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);
 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);
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         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1679         if (vma_is_special_huge(vma)) {
1680                 if (arch_needs_pgtable_deposit())
1681                         zap_deposited_table(tlb->mm, pmd);
1682                 spin_unlock(ptl);
1683         } else if (is_huge_zero_pmd(orig_pmd)) {
1684                 zap_deposited_table(tlb->mm, pmd);
1685                 spin_unlock(ptl);
1686         } else {
1687                 struct page *page = NULL;
1688                 int flush_needed = 1;
1689
1690                 if (pmd_present(orig_pmd)) {
1691                         page = pmd_page(orig_pmd);
1692                         page_remove_rmap(page, vma, true);
1693                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1694                         VM_BUG_ON_PAGE(!PageHead(page), page);
1695                 } else if (thp_migration_supported()) {
1696                         swp_entry_t entry;
1697
1698                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1699                         entry = pmd_to_swp_entry(orig_pmd);
1700                         page = pfn_swap_entry_to_page(entry);
1701                         flush_needed = 0;
1702                 } else
1703                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1704
1705                 if (PageAnon(page)) {
1706                         zap_deposited_table(tlb->mm, pmd);
1707                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1708                 } else {
1709                         if (arch_needs_pgtable_deposit())
1710                                 zap_deposited_table(tlb->mm, pmd);
1711                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1712                 }
1713
1714                 spin_unlock(ptl);
1715                 if (flush_needed)
1716                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1717         }
1718         return 1;
1719 }
1720
1721 #ifndef pmd_move_must_withdraw
1722 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1723                                          spinlock_t *old_pmd_ptl,
1724                                          struct vm_area_struct *vma)
1725 {
1726         /*
1727          * With split pmd lock we also need to move preallocated
1728          * PTE page table if new_pmd is on different PMD page table.
1729          *
1730          * We also don't deposit and withdraw tables for file pages.
1731          */
1732         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1733 }
1734 #endif
1735
1736 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1737 {
1738 #ifdef CONFIG_MEM_SOFT_DIRTY
1739         if (unlikely(is_pmd_migration_entry(pmd)))
1740                 pmd = pmd_swp_mksoft_dirty(pmd);
1741         else if (pmd_present(pmd))
1742                 pmd = pmd_mksoft_dirty(pmd);
1743 #endif
1744         return pmd;
1745 }
1746
1747 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1748                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1749 {
1750         spinlock_t *old_ptl, *new_ptl;
1751         pmd_t pmd;
1752         struct mm_struct *mm = vma->vm_mm;
1753         bool force_flush = false;
1754
1755         /*
1756          * The destination pmd shouldn't be established, free_pgtables()
1757          * should have released it; but move_page_tables() might have already
1758          * inserted a page table, if racing against shmem/file collapse.
1759          */
1760         if (!pmd_none(*new_pmd)) {
1761                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1762                 return false;
1763         }
1764
1765         /*
1766          * We don't have to worry about the ordering of src and dst
1767          * ptlocks because exclusive mmap_lock prevents deadlock.
1768          */
1769         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1770         if (old_ptl) {
1771                 new_ptl = pmd_lockptr(mm, new_pmd);
1772                 if (new_ptl != old_ptl)
1773                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1774                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1775                 if (pmd_present(pmd))
1776                         force_flush = true;
1777                 VM_BUG_ON(!pmd_none(*new_pmd));
1778
1779                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1780                         pgtable_t pgtable;
1781                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1782                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1783                 }
1784                 pmd = move_soft_dirty_pmd(pmd);
1785                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1786                 if (force_flush)
1787                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1788                 if (new_ptl != old_ptl)
1789                         spin_unlock(new_ptl);
1790                 spin_unlock(old_ptl);
1791                 return true;
1792         }
1793         return false;
1794 }
1795
1796 /*
1797  * Returns
1798  *  - 0 if PMD could not be locked
1799  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1800  *      or if prot_numa but THP migration is not supported
1801  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1802  */
1803 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1804                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1805                     unsigned long cp_flags)
1806 {
1807         struct mm_struct *mm = vma->vm_mm;
1808         spinlock_t *ptl;
1809         pmd_t oldpmd, entry;
1810         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1811         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1812         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1813         int ret = 1;
1814
1815         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1816
1817         if (prot_numa && !thp_migration_supported())
1818                 return 1;
1819
1820         ptl = __pmd_trans_huge_lock(pmd, vma);
1821         if (!ptl)
1822                 return 0;
1823
1824 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1825         if (is_swap_pmd(*pmd)) {
1826                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1827                 struct page *page = pfn_swap_entry_to_page(entry);
1828                 pmd_t newpmd;
1829
1830                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1831                 if (is_writable_migration_entry(entry)) {
1832                         /*
1833                          * A protection check is difficult so
1834                          * just be safe and disable write
1835                          */
1836                         if (PageAnon(page))
1837                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1838                         else
1839                                 entry = make_readable_migration_entry(swp_offset(entry));
1840                         newpmd = swp_entry_to_pmd(entry);
1841                         if (pmd_swp_soft_dirty(*pmd))
1842                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1843                 } else {
1844                         newpmd = *pmd;
1845                 }
1846
1847                 if (uffd_wp)
1848                         newpmd = pmd_swp_mkuffd_wp(newpmd);
1849                 else if (uffd_wp_resolve)
1850                         newpmd = pmd_swp_clear_uffd_wp(newpmd);
1851                 if (!pmd_same(*pmd, newpmd))
1852                         set_pmd_at(mm, addr, pmd, newpmd);
1853                 goto unlock;
1854         }
1855 #endif
1856
1857         if (prot_numa) {
1858                 struct page *page;
1859                 bool toptier;
1860                 /*
1861                  * Avoid trapping faults against the zero page. The read-only
1862                  * data is likely to be read-cached on the local CPU and
1863                  * local/remote hits to the zero page are not interesting.
1864                  */
1865                 if (is_huge_zero_pmd(*pmd))
1866                         goto unlock;
1867
1868                 if (pmd_protnone(*pmd))
1869                         goto unlock;
1870
1871                 page = pmd_page(*pmd);
1872                 toptier = node_is_toptier(page_to_nid(page));
1873                 /*
1874                  * Skip scanning top tier node if normal numa
1875                  * balancing is disabled
1876                  */
1877                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1878                     toptier)
1879                         goto unlock;
1880
1881                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1882                     !toptier)
1883                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1884         }
1885         /*
1886          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1887          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1888          * which is also under mmap_read_lock(mm):
1889          *
1890          *      CPU0:                           CPU1:
1891          *                              change_huge_pmd(prot_numa=1)
1892          *                               pmdp_huge_get_and_clear_notify()
1893          * madvise_dontneed()
1894          *  zap_pmd_range()
1895          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1896          *   // skip the pmd
1897          *                               set_pmd_at();
1898          *                               // pmd is re-established
1899          *
1900          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1901          * which may break userspace.
1902          *
1903          * pmdp_invalidate_ad() is required to make sure we don't miss
1904          * dirty/young flags set by hardware.
1905          */
1906         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1907
1908         entry = pmd_modify(oldpmd, newprot);
1909         if (uffd_wp)
1910                 entry = pmd_mkuffd_wp(entry);
1911         else if (uffd_wp_resolve)
1912                 /*
1913                  * Leave the write bit to be handled by PF interrupt
1914                  * handler, then things like COW could be properly
1915                  * handled.
1916                  */
1917                 entry = pmd_clear_uffd_wp(entry);
1918
1919         /* See change_pte_range(). */
1920         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1921             can_change_pmd_writable(vma, addr, entry))
1922                 entry = pmd_mkwrite(entry);
1923
1924         ret = HPAGE_PMD_NR;
1925         set_pmd_at(mm, addr, pmd, entry);
1926
1927         if (huge_pmd_needs_flush(oldpmd, entry))
1928                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1929 unlock:
1930         spin_unlock(ptl);
1931         return ret;
1932 }
1933
1934 /*
1935  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1936  *
1937  * Note that if it returns page table lock pointer, this routine returns without
1938  * unlocking page table lock. So callers must unlock it.
1939  */
1940 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1941 {
1942         spinlock_t *ptl;
1943         ptl = pmd_lock(vma->vm_mm, pmd);
1944         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1945                         pmd_devmap(*pmd)))
1946                 return ptl;
1947         spin_unlock(ptl);
1948         return NULL;
1949 }
1950
1951 /*
1952  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1953  *
1954  * Note that if it returns page table lock pointer, this routine returns without
1955  * unlocking page table lock. So callers must unlock it.
1956  */
1957 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1958 {
1959         spinlock_t *ptl;
1960
1961         ptl = pud_lock(vma->vm_mm, pud);
1962         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1963                 return ptl;
1964         spin_unlock(ptl);
1965         return NULL;
1966 }
1967
1968 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1969 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1970                  pud_t *pud, unsigned long addr)
1971 {
1972         spinlock_t *ptl;
1973
1974         ptl = __pud_trans_huge_lock(pud, vma);
1975         if (!ptl)
1976                 return 0;
1977
1978         pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
1979         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1980         if (vma_is_special_huge(vma)) {
1981                 spin_unlock(ptl);
1982                 /* No zero page support yet */
1983         } else {
1984                 /* No support for anonymous PUD pages yet */
1985                 BUG();
1986         }
1987         return 1;
1988 }
1989
1990 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1991                 unsigned long haddr)
1992 {
1993         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1994         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1995         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1996         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1997
1998         count_vm_event(THP_SPLIT_PUD);
1999
2000         pudp_huge_clear_flush(vma, haddr, pud);
2001 }
2002
2003 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2004                 unsigned long address)
2005 {
2006         spinlock_t *ptl;
2007         struct mmu_notifier_range range;
2008
2009         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2010                                 address & HPAGE_PUD_MASK,
2011                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2012         mmu_notifier_invalidate_range_start(&range);
2013         ptl = pud_lock(vma->vm_mm, pud);
2014         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2015                 goto out;
2016         __split_huge_pud_locked(vma, pud, range.start);
2017
2018 out:
2019         spin_unlock(ptl);
2020         mmu_notifier_invalidate_range_end(&range);
2021 }
2022 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2023
2024 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2025                 unsigned long haddr, pmd_t *pmd)
2026 {
2027         struct mm_struct *mm = vma->vm_mm;
2028         pgtable_t pgtable;
2029         pmd_t _pmd, old_pmd;
2030         unsigned long addr;
2031         pte_t *pte;
2032         int i;
2033
2034         /*
2035          * Leave pmd empty until pte is filled note that it is fine to delay
2036          * notification until mmu_notifier_invalidate_range_end() as we are
2037          * replacing a zero pmd write protected page with a zero pte write
2038          * protected page.
2039          *
2040          * See Documentation/mm/mmu_notifier.rst
2041          */
2042         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2043
2044         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2045         pmd_populate(mm, &_pmd, pgtable);
2046
2047         pte = pte_offset_map(&_pmd, haddr);
2048         VM_BUG_ON(!pte);
2049         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2050                 pte_t entry;
2051
2052                 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2053                 entry = pte_mkspecial(entry);
2054                 if (pmd_uffd_wp(old_pmd))
2055                         entry = pte_mkuffd_wp(entry);
2056                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2057                 set_pte_at(mm, addr, pte, entry);
2058                 pte++;
2059         }
2060         pte_unmap(pte - 1);
2061         smp_wmb(); /* make pte visible before pmd */
2062         pmd_populate(mm, pmd, pgtable);
2063 }
2064
2065 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2066                 unsigned long haddr, bool freeze)
2067 {
2068         struct mm_struct *mm = vma->vm_mm;
2069         struct page *page;
2070         pgtable_t pgtable;
2071         pmd_t old_pmd, _pmd;
2072         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2073         bool anon_exclusive = false, dirty = false;
2074         unsigned long addr;
2075         pte_t *pte;
2076         int i;
2077
2078         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2079         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2080         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2081         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2082                                 && !pmd_devmap(*pmd));
2083
2084         count_vm_event(THP_SPLIT_PMD);
2085
2086         if (!vma_is_anonymous(vma)) {
2087                 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2088                 /*
2089                  * We are going to unmap this huge page. So
2090                  * just go ahead and zap it
2091                  */
2092                 if (arch_needs_pgtable_deposit())
2093                         zap_deposited_table(mm, pmd);
2094                 if (vma_is_special_huge(vma))
2095                         return;
2096                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2097                         swp_entry_t entry;
2098
2099                         entry = pmd_to_swp_entry(old_pmd);
2100                         page = pfn_swap_entry_to_page(entry);
2101                 } else {
2102                         page = pmd_page(old_pmd);
2103                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2104                                 set_page_dirty(page);
2105                         if (!PageReferenced(page) && pmd_young(old_pmd))
2106                                 SetPageReferenced(page);
2107                         page_remove_rmap(page, vma, true);
2108                         put_page(page);
2109                 }
2110                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2111                 return;
2112         }
2113
2114         if (is_huge_zero_pmd(*pmd)) {
2115                 /*
2116                  * FIXME: Do we want to invalidate secondary mmu by calling
2117                  * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2118                  * inside __split_huge_pmd() ?
2119                  *
2120                  * We are going from a zero huge page write protected to zero
2121                  * small page also write protected so it does not seems useful
2122                  * to invalidate secondary mmu at this time.
2123                  */
2124                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2125         }
2126
2127         /*
2128          * Up to this point the pmd is present and huge and userland has the
2129          * whole access to the hugepage during the split (which happens in
2130          * place). If we overwrite the pmd with the not-huge version pointing
2131          * to the pte here (which of course we could if all CPUs were bug
2132          * free), userland could trigger a small page size TLB miss on the
2133          * small sized TLB while the hugepage TLB entry is still established in
2134          * the huge TLB. Some CPU doesn't like that.
2135          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2136          * 383 on page 105. Intel should be safe but is also warns that it's
2137          * only safe if the permission and cache attributes of the two entries
2138          * loaded in the two TLB is identical (which should be the case here).
2139          * But it is generally safer to never allow small and huge TLB entries
2140          * for the same virtual address to be loaded simultaneously. So instead
2141          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2142          * current pmd notpresent (atomically because here the pmd_trans_huge
2143          * must remain set at all times on the pmd until the split is complete
2144          * for this pmd), then we flush the SMP TLB and finally we write the
2145          * non-huge version of the pmd entry with pmd_populate.
2146          */
2147         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2148
2149         pmd_migration = is_pmd_migration_entry(old_pmd);
2150         if (unlikely(pmd_migration)) {
2151                 swp_entry_t entry;
2152
2153                 entry = pmd_to_swp_entry(old_pmd);
2154                 page = pfn_swap_entry_to_page(entry);
2155                 write = is_writable_migration_entry(entry);
2156                 if (PageAnon(page))
2157                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2158                 young = is_migration_entry_young(entry);
2159                 dirty = is_migration_entry_dirty(entry);
2160                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2161                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2162         } else {
2163                 page = pmd_page(old_pmd);
2164                 if (pmd_dirty(old_pmd)) {
2165                         dirty = true;
2166                         SetPageDirty(page);
2167                 }
2168                 write = pmd_write(old_pmd);
2169                 young = pmd_young(old_pmd);
2170                 soft_dirty = pmd_soft_dirty(old_pmd);
2171                 uffd_wp = pmd_uffd_wp(old_pmd);
2172
2173                 VM_BUG_ON_PAGE(!page_count(page), page);
2174
2175                 /*
2176                  * Without "freeze", we'll simply split the PMD, propagating the
2177                  * PageAnonExclusive() flag for each PTE by setting it for
2178                  * each subpage -- no need to (temporarily) clear.
2179                  *
2180                  * With "freeze" we want to replace mapped pages by
2181                  * migration entries right away. This is only possible if we
2182                  * managed to clear PageAnonExclusive() -- see
2183                  * set_pmd_migration_entry().
2184                  *
2185                  * In case we cannot clear PageAnonExclusive(), split the PMD
2186                  * only and let try_to_migrate_one() fail later.
2187                  *
2188                  * See page_try_share_anon_rmap(): invalidate PMD first.
2189                  */
2190                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2191                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2192                         freeze = false;
2193                 if (!freeze)
2194                         page_ref_add(page, HPAGE_PMD_NR - 1);
2195         }
2196
2197         /*
2198          * Withdraw the table only after we mark the pmd entry invalid.
2199          * This's critical for some architectures (Power).
2200          */
2201         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2202         pmd_populate(mm, &_pmd, pgtable);
2203
2204         pte = pte_offset_map(&_pmd, haddr);
2205         VM_BUG_ON(!pte);
2206         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2207                 pte_t entry;
2208                 /*
2209                  * Note that NUMA hinting access restrictions are not
2210                  * transferred to avoid any possibility of altering
2211                  * permissions across VMAs.
2212                  */
2213                 if (freeze || pmd_migration) {
2214                         swp_entry_t swp_entry;
2215                         if (write)
2216                                 swp_entry = make_writable_migration_entry(
2217                                                         page_to_pfn(page + i));
2218                         else if (anon_exclusive)
2219                                 swp_entry = make_readable_exclusive_migration_entry(
2220                                                         page_to_pfn(page + i));
2221                         else
2222                                 swp_entry = make_readable_migration_entry(
2223                                                         page_to_pfn(page + i));
2224                         if (young)
2225                                 swp_entry = make_migration_entry_young(swp_entry);
2226                         if (dirty)
2227                                 swp_entry = make_migration_entry_dirty(swp_entry);
2228                         entry = swp_entry_to_pte(swp_entry);
2229                         if (soft_dirty)
2230                                 entry = pte_swp_mksoft_dirty(entry);
2231                         if (uffd_wp)
2232                                 entry = pte_swp_mkuffd_wp(entry);
2233                 } else {
2234                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2235                         if (write)
2236                                 entry = pte_mkwrite(entry);
2237                         if (anon_exclusive)
2238                                 SetPageAnonExclusive(page + i);
2239                         if (!young)
2240                                 entry = pte_mkold(entry);
2241                         /* NOTE: this may set soft-dirty too on some archs */
2242                         if (dirty)
2243                                 entry = pte_mkdirty(entry);
2244                         if (soft_dirty)
2245                                 entry = pte_mksoft_dirty(entry);
2246                         if (uffd_wp)
2247                                 entry = pte_mkuffd_wp(entry);
2248                         page_add_anon_rmap(page + i, vma, addr, RMAP_NONE);
2249                 }
2250                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2251                 set_pte_at(mm, addr, pte, entry);
2252                 pte++;
2253         }
2254         pte_unmap(pte - 1);
2255
2256         if (!pmd_migration)
2257                 page_remove_rmap(page, vma, true);
2258         if (freeze)
2259                 put_page(page);
2260
2261         smp_wmb(); /* make pte visible before pmd */
2262         pmd_populate(mm, pmd, pgtable);
2263 }
2264
2265 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2266                 unsigned long address, bool freeze, struct folio *folio)
2267 {
2268         spinlock_t *ptl;
2269         struct mmu_notifier_range range;
2270
2271         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2272                                 address & HPAGE_PMD_MASK,
2273                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2274         mmu_notifier_invalidate_range_start(&range);
2275         ptl = pmd_lock(vma->vm_mm, pmd);
2276
2277         /*
2278          * If caller asks to setup a migration entry, we need a folio to check
2279          * pmd against. Otherwise we can end up replacing wrong folio.
2280          */
2281         VM_BUG_ON(freeze && !folio);
2282         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2283
2284         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2285             is_pmd_migration_entry(*pmd)) {
2286                 /*
2287                  * It's safe to call pmd_page when folio is set because it's
2288                  * guaranteed that pmd is present.
2289                  */
2290                 if (folio && folio != page_folio(pmd_page(*pmd)))
2291                         goto out;
2292                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2293         }
2294
2295 out:
2296         spin_unlock(ptl);
2297         mmu_notifier_invalidate_range_end(&range);
2298 }
2299
2300 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2301                 bool freeze, struct folio *folio)
2302 {
2303         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2304
2305         if (!pmd)
2306                 return;
2307
2308         __split_huge_pmd(vma, pmd, address, freeze, folio);
2309 }
2310
2311 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2312 {
2313         /*
2314          * If the new address isn't hpage aligned and it could previously
2315          * contain an hugepage: check if we need to split an huge pmd.
2316          */
2317         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2318             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2319                          ALIGN(address, HPAGE_PMD_SIZE)))
2320                 split_huge_pmd_address(vma, address, false, NULL);
2321 }
2322
2323 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2324                              unsigned long start,
2325                              unsigned long end,
2326                              long adjust_next)
2327 {
2328         /* Check if we need to split start first. */
2329         split_huge_pmd_if_needed(vma, start);
2330
2331         /* Check if we need to split end next. */
2332         split_huge_pmd_if_needed(vma, end);
2333
2334         /*
2335          * If we're also updating the next vma vm_start,
2336          * check if we need to split it.
2337          */
2338         if (adjust_next > 0) {
2339                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2340                 unsigned long nstart = next->vm_start;
2341                 nstart += adjust_next;
2342                 split_huge_pmd_if_needed(next, nstart);
2343         }
2344 }
2345
2346 static void unmap_folio(struct folio *folio)
2347 {
2348         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2349                 TTU_SYNC;
2350
2351         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2352
2353         /*
2354          * Anon pages need migration entries to preserve them, but file
2355          * pages can simply be left unmapped, then faulted back on demand.
2356          * If that is ever changed (perhaps for mlock), update remap_page().
2357          */
2358         if (folio_test_anon(folio))
2359                 try_to_migrate(folio, ttu_flags);
2360         else
2361                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2362 }
2363
2364 static void remap_page(struct folio *folio, unsigned long nr)
2365 {
2366         int i = 0;
2367
2368         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2369         if (!folio_test_anon(folio))
2370                 return;
2371         for (;;) {
2372                 remove_migration_ptes(folio, folio, true);
2373                 i += folio_nr_pages(folio);
2374                 if (i >= nr)
2375                         break;
2376                 folio = folio_next(folio);
2377         }
2378 }
2379
2380 static void lru_add_page_tail(struct page *head, struct page *tail,
2381                 struct lruvec *lruvec, struct list_head *list)
2382 {
2383         VM_BUG_ON_PAGE(!PageHead(head), head);
2384         VM_BUG_ON_PAGE(PageCompound(tail), head);
2385         VM_BUG_ON_PAGE(PageLRU(tail), head);
2386         lockdep_assert_held(&lruvec->lru_lock);
2387
2388         if (list) {
2389                 /* page reclaim is reclaiming a huge page */
2390                 VM_WARN_ON(PageLRU(head));
2391                 get_page(tail);
2392                 list_add_tail(&tail->lru, list);
2393         } else {
2394                 /* head is still on lru (and we have it frozen) */
2395                 VM_WARN_ON(!PageLRU(head));
2396                 if (PageUnevictable(tail))
2397                         tail->mlock_count = 0;
2398                 else
2399                         list_add_tail(&tail->lru, &head->lru);
2400                 SetPageLRU(tail);
2401         }
2402 }
2403
2404 static void __split_huge_page_tail(struct folio *folio, int tail,
2405                 struct lruvec *lruvec, struct list_head *list)
2406 {
2407         struct page *head = &folio->page;
2408         struct page *page_tail = head + tail;
2409         /*
2410          * Careful: new_folio is not a "real" folio before we cleared PageTail.
2411          * Don't pass it around before clear_compound_head().
2412          */
2413         struct folio *new_folio = (struct folio *)page_tail;
2414
2415         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2416
2417         /*
2418          * Clone page flags before unfreezing refcount.
2419          *
2420          * After successful get_page_unless_zero() might follow flags change,
2421          * for example lock_page() which set PG_waiters.
2422          *
2423          * Note that for mapped sub-pages of an anonymous THP,
2424          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2425          * the migration entry instead from where remap_page() will restore it.
2426          * We can still have PG_anon_exclusive set on effectively unmapped and
2427          * unreferenced sub-pages of an anonymous THP: we can simply drop
2428          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2429          */
2430         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2431         page_tail->flags |= (head->flags &
2432                         ((1L << PG_referenced) |
2433                          (1L << PG_swapbacked) |
2434                          (1L << PG_swapcache) |
2435                          (1L << PG_mlocked) |
2436                          (1L << PG_uptodate) |
2437                          (1L << PG_active) |
2438                          (1L << PG_workingset) |
2439                          (1L << PG_locked) |
2440                          (1L << PG_unevictable) |
2441 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2442                          (1L << PG_arch_2) |
2443                          (1L << PG_arch_3) |
2444 #endif
2445                          (1L << PG_dirty) |
2446                          LRU_GEN_MASK | LRU_REFS_MASK));
2447
2448         /* ->mapping in first and second tail page is replaced by other uses */
2449         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2450                         page_tail);
2451         page_tail->mapping = head->mapping;
2452         page_tail->index = head->index + tail;
2453
2454         /*
2455          * page->private should not be set in tail pages. Fix up and warn once
2456          * if private is unexpectedly set.
2457          */
2458         if (unlikely(page_tail->private)) {
2459                 VM_WARN_ON_ONCE_PAGE(true, page_tail);
2460                 page_tail->private = 0;
2461         }
2462         if (folio_test_swapcache(folio))
2463                 new_folio->swap.val = folio->swap.val + tail;
2464
2465         /* Page flags must be visible before we make the page non-compound. */
2466         smp_wmb();
2467
2468         /*
2469          * Clear PageTail before unfreezing page refcount.
2470          *
2471          * After successful get_page_unless_zero() might follow put_page()
2472          * which needs correct compound_head().
2473          */
2474         clear_compound_head(page_tail);
2475
2476         /* Finally unfreeze refcount. Additional reference from page cache. */
2477         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2478                                           PageSwapCache(head)));
2479
2480         if (page_is_young(head))
2481                 set_page_young(page_tail);
2482         if (page_is_idle(head))
2483                 set_page_idle(page_tail);
2484
2485         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2486
2487         /*
2488          * always add to the tail because some iterators expect new
2489          * pages to show after the currently processed elements - e.g.
2490          * migrate_pages
2491          */
2492         lru_add_page_tail(head, page_tail, lruvec, list);
2493 }
2494
2495 static void __split_huge_page(struct page *page, struct list_head *list,
2496                 pgoff_t end)
2497 {
2498         struct folio *folio = page_folio(page);
2499         struct page *head = &folio->page;
2500         struct lruvec *lruvec;
2501         struct address_space *swap_cache = NULL;
2502         unsigned long offset = 0;
2503         unsigned int nr = thp_nr_pages(head);
2504         int i, nr_dropped = 0;
2505
2506         /* complete memcg works before add pages to LRU */
2507         split_page_memcg(head, nr);
2508
2509         if (folio_test_anon(folio) && folio_test_swapcache(folio)) {
2510                 offset = swp_offset(folio->swap);
2511                 swap_cache = swap_address_space(folio->swap);
2512                 xa_lock(&swap_cache->i_pages);
2513         }
2514
2515         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2516         lruvec = folio_lruvec_lock(folio);
2517
2518         ClearPageHasHWPoisoned(head);
2519
2520         for (i = nr - 1; i >= 1; i--) {
2521                 __split_huge_page_tail(folio, i, lruvec, list);
2522                 /* Some pages can be beyond EOF: drop them from page cache */
2523                 if (head[i].index >= end) {
2524                         struct folio *tail = page_folio(head + i);
2525
2526                         if (shmem_mapping(head->mapping))
2527                                 nr_dropped++;
2528                         else if (folio_test_clear_dirty(tail))
2529                                 folio_account_cleaned(tail,
2530                                         inode_to_wb(folio->mapping->host));
2531                         __filemap_remove_folio(tail, NULL);
2532                         folio_put(tail);
2533                 } else if (!PageAnon(page)) {
2534                         __xa_store(&head->mapping->i_pages, head[i].index,
2535                                         head + i, 0);
2536                 } else if (swap_cache) {
2537                         __xa_store(&swap_cache->i_pages, offset + i,
2538                                         head + i, 0);
2539                 }
2540         }
2541
2542         ClearPageCompound(head);
2543         unlock_page_lruvec(lruvec);
2544         /* Caller disabled irqs, so they are still disabled here */
2545
2546         split_page_owner(head, nr);
2547
2548         /* See comment in __split_huge_page_tail() */
2549         if (PageAnon(head)) {
2550                 /* Additional pin to swap cache */
2551                 if (PageSwapCache(head)) {
2552                         page_ref_add(head, 2);
2553                         xa_unlock(&swap_cache->i_pages);
2554                 } else {
2555                         page_ref_inc(head);
2556                 }
2557         } else {
2558                 /* Additional pin to page cache */
2559                 page_ref_add(head, 2);
2560                 xa_unlock(&head->mapping->i_pages);
2561         }
2562         local_irq_enable();
2563
2564         if (nr_dropped)
2565                 shmem_uncharge(head->mapping->host, nr_dropped);
2566         remap_page(folio, nr);
2567
2568         if (folio_test_swapcache(folio))
2569                 split_swap_cluster(folio->swap);
2570
2571         for (i = 0; i < nr; i++) {
2572                 struct page *subpage = head + i;
2573                 if (subpage == page)
2574                         continue;
2575                 unlock_page(subpage);
2576
2577                 /*
2578                  * Subpages may be freed if there wasn't any mapping
2579                  * like if add_to_swap() is running on a lru page that
2580                  * had its mapping zapped. And freeing these pages
2581                  * requires taking the lru_lock so we do the put_page
2582                  * of the tail pages after the split is complete.
2583                  */
2584                 free_page_and_swap_cache(subpage);
2585         }
2586 }
2587
2588 /* Racy check whether the huge page can be split */
2589 bool can_split_folio(struct folio *folio, int *pextra_pins)
2590 {
2591         int extra_pins;
2592
2593         /* Additional pins from page cache */
2594         if (folio_test_anon(folio))
2595                 extra_pins = folio_test_swapcache(folio) ?
2596                                 folio_nr_pages(folio) : 0;
2597         else
2598                 extra_pins = folio_nr_pages(folio);
2599         if (pextra_pins)
2600                 *pextra_pins = extra_pins;
2601         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2602 }
2603
2604 /*
2605  * This function splits huge page into normal pages. @page can point to any
2606  * subpage of huge page to split. Split doesn't change the position of @page.
2607  *
2608  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2609  * The huge page must be locked.
2610  *
2611  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2612  *
2613  * Both head page and tail pages will inherit mapping, flags, and so on from
2614  * the hugepage.
2615  *
2616  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2617  * they are not mapped.
2618  *
2619  * Returns 0 if the hugepage is split successfully.
2620  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2621  * us.
2622  */
2623 int split_huge_page_to_list(struct page *page, struct list_head *list)
2624 {
2625         struct folio *folio = page_folio(page);
2626         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2627         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2628         struct anon_vma *anon_vma = NULL;
2629         struct address_space *mapping = NULL;
2630         int extra_pins, ret;
2631         pgoff_t end;
2632         bool is_hzp;
2633
2634         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2635         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2636
2637         is_hzp = is_huge_zero_page(&folio->page);
2638         if (is_hzp) {
2639                 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2640                 return -EBUSY;
2641         }
2642
2643         if (folio_test_writeback(folio))
2644                 return -EBUSY;
2645
2646         if (folio_test_anon(folio)) {
2647                 /*
2648                  * The caller does not necessarily hold an mmap_lock that would
2649                  * prevent the anon_vma disappearing so we first we take a
2650                  * reference to it and then lock the anon_vma for write. This
2651                  * is similar to folio_lock_anon_vma_read except the write lock
2652                  * is taken to serialise against parallel split or collapse
2653                  * operations.
2654                  */
2655                 anon_vma = folio_get_anon_vma(folio);
2656                 if (!anon_vma) {
2657                         ret = -EBUSY;
2658                         goto out;
2659                 }
2660                 end = -1;
2661                 mapping = NULL;
2662                 anon_vma_lock_write(anon_vma);
2663         } else {
2664                 gfp_t gfp;
2665
2666                 mapping = folio->mapping;
2667
2668                 /* Truncated ? */
2669                 if (!mapping) {
2670                         ret = -EBUSY;
2671                         goto out;
2672                 }
2673
2674                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2675                                                         GFP_RECLAIM_MASK);
2676
2677                 if (!filemap_release_folio(folio, gfp)) {
2678                         ret = -EBUSY;
2679                         goto out;
2680                 }
2681
2682                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2683                 if (xas_error(&xas)) {
2684                         ret = xas_error(&xas);
2685                         goto out;
2686                 }
2687
2688                 anon_vma = NULL;
2689                 i_mmap_lock_read(mapping);
2690
2691                 /*
2692                  *__split_huge_page() may need to trim off pages beyond EOF:
2693                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2694                  * which cannot be nested inside the page tree lock. So note
2695                  * end now: i_size itself may be changed at any moment, but
2696                  * folio lock is good enough to serialize the trimming.
2697                  */
2698                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2699                 if (shmem_mapping(mapping))
2700                         end = shmem_fallocend(mapping->host, end);
2701         }
2702
2703         /*
2704          * Racy check if we can split the page, before unmap_folio() will
2705          * split PMDs
2706          */
2707         if (!can_split_folio(folio, &extra_pins)) {
2708                 ret = -EAGAIN;
2709                 goto out_unlock;
2710         }
2711
2712         unmap_folio(folio);
2713
2714         /* block interrupt reentry in xa_lock and spinlock */
2715         local_irq_disable();
2716         if (mapping) {
2717                 /*
2718                  * Check if the folio is present in page cache.
2719                  * We assume all tail are present too, if folio is there.
2720                  */
2721                 xas_lock(&xas);
2722                 xas_reset(&xas);
2723                 if (xas_load(&xas) != folio)
2724                         goto fail;
2725         }
2726
2727         /* Prevent deferred_split_scan() touching ->_refcount */
2728         spin_lock(&ds_queue->split_queue_lock);
2729         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2730                 if (!list_empty(&folio->_deferred_list)) {
2731                         ds_queue->split_queue_len--;
2732                         list_del(&folio->_deferred_list);
2733                 }
2734                 spin_unlock(&ds_queue->split_queue_lock);
2735                 if (mapping) {
2736                         int nr = folio_nr_pages(folio);
2737
2738                         xas_split(&xas, folio, folio_order(folio));
2739                         if (folio_test_swapbacked(folio)) {
2740                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2741                                                         -nr);
2742                         } else {
2743                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2744                                                         -nr);
2745                                 filemap_nr_thps_dec(mapping);
2746                         }
2747                 }
2748
2749                 __split_huge_page(page, list, end);
2750                 ret = 0;
2751         } else {
2752                 spin_unlock(&ds_queue->split_queue_lock);
2753 fail:
2754                 if (mapping)
2755                         xas_unlock(&xas);
2756                 local_irq_enable();
2757                 remap_page(folio, folio_nr_pages(folio));
2758                 ret = -EAGAIN;
2759         }
2760
2761 out_unlock:
2762         if (anon_vma) {
2763                 anon_vma_unlock_write(anon_vma);
2764                 put_anon_vma(anon_vma);
2765         }
2766         if (mapping)
2767                 i_mmap_unlock_read(mapping);
2768 out:
2769         xas_destroy(&xas);
2770         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2771         return ret;
2772 }
2773
2774 void folio_undo_large_rmappable(struct folio *folio)
2775 {
2776         struct deferred_split *ds_queue;
2777         unsigned long flags;
2778
2779         /*
2780          * At this point, there is no one trying to add the folio to
2781          * deferred_list. If folio is not in deferred_list, it's safe
2782          * to check without acquiring the split_queue_lock.
2783          */
2784         if (data_race(list_empty(&folio->_deferred_list)))
2785                 return;
2786
2787         ds_queue = get_deferred_split_queue(folio);
2788         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2789         if (!list_empty(&folio->_deferred_list)) {
2790                 ds_queue->split_queue_len--;
2791                 list_del(&folio->_deferred_list);
2792         }
2793         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2794 }
2795
2796 void deferred_split_folio(struct folio *folio)
2797 {
2798         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2799 #ifdef CONFIG_MEMCG
2800         struct mem_cgroup *memcg = folio_memcg(folio);
2801 #endif
2802         unsigned long flags;
2803
2804         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2805
2806         /*
2807          * The try_to_unmap() in page reclaim path might reach here too,
2808          * this may cause a race condition to corrupt deferred split queue.
2809          * And, if page reclaim is already handling the same folio, it is
2810          * unnecessary to handle it again in shrinker.
2811          *
2812          * Check the swapcache flag to determine if the folio is being
2813          * handled by page reclaim since THP swap would add the folio into
2814          * swap cache before calling try_to_unmap().
2815          */
2816         if (folio_test_swapcache(folio))
2817                 return;
2818
2819         if (!list_empty(&folio->_deferred_list))
2820                 return;
2821
2822         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2823         if (list_empty(&folio->_deferred_list)) {
2824                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2825                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2826                 ds_queue->split_queue_len++;
2827 #ifdef CONFIG_MEMCG
2828                 if (memcg)
2829                         set_shrinker_bit(memcg, folio_nid(folio),
2830                                          deferred_split_shrinker.id);
2831 #endif
2832         }
2833         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2834 }
2835
2836 static unsigned long deferred_split_count(struct shrinker *shrink,
2837                 struct shrink_control *sc)
2838 {
2839         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2840         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2841
2842 #ifdef CONFIG_MEMCG
2843         if (sc->memcg)
2844                 ds_queue = &sc->memcg->deferred_split_queue;
2845 #endif
2846         return READ_ONCE(ds_queue->split_queue_len);
2847 }
2848
2849 static unsigned long deferred_split_scan(struct shrinker *shrink,
2850                 struct shrink_control *sc)
2851 {
2852         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2853         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2854         unsigned long flags;
2855         LIST_HEAD(list);
2856         struct folio *folio, *next;
2857         int split = 0;
2858
2859 #ifdef CONFIG_MEMCG
2860         if (sc->memcg)
2861                 ds_queue = &sc->memcg->deferred_split_queue;
2862 #endif
2863
2864         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2865         /* Take pin on all head pages to avoid freeing them under us */
2866         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2867                                                         _deferred_list) {
2868                 if (folio_try_get(folio)) {
2869                         list_move(&folio->_deferred_list, &list);
2870                 } else {
2871                         /* We lost race with folio_put() */
2872                         list_del_init(&folio->_deferred_list);
2873                         ds_queue->split_queue_len--;
2874                 }
2875                 if (!--sc->nr_to_scan)
2876                         break;
2877         }
2878         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2879
2880         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2881                 if (!folio_trylock(folio))
2882                         goto next;
2883                 /* split_huge_page() removes page from list on success */
2884                 if (!split_folio(folio))
2885                         split++;
2886                 folio_unlock(folio);
2887 next:
2888                 folio_put(folio);
2889         }
2890
2891         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2892         list_splice_tail(&list, &ds_queue->split_queue);
2893         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2894
2895         /*
2896          * Stop shrinker if we didn't split any page, but the queue is empty.
2897          * This can happen if pages were freed under us.
2898          */
2899         if (!split && list_empty(&ds_queue->split_queue))
2900                 return SHRINK_STOP;
2901         return split;
2902 }
2903
2904 static struct shrinker deferred_split_shrinker = {
2905         .count_objects = deferred_split_count,
2906         .scan_objects = deferred_split_scan,
2907         .seeks = DEFAULT_SEEKS,
2908         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2909                  SHRINKER_NONSLAB,
2910 };
2911
2912 #ifdef CONFIG_DEBUG_FS
2913 static void split_huge_pages_all(void)
2914 {
2915         struct zone *zone;
2916         struct page *page;
2917         struct folio *folio;
2918         unsigned long pfn, max_zone_pfn;
2919         unsigned long total = 0, split = 0;
2920
2921         pr_debug("Split all THPs\n");
2922         for_each_zone(zone) {
2923                 if (!managed_zone(zone))
2924                         continue;
2925                 max_zone_pfn = zone_end_pfn(zone);
2926                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2927                         int nr_pages;
2928
2929                         page = pfn_to_online_page(pfn);
2930                         if (!page || PageTail(page))
2931                                 continue;
2932                         folio = page_folio(page);
2933                         if (!folio_try_get(folio))
2934                                 continue;
2935
2936                         if (unlikely(page_folio(page) != folio))
2937                                 goto next;
2938
2939                         if (zone != folio_zone(folio))
2940                                 goto next;
2941
2942                         if (!folio_test_large(folio)
2943                                 || folio_test_hugetlb(folio)
2944                                 || !folio_test_lru(folio))
2945                                 goto next;
2946
2947                         total++;
2948                         folio_lock(folio);
2949                         nr_pages = folio_nr_pages(folio);
2950                         if (!split_folio(folio))
2951                                 split++;
2952                         pfn += nr_pages - 1;
2953                         folio_unlock(folio);
2954 next:
2955                         folio_put(folio);
2956                         cond_resched();
2957                 }
2958         }
2959
2960         pr_debug("%lu of %lu THP split\n", split, total);
2961 }
2962
2963 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2964 {
2965         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2966                     is_vm_hugetlb_page(vma);
2967 }
2968
2969 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2970                                 unsigned long vaddr_end)
2971 {
2972         int ret = 0;
2973         struct task_struct *task;
2974         struct mm_struct *mm;
2975         unsigned long total = 0, split = 0;
2976         unsigned long addr;
2977
2978         vaddr_start &= PAGE_MASK;
2979         vaddr_end &= PAGE_MASK;
2980
2981         /* Find the task_struct from pid */
2982         rcu_read_lock();
2983         task = find_task_by_vpid(pid);
2984         if (!task) {
2985                 rcu_read_unlock();
2986                 ret = -ESRCH;
2987                 goto out;
2988         }
2989         get_task_struct(task);
2990         rcu_read_unlock();
2991
2992         /* Find the mm_struct */
2993         mm = get_task_mm(task);
2994         put_task_struct(task);
2995
2996         if (!mm) {
2997                 ret = -EINVAL;
2998                 goto out;
2999         }
3000
3001         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3002                  pid, vaddr_start, vaddr_end);
3003
3004         mmap_read_lock(mm);
3005         /*
3006          * always increase addr by PAGE_SIZE, since we could have a PTE page
3007          * table filled with PTE-mapped THPs, each of which is distinct.
3008          */
3009         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3010                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3011                 struct page *page;
3012                 struct folio *folio;
3013
3014                 if (!vma)
3015                         break;
3016
3017                 /* skip special VMA and hugetlb VMA */
3018                 if (vma_not_suitable_for_thp_split(vma)) {
3019                         addr = vma->vm_end;
3020                         continue;
3021                 }
3022
3023                 /* FOLL_DUMP to ignore special (like zero) pages */
3024                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3025
3026                 if (IS_ERR_OR_NULL(page))
3027                         continue;
3028
3029                 folio = page_folio(page);
3030                 if (!is_transparent_hugepage(folio))
3031                         goto next;
3032
3033                 total++;
3034                 if (!can_split_folio(folio, NULL))
3035                         goto next;
3036
3037                 if (!folio_trylock(folio))
3038                         goto next;
3039
3040                 if (!split_folio(folio))
3041                         split++;
3042
3043                 folio_unlock(folio);
3044 next:
3045                 folio_put(folio);
3046                 cond_resched();
3047         }
3048         mmap_read_unlock(mm);
3049         mmput(mm);
3050
3051         pr_debug("%lu of %lu THP split\n", split, total);
3052
3053 out:
3054         return ret;
3055 }
3056
3057 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3058                                 pgoff_t off_end)
3059 {
3060         struct filename *file;
3061         struct file *candidate;
3062         struct address_space *mapping;
3063         int ret = -EINVAL;
3064         pgoff_t index;
3065         int nr_pages = 1;
3066         unsigned long total = 0, split = 0;
3067
3068         file = getname_kernel(file_path);
3069         if (IS_ERR(file))
3070                 return ret;
3071
3072         candidate = file_open_name(file, O_RDONLY, 0);
3073         if (IS_ERR(candidate))
3074                 goto out;
3075
3076         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3077                  file_path, off_start, off_end);
3078
3079         mapping = candidate->f_mapping;
3080
3081         for (index = off_start; index < off_end; index += nr_pages) {
3082                 struct folio *folio = filemap_get_folio(mapping, index);
3083
3084                 nr_pages = 1;
3085                 if (IS_ERR(folio))
3086                         continue;
3087
3088                 if (!folio_test_large(folio))
3089                         goto next;
3090
3091                 total++;
3092                 nr_pages = folio_nr_pages(folio);
3093
3094                 if (!folio_trylock(folio))
3095                         goto next;
3096
3097                 if (!split_folio(folio))
3098                         split++;
3099
3100                 folio_unlock(folio);
3101 next:
3102                 folio_put(folio);
3103                 cond_resched();
3104         }
3105
3106         filp_close(candidate, NULL);
3107         ret = 0;
3108
3109         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3110 out:
3111         putname(file);
3112         return ret;
3113 }
3114
3115 #define MAX_INPUT_BUF_SZ 255
3116
3117 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3118                                 size_t count, loff_t *ppops)
3119 {
3120         static DEFINE_MUTEX(split_debug_mutex);
3121         ssize_t ret;
3122         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3123         char input_buf[MAX_INPUT_BUF_SZ];
3124         int pid;
3125         unsigned long vaddr_start, vaddr_end;
3126
3127         ret = mutex_lock_interruptible(&split_debug_mutex);
3128         if (ret)
3129                 return ret;
3130
3131         ret = -EFAULT;
3132
3133         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3134         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3135                 goto out;
3136
3137         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3138
3139         if (input_buf[0] == '/') {
3140                 char *tok;
3141                 char *buf = input_buf;
3142                 char file_path[MAX_INPUT_BUF_SZ];
3143                 pgoff_t off_start = 0, off_end = 0;
3144                 size_t input_len = strlen(input_buf);
3145
3146                 tok = strsep(&buf, ",");
3147                 if (tok) {
3148                         strcpy(file_path, tok);
3149                 } else {
3150                         ret = -EINVAL;
3151                         goto out;
3152                 }
3153
3154                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3155                 if (ret != 2) {
3156                         ret = -EINVAL;
3157                         goto out;
3158                 }
3159                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3160                 if (!ret)
3161                         ret = input_len;
3162
3163                 goto out;
3164         }
3165
3166         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3167         if (ret == 1 && pid == 1) {
3168                 split_huge_pages_all();
3169                 ret = strlen(input_buf);
3170                 goto out;
3171         } else if (ret != 3) {
3172                 ret = -EINVAL;
3173                 goto out;
3174         }
3175
3176         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3177         if (!ret)
3178                 ret = strlen(input_buf);
3179 out:
3180         mutex_unlock(&split_debug_mutex);
3181         return ret;
3182
3183 }
3184
3185 static const struct file_operations split_huge_pages_fops = {
3186         .owner   = THIS_MODULE,
3187         .write   = split_huge_pages_write,
3188         .llseek  = no_llseek,
3189 };
3190
3191 static int __init split_huge_pages_debugfs(void)
3192 {
3193         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3194                             &split_huge_pages_fops);
3195         return 0;
3196 }
3197 late_initcall(split_huge_pages_debugfs);
3198 #endif
3199
3200 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3201 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3202                 struct page *page)
3203 {
3204         struct vm_area_struct *vma = pvmw->vma;
3205         struct mm_struct *mm = vma->vm_mm;
3206         unsigned long address = pvmw->address;
3207         bool anon_exclusive;
3208         pmd_t pmdval;
3209         swp_entry_t entry;
3210         pmd_t pmdswp;
3211
3212         if (!(pvmw->pmd && !pvmw->pte))
3213                 return 0;
3214
3215         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3216         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3217
3218         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3219         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3220         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3221                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3222                 return -EBUSY;
3223         }
3224
3225         if (pmd_dirty(pmdval))
3226                 set_page_dirty(page);
3227         if (pmd_write(pmdval))
3228                 entry = make_writable_migration_entry(page_to_pfn(page));
3229         else if (anon_exclusive)
3230                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3231         else
3232                 entry = make_readable_migration_entry(page_to_pfn(page));
3233         if (pmd_young(pmdval))
3234                 entry = make_migration_entry_young(entry);
3235         if (pmd_dirty(pmdval))
3236                 entry = make_migration_entry_dirty(entry);
3237         pmdswp = swp_entry_to_pmd(entry);
3238         if (pmd_soft_dirty(pmdval))
3239                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3240         if (pmd_uffd_wp(pmdval))
3241                 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3242         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3243         page_remove_rmap(page, vma, true);
3244         put_page(page);
3245         trace_set_migration_pmd(address, pmd_val(pmdswp));
3246
3247         return 0;
3248 }
3249
3250 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3251 {
3252         struct vm_area_struct *vma = pvmw->vma;
3253         struct mm_struct *mm = vma->vm_mm;
3254         unsigned long address = pvmw->address;
3255         unsigned long haddr = address & HPAGE_PMD_MASK;
3256         pmd_t pmde;
3257         swp_entry_t entry;
3258
3259         if (!(pvmw->pmd && !pvmw->pte))
3260                 return;
3261
3262         entry = pmd_to_swp_entry(*pvmw->pmd);
3263         get_page(new);
3264         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3265         if (pmd_swp_soft_dirty(*pvmw->pmd))
3266                 pmde = pmd_mksoft_dirty(pmde);
3267         if (is_writable_migration_entry(entry))
3268                 pmde = pmd_mkwrite(pmde);
3269         if (pmd_swp_uffd_wp(*pvmw->pmd))
3270                 pmde = pmd_mkuffd_wp(pmde);
3271         if (!is_migration_entry_young(entry))
3272                 pmde = pmd_mkold(pmde);
3273         /* NOTE: this may contain setting soft-dirty on some archs */
3274         if (PageDirty(new) && is_migration_entry_dirty(entry))
3275                 pmde = pmd_mkdirty(pmde);
3276
3277         if (PageAnon(new)) {
3278                 rmap_t rmap_flags = RMAP_COMPOUND;
3279
3280                 if (!is_readable_migration_entry(entry))
3281                         rmap_flags |= RMAP_EXCLUSIVE;
3282
3283                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3284         } else {
3285                 page_add_file_rmap(new, vma, true);
3286         }
3287         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3288         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3289
3290         /* No need to invalidate - it was non-present before */
3291         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3292         trace_remove_migration_pmd(address, pmd_val(pmde));
3293 }
3294 #endif