          * Check PageSwapCache to determine if the page is being
2820          * handled by page reclaim since THP swap would add the page into
2821          * swap cache before calling try_to_unmap().
2822          */
2823         if (PageSwapCache(page))
2824                 return;
2825
2826         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2827         if (list_empty(page_deferred_list(page))) {
2828                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2829                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2830                 ds_queue->split_queue_len++;
2831 #ifdef CONFIG_MEMCG
2832                 if (memcg)
2833                         set_shrinker_bit(memcg, page_to_nid(page),
2834                                          deferred_split_shrinker.id);
2835 #endif
2836         }
2837         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2838 }
2839
2840 static unsigned long deferred_split_count(struct shrinker *shrink,
2841                 struct shrink_control *sc)
2842 {
2843         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2844         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2845
2846 #ifdef CONFIG_MEMCG
2847         if (sc->memcg)
2848                 ds_queue = &sc->memcg->deferred_split_queue;
2849 #endif
2850         return READ_ONCE(ds_queue->split_queue_len);
2851 }
2852
2853 static unsigned long deferred_split_scan(struct shrinker *shrink,
2854                 struct shrink_control *sc)
2855 {
2856         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2857         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2858         unsigned long flags;
2859         LIST_HEAD(list), *pos, *next;
2860         struct page *page;
2861         int split = 0;
2862
2863 #ifdef CONFIG_MEMCG
2864         if (sc->memcg)
2865                 ds_queue = &sc->memcg->deferred_split_queue;
2866 #endif
2867
2868         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2869         /* Take pin on all head pages to avoid freeing them under us */
2870         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2871                 page = list_entry((void *)pos, struct page, mapping);
2872                 page = compound_head(page);
2873                 if (get_page_unless_zero(page)) {
2874                         list_move(page_deferred_list(page), &list);
2875                 } else {
2876                         /* We lost race with put_compound_page() */
2877                         list_del_init(page_deferred_list(page));
2878                         ds_queue->split_queue_len--;
2879                 }
2880                 if (!--sc->nr_to_scan)
2881                         break;
2882         }
2883         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2884
2885         list_for_each_safe(pos, next, &list) {
2886                 page = list_entry((void *)pos, struct page, mapping);
2887                 if (!trylock_page(page))
2888                         goto next;
2889                 /* split_huge_page() removes page from list on success */
2890                 if (!split_huge_page(page))
2891                         split++;
2892                 unlock_page(page);
2893 next:
2894                 put_page(page);
2895         }
2896
2897         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2898         list_splice_tail(&list, &ds_queue->split_queue);
2899         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2900
2901         /*
2902          * Stop shrinker if we didn't split any page, but the queue is empty.
2903          * This can happen if pages were freed under us.
2904          */
2905         if (!split && list_empty(&ds_queue->split_queue))
2906                 return SHRINK_STOP;
2907         return split;
2908 }
2909
2910 static struct shrinker deferred_split_shrinker = {
2911         .count_objects = deferred_split_count,
2912         .scan_objects = deferred_split_scan,
2913         .seeks = DEFAULT_SEEKS,
2914         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2915                  SHRINKER_NONSLAB,
2916 };
2917
2918 #ifdef CONFIG_DEBUG_FS
2919 static void split_huge_pages_all(void)
2920 {
2921         struct zone *zone;
2922         struct page *page;
2923         unsigned long pfn, max_zone_pfn;
2924         unsigned long total = 0, split = 0;
2925
2926         pr_debug("Split all THPs\n");
2927         for_each_populated_zone(zone) {
2928                 max_zone_pfn = zone_end_pfn(zone);
2929                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2930                         if (!pfn_valid(pfn))
2931                                 continue;
2932
2933                         page = pfn_to_page(pfn);
2934                         if (!get_page_unless_zero(page))
2935                                 continue;
2936
2937                         if (zone != page_zone(page))
2938                                 goto next;
2939
2940                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2941                                 goto next;
2942
2943                         total++;
2944                         lock_page(page);
2945                         if (!split_huge_page(page))
2946                                 split++;
2947                         unlock_page(page);
2948 next:
2949                         put_page(page);
2950                         cond_resched();
2951                 }
2952         }
2953
2954         pr_debug("%lu of %lu THP split\n", split, total);
2955 }
2956
2957 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2958 {
2959         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2960                     is_vm_hugetlb_page(vma);
2961 }
2962
2963 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2964                                 unsigned long vaddr_end)
2965 {
2966         int ret = 0;
2967         struct task_struct *task;
2968         struct mm_struct *mm;
2969         unsigned long total = 0, split = 0;
2970         unsigned long addr;
2971
2972         vaddr_start &= PAGE_MASK;
2973         vaddr_end &= PAGE_MASK;
2974
2975         /* Find the task_struct from pid */
2976         rcu_read_lock();
2977         task = find_task_by_vpid(pid);
2978         if (!task) {
2979                 rcu_read_unlock();
2980                 ret = -ESRCH;
2981                 goto out;
2982         }
2983         get_task_struct(task);
2984         rcu_read_unlock();
2985
2986         /* Find the mm_struct */
2987         mm = get_task_mm(task);
2988         put_task_struct(task);
2989
2990         if (!mm) {
2991                 ret = -EINVAL;
2992                 goto out;
2993         }
2994
2995         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2996                  pid, vaddr_start, vaddr_end);
2997
2998         mmap_read_lock(mm);
2999         /*
3000          * always increase addr by PAGE_SIZE, since we could have a PTE page
3001          * table filled with PTE-mapped THPs, each of which is distinct.
3002          */
3003         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3004                 struct vm_area_struct *vma = find_vma(mm, addr);
3005                 unsigned int follflags;
3006                 struct page *page;
3007
3008                 if (!vma || addr < vma->vm_start)
3009                         break;
3010
3011                 /* skip special VMA and hugetlb VMA */
3012                 if (vma_not_suitable_for_thp_split(vma)) {
3013                         addr = vma->vm_end;
3014                         continue;
3015                 }
3016
3017                 /* FOLL_DUMP to ignore special (like zero) pages */
3018                 follflags = FOLL_GET | FOLL_DUMP;
3019                 page = follow_page(vma, addr, follflags);
3020
3021                 if (IS_ERR(page))
3022                         continue;
3023                 if (!page)
3024                         continue;
3025
3026                 if (!is_transparent_hugepage(page))
3027                         goto next;
3028
3029                 total++;
3030                 if (!can_split_huge_page(compound_head(page), NULL))
3031                         goto next;
3032
3033                 if (!trylock_page(page))
3034                         goto next;
3035
3036                 if (!split_huge_page(page))
3037                         split++;
3038
3039                 unlock_page(page);
3040 next:
3041                 put_page(page);
3042                 cond_resched();
3043         }
3044         mmap_read_unlock(mm);
3045         mmput(mm);
3046
3047         pr_debug("%lu of %lu THP split\n", split, total);
3048
3049 out:
3050         return ret;
3051 }
3052
3053 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3054                                 pgoff_t off_end)
3055 {
3056         struct filename *file;
3057         struct file *candidate;
3058         struct address_space *mapping;
3059         int ret = -EINVAL;
3060         pgoff_t index;
3061         int nr_pages = 1;
3062         unsigned long total = 0, split = 0;
3063
3064         file = getname_kernel(file_path);
3065         if (IS_ERR(file))
3066                 return ret;
3067
3068         candidate = file_open_name(file, O_RDONLY, 0);
3069         if (IS_ERR(candidate))
3070                 goto out;
3071
3072         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3073                  file_path, off_start, off_end);
3074
3075         mapping = candidate->f_mapping;
3076
3077         for (index = off_start; index < off_end; index += nr_pages) {
3078                 struct page *fpage = pagecache_get_page(mapping, index,
3079                                                 FGP_ENTRY | FGP_HEAD, 0);
3080
3081                 nr_pages = 1;
3082                 if (xa_is_value(fpage) || !fpage)
3083                         continue;
3084
3085                 if (!is_transparent_hugepage(fpage))
3086                         goto next;
3087
3088                 total++;
3089                 nr_pages = thp_nr_pages(fpage);
3090
3091                 if (!trylock_page(fpage))
3092                         goto next;
3093
3094                 if (!split_huge_page(fpage))
3095                         split++;
3096
3097                 unlock_page(fpage);
3098 next:
3099                 put_page(fpage);
3100                 cond_resched();
3101         }
3102
3103         filp_close(candidate, NULL);
3104         ret = 0;
3105
3106         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3107 out:
3108         putname(file);
3109         return ret;
3110 }
3111
3112 #define MAX_INPUT_BUF_SZ 255
3113
3114 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3115                                 size_t count, loff_t *ppops)
3116 {
3117         static DEFINE_MUTEX(split_debug_mutex);
3118         ssize_t ret;
3119         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3120         char input_buf[MAX_INPUT_BUF_SZ];
3121         int pid;
3122         unsigned long vaddr_start, vaddr_end;
3123
3124         ret = mutex_lock_interruptible(&split_debug_mutex);
3125         if (ret)
3126                 return ret;
3127
3128         ret = -EFAULT;
3129
3130         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3131         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3132                 goto out;
3133
3134         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3135
3136         if (input_buf[0] == '/') {
3137                 char *tok;
3138                 char *buf = input_buf;
3139                 char file_path[MAX_INPUT_BUF_SZ];
3140                 pgoff_t off_start = 0, off_end = 0;
3141                 size_t input_len = strlen(input_buf);
3142
3143                 tok = strsep(&buf, ",");
3144                 if (tok) {
3145                         strncpy(file_path, tok, MAX_INPUT_BUF_SZ);
3146                 } else {
3147                         ret = -EINVAL;
3148                         goto out;
3149                 }
3150
3151                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3152                 if (ret != 2) {
3153                         ret = -EINVAL;
3154                         goto out;
3155                 }
3156                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3157                 if (!ret)
3158                         ret = input_len;
3159
3160                 goto out;
3161         }
3162
3163         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3164         if (ret == 1 && pid == 1) {
3165                 split_huge_pages_all();
3166                 ret = strlen(input_buf);
3167                 goto out;
3168         } else if (ret != 3) {
3169                 ret = -EINVAL;
3170                 goto out;
3171         }
3172
3173         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3174         if (!ret)
3175                 ret = strlen(input_buf);
3176 out:
3177         mutex_unlock(&split_debug_mutex);
3178         return ret;
3179
3180 }
3181
3182 static const struct file_operations split_huge_pages_fops = {
3183         .owner   = THIS_MODULE,
3184         .write   = split_huge_pages_write,
3185         .llseek  = no_llseek,
3186 };
3187
3188 static int __init split_huge_pages_debugfs(void)
3189 {
3190         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3191                             &split_huge_pages_fops);
3192         return 0;
3193 }
3194 late_initcall(split_huge_pages_debugfs);
3195 #endif
3196
3197 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3198 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3199                 struct page *page)
3200 {
3201         struct vm_area_struct *vma = pvmw->vma;
3202         struct mm_struct *mm = vma->vm_mm;
3203         unsigned long address = pvmw->address;
3204         pmd_t pmdval;
3205         swp_entry_t entry;
3206         pmd_t pmdswp;
3207
3208         if (!(pvmw->pmd && !pvmw->pte))
3209                 return;
3210
3211         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3212         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3213         if (pmd_dirty(pmdval))
3214                 set_page_dirty(page);
3215         entry = make_migration_entry(page, pmd_write(pmdval));
3216         pmdswp = swp_entry_to_pmd(entry);
3217         if (pmd_soft_dirty(pmdval))
3218                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3219         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3220         page_remove_rmap(page, true);
3221         put_page(page);
3222 }
3223
3224 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3225 {
3226         struct vm_area_struct *vma = pvmw->vma;
3227         struct mm_struct *mm = vma->vm_mm;
3228         unsigned long address = pvmw->address;
3229         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3230         pmd_t pmde;
3231         swp_entry_t entry;
3232
3233         if (!(pvmw->pmd && !pvmw->pte))
3234                 return;
3235
3236         entry = pmd_to_swp_entry(*pvmw->pmd);
3237         get_page(new);
3238         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3239         if (pmd_swp_soft_dirty(*pvmw->pmd))
3240                 pmde = pmd_mksoft_dirty(pmde);
3241         if (is_write_migration_entry(entry))
3242                 pmde = maybe_pmd_mkwrite(pmde, vma);
3243
3244         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3245         if (PageAnon(new))
3246                 page_add_anon_rmap(new, vma, mmun_start, true);
3247         else
3248                 page_add_file_rmap(new, true);
3249         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3250         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3251                 mlock_vma_page(new);
3252         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3253 }
3254 #endif