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/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37 #include <linux/page_owner.h>
  38 #include <linux/sched/sysctl.h>
  39 #include <linux/memory-tiers.h>
  40 #include <linux/compat.h>
  41
  42 #include <asm/tlb.h>
  43 #include <asm/pgalloc.h>
  44 #include "internal.h"
  45 #include "swap.h"
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/thp.h>
  49
  50 /*
  51  * By default, transparent hugepage support is disabled in order to avoid
  52  * risking an increased memory footprint for applications that are not
  53  * guaranteed to benefit from it. When transparent hugepage support is
  54  * enabled, it is for all mappings, and khugepaged scans all mappings.
  55  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  56  * for all hugepage allocations.
  57  */
  58 unsigned long transparent_hugepage_flags __read_mostly =
  59 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  60         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  61 #endif
  62 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  63         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  64 #endif
  65         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  66         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  67         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  68
  69 static struct shrinker *deferred_split_shrinker;
  70 static unsigned long deferred_split_count(struct shrinker *shrink,
  71                                           struct shrink_control *sc);
  72 static unsigned long deferred_split_scan(struct shrinker *shrink,
  73                                          struct shrink_control *sc);
  74
  75 static atomic_t huge_zero_refcount;
  76 struct page *huge_zero_page __read_mostly;
  77 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  78 unsigned long huge_anon_orders_always __read_mostly;
  79 unsigned long huge_anon_orders_madvise __read_mostly;
  80 unsigned long huge_anon_orders_inherit __read_mostly;
  81
  82 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
  83                                          unsigned long vm_flags, bool smaps,
  84                                          bool in_pf, bool enforce_sysfs,
  85                                          unsigned long orders)
  86 {
  87         /* Check the intersection of requested and supported orders. */
  88         orders &= vma_is_anonymous(vma) ?
  89                         THP_ORDERS_ALL_ANON : THP_ORDERS_ALL_FILE;
  90         if (!orders)
  91                 return 0;
  92
  93         if (!vma->vm_mm)                /* vdso */
  94                 return 0;
  95
  96         /*
  97          * Explicitly disabled through madvise or prctl, or some
  98          * architectures may disable THP for some mappings, for
  99          * example, s390 kvm.
 100          * */
 101         if ((vm_flags & VM_NOHUGEPAGE) ||
 102             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 103                 return 0;
 104         /*
 105          * If the hardware/firmware marked hugepage support disabled.
 106          */
 107         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED))
 108                 return 0;
 109
 110         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
 111         if (vma_is_dax(vma))
 112                 return in_pf ? orders : 0;
 113
 114         /*
 115          * khugepaged special VMA and hugetlb VMA.
 116          * Must be checked after dax since some dax mappings may have
 117          * VM_MIXEDMAP set.
 118          */
 119         if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
 120                 return 0;
 121
 122         /*
 123          * Check alignment for file vma and size for both file and anon vma by
 124          * filtering out the unsuitable orders.
 125          *
 126          * Skip the check for page fault. Huge fault does the check in fault
 127          * handlers.
 128          */
 129         if (!in_pf) {
 130                 int order = highest_order(orders);
 131                 unsigned long addr;
 132
 133                 while (orders) {
 134                         addr = vma->vm_end - (PAGE_SIZE << order);
 135                         if (thp_vma_suitable_order(vma, addr, order))
 136                                 break;
 137                         order = next_order(&orders, order);
 138                 }
 139
 140                 if (!orders)
 141                         return 0;
 142         }
 143
 144         /*
 145          * Enabled via shmem mount options or sysfs settings.
 146          * Must be done before hugepage flags check since shmem has its
 147          * own flags.
 148          */
 149         if (!in_pf && shmem_file(vma->vm_file))
 150                 return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
 151                                      !enforce_sysfs, vma->vm_mm, vm_flags)
 152                         ? orders : 0;
 153
 154         if (!vma_is_anonymous(vma)) {
 155                 /*
 156                  * Enforce sysfs THP requirements as necessary. Anonymous vmas
 157                  * were already handled in thp_vma_allowable_orders().
 158                  */
 159                 if (enforce_sysfs &&
 160                     (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 161                                                     !hugepage_global_always())))
 162                         return 0;
 163
 164                 /*
 165                  * Trust that ->huge_fault() handlers know what they are doing
 166                  * in fault path.
 167                  */
 168                 if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
 169                         return orders;
 170                 /* Only regular file is valid in collapse path */
 171                 if (((!in_pf || smaps)) && file_thp_enabled(vma))
 172                         return orders;
 173                 return 0;
 174         }
 175
 176         if (vma_is_temporary_stack(vma))
 177                 return 0;
 178
 179         /*
 180          * THPeligible bit of smaps should show 1 for proper VMAs even
 181          * though anon_vma is not initialized yet.
 182          *
 183          * Allow page fault since anon_vma may be not initialized until
 184          * the first page fault.
 185          */
 186         if (!vma->anon_vma)
 187                 return (smaps || in_pf) ? orders : 0;
 188
 189         return orders;
 190 }
 191
 192 static bool get_huge_zero_page(void)
 193 {
 194         struct page *zero_page;
 195 retry:
 196         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 197                 return true;
 198
 199         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 200                         HPAGE_PMD_ORDER);
 201         if (!zero_page) {
 202                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 203                 return false;
 204         }
 205         preempt_disable();
 206         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 207                 preempt_enable();
 208                 __free_pages(zero_page, compound_order(zero_page));
 209                 goto retry;
 210         }
 211         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 212
 213         /* We take additional reference here. It will be put back by shrinker */
 214         atomic_set(&huge_zero_refcount, 2);
 215         preempt_enable();
 216         count_vm_event(THP_ZERO_PAGE_ALLOC);
 217         return true;
 218 }
 219
 220 static void put_huge_zero_page(void)
 221 {
 222         /*
 223          * Counter should never go to zero here. Only shrinker can put
 224          * last reference.
 225          */
 226         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 227 }
 228
 229 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 230 {
 231         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 232                 return READ_ONCE(huge_zero_page);
 233
 234         if (!get_huge_zero_page())
 235                 return NULL;
 236
 237         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 238                 put_huge_zero_page();
 239
 240         return READ_ONCE(huge_zero_page);
 241 }
 242
 243 void mm_put_huge_zero_page(struct mm_struct *mm)
 244 {
 245         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 246                 put_huge_zero_page();
 247 }
 248
 249 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 250                                         struct shrink_control *sc)
 251 {
 252         /* we can free zero page only if last reference remains */
 253         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 254 }
 255
 256 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 257                                        struct shrink_control *sc)
 258 {
 259         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 260                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 261                 BUG_ON(zero_page == NULL);
 262                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 263                 __free_pages(zero_page, compound_order(zero_page));
 264                 return HPAGE_PMD_NR;
 265         }
 266
 267         return 0;
 268 }
 269
 270 static struct shrinker *huge_zero_page_shrinker;
 271
 272 #ifdef CONFIG_SYSFS
 273 static ssize_t enabled_show(struct kobject *kobj,
 274                             struct kobj_attribute *attr, char *buf)
 275 {
 276         const char *output;
 277
 278         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 279                 output = "[always] madvise never";
 280         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 281                           &transparent_hugepage_flags))
 282                 output = "always [madvise] never";
 283         else
 284                 output = "always madvise [never]";
 285
 286         return sysfs_emit(buf, "%s\n", output);
 287 }
 288
 289 static ssize_t enabled_store(struct kobject *kobj,
 290                              struct kobj_attribute *attr,
 291                              const char *buf, size_t count)
 292 {
 293         ssize_t ret = count;
 294
 295         if (sysfs_streq(buf, "always")) {
 296                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 297                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 298         } else if (sysfs_streq(buf, "madvise")) {
 299                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 300                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 301         } else if (sysfs_streq(buf, "never")) {
 302                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 303                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 304         } else
 305                 ret = -EINVAL;
 306
 307         if (ret > 0) {
 308                 int err = start_stop_khugepaged();
 309                 if (err)
 310                         ret = err;
 311         }
 312         return ret;
 313 }
 314
 315 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 316
 317 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 318                                   struct kobj_attribute *attr, char *buf,
 319                                   enum transparent_hugepage_flag flag)
 320 {
 321         return sysfs_emit(buf, "%d\n",
 322                           !!test_bit(flag, &transparent_hugepage_flags));
 323 }
 324
 325 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 326                                  struct kobj_attribute *attr,
 327                                  const char *buf, size_t count,
 328                                  enum transparent_hugepage_flag flag)
 329 {
 330         unsigned long value;
 331         int ret;
 332
 333         ret = kstrtoul(buf, 10, &value);
 334         if (ret < 0)
 335                 return ret;
 336         if (value > 1)
 337                 return -EINVAL;
 338
 339         if (value)
 340                 set_bit(flag, &transparent_hugepage_flags);
 341         else
 342                 clear_bit(flag, &transparent_hugepage_flags);
 343
 344         return count;
 345 }
 346
 347 static ssize_t defrag_show(struct kobject *kobj,
 348                            struct kobj_attribute *attr, char *buf)
 349 {
 350         const char *output;
 351
 352         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 353                      &transparent_hugepage_flags))
 354                 output = "[always] defer defer+madvise madvise never";
 355         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 356                           &transparent_hugepage_flags))
 357                 output = "always [defer] defer+madvise madvise never";
 358         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 359                           &transparent_hugepage_flags))
 360                 output = "always defer [defer+madvise] madvise never";
 361         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 362                           &transparent_hugepage_flags))
 363                 output = "always defer defer+madvise [madvise] never";
 364         else
 365                 output = "always defer defer+madvise madvise [never]";
 366
 367         return sysfs_emit(buf, "%s\n", output);
 368 }
 369
 370 static ssize_t defrag_store(struct kobject *kobj,
 371                             struct kobj_attribute *attr,
 372                             const char *buf, size_t count)
 373 {
 374         if (sysfs_streq(buf, "always")) {
 375                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 376                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 377                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 378                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 379         } else if (sysfs_streq(buf, "defer+madvise")) {
 380                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 381                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 382                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 383                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 384         } else if (sysfs_streq(buf, "defer")) {
 385                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 386                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 387                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 388                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 389         } else if (sysfs_streq(buf, "madvise")) {
 390                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 391                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 392                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 393                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 394         } else if (sysfs_streq(buf, "never")) {
 395                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 396                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 397                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 398                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 399         } else
 400                 return -EINVAL;
 401
 402         return count;
 403 }
 404 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 405
 406 static ssize_t use_zero_page_show(struct kobject *kobj,
 407                                   struct kobj_attribute *attr, char *buf)
 408 {
 409         return single_hugepage_flag_show(kobj, attr, buf,
 410                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 411 }
 412 static ssize_t use_zero_page_store(struct kobject *kobj,
 413                 struct kobj_attribute *attr, const char *buf, size_t count)
 414 {
 415         return single_hugepage_flag_store(kobj, attr, buf, count,
 416                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 417 }
 418 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 419
 420 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 421                                    struct kobj_attribute *attr, char *buf)
 422 {
 423         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 424 }
 425 static struct kobj_attribute hpage_pmd_size_attr =
 426         __ATTR_RO(hpage_pmd_size);
 427
 428 static struct attribute *hugepage_attr[] = {
 429         &enabled_attr.attr,
 430         &defrag_attr.attr,
 431         &use_zero_page_attr.attr,
 432         &hpage_pmd_size_attr.attr,
 433 #ifdef CONFIG_SHMEM
 434         &shmem_enabled_attr.attr,
 435 #endif
 436         NULL,
 437 };
 438
 439 static const struct attribute_group hugepage_attr_group = {
 440         .attrs = hugepage_attr,
 441 };
 442
 443 static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);
 444 static void thpsize_release(struct kobject *kobj);
 445 static DEFINE_SPINLOCK(huge_anon_orders_lock);
 446 static LIST_HEAD(thpsize_list);
 447
 448 struct thpsize {
 449         struct kobject kobj;
 450         struct list_head node;
 451         int order;
 452 };
 453
 454 #define to_thpsize(kobj) container_of(kobj, struct thpsize, kobj)
 455
 456 static ssize_t thpsize_enabled_show(struct kobject *kobj,
 457                                     struct kobj_attribute *attr, char *buf)
 458 {
 459         int order = to_thpsize(kobj)->order;
 460         const char *output;
 461
 462         if (test_bit(order, &huge_anon_orders_always))
 463                 output = "[always] inherit madvise never";
 464         else if (test_bit(order, &huge_anon_orders_inherit))
 465                 output = "always [inherit] madvise never";
 466         else if (test_bit(order, &huge_anon_orders_madvise))
 467                 output = "always inherit [madvise] never";
 468         else
 469                 output = "always inherit madvise [never]";
 470
 471         return sysfs_emit(buf, "%s\n", output);
 472 }
 473
 474 static ssize_t thpsize_enabled_store(struct kobject *kobj,
 475                                      struct kobj_attribute *attr,
 476                                      const char *buf, size_t count)
 477 {
 478         int order = to_thpsize(kobj)->order;
 479         ssize_t ret = count;
 480
 481         if (sysfs_streq(buf, "always")) {
 482                 spin_lock(&huge_anon_orders_lock);
 483                 clear_bit(order, &huge_anon_orders_inherit);
 484                 clear_bit(order, &huge_anon_orders_madvise);
 485                 set_bit(order, &huge_anon_orders_always);
 486                 spin_unlock(&huge_anon_orders_lock);
 487         } else if (sysfs_streq(buf, "inherit")) {
 488                 spin_lock(&huge_anon_orders_lock);
 489                 clear_bit(order, &huge_anon_orders_always);
 490                 clear_bit(order, &huge_anon_orders_madvise);
 491                 set_bit(order, &huge_anon_orders_inherit);
 492                 spin_unlock(&huge_anon_orders_lock);
 493         } else if (sysfs_streq(buf, "madvise")) {
 494                 spin_lock(&huge_anon_orders_lock);
 495                 clear_bit(order, &huge_anon_orders_always);
 496                 clear_bit(order, &huge_anon_orders_inherit);
 497                 set_bit(order, &huge_anon_orders_madvise);
 498                 spin_unlock(&huge_anon_orders_lock);
 499         } else if (sysfs_streq(buf, "never")) {
 500                 spin_lock(&huge_anon_orders_lock);
 501                 clear_bit(order, &huge_anon_orders_always);
 502                 clear_bit(order, &huge_anon_orders_inherit);
 503                 clear_bit(order, &huge_anon_orders_madvise);
 504                 spin_unlock(&huge_anon_orders_lock);
 505         } else
 506                 ret = -EINVAL;
 507
 508         return ret;
 509 }
 510
 511 static struct kobj_attribute thpsize_enabled_attr =
 512         __ATTR(enabled, 0644, thpsize_enabled_show, thpsize_enabled_store);
 513
 514 static struct attribute *thpsize_attrs[] = {
 515         &thpsize_enabled_attr.attr,
 516         NULL,
 517 };
 518
 519 static const struct attribute_group thpsize_attr_group = {
 520         .attrs = thpsize_attrs,
 521 };
 522
 523 static const struct kobj_type thpsize_ktype = {
 524         .release = &thpsize_release,
 525         .sysfs_ops = &kobj_sysfs_ops,
 526 };
 527
 528 static struct thpsize *thpsize_create(int order, struct kobject *parent)
 529 {
 530         unsigned long size = (PAGE_SIZE << order) / SZ_1K;
 531         struct thpsize *thpsize;
 532         int ret;
 533
 534         thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);
 535         if (!thpsize)
 536                 return ERR_PTR(-ENOMEM);
 537
 538         ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,
 539                                    "hugepages-%lukB", size);
 540         if (ret) {
 541                 kfree(thpsize);
 542                 return ERR_PTR(ret);
 543         }
 544
 545         ret = sysfs_create_group(&thpsize->kobj, &thpsize_attr_group);
 546         if (ret) {
 547                 kobject_put(&thpsize->kobj);
 548                 return ERR_PTR(ret);
 549         }
 550
 551         thpsize->order = order;
 552         return thpsize;
 553 }
 554
 555 static void thpsize_release(struct kobject *kobj)
 556 {
 557         kfree(to_thpsize(kobj));
 558 }
 559
 560 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 561 {
 562         int err;
 563         struct thpsize *thpsize;
 564         unsigned long orders;
 565         int order;
 566
 567         /*
 568          * Default to setting PMD-sized THP to inherit the global setting and
 569          * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time
 570          * constant so we have to do this here.
 571          */
 572         huge_anon_orders_inherit = BIT(PMD_ORDER);
 573
 574         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 575         if (unlikely(!*hugepage_kobj)) {
 576                 pr_err("failed to create transparent hugepage kobject\n");
 577                 return -ENOMEM;
 578         }
 579
 580         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 581         if (err) {
 582                 pr_err("failed to register transparent hugepage group\n");
 583                 goto delete_obj;
 584         }
 585
 586         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 587         if (err) {
 588                 pr_err("failed to register transparent hugepage group\n");
 589                 goto remove_hp_group;
 590         }
 591
 592         orders = THP_ORDERS_ALL_ANON;
 593         order = highest_order(orders);
 594         while (orders) {
 595                 thpsize = thpsize_create(order, *hugepage_kobj);
 596                 if (IS_ERR(thpsize)) {
 597                         pr_err("failed to create thpsize for order %d\n", order);
 598                         err = PTR_ERR(thpsize);
 599                         goto remove_all;
 600                 }
 601                 list_add(&thpsize->node, &thpsize_list);
 602                 order = next_order(&orders, order);
 603         }
 604
 605         return 0;
 606
 607 remove_all:
 608         hugepage_exit_sysfs(*hugepage_kobj);
 609         return err;
 610 remove_hp_group:
 611         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 612 delete_obj:
 613         kobject_put(*hugepage_kobj);
 614         return err;
 615 }
 616
 617 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 618 {
 619         struct thpsize *thpsize, *tmp;
 620
 621         list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {
 622                 list_del(&thpsize->node);
 623                 kobject_put(&thpsize->kobj);
 624         }
 625
 626         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 627         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 628         kobject_put(hugepage_kobj);
 629 }
 630 #else
 631 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 632 {
 633         return 0;
 634 }
 635
 636 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 637 {
 638 }
 639 #endif /* CONFIG_SYSFS */
 640
 641 static int __init thp_shrinker_init(void)
 642 {
 643         huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");
 644         if (!huge_zero_page_shrinker)
 645                 return -ENOMEM;
 646
 647         deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
 648                                                  SHRINKER_MEMCG_AWARE |
 649                                                  SHRINKER_NONSLAB,
 650                                                  "thp-deferred_split");
 651         if (!deferred_split_shrinker) {
 652                 shrinker_free(huge_zero_page_shrinker);
 653                 return -ENOMEM;
 654         }
 655
 656         huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;
 657         huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;
 658         shrinker_register(huge_zero_page_shrinker);
 659
 660         deferred_split_shrinker->count_objects = deferred_split_count;
 661         deferred_split_shrinker->scan_objects = deferred_split_scan;
 662         shrinker_register(deferred_split_shrinker);
 663
 664         return 0;
 665 }
 666
 667 static void __init thp_shrinker_exit(void)
 668 {
 669         shrinker_free(huge_zero_page_shrinker);
 670         shrinker_free(deferred_split_shrinker);
 671 }
 672
 673 static int __init hugepage_init(void)
 674 {
 675         int err;
 676         struct kobject *hugepage_kobj;
 677
 678         if (!has_transparent_hugepage()) {
 679                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
 680                 return -EINVAL;
 681         }
 682
 683         /*
 684          * hugepages can't be allocated by the buddy allocator
 685          */
 686         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);
 687         /*
 688          * we use page->mapping and page->index in second tail page
 689          * as list_head: assuming THP order >= 2
 690          */
 691         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 692
 693         err = hugepage_init_sysfs(&hugepage_kobj);
 694         if (err)
 695                 goto err_sysfs;
 696
 697         err = khugepaged_init();
 698         if (err)
 699                 goto err_slab;
 700
 701         err = thp_shrinker_init();
 702         if (err)
 703                 goto err_shrinker;
 704
 705         /*
 706          * By default disable transparent hugepages on smaller systems,
 707          * where the extra memory used could hurt more than TLB overhead
 708          * is likely to save.  The admin can still enable it through /sys.
 709          */
 710         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 711                 transparent_hugepage_flags = 0;
 712                 return 0;
 713         }
 714
 715         err = start_stop_khugepaged();
 716         if (err)
 717                 goto err_khugepaged;
 718
 719         return 0;
 720 err_khugepaged:
 721         thp_shrinker_exit();
 722 err_shrinker:
 723         khugepaged_destroy();
 724 err_slab:
 725         hugepage_exit_sysfs(hugepage_kobj);
 726 err_sysfs:
 727         return err;
 728 }
 729 subsys_initcall(hugepage_init);
 730
 731 static int __init setup_transparent_hugepage(char *str)
 732 {
 733         int ret = 0;
 734         if (!str)
 735                 goto out;
 736         if (!strcmp(str, "always")) {
 737                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 738                         &transparent_hugepage_flags);
 739                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 740                           &transparent_hugepage_flags);
 741                 ret = 1;
 742         } else if (!strcmp(str, "madvise")) {
 743                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 744                           &transparent_hugepage_flags);
 745                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 746                         &transparent_hugepage_flags);
 747                 ret = 1;
 748         } else if (!strcmp(str, "never")) {
 749                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 750                           &transparent_hugepage_flags);
 751                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 752                           &transparent_hugepage_flags);
 753                 ret = 1;
 754         }
 755 out:
 756         if (!ret)
 757                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 758         return ret;
 759 }
 760 __setup("transparent_hugepage=", setup_transparent_hugepage);
 761
 762 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 763 {
 764         if (likely(vma->vm_flags & VM_WRITE))
 765                 pmd = pmd_mkwrite(pmd, vma);
 766         return pmd;
 767 }
 768
 769 #ifdef CONFIG_MEMCG
 770 static inline
 771 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 772 {
 773         struct mem_cgroup *memcg = folio_memcg(folio);
 774         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 775
 776         if (memcg)
 777                 return &memcg->deferred_split_queue;
 778         else
 779                 return &pgdat->deferred_split_queue;
 780 }
 781 #else
 782 static inline
 783 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 784 {
 785         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 786
 787         return &pgdat->deferred_split_queue;
 788 }
 789 #endif
 790
 791 void folio_prep_large_rmappable(struct folio *folio)
 792 {
 793         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
 794         INIT_LIST_HEAD(&folio->_deferred_list);
 795         folio_set_large_rmappable(folio);
 796 }
 797
 798 static inline bool is_transparent_hugepage(struct folio *folio)
 799 {
 800         if (!folio_test_large(folio))
 801                 return false;
 802
 803         return is_huge_zero_page(&folio->page) ||
 804                 folio_test_large_rmappable(folio);
 805 }
 806
 807 static unsigned long __thp_get_unmapped_area(struct file *filp,
 808                 unsigned long addr, unsigned long len,
 809                 loff_t off, unsigned long flags, unsigned long size)
 810 {
 811         loff_t off_end = off + len;
 812         loff_t off_align = round_up(off, size);
 813         unsigned long len_pad, ret;
 814
 815         if (IS_ENABLED(CONFIG_32BIT) || in_compat_syscall())
 816                 return 0;
 817
 818         if (off_end <= off_align || (off_end - off_align) < size)
 819                 return 0;
 820
 821         len_pad = len + size;
 822         if (len_pad < len || (off + len_pad) < off)
 823                 return 0;
 824
 825         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 826                                               off >> PAGE_SHIFT, flags);
 827
 828         /*
 829          * The failure might be due to length padding. The caller will retry
 830          * without the padding.
 831          */
 832         if (IS_ERR_VALUE(ret))
 833                 return 0;
 834
 835         /*
 836          * Do not try to align to THP boundary if allocation at the address
 837          * hint succeeds.
 838          */
 839         if (ret == addr)
 840                 return addr;
 841
 842         ret += (off - ret) & (size - 1);
 843         return ret;
 844 }
 845
 846 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 847                 unsigned long len, unsigned long pgoff, unsigned long flags)
 848 {
 849         unsigned long ret;
 850         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 851
 852         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 853         if (ret)
 854                 return ret;
 855
 856         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 857 }
 858 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 859
 860 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 861                         struct page *page, gfp_t gfp)
 862 {
 863         struct vm_area_struct *vma = vmf->vma;
 864         struct folio *folio = page_folio(page);
 865         pgtable_t pgtable;
 866         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 867         vm_fault_t ret = 0;
 868
 869         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
 870
 871         if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
 872                 folio_put(folio);
 873                 count_vm_event(THP_FAULT_FALLBACK);
 874                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 875                 return VM_FAULT_FALLBACK;
 876         }
 877         folio_throttle_swaprate(folio, gfp);
 878
 879         pgtable = pte_alloc_one(vma->vm_mm);
 880         if (unlikely(!pgtable)) {
 881                 ret = VM_FAULT_OOM;
 882                 goto release;
 883         }
 884
 885         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 886         /*
 887          * The memory barrier inside __folio_mark_uptodate makes sure that
 888          * clear_huge_page writes become visible before the set_pmd_at()
 889          * write.
 890          */
 891         __folio_mark_uptodate(folio);
 892
 893         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 894         if (unlikely(!pmd_none(*vmf->pmd))) {
 895                 goto unlock_release;
 896         } else {
 897                 pmd_t entry;
 898
 899                 ret = check_stable_address_space(vma->vm_mm);
 900                 if (ret)
 901                         goto unlock_release;
 902
 903                 /* Deliver the page fault to userland */
 904                 if (userfaultfd_missing(vma)) {
 905                         spin_unlock(vmf->ptl);
 906                         folio_put(folio);
 907                         pte_free(vma->vm_mm, pgtable);
 908                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 909                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 910                         return ret;
 911                 }
 912
 913                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 914                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 915                 folio_add_new_anon_rmap(folio, vma, haddr);
 916                 folio_add_lru_vma(folio, vma);
 917                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 918                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 919                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 920                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 921                 mm_inc_nr_ptes(vma->vm_mm);
 922                 spin_unlock(vmf->ptl);
 923                 count_vm_event(THP_FAULT_ALLOC);
 924                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 925         }
 926
 927         return 0;
 928 unlock_release:
 929         spin_unlock(vmf->ptl);
 930 release:
 931         if (pgtable)
 932                 pte_free(vma->vm_mm, pgtable);
 933         folio_put(folio);
 934         return ret;
 935
 936 }
 937
 938 /*
 939  * always: directly stall for all thp allocations
 940  * defer: wake kswapd and fail if not immediately available
 941  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 942  *                fail if not immediately available
 943  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 944  *          available
 945  * never: never stall for any thp allocation
 946  */
 947 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 948 {
 949         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 950
 951         /* Always do synchronous compaction */
 952         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 953                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 954
 955         /* Kick kcompactd and fail quickly */
 956         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 957                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 958
 959         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 960         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 961                 return GFP_TRANSHUGE_LIGHT |
 962                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 963                                         __GFP_KSWAPD_RECLAIM);
 964
 965         /* Only do synchronous compaction if madvised */
 966         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 967                 return GFP_TRANSHUGE_LIGHT |
 968                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 969
 970         return GFP_TRANSHUGE_LIGHT;
 971 }
 972
 973 /* Caller must hold page table lock. */
 974 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 975                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 976                 struct page *zero_page)
 977 {
 978         pmd_t entry;
 979         if (!pmd_none(*pmd))
 980                 return;
 981         entry = mk_pmd(zero_page, vma->vm_page_prot);
 982         entry = pmd_mkhuge(entry);
 983         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 984         set_pmd_at(mm, haddr, pmd, entry);
 985         mm_inc_nr_ptes(mm);
 986 }
 987
 988 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 989 {
 990         struct vm_area_struct *vma = vmf->vma;
 991         gfp_t gfp;
 992         struct folio *folio;
 993         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 994
 995         if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
 996                 return VM_FAULT_FALLBACK;
 997         if (unlikely(anon_vma_prepare(vma)))
 998                 return VM_FAULT_OOM;
 999         khugepaged_enter_vma(vma, vma->vm_flags);
1000
1001         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
1002                         !mm_forbids_zeropage(vma->vm_mm) &&
1003                         transparent_hugepage_use_zero_page()) {
1004                 pgtable_t pgtable;
1005                 struct page *zero_page;
1006                 vm_fault_t ret;
1007                 pgtable = pte_alloc_one(vma->vm_mm);
1008                 if (unlikely(!pgtable))
1009                         return VM_FAULT_OOM;
1010                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
1011                 if (unlikely(!zero_page)) {
1012                         pte_free(vma->vm_mm, pgtable);
1013                         count_vm_event(THP_FAULT_FALLBACK);
1014                         return VM_FAULT_FALLBACK;
1015                 }
1016                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1017                 ret = 0;
1018                 if (pmd_none(*vmf->pmd)) {
1019                         ret = check_stable_address_space(vma->vm_mm);
1020                         if (ret) {
1021                                 spin_unlock(vmf->ptl);
1022                                 pte_free(vma->vm_mm, pgtable);
1023                         } else if (userfaultfd_missing(vma)) {
1024                                 spin_unlock(vmf->ptl);
1025                                 pte_free(vma->vm_mm, pgtable);
1026                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
1027                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1028                         } else {
1029                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
1030                                                    haddr, vmf->pmd, zero_page);
1031                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1032                                 spin_unlock(vmf->ptl);
1033                         }
1034                 } else {
1035                         spin_unlock(vmf->ptl);
1036                         pte_free(vma->vm_mm, pgtable);
1037                 }
1038                 return ret;
1039         }
1040         gfp = vma_thp_gfp_mask(vma);
1041         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
1042         if (unlikely(!folio)) {
1043                 count_vm_event(THP_FAULT_FALLBACK);
1044                 return VM_FAULT_FALLBACK;
1045         }
1046         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
1047 }
1048
1049 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1050                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
1051                 pgtable_t pgtable)
1052 {
1053         struct mm_struct *mm = vma->vm_mm;
1054         pmd_t entry;
1055         spinlock_t *ptl;
1056
1057         ptl = pmd_lock(mm, pmd);
1058         if (!pmd_none(*pmd)) {
1059                 if (write) {
1060                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
1061                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
1062                                 goto out_unlock;
1063                         }
1064                         entry = pmd_mkyoung(*pmd);
1065                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1066                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
1067                                 update_mmu_cache_pmd(vma, addr, pmd);
1068                 }
1069
1070                 goto out_unlock;
1071         }
1072
1073         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1074         if (pfn_t_devmap(pfn))
1075                 entry = pmd_mkdevmap(entry);
1076         if (write) {
1077                 entry = pmd_mkyoung(pmd_mkdirty(entry));
1078                 entry = maybe_pmd_mkwrite(entry, vma);
1079         }
1080
1081         if (pgtable) {
1082                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1083                 mm_inc_nr_ptes(mm);
1084                 pgtable = NULL;
1085         }
1086
1087         set_pmd_at(mm, addr, pmd, entry);
1088         update_mmu_cache_pmd(vma, addr, pmd);
1089
1090 out_unlock:
1091         spin_unlock(ptl);
1092         if (pgtable)
1093                 pte_free(mm, pgtable);
1094 }
1095
1096 /**
1097  * vmf_insert_pfn_pmd - insert a pmd size pfn
1098  * @vmf: Structure describing the fault
1099  * @pfn: pfn to insert
1100  * @write: whether it's a write fault
1101  *
1102  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
1103  *
1104  * Return: vm_fault_t value.
1105  */
1106 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
1107 {
1108         unsigned long addr = vmf->address & PMD_MASK;
1109         struct vm_area_struct *vma = vmf->vma;
1110         pgprot_t pgprot = vma->vm_page_prot;
1111         pgtable_t pgtable = NULL;
1112
1113         /*
1114          * If we had pmd_special, we could avoid all these restrictions,
1115          * but we need to be consistent with PTEs and architectures that
1116          * can't support a 'special' bit.
1117          */
1118         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1119                         !pfn_t_devmap(pfn));
1120         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1121                                                 (VM_PFNMAP|VM_MIXEDMAP));
1122         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1123
1124         if (addr < vma->vm_start || addr >= vma->vm_end)
1125                 return VM_FAULT_SIGBUS;
1126
1127         if (arch_needs_pgtable_deposit()) {
1128                 pgtable = pte_alloc_one(vma->vm_mm);
1129                 if (!pgtable)
1130                         return VM_FAULT_OOM;
1131         }
1132
1133         track_pfn_insert(vma, &pgprot, pfn);
1134
1135         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
1136         return VM_FAULT_NOPAGE;
1137 }
1138 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1139
1140 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1141 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
1142 {
1143         if (likely(vma->vm_flags & VM_WRITE))
1144                 pud = pud_mkwrite(pud);
1145         return pud;
1146 }
1147
1148 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
1149                 pud_t *pud, pfn_t pfn, bool write)
1150 {
1151         struct mm_struct *mm = vma->vm_mm;
1152         pgprot_t prot = vma->vm_page_prot;
1153         pud_t entry;
1154         spinlock_t *ptl;
1155
1156         ptl = pud_lock(mm, pud);
1157         if (!pud_none(*pud)) {
1158                 if (write) {
1159                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
1160                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
1161                                 goto out_unlock;
1162                         }
1163                         entry = pud_mkyoung(*pud);
1164                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
1165                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
1166                                 update_mmu_cache_pud(vma, addr, pud);
1167                 }
1168                 goto out_unlock;
1169         }
1170
1171         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
1172         if (pfn_t_devmap(pfn))
1173                 entry = pud_mkdevmap(entry);
1174         if (write) {
1175                 entry = pud_mkyoung(pud_mkdirty(entry));
1176                 entry = maybe_pud_mkwrite(entry, vma);
1177         }
1178         set_pud_at(mm, addr, pud, entry);
1179         update_mmu_cache_pud(vma, addr, pud);
1180
1181 out_unlock:
1182         spin_unlock(ptl);
1183 }
1184
1185 /**
1186  * vmf_insert_pfn_pud - insert a pud size pfn
1187  * @vmf: Structure describing the fault
1188  * @pfn: pfn to insert
1189  * @write: whether it's a write fault
1190  *
1191  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1192  *
1193  * Return: vm_fault_t value.
1194  */
1195 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
1196 {
1197         unsigned long addr = vmf->address & PUD_MASK;
1198         struct vm_area_struct *vma = vmf->vma;
1199         pgprot_t pgprot = vma->vm_page_prot;
1200
1201         /*
1202          * If we had pud_special, we could avoid all these restrictions,
1203          * but we need to be consistent with PTEs and architectures that
1204          * can't support a 'special' bit.
1205          */
1206         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1207                         !pfn_t_devmap(pfn));
1208         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1209                                                 (VM_PFNMAP|VM_MIXEDMAP));
1210         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1211
1212         if (addr < vma->vm_start || addr >= vma->vm_end)
1213                 return VM_FAULT_SIGBUS;
1214
1215         track_pfn_insert(vma, &pgprot, pfn);
1216
1217         insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1218         return VM_FAULT_NOPAGE;
1219 }
1220 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1221 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1222
1223 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1224                       pmd_t *pmd, bool write)
1225 {
1226         pmd_t _pmd;
1227
1228         _pmd = pmd_mkyoung(*pmd);
1229         if (write)
1230                 _pmd = pmd_mkdirty(_pmd);
1231         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1232                                   pmd, _pmd, write))
1233                 update_mmu_cache_pmd(vma, addr, pmd);
1234 }
1235
1236 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1237                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1238 {
1239         unsigned long pfn = pmd_pfn(*pmd);
1240         struct mm_struct *mm = vma->vm_mm;
1241         struct page *page;
1242         int ret;
1243
1244         assert_spin_locked(pmd_lockptr(mm, pmd));
1245
1246         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1247                 return NULL;
1248
1249         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1250                 /* pass */;
1251         else
1252                 return NULL;
1253
1254         if (flags & FOLL_TOUCH)
1255                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1256
1257         /*
1258          * device mapped pages can only be returned if the
1259          * caller will manage the page reference count.
1260          */
1261         if (!(flags & (FOLL_GET | FOLL_PIN)))
1262                 return ERR_PTR(-EEXIST);
1263
1264         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1265         *pgmap = get_dev_pagemap(pfn, *pgmap);
1266         if (!*pgmap)
1267                 return ERR_PTR(-EFAULT);
1268         page = pfn_to_page(pfn);
1269         ret = try_grab_page(page, flags);
1270         if (ret)
1271                 page = ERR_PTR(ret);
1272
1273         return page;
1274 }
1275
1276 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1277                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1278                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1279 {
1280         spinlock_t *dst_ptl, *src_ptl;
1281         struct page *src_page;
1282         struct folio *src_folio;
1283         pmd_t pmd;
1284         pgtable_t pgtable = NULL;
1285         int ret = -ENOMEM;
1286
1287         /* Skip if can be re-fill on fault */
1288         if (!vma_is_anonymous(dst_vma))
1289                 return 0;
1290
1291         pgtable = pte_alloc_one(dst_mm);
1292         if (unlikely(!pgtable))
1293                 goto out;
1294
1295         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1296         src_ptl = pmd_lockptr(src_mm, src_pmd);
1297         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1298
1299         ret = -EAGAIN;
1300         pmd = *src_pmd;
1301
1302 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1303         if (unlikely(is_swap_pmd(pmd))) {
1304                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1305
1306                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1307                 if (!is_readable_migration_entry(entry)) {
1308                         entry = make_readable_migration_entry(
1309                                                         swp_offset(entry));
1310                         pmd = swp_entry_to_pmd(entry);
1311                         if (pmd_swp_soft_dirty(*src_pmd))
1312                                 pmd = pmd_swp_mksoft_dirty(pmd);
1313                         if (pmd_swp_uffd_wp(*src_pmd))
1314                                 pmd = pmd_swp_mkuffd_wp(pmd);
1315                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1316                 }
1317                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1318                 mm_inc_nr_ptes(dst_mm);
1319                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1320                 if (!userfaultfd_wp(dst_vma))
1321                         pmd = pmd_swp_clear_uffd_wp(pmd);
1322                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1323                 ret = 0;
1324                 goto out_unlock;
1325         }
1326 #endif
1327
1328         if (unlikely(!pmd_trans_huge(pmd))) {
1329                 pte_free(dst_mm, pgtable);
1330                 goto out_unlock;
1331         }
1332         /*
1333          * When page table lock is held, the huge zero pmd should not be
1334          * under splitting since we don't split the page itself, only pmd to
1335          * a page table.
1336          */
1337         if (is_huge_zero_pmd(pmd)) {
1338                 /*
1339                  * get_huge_zero_page() will never allocate a new page here,
1340                  * since we already have a zero page to copy. It just takes a
1341                  * reference.
1342                  */
1343                 mm_get_huge_zero_page(dst_mm);
1344                 goto out_zero_page;
1345         }
1346
1347         src_page = pmd_page(pmd);
1348         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1349         src_folio = page_folio(src_page);
1350
1351         folio_get(src_folio);
1352         if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) {
1353                 /* Page maybe pinned: split and retry the fault on PTEs. */
1354                 folio_put(src_folio);
1355                 pte_free(dst_mm, pgtable);
1356                 spin_unlock(src_ptl);
1357                 spin_unlock(dst_ptl);
1358                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1359                 return -EAGAIN;
1360         }
1361         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1362 out_zero_page:
1363         mm_inc_nr_ptes(dst_mm);
1364         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1365         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1366         if (!userfaultfd_wp(dst_vma))
1367                 pmd = pmd_clear_uffd_wp(pmd);
1368         pmd = pmd_mkold(pmd_wrprotect(pmd));
1369         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1370
1371         ret = 0;
1372 out_unlock:
1373         spin_unlock(src_ptl);
1374         spin_unlock(dst_ptl);
1375 out:
1376         return ret;
1377 }
1378
1379 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1380 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1381                       pud_t *pud, bool write)
1382 {
1383         pud_t _pud;
1384
1385         _pud = pud_mkyoung(*pud);
1386         if (write)
1387                 _pud = pud_mkdirty(_pud);
1388         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1389                                   pud, _pud, write))
1390                 update_mmu_cache_pud(vma, addr, pud);
1391 }
1392
1393 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1394                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1395 {
1396         unsigned long pfn = pud_pfn(*pud);
1397         struct mm_struct *mm = vma->vm_mm;
1398         struct page *page;
1399         int ret;
1400
1401         assert_spin_locked(pud_lockptr(mm, pud));
1402
1403         if (flags & FOLL_WRITE && !pud_write(*pud))
1404                 return NULL;
1405
1406         if (pud_present(*pud) && pud_devmap(*pud))
1407                 /* pass */;
1408         else
1409                 return NULL;
1410
1411         if (flags & FOLL_TOUCH)
1412                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1413
1414         /*
1415          * device mapped pages can only be returned if the
1416          * caller will manage the page reference count.
1417          *
1418          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1419          */
1420         if (!(flags & (FOLL_GET | FOLL_PIN)))
1421                 return ERR_PTR(-EEXIST);
1422
1423         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1424         *pgmap = get_dev_pagemap(pfn, *pgmap);
1425         if (!*pgmap)
1426                 return ERR_PTR(-EFAULT);
1427         page = pfn_to_page(pfn);
1428
1429         ret = try_grab_page(page, flags);
1430         if (ret)
1431                 page = ERR_PTR(ret);
1432
1433         return page;
1434 }
1435
1436 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1437                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1438                   struct vm_area_struct *vma)
1439 {
1440         spinlock_t *dst_ptl, *src_ptl;
1441         pud_t pud;
1442         int ret;
1443
1444         dst_ptl = pud_lock(dst_mm, dst_pud);
1445         src_ptl = pud_lockptr(src_mm, src_pud);
1446         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1447
1448         ret = -EAGAIN;
1449         pud = *src_pud;
1450         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1451                 goto out_unlock;
1452
1453         /*
1454          * When page table lock is held, the huge zero pud should not be
1455          * under splitting since we don't split the page itself, only pud to
1456          * a page table.
1457          */
1458         if (is_huge_zero_pud(pud)) {
1459                 /* No huge zero pud yet */
1460         }
1461
1462         /*
1463          * TODO: once we support anonymous pages, use
1464          * folio_try_dup_anon_rmap_*() and split if duplicating fails.
1465          */
1466         pudp_set_wrprotect(src_mm, addr, src_pud);
1467         pud = pud_mkold(pud_wrprotect(pud));
1468         set_pud_at(dst_mm, addr, dst_pud, pud);
1469
1470         ret = 0;
1471 out_unlock:
1472         spin_unlock(src_ptl);
1473         spin_unlock(dst_ptl);
1474         return ret;
1475 }
1476
1477 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1478 {
1479         bool write = vmf->flags & FAULT_FLAG_WRITE;
1480
1481         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1482         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1483                 goto unlock;
1484
1485         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1486 unlock:
1487         spin_unlock(vmf->ptl);
1488 }
1489 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1490
1491 void huge_pmd_set_accessed(struct vm_fault *vmf)
1492 {
1493         bool write = vmf->flags & FAULT_FLAG_WRITE;
1494
1495         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1496         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1497                 goto unlock;
1498
1499         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1500
1501 unlock:
1502         spin_unlock(vmf->ptl);
1503 }
1504
1505 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1506 {
1507         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1508         struct vm_area_struct *vma = vmf->vma;
1509         struct folio *folio;
1510         struct page *page;
1511         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1512         pmd_t orig_pmd = vmf->orig_pmd;
1513
1514         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1515         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1516
1517         if (is_huge_zero_pmd(orig_pmd))
1518                 goto fallback;
1519
1520         spin_lock(vmf->ptl);
1521
1522         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1523                 spin_unlock(vmf->ptl);
1524                 return 0;
1525         }
1526
1527         page = pmd_page(orig_pmd);
1528         folio = page_folio(page);
1529         VM_BUG_ON_PAGE(!PageHead(page), page);
1530
1531         /* Early check when only holding the PT lock. */
1532         if (PageAnonExclusive(page))
1533                 goto reuse;
1534
1535         if (!folio_trylock(folio)) {
1536                 folio_get(folio);
1537                 spin_unlock(vmf->ptl);
1538                 folio_lock(folio);
1539                 spin_lock(vmf->ptl);
1540                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1541                         spin_unlock(vmf->ptl);
1542                         folio_unlock(folio);
1543                         folio_put(folio);
1544                         return 0;
1545                 }
1546                 folio_put(folio);
1547         }
1548
1549         /* Recheck after temporarily dropping the PT lock. */
1550         if (PageAnonExclusive(page)) {
1551                 folio_unlock(folio);
1552                 goto reuse;
1553         }
1554
1555         /*
1556          * See do_wp_page(): we can only reuse the folio exclusively if
1557          * there are no additional references. Note that we always drain
1558          * the LRU cache immediately after adding a THP.
1559          */
1560         if (folio_ref_count(folio) >
1561                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1562                 goto unlock_fallback;
1563         if (folio_test_swapcache(folio))
1564                 folio_free_swap(folio);
1565         if (folio_ref_count(folio) == 1) {
1566                 pmd_t entry;
1567
1568                 folio_move_anon_rmap(folio, vma);
1569                 SetPageAnonExclusive(page);
1570                 folio_unlock(folio);
1571 reuse:
1572                 if (unlikely(unshare)) {
1573                         spin_unlock(vmf->ptl);
1574                         return 0;
1575                 }
1576                 entry = pmd_mkyoung(orig_pmd);
1577                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1578                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1579                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1580                 spin_unlock(vmf->ptl);
1581                 return 0;
1582         }
1583
1584 unlock_fallback:
1585         folio_unlock(folio);
1586         spin_unlock(vmf->ptl);
1587 fallback:
1588         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1589         return VM_FAULT_FALLBACK;
1590 }
1591
1592 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1593                                            unsigned long addr, pmd_t pmd)
1594 {
1595         struct page *page;
1596
1597         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1598                 return false;
1599
1600         /* Don't touch entries that are not even readable (NUMA hinting). */
1601         if (pmd_protnone(pmd))
1602                 return false;
1603
1604         /* Do we need write faults for softdirty tracking? */
1605         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1606                 return false;
1607
1608         /* Do we need write faults for uffd-wp tracking? */
1609         if (userfaultfd_huge_pmd_wp(vma, pmd))
1610                 return false;
1611
1612         if (!(vma->vm_flags & VM_SHARED)) {
1613                 /* See can_change_pte_writable(). */
1614                 page = vm_normal_page_pmd(vma, addr, pmd);
1615                 return page && PageAnon(page) && PageAnonExclusive(page);
1616         }
1617
1618         /* See can_change_pte_writable(). */
1619         return pmd_dirty(pmd);
1620 }
1621
1622 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1623 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1624                                         struct vm_area_struct *vma,
1625                                         unsigned int flags)
1626 {
1627         /* If the pmd is writable, we can write to the page. */
1628         if (pmd_write(pmd))
1629                 return true;
1630
1631         /* Maybe FOLL_FORCE is set to override it? */
1632         if (!(flags & FOLL_FORCE))
1633                 return false;
1634
1635         /* But FOLL_FORCE has no effect on shared mappings */
1636         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1637                 return false;
1638
1639         /* ... or read-only private ones */
1640         if (!(vma->vm_flags & VM_MAYWRITE))
1641                 return false;
1642
1643         /* ... or already writable ones that just need to take a write fault */
1644         if (vma->vm_flags & VM_WRITE)
1645                 return false;
1646
1647         /*
1648          * See can_change_pte_writable(): we broke COW and could map the page
1649          * writable if we have an exclusive anonymous page ...
1650          */
1651         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1652                 return false;
1653
1654         /* ... and a write-fault isn't required for other reasons. */
1655         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1656                 return false;
1657         return !userfaultfd_huge_pmd_wp(vma, pmd);
1658 }
1659
1660 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1661                                    unsigned long addr,
1662                                    pmd_t *pmd,
1663                                    unsigned int flags)
1664 {
1665         struct mm_struct *mm = vma->vm_mm;
1666         struct page *page;
1667         int ret;
1668
1669         assert_spin_locked(pmd_lockptr(mm, pmd));
1670
1671         page = pmd_page(*pmd);
1672         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1673
1674         if ((flags & FOLL_WRITE) &&
1675             !can_follow_write_pmd(*pmd, page, vma, flags))
1676                 return NULL;
1677
1678         /* Avoid dumping huge zero page */
1679         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1680                 return ERR_PTR(-EFAULT);
1681
1682         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
1683                 return NULL;
1684
1685         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1686                 return ERR_PTR(-EMLINK);
1687
1688         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1689                         !PageAnonExclusive(page), page);
1690
1691         ret = try_grab_page(page, flags);
1692         if (ret)
1693                 return ERR_PTR(ret);
1694
1695         if (flags & FOLL_TOUCH)
1696                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1697
1698         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1699         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1700
1701         return page;
1702 }
1703
1704 /* NUMA hinting page fault entry point for trans huge pmds */
1705 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1706 {
1707         struct vm_area_struct *vma = vmf->vma;
1708         pmd_t oldpmd = vmf->orig_pmd;
1709         pmd_t pmd;
1710         struct folio *folio;
1711         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1712         int nid = NUMA_NO_NODE;
1713         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1714         bool migrated = false, writable = false;
1715         int flags = 0;
1716
1717         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1718         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1719                 spin_unlock(vmf->ptl);
1720                 goto out;
1721         }
1722
1723         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1724
1725         /*
1726          * Detect now whether the PMD could be writable; this information
1727          * is only valid while holding the PT lock.
1728          */
1729         writable = pmd_write(pmd);
1730         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1731             can_change_pmd_writable(vma, vmf->address, pmd))
1732                 writable = true;
1733
1734         folio = vm_normal_folio_pmd(vma, haddr, pmd);
1735         if (!folio)
1736                 goto out_map;
1737
1738         /* See similar comment in do_numa_page for explanation */
1739         if (!writable)
1740                 flags |= TNF_NO_GROUP;
1741
1742         nid = folio_nid(folio);
1743         /*
1744          * For memory tiering mode, cpupid of slow memory page is used
1745          * to record page access time.  So use default value.
1746          */
1747         if (node_is_toptier(nid))
1748                 last_cpupid = folio_last_cpupid(folio);
1749         target_nid = numa_migrate_prep(folio, vma, haddr, nid, &flags);
1750         if (target_nid == NUMA_NO_NODE) {
1751                 folio_put(folio);
1752                 goto out_map;
1753         }
1754
1755         spin_unlock(vmf->ptl);
1756         writable = false;
1757
1758         migrated = migrate_misplaced_folio(folio, vma, target_nid);
1759         if (migrated) {
1760                 flags |= TNF_MIGRATED;
1761                 nid = target_nid;
1762         } else {
1763                 flags |= TNF_MIGRATE_FAIL;
1764                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1765                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1766                         spin_unlock(vmf->ptl);
1767                         goto out;
1768                 }
1769                 goto out_map;
1770         }
1771
1772 out:
1773         if (nid != NUMA_NO_NODE)
1774                 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
1775
1776         return 0;
1777
1778 out_map:
1779         /* Restore the PMD */
1780         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1781         pmd = pmd_mkyoung(pmd);
1782         if (writable)
1783                 pmd = pmd_mkwrite(pmd, vma);
1784         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1785         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1786         spin_unlock(vmf->ptl);
1787         goto out;
1788 }
1789
1790 /*
1791  * Return true if we do MADV_FREE successfully on entire pmd page.
1792  * Otherwise, return false.
1793  */
1794 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1795                 pmd_t *pmd, unsigned long addr, unsigned long next)
1796 {
1797         spinlock_t *ptl;
1798         pmd_t orig_pmd;
1799         struct folio *folio;
1800         struct mm_struct *mm = tlb->mm;
1801         bool ret = false;
1802
1803         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1804
1805         ptl = pmd_trans_huge_lock(pmd, vma);
1806         if (!ptl)
1807                 goto out_unlocked;
1808
1809         orig_pmd = *pmd;
1810         if (is_huge_zero_pmd(orig_pmd))
1811                 goto out;
1812
1813         if (unlikely(!pmd_present(orig_pmd))) {
1814                 VM_BUG_ON(thp_migration_supported() &&
1815                                   !is_pmd_migration_entry(orig_pmd));
1816                 goto out;
1817         }
1818
1819         folio = pfn_folio(pmd_pfn(orig_pmd));
1820         /*
1821          * If other processes are mapping this folio, we couldn't discard
1822          * the folio unless they all do MADV_FREE so let's skip the folio.
1823          */
1824         if (folio_estimated_sharers(folio) != 1)
1825                 goto out;
1826
1827         if (!folio_trylock(folio))
1828                 goto out;
1829
1830         /*
1831          * If user want to discard part-pages of THP, split it so MADV_FREE
1832          * will deactivate only them.
1833          */
1834         if (next - addr != HPAGE_PMD_SIZE) {
1835                 folio_get(folio);
1836                 spin_unlock(ptl);
1837                 split_folio(folio);
1838                 folio_unlock(folio);
1839                 folio_put(folio);
1840                 goto out_unlocked;
1841         }
1842
1843         if (folio_test_dirty(folio))
1844                 folio_clear_dirty(folio);
1845         folio_unlock(folio);
1846
1847         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1848                 pmdp_invalidate(vma, addr, pmd);
1849                 orig_pmd = pmd_mkold(orig_pmd);
1850                 orig_pmd = pmd_mkclean(orig_pmd);
1851
1852                 set_pmd_at(mm, addr, pmd, orig_pmd);
1853                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1854         }
1855
1856         folio_mark_lazyfree(folio);
1857         ret = true;
1858 out:
1859         spin_unlock(ptl);
1860 out_unlocked:
1861         return ret;
1862 }
1863
1864 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1865 {
1866         pgtable_t pgtable;
1867
1868         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1869         pte_free(mm, pgtable);
1870         mm_dec_nr_ptes(mm);
1871 }
1872
1873 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1874                  pmd_t *pmd, unsigned long addr)
1875 {
1876         pmd_t orig_pmd;
1877         spinlock_t *ptl;
1878
1879         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1880
1881         ptl = __pmd_trans_huge_lock(pmd, vma);
1882         if (!ptl)
1883                 return 0;
1884         /*
1885          * For architectures like ppc64 we look at deposited pgtable
1886          * when calling pmdp_huge_get_and_clear. So do the
1887          * pgtable_trans_huge_withdraw after finishing pmdp related
1888          * operations.
1889          */
1890         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1891                                                 tlb->fullmm);
1892         arch_check_zapped_pmd(vma, orig_pmd);
1893         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1894         if (vma_is_special_huge(vma)) {
1895                 if (arch_needs_pgtable_deposit())
1896                         zap_deposited_table(tlb->mm, pmd);
1897                 spin_unlock(ptl);
1898         } else if (is_huge_zero_pmd(orig_pmd)) {
1899                 zap_deposited_table(tlb->mm, pmd);
1900                 spin_unlock(ptl);
1901         } else {
1902                 struct page *page = NULL;
1903                 int flush_needed = 1;
1904
1905                 if (pmd_present(orig_pmd)) {
1906                         page = pmd_page(orig_pmd);
1907                         folio_remove_rmap_pmd(page_folio(page), page, vma);
1908                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1909                         VM_BUG_ON_PAGE(!PageHead(page), page);
1910                 } else if (thp_migration_supported()) {
1911                         swp_entry_t entry;
1912
1913                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1914                         entry = pmd_to_swp_entry(orig_pmd);
1915                         page = pfn_swap_entry_to_page(entry);
1916                         flush_needed = 0;
1917                 } else
1918                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1919
1920                 if (PageAnon(page)) {
1921                         zap_deposited_table(tlb->mm, pmd);
1922                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1923                 } else {
1924                         if (arch_needs_pgtable_deposit())
1925                                 zap_deposited_table(tlb->mm, pmd);
1926                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1927                 }
1928
1929                 spin_unlock(ptl);
1930                 if (flush_needed)
1931                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1932         }
1933         return 1;
1934 }
1935
1936 #ifndef pmd_move_must_withdraw
1937 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1938                                          spinlock_t *old_pmd_ptl,
1939                                          struct vm_area_struct *vma)
1940 {
1941         /*
1942          * With split pmd lock we also need to move preallocated
1943          * PTE page table if new_pmd is on different PMD page table.
1944          *
1945          * We also don't deposit and withdraw tables for file pages.
1946          */
1947         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1948 }
1949 #endif
1950
1951 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1952 {
1953 #ifdef CONFIG_MEM_SOFT_DIRTY
1954         if (unlikely(is_pmd_migration_entry(pmd)))
1955                 pmd = pmd_swp_mksoft_dirty(pmd);
1956         else if (pmd_present(pmd))
1957                 pmd = pmd_mksoft_dirty(pmd);
1958 #endif
1959         return pmd;
1960 }
1961
1962 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1963                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1964 {
1965         spinlock_t *old_ptl, *new_ptl;
1966         pmd_t pmd;
1967         struct mm_struct *mm = vma->vm_mm;
1968         bool force_flush = false;
1969
1970         /*
1971          * The destination pmd shouldn't be established, free_pgtables()
1972          * should have released it; but move_page_tables() might have already
1973          * inserted a page table, if racing against shmem/file collapse.
1974          */
1975         if (!pmd_none(*new_pmd)) {
1976                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1977                 return false;
1978         }
1979
1980         /*
1981          * We don't have to worry about the ordering of src and dst
1982          * ptlocks because exclusive mmap_lock prevents deadlock.
1983          */
1984         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1985         if (old_ptl) {
1986                 new_ptl = pmd_lockptr(mm, new_pmd);
1987                 if (new_ptl != old_ptl)
1988                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1989                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1990                 if (pmd_present(pmd))
1991                         force_flush = true;
1992                 VM_BUG_ON(!pmd_none(*new_pmd));
1993
1994                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1995                         pgtable_t pgtable;
1996                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1997                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1998                 }
1999                 pmd = move_soft_dirty_pmd(pmd);
2000                 set_pmd_at(mm, new_addr, new_pmd, pmd);
2001                 if (force_flush)
2002                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
2003                 if (new_ptl != old_ptl)
2004                         spin_unlock(new_ptl);
2005                 spin_unlock(old_ptl);
2006                 return true;
2007         }
2008         return false;
2009 }
2010
2011 /*
2012  * Returns
2013  *  - 0 if PMD could not be locked
2014  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
2015  *      or if prot_numa but THP migration is not supported
2016  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
2017  */
2018 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2019                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
2020                     unsigned long cp_flags)
2021 {
2022         struct mm_struct *mm = vma->vm_mm;
2023         spinlock_t *ptl;
2024         pmd_t oldpmd, entry;
2025         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
2026         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
2027         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
2028         int ret = 1;
2029
2030         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2031
2032         if (prot_numa && !thp_migration_supported())
2033                 return 1;
2034
2035         ptl = __pmd_trans_huge_lock(pmd, vma);
2036         if (!ptl)
2037                 return 0;
2038
2039 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2040         if (is_swap_pmd(*pmd)) {
2041                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2042                 struct folio *folio = page_folio(pfn_swap_entry_to_page(entry));
2043                 pmd_t newpmd;
2044
2045                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2046                 if (is_writable_migration_entry(entry)) {
2047                         /*
2048                          * A protection check is difficult so
2049                          * just be safe and disable write
2050                          */
2051                         if (folio_test_anon(folio))
2052                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
2053                         else
2054                                 entry = make_readable_migration_entry(swp_offset(entry));
2055                         newpmd = swp_entry_to_pmd(entry);
2056                         if (pmd_swp_soft_dirty(*pmd))
2057                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
2058                 } else {
2059                         newpmd = *pmd;
2060                 }
2061
2062                 if (uffd_wp)
2063                         newpmd = pmd_swp_mkuffd_wp(newpmd);
2064                 else if (uffd_wp_resolve)
2065                         newpmd = pmd_swp_clear_uffd_wp(newpmd);
2066                 if (!pmd_same(*pmd, newpmd))
2067                         set_pmd_at(mm, addr, pmd, newpmd);
2068                 goto unlock;
2069         }
2070 #endif
2071
2072         if (prot_numa) {
2073                 struct folio *folio;
2074                 bool toptier;
2075                 /*
2076                  * Avoid trapping faults against the zero page. The read-only
2077                  * data is likely to be read-cached on the local CPU and
2078                  * local/remote hits to the zero page are not interesting.
2079                  */
2080                 if (is_huge_zero_pmd(*pmd))
2081                         goto unlock;
2082
2083                 if (pmd_protnone(*pmd))
2084                         goto unlock;
2085
2086                 folio = page_folio(pmd_page(*pmd));
2087                 toptier = node_is_toptier(folio_nid(folio));
2088                 /*
2089                  * Skip scanning top tier node if normal numa
2090                  * balancing is disabled
2091                  */
2092                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
2093                     toptier)
2094                         goto unlock;
2095
2096                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
2097                     !toptier)
2098                         folio_xchg_access_time(folio,
2099                                                jiffies_to_msecs(jiffies));
2100         }
2101         /*
2102          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
2103          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2104          * which is also under mmap_read_lock(mm):
2105          *
2106          *      CPU0:                           CPU1:
2107          *                              change_huge_pmd(prot_numa=1)
2108          *                               pmdp_huge_get_and_clear_notify()
2109          * madvise_dontneed()
2110          *  zap_pmd_range()
2111          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2112          *   // skip the pmd
2113          *                               set_pmd_at();
2114          *                               // pmd is re-established
2115          *
2116          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2117          * which may break userspace.
2118          *
2119          * pmdp_invalidate_ad() is required to make sure we don't miss
2120          * dirty/young flags set by hardware.
2121          */
2122         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
2123
2124         entry = pmd_modify(oldpmd, newprot);
2125         if (uffd_wp)
2126                 entry = pmd_mkuffd_wp(entry);
2127         else if (uffd_wp_resolve)
2128                 /*
2129                  * Leave the write bit to be handled by PF interrupt
2130                  * handler, then things like COW could be properly
2131                  * handled.
2132                  */
2133                 entry = pmd_clear_uffd_wp(entry);
2134
2135         /* See change_pte_range(). */
2136         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
2137             can_change_pmd_writable(vma, addr, entry))
2138                 entry = pmd_mkwrite(entry, vma);
2139
2140         ret = HPAGE_PMD_NR;
2141         set_pmd_at(mm, addr, pmd, entry);
2142
2143         if (huge_pmd_needs_flush(oldpmd, entry))
2144                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
2145 unlock:
2146         spin_unlock(ptl);
2147         return ret;
2148 }
2149
2150 #ifdef CONFIG_USERFAULTFD
2151 /*
2152  * The PT lock for src_pmd and the mmap_lock for reading are held by
2153  * the caller, but it must return after releasing the page_table_lock.
2154  * Just move the page from src_pmd to dst_pmd if possible.
2155  * Return zero if succeeded in moving the page, -EAGAIN if it needs to be
2156  * repeated by the caller, or other errors in case of failure.
2157  */
2158 int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,
2159                         struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
2160                         unsigned long dst_addr, unsigned long src_addr)
2161 {
2162         pmd_t _dst_pmd, src_pmdval;
2163         struct page *src_page;
2164         struct folio *src_folio;
2165         struct anon_vma *src_anon_vma;
2166         spinlock_t *src_ptl, *dst_ptl;
2167         pgtable_t src_pgtable;
2168         struct mmu_notifier_range range;
2169         int err = 0;
2170
2171         src_pmdval = *src_pmd;
2172         src_ptl = pmd_lockptr(mm, src_pmd);
2173
2174         lockdep_assert_held(src_ptl);
2175         mmap_assert_locked(mm);
2176
2177         /* Sanity checks before the operation */
2178         if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||
2179             WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {
2180                 spin_unlock(src_ptl);
2181                 return -EINVAL;
2182         }
2183
2184         if (!pmd_trans_huge(src_pmdval)) {
2185                 spin_unlock(src_ptl);
2186                 if (is_pmd_migration_entry(src_pmdval)) {
2187                         pmd_migration_entry_wait(mm, &src_pmdval);
2188                         return -EAGAIN;
2189                 }
2190                 return -ENOENT;
2191         }
2192
2193         src_page = pmd_page(src_pmdval);
2194         if (unlikely(!PageAnonExclusive(src_page))) {
2195                 spin_unlock(src_ptl);
2196                 return -EBUSY;
2197         }
2198
2199         src_folio = page_folio(src_page);
2200         folio_get(src_folio);
2201         spin_unlock(src_ptl);
2202
2203         flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);
2204         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,
2205                                 src_addr + HPAGE_PMD_SIZE);
2206         mmu_notifier_invalidate_range_start(&range);
2207
2208         folio_lock(src_folio);
2209
2210         /*
2211          * split_huge_page walks the anon_vma chain without the page
2212          * lock. Serialize against it with the anon_vma lock, the page
2213          * lock is not enough.
2214          */
2215         src_anon_vma = folio_get_anon_vma(src_folio);
2216         if (!src_anon_vma) {
2217                 err = -EAGAIN;
2218                 goto unlock_folio;
2219         }
2220         anon_vma_lock_write(src_anon_vma);
2221
2222         dst_ptl = pmd_lockptr(mm, dst_pmd);
2223         double_pt_lock(src_ptl, dst_ptl);
2224         if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||
2225                      !pmd_same(*dst_pmd, dst_pmdval))) {
2226                 err = -EAGAIN;
2227                 goto unlock_ptls;
2228         }
2229         if (folio_maybe_dma_pinned(src_folio) ||
2230             !PageAnonExclusive(&src_folio->page)) {
2231                 err = -EBUSY;
2232                 goto unlock_ptls;
2233         }
2234
2235         if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||
2236             WARN_ON_ONCE(!folio_test_anon(src_folio))) {
2237                 err = -EBUSY;
2238                 goto unlock_ptls;
2239         }
2240
2241         folio_move_anon_rmap(src_folio, dst_vma);
2242         WRITE_ONCE(src_folio->index, linear_page_index(dst_vma, dst_addr));
2243
2244         src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2245         /* Folio got pinned from under us. Put it back and fail the move. */
2246         if (folio_maybe_dma_pinned(src_folio)) {
2247                 set_pmd_at(mm, src_addr, src_pmd, src_pmdval);
2248                 err = -EBUSY;
2249                 goto unlock_ptls;
2250         }
2251
2252         _dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot);
2253         /* Follow mremap() behavior and treat the entry dirty after the move */
2254         _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);
2255         set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);
2256
2257         src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);
2258         pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);
2259 unlock_ptls:
2260         double_pt_unlock(src_ptl, dst_ptl);
2261         anon_vma_unlock_write(src_anon_vma);
2262         put_anon_vma(src_anon_vma);
2263 unlock_folio:
2264         /* unblock rmap walks */
2265         folio_unlock(src_folio);
2266         mmu_notifier_invalidate_range_end(&range);
2267         folio_put(src_folio);
2268         return err;
2269 }
2270 #endif /* CONFIG_USERFAULTFD */
2271
2272 /*
2273  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2274  *
2275  * Note that if it returns page table lock pointer, this routine returns without
2276  * unlocking page table lock. So callers must unlock it.
2277  */
2278 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2279 {
2280         spinlock_t *ptl;
2281         ptl = pmd_lock(vma->vm_mm, pmd);
2282         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2283                         pmd_devmap(*pmd)))
2284                 return ptl;
2285         spin_unlock(ptl);
2286         return NULL;
2287 }
2288
2289 /*
2290  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
2291  *
2292  * Note that if it returns page table lock pointer, this routine returns without
2293  * unlocking page table lock. So callers must unlock it.
2294  */
2295 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2296 {
2297         spinlock_t *ptl;
2298
2299         ptl = pud_lock(vma->vm_mm, pud);
2300         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2301                 return ptl;
2302         spin_unlock(ptl);
2303         return NULL;
2304 }
2305
2306 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2307 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2308                  pud_t *pud, unsigned long addr)
2309 {
2310         spinlock_t *ptl;
2311
2312         ptl = __pud_trans_huge_lock(pud, vma);
2313         if (!ptl)
2314                 return 0;
2315
2316         pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2317         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2318         if (vma_is_special_huge(vma)) {
2319                 spin_unlock(ptl);
2320                 /* No zero page support yet */
2321         } else {
2322                 /* No support for anonymous PUD pages yet */
2323                 BUG();
2324         }
2325         return 1;
2326 }
2327
2328 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2329                 unsigned long haddr)
2330 {
2331         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2332         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2333         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2334         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2335
2336         count_vm_event(THP_SPLIT_PUD);
2337
2338         pudp_huge_clear_flush(vma, haddr, pud);
2339 }
2340
2341 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2342                 unsigned long address)
2343 {
2344         spinlock_t *ptl;
2345         struct mmu_notifier_range range;
2346
2347         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2348                                 address & HPAGE_PUD_MASK,
2349                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2350         mmu_notifier_invalidate_range_start(&range);
2351         ptl = pud_lock(vma->vm_mm, pud);
2352         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2353                 goto out;
2354         __split_huge_pud_locked(vma, pud, range.start);
2355
2356 out:
2357         spin_unlock(ptl);
2358         mmu_notifier_invalidate_range_end(&range);
2359 }
2360 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2361
2362 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2363                 unsigned long haddr, pmd_t *pmd)
2364 {
2365         struct mm_struct *mm = vma->vm_mm;
2366         pgtable_t pgtable;
2367         pmd_t _pmd, old_pmd;
2368         unsigned long addr;
2369         pte_t *pte;
2370         int i;
2371
2372         /*
2373          * Leave pmd empty until pte is filled note that it is fine to delay
2374          * notification until mmu_notifier_invalidate_range_end() as we are
2375          * replacing a zero pmd write protected page with a zero pte write
2376          * protected page.
2377          *
2378          * See Documentation/mm/mmu_notifier.rst
2379          */
2380         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2381
2382         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2383         pmd_populate(mm, &_pmd, pgtable);
2384
2385         pte = pte_offset_map(&_pmd, haddr);
2386         VM_BUG_ON(!pte);
2387         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2388                 pte_t entry;
2389
2390                 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2391                 entry = pte_mkspecial(entry);
2392                 if (pmd_uffd_wp(old_pmd))
2393                         entry = pte_mkuffd_wp(entry);
2394                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2395                 set_pte_at(mm, addr, pte, entry);
2396                 pte++;
2397         }
2398         pte_unmap(pte - 1);
2399         smp_wmb(); /* make pte visible before pmd */
2400         pmd_populate(mm, pmd, pgtable);
2401 }
2402
2403 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2404                 unsigned long haddr, bool freeze)
2405 {
2406         struct mm_struct *mm = vma->vm_mm;
2407         struct folio *folio;
2408         struct page *page;
2409         pgtable_t pgtable;
2410         pmd_t old_pmd, _pmd;
2411         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2412         bool anon_exclusive = false, dirty = false;
2413         unsigned long addr;
2414         pte_t *pte;
2415         int i;
2416
2417         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2418         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2419         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2420         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2421                                 && !pmd_devmap(*pmd));
2422
2423         count_vm_event(THP_SPLIT_PMD);
2424
2425         if (!vma_is_anonymous(vma)) {
2426                 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2427                 /*
2428                  * We are going to unmap this huge page. So
2429                  * just go ahead and zap it
2430                  */
2431                 if (arch_needs_pgtable_deposit())
2432                         zap_deposited_table(mm, pmd);
2433                 if (vma_is_special_huge(vma))
2434                         return;
2435                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2436                         swp_entry_t entry;
2437
2438                         entry = pmd_to_swp_entry(old_pmd);
2439                         page = pfn_swap_entry_to_page(entry);
2440                 } else {
2441                         page = pmd_page(old_pmd);
2442                         folio = page_folio(page);
2443                         if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))
2444                                 folio_mark_dirty(folio);
2445                         if (!folio_test_referenced(folio) && pmd_young(old_pmd))
2446                                 folio_set_referenced(folio);
2447                         folio_remove_rmap_pmd(folio, page, vma);
2448                         folio_put(folio);
2449                 }
2450                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2451                 return;
2452         }
2453
2454         if (is_huge_zero_pmd(*pmd)) {
2455                 /*
2456                  * FIXME: Do we want to invalidate secondary mmu by calling
2457                  * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2458                  * inside __split_huge_pmd() ?
2459                  *
2460                  * We are going from a zero huge page write protected to zero
2461                  * small page also write protected so it does not seems useful
2462                  * to invalidate secondary mmu at this time.
2463                  */
2464                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2465         }
2466
2467         /*
2468          * Up to this point the pmd is present and huge and userland has the
2469          * whole access to the hugepage during the split (which happens in
2470          * place). If we overwrite the pmd with the not-huge version pointing
2471          * to the pte here (which of course we could if all CPUs were bug
2472          * free), userland could trigger a small page size TLB miss on the
2473          * small sized TLB while the hugepage TLB entry is still established in
2474          * the huge TLB. Some CPU doesn't like that.
2475          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2476          * 383 on page 105. Intel should be safe but is also warns that it's
2477          * only safe if the permission and cache attributes of the two entries
2478          * loaded in the two TLB is identical (which should be the case here).
2479          * But it is generally safer to never allow small and huge TLB entries
2480          * for the same virtual address to be loaded simultaneously. So instead
2481          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2482          * current pmd notpresent (atomically because here the pmd_trans_huge
2483          * must remain set at all times on the pmd until the split is complete
2484          * for this pmd), then we flush the SMP TLB and finally we write the
2485          * non-huge version of the pmd entry with pmd_populate.
2486          */
2487         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2488
2489         pmd_migration = is_pmd_migration_entry(old_pmd);
2490         if (unlikely(pmd_migration)) {
2491                 swp_entry_t entry;
2492
2493                 entry = pmd_to_swp_entry(old_pmd);
2494                 page = pfn_swap_entry_to_page(entry);
2495                 write = is_writable_migration_entry(entry);
2496                 if (PageAnon(page))
2497                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2498                 young = is_migration_entry_young(entry);
2499                 dirty = is_migration_entry_dirty(entry);
2500                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2501                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2502         } else {
2503                 page = pmd_page(old_pmd);
2504                 folio = page_folio(page);
2505                 if (pmd_dirty(old_pmd)) {
2506                         dirty = true;
2507                         folio_set_dirty(folio);
2508                 }
2509                 write = pmd_write(old_pmd);
2510                 young = pmd_young(old_pmd);
2511                 soft_dirty = pmd_soft_dirty(old_pmd);
2512                 uffd_wp = pmd_uffd_wp(old_pmd);
2513
2514                 VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);
2515                 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2516
2517                 /*
2518                  * Without "freeze", we'll simply split the PMD, propagating the
2519                  * PageAnonExclusive() flag for each PTE by setting it for
2520                  * each subpage -- no need to (temporarily) clear.
2521                  *
2522                  * With "freeze" we want to replace mapped pages by
2523                  * migration entries right away. This is only possible if we
2524                  * managed to clear PageAnonExclusive() -- see
2525                  * set_pmd_migration_entry().
2526                  *
2527                  * In case we cannot clear PageAnonExclusive(), split the PMD
2528                  * only and let try_to_migrate_one() fail later.
2529                  *
2530                  * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.
2531                  */
2532                 anon_exclusive = PageAnonExclusive(page);
2533                 if (freeze && anon_exclusive &&
2534                     folio_try_share_anon_rmap_pmd(folio, page))
2535                         freeze = false;
2536                 if (!freeze) {
2537                         rmap_t rmap_flags = RMAP_NONE;
2538
2539                         folio_ref_add(folio, HPAGE_PMD_NR - 1);
2540                         if (anon_exclusive)
2541                                 rmap_flags |= RMAP_EXCLUSIVE;
2542                         folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
2543                                                  vma, haddr, rmap_flags);
2544                 }
2545         }
2546
2547         /*
2548          * Withdraw the table only after we mark the pmd entry invalid.
2549          * This's critical for some architectures (Power).
2550          */
2551         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2552         pmd_populate(mm, &_pmd, pgtable);
2553
2554         pte = pte_offset_map(&_pmd, haddr);
2555         VM_BUG_ON(!pte);
2556         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2557                 pte_t entry;
2558                 /*
2559                  * Note that NUMA hinting access restrictions are not
2560                  * transferred to avoid any possibility of altering
2561                  * permissions across VMAs.
2562                  */
2563                 if (freeze || pmd_migration) {
2564                         swp_entry_t swp_entry;
2565                         if (write)
2566                                 swp_entry = make_writable_migration_entry(
2567                                                         page_to_pfn(page + i));
2568                         else if (anon_exclusive)
2569                                 swp_entry = make_readable_exclusive_migration_entry(
2570                                                         page_to_pfn(page + i));
2571                         else
2572                                 swp_entry = make_readable_migration_entry(
2573                                                         page_to_pfn(page + i));
2574                         if (young)
2575                                 swp_entry = make_migration_entry_young(swp_entry);
2576                         if (dirty)
2577                                 swp_entry = make_migration_entry_dirty(swp_entry);
2578                         entry = swp_entry_to_pte(swp_entry);
2579                         if (soft_dirty)
2580                                 entry = pte_swp_mksoft_dirty(entry);
2581                         if (uffd_wp)
2582                                 entry = pte_swp_mkuffd_wp(entry);
2583                 } else {
2584                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2585                         if (write)
2586                                 entry = pte_mkwrite(entry, vma);
2587                         if (!young)
2588                                 entry = pte_mkold(entry);
2589                         /* NOTE: this may set soft-dirty too on some archs */
2590                         if (dirty)
2591                                 entry = pte_mkdirty(entry);
2592                         if (soft_dirty)
2593                                 entry = pte_mksoft_dirty(entry);
2594                         if (uffd_wp)
2595                                 entry = pte_mkuffd_wp(entry);
2596                 }
2597                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2598                 set_pte_at(mm, addr, pte, entry);
2599                 pte++;
2600         }
2601         pte_unmap(pte - 1);
2602
2603         if (!pmd_migration)
2604                 folio_remove_rmap_pmd(folio, page, vma);
2605         if (freeze)
2606                 put_page(page);
2607
2608         smp_wmb(); /* make pte visible before pmd */
2609         pmd_populate(mm, pmd, pgtable);
2610 }
2611
2612 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2613                 unsigned long address, bool freeze, struct folio *folio)
2614 {
2615         spinlock_t *ptl;
2616         struct mmu_notifier_range range;
2617
2618         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2619                                 address & HPAGE_PMD_MASK,
2620                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2621         mmu_notifier_invalidate_range_start(&range);
2622         ptl = pmd_lock(vma->vm_mm, pmd);
2623
2624         /*
2625          * If caller asks to setup a migration entry, we need a folio to check
2626          * pmd against. Otherwise we can end up replacing wrong folio.
2627          */
2628         VM_BUG_ON(freeze && !folio);
2629         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2630
2631         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2632             is_pmd_migration_entry(*pmd)) {
2633                 /*
2634                  * It's safe to call pmd_page when folio is set because it's
2635                  * guaranteed that pmd is present.
2636                  */
2637                 if (folio && folio != page_folio(pmd_page(*pmd)))
2638                         goto out;
2639                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2640         }
2641
2642 out:
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_SPLIT_HUGE_PMD |
2696                 TTU_SYNC | TTU_BATCH_FLUSH;
2697
2698         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2699
2700         /*
2701          * Anon pages need migration entries to preserve them, but file
2702          * pages can simply be left unmapped, then faulted back on demand.
2703          * If that is ever changed (perhaps for mlock), update remap_page().
2704          */
2705         if (folio_test_anon(folio))
2706                 try_to_migrate(folio, ttu_flags);
2707         else
2708                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2709
2710         try_to_unmap_flush();
2711 }
2712
2713 static void remap_page(struct folio *folio, unsigned long nr)
2714 {
2715         int i = 0;
2716
2717         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2718         if (!folio_test_anon(folio))
2719                 return;
2720         for (;;) {
2721                 remove_migration_ptes(folio, folio, true);
2722                 i += folio_nr_pages(folio);
2723                 if (i >= nr)
2724                         break;
2725                 folio = folio_next(folio);
2726         }
2727 }
2728
2729 static void lru_add_page_tail(struct page *head, struct page *tail,
2730                 struct lruvec *lruvec, struct list_head *list)
2731 {
2732         VM_BUG_ON_PAGE(!PageHead(head), head);
2733         VM_BUG_ON_PAGE(PageCompound(tail), head);
2734         VM_BUG_ON_PAGE(PageLRU(tail), head);
2735         lockdep_assert_held(&lruvec->lru_lock);
2736
2737         if (list) {
2738                 /* page reclaim is reclaiming a huge page */
2739                 VM_WARN_ON(PageLRU(head));
2740                 get_page(tail);
2741                 list_add_tail(&tail->lru, list);
2742         } else {
2743                 /* head is still on lru (and we have it frozen) */
2744                 VM_WARN_ON(!PageLRU(head));
2745                 if (PageUnevictable(tail))
2746                         tail->mlock_count = 0;
2747                 else
2748                         list_add_tail(&tail->lru, &head->lru);
2749                 SetPageLRU(tail);
2750         }
2751 }
2752
2753 static void __split_huge_page_tail(struct folio *folio, int tail,
2754                 struct lruvec *lruvec, struct list_head *list)
2755 {
2756         struct page *head = &folio->page;
2757         struct page *page_tail = head + tail;
2758         /*
2759          * Careful: new_folio is not a "real" folio before we cleared PageTail.
2760          * Don't pass it around before clear_compound_head().
2761          */
2762         struct folio *new_folio = (struct folio *)page_tail;
2763
2764         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2765
2766         /*
2767          * Clone page flags before unfreezing refcount.
2768          *
2769          * After successful get_page_unless_zero() might follow flags change,
2770          * for example lock_page() which set PG_waiters.
2771          *
2772          * Note that for mapped sub-pages of an anonymous THP,
2773          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2774          * the migration entry instead from where remap_page() will restore it.
2775          * We can still have PG_anon_exclusive set on effectively unmapped and
2776          * unreferenced sub-pages of an anonymous THP: we can simply drop
2777          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2778          */
2779         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2780         page_tail->flags |= (head->flags &
2781                         ((1L << PG_referenced) |
2782                          (1L << PG_swapbacked) |
2783                          (1L << PG_swapcache) |
2784                          (1L << PG_mlocked) |
2785                          (1L << PG_uptodate) |
2786                          (1L << PG_active) |
2787                          (1L << PG_workingset) |
2788                          (1L << PG_locked) |
2789                          (1L << PG_unevictable) |
2790 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2791                          (1L << PG_arch_2) |
2792                          (1L << PG_arch_3) |
2793 #endif
2794                          (1L << PG_dirty) |
2795                          LRU_GEN_MASK | LRU_REFS_MASK));
2796
2797         /* ->mapping in first and second tail page is replaced by other uses */
2798         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2799                         page_tail);
2800         page_tail->mapping = head->mapping;
2801         page_tail->index = head->index + tail;
2802
2803         /*
2804          * page->private should not be set in tail pages. Fix up and warn once
2805          * if private is unexpectedly set.
2806          */
2807         if (unlikely(page_tail->private)) {
2808                 VM_WARN_ON_ONCE_PAGE(true, page_tail);
2809                 page_tail->private = 0;
2810         }
2811         if (folio_test_swapcache(folio))
2812                 new_folio->swap.val = folio->swap.val + tail;
2813
2814         /* Page flags must be visible before we make the page non-compound. */
2815         smp_wmb();
2816
2817         /*
2818          * Clear PageTail before unfreezing page refcount.
2819          *
2820          * After successful get_page_unless_zero() might follow put_page()
2821          * which needs correct compound_head().
2822          */
2823         clear_compound_head(page_tail);
2824
2825         /* Finally unfreeze refcount. Additional reference from page cache. */
2826         page_ref_unfreeze(page_tail, 1 + (!folio_test_anon(folio) ||
2827                                           folio_test_swapcache(folio)));
2828
2829         if (folio_test_young(folio))
2830                 folio_set_young(new_folio);
2831         if (folio_test_idle(folio))
2832                 folio_set_idle(new_folio);
2833
2834         folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
2835
2836         /*
2837          * always add to the tail because some iterators expect new
2838          * pages to show after the currently processed elements - e.g.
2839          * migrate_pages
2840          */
2841         lru_add_page_tail(head, page_tail, lruvec, list);
2842 }
2843
2844 static void __split_huge_page(struct page *page, struct list_head *list,
2845                 pgoff_t end)
2846 {
2847         struct folio *folio = page_folio(page);
2848         struct page *head = &folio->page;
2849         struct lruvec *lruvec;
2850         struct address_space *swap_cache = NULL;
2851         unsigned long offset = 0;
2852         unsigned int nr = thp_nr_pages(head);
2853         int i, nr_dropped = 0;
2854
2855         /* complete memcg works before add pages to LRU */
2856         split_page_memcg(head, nr);
2857
2858         if (folio_test_anon(folio) && folio_test_swapcache(folio)) {
2859                 offset = swp_offset(folio->swap);
2860                 swap_cache = swap_address_space(folio->swap);
2861                 xa_lock(&swap_cache->i_pages);
2862         }
2863
2864         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2865         lruvec = folio_lruvec_lock(folio);
2866
2867         ClearPageHasHWPoisoned(head);
2868
2869         for (i = nr - 1; i >= 1; i--) {
2870                 __split_huge_page_tail(folio, i, lruvec, list);
2871                 /* Some pages can be beyond EOF: drop them from page cache */
2872                 if (head[i].index >= end) {
2873                         struct folio *tail = page_folio(head + i);
2874
2875                         if (shmem_mapping(head->mapping))
2876                                 nr_dropped++;
2877                         else if (folio_test_clear_dirty(tail))
2878                                 folio_account_cleaned(tail,
2879                                         inode_to_wb(folio->mapping->host));
2880                         __filemap_remove_folio(tail, NULL);
2881                         folio_put(tail);
2882                 } else if (!PageAnon(page)) {
2883                         __xa_store(&head->mapping->i_pages, head[i].index,
2884                                         head + i, 0);
2885                 } else if (swap_cache) {
2886                         __xa_store(&swap_cache->i_pages, offset + i,
2887                                         head + i, 0);
2888                 }
2889         }
2890
2891         ClearPageCompound(head);
2892         unlock_page_lruvec(lruvec);
2893         /* Caller disabled irqs, so they are still disabled here */
2894
2895         split_page_owner(head, nr);
2896
2897         /* See comment in __split_huge_page_tail() */
2898         if (PageAnon(head)) {
2899                 /* Additional pin to swap cache */
2900                 if (PageSwapCache(head)) {
2901                         page_ref_add(head, 2);
2902                         xa_unlock(&swap_cache->i_pages);
2903                 } else {
2904                         page_ref_inc(head);
2905                 }
2906         } else {
2907                 /* Additional pin to page cache */
2908                 page_ref_add(head, 2);
2909                 xa_unlock(&head->mapping->i_pages);
2910         }
2911         local_irq_enable();
2912
2913         if (nr_dropped)
2914                 shmem_uncharge(head->mapping->host, nr_dropped);
2915         remap_page(folio, nr);
2916
2917         if (folio_test_swapcache(folio))
2918                 split_swap_cluster(folio->swap);
2919
2920         for (i = 0; i < nr; i++) {
2921                 struct page *subpage = head + i;
2922                 if (subpage == page)
2923                         continue;
2924                 unlock_page(subpage);
2925
2926                 /*
2927                  * Subpages may be freed if there wasn't any mapping
2928                  * like if add_to_swap() is running on a lru page that
2929                  * had its mapping zapped. And freeing these pages
2930                  * requires taking the lru_lock so we do the put_page
2931                  * of the tail pages after the split is complete.
2932                  */
2933                 free_page_and_swap_cache(subpage);
2934         }
2935 }
2936
2937 /* Racy check whether the huge page can be split */
2938 bool can_split_folio(struct folio *folio, int *pextra_pins)
2939 {
2940         int extra_pins;
2941
2942         /* Additional pins from page cache */
2943         if (folio_test_anon(folio))
2944                 extra_pins = folio_test_swapcache(folio) ?
2945                                 folio_nr_pages(folio) : 0;
2946         else
2947                 extra_pins = folio_nr_pages(folio);
2948         if (pextra_pins)
2949                 *pextra_pins = extra_pins;
2950         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2951 }
2952
2953 /*
2954  * This function splits huge page into normal pages. @page can point to any
2955  * subpage of huge page to split. Split doesn't change the position of @page.
2956  *
2957  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2958  * The huge page must be locked.
2959  *
2960  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2961  *
2962  * Both head page and tail pages will inherit mapping, flags, and so on from
2963  * the hugepage.
2964  *
2965  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2966  * they are not mapped.
2967  *
2968  * Returns 0 if the hugepage is split successfully.
2969  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2970  * us.
2971  */
2972 int split_huge_page_to_list(struct page *page, struct list_head *list)
2973 {
2974         struct folio *folio = page_folio(page);
2975         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2976         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2977         struct anon_vma *anon_vma = NULL;
2978         struct address_space *mapping = NULL;
2979         int extra_pins, ret;
2980         pgoff_t end;
2981         bool is_hzp;
2982
2983         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2984         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2985
2986         is_hzp = is_huge_zero_page(&folio->page);
2987         if (is_hzp) {
2988                 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2989                 return -EBUSY;
2990         }
2991
2992         if (folio_test_writeback(folio))
2993                 return -EBUSY;
2994
2995         if (folio_test_anon(folio)) {
2996                 /*
2997                  * The caller does not necessarily hold an mmap_lock that would
2998                  * prevent the anon_vma disappearing so we first we take a
2999                  * reference to it and then lock the anon_vma for write. This
3000                  * is similar to folio_lock_anon_vma_read except the write lock
3001                  * is taken to serialise against parallel split or collapse
3002                  * operations.
3003                  */
3004                 anon_vma = folio_get_anon_vma(folio);
3005                 if (!anon_vma) {
3006                         ret = -EBUSY;
3007                         goto out;
3008                 }
3009                 end = -1;
3010                 mapping = NULL;
3011                 anon_vma_lock_write(anon_vma);
3012         } else {
3013                 gfp_t gfp;
3014
3015                 mapping = folio->mapping;
3016
3017                 /* Truncated ? */
3018                 if (!mapping) {
3019                         ret = -EBUSY;
3020                         goto out;
3021                 }
3022
3023                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
3024                                                         GFP_RECLAIM_MASK);
3025
3026                 if (!filemap_release_folio(folio, gfp)) {
3027                         ret = -EBUSY;
3028                         goto out;
3029                 }
3030
3031                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
3032                 if (xas_error(&xas)) {
3033                         ret = xas_error(&xas);
3034                         goto out;
3035                 }
3036
3037                 anon_vma = NULL;
3038                 i_mmap_lock_read(mapping);
3039
3040                 /*
3041                  *__split_huge_page() may need to trim off pages beyond EOF:
3042                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
3043                  * which cannot be nested inside the page tree lock. So note
3044                  * end now: i_size itself may be changed at any moment, but
3045                  * folio lock is good enough to serialize the trimming.
3046                  */
3047                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3048                 if (shmem_mapping(mapping))
3049                         end = shmem_fallocend(mapping->host, end);
3050         }
3051
3052         /*
3053          * Racy check if we can split the page, before unmap_folio() will
3054          * split PMDs
3055          */
3056         if (!can_split_folio(folio, &extra_pins)) {
3057                 ret = -EAGAIN;
3058                 goto out_unlock;
3059         }
3060
3061         unmap_folio(folio);
3062
3063         /* block interrupt reentry in xa_lock and spinlock */
3064         local_irq_disable();
3065         if (mapping) {
3066                 /*
3067                  * Check if the folio is present in page cache.
3068                  * We assume all tail are present too, if folio is there.
3069                  */
3070                 xas_lock(&xas);
3071                 xas_reset(&xas);
3072                 if (xas_load(&xas) != folio)
3073                         goto fail;
3074         }
3075
3076         /* Prevent deferred_split_scan() touching ->_refcount */
3077         spin_lock(&ds_queue->split_queue_lock);
3078         if (folio_ref_freeze(folio, 1 + extra_pins)) {
3079                 if (!list_empty(&folio->_deferred_list)) {
3080                         ds_queue->split_queue_len--;
3081                         list_del(&folio->_deferred_list);
3082                 }
3083                 spin_unlock(&ds_queue->split_queue_lock);
3084                 if (mapping) {
3085                         int nr = folio_nr_pages(folio);
3086
3087                         xas_split(&xas, folio, folio_order(folio));
3088                         if (folio_test_pmd_mappable(folio)) {
3089                                 if (folio_test_swapbacked(folio)) {
3090                                         __lruvec_stat_mod_folio(folio,
3091                                                         NR_SHMEM_THPS, -nr);
3092                                 } else {
3093                                         __lruvec_stat_mod_folio(folio,
3094                                                         NR_FILE_THPS, -nr);
3095                                         filemap_nr_thps_dec(mapping);
3096                                 }
3097                         }
3098                 }
3099
3100                 __split_huge_page(page, list, end);
3101                 ret = 0;
3102         } else {
3103                 spin_unlock(&ds_queue->split_queue_lock);
3104 fail:
3105                 if (mapping)
3106                         xas_unlock(&xas);
3107                 local_irq_enable();
3108                 remap_page(folio, folio_nr_pages(folio));
3109                 ret = -EAGAIN;
3110         }
3111
3112 out_unlock:
3113         if (anon_vma) {
3114                 anon_vma_unlock_write(anon_vma);
3115                 put_anon_vma(anon_vma);
3116         }
3117         if (mapping)
3118                 i_mmap_unlock_read(mapping);
3119 out:
3120         xas_destroy(&xas);
3121         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3122         return ret;
3123 }
3124
3125 void folio_undo_large_rmappable(struct folio *folio)
3126 {
3127         struct deferred_split *ds_queue;
3128         unsigned long flags;
3129
3130         /*
3131          * At this point, there is no one trying to add the folio to
3132          * deferred_list. If folio is not in deferred_list, it's safe
3133          * to check without acquiring the split_queue_lock.
3134          */
3135         if (data_race(list_empty(&folio->_deferred_list)))
3136                 return;
3137
3138         ds_queue = get_deferred_split_queue(folio);
3139         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3140         if (!list_empty(&folio->_deferred_list)) {
3141                 ds_queue->split_queue_len--;
3142                 list_del_init(&folio->_deferred_list);
3143         }
3144         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3145 }
3146
3147 void deferred_split_folio(struct folio *folio)
3148 {
3149         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3150 #ifdef CONFIG_MEMCG
3151         struct mem_cgroup *memcg = folio_memcg(folio);
3152 #endif
3153         unsigned long flags;
3154
3155         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
3156
3157         /*
3158          * The try_to_unmap() in page reclaim path might reach here too,
3159          * this may cause a race condition to corrupt deferred split queue.
3160          * And, if page reclaim is already handling the same folio, it is
3161          * unnecessary to handle it again in shrinker.
3162          *
3163          * Check the swapcache flag to determine if the folio is being
3164          * handled by page reclaim since THP swap would add the folio into
3165          * swap cache before calling try_to_unmap().
3166          */
3167         if (folio_test_swapcache(folio))
3168                 return;
3169
3170         if (!list_empty(&folio->_deferred_list))
3171                 return;
3172
3173         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3174         if (list_empty(&folio->_deferred_list)) {
3175                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3176                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
3177                 ds_queue->split_queue_len++;
3178 #ifdef CONFIG_MEMCG
3179                 if (memcg)
3180                         set_shrinker_bit(memcg, folio_nid(folio),
3181                                          deferred_split_shrinker->id);
3182 #endif
3183         }
3184         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3185 }
3186
3187 static unsigned long deferred_split_count(struct shrinker *shrink,
3188                 struct shrink_control *sc)
3189 {
3190         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3191         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3192
3193 #ifdef CONFIG_MEMCG
3194         if (sc->memcg)
3195                 ds_queue = &sc->memcg->deferred_split_queue;
3196 #endif
3197         return READ_ONCE(ds_queue->split_queue_len);
3198 }
3199
3200 static unsigned long deferred_split_scan(struct shrinker *shrink,
3201                 struct shrink_control *sc)
3202 {
3203         struct pglist_data *pgdata = NODE_DATA(sc->nid);
3204         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3205         unsigned long flags;
3206         LIST_HEAD(list);
3207         struct folio *folio, *next;
3208         int split = 0;
3209
3210 #ifdef CONFIG_MEMCG
3211         if (sc->memcg)
3212                 ds_queue = &sc->memcg->deferred_split_queue;
3213 #endif
3214
3215         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3216         /* Take pin on all head pages to avoid freeing them under us */
3217         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
3218                                                         _deferred_list) {
3219                 if (folio_try_get(folio)) {
3220                         list_move(&folio->_deferred_list, &list);
3221                 } else {
3222                         /* We lost race with folio_put() */
3223                         list_del_init(&folio->_deferred_list);
3224                         ds_queue->split_queue_len--;
3225                 }
3226                 if (!--sc->nr_to_scan)
3227                         break;
3228         }
3229         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3230
3231         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
3232                 if (!folio_trylock(folio))
3233                         goto next;
3234                 /* split_huge_page() removes page from list on success */
3235                 if (!split_folio(folio))
3236                         split++;
3237                 folio_unlock(folio);
3238 next:
3239                 folio_put(folio);
3240         }
3241
3242         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3243         list_splice_tail(&list, &ds_queue->split_queue);
3244         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3245
3246         /*
3247          * Stop shrinker if we didn't split any page, but the queue is empty.
3248          * This can happen if pages were freed under us.
3249          */
3250         if (!split && list_empty(&ds_queue->split_queue))
3251                 return SHRINK_STOP;
3252         return split;
3253 }
3254
3255 #ifdef CONFIG_DEBUG_FS
3256 static void split_huge_pages_all(void)
3257 {
3258         struct zone *zone;
3259         struct page *page;
3260         struct folio *folio;
3261         unsigned long pfn, max_zone_pfn;
3262         unsigned long total = 0, split = 0;
3263
3264         pr_debug("Split all THPs\n");
3265         for_each_zone(zone) {
3266                 if (!managed_zone(zone))
3267                         continue;
3268                 max_zone_pfn = zone_end_pfn(zone);
3269                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3270                         int nr_pages;
3271
3272                         page = pfn_to_online_page(pfn);
3273                         if (!page || PageTail(page))
3274                                 continue;
3275                         folio = page_folio(page);
3276                         if (!folio_try_get(folio))
3277                                 continue;
3278
3279                         if (unlikely(page_folio(page) != folio))
3280                                 goto next;
3281
3282                         if (zone != folio_zone(folio))
3283                                 goto next;
3284
3285                         if (!folio_test_large(folio)
3286                                 || folio_test_hugetlb(folio)
3287                                 || !folio_test_lru(folio))
3288                                 goto next;
3289
3290                         total++;
3291                         folio_lock(folio);
3292                         nr_pages = folio_nr_pages(folio);
3293                         if (!split_folio(folio))
3294                                 split++;
3295                         pfn += nr_pages - 1;
3296                         folio_unlock(folio);
3297 next:
3298                         folio_put(folio);
3299                         cond_resched();
3300                 }
3301         }
3302
3303         pr_debug("%lu of %lu THP split\n", split, total);
3304 }
3305
3306 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
3307 {
3308         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
3309                     is_vm_hugetlb_page(vma);
3310 }
3311
3312 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
3313                                 unsigned long vaddr_end)
3314 {
3315         int ret = 0;
3316         struct task_struct *task;
3317         struct mm_struct *mm;
3318         unsigned long total = 0, split = 0;
3319         unsigned long addr;
3320
3321         vaddr_start &= PAGE_MASK;
3322         vaddr_end &= PAGE_MASK;
3323
3324         /* Find the task_struct from pid */
3325         rcu_read_lock();
3326         task = find_task_by_vpid(pid);
3327         if (!task) {
3328                 rcu_read_unlock();
3329                 ret = -ESRCH;
3330                 goto out;
3331         }
3332         get_task_struct(task);
3333         rcu_read_unlock();
3334
3335         /* Find the mm_struct */
3336         mm = get_task_mm(task);
3337         put_task_struct(task);
3338
3339         if (!mm) {
3340                 ret = -EINVAL;
3341                 goto out;
3342         }
3343
3344         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3345                  pid, vaddr_start, vaddr_end);
3346
3347         mmap_read_lock(mm);
3348         /*
3349          * always increase addr by PAGE_SIZE, since we could have a PTE page
3350          * table filled with PTE-mapped THPs, each of which is distinct.
3351          */
3352         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3353                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3354                 struct page *page;
3355                 struct folio *folio;
3356
3357                 if (!vma)
3358                         break;
3359
3360                 /* skip special VMA and hugetlb VMA */
3361                 if (vma_not_suitable_for_thp_split(vma)) {
3362                         addr = vma->vm_end;
3363                         continue;
3364                 }
3365
3366                 /* FOLL_DUMP to ignore special (like zero) pages */
3367                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3368
3369                 if (IS_ERR_OR_NULL(page))
3370                         continue;
3371
3372                 folio = page_folio(page);
3373                 if (!is_transparent_hugepage(folio))
3374                         goto next;
3375
3376                 total++;
3377                 if (!can_split_folio(folio, NULL))
3378                         goto next;
3379
3380                 if (!folio_trylock(folio))
3381                         goto next;
3382
3383                 if (!split_folio(folio))
3384                         split++;
3385
3386                 folio_unlock(folio);
3387 next:
3388                 folio_put(folio);
3389                 cond_resched();
3390         }
3391         mmap_read_unlock(mm);
3392         mmput(mm);
3393
3394         pr_debug("%lu of %lu THP split\n", split, total);
3395
3396 out:
3397         return ret;
3398 }
3399
3400 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3401                                 pgoff_t off_end)
3402 {
3403         struct filename *file;
3404         struct file *candidate;
3405         struct address_space *mapping;
3406         int ret = -EINVAL;
3407         pgoff_t index;
3408         int nr_pages = 1;
3409         unsigned long total = 0, split = 0;
3410
3411         file = getname_kernel(file_path);
3412         if (IS_ERR(file))
3413                 return ret;
3414
3415         candidate = file_open_name(file, O_RDONLY, 0);
3416         if (IS_ERR(candidate))
3417                 goto out;
3418
3419         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3420                  file_path, off_start, off_end);
3421
3422         mapping = candidate->f_mapping;
3423
3424         for (index = off_start; index < off_end; index += nr_pages) {
3425                 struct folio *folio = filemap_get_folio(mapping, index);
3426
3427                 nr_pages = 1;
3428                 if (IS_ERR(folio))
3429                         continue;
3430
3431                 if (!folio_test_large(folio))
3432                         goto next;
3433
3434                 total++;
3435                 nr_pages = folio_nr_pages(folio);
3436
3437                 if (!folio_trylock(folio))
3438                         goto next;
3439
3440                 if (!split_folio(folio))
3441                         split++;
3442
3443                 folio_unlock(folio);
3444 next:
3445                 folio_put(folio);
3446                 cond_resched();
3447         }
3448
3449         filp_close(candidate, NULL);
3450         ret = 0;
3451
3452         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3453 out:
3454         putname(file);
3455         return ret;
3456 }
3457
3458 #define MAX_INPUT_BUF_SZ 255
3459
3460 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3461                                 size_t count, loff_t *ppops)
3462 {
3463         static DEFINE_MUTEX(split_debug_mutex);
3464         ssize_t ret;
3465         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3466         char input_buf[MAX_INPUT_BUF_SZ];
3467         int pid;
3468         unsigned long vaddr_start, vaddr_end;
3469
3470         ret = mutex_lock_interruptible(&split_debug_mutex);
3471         if (ret)
3472                 return ret;
3473
3474         ret = -EFAULT;
3475
3476         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3477         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3478                 goto out;
3479
3480         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3481
3482         if (input_buf[0] == '/') {
3483                 char *tok;
3484                 char *buf = input_buf;
3485                 char file_path[MAX_INPUT_BUF_SZ];
3486                 pgoff_t off_start = 0, off_end = 0;
3487                 size_t input_len = strlen(input_buf);
3488
3489                 tok = strsep(&buf, ",");
3490                 if (tok) {
3491                         strcpy(file_path, tok);
3492                 } else {
3493                         ret = -EINVAL;
3494                         goto out;
3495                 }
3496
3497                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3498                 if (ret != 2) {
3499                         ret = -EINVAL;
3500                         goto out;
3501                 }
3502                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3503                 if (!ret)
3504                         ret = input_len;
3505
3506                 goto out;
3507         }
3508
3509         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3510         if (ret == 1 && pid == 1) {
3511                 split_huge_pages_all();
3512                 ret = strlen(input_buf);
3513                 goto out;
3514         } else if (ret != 3) {
3515                 ret = -EINVAL;
3516                 goto out;
3517         }
3518
3519         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3520         if (!ret)
3521                 ret = strlen(input_buf);
3522 out:
3523         mutex_unlock(&split_debug_mutex);
3524         return ret;
3525
3526 }
3527
3528 static const struct file_operations split_huge_pages_fops = {
3529         .owner   = THIS_MODULE,
3530         .write   = split_huge_pages_write,
3531         .llseek  = no_llseek,
3532 };
3533
3534 static int __init split_huge_pages_debugfs(void)
3535 {
3536         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3537                             &split_huge_pages_fops);
3538         return 0;
3539 }
3540 late_initcall(split_huge_pages_debugfs);
3541 #endif
3542
3543 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3544 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3545                 struct page *page)
3546 {
3547         struct folio *folio = page_folio(page);
3548         struct vm_area_struct *vma = pvmw->vma;
3549         struct mm_struct *mm = vma->vm_mm;
3550         unsigned long address = pvmw->address;
3551         bool anon_exclusive;
3552         pmd_t pmdval;
3553         swp_entry_t entry;
3554         pmd_t pmdswp;
3555
3556         if (!(pvmw->pmd && !pvmw->pte))
3557                 return 0;
3558
3559         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3560         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3561
3562         /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
3563         anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
3564         if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {
3565                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3566                 return -EBUSY;
3567         }
3568
3569         if (pmd_dirty(pmdval))
3570                 folio_mark_dirty(folio);
3571         if (pmd_write(pmdval))
3572                 entry = make_writable_migration_entry(page_to_pfn(page));
3573         else if (anon_exclusive)
3574                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3575         else
3576                 entry = make_readable_migration_entry(page_to_pfn(page));
3577         if (pmd_young(pmdval))
3578                 entry = make_migration_entry_young(entry);
3579         if (pmd_dirty(pmdval))
3580                 entry = make_migration_entry_dirty(entry);
3581         pmdswp = swp_entry_to_pmd(entry);
3582         if (pmd_soft_dirty(pmdval))
3583                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3584         if (pmd_uffd_wp(pmdval))
3585                 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3586         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3587         folio_remove_rmap_pmd(folio, page, vma);
3588         folio_put(folio);
3589         trace_set_migration_pmd(address, pmd_val(pmdswp));
3590
3591         return 0;
3592 }
3593
3594 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3595 {
3596         struct folio *folio = page_folio(new);
3597         struct vm_area_struct *vma = pvmw->vma;
3598         struct mm_struct *mm = vma->vm_mm;
3599         unsigned long address = pvmw->address;
3600         unsigned long haddr = address & HPAGE_PMD_MASK;
3601         pmd_t pmde;
3602         swp_entry_t entry;
3603
3604         if (!(pvmw->pmd && !pvmw->pte))
3605                 return;
3606
3607         entry = pmd_to_swp_entry(*pvmw->pmd);
3608         folio_get(folio);
3609         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3610         if (pmd_swp_soft_dirty(*pvmw->pmd))
3611                 pmde = pmd_mksoft_dirty(pmde);
3612         if (is_writable_migration_entry(entry))
3613                 pmde = pmd_mkwrite(pmde, vma);
3614         if (pmd_swp_uffd_wp(*pvmw->pmd))
3615                 pmde = pmd_mkuffd_wp(pmde);
3616         if (!is_migration_entry_young(entry))
3617                 pmde = pmd_mkold(pmde);
3618         /* NOTE: this may contain setting soft-dirty on some archs */
3619         if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
3620                 pmde = pmd_mkdirty(pmde);
3621
3622         if (folio_test_anon(folio)) {
3623                 rmap_t rmap_flags = RMAP_NONE;
3624
3625                 if (!is_readable_migration_entry(entry))
3626                         rmap_flags |= RMAP_EXCLUSIVE;
3627
3628                 folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);
3629         } else {
3630                 folio_add_file_rmap_pmd(folio, new, vma);
3631         }
3632         VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));
3633         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3634
3635         /* No need to invalidate - it was non-present before */
3636         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3637         trace_remove_migration_pmd(address, pmd_val(pmde));
3638 }
3639 #endif