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/backing-dev.h>
  22 #include <linux/dax.h>
  23 #include <linux/mm_types.h>
  24 #include <linux/khugepaged.h>
  25 #include <linux/freezer.h>
  26 #include <linux/pfn_t.h>
  27 #include <linux/mman.h>
  28 #include <linux/memremap.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/debugfs.h>
  31 #include <linux/migrate.h>
  32 #include <linux/hashtable.h>
  33 #include <linux/userfaultfd_k.h>
  34 #include <linux/page_idle.h>
  35 #include <linux/shmem_fs.h>
  36 #include <linux/oom.h>
  37 #include <linux/numa.h>
  38 #include <linux/page_owner.h>
  39 #include <linux/sched/sysctl.h>
  40 #include <linux/memory-tiers.h>
  41 #include <linux/compat.h>
  42 #include <linux/pgalloc_tag.h>
  43
  44 #include <asm/tlb.h>
  45 #include <asm/pgalloc.h>
  46 #include "internal.h"
  47 #include "swap.h"
  48
  49 #define CREATE_TRACE_POINTS
  50 #include <trace/events/thp.h>
  51
  52 /*
  53  * By default, transparent hugepage support is disabled in order to avoid
  54  * risking an increased memory footprint for applications that are not
  55  * guaranteed to benefit from it. When transparent hugepage support is
  56  * enabled, it is for all mappings, and khugepaged scans all mappings.
  57  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  58  * for all hugepage allocations.
  59  */
  60 unsigned long transparent_hugepage_flags __read_mostly =
  61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  62         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  63 #endif
  64 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  65         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  66 #endif
  67         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  68         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  69         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  70
  71 static struct shrinker *deferred_split_shrinker;
  72 static unsigned long deferred_split_count(struct shrinker *shrink,
  73                                           struct shrink_control *sc);
  74 static unsigned long deferred_split_scan(struct shrinker *shrink,
  75                                          struct shrink_control *sc);
  76
  77 static atomic_t huge_zero_refcount;
  78 struct folio *huge_zero_folio __read_mostly;
  79 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  80 unsigned long huge_anon_orders_always __read_mostly;
  81 unsigned long huge_anon_orders_madvise __read_mostly;
  82 unsigned long huge_anon_orders_inherit __read_mostly;
  83
  84 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
  85                                          unsigned long vm_flags,
  86                                          unsigned long tva_flags,
  87                                          unsigned long orders)
  88 {
  89         bool smaps = tva_flags & TVA_SMAPS;
  90         bool in_pf = tva_flags & TVA_IN_PF;
  91         bool enforce_sysfs = tva_flags & TVA_ENFORCE_SYSFS;
  92         unsigned long supported_orders;
  93
  94         /* Check the intersection of requested and supported orders. */
  95         if (vma_is_anonymous(vma))
  96                 supported_orders = THP_ORDERS_ALL_ANON;
  97         else if (vma_is_dax(vma))
  98                 supported_orders = THP_ORDERS_ALL_FILE_DAX;
  99         else
 100                 supported_orders = THP_ORDERS_ALL_FILE_DEFAULT;
 101
 102         orders &= supported_orders;
 103         if (!orders)
 104                 return 0;
 105
 106         if (!vma->vm_mm)                /* vdso */
 107                 return 0;
 108
 109         /*
 110          * Explicitly disabled through madvise or prctl, or some
 111          * architectures may disable THP for some mappings, for
 112          * example, s390 kvm.
 113          * */
 114         if ((vm_flags & VM_NOHUGEPAGE) ||
 115             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 116                 return 0;
 117         /*
 118          * If the hardware/firmware marked hugepage support disabled.
 119          */
 120         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED))
 121                 return 0;
 122
 123         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
 124         if (vma_is_dax(vma))
 125                 return in_pf ? orders : 0;
 126
 127         /*
 128          * khugepaged special VMA and hugetlb VMA.
 129          * Must be checked after dax since some dax mappings may have
 130          * VM_MIXEDMAP set.
 131          */
 132         if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
 133                 return 0;
 134
 135         /*
 136          * Check alignment for file vma and size for both file and anon vma by
 137          * filtering out the unsuitable orders.
 138          *
 139          * Skip the check for page fault. Huge fault does the check in fault
 140          * handlers.
 141          */
 142         if (!in_pf) {
 143                 int order = highest_order(orders);
 144                 unsigned long addr;
 145
 146                 while (orders) {
 147                         addr = vma->vm_end - (PAGE_SIZE << order);
 148                         if (thp_vma_suitable_order(vma, addr, order))
 149                                 break;
 150                         order = next_order(&orders, order);
 151                 }
 152
 153                 if (!orders)
 154                         return 0;
 155         }
 156
 157         /*
 158          * Enabled via shmem mount options or sysfs settings.
 159          * Must be done before hugepage flags check since shmem has its
 160          * own flags.
 161          */
 162         if (!in_pf && shmem_file(vma->vm_file)) {
 163                 bool global_huge = shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
 164                                                         !enforce_sysfs, vma->vm_mm, vm_flags);
 165
 166                 if (!vma_is_anon_shmem(vma))
 167                         return global_huge ? orders : 0;
 168                 return shmem_allowable_huge_orders(file_inode(vma->vm_file),
 169                                                         vma, vma->vm_pgoff, global_huge);
 170         }
 171
 172         if (!vma_is_anonymous(vma)) {
 173                 /*
 174                  * Enforce sysfs THP requirements as necessary. Anonymous vmas
 175                  * were already handled in thp_vma_allowable_orders().
 176                  */
 177                 if (enforce_sysfs &&
 178                     (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 179                                                     !hugepage_global_always())))
 180                         return 0;
 181
 182                 /*
 183                  * Trust that ->huge_fault() handlers know what they are doing
 184                  * in fault path.
 185                  */
 186                 if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
 187                         return orders;
 188                 /* Only regular file is valid in collapse path */
 189                 if (((!in_pf || smaps)) && file_thp_enabled(vma))
 190                         return orders;
 191                 return 0;
 192         }
 193
 194         if (vma_is_temporary_stack(vma))
 195                 return 0;
 196
 197         /*
 198          * THPeligible bit of smaps should show 1 for proper VMAs even
 199          * though anon_vma is not initialized yet.
 200          *
 201          * Allow page fault since anon_vma may be not initialized until
 202          * the first page fault.
 203          */
 204         if (!vma->anon_vma)
 205                 return (smaps || in_pf) ? orders : 0;
 206
 207         return orders;
 208 }
 209
 210 static bool get_huge_zero_page(void)
 211 {
 212         struct folio *zero_folio;
 213 retry:
 214         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 215                 return true;
 216
 217         zero_folio = folio_alloc((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 218                         HPAGE_PMD_ORDER);
 219         if (!zero_folio) {
 220                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 221                 return false;
 222         }
 223         preempt_disable();
 224         if (cmpxchg(&huge_zero_folio, NULL, zero_folio)) {
 225                 preempt_enable();
 226                 folio_put(zero_folio);
 227                 goto retry;
 228         }
 229         WRITE_ONCE(huge_zero_pfn, folio_pfn(zero_folio));
 230
 231         /* We take additional reference here. It will be put back by shrinker */
 232         atomic_set(&huge_zero_refcount, 2);
 233         preempt_enable();
 234         count_vm_event(THP_ZERO_PAGE_ALLOC);
 235         return true;
 236 }
 237
 238 static void put_huge_zero_page(void)
 239 {
 240         /*
 241          * Counter should never go to zero here. Only shrinker can put
 242          * last reference.
 243          */
 244         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 245 }
 246
 247 struct folio *mm_get_huge_zero_folio(struct mm_struct *mm)
 248 {
 249         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 250                 return READ_ONCE(huge_zero_folio);
 251
 252         if (!get_huge_zero_page())
 253                 return NULL;
 254
 255         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 256                 put_huge_zero_page();
 257
 258         return READ_ONCE(huge_zero_folio);
 259 }
 260
 261 void mm_put_huge_zero_folio(struct mm_struct *mm)
 262 {
 263         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 264                 put_huge_zero_page();
 265 }
 266
 267 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 268                                         struct shrink_control *sc)
 269 {
 270         /* we can free zero page only if last reference remains */
 271         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 272 }
 273
 274 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 275                                        struct shrink_control *sc)
 276 {
 277         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 278                 struct folio *zero_folio = xchg(&huge_zero_folio, NULL);
 279                 BUG_ON(zero_folio == NULL);
 280                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 281                 folio_put(zero_folio);
 282                 return HPAGE_PMD_NR;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static struct shrinker *huge_zero_page_shrinker;
 289
 290 #ifdef CONFIG_SYSFS
 291 static ssize_t enabled_show(struct kobject *kobj,
 292                             struct kobj_attribute *attr, char *buf)
 293 {
 294         const char *output;
 295
 296         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 297                 output = "[always] madvise never";
 298         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 299                           &transparent_hugepage_flags))
 300                 output = "always [madvise] never";
 301         else
 302                 output = "always madvise [never]";
 303
 304         return sysfs_emit(buf, "%s\n", output);
 305 }
 306
 307 static ssize_t enabled_store(struct kobject *kobj,
 308                              struct kobj_attribute *attr,
 309                              const char *buf, size_t count)
 310 {
 311         ssize_t ret = count;
 312
 313         if (sysfs_streq(buf, "always")) {
 314                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 315                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 316         } else if (sysfs_streq(buf, "madvise")) {
 317                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 318                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 319         } else if (sysfs_streq(buf, "never")) {
 320                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 321                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 322         } else
 323                 ret = -EINVAL;
 324
 325         if (ret > 0) {
 326                 int err = start_stop_khugepaged();
 327                 if (err)
 328                         ret = err;
 329         }
 330         return ret;
 331 }
 332
 333 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 334
 335 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 336                                   struct kobj_attribute *attr, char *buf,
 337                                   enum transparent_hugepage_flag flag)
 338 {
 339         return sysfs_emit(buf, "%d\n",
 340                           !!test_bit(flag, &transparent_hugepage_flags));
 341 }
 342
 343 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 344                                  struct kobj_attribute *attr,
 345                                  const char *buf, size_t count,
 346                                  enum transparent_hugepage_flag flag)
 347 {
 348         unsigned long value;
 349         int ret;
 350
 351         ret = kstrtoul(buf, 10, &value);
 352         if (ret < 0)
 353                 return ret;
 354         if (value > 1)
 355                 return -EINVAL;
 356
 357         if (value)
 358                 set_bit(flag, &transparent_hugepage_flags);
 359         else
 360                 clear_bit(flag, &transparent_hugepage_flags);
 361
 362         return count;
 363 }
 364
 365 static ssize_t defrag_show(struct kobject *kobj,
 366                            struct kobj_attribute *attr, char *buf)
 367 {
 368         const char *output;
 369
 370         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 371                      &transparent_hugepage_flags))
 372                 output = "[always] defer defer+madvise madvise never";
 373         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 374                           &transparent_hugepage_flags))
 375                 output = "always [defer] defer+madvise madvise never";
 376         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 377                           &transparent_hugepage_flags))
 378                 output = "always defer [defer+madvise] madvise never";
 379         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 380                           &transparent_hugepage_flags))
 381                 output = "always defer defer+madvise [madvise] never";
 382         else
 383                 output = "always defer defer+madvise madvise [never]";
 384
 385         return sysfs_emit(buf, "%s\n", output);
 386 }
 387
 388 static ssize_t defrag_store(struct kobject *kobj,
 389                             struct kobj_attribute *attr,
 390                             const char *buf, size_t count)
 391 {
 392         if (sysfs_streq(buf, "always")) {
 393                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 394                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 395                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 396                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 397         } else if (sysfs_streq(buf, "defer+madvise")) {
 398                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 399                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 400                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 401                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 402         } else if (sysfs_streq(buf, "defer")) {
 403                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 404                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 405                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 406                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 407         } else if (sysfs_streq(buf, "madvise")) {
 408                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 409                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 410                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 411                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 412         } else if (sysfs_streq(buf, "never")) {
 413                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 414                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 415                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 416                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 417         } else
 418                 return -EINVAL;
 419
 420         return count;
 421 }
 422 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 423
 424 static ssize_t use_zero_page_show(struct kobject *kobj,
 425                                   struct kobj_attribute *attr, char *buf)
 426 {
 427         return single_hugepage_flag_show(kobj, attr, buf,
 428                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 429 }
 430 static ssize_t use_zero_page_store(struct kobject *kobj,
 431                 struct kobj_attribute *attr, const char *buf, size_t count)
 432 {
 433         return single_hugepage_flag_store(kobj, attr, buf, count,
 434                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 435 }
 436 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 437
 438 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 439                                    struct kobj_attribute *attr, char *buf)
 440 {
 441         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 442 }
 443 static struct kobj_attribute hpage_pmd_size_attr =
 444         __ATTR_RO(hpage_pmd_size);
 445
 446 static struct attribute *hugepage_attr[] = {
 447         &enabled_attr.attr,
 448         &defrag_attr.attr,
 449         &use_zero_page_attr.attr,
 450         &hpage_pmd_size_attr.attr,
 451 #ifdef CONFIG_SHMEM
 452         &shmem_enabled_attr.attr,
 453 #endif
 454         NULL,
 455 };
 456
 457 static const struct attribute_group hugepage_attr_group = {
 458         .attrs = hugepage_attr,
 459 };
 460
 461 static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);
 462 static void thpsize_release(struct kobject *kobj);
 463 static DEFINE_SPINLOCK(huge_anon_orders_lock);
 464 static LIST_HEAD(thpsize_list);
 465
 466 static ssize_t thpsize_enabled_show(struct kobject *kobj,
 467                                     struct kobj_attribute *attr, char *buf)
 468 {
 469         int order = to_thpsize(kobj)->order;
 470         const char *output;
 471
 472         if (test_bit(order, &huge_anon_orders_always))
 473                 output = "[always] inherit madvise never";
 474         else if (test_bit(order, &huge_anon_orders_inherit))
 475                 output = "always [inherit] madvise never";
 476         else if (test_bit(order, &huge_anon_orders_madvise))
 477                 output = "always inherit [madvise] never";
 478         else
 479                 output = "always inherit madvise [never]";
 480
 481         return sysfs_emit(buf, "%s\n", output);
 482 }
 483
 484 static ssize_t thpsize_enabled_store(struct kobject *kobj,
 485                                      struct kobj_attribute *attr,
 486                                      const char *buf, size_t count)
 487 {
 488         int order = to_thpsize(kobj)->order;
 489         ssize_t ret = count;
 490
 491         if (sysfs_streq(buf, "always")) {
 492                 spin_lock(&huge_anon_orders_lock);
 493                 clear_bit(order, &huge_anon_orders_inherit);
 494                 clear_bit(order, &huge_anon_orders_madvise);
 495                 set_bit(order, &huge_anon_orders_always);
 496                 spin_unlock(&huge_anon_orders_lock);
 497         } else if (sysfs_streq(buf, "inherit")) {
 498                 spin_lock(&huge_anon_orders_lock);
 499                 clear_bit(order, &huge_anon_orders_always);
 500                 clear_bit(order, &huge_anon_orders_madvise);
 501                 set_bit(order, &huge_anon_orders_inherit);
 502                 spin_unlock(&huge_anon_orders_lock);
 503         } else if (sysfs_streq(buf, "madvise")) {
 504                 spin_lock(&huge_anon_orders_lock);
 505                 clear_bit(order, &huge_anon_orders_always);
 506                 clear_bit(order, &huge_anon_orders_inherit);
 507                 set_bit(order, &huge_anon_orders_madvise);
 508                 spin_unlock(&huge_anon_orders_lock);
 509         } else if (sysfs_streq(buf, "never")) {
 510                 spin_lock(&huge_anon_orders_lock);
 511                 clear_bit(order, &huge_anon_orders_always);
 512                 clear_bit(order, &huge_anon_orders_inherit);
 513                 clear_bit(order, &huge_anon_orders_madvise);
 514                 spin_unlock(&huge_anon_orders_lock);
 515         } else
 516                 ret = -EINVAL;
 517
 518         if (ret > 0) {
 519                 int err;
 520
 521                 err = start_stop_khugepaged();
 522                 if (err)
 523                         ret = err;
 524         }
 525         return ret;
 526 }
 527
 528 static struct kobj_attribute thpsize_enabled_attr =
 529         __ATTR(enabled, 0644, thpsize_enabled_show, thpsize_enabled_store);
 530
 531 static struct attribute *thpsize_attrs[] = {
 532         &thpsize_enabled_attr.attr,
 533 #ifdef CONFIG_SHMEM
 534         &thpsize_shmem_enabled_attr.attr,
 535 #endif
 536         NULL,
 537 };
 538
 539 static const struct attribute_group thpsize_attr_group = {
 540         .attrs = thpsize_attrs,
 541 };
 542
 543 static const struct kobj_type thpsize_ktype = {
 544         .release = &thpsize_release,
 545         .sysfs_ops = &kobj_sysfs_ops,
 546 };
 547
 548 DEFINE_PER_CPU(struct mthp_stat, mthp_stats) = {{{0}}};
 549
 550 static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item)
 551 {
 552         unsigned long sum = 0;
 553         int cpu;
 554
 555         for_each_possible_cpu(cpu) {
 556                 struct mthp_stat *this = &per_cpu(mthp_stats, cpu);
 557
 558                 sum += this->stats[order][item];
 559         }
 560
 561         return sum;
 562 }
 563
 564 #define DEFINE_MTHP_STAT_ATTR(_name, _index)                            \
 565 static ssize_t _name##_show(struct kobject *kobj,                       \
 566                         struct kobj_attribute *attr, char *buf)         \
 567 {                                                                       \
 568         int order = to_thpsize(kobj)->order;                            \
 569                                                                         \
 570         return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index));  \
 571 }                                                                       \
 572 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
 573
 574 DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC);
 575 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK);
 576 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
 577 DEFINE_MTHP_STAT_ATTR(swpout, MTHP_STAT_SWPOUT);
 578 DEFINE_MTHP_STAT_ATTR(swpout_fallback, MTHP_STAT_SWPOUT_FALLBACK);
 579 DEFINE_MTHP_STAT_ATTR(shmem_alloc, MTHP_STAT_SHMEM_ALLOC);
 580 DEFINE_MTHP_STAT_ATTR(shmem_fallback, MTHP_STAT_SHMEM_FALLBACK);
 581 DEFINE_MTHP_STAT_ATTR(shmem_fallback_charge, MTHP_STAT_SHMEM_FALLBACK_CHARGE);
 582 DEFINE_MTHP_STAT_ATTR(split, MTHP_STAT_SPLIT);
 583 DEFINE_MTHP_STAT_ATTR(split_failed, MTHP_STAT_SPLIT_FAILED);
 584 DEFINE_MTHP_STAT_ATTR(split_deferred, MTHP_STAT_SPLIT_DEFERRED);
 585
 586 static struct attribute *stats_attrs[] = {
 587         &anon_fault_alloc_attr.attr,
 588         &anon_fault_fallback_attr.attr,
 589         &anon_fault_fallback_charge_attr.attr,
 590         &swpout_attr.attr,
 591         &swpout_fallback_attr.attr,
 592         &shmem_alloc_attr.attr,
 593         &shmem_fallback_attr.attr,
 594         &shmem_fallback_charge_attr.attr,
 595         &split_attr.attr,
 596         &split_failed_attr.attr,
 597         &split_deferred_attr.attr,
 598         NULL,
 599 };
 600
 601 static struct attribute_group stats_attr_group = {
 602         .name = "stats",
 603         .attrs = stats_attrs,
 604 };
 605
 606 static struct thpsize *thpsize_create(int order, struct kobject *parent)
 607 {
 608         unsigned long size = (PAGE_SIZE << order) / SZ_1K;
 609         struct thpsize *thpsize;
 610         int ret;
 611
 612         thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);
 613         if (!thpsize)
 614                 return ERR_PTR(-ENOMEM);
 615
 616         ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,
 617                                    "hugepages-%lukB", size);
 618         if (ret) {
 619                 kfree(thpsize);
 620                 return ERR_PTR(ret);
 621         }
 622
 623         ret = sysfs_create_group(&thpsize->kobj, &thpsize_attr_group);
 624         if (ret) {
 625                 kobject_put(&thpsize->kobj);
 626                 return ERR_PTR(ret);
 627         }
 628
 629         ret = sysfs_create_group(&thpsize->kobj, &stats_attr_group);
 630         if (ret) {
 631                 kobject_put(&thpsize->kobj);
 632                 return ERR_PTR(ret);
 633         }
 634
 635         thpsize->order = order;
 636         return thpsize;
 637 }
 638
 639 static void thpsize_release(struct kobject *kobj)
 640 {
 641         kfree(to_thpsize(kobj));
 642 }
 643
 644 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 645 {
 646         int err;
 647         struct thpsize *thpsize;
 648         unsigned long orders;
 649         int order;
 650
 651         /*
 652          * Default to setting PMD-sized THP to inherit the global setting and
 653          * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time
 654          * constant so we have to do this here.
 655          */
 656         huge_anon_orders_inherit = BIT(PMD_ORDER);
 657
 658         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 659         if (unlikely(!*hugepage_kobj)) {
 660                 pr_err("failed to create transparent hugepage kobject\n");
 661                 return -ENOMEM;
 662         }
 663
 664         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 665         if (err) {
 666                 pr_err("failed to register transparent hugepage group\n");
 667                 goto delete_obj;
 668         }
 669
 670         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 671         if (err) {
 672                 pr_err("failed to register transparent hugepage group\n");
 673                 goto remove_hp_group;
 674         }
 675
 676         orders = THP_ORDERS_ALL_ANON;
 677         order = highest_order(orders);
 678         while (orders) {
 679                 thpsize = thpsize_create(order, *hugepage_kobj);
 680                 if (IS_ERR(thpsize)) {
 681                         pr_err("failed to create thpsize for order %d\n", order);
 682                         err = PTR_ERR(thpsize);
 683                         goto remove_all;
 684                 }
 685                 list_add(&thpsize->node, &thpsize_list);
 686                 order = next_order(&orders, order);
 687         }
 688
 689         return 0;
 690
 691 remove_all:
 692         hugepage_exit_sysfs(*hugepage_kobj);
 693         return err;
 694 remove_hp_group:
 695         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 696 delete_obj:
 697         kobject_put(*hugepage_kobj);
 698         return err;
 699 }
 700
 701 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 702 {
 703         struct thpsize *thpsize, *tmp;
 704
 705         list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {
 706                 list_del(&thpsize->node);
 707                 kobject_put(&thpsize->kobj);
 708         }
 709
 710         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 711         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 712         kobject_put(hugepage_kobj);
 713 }
 714 #else
 715 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 716 {
 717         return 0;
 718 }
 719
 720 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 721 {
 722 }
 723 #endif /* CONFIG_SYSFS */
 724
 725 static int __init thp_shrinker_init(void)
 726 {
 727         huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");
 728         if (!huge_zero_page_shrinker)
 729                 return -ENOMEM;
 730
 731         deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
 732                                                  SHRINKER_MEMCG_AWARE |
 733                                                  SHRINKER_NONSLAB,
 734                                                  "thp-deferred_split");
 735         if (!deferred_split_shrinker) {
 736                 shrinker_free(huge_zero_page_shrinker);
 737                 return -ENOMEM;
 738         }
 739
 740         huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;
 741         huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;
 742         shrinker_register(huge_zero_page_shrinker);
 743
 744         deferred_split_shrinker->count_objects = deferred_split_count;
 745         deferred_split_shrinker->scan_objects = deferred_split_scan;
 746         shrinker_register(deferred_split_shrinker);
 747
 748         return 0;
 749 }
 750
 751 static void __init thp_shrinker_exit(void)
 752 {
 753         shrinker_free(huge_zero_page_shrinker);
 754         shrinker_free(deferred_split_shrinker);
 755 }
 756
 757 static int __init hugepage_init(void)
 758 {
 759         int err;
 760         struct kobject *hugepage_kobj;
 761
 762         if (!has_transparent_hugepage()) {
 763                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
 764                 return -EINVAL;
 765         }
 766
 767         /*
 768          * hugepages can't be allocated by the buddy allocator
 769          */
 770         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);
 771
 772         err = hugepage_init_sysfs(&hugepage_kobj);
 773         if (err)
 774                 goto err_sysfs;
 775
 776         err = khugepaged_init();
 777         if (err)
 778                 goto err_slab;
 779
 780         err = thp_shrinker_init();
 781         if (err)
 782                 goto err_shrinker;
 783
 784         /*
 785          * By default disable transparent hugepages on smaller systems,
 786          * where the extra memory used could hurt more than TLB overhead
 787          * is likely to save.  The admin can still enable it through /sys.
 788          */
 789         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 790                 transparent_hugepage_flags = 0;
 791                 return 0;
 792         }
 793
 794         err = start_stop_khugepaged();
 795         if (err)
 796                 goto err_khugepaged;
 797
 798         return 0;
 799 err_khugepaged:
 800         thp_shrinker_exit();
 801 err_shrinker:
 802         khugepaged_destroy();
 803 err_slab:
 804         hugepage_exit_sysfs(hugepage_kobj);
 805 err_sysfs:
 806         return err;
 807 }
 808 subsys_initcall(hugepage_init);
 809
 810 static int __init setup_transparent_hugepage(char *str)
 811 {
 812         int ret = 0;
 813         if (!str)
 814                 goto out;
 815         if (!strcmp(str, "always")) {
 816                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 817                         &transparent_hugepage_flags);
 818                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 819                           &transparent_hugepage_flags);
 820                 ret = 1;
 821         } else if (!strcmp(str, "madvise")) {
 822                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 823                           &transparent_hugepage_flags);
 824                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 825                         &transparent_hugepage_flags);
 826                 ret = 1;
 827         } else if (!strcmp(str, "never")) {
 828                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 829                           &transparent_hugepage_flags);
 830                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 831                           &transparent_hugepage_flags);
 832                 ret = 1;
 833         }
 834 out:
 835         if (!ret)
 836                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 837         return ret;
 838 }
 839 __setup("transparent_hugepage=", setup_transparent_hugepage);
 840
 841 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 842 {
 843         if (likely(vma->vm_flags & VM_WRITE))
 844                 pmd = pmd_mkwrite(pmd, vma);
 845         return pmd;
 846 }
 847
 848 #ifdef CONFIG_MEMCG
 849 static inline
 850 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 851 {
 852         struct mem_cgroup *memcg = folio_memcg(folio);
 853         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 854
 855         if (memcg)
 856                 return &memcg->deferred_split_queue;
 857         else
 858                 return &pgdat->deferred_split_queue;
 859 }
 860 #else
 861 static inline
 862 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 863 {
 864         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 865
 866         return &pgdat->deferred_split_queue;
 867 }
 868 #endif
 869
 870 static inline bool is_transparent_hugepage(const struct folio *folio)
 871 {
 872         if (!folio_test_large(folio))
 873                 return false;
 874
 875         return is_huge_zero_folio(folio) ||
 876                 folio_test_large_rmappable(folio);
 877 }
 878
 879 static unsigned long __thp_get_unmapped_area(struct file *filp,
 880                 unsigned long addr, unsigned long len,
 881                 loff_t off, unsigned long flags, unsigned long size,
 882                 vm_flags_t vm_flags)
 883 {
 884         loff_t off_end = off + len;
 885         loff_t off_align = round_up(off, size);
 886         unsigned long len_pad, ret, off_sub;
 887
 888         if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())
 889                 return 0;
 890
 891         if (off_end <= off_align || (off_end - off_align) < size)
 892                 return 0;
 893
 894         len_pad = len + size;
 895         if (len_pad < len || (off + len_pad) < off)
 896                 return 0;
 897
 898         ret = mm_get_unmapped_area_vmflags(current->mm, filp, addr, len_pad,
 899                                            off >> PAGE_SHIFT, flags, vm_flags);
 900
 901         /*
 902          * The failure might be due to length padding. The caller will retry
 903          * without the padding.
 904          */
 905         if (IS_ERR_VALUE(ret))
 906                 return 0;
 907
 908         /*
 909          * Do not try to align to THP boundary if allocation at the address
 910          * hint succeeds.
 911          */
 912         if (ret == addr)
 913                 return addr;
 914
 915         off_sub = (off - ret) & (size - 1);
 916
 917         if (test_bit(MMF_TOPDOWN, &current->mm->flags) && !off_sub)
 918                 return ret + size;
 919
 920         ret += off_sub;
 921         return ret;
 922 }
 923
 924 unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,
 925                 unsigned long len, unsigned long pgoff, unsigned long flags,
 926                 vm_flags_t vm_flags)
 927 {
 928         unsigned long ret;
 929         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 930
 931         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE, vm_flags);
 932         if (ret)
 933                 return ret;
 934
 935         return mm_get_unmapped_area_vmflags(current->mm, filp, addr, len, pgoff, flags,
 936                                             vm_flags);
 937 }
 938
 939 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 940                 unsigned long len, unsigned long pgoff, unsigned long flags)
 941 {
 942         return thp_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags, 0);
 943 }
 944 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 945
 946 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 947                         struct page *page, gfp_t gfp)
 948 {
 949         struct vm_area_struct *vma = vmf->vma;
 950         struct folio *folio = page_folio(page);
 951         pgtable_t pgtable;
 952         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 953         vm_fault_t ret = 0;
 954
 955         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
 956
 957         if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
 958                 folio_put(folio);
 959                 count_vm_event(THP_FAULT_FALLBACK);
 960                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 961                 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK);
 962                 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
 963                 return VM_FAULT_FALLBACK;
 964         }
 965         folio_throttle_swaprate(folio, gfp);
 966
 967         pgtable = pte_alloc_one(vma->vm_mm);
 968         if (unlikely(!pgtable)) {
 969                 ret = VM_FAULT_OOM;
 970                 goto release;
 971         }
 972
 973         folio_zero_user(folio, vmf->address);
 974         /*
 975          * The memory barrier inside __folio_mark_uptodate makes sure that
 976          * folio_zero_user writes become visible before the set_pmd_at()
 977          * write.
 978          */
 979         __folio_mark_uptodate(folio);
 980
 981         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 982         if (unlikely(!pmd_none(*vmf->pmd))) {
 983                 goto unlock_release;
 984         } else {
 985                 pmd_t entry;
 986
 987                 ret = check_stable_address_space(vma->vm_mm);
 988                 if (ret)
 989                         goto unlock_release;
 990
 991                 /* Deliver the page fault to userland */
 992                 if (userfaultfd_missing(vma)) {
 993                         spin_unlock(vmf->ptl);
 994                         folio_put(folio);
 995                         pte_free(vma->vm_mm, pgtable);
 996                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 997                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 998                         return ret;
 999                 }
1000
1001                 entry = mk_huge_pmd(page, vma->vm_page_prot);
1002                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1003                 folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE);
1004                 folio_add_lru_vma(folio, vma);
1005                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1006                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1007                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1008                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1009                 mm_inc_nr_ptes(vma->vm_mm);
1010                 spin_unlock(vmf->ptl);
1011                 count_vm_event(THP_FAULT_ALLOC);
1012                 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
1013                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
1014         }
1015
1016         return 0;
1017 unlock_release:
1018         spin_unlock(vmf->ptl);
1019 release:
1020         if (pgtable)
1021                 pte_free(vma->vm_mm, pgtable);
1022         folio_put(folio);
1023         return ret;
1024
1025 }
1026
1027 /*
1028  * always: directly stall for all thp allocations
1029  * defer: wake kswapd and fail if not immediately available
1030  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
1031  *                fail if not immediately available
1032  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
1033  *          available
1034  * never: never stall for any thp allocation
1035  */
1036 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
1037 {
1038         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
1039
1040         /* Always do synchronous compaction */
1041         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
1042                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
1043
1044         /* Kick kcompactd and fail quickly */
1045         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
1046                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
1047
1048         /* Synchronous compaction if madvised, otherwise kick kcompactd */
1049         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
1050                 return GFP_TRANSHUGE_LIGHT |
1051                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
1052                                         __GFP_KSWAPD_RECLAIM);
1053
1054         /* Only do synchronous compaction if madvised */
1055         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
1056                 return GFP_TRANSHUGE_LIGHT |
1057                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
1058
1059         return GFP_TRANSHUGE_LIGHT;
1060 }
1061
1062 /* Caller must hold page table lock. */
1063 static void set_huge_zero_folio(pgtable_t pgtable, struct mm_struct *mm,
1064                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
1065                 struct folio *zero_folio)
1066 {
1067         pmd_t entry;
1068         if (!pmd_none(*pmd))
1069                 return;
1070         entry = mk_pmd(&zero_folio->page, vma->vm_page_prot);
1071         entry = pmd_mkhuge(entry);
1072         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1073         set_pmd_at(mm, haddr, pmd, entry);
1074         mm_inc_nr_ptes(mm);
1075 }
1076
1077 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
1078 {
1079         struct vm_area_struct *vma = vmf->vma;
1080         gfp_t gfp;
1081         struct folio *folio;
1082         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1083         vm_fault_t ret;
1084
1085         if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
1086                 return VM_FAULT_FALLBACK;
1087         ret = vmf_anon_prepare(vmf);
1088         if (ret)
1089                 return ret;
1090         khugepaged_enter_vma(vma, vma->vm_flags);
1091
1092         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
1093                         !mm_forbids_zeropage(vma->vm_mm) &&
1094                         transparent_hugepage_use_zero_page()) {
1095                 pgtable_t pgtable;
1096                 struct folio *zero_folio;
1097                 vm_fault_t ret;
1098
1099                 pgtable = pte_alloc_one(vma->vm_mm);
1100                 if (unlikely(!pgtable))
1101                         return VM_FAULT_OOM;
1102                 zero_folio = mm_get_huge_zero_folio(vma->vm_mm);
1103                 if (unlikely(!zero_folio)) {
1104                         pte_free(vma->vm_mm, pgtable);
1105                         count_vm_event(THP_FAULT_FALLBACK);
1106                         return VM_FAULT_FALLBACK;
1107                 }
1108                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1109                 ret = 0;
1110                 if (pmd_none(*vmf->pmd)) {
1111                         ret = check_stable_address_space(vma->vm_mm);
1112                         if (ret) {
1113                                 spin_unlock(vmf->ptl);
1114                                 pte_free(vma->vm_mm, pgtable);
1115                         } else if (userfaultfd_missing(vma)) {
1116                                 spin_unlock(vmf->ptl);
1117                                 pte_free(vma->vm_mm, pgtable);
1118                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
1119                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1120                         } else {
1121                                 set_huge_zero_folio(pgtable, vma->vm_mm, vma,
1122                                                    haddr, vmf->pmd, zero_folio);
1123                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1124                                 spin_unlock(vmf->ptl);
1125                         }
1126                 } else {
1127                         spin_unlock(vmf->ptl);
1128                         pte_free(vma->vm_mm, pgtable);
1129                 }
1130                 return ret;
1131         }
1132         gfp = vma_thp_gfp_mask(vma);
1133         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
1134         if (unlikely(!folio)) {
1135                 count_vm_event(THP_FAULT_FALLBACK);
1136                 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK);
1137                 return VM_FAULT_FALLBACK;
1138         }
1139         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
1140 }
1141
1142 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1143                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
1144                 pgtable_t pgtable)
1145 {
1146         struct mm_struct *mm = vma->vm_mm;
1147         pmd_t entry;
1148         spinlock_t *ptl;
1149
1150         ptl = pmd_lock(mm, pmd);
1151         if (!pmd_none(*pmd)) {
1152                 if (write) {
1153                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
1154                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
1155                                 goto out_unlock;
1156                         }
1157                         entry = pmd_mkyoung(*pmd);
1158                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1159                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
1160                                 update_mmu_cache_pmd(vma, addr, pmd);
1161                 }
1162
1163                 goto out_unlock;
1164         }
1165
1166         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1167         if (pfn_t_devmap(pfn))
1168                 entry = pmd_mkdevmap(entry);
1169         if (write) {
1170                 entry = pmd_mkyoung(pmd_mkdirty(entry));
1171                 entry = maybe_pmd_mkwrite(entry, vma);
1172         }
1173
1174         if (pgtable) {
1175                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1176                 mm_inc_nr_ptes(mm);
1177                 pgtable = NULL;
1178         }
1179
1180         set_pmd_at(mm, addr, pmd, entry);
1181         update_mmu_cache_pmd(vma, addr, pmd);
1182
1183 out_unlock:
1184         spin_unlock(ptl);
1185         if (pgtable)
1186                 pte_free(mm, pgtable);
1187 }
1188
1189 /**
1190  * vmf_insert_pfn_pmd - insert a pmd size pfn
1191  * @vmf: Structure describing the fault
1192  * @pfn: pfn to insert
1193  * @write: whether it's a write fault
1194  *
1195  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
1196  *
1197  * Return: vm_fault_t value.
1198  */
1199 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
1200 {
1201         unsigned long addr = vmf->address & PMD_MASK;
1202         struct vm_area_struct *vma = vmf->vma;
1203         pgprot_t pgprot = vma->vm_page_prot;
1204         pgtable_t pgtable = NULL;
1205
1206         /*
1207          * If we had pmd_special, we could avoid all these restrictions,
1208          * but we need to be consistent with PTEs and architectures that
1209          * can't support a 'special' bit.
1210          */
1211         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1212                         !pfn_t_devmap(pfn));
1213         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1214                                                 (VM_PFNMAP|VM_MIXEDMAP));
1215         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1216
1217         if (addr < vma->vm_start || addr >= vma->vm_end)
1218                 return VM_FAULT_SIGBUS;
1219
1220         if (arch_needs_pgtable_deposit()) {
1221                 pgtable = pte_alloc_one(vma->vm_mm);
1222                 if (!pgtable)
1223                         return VM_FAULT_OOM;
1224         }
1225
1226         track_pfn_insert(vma, &pgprot, pfn);
1227
1228         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
1229         return VM_FAULT_NOPAGE;
1230 }
1231 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1232
1233 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1234 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
1235 {
1236         if (likely(vma->vm_flags & VM_WRITE))
1237                 pud = pud_mkwrite(pud);
1238         return pud;
1239 }
1240
1241 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
1242                 pud_t *pud, pfn_t pfn, bool write)
1243 {
1244         struct mm_struct *mm = vma->vm_mm;
1245         pgprot_t prot = vma->vm_page_prot;
1246         pud_t entry;
1247         spinlock_t *ptl;
1248
1249         ptl = pud_lock(mm, pud);
1250         if (!pud_none(*pud)) {
1251                 if (write) {
1252                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
1253                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
1254                                 goto out_unlock;
1255                         }
1256                         entry = pud_mkyoung(*pud);
1257                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
1258                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
1259                                 update_mmu_cache_pud(vma, addr, pud);
1260                 }
1261                 goto out_unlock;
1262         }
1263
1264         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
1265         if (pfn_t_devmap(pfn))
1266                 entry = pud_mkdevmap(entry);
1267         if (write) {
1268                 entry = pud_mkyoung(pud_mkdirty(entry));
1269                 entry = maybe_pud_mkwrite(entry, vma);
1270         }
1271         set_pud_at(mm, addr, pud, entry);
1272         update_mmu_cache_pud(vma, addr, pud);
1273
1274 out_unlock:
1275         spin_unlock(ptl);
1276 }
1277
1278 /**
1279  * vmf_insert_pfn_pud - insert a pud size pfn
1280  * @vmf: Structure describing the fault
1281  * @pfn: pfn to insert
1282  * @write: whether it's a write fault
1283  *
1284  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1285  *
1286  * Return: vm_fault_t value.
1287  */
1288 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
1289 {
1290         unsigned long addr = vmf->address & PUD_MASK;
1291         struct vm_area_struct *vma = vmf->vma;
1292         pgprot_t pgprot = vma->vm_page_prot;
1293
1294         /*
1295          * If we had pud_special, we could avoid all these restrictions,
1296          * but we need to be consistent with PTEs and architectures that
1297          * can't support a 'special' bit.
1298          */
1299         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1300                         !pfn_t_devmap(pfn));
1301         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1302                                                 (VM_PFNMAP|VM_MIXEDMAP));
1303         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1304
1305         if (addr < vma->vm_start || addr >= vma->vm_end)
1306                 return VM_FAULT_SIGBUS;
1307
1308         track_pfn_insert(vma, &pgprot, pfn);
1309
1310         insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1311         return VM_FAULT_NOPAGE;
1312 }
1313 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1314 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1315
1316 void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1317                pmd_t *pmd, bool write)
1318 {
1319         pmd_t _pmd;
1320
1321         _pmd = pmd_mkyoung(*pmd);
1322         if (write)
1323                 _pmd = pmd_mkdirty(_pmd);
1324         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1325                                   pmd, _pmd, write))
1326                 update_mmu_cache_pmd(vma, addr, pmd);
1327 }
1328
1329 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1330                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1331 {
1332         unsigned long pfn = pmd_pfn(*pmd);
1333         struct mm_struct *mm = vma->vm_mm;
1334         struct page *page;
1335         int ret;
1336
1337         assert_spin_locked(pmd_lockptr(mm, pmd));
1338
1339         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1340                 return NULL;
1341
1342         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1343                 /* pass */;
1344         else
1345                 return NULL;
1346
1347         if (flags & FOLL_TOUCH)
1348                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1349
1350         /*
1351          * device mapped pages can only be returned if the
1352          * caller will manage the page reference count.
1353          */
1354         if (!(flags & (FOLL_GET | FOLL_PIN)))
1355                 return ERR_PTR(-EEXIST);
1356
1357         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1358         *pgmap = get_dev_pagemap(pfn, *pgmap);
1359         if (!*pgmap)
1360                 return ERR_PTR(-EFAULT);
1361         page = pfn_to_page(pfn);
1362         ret = try_grab_folio(page_folio(page), 1, flags);
1363         if (ret)
1364                 page = ERR_PTR(ret);
1365
1366         return page;
1367 }
1368
1369 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1370                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1371                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1372 {
1373         spinlock_t *dst_ptl, *src_ptl;
1374         struct page *src_page;
1375         struct folio *src_folio;
1376         pmd_t pmd;
1377         pgtable_t pgtable = NULL;
1378         int ret = -ENOMEM;
1379
1380         /* Skip if can be re-fill on fault */
1381         if (!vma_is_anonymous(dst_vma))
1382                 return 0;
1383
1384         pgtable = pte_alloc_one(dst_mm);
1385         if (unlikely(!pgtable))
1386                 goto out;
1387
1388         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1389         src_ptl = pmd_lockptr(src_mm, src_pmd);
1390         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1391
1392         ret = -EAGAIN;
1393         pmd = *src_pmd;
1394
1395 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1396         if (unlikely(is_swap_pmd(pmd))) {
1397                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1398
1399                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1400                 if (!is_readable_migration_entry(entry)) {
1401                         entry = make_readable_migration_entry(
1402                                                         swp_offset(entry));
1403                         pmd = swp_entry_to_pmd(entry);
1404                         if (pmd_swp_soft_dirty(*src_pmd))
1405                                 pmd = pmd_swp_mksoft_dirty(pmd);
1406                         if (pmd_swp_uffd_wp(*src_pmd))
1407                                 pmd = pmd_swp_mkuffd_wp(pmd);
1408                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1409                 }
1410                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1411                 mm_inc_nr_ptes(dst_mm);
1412                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1413                 if (!userfaultfd_wp(dst_vma))
1414                         pmd = pmd_swp_clear_uffd_wp(pmd);
1415                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1416                 ret = 0;
1417                 goto out_unlock;
1418         }
1419 #endif
1420
1421         if (unlikely(!pmd_trans_huge(pmd))) {
1422                 pte_free(dst_mm, pgtable);
1423                 goto out_unlock;
1424         }
1425         /*
1426          * When page table lock is held, the huge zero pmd should not be
1427          * under splitting since we don't split the page itself, only pmd to
1428          * a page table.
1429          */
1430         if (is_huge_zero_pmd(pmd)) {
1431                 /*
1432                  * mm_get_huge_zero_folio() will never allocate a new
1433                  * folio here, since we already have a zero page to
1434                  * copy. It just takes a reference.
1435                  */
1436                 mm_get_huge_zero_folio(dst_mm);
1437                 goto out_zero_page;
1438         }
1439
1440         src_page = pmd_page(pmd);
1441         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1442         src_folio = page_folio(src_page);
1443
1444         folio_get(src_folio);
1445         if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) {
1446                 /* Page maybe pinned: split and retry the fault on PTEs. */
1447                 folio_put(src_folio);
1448                 pte_free(dst_mm, pgtable);
1449                 spin_unlock(src_ptl);
1450                 spin_unlock(dst_ptl);
1451                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1452                 return -EAGAIN;
1453         }
1454         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1455 out_zero_page:
1456         mm_inc_nr_ptes(dst_mm);
1457         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1458         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1459         if (!userfaultfd_wp(dst_vma))
1460                 pmd = pmd_clear_uffd_wp(pmd);
1461         pmd = pmd_mkold(pmd_wrprotect(pmd));
1462         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1463
1464         ret = 0;
1465 out_unlock:
1466         spin_unlock(src_ptl);
1467         spin_unlock(dst_ptl);
1468 out:
1469         return ret;
1470 }
1471
1472 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1473 void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1474                pud_t *pud, bool write)
1475 {
1476         pud_t _pud;
1477
1478         _pud = pud_mkyoung(*pud);
1479         if (write)
1480                 _pud = pud_mkdirty(_pud);
1481         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1482                                   pud, _pud, write))
1483                 update_mmu_cache_pud(vma, addr, pud);
1484 }
1485
1486 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1487                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1488                   struct vm_area_struct *vma)
1489 {
1490         spinlock_t *dst_ptl, *src_ptl;
1491         pud_t pud;
1492         int ret;
1493
1494         dst_ptl = pud_lock(dst_mm, dst_pud);
1495         src_ptl = pud_lockptr(src_mm, src_pud);
1496         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1497
1498         ret = -EAGAIN;
1499         pud = *src_pud;
1500         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1501                 goto out_unlock;
1502
1503         /*
1504          * When page table lock is held, the huge zero pud should not be
1505          * under splitting since we don't split the page itself, only pud to
1506          * a page table.
1507          */
1508         if (is_huge_zero_pud(pud)) {
1509                 /* No huge zero pud yet */
1510         }
1511
1512         /*
1513          * TODO: once we support anonymous pages, use
1514          * folio_try_dup_anon_rmap_*() and split if duplicating fails.
1515          */
1516         pudp_set_wrprotect(src_mm, addr, src_pud);
1517         pud = pud_mkold(pud_wrprotect(pud));
1518         set_pud_at(dst_mm, addr, dst_pud, pud);
1519
1520         ret = 0;
1521 out_unlock:
1522         spin_unlock(src_ptl);
1523         spin_unlock(dst_ptl);
1524         return ret;
1525 }
1526
1527 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1528 {
1529         bool write = vmf->flags & FAULT_FLAG_WRITE;
1530
1531         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1532         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1533                 goto unlock;
1534
1535         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1536 unlock:
1537         spin_unlock(vmf->ptl);
1538 }
1539 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1540
1541 void huge_pmd_set_accessed(struct vm_fault *vmf)
1542 {
1543         bool write = vmf->flags & FAULT_FLAG_WRITE;
1544
1545         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1546         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1547                 goto unlock;
1548
1549         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1550
1551 unlock:
1552         spin_unlock(vmf->ptl);
1553 }
1554
1555 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1556 {
1557         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1558         struct vm_area_struct *vma = vmf->vma;
1559         struct folio *folio;
1560         struct page *page;
1561         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1562         pmd_t orig_pmd = vmf->orig_pmd;
1563
1564         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1565         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1566
1567         if (is_huge_zero_pmd(orig_pmd))
1568                 goto fallback;
1569
1570         spin_lock(vmf->ptl);
1571
1572         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1573                 spin_unlock(vmf->ptl);
1574                 return 0;
1575         }
1576
1577         page = pmd_page(orig_pmd);
1578         folio = page_folio(page);
1579         VM_BUG_ON_PAGE(!PageHead(page), page);
1580
1581         /* Early check when only holding the PT lock. */
1582         if (PageAnonExclusive(page))
1583                 goto reuse;
1584
1585         if (!folio_trylock(folio)) {
1586                 folio_get(folio);
1587                 spin_unlock(vmf->ptl);
1588                 folio_lock(folio);
1589                 spin_lock(vmf->ptl);
1590                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1591                         spin_unlock(vmf->ptl);
1592                         folio_unlock(folio);
1593                         folio_put(folio);
1594                         return 0;
1595                 }
1596                 folio_put(folio);
1597         }
1598
1599         /* Recheck after temporarily dropping the PT lock. */
1600         if (PageAnonExclusive(page)) {
1601                 folio_unlock(folio);
1602                 goto reuse;
1603         }
1604
1605         /*
1606          * See do_wp_page(): we can only reuse the folio exclusively if
1607          * there are no additional references. Note that we always drain
1608          * the LRU cache immediately after adding a THP.
1609          */
1610         if (folio_ref_count(folio) >
1611                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1612                 goto unlock_fallback;
1613         if (folio_test_swapcache(folio))
1614                 folio_free_swap(folio);
1615         if (folio_ref_count(folio) == 1) {
1616                 pmd_t entry;
1617
1618                 folio_move_anon_rmap(folio, vma);
1619                 SetPageAnonExclusive(page);
1620                 folio_unlock(folio);
1621 reuse:
1622                 if (unlikely(unshare)) {
1623                         spin_unlock(vmf->ptl);
1624                         return 0;
1625                 }
1626                 entry = pmd_mkyoung(orig_pmd);
1627                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1628                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1629                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1630                 spin_unlock(vmf->ptl);
1631                 return 0;
1632         }
1633
1634 unlock_fallback:
1635         folio_unlock(folio);
1636         spin_unlock(vmf->ptl);
1637 fallback:
1638         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1639         return VM_FAULT_FALLBACK;
1640 }
1641
1642 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1643                                            unsigned long addr, pmd_t pmd)
1644 {
1645         struct page *page;
1646
1647         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1648                 return false;
1649
1650         /* Don't touch entries that are not even readable (NUMA hinting). */
1651         if (pmd_protnone(pmd))
1652                 return false;
1653
1654         /* Do we need write faults for softdirty tracking? */
1655         if (pmd_needs_soft_dirty_wp(vma, pmd))
1656                 return false;
1657
1658         /* Do we need write faults for uffd-wp tracking? */
1659         if (userfaultfd_huge_pmd_wp(vma, pmd))
1660                 return false;
1661
1662         if (!(vma->vm_flags & VM_SHARED)) {
1663                 /* See can_change_pte_writable(). */
1664                 page = vm_normal_page_pmd(vma, addr, pmd);
1665                 return page && PageAnon(page) && PageAnonExclusive(page);
1666         }
1667
1668         /* See can_change_pte_writable(). */
1669         return pmd_dirty(pmd);
1670 }
1671
1672 /* NUMA hinting page fault entry point for trans huge pmds */
1673 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1674 {
1675         struct vm_area_struct *vma = vmf->vma;
1676         pmd_t oldpmd = vmf->orig_pmd;
1677         pmd_t pmd;
1678         struct folio *folio;
1679         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1680         int nid = NUMA_NO_NODE;
1681         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1682         bool writable = false;
1683         int flags = 0;
1684
1685         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1686         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1687                 spin_unlock(vmf->ptl);
1688                 return 0;
1689         }
1690
1691         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1692
1693         /*
1694          * Detect now whether the PMD could be writable; this information
1695          * is only valid while holding the PT lock.
1696          */
1697         writable = pmd_write(pmd);
1698         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1699             can_change_pmd_writable(vma, vmf->address, pmd))
1700                 writable = true;
1701
1702         folio = vm_normal_folio_pmd(vma, haddr, pmd);
1703         if (!folio)
1704                 goto out_map;
1705
1706         /* See similar comment in do_numa_page for explanation */
1707         if (!writable)
1708                 flags |= TNF_NO_GROUP;
1709
1710         nid = folio_nid(folio);
1711         /*
1712          * For memory tiering mode, cpupid of slow memory page is used
1713          * to record page access time.  So use default value.
1714          */
1715         if (node_is_toptier(nid))
1716                 last_cpupid = folio_last_cpupid(folio);
1717         target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
1718         if (target_nid == NUMA_NO_NODE)
1719                 goto out_map;
1720         if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {
1721                 flags |= TNF_MIGRATE_FAIL;
1722                 goto out_map;
1723         }
1724         /* The folio is isolated and isolation code holds a folio reference. */
1725         spin_unlock(vmf->ptl);
1726         writable = false;
1727
1728         if (!migrate_misplaced_folio(folio, vma, target_nid)) {
1729                 flags |= TNF_MIGRATED;
1730                 nid = target_nid;
1731                 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
1732                 return 0;
1733         }
1734
1735         flags |= TNF_MIGRATE_FAIL;
1736         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1737         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1738                 spin_unlock(vmf->ptl);
1739                 return 0;
1740         }
1741 out_map:
1742         /* Restore the PMD */
1743         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1744         pmd = pmd_mkyoung(pmd);
1745         if (writable)
1746                 pmd = pmd_mkwrite(pmd, vma);
1747         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1748         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1749         spin_unlock(vmf->ptl);
1750
1751         if (nid != NUMA_NO_NODE)
1752                 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
1753         return 0;
1754 }
1755
1756 /*
1757  * Return true if we do MADV_FREE successfully on entire pmd page.
1758  * Otherwise, return false.
1759  */
1760 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1761                 pmd_t *pmd, unsigned long addr, unsigned long next)
1762 {
1763         spinlock_t *ptl;
1764         pmd_t orig_pmd;
1765         struct folio *folio;
1766         struct mm_struct *mm = tlb->mm;
1767         bool ret = false;
1768
1769         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1770
1771         ptl = pmd_trans_huge_lock(pmd, vma);
1772         if (!ptl)
1773                 goto out_unlocked;
1774
1775         orig_pmd = *pmd;
1776         if (is_huge_zero_pmd(orig_pmd))
1777                 goto out;
1778
1779         if (unlikely(!pmd_present(orig_pmd))) {
1780                 VM_BUG_ON(thp_migration_supported() &&
1781                                   !is_pmd_migration_entry(orig_pmd));
1782                 goto out;
1783         }
1784
1785         folio = pmd_folio(orig_pmd);
1786         /*
1787          * If other processes are mapping this folio, we couldn't discard
1788          * the folio unless they all do MADV_FREE so let's skip the folio.
1789          */
1790         if (folio_likely_mapped_shared(folio))
1791                 goto out;
1792
1793         if (!folio_trylock(folio))
1794                 goto out;
1795
1796         /*
1797          * If user want to discard part-pages of THP, split it so MADV_FREE
1798          * will deactivate only them.
1799          */
1800         if (next - addr != HPAGE_PMD_SIZE) {
1801                 folio_get(folio);
1802                 spin_unlock(ptl);
1803                 split_folio(folio);
1804                 folio_unlock(folio);
1805                 folio_put(folio);
1806                 goto out_unlocked;
1807         }
1808
1809         if (folio_test_dirty(folio))
1810                 folio_clear_dirty(folio);
1811         folio_unlock(folio);
1812
1813         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1814                 pmdp_invalidate(vma, addr, pmd);
1815                 orig_pmd = pmd_mkold(orig_pmd);
1816                 orig_pmd = pmd_mkclean(orig_pmd);
1817
1818                 set_pmd_at(mm, addr, pmd, orig_pmd);
1819                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1820         }
1821
1822         folio_mark_lazyfree(folio);
1823         ret = true;
1824 out:
1825         spin_unlock(ptl);
1826 out_unlocked:
1827         return ret;
1828 }
1829
1830 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1831 {
1832         pgtable_t pgtable;
1833
1834         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1835         pte_free(mm, pgtable);
1836         mm_dec_nr_ptes(mm);
1837 }
1838
1839 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1840                  pmd_t *pmd, unsigned long addr)
1841 {
1842         pmd_t orig_pmd;
1843         spinlock_t *ptl;
1844
1845         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1846
1847         ptl = __pmd_trans_huge_lock(pmd, vma);
1848         if (!ptl)
1849                 return 0;
1850         /*
1851          * For architectures like ppc64 we look at deposited pgtable
1852          * when calling pmdp_huge_get_and_clear. So do the
1853          * pgtable_trans_huge_withdraw after finishing pmdp related
1854          * operations.
1855          */
1856         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1857                                                 tlb->fullmm);
1858         arch_check_zapped_pmd(vma, orig_pmd);
1859         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1860         if (vma_is_special_huge(vma)) {
1861                 if (arch_needs_pgtable_deposit())
1862                         zap_deposited_table(tlb->mm, pmd);
1863                 spin_unlock(ptl);
1864         } else if (is_huge_zero_pmd(orig_pmd)) {
1865                 zap_deposited_table(tlb->mm, pmd);
1866                 spin_unlock(ptl);
1867         } else {
1868                 struct folio *folio = NULL;
1869                 int flush_needed = 1;
1870
1871                 if (pmd_present(orig_pmd)) {
1872                         struct page *page = pmd_page(orig_pmd);
1873
1874                         folio = page_folio(page);
1875                         folio_remove_rmap_pmd(folio, page, vma);
1876                         WARN_ON_ONCE(folio_mapcount(folio) < 0);
1877                         VM_BUG_ON_PAGE(!PageHead(page), page);
1878                 } else if (thp_migration_supported()) {
1879                         swp_entry_t entry;
1880
1881                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1882                         entry = pmd_to_swp_entry(orig_pmd);
1883                         folio = pfn_swap_entry_folio(entry);
1884                         flush_needed = 0;
1885                 } else
1886                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1887
1888                 if (folio_test_anon(folio)) {
1889                         zap_deposited_table(tlb->mm, pmd);
1890                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1891                 } else {
1892                         if (arch_needs_pgtable_deposit())
1893                                 zap_deposited_table(tlb->mm, pmd);
1894                         add_mm_counter(tlb->mm, mm_counter_file(folio),
1895                                        -HPAGE_PMD_NR);
1896                 }
1897
1898                 spin_unlock(ptl);
1899                 if (flush_needed)
1900                         tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE);
1901         }
1902         return 1;
1903 }
1904
1905 #ifndef pmd_move_must_withdraw
1906 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1907                                          spinlock_t *old_pmd_ptl,
1908                                          struct vm_area_struct *vma)
1909 {
1910         /*
1911          * With split pmd lock we also need to move preallocated
1912          * PTE page table if new_pmd is on different PMD page table.
1913          *
1914          * We also don't deposit and withdraw tables for file pages.
1915          */
1916         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1917 }
1918 #endif
1919
1920 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1921 {
1922 #ifdef CONFIG_MEM_SOFT_DIRTY
1923         if (unlikely(is_pmd_migration_entry(pmd)))
1924                 pmd = pmd_swp_mksoft_dirty(pmd);
1925         else if (pmd_present(pmd))
1926                 pmd = pmd_mksoft_dirty(pmd);
1927 #endif
1928         return pmd;
1929 }
1930
1931 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1932                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1933 {
1934         spinlock_t *old_ptl, *new_ptl;
1935         pmd_t pmd;
1936         struct mm_struct *mm = vma->vm_mm;
1937         bool force_flush = false;
1938
1939         /*
1940          * The destination pmd shouldn't be established, free_pgtables()
1941          * should have released it; but move_page_tables() might have already
1942          * inserted a page table, if racing against shmem/file collapse.
1943          */
1944         if (!pmd_none(*new_pmd)) {
1945                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1946                 return false;
1947         }
1948
1949         /*
1950          * We don't have to worry about the ordering of src and dst
1951          * ptlocks because exclusive mmap_lock prevents deadlock.
1952          */
1953         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1954         if (old_ptl) {
1955                 new_ptl = pmd_lockptr(mm, new_pmd);
1956                 if (new_ptl != old_ptl)
1957                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1958                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1959                 if (pmd_present(pmd))
1960                         force_flush = true;
1961                 VM_BUG_ON(!pmd_none(*new_pmd));
1962
1963                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1964                         pgtable_t pgtable;
1965                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1966                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1967                 }
1968                 pmd = move_soft_dirty_pmd(pmd);
1969                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1970                 if (force_flush)
1971                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1972                 if (new_ptl != old_ptl)
1973                         spin_unlock(new_ptl);
1974                 spin_unlock(old_ptl);
1975                 return true;
1976         }
1977         return false;
1978 }
1979
1980 /*
1981  * Returns
1982  *  - 0 if PMD could not be locked
1983  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1984  *      or if prot_numa but THP migration is not supported
1985  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1986  */
1987 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1988                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1989                     unsigned long cp_flags)
1990 {
1991         struct mm_struct *mm = vma->vm_mm;
1992         spinlock_t *ptl;
1993         pmd_t oldpmd, entry;
1994         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1995         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1996         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1997         int ret = 1;
1998
1999         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2000
2001         if (prot_numa && !thp_migration_supported())
2002                 return 1;
2003
2004         ptl = __pmd_trans_huge_lock(pmd, vma);
2005         if (!ptl)
2006                 return 0;
2007
2008 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2009         if (is_swap_pmd(*pmd)) {
2010                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2011                 struct folio *folio = pfn_swap_entry_folio(entry);
2012                 pmd_t newpmd;
2013
2014                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2015                 if (is_writable_migration_entry(entry)) {
2016                         /*
2017                          * A protection check is difficult so
2018                          * just be safe and disable write
2019                          */
2020                         if (folio_test_anon(folio))
2021                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
2022                         else
2023                                 entry = make_readable_migration_entry(swp_offset(entry));
2024                         newpmd = swp_entry_to_pmd(entry);
2025                         if (pmd_swp_soft_dirty(*pmd))
2026                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
2027                 } else {
2028                         newpmd = *pmd;
2029                 }
2030
2031                 if (uffd_wp)
2032                         newpmd = pmd_swp_mkuffd_wp(newpmd);
2033                 else if (uffd_wp_resolve)
2034                         newpmd = pmd_swp_clear_uffd_wp(newpmd);
2035                 if (!pmd_same(*pmd, newpmd))
2036                         set_pmd_at(mm, addr, pmd, newpmd);
2037                 goto unlock;
2038         }
2039 #endif
2040
2041         if (prot_numa) {
2042                 struct folio *folio;
2043                 bool toptier;
2044                 /*
2045                  * Avoid trapping faults against the zero page. The read-only
2046                  * data is likely to be read-cached on the local CPU and
2047                  * local/remote hits to the zero page are not interesting.
2048                  */
2049                 if (is_huge_zero_pmd(*pmd))
2050                         goto unlock;
2051
2052                 if (pmd_protnone(*pmd))
2053                         goto unlock;
2054
2055                 folio = pmd_folio(*pmd);
2056                 toptier = node_is_toptier(folio_nid(folio));
2057                 /*
2058                  * Skip scanning top tier node if normal numa
2059                  * balancing is disabled
2060                  */
2061                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
2062                     toptier)
2063                         goto unlock;
2064
2065                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
2066                     !toptier)
2067                         folio_xchg_access_time(folio,
2068                                                jiffies_to_msecs(jiffies));
2069         }
2070         /*
2071          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
2072          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2073          * which is also under mmap_read_lock(mm):
2074          *
2075          *      CPU0:                           CPU1:
2076          *                              change_huge_pmd(prot_numa=1)
2077          *                               pmdp_huge_get_and_clear_notify()
2078          * madvise_dontneed()
2079          *  zap_pmd_range()
2080          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2081          *   // skip the pmd
2082          *                               set_pmd_at();
2083          *                               // pmd is re-established
2084          *
2085          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2086          * which may break userspace.
2087          *
2088          * pmdp_invalidate_ad() is required to make sure we don't miss
2089          * dirty/young flags set by hardware.
2090          */
2091         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
2092
2093         entry = pmd_modify(oldpmd, newprot);
2094         if (uffd_wp)
2095                 entry = pmd_mkuffd_wp(entry);
2096         else if (uffd_wp_resolve)
2097                 /*
2098                  * Leave the write bit to be handled by PF interrupt
2099                  * handler, then things like COW could be properly
2100                  * handled.
2101                  */
2102                 entry = pmd_clear_uffd_wp(entry);
2103
2104         /* See change_pte_range(). */
2105         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
2106             can_change_pmd_writable(vma, addr, entry))
2107                 entry = pmd_mkwrite(entry, vma);
2108
2109         ret = HPAGE_PMD_NR;
2110         set_pmd_at(mm, addr, pmd, entry);
2111
2112         if (huge_pmd_needs_flush(oldpmd, entry))
2113                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
2114 unlock:
2115         spin_unlock(ptl);
2116         return ret;
2117 }
2118
2119 #ifdef CONFIG_USERFAULTFD
2120 /*
2121  * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by
2122  * the caller, but it must return after releasing the page_table_lock.
2123  * Just move the page from src_pmd to dst_pmd if possible.
2124  * Return zero if succeeded in moving the page, -EAGAIN if it needs to be
2125  * repeated by the caller, or other errors in case of failure.
2126  */
2127 int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,
2128                         struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
2129                         unsigned long dst_addr, unsigned long src_addr)
2130 {
2131         pmd_t _dst_pmd, src_pmdval;
2132         struct page *src_page;
2133         struct folio *src_folio;
2134         struct anon_vma *src_anon_vma;
2135         spinlock_t *src_ptl, *dst_ptl;
2136         pgtable_t src_pgtable;
2137         struct mmu_notifier_range range;
2138         int err = 0;
2139
2140         src_pmdval = *src_pmd;
2141         src_ptl = pmd_lockptr(mm, src_pmd);
2142
2143         lockdep_assert_held(src_ptl);
2144         vma_assert_locked(src_vma);
2145         vma_assert_locked(dst_vma);
2146
2147         /* Sanity checks before the operation */
2148         if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||
2149             WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {
2150                 spin_unlock(src_ptl);
2151                 return -EINVAL;
2152         }
2153
2154         if (!pmd_trans_huge(src_pmdval)) {
2155                 spin_unlock(src_ptl);
2156                 if (is_pmd_migration_entry(src_pmdval)) {
2157                         pmd_migration_entry_wait(mm, &src_pmdval);
2158                         return -EAGAIN;
2159                 }
2160                 return -ENOENT;
2161         }
2162
2163         src_page = pmd_page(src_pmdval);
2164
2165         if (!is_huge_zero_pmd(src_pmdval)) {
2166                 if (unlikely(!PageAnonExclusive(src_page))) {
2167                         spin_unlock(src_ptl);
2168                         return -EBUSY;
2169                 }
2170
2171                 src_folio = page_folio(src_page);
2172                 folio_get(src_folio);
2173         } else
2174                 src_folio = NULL;
2175
2176         spin_unlock(src_ptl);
2177
2178         flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);
2179         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,
2180                                 src_addr + HPAGE_PMD_SIZE);
2181         mmu_notifier_invalidate_range_start(&range);
2182
2183         if (src_folio) {
2184                 folio_lock(src_folio);
2185
2186                 /*
2187                  * split_huge_page walks the anon_vma chain without the page
2188                  * lock. Serialize against it with the anon_vma lock, the page
2189                  * lock is not enough.
2190                  */
2191                 src_anon_vma = folio_get_anon_vma(src_folio);
2192                 if (!src_anon_vma) {
2193                         err = -EAGAIN;
2194                         goto unlock_folio;
2195                 }
2196                 anon_vma_lock_write(src_anon_vma);
2197         } else
2198                 src_anon_vma = NULL;
2199
2200         dst_ptl = pmd_lockptr(mm, dst_pmd);
2201         double_pt_lock(src_ptl, dst_ptl);
2202         if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||
2203                      !pmd_same(*dst_pmd, dst_pmdval))) {
2204                 err = -EAGAIN;
2205                 goto unlock_ptls;
2206         }
2207         if (src_folio) {
2208                 if (folio_maybe_dma_pinned(src_folio) ||
2209                     !PageAnonExclusive(&src_folio->page)) {
2210                         err = -EBUSY;
2211                         goto unlock_ptls;
2212                 }
2213
2214                 if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||
2215                     WARN_ON_ONCE(!folio_test_anon(src_folio))) {
2216                         err = -EBUSY;
2217                         goto unlock_ptls;
2218                 }
2219
2220                 src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2221                 /* Folio got pinned from under us. Put it back and fail the move. */
2222                 if (folio_maybe_dma_pinned(src_folio)) {
2223                         set_pmd_at(mm, src_addr, src_pmd, src_pmdval);
2224                         err = -EBUSY;
2225                         goto unlock_ptls;
2226                 }
2227
2228                 folio_move_anon_rmap(src_folio, dst_vma);
2229                 src_folio->index = linear_page_index(dst_vma, dst_addr);
2230
2231                 _dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot);
2232                 /* Follow mremap() behavior and treat the entry dirty after the move */
2233                 _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);
2234         } else {
2235                 src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2236                 _dst_pmd = mk_huge_pmd(src_page, dst_vma->vm_page_prot);
2237         }
2238         set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);
2239
2240         src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);
2241         pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);
2242 unlock_ptls:
2243         double_pt_unlock(src_ptl, dst_ptl);
2244         if (src_anon_vma) {
2245                 anon_vma_unlock_write(src_anon_vma);
2246                 put_anon_vma(src_anon_vma);
2247         }
2248 unlock_folio:
2249         /* unblock rmap walks */
2250         if (src_folio)
2251                 folio_unlock(src_folio);
2252         mmu_notifier_invalidate_range_end(&range);
2253         if (src_folio)
2254                 folio_put(src_folio);
2255         return err;
2256 }
2257 #endif /* CONFIG_USERFAULTFD */
2258
2259 /*
2260  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2261  *
2262  * Note that if it returns page table lock pointer, this routine returns without
2263  * unlocking page table lock. So callers must unlock it.
2264  */
2265 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2266 {
2267         spinlock_t *ptl;
2268         ptl = pmd_lock(vma->vm_mm, pmd);
2269         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2270                         pmd_devmap(*pmd)))
2271                 return ptl;
2272         spin_unlock(ptl);
2273         return NULL;
2274 }
2275
2276 /*
2277  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
2278  *
2279  * Note that if it returns page table lock pointer, this routine returns without
2280  * unlocking page table lock. So callers must unlock it.
2281  */
2282 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2283 {
2284         spinlock_t *ptl;
2285
2286         ptl = pud_lock(vma->vm_mm, pud);
2287         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2288                 return ptl;
2289         spin_unlock(ptl);
2290         return NULL;
2291 }
2292
2293 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2294 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2295                  pud_t *pud, unsigned long addr)
2296 {
2297         spinlock_t *ptl;
2298
2299         ptl = __pud_trans_huge_lock(pud, vma);
2300         if (!ptl)
2301                 return 0;
2302
2303         pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2304         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2305         if (vma_is_special_huge(vma)) {
2306                 spin_unlock(ptl);
2307                 /* No zero page support yet */
2308         } else {
2309                 /* No support for anonymous PUD pages yet */
2310                 BUG();
2311         }
2312         return 1;
2313 }
2314
2315 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2316                 unsigned long haddr)
2317 {
2318         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2319         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2320         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2321         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2322
2323         count_vm_event(THP_SPLIT_PUD);
2324
2325         pudp_huge_clear_flush(vma, haddr, pud);
2326 }
2327
2328 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2329                 unsigned long address)
2330 {
2331         spinlock_t *ptl;
2332         struct mmu_notifier_range range;
2333
2334         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2335                                 address & HPAGE_PUD_MASK,
2336                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2337         mmu_notifier_invalidate_range_start(&range);
2338         ptl = pud_lock(vma->vm_mm, pud);
2339         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2340                 goto out;
2341         __split_huge_pud_locked(vma, pud, range.start);
2342
2343 out:
2344         spin_unlock(ptl);
2345         mmu_notifier_invalidate_range_end(&range);
2346 }
2347 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2348
2349 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2350                 unsigned long haddr, pmd_t *pmd)
2351 {
2352         struct mm_struct *mm = vma->vm_mm;
2353         pgtable_t pgtable;
2354         pmd_t _pmd, old_pmd;
2355         unsigned long addr;
2356         pte_t *pte;
2357         int i;
2358
2359         /*
2360          * Leave pmd empty until pte is filled note that it is fine to delay
2361          * notification until mmu_notifier_invalidate_range_end() as we are
2362          * replacing a zero pmd write protected page with a zero pte write
2363          * protected page.
2364          *
2365          * See Documentation/mm/mmu_notifier.rst
2366          */
2367         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2368
2369         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2370         pmd_populate(mm, &_pmd, pgtable);
2371
2372         pte = pte_offset_map(&_pmd, haddr);
2373         VM_BUG_ON(!pte);
2374         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2375                 pte_t entry;
2376
2377                 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2378                 entry = pte_mkspecial(entry);
2379                 if (pmd_uffd_wp(old_pmd))
2380                         entry = pte_mkuffd_wp(entry);
2381                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2382                 set_pte_at(mm, addr, pte, entry);
2383                 pte++;
2384         }
2385         pte_unmap(pte - 1);
2386         smp_wmb(); /* make pte visible before pmd */
2387         pmd_populate(mm, pmd, pgtable);
2388 }
2389
2390 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2391                 unsigned long haddr, bool freeze)
2392 {
2393         struct mm_struct *mm = vma->vm_mm;
2394         struct folio *folio;
2395         struct page *page;
2396         pgtable_t pgtable;
2397         pmd_t old_pmd, _pmd;
2398         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2399         bool anon_exclusive = false, dirty = false;
2400         unsigned long addr;
2401         pte_t *pte;
2402         int i;
2403
2404         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2405         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2406         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2407         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2408                                 && !pmd_devmap(*pmd));
2409
2410         count_vm_event(THP_SPLIT_PMD);
2411
2412         if (!vma_is_anonymous(vma)) {
2413                 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2414                 /*
2415                  * We are going to unmap this huge page. So
2416                  * just go ahead and zap it
2417                  */
2418                 if (arch_needs_pgtable_deposit())
2419                         zap_deposited_table(mm, pmd);
2420                 if (vma_is_special_huge(vma))
2421                         return;
2422                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2423                         swp_entry_t entry;
2424
2425                         entry = pmd_to_swp_entry(old_pmd);
2426                         folio = pfn_swap_entry_folio(entry);
2427                 } else {
2428                         page = pmd_page(old_pmd);
2429                         folio = page_folio(page);
2430                         if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))
2431                                 folio_mark_dirty(folio);
2432                         if (!folio_test_referenced(folio) && pmd_young(old_pmd))
2433                                 folio_set_referenced(folio);
2434                         folio_remove_rmap_pmd(folio, page, vma);
2435                         folio_put(folio);
2436                 }
2437                 add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR);
2438                 return;
2439         }
2440
2441         if (is_huge_zero_pmd(*pmd)) {
2442                 /*
2443                  * FIXME: Do we want to invalidate secondary mmu by calling
2444                  * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2445                  * inside __split_huge_pmd() ?
2446                  *
2447                  * We are going from a zero huge page write protected to zero
2448                  * small page also write protected so it does not seems useful
2449                  * to invalidate secondary mmu at this time.
2450                  */
2451                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2452         }
2453
2454         pmd_migration = is_pmd_migration_entry(*pmd);
2455         if (unlikely(pmd_migration)) {
2456                 swp_entry_t entry;
2457
2458                 old_pmd = *pmd;
2459                 entry = pmd_to_swp_entry(old_pmd);
2460                 page = pfn_swap_entry_to_page(entry);
2461                 write = is_writable_migration_entry(entry);
2462                 if (PageAnon(page))
2463                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2464                 young = is_migration_entry_young(entry);
2465                 dirty = is_migration_entry_dirty(entry);
2466                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2467                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2468         } else {
2469                 /*
2470                  * Up to this point the pmd is present and huge and userland has
2471                  * the whole access to the hugepage during the split (which
2472                  * happens in place). If we overwrite the pmd with the not-huge
2473                  * version pointing to the pte here (which of course we could if
2474                  * all CPUs were bug free), userland could trigger a small page
2475                  * size TLB miss on the small sized TLB while the hugepage TLB
2476                  * entry is still established in the huge TLB. Some CPU doesn't
2477                  * like that. See
2478                  * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2479                  * 383 on page 105. Intel should be safe but is also warns that
2480                  * it's only safe if the permission and cache attributes of the
2481                  * two entries loaded in the two TLB is identical (which should
2482                  * be the case here). But it is generally safer to never allow
2483                  * small and huge TLB entries for the same virtual address to be
2484                  * loaded simultaneously. So instead of doing "pmd_populate();
2485                  * flush_pmd_tlb_range();" we first mark the current pmd
2486                  * notpresent (atomically because here the pmd_trans_huge must
2487                  * remain set at all times on the pmd until the split is
2488                  * complete for this pmd), then we flush the SMP TLB and finally
2489                  * we write the non-huge version of the pmd entry with
2490                  * pmd_populate.
2491                  */
2492                 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2493                 page = pmd_page(old_pmd);
2494                 folio = page_folio(page);
2495                 if (pmd_dirty(old_pmd)) {
2496                         dirty = true;
2497                         folio_set_dirty(folio);
2498                 }
2499                 write = pmd_write(old_pmd);
2500                 young = pmd_young(old_pmd);
2501                 soft_dirty = pmd_soft_dirty(old_pmd);
2502                 uffd_wp = pmd_uffd_wp(old_pmd);
2503
2504                 VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);
2505                 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2506
2507                 /*
2508                  * Without "freeze", we'll simply split the PMD, propagating the
2509                  * PageAnonExclusive() flag for each PTE by setting it for
2510                  * each subpage -- no need to (temporarily) clear.
2511                  *
2512                  * With "freeze" we want to replace mapped pages by
2513                  * migration entries right away. This is only possible if we
2514                  * managed to clear PageAnonExclusive() -- see
2515                  * set_pmd_migration_entry().
2516                  *
2517                  * In case we cannot clear PageAnonExclusive(), split the PMD
2518                  * only and let try_to_migrate_one() fail later.
2519                  *
2520                  * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.
2521                  */
2522                 anon_exclusive = PageAnonExclusive(page);
2523                 if (freeze && anon_exclusive &&
2524                     folio_try_share_anon_rmap_pmd(folio, page))
2525                         freeze = false;
2526                 if (!freeze) {
2527                         rmap_t rmap_flags = RMAP_NONE;
2528
2529                         folio_ref_add(folio, HPAGE_PMD_NR - 1);
2530                         if (anon_exclusive)
2531                                 rmap_flags |= RMAP_EXCLUSIVE;
2532                         folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
2533                                                  vma, haddr, rmap_flags);
2534                 }
2535         }
2536
2537         /*
2538          * Withdraw the table only after we mark the pmd entry invalid.
2539          * This's critical for some architectures (Power).
2540          */
2541         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2542         pmd_populate(mm, &_pmd, pgtable);
2543
2544         pte = pte_offset_map(&_pmd, haddr);
2545         VM_BUG_ON(!pte);
2546
2547         /*
2548          * Note that NUMA hinting access restrictions are not transferred to
2549          * avoid any possibility of altering permissions across VMAs.
2550          */
2551         if (freeze || pmd_migration) {
2552                 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2553                         pte_t entry;
2554                         swp_entry_t swp_entry;
2555
2556                         if (write)
2557                                 swp_entry = make_writable_migration_entry(
2558                                                         page_to_pfn(page + i));
2559                         else if (anon_exclusive)
2560                                 swp_entry = make_readable_exclusive_migration_entry(
2561                                                         page_to_pfn(page + i));
2562                         else
2563                                 swp_entry = make_readable_migration_entry(
2564                                                         page_to_pfn(page + i));
2565                         if (young)
2566                                 swp_entry = make_migration_entry_young(swp_entry);
2567                         if (dirty)
2568                                 swp_entry = make_migration_entry_dirty(swp_entry);
2569                         entry = swp_entry_to_pte(swp_entry);
2570                         if (soft_dirty)
2571                                 entry = pte_swp_mksoft_dirty(entry);
2572                         if (uffd_wp)
2573                                 entry = pte_swp_mkuffd_wp(entry);
2574
2575                         VM_WARN_ON(!pte_none(ptep_get(pte + i)));
2576                         set_pte_at(mm, addr, pte + i, entry);
2577                 }
2578         } else {
2579                 pte_t entry;
2580
2581                 entry = mk_pte(page, READ_ONCE(vma->vm_page_prot));
2582                 if (write)
2583                         entry = pte_mkwrite(entry, vma);
2584                 if (!young)
2585                         entry = pte_mkold(entry);
2586                 /* NOTE: this may set soft-dirty too on some archs */
2587                 if (dirty)
2588                         entry = pte_mkdirty(entry);
2589                 if (soft_dirty)
2590                         entry = pte_mksoft_dirty(entry);
2591                 if (uffd_wp)
2592                         entry = pte_mkuffd_wp(entry);
2593
2594                 for (i = 0; i < HPAGE_PMD_NR; i++)
2595                         VM_WARN_ON(!pte_none(ptep_get(pte + i)));
2596
2597                 set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR);
2598         }
2599         pte_unmap(pte);
2600
2601         if (!pmd_migration)
2602                 folio_remove_rmap_pmd(folio, page, vma);
2603         if (freeze)
2604                 put_page(page);
2605
2606         smp_wmb(); /* make pte visible before pmd */
2607         pmd_populate(mm, pmd, pgtable);
2608 }
2609
2610 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
2611                            pmd_t *pmd, bool freeze, struct folio *folio)
2612 {
2613         VM_WARN_ON_ONCE(folio && !folio_test_pmd_mappable(folio));
2614         VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));
2615         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2616         VM_BUG_ON(freeze && !folio);
2617
2618         /*
2619          * When the caller requests to set up a migration entry, we
2620          * require a folio to check the PMD against. Otherwise, there
2621          * is a risk of replacing the wrong folio.
2622          */
2623         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2624             is_pmd_migration_entry(*pmd)) {
2625                 if (folio && folio != pmd_folio(*pmd))
2626                         return;
2627                 __split_huge_pmd_locked(vma, pmd, address, freeze);
2628         }
2629 }
2630
2631 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2632                 unsigned long address, bool freeze, struct folio *folio)
2633 {
2634         spinlock_t *ptl;
2635         struct mmu_notifier_range range;
2636
2637         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2638                                 address & HPAGE_PMD_MASK,
2639                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2640         mmu_notifier_invalidate_range_start(&range);
2641         ptl = pmd_lock(vma->vm_mm, pmd);
2642         split_huge_pmd_locked(vma, range.start, pmd, freeze, folio);
2643         spin_unlock(ptl);
2644         mmu_notifier_invalidate_range_end(&range);
2645 }
2646
2647 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2648                 bool freeze, struct folio *folio)
2649 {
2650         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2651
2652         if (!pmd)
2653                 return;
2654
2655         __split_huge_pmd(vma, pmd, address, freeze, folio);
2656 }
2657
2658 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2659 {
2660         /*
2661          * If the new address isn't hpage aligned and it could previously
2662          * contain an hugepage: check if we need to split an huge pmd.
2663          */
2664         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2665             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2666                          ALIGN(address, HPAGE_PMD_SIZE)))
2667                 split_huge_pmd_address(vma, address, false, NULL);
2668 }
2669
2670 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2671                              unsigned long start,
2672                              unsigned long end,
2673                              long adjust_next)
2674 {
2675         /* Check if we need to split start first. */
2676         split_huge_pmd_if_needed(vma, start);
2677
2678         /* Check if we need to split end next. */
2679         split_huge_pmd_if_needed(vma, end);
2680
2681         /*
2682          * If we're also updating the next vma vm_start,
2683          * check if we need to split it.
2684          */
2685         if (adjust_next > 0) {
2686                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2687                 unsigned long nstart = next->vm_start;
2688                 nstart += adjust_next;
2689                 split_huge_pmd_if_needed(next, nstart);
2690         }
2691 }
2692
2693 static void unmap_folio(struct folio *folio)
2694 {
2695         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC |
2696                 TTU_BATCH_FLUSH;
2697
2698         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2699
2700         if (folio_test_pmd_mappable(folio))
2701                 ttu_flags |= TTU_SPLIT_HUGE_PMD;
2702
2703         /*
2704          * Anon pages need migration entries to preserve them, but file
2705          * pages can simply be left unmapped, then faulted back on demand.
2706          * If that is ever changed (perhaps for mlock), update remap_page().
2707          */
2708         if (folio_test_anon(folio))
2709                 try_to_migrate(folio, ttu_flags);
2710         else
2711                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2712
2713         try_to_unmap_flush();
2714 }
2715
2716 static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
2717                                             unsigned long addr, pmd_t *pmdp,
2718                                             struct folio *folio)
2719 {
2720         struct mm_struct *mm = vma->vm_mm;
2721         int ref_count, map_count;
2722         pmd_t orig_pmd = *pmdp;
2723
2724         if (folio_test_dirty(folio) || pmd_dirty(orig_pmd))
2725                 return false;
2726
2727         orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);
2728
2729         /*
2730          * Syncing against concurrent GUP-fast:
2731          * - clear PMD; barrier; read refcount
2732          * - inc refcount; barrier; read PMD
2733          */
2734         smp_mb();
2735
2736         ref_count = folio_ref_count(folio);
2737         map_count = folio_mapcount(folio);
2738
2739         /*
2740          * Order reads for folio refcount and dirty flag
2741          * (see comments in __remove_mapping()).
2742          */
2743         smp_rmb();
2744
2745         /*
2746          * If the folio or its PMD is redirtied at this point, or if there
2747          * are unexpected references, we will give up to discard this folio
2748          * and remap it.
2749          *
2750          * The only folio refs must be one from isolation plus the rmap(s).
2751          */
2752         if (folio_test_dirty(folio) || pmd_dirty(orig_pmd) ||
2753             ref_count != map_count + 1) {
2754                 set_pmd_at(mm, addr, pmdp, orig_pmd);
2755                 return false;
2756         }
2757
2758         folio_remove_rmap_pmd(folio, pmd_page(orig_pmd), vma);
2759         zap_deposited_table(mm, pmdp);
2760         add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);
2761         if (vma->vm_flags & VM_LOCKED)
2762                 mlock_drain_local();
2763         folio_put(folio);
2764
2765         return true;
2766 }
2767
2768 bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
2769                            pmd_t *pmdp, struct folio *folio)
2770 {
2771         VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);
2772         VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
2773         VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));
2774
2775         if (folio_test_anon(folio) && !folio_test_swapbacked(folio))
2776                 return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);
2777
2778         return false;
2779 }
2780
2781 static void remap_page(struct folio *folio, unsigned long nr)
2782 {
2783         int i = 0;
2784
2785         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2786         if (!folio_test_anon(folio))
2787                 return;
2788         for (;;) {
2789                 remove_migration_ptes(folio, folio, true);
2790                 i += folio_nr_pages(folio);
2791                 if (i >= nr)
2792                         break;
2793                 folio = folio_next(folio);
2794         }
2795 }
2796
2797 static void lru_add_page_tail(struct page *head, struct page *tail,
2798                 struct lruvec *lruvec, struct list_head *list)
2799 {
2800         VM_BUG_ON_PAGE(!PageHead(head), head);
2801         VM_BUG_ON_PAGE(PageLRU(tail), head);
2802         lockdep_assert_held(&lruvec->lru_lock);
2803
2804         if (list) {
2805                 /* page reclaim is reclaiming a huge page */
2806                 VM_WARN_ON(PageLRU(head));
2807                 get_page(tail);
2808                 list_add_tail(&tail->lru, list);
2809         } else {
2810                 /* head is still on lru (and we have it frozen) */
2811                 VM_WARN_ON(!PageLRU(head));
2812                 if (PageUnevictable(tail))
2813                         tail->mlock_count = 0;
2814                 else
2815                         list_add_tail(&tail->lru, &head->lru);
2816                 SetPageLRU(tail);
2817         }
2818 }
2819
2820 static void __split_huge_page_tail(struct folio *folio, int tail,
2821                 struct lruvec *lruvec, struct list_head *list,
2822                 unsigned int new_order)
2823 {
2824         struct page *head = &folio->page;
2825         struct page *page_tail = head + tail;
2826         /*
2827          * Careful: new_folio is not a "real" folio before we cleared PageTail.
2828          * Don't pass it around before clear_compound_head().
2829          */
2830         struct folio *new_folio = (struct folio *)page_tail;
2831
2832         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2833
2834         /*
2835          * Clone page flags before unfreezing refcount.
2836          *
2837          * After successful get_page_unless_zero() might follow flags change,
2838          * for example lock_page() which set PG_waiters.
2839          *
2840          * Note that for mapped sub-pages of an anonymous THP,
2841          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2842          * the migration entry instead from where remap_page() will restore it.
2843          * We can still have PG_anon_exclusive set on effectively unmapped and
2844          * unreferenced sub-pages of an anonymous THP: we can simply drop
2845          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2846          */
2847         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2848         page_tail->flags |= (head->flags &
2849                         ((1L << PG_referenced) |
2850                          (1L << PG_swapbacked) |
2851                          (1L << PG_swapcache) |
2852                          (1L << PG_mlocked) |
2853                          (1L << PG_uptodate) |
2854                          (1L << PG_active) |
2855                          (1L << PG_workingset) |
2856                          (1L << PG_locked) |
2857                          (1L << PG_unevictable) |
2858 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2859                          (1L << PG_arch_2) |
2860                          (1L << PG_arch_3) |
2861 #endif
2862                          (1L << PG_dirty) |
2863                          LRU_GEN_MASK | LRU_REFS_MASK));
2864
2865         /* ->mapping in first and second tail page is replaced by other uses */
2866         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2867                         page_tail);
2868         page_tail->mapping = head->mapping;
2869         page_tail->index = head->index + tail;
2870
2871         /*
2872          * page->private should not be set in tail pages. Fix up and warn once
2873          * if private is unexpectedly set.
2874          */
2875         if (unlikely(page_tail->private)) {
2876                 VM_WARN_ON_ONCE_PAGE(true, page_tail);
2877                 page_tail->private = 0;
2878         }
2879         if (folio_test_swapcache(folio))
2880                 new_folio->swap.val = folio->swap.val + tail;
2881
2882         /* Page flags must be visible before we make the page non-compound. */
2883         smp_wmb();
2884
2885         /*
2886          * Clear PageTail before unfreezing page refcount.
2887          *
2888          * After successful get_page_unless_zero() might follow put_page()
2889          * which needs correct compound_head().
2890          */
2891         clear_compound_head(page_tail);
2892         if (new_order) {
2893                 prep_compound_page(page_tail, new_order);
2894                 folio_set_large_rmappable(new_folio);
2895         }
2896
2897         /* Finally unfreeze refcount. Additional reference from page cache. */
2898         page_ref_unfreeze(page_tail,
2899                 1 + ((!folio_test_anon(folio) || folio_test_swapcache(folio)) ?
2900                              folio_nr_pages(new_folio) : 0));
2901
2902         if (folio_test_young(folio))
2903                 folio_set_young(new_folio);
2904         if (folio_test_idle(folio))
2905                 folio_set_idle(new_folio);
2906
2907         folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
2908
2909         /*
2910          * always add to the tail because some iterators expect new
2911          * pages to show after the currently processed elements - e.g.
2912          * migrate_pages
2913          */
2914         lru_add_page_tail(head, page_tail, lruvec, list);
2915 }
2916
2917 static void __split_huge_page(struct page *page, struct list_head *list,
2918                 pgoff_t end, unsigned int new_order)
2919 {
2920         struct folio *folio = page_folio(page);
2921         struct page *head = &folio->page;
2922         struct lruvec *lruvec;
2923         struct address_space *swap_cache = NULL;
2924         unsigned long offset = 0;
2925         int i, nr_dropped = 0;
2926         unsigned int new_nr = 1 << new_order;
2927         int order = folio_order(folio);
2928         unsigned int nr = 1 << order;
2929
2930         /* complete memcg works before add pages to LRU */
2931         split_page_memcg(head, order, new_order);
2932
2933         if (folio_test_anon(folio) && folio_test_swapcache(folio)) {
2934                 offset = swap_cache_index(folio->swap);
2935                 swap_cache = swap_address_space(folio->swap);
2936                 xa_lock(&swap_cache->i_pages);
2937         }
2938
2939         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2940         lruvec = folio_lruvec_lock(folio);
2941
2942         ClearPageHasHWPoisoned(head);
2943
2944         for (i = nr - new_nr; i >= new_nr; i -= new_nr) {
2945                 __split_huge_page_tail(folio, i, lruvec, list, new_order);
2946                 /* Some pages can be beyond EOF: drop them from page cache */
2947                 if (head[i].index >= end) {
2948                         struct folio *tail = page_folio(head + i);
2949
2950                         if (shmem_mapping(folio->mapping))
2951                                 nr_dropped++;
2952                         else if (folio_test_clear_dirty(tail))
2953                                 folio_account_cleaned(tail,
2954                                         inode_to_wb(folio->mapping->host));
2955                         __filemap_remove_folio(tail, NULL);
2956                         folio_put(tail);
2957                 } else if (!PageAnon(page)) {
2958                         __xa_store(&folio->mapping->i_pages, head[i].index,
2959                                         head + i, 0);
2960                 } else if (swap_cache) {
2961                         __xa_store(&swap_cache->i_pages, offset + i,
2962                                         head + i, 0);
2963                 }
2964         }
2965
2966         if (!new_order)
2967                 ClearPageCompound(head);
2968         else {
2969                 struct folio *new_folio = (struct folio *)head;
2970
2971                 folio_set_order(new_folio, new_order);
2972         }
2973         unlock_page_lruvec(lruvec);
2974         /* Caller disabled irqs, so they are still disabled here */
2975
2976         split_page_owner(head, order, new_order);
2977         pgalloc_tag_split(head, 1 << order);
2978
2979         /* See comment in __split_huge_page_tail() */
2980         if (folio_test_anon(folio)) {
2981                 /* Additional pin to swap cache */
2982                 if (folio_test_swapcache(folio)) {
2983                         folio_ref_add(folio, 1 + new_nr);
2984                         xa_unlock(&swap_cache->i_pages);
2985                 } else {
2986                         folio_ref_inc(folio);
2987                 }
2988         } else {
2989                 /* Additional pin to page cache */
2990                 folio_ref_add(folio, 1 + new_nr);
2991                 xa_unlock(&folio->mapping->i_pages);
2992         }
2993         local_irq_enable();
2994
2995         if (nr_dropped)
2996                 shmem_uncharge(folio->mapping->host, nr_dropped);
2997         remap_page(folio, nr);
2998
2999         /*
3000          * set page to its compound_head when split to non order-0 pages, so
3001          * we can skip unlocking it below, since PG_locked is transferred to
3002          * the compound_head of the page and the caller will unlock it.
3003          */
3004         if (new_order)
3005                 page = compound_head(page);
3006
3007         for (i = 0; i < nr; i += new_nr) {
3008                 struct page *subpage = head + i;
3009                 struct folio *new_folio = page_folio(subpage);
3010                 if (subpage == page)
3011                         continue;
3012                 folio_unlock(new_folio);
3013
3014                 /*
3015                  * Subpages may be freed if there wasn't any mapping
3016                  * like if add_to_swap() is running on a lru page that
3017                  * had its mapping zapped. And freeing these pages
3018                  * requires taking the lru_lock so we do the put_page
3019                  * of the tail pages after the split is complete.
3020                  */
3021                 free_page_and_swap_cache(subpage);
3022         }
3023 }
3024
3025 /* Racy check whether the huge page can be split */
3026 bool can_split_folio(struct folio *folio, int *pextra_pins)
3027 {
3028         int extra_pins;
3029
3030         /* Additional pins from page cache */
3031         if (folio_test_anon(folio))
3032                 extra_pins = folio_test_swapcache(folio) ?
3033                                 folio_nr_pages(folio) : 0;
3034         else
3035                 extra_pins = folio_nr_pages(folio);
3036         if (pextra_pins)
3037                 *pextra_pins = extra_pins;
3038         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
3039 }
3040
3041 /*
3042  * This function splits a large folio into smaller folios of order @new_order.
3043  * @page can point to any page of the large folio to split. The split operation
3044  * does not change the position of @page.
3045  *
3046  * Prerequisites:
3047  *
3048  * 1) The caller must hold a reference on the @page's owning folio, also known
3049  *    as the large folio.
3050  *
3051  * 2) The large folio must be locked.
3052  *
3053  * 3) The folio must not be pinned. Any unexpected folio references, including
3054  *    GUP pins, will result in the folio not getting split; instead, the caller
3055  *    will receive an -EAGAIN.
3056  *
3057  * 4) @new_order > 1, usually. Splitting to order-1 anonymous folios is not
3058  *    supported for non-file-backed folios, because folio->_deferred_list, which
3059  *    is used by partially mapped folios, is stored in subpage 2, but an order-1
3060  *    folio only has subpages 0 and 1. File-backed order-1 folios are supported,
3061  *    since they do not use _deferred_list.
3062  *
3063  * After splitting, the caller's folio reference will be transferred to @page,
3064  * resulting in a raised refcount of @page after this call. The other pages may
3065  * be freed if they are not mapped.
3066  *
3067  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3068  *
3069  * Pages in @new_order will inherit the mapping, flags, and so on from the
3070  * huge page.
3071  *
3072  * Returns 0 if the huge page was split successfully.
3073  *
3074  * Returns -EAGAIN if the folio has unexpected reference (e.g., GUP) or if
3075  * the folio was concurrently removed from the page cache.
3076  *
3077  * Returns -EBUSY when trying to split the huge zeropage, if the folio is
3078  * under writeback, if fs-specific folio metadata cannot currently be
3079  * released, or if some unexpected race happened (e.g., anon VMA disappeared,
3080  * truncation).
3081  *
3082  * Returns -EINVAL when trying to split to an order that is incompatible
3083  * with the folio. Splitting to order 0 is compatible with all folios.
3084  */
3085 int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
3086                                      unsigned int new_order)
3087 {
3088         struct folio *folio = page_folio(page);
3089         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3090         /* reset xarray order to new order after split */
3091         XA_STATE_ORDER(xas, &folio->mapping->i_pages, folio->index, new_order);
3092         struct anon_vma *anon_vma = NULL;
3093         struct address_space *mapping = NULL;
3094         int order = folio_order(folio);
3095         int extra_pins, ret;
3096         pgoff_t end;
3097         bool is_hzp;
3098
3099         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3100         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
3101
3102         if (new_order >= folio_order(folio))
3103                 return -EINVAL;
3104
3105         if (folio_test_anon(folio)) {
3106                 /* order-1 is not supported for anonymous THP. */
3107                 if (new_order == 1) {
3108                         VM_WARN_ONCE(1, "Cannot split to order-1 folio");
3109                         return -EINVAL;
3110                 }
3111         } else if (new_order) {
3112                 /* Split shmem folio to non-zero order not supported */
3113                 if (shmem_mapping(folio->mapping)) {
3114                         VM_WARN_ONCE(1,
3115                                 "Cannot split shmem folio to non-0 order");
3116                         return -EINVAL;
3117                 }
3118                 /*
3119                  * No split if the file system does not support large folio.
3120                  * Note that we might still have THPs in such mappings due to
3121                  * CONFIG_READ_ONLY_THP_FOR_FS. But in that case, the mapping
3122                  * does not actually support large folios properly.
3123                  */
3124                 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
3125                     !mapping_large_folio_support(folio->mapping)) {
3126                         VM_WARN_ONCE(1,
3127                                 "Cannot split file folio to non-0 order");
3128                         return -EINVAL;
3129                 }
3130         }
3131
3132         /* Only swapping a whole PMD-mapped folio is supported */
3133         if (folio_test_swapcache(folio) && new_order)
3134                 return -EINVAL;
3135
3136         is_hzp = is_huge_zero_folio(folio);
3137         if (is_hzp) {
3138                 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
3139                 return -EBUSY;
3140         }
3141
3142         if (folio_test_writeback(folio))
3143                 return -EBUSY;
3144
3145         if (folio_test_anon(folio)) {
3146                 /*
3147                  * The caller does not necessarily hold an mmap_lock that would
3148                  * prevent the anon_vma disappearing so we first we take a
3149                  * reference to it and then lock the anon_vma for write. This
3150                  * is similar to folio_lock_anon_vma_read except the write lock
3151                  * is taken to serialise against parallel split or collapse
3152                  * operations.
3153                  */
3154                 anon_vma = folio_get_anon_vma(folio);
3155                 if (!anon_vma) {
3156                         ret = -EBUSY;
3157                         goto out;
3158                 }
3159                 end = -1;
3160                 mapping = NULL;
3161                 anon_vma_lock_write(anon_vma);
3162         } else {
3163                 gfp_t gfp;
3164
3165                 mapping = folio->mapping;
3166
3167                 /* Truncated ? */
3168                 if (!mapping) {
3169                         ret = -EBUSY;
3170                         goto out;
3171                 }
3172
3173                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
3174                                                         GFP_RECLAIM_MASK);
3175
3176                 if (!filemap_release_folio(folio, gfp)) {
3177                         ret = -EBUSY;
3178                         goto out;
3179                 }
3180
3181                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
3182                 if (xas_error(&xas)) {
3183                         ret = xas_error(&xas);
3184                         goto out;
3185                 }
3186
3187                 anon_vma = NULL;
3188                 i_mmap_lock_read(mapping);
3189
3190                 /*
3191                  *__split_huge_page() may need to trim off pages beyond EOF:
3192                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
3193                  * which cannot be nested inside the page tree lock. So note
3194                  * end now: i_size itself may be changed at any moment, but
3195                  * folio lock is good enough to serialize the trimming.
3196                  */
3197                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3198                 if (shmem_mapping(mapping))
3199                         end = shmem_fallocend(mapping->host, end);
3200         }
3201
3202         /*
3203          * Racy check if we can split the page, before unmap_folio() will
3204          * split PMDs
3205          */
3206         if (!can_split_folio(folio, &extra_pins)) {
3207                 ret = -EAGAIN;
3208                 goto out_unlock;
3209         }
3210
3211         unmap_folio(folio);
3212
3213         /* block interrupt reentry in xa_lock and spinlock */
3214         local_irq_disable();
3215         if (mapping) {
3216                 /*
3217                  * Check if the folio is present in page cache.
3218                  * We assume all tail are present too, if folio is there.
3219                  */
3220                 xas_lock(&xas);
3221                 xas_reset(&xas);
3222                 if (xas_load(&xas) != folio)
3223                         goto fail;
3224         }
3225
3226         /* Prevent deferred_split_scan() touching ->_refcount */
3227         spin_lock(&ds_queue->split_queue_lock);
3228         if (folio_ref_freeze(folio, 1 + extra_pins)) {
3229                 if (folio_order(folio) > 1 &&
3230                     !list_empty(&folio->_deferred_list)) {
3231                         ds_queue->split_queue_len--;
3232                         /*
3233                          * Reinitialize page_deferred_list after removing the
3234                          * page from the split_queue, otherwise a subsequent
3235                          * split will see list corruption when checking the
3236                          * page_deferred_list.
3237                          */
3238                         list_del_init(&folio->_deferred_list);
3239                 }
3240                 spin_unlock(&ds_queue->split_queue_lock);
3241                 if (mapping) {
3242                         int nr = folio_nr_pages(folio);
3243
3244                         xas_split(&xas, folio, folio_order(folio));
3245                         if (folio_test_pmd_mappable(folio) &&
3246                             new_order < HPAGE_PMD_ORDER) {
3247                                 if (folio_test_swapbacked(folio)) {
3248                                         __lruvec_stat_mod_folio(folio,
3249                                                         NR_SHMEM_THPS, -nr);
3250                                 } else {
3251                                         __lruvec_stat_mod_folio(folio,
3252                                                         NR_FILE_THPS, -nr);
3253                                         filemap_nr_thps_dec(mapping);
3254                                 }
3255                         }
3256                 }
3257
3258                 __split_huge_page(page, list, end, new_order);
3259                 ret = 0;
3260         } else {
3261                 spin_unlock(&ds_queue->split_queue_lock);
3262 fail:
3263                 if (mapping)
3264                         xas_unlock(&xas);
3265                 local_irq_enable();
3266                 remap_page(folio, folio_nr_pages(folio));
3267                 ret = -EAGAIN;
3268         }
3269
3270 out_unlock:
3271         if (anon_vma) {
3272                 anon_vma_unlock_write(anon_vma);
3273                 put_anon_vma(anon_vma);
3274         }
3275         if (mapping)
3276                 i_mmap_unlock_read(mapping);
3277 out:
3278         xas_destroy(&xas);
3279         if (order == HPAGE_PMD_ORDER)
3280                 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3281         count_mthp_stat(order, !ret ? MTHP_STAT_SPLIT : MTHP_STAT_SPLIT_FAILED);
3282         return ret;
3283 }
3284
3285 void __folio_undo_large_rmappable(struct folio *folio)
3286 {
3287         struct deferred_split *ds_queue;
3288         unsigned long flags;
3289
3290         ds_queue = get_deferred_split_queue(folio);
3291         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3292         if (!list_empty(&folio->_deferred_list)) {
3293                 ds_queue->split_queue_len--;
3294                 list_del_init(&folio->_deferred_list);
3295         }
3296         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3297 }
3298
3299 void deferred_split_folio(struct folio *folio)
3300 {
3301         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3302 #ifdef CONFIG_MEMCG
3303         struct mem_cgroup *memcg = folio_memcg(folio);
3304 #endif
3305         unsigned long flags;
3306
3307         /*
3308          * Order 1 folios have no space for a deferred list, but we also
3309          * won't waste much memory by not adding them to the deferred list.
3310          */
3311         if (folio_order(folio) <= 1)
3312                 return;
3313
3314         /*
3315          * The try_to_unmap() in page reclaim path might reach here too,
3316          * this may cause a race condition to corrupt deferred split queue.
3317          * And, if page reclaim is already handling the same folio, it is
3318          * unnecessary to handle it again in shrinker.
3319          *
3320          * Check the swapcache flag to determine if the folio is being
3321          * handled by page reclaim since THP swap would add the folio into
3322          * swap cache before calling try_to_unmap().
3323          */
3324         if (folio_test_swapcache(folio))
3325                 return;
3326
3327         if (!list_empty(&folio->_deferred_list))
3328                 return;
3329
3330         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3331         if (list_empty(&folio->_deferred_list)) {
3332                 if (folio_test_pmd_mappable(folio))
3333                         count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3334                 count_mthp_stat(folio_order(folio), MTHP_STAT_SPLIT_DEFERRED);
3335                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
3336                 ds_queue->split_queue_len++;
3337 #ifdef CONFIG_MEMCG
3338                 if (memcg)
3339                         set_shrinker_bit(memcg, folio_nid(folio),
3340                                          deferred_split_shrinker->id);
3341 #endif
3342         }
3343         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3344 }
3345
3346 static unsigned long deferred_split_count(struct shrinker *shrink,
3347                 struct shrink_control *sc)
3348 {
3349         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3350         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3351
3352 #ifdef CONFIG_MEMCG
3353         if (sc->memcg)
3354                 ds_queue = &sc->memcg->deferred_split_queue;
3355 #endif
3356         return READ_ONCE(ds_queue->split_queue_len);
3357 }
3358
3359 static unsigned long deferred_split_scan(struct shrinker *shrink,
3360                 struct shrink_control *sc)
3361 {
3362         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3363         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3364         unsigned long flags;
3365         LIST_HEAD(list);
3366         struct folio *folio, *next;
3367         int split = 0;
3368
3369 #ifdef CONFIG_MEMCG
3370         if (sc->memcg)
3371                 ds_queue = &sc->memcg->deferred_split_queue;
3372 #endif
3373
3374         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3375         /* Take pin on all head pages to avoid freeing them under us */
3376         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
3377                                                         _deferred_list) {
3378                 if (folio_try_get(folio)) {
3379                         list_move(&folio->_deferred_list, &list);
3380                 } else {
3381                         /* We lost race with folio_put() */
3382                         list_del_init(&folio->_deferred_list);
3383                         ds_queue->split_queue_len--;
3384                 }
3385                 if (!--sc->nr_to_scan)
3386                         break;
3387         }
3388         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3389
3390         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
3391                 if (!folio_trylock(folio))
3392                         goto next;
3393                 /* split_huge_page() removes page from list on success */
3394                 if (!split_folio(folio))
3395                         split++;
3396                 folio_unlock(folio);
3397 next:
3398                 folio_put(folio);
3399         }
3400
3401         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3402         list_splice_tail(&list, &ds_queue->split_queue);
3403         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3404
3405         /*
3406          * Stop shrinker if we didn't split any page, but the queue is empty.
3407          * This can happen if pages were freed under us.
3408          */
3409         if (!split && list_empty(&ds_queue->split_queue))
3410                 return SHRINK_STOP;
3411         return split;
3412 }
3413
3414 #ifdef CONFIG_DEBUG_FS
3415 static void split_huge_pages_all(void)
3416 {
3417         struct zone *zone;
3418         struct page *page;
3419         struct folio *folio;
3420         unsigned long pfn, max_zone_pfn;
3421         unsigned long total = 0, split = 0;
3422
3423         pr_debug("Split all THPs\n");
3424         for_each_zone(zone) {
3425                 if (!managed_zone(zone))
3426                         continue;
3427                 max_zone_pfn = zone_end_pfn(zone);
3428                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3429                         int nr_pages;
3430
3431                         page = pfn_to_online_page(pfn);
3432                         if (!page || PageTail(page))
3433                                 continue;
3434                         folio = page_folio(page);
3435                         if (!folio_try_get(folio))
3436                                 continue;
3437
3438                         if (unlikely(page_folio(page) != folio))
3439                                 goto next;
3440
3441                         if (zone != folio_zone(folio))
3442                                 goto next;
3443
3444                         if (!folio_test_large(folio)
3445                                 || folio_test_hugetlb(folio)
3446                                 || !folio_test_lru(folio))
3447                                 goto next;
3448
3449                         total++;
3450                         folio_lock(folio);
3451                         nr_pages = folio_nr_pages(folio);
3452                         if (!split_folio(folio))
3453                                 split++;
3454                         pfn += nr_pages - 1;
3455                         folio_unlock(folio);
3456 next:
3457                         folio_put(folio);
3458                         cond_resched();
3459                 }
3460         }
3461
3462         pr_debug("%lu of %lu THP split\n", split, total);
3463 }
3464
3465 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
3466 {
3467         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
3468                     is_vm_hugetlb_page(vma);
3469 }
3470
3471 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
3472                                 unsigned long vaddr_end, unsigned int new_order)
3473 {
3474         int ret = 0;
3475         struct task_struct *task;
3476         struct mm_struct *mm;
3477         unsigned long total = 0, split = 0;
3478         unsigned long addr;
3479
3480         vaddr_start &= PAGE_MASK;
3481         vaddr_end &= PAGE_MASK;
3482
3483         /* Find the task_struct from pid */
3484         rcu_read_lock();
3485         task = find_task_by_vpid(pid);
3486         if (!task) {
3487                 rcu_read_unlock();
3488                 ret = -ESRCH;
3489                 goto out;
3490         }
3491         get_task_struct(task);
3492         rcu_read_unlock();
3493
3494         /* Find the mm_struct */
3495         mm = get_task_mm(task);
3496         put_task_struct(task);
3497
3498         if (!mm) {
3499                 ret = -EINVAL;
3500                 goto out;
3501         }
3502
3503         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3504                  pid, vaddr_start, vaddr_end);
3505
3506         mmap_read_lock(mm);
3507         /*
3508          * always increase addr by PAGE_SIZE, since we could have a PTE page
3509          * table filled with PTE-mapped THPs, each of which is distinct.
3510          */
3511         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3512                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3513                 struct page *page;
3514                 struct folio *folio;
3515
3516                 if (!vma)
3517                         break;
3518
3519                 /* skip special VMA and hugetlb VMA */
3520                 if (vma_not_suitable_for_thp_split(vma)) {
3521                         addr = vma->vm_end;
3522                         continue;
3523                 }
3524
3525                 /* FOLL_DUMP to ignore special (like zero) pages */
3526                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3527
3528                 if (IS_ERR_OR_NULL(page))
3529                         continue;
3530
3531                 folio = page_folio(page);
3532                 if (!is_transparent_hugepage(folio))
3533                         goto next;
3534
3535                 if (new_order >= folio_order(folio))
3536                         goto next;
3537
3538                 total++;
3539                 /*
3540                  * For folios with private, split_huge_page_to_list_to_order()
3541                  * will try to drop it before split and then check if the folio
3542                  * can be split or not. So skip the check here.
3543                  */
3544                 if (!folio_test_private(folio) &&
3545                     !can_split_folio(folio, NULL))
3546                         goto next;
3547
3548                 if (!folio_trylock(folio))
3549                         goto next;
3550
3551                 if (!split_folio_to_order(folio, new_order))
3552                         split++;
3553
3554                 folio_unlock(folio);
3555 next:
3556                 folio_put(folio);
3557                 cond_resched();
3558         }
3559         mmap_read_unlock(mm);
3560         mmput(mm);
3561
3562         pr_debug("%lu of %lu THP split\n", split, total);
3563
3564 out:
3565         return ret;
3566 }
3567
3568 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3569                                 pgoff_t off_end, unsigned int new_order)
3570 {
3571         struct filename *file;
3572         struct file *candidate;
3573         struct address_space *mapping;
3574         int ret = -EINVAL;
3575         pgoff_t index;
3576         int nr_pages = 1;
3577         unsigned long total = 0, split = 0;
3578
3579         file = getname_kernel(file_path);
3580         if (IS_ERR(file))
3581                 return ret;
3582
3583         candidate = file_open_name(file, O_RDONLY, 0);
3584         if (IS_ERR(candidate))
3585                 goto out;
3586
3587         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3588                  file_path, off_start, off_end);
3589
3590         mapping = candidate->f_mapping;
3591
3592         for (index = off_start; index < off_end; index += nr_pages) {
3593                 struct folio *folio = filemap_get_folio(mapping, index);
3594
3595                 nr_pages = 1;
3596                 if (IS_ERR(folio))
3597                         continue;
3598
3599                 if (!folio_test_large(folio))
3600                         goto next;
3601
3602                 total++;
3603                 nr_pages = folio_nr_pages(folio);
3604
3605                 if (new_order >= folio_order(folio))
3606                         goto next;
3607
3608                 if (!folio_trylock(folio))
3609                         goto next;
3610
3611                 if (!split_folio_to_order(folio, new_order))
3612                         split++;
3613
3614                 folio_unlock(folio);
3615 next:
3616                 folio_put(folio);
3617                 cond_resched();
3618         }
3619
3620         filp_close(candidate, NULL);
3621         ret = 0;
3622
3623         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3624 out:
3625         putname(file);
3626         return ret;
3627 }
3628
3629 #define MAX_INPUT_BUF_SZ 255
3630
3631 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3632                                 size_t count, loff_t *ppops)
3633 {
3634         static DEFINE_MUTEX(split_debug_mutex);
3635         ssize_t ret;
3636         /*
3637          * hold pid, start_vaddr, end_vaddr, new_order or
3638          * file_path, off_start, off_end, new_order
3639          */
3640         char input_buf[MAX_INPUT_BUF_SZ];
3641         int pid;
3642         unsigned long vaddr_start, vaddr_end;
3643         unsigned int new_order = 0;
3644
3645         ret = mutex_lock_interruptible(&split_debug_mutex);
3646         if (ret)
3647                 return ret;
3648
3649         ret = -EFAULT;
3650
3651         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3652         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3653                 goto out;
3654
3655         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3656
3657         if (input_buf[0] == '/') {
3658                 char *tok;
3659                 char *buf = input_buf;
3660                 char file_path[MAX_INPUT_BUF_SZ];
3661                 pgoff_t off_start = 0, off_end = 0;
3662                 size_t input_len = strlen(input_buf);
3663
3664                 tok = strsep(&buf, ",");
3665                 if (tok) {
3666                         strcpy(file_path, tok);
3667                 } else {
3668                         ret = -EINVAL;
3669                         goto out;
3670                 }
3671
3672                 ret = sscanf(buf, "0x%lx,0x%lx,%d", &off_start, &off_end, &new_order);
3673                 if (ret != 2 && ret != 3) {
3674                         ret = -EINVAL;
3675                         goto out;
3676                 }
3677                 ret = split_huge_pages_in_file(file_path, off_start, off_end, new_order);
3678                 if (!ret)
3679                         ret = input_len;
3680
3681                 goto out;
3682         }
3683
3684         ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d", &pid, &vaddr_start, &vaddr_end, &new_order);
3685         if (ret == 1 && pid == 1) {
3686                 split_huge_pages_all();
3687                 ret = strlen(input_buf);
3688                 goto out;
3689         } else if (ret != 3 && ret != 4) {
3690                 ret = -EINVAL;
3691                 goto out;
3692         }
3693
3694         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order);
3695         if (!ret)
3696                 ret = strlen(input_buf);
3697 out:
3698         mutex_unlock(&split_debug_mutex);
3699         return ret;
3700
3701 }
3702
3703 static const struct file_operations split_huge_pages_fops = {
3704         .owner   = THIS_MODULE,
3705         .write   = split_huge_pages_write,
3706         .llseek  = no_llseek,
3707 };
3708
3709 static int __init split_huge_pages_debugfs(void)
3710 {
3711         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3712                             &split_huge_pages_fops);
3713         return 0;
3714 }
3715 late_initcall(split_huge_pages_debugfs);
3716 #endif
3717
3718 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3719 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3720                 struct page *page)
3721 {
3722         struct folio *folio = page_folio(page);
3723         struct vm_area_struct *vma = pvmw->vma;
3724         struct mm_struct *mm = vma->vm_mm;
3725         unsigned long address = pvmw->address;
3726         bool anon_exclusive;
3727         pmd_t pmdval;
3728         swp_entry_t entry;
3729         pmd_t pmdswp;
3730
3731         if (!(pvmw->pmd && !pvmw->pte))
3732                 return 0;
3733
3734         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3735         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3736
3737         /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
3738         anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
3739         if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {
3740                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3741                 return -EBUSY;
3742         }
3743
3744         if (pmd_dirty(pmdval))
3745                 folio_mark_dirty(folio);
3746         if (pmd_write(pmdval))
3747                 entry = make_writable_migration_entry(page_to_pfn(page));
3748         else if (anon_exclusive)
3749                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3750         else
3751                 entry = make_readable_migration_entry(page_to_pfn(page));
3752         if (pmd_young(pmdval))
3753                 entry = make_migration_entry_young(entry);
3754         if (pmd_dirty(pmdval))
3755                 entry = make_migration_entry_dirty(entry);
3756         pmdswp = swp_entry_to_pmd(entry);
3757         if (pmd_soft_dirty(pmdval))
3758                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3759         if (pmd_uffd_wp(pmdval))
3760                 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3761         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3762         folio_remove_rmap_pmd(folio, page, vma);
3763         folio_put(folio);
3764         trace_set_migration_pmd(address, pmd_val(pmdswp));
3765
3766         return 0;
3767 }
3768
3769 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3770 {
3771         struct folio *folio = page_folio(new);
3772         struct vm_area_struct *vma = pvmw->vma;
3773         struct mm_struct *mm = vma->vm_mm;
3774         unsigned long address = pvmw->address;
3775         unsigned long haddr = address & HPAGE_PMD_MASK;
3776         pmd_t pmde;
3777         swp_entry_t entry;
3778
3779         if (!(pvmw->pmd && !pvmw->pte))
3780                 return;
3781
3782         entry = pmd_to_swp_entry(*pvmw->pmd);
3783         folio_get(folio);
3784         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3785         if (pmd_swp_soft_dirty(*pvmw->pmd))
3786                 pmde = pmd_mksoft_dirty(pmde);
3787         if (is_writable_migration_entry(entry))
3788                 pmde = pmd_mkwrite(pmde, vma);
3789         if (pmd_swp_uffd_wp(*pvmw->pmd))
3790                 pmde = pmd_mkuffd_wp(pmde);
3791         if (!is_migration_entry_young(entry))
3792                 pmde = pmd_mkold(pmde);
3793         /* NOTE: this may contain setting soft-dirty on some archs */
3794         if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
3795                 pmde = pmd_mkdirty(pmde);
3796
3797         if (folio_test_anon(folio)) {
3798                 rmap_t rmap_flags = RMAP_NONE;
3799
3800                 if (!is_readable_migration_entry(entry))
3801                         rmap_flags |= RMAP_EXCLUSIVE;
3802
3803                 folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);
3804         } else {
3805                 folio_add_file_rmap_pmd(folio, new, vma);
3806         }
3807         VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));
3808         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3809
3810         /* No need to invalidate - it was non-present before */
3811         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3812         trace_remove_migration_pmd(address, pmd_val(pmde));
3813 }
3814 #endif