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