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_NEVER_DAX))
  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_huge_enabled(vma, !enforce_sysfs);
 123
 124         /* Enforce sysfs THP requirements as necessary */
 125         if (enforce_sysfs &&
 126             (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 127                                            !hugepage_flags_always())))
 128                 return false;
 129
 130         /* Only regular file is valid */
 131         if (!in_pf && file_thp_enabled(vma))
 132                 return true;
 133
 134         if (!vma_is_anonymous(vma))
 135                 return false;
 136
 137         if (vma_is_temporary_stack(vma))
 138                 return false;
 139
 140         /*
 141          * THPeligible bit of smaps should show 1 for proper VMAs even
 142          * though anon_vma is not initialized yet.
 143          *
 144          * Allow page fault since anon_vma may be not initialized until
 145          * the first page fault.
 146          */
 147         if (!vma->anon_vma)
 148                 return (smaps || in_pf);
 149
 150         return true;
 151 }
 152
 153 static bool get_huge_zero_page(void)
 154 {
 155         struct page *zero_page;
 156 retry:
 157         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 158                 return true;
 159
 160         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 161                         HPAGE_PMD_ORDER);
 162         if (!zero_page) {
 163                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 164                 return false;
 165         }
 166         preempt_disable();
 167         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 168                 preempt_enable();
 169                 __free_pages(zero_page, compound_order(zero_page));
 170                 goto retry;
 171         }
 172         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 173
 174         /* We take additional reference here. It will be put back by shrinker */
 175         atomic_set(&huge_zero_refcount, 2);
 176         preempt_enable();
 177         count_vm_event(THP_ZERO_PAGE_ALLOC);
 178         return true;
 179 }
 180
 181 static void put_huge_zero_page(void)
 182 {
 183         /*
 184          * Counter should never go to zero here. Only shrinker can put
 185          * last reference.
 186          */
 187         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 188 }
 189
 190 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 191 {
 192         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 193                 return READ_ONCE(huge_zero_page);
 194
 195         if (!get_huge_zero_page())
 196                 return NULL;
 197
 198         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 199                 put_huge_zero_page();
 200
 201         return READ_ONCE(huge_zero_page);
 202 }
 203
 204 void mm_put_huge_zero_page(struct mm_struct *mm)
 205 {
 206         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 207                 put_huge_zero_page();
 208 }
 209
 210 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 211                                         struct shrink_control *sc)
 212 {
 213         /* we can free zero page only if last reference remains */
 214         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 215 }
 216
 217 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 218                                        struct shrink_control *sc)
 219 {
 220         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 221                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 222                 BUG_ON(zero_page == NULL);
 223                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 224                 __free_pages(zero_page, compound_order(zero_page));
 225                 return HPAGE_PMD_NR;
 226         }
 227
 228         return 0;
 229 }
 230
 231 static struct shrinker huge_zero_page_shrinker = {
 232         .count_objects = shrink_huge_zero_page_count,
 233         .scan_objects = shrink_huge_zero_page_scan,
 234         .seeks = DEFAULT_SEEKS,
 235 };
 236
 237 #ifdef CONFIG_SYSFS
 238 static ssize_t enabled_show(struct kobject *kobj,
 239                             struct kobj_attribute *attr, char *buf)
 240 {
 241         const char *output;
 242
 243         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 244                 output = "[always] madvise never";
 245         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 246                           &transparent_hugepage_flags))
 247                 output = "always [madvise] never";
 248         else
 249                 output = "always madvise [never]";
 250
 251         return sysfs_emit(buf, "%s\n", output);
 252 }
 253
 254 static ssize_t enabled_store(struct kobject *kobj,
 255                              struct kobj_attribute *attr,
 256                              const char *buf, size_t count)
 257 {
 258         ssize_t ret = count;
 259
 260         if (sysfs_streq(buf, "always")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 263         } else if (sysfs_streq(buf, "madvise")) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 265                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 266         } else if (sysfs_streq(buf, "never")) {
 267                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 269         } else
 270                 ret = -EINVAL;
 271
 272         if (ret > 0) {
 273                 int err = start_stop_khugepaged();
 274                 if (err)
 275                         ret = err;
 276         }
 277         return ret;
 278 }
 279
 280 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 281
 282 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 283                                   struct kobj_attribute *attr, char *buf,
 284                                   enum transparent_hugepage_flag flag)
 285 {
 286         return sysfs_emit(buf, "%d\n",
 287                           !!test_bit(flag, &transparent_hugepage_flags));
 288 }
 289
 290 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 291                                  struct kobj_attribute *attr,
 292                                  const char *buf, size_t count,
 293                                  enum transparent_hugepage_flag flag)
 294 {
 295         unsigned long value;
 296         int ret;
 297
 298         ret = kstrtoul(buf, 10, &value);
 299         if (ret < 0)
 300                 return ret;
 301         if (value > 1)
 302                 return -EINVAL;
 303
 304         if (value)
 305                 set_bit(flag, &transparent_hugepage_flags);
 306         else
 307                 clear_bit(flag, &transparent_hugepage_flags);
 308
 309         return count;
 310 }
 311
 312 static ssize_t defrag_show(struct kobject *kobj,
 313                            struct kobj_attribute *attr, char *buf)
 314 {
 315         const char *output;
 316
 317         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 318                      &transparent_hugepage_flags))
 319                 output = "[always] defer defer+madvise madvise never";
 320         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 321                           &transparent_hugepage_flags))
 322                 output = "always [defer] defer+madvise madvise never";
 323         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 324                           &transparent_hugepage_flags))
 325                 output = "always defer [defer+madvise] madvise never";
 326         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 327                           &transparent_hugepage_flags))
 328                 output = "always defer defer+madvise [madvise] never";
 329         else
 330                 output = "always defer defer+madvise madvise [never]";
 331
 332         return sysfs_emit(buf, "%s\n", output);
 333 }
 334
 335 static ssize_t defrag_store(struct kobject *kobj,
 336                             struct kobj_attribute *attr,
 337                             const char *buf, size_t count)
 338 {
 339         if (sysfs_streq(buf, "always")) {
 340                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 341                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 342                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 343                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 344         } else if (sysfs_streq(buf, "defer+madvise")) {
 345                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 346                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 347                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 348                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 349         } else if (sysfs_streq(buf, "defer")) {
 350                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 351                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 352                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 353                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 354         } else if (sysfs_streq(buf, "madvise")) {
 355                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 356                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 357                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 358                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 359         } else if (sysfs_streq(buf, "never")) {
 360                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 361                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 362                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 363                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 364         } else
 365                 return -EINVAL;
 366
 367         return count;
 368 }
 369 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 370
 371 static ssize_t use_zero_page_show(struct kobject *kobj,
 372                                   struct kobj_attribute *attr, char *buf)
 373 {
 374         return single_hugepage_flag_show(kobj, attr, buf,
 375                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 376 }
 377 static ssize_t use_zero_page_store(struct kobject *kobj,
 378                 struct kobj_attribute *attr, const char *buf, size_t count)
 379 {
 380         return single_hugepage_flag_store(kobj, attr, buf, count,
 381                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 382 }
 383 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 384
 385 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 386                                    struct kobj_attribute *attr, char *buf)
 387 {
 388         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 389 }
 390 static struct kobj_attribute hpage_pmd_size_attr =
 391         __ATTR_RO(hpage_pmd_size);
 392
 393 static struct attribute *hugepage_attr[] = {
 394         &enabled_attr.attr,
 395         &defrag_attr.attr,
 396         &use_zero_page_attr.attr,
 397         &hpage_pmd_size_attr.attr,
 398 #ifdef CONFIG_SHMEM
 399         &shmem_enabled_attr.attr,
 400 #endif
 401         NULL,
 402 };
 403
 404 static const struct attribute_group hugepage_attr_group = {
 405         .attrs = hugepage_attr,
 406 };
 407
 408 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 409 {
 410         int err;
 411
 412         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 413         if (unlikely(!*hugepage_kobj)) {
 414                 pr_err("failed to create transparent hugepage kobject\n");
 415                 return -ENOMEM;
 416         }
 417
 418         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 419         if (err) {
 420                 pr_err("failed to register transparent hugepage group\n");
 421                 goto delete_obj;
 422         }
 423
 424         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 425         if (err) {
 426                 pr_err("failed to register transparent hugepage group\n");
 427                 goto remove_hp_group;
 428         }
 429
 430         return 0;
 431
 432 remove_hp_group:
 433         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 434 delete_obj:
 435         kobject_put(*hugepage_kobj);
 436         return err;
 437 }
 438
 439 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 440 {
 441         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 442         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 443         kobject_put(hugepage_kobj);
 444 }
 445 #else
 446 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 447 {
 448         return 0;
 449 }
 450
 451 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 452 {
 453 }
 454 #endif /* CONFIG_SYSFS */
 455
 456 static int __init hugepage_init(void)
 457 {
 458         int err;
 459         struct kobject *hugepage_kobj;
 460
 461         if (!has_transparent_hugepage()) {
 462                 /*
 463                  * Hardware doesn't support hugepages, hence disable
 464                  * DAX PMD support.
 465                  */
 466                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 467                 return -EINVAL;
 468         }
 469
 470         /*
 471          * hugepages can't be allocated by the buddy allocator
 472          */
 473         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 474         /*
 475          * we use page->mapping and page->index in second tail page
 476          * as list_head: assuming THP order >= 2
 477          */
 478         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 479
 480         err = hugepage_init_sysfs(&hugepage_kobj);
 481         if (err)
 482                 goto err_sysfs;
 483
 484         err = khugepaged_init();
 485         if (err)
 486                 goto err_slab;
 487
 488         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 489         if (err)
 490                 goto err_hzp_shrinker;
 491         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 492         if (err)
 493                 goto err_split_shrinker;
 494
 495         /*
 496          * By default disable transparent hugepages on smaller systems,
 497          * where the extra memory used could hurt more than TLB overhead
 498          * is likely to save.  The admin can still enable it through /sys.
 499          */
 500         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 501                 transparent_hugepage_flags = 0;
 502                 return 0;
 503         }
 504
 505         err = start_stop_khugepaged();
 506         if (err)
 507                 goto err_khugepaged;
 508
 509         return 0;
 510 err_khugepaged:
 511         unregister_shrinker(&deferred_split_shrinker);
 512 err_split_shrinker:
 513         unregister_shrinker(&huge_zero_page_shrinker);
 514 err_hzp_shrinker:
 515         khugepaged_destroy();
 516 err_slab:
 517         hugepage_exit_sysfs(hugepage_kobj);
 518 err_sysfs:
 519         return err;
 520 }
 521 subsys_initcall(hugepage_init);
 522
 523 static int __init setup_transparent_hugepage(char *str)
 524 {
 525         int ret = 0;
 526         if (!str)
 527                 goto out;
 528         if (!strcmp(str, "always")) {
 529                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 530                         &transparent_hugepage_flags);
 531                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 532                           &transparent_hugepage_flags);
 533                 ret = 1;
 534         } else if (!strcmp(str, "madvise")) {
 535                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 536                           &transparent_hugepage_flags);
 537                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 538                         &transparent_hugepage_flags);
 539                 ret = 1;
 540         } else if (!strcmp(str, "never")) {
 541                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 542                           &transparent_hugepage_flags);
 543                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 544                           &transparent_hugepage_flags);
 545                 ret = 1;
 546         }
 547 out:
 548         if (!ret)
 549                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 550         return ret;
 551 }
 552 __setup("transparent_hugepage=", setup_transparent_hugepage);
 553
 554 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 555 {
 556         if (likely(vma->vm_flags & VM_WRITE))
 557                 pmd = pmd_mkwrite(pmd);
 558         return pmd;
 559 }
 560
 561 #ifdef CONFIG_MEMCG
 562 static inline
 563 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 564 {
 565         struct mem_cgroup *memcg = folio_memcg(folio);
 566         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 567
 568         if (memcg)
 569                 return &memcg->deferred_split_queue;
 570         else
 571                 return &pgdat->deferred_split_queue;
 572 }
 573 #else
 574 static inline
 575 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 576 {
 577         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 578
 579         return &pgdat->deferred_split_queue;
 580 }
 581 #endif
 582
 583 void prep_transhuge_page(struct page *page)
 584 {
 585         struct folio *folio = (struct folio *)page;
 586
 587         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
 588         INIT_LIST_HEAD(&folio->_deferred_list);
 589         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 590 }
 591
 592 static inline bool is_transparent_hugepage(struct page *page)
 593 {
 594         struct folio *folio;
 595
 596         if (!PageCompound(page))
 597                 return false;
 598
 599         folio = page_folio(page);
 600         return is_huge_zero_page(&folio->page) ||
 601                folio->_folio_dtor == TRANSHUGE_PAGE_DTOR;
 602 }
 603
 604 static unsigned long __thp_get_unmapped_area(struct file *filp,
 605                 unsigned long addr, unsigned long len,
 606                 loff_t off, unsigned long flags, unsigned long size)
 607 {
 608         loff_t off_end = off + len;
 609         loff_t off_align = round_up(off, size);
 610         unsigned long len_pad, ret;
 611
 612         if (off_end <= off_align || (off_end - off_align) < size)
 613                 return 0;
 614
 615         len_pad = len + size;
 616         if (len_pad < len || (off + len_pad) < off)
 617                 return 0;
 618
 619         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 620                                               off >> PAGE_SHIFT, flags);
 621
 622         /*
 623          * The failure might be due to length padding. The caller will retry
 624          * without the padding.
 625          */
 626         if (IS_ERR_VALUE(ret))
 627                 return 0;
 628
 629         /*
 630          * Do not try to align to THP boundary if allocation at the address
 631          * hint succeeds.
 632          */
 633         if (ret == addr)
 634                 return addr;
 635
 636         ret += (off - ret) & (size - 1);
 637         return ret;
 638 }
 639
 640 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 641                 unsigned long len, unsigned long pgoff, unsigned long flags)
 642 {
 643         unsigned long ret;
 644         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 645
 646         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 647         if (ret)
 648                 return ret;
 649
 650         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 651 }
 652 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 653
 654 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 655                         struct page *page, gfp_t gfp)
 656 {
 657         struct vm_area_struct *vma = vmf->vma;
 658         pgtable_t pgtable;
 659         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 660         vm_fault_t ret = 0;
 661
 662         VM_BUG_ON_PAGE(!PageCompound(page), page);
 663
 664         if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
 665                 put_page(page);
 666                 count_vm_event(THP_FAULT_FALLBACK);
 667                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 668                 return VM_FAULT_FALLBACK;
 669         }
 670         cgroup_throttle_swaprate(page, gfp);
 671
 672         pgtable = pte_alloc_one(vma->vm_mm);
 673         if (unlikely(!pgtable)) {
 674                 ret = VM_FAULT_OOM;
 675                 goto release;
 676         }
 677
 678         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 679         /*
 680          * The memory barrier inside __SetPageUptodate makes sure that
 681          * clear_huge_page writes become visible before the set_pmd_at()
 682          * write.
 683          */
 684         __SetPageUptodate(page);
 685
 686         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 687         if (unlikely(!pmd_none(*vmf->pmd))) {
 688                 goto unlock_release;
 689         } else {
 690                 pmd_t entry;
 691
 692                 ret = check_stable_address_space(vma->vm_mm);
 693                 if (ret)
 694                         goto unlock_release;
 695
 696                 /* Deliver the page fault to userland */
 697                 if (userfaultfd_missing(vma)) {
 698                         spin_unlock(vmf->ptl);
 699                         put_page(page);
 700                         pte_free(vma->vm_mm, pgtable);
 701                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 702                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 703                         return ret;
 704                 }
 705
 706                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 707                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 708                 page_add_new_anon_rmap(page, vma, haddr);
 709                 lru_cache_add_inactive_or_unevictable(page, vma);
 710                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 711                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 712                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 713                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 714                 mm_inc_nr_ptes(vma->vm_mm);
 715                 spin_unlock(vmf->ptl);
 716                 count_vm_event(THP_FAULT_ALLOC);
 717                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 718         }
 719
 720         return 0;
 721 unlock_release:
 722         spin_unlock(vmf->ptl);
 723 release:
 724         if (pgtable)
 725                 pte_free(vma->vm_mm, pgtable);
 726         put_page(page);
 727         return ret;
 728
 729 }
 730
 731 /*
 732  * always: directly stall for all thp allocations
 733  * defer: wake kswapd and fail if not immediately available
 734  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 735  *                fail if not immediately available
 736  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 737  *          available
 738  * never: never stall for any thp allocation
 739  */
 740 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 741 {
 742         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 743
 744         /* Always do synchronous compaction */
 745         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 746                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 747
 748         /* Kick kcompactd and fail quickly */
 749         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 750                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 751
 752         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 753         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 754                 return GFP_TRANSHUGE_LIGHT |
 755                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 756                                         __GFP_KSWAPD_RECLAIM);
 757
 758         /* Only do synchronous compaction if madvised */
 759         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 760                 return GFP_TRANSHUGE_LIGHT |
 761                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 762
 763         return GFP_TRANSHUGE_LIGHT;
 764 }
 765
 766 /* Caller must hold page table lock. */
 767 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 768                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 769                 struct page *zero_page)
 770 {
 771         pmd_t entry;
 772         if (!pmd_none(*pmd))
 773                 return;
 774         entry = mk_pmd(zero_page, vma->vm_page_prot);
 775         entry = pmd_mkhuge(entry);
 776         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 777         set_pmd_at(mm, haddr, pmd, entry);
 778         mm_inc_nr_ptes(mm);
 779 }
 780
 781 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 782 {
 783         struct vm_area_struct *vma = vmf->vma;
 784         gfp_t gfp;
 785         struct folio *folio;
 786         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 787
 788         if (!transhuge_vma_suitable(vma, haddr))
 789                 return VM_FAULT_FALLBACK;
 790         if (unlikely(anon_vma_prepare(vma)))
 791                 return VM_FAULT_OOM;
 792         khugepaged_enter_vma(vma, vma->vm_flags);
 793
 794         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 795                         !mm_forbids_zeropage(vma->vm_mm) &&
 796                         transparent_hugepage_use_zero_page()) {
 797                 pgtable_t pgtable;
 798                 struct page *zero_page;
 799                 vm_fault_t ret;
 800                 pgtable = pte_alloc_one(vma->vm_mm);
 801                 if (unlikely(!pgtable))
 802                         return VM_FAULT_OOM;
 803                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 804                 if (unlikely(!zero_page)) {
 805                         pte_free(vma->vm_mm, pgtable);
 806                         count_vm_event(THP_FAULT_FALLBACK);
 807                         return VM_FAULT_FALLBACK;
 808                 }
 809                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 810                 ret = 0;
 811                 if (pmd_none(*vmf->pmd)) {
 812                         ret = check_stable_address_space(vma->vm_mm);
 813                         if (ret) {
 814                                 spin_unlock(vmf->ptl);
 815                                 pte_free(vma->vm_mm, pgtable);
 816                         } else if (userfaultfd_missing(vma)) {
 817                                 spin_unlock(vmf->ptl);
 818                                 pte_free(vma->vm_mm, pgtable);
 819                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 820                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 821                         } else {
 822                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 823                                                    haddr, vmf->pmd, zero_page);
 824                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 825                                 spin_unlock(vmf->ptl);
 826                         }
 827                 } else {
 828                         spin_unlock(vmf->ptl);
 829                         pte_free(vma->vm_mm, pgtable);
 830                 }
 831                 return ret;
 832         }
 833         gfp = vma_thp_gfp_mask(vma);
 834         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 835         if (unlikely(!folio)) {
 836                 count_vm_event(THP_FAULT_FALLBACK);
 837                 return VM_FAULT_FALLBACK;
 838         }
 839         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 840 }
 841
 842 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 843                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 844                 pgtable_t pgtable)
 845 {
 846         struct mm_struct *mm = vma->vm_mm;
 847         pmd_t entry;
 848         spinlock_t *ptl;
 849
 850         ptl = pmd_lock(mm, pmd);
 851         if (!pmd_none(*pmd)) {
 852                 if (write) {
 853                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 854                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 855                                 goto out_unlock;
 856                         }
 857                         entry = pmd_mkyoung(*pmd);
 858                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 859                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 860                                 update_mmu_cache_pmd(vma, addr, pmd);
 861                 }
 862
 863                 goto out_unlock;
 864         }
 865
 866         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 867         if (pfn_t_devmap(pfn))
 868                 entry = pmd_mkdevmap(entry);
 869         if (write) {
 870                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 871                 entry = maybe_pmd_mkwrite(entry, vma);
 872         }
 873
 874         if (pgtable) {
 875                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 876                 mm_inc_nr_ptes(mm);
 877                 pgtable = NULL;
 878         }
 879
 880         set_pmd_at(mm, addr, pmd, entry);
 881         update_mmu_cache_pmd(vma, addr, pmd);
 882
 883 out_unlock:
 884         spin_unlock(ptl);
 885         if (pgtable)
 886                 pte_free(mm, pgtable);
 887 }
 888
 889 /**
 890  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 891  * @vmf: Structure describing the fault
 892  * @pfn: pfn to insert
 893  * @pgprot: page protection to use
 894  * @write: whether it's a write fault
 895  *
 896  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 897  * also consult the vmf_insert_mixed_prot() documentation when
 898  * @pgprot != @vmf->vma->vm_page_prot.
 899  *
 900  * Return: vm_fault_t value.
 901  */
 902 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 903                                    pgprot_t pgprot, bool write)
 904 {
 905         unsigned long addr = vmf->address & PMD_MASK;
 906         struct vm_area_struct *vma = vmf->vma;
 907         pgtable_t pgtable = NULL;
 908
 909         /*
 910          * If we had pmd_special, we could avoid all these restrictions,
 911          * but we need to be consistent with PTEs and architectures that
 912          * can't support a 'special' bit.
 913          */
 914         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 915                         !pfn_t_devmap(pfn));
 916         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 917                                                 (VM_PFNMAP|VM_MIXEDMAP));
 918         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 919
 920         if (addr < vma->vm_start || addr >= vma->vm_end)
 921                 return VM_FAULT_SIGBUS;
 922
 923         if (arch_needs_pgtable_deposit()) {
 924                 pgtable = pte_alloc_one(vma->vm_mm);
 925                 if (!pgtable)
 926                         return VM_FAULT_OOM;
 927         }
 928
 929         track_pfn_insert(vma, &pgprot, pfn);
 930
 931         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 932         return VM_FAULT_NOPAGE;
 933 }
 934 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 935
 936 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 937 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 938 {
 939         if (likely(vma->vm_flags & VM_WRITE))
 940                 pud = pud_mkwrite(pud);
 941         return pud;
 942 }
 943
 944 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 945                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 946 {
 947         struct mm_struct *mm = vma->vm_mm;
 948         pud_t entry;
 949         spinlock_t *ptl;
 950
 951         ptl = pud_lock(mm, pud);
 952         if (!pud_none(*pud)) {
 953                 if (write) {
 954                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 955                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 956                                 goto out_unlock;
 957                         }
 958                         entry = pud_mkyoung(*pud);
 959                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 960                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 961                                 update_mmu_cache_pud(vma, addr, pud);
 962                 }
 963                 goto out_unlock;
 964         }
 965
 966         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 967         if (pfn_t_devmap(pfn))
 968                 entry = pud_mkdevmap(entry);
 969         if (write) {
 970                 entry = pud_mkyoung(pud_mkdirty(entry));
 971                 entry = maybe_pud_mkwrite(entry, vma);
 972         }
 973         set_pud_at(mm, addr, pud, entry);
 974         update_mmu_cache_pud(vma, addr, pud);
 975
 976 out_unlock:
 977         spin_unlock(ptl);
 978 }
 979
 980 /**
 981  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 982  * @vmf: Structure describing the fault
 983  * @pfn: pfn to insert
 984  * @pgprot: page protection to use
 985  * @write: whether it's a write fault
 986  *
 987  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 988  * also consult the vmf_insert_mixed_prot() documentation when
 989  * @pgprot != @vmf->vma->vm_page_prot.
 990  *
 991  * Return: vm_fault_t value.
 992  */
 993 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 994                                    pgprot_t pgprot, bool write)
 995 {
 996         unsigned long addr = vmf->address & PUD_MASK;
 997         struct vm_area_struct *vma = vmf->vma;
 998
 999         /*
1000          * If we had pud_special, we could avoid all these restrictions,
1001          * but we need to be consistent with PTEs and architectures that
1002          * can't support a 'special' bit.
1003          */
1004         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1005                         !pfn_t_devmap(pfn));
1006         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1007                                                 (VM_PFNMAP|VM_MIXEDMAP));
1008         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1009
1010         if (addr < vma->vm_start || addr >= vma->vm_end)
1011                 return VM_FAULT_SIGBUS;
1012
1013         track_pfn_insert(vma, &pgprot, pfn);
1014
1015         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1016         return VM_FAULT_NOPAGE;
1017 }
1018 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1019 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1020
1021 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1022                       pmd_t *pmd, bool write)
1023 {
1024         pmd_t _pmd;
1025
1026         _pmd = pmd_mkyoung(*pmd);
1027         if (write)
1028                 _pmd = pmd_mkdirty(_pmd);
1029         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1030                                   pmd, _pmd, write))
1031                 update_mmu_cache_pmd(vma, addr, pmd);
1032 }
1033
1034 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1035                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1036 {
1037         unsigned long pfn = pmd_pfn(*pmd);
1038         struct mm_struct *mm = vma->vm_mm;
1039         struct page *page;
1040         int ret;
1041
1042         assert_spin_locked(pmd_lockptr(mm, pmd));
1043
1044         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1045         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1046                          (FOLL_PIN | FOLL_GET)))
1047                 return NULL;
1048
1049         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1050                 return NULL;
1051
1052         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1053                 /* pass */;
1054         else
1055                 return NULL;
1056
1057         if (flags & FOLL_TOUCH)
1058                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1059
1060         /*
1061          * device mapped pages can only be returned if the
1062          * caller will manage the page reference count.
1063          */
1064         if (!(flags & (FOLL_GET | FOLL_PIN)))
1065                 return ERR_PTR(-EEXIST);
1066
1067         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1068         *pgmap = get_dev_pagemap(pfn, *pgmap);
1069         if (!*pgmap)
1070                 return ERR_PTR(-EFAULT);
1071         page = pfn_to_page(pfn);
1072         ret = try_grab_page(page, flags);
1073         if (ret)
1074                 page = ERR_PTR(ret);
1075
1076         return page;
1077 }
1078
1079 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1080                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1081                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1082 {
1083         spinlock_t *dst_ptl, *src_ptl;
1084         struct page *src_page;
1085         pmd_t pmd;
1086         pgtable_t pgtable = NULL;
1087         int ret = -ENOMEM;
1088
1089         /* Skip if can be re-fill on fault */
1090         if (!vma_is_anonymous(dst_vma))
1091                 return 0;
1092
1093         pgtable = pte_alloc_one(dst_mm);
1094         if (unlikely(!pgtable))
1095                 goto out;
1096
1097         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1098         src_ptl = pmd_lockptr(src_mm, src_pmd);
1099         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1100
1101         ret = -EAGAIN;
1102         pmd = *src_pmd;
1103
1104 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1105         if (unlikely(is_swap_pmd(pmd))) {
1106                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1107
1108                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1109                 if (!is_readable_migration_entry(entry)) {
1110                         entry = make_readable_migration_entry(
1111                                                         swp_offset(entry));
1112                         pmd = swp_entry_to_pmd(entry);
1113                         if (pmd_swp_soft_dirty(*src_pmd))
1114                                 pmd = pmd_swp_mksoft_dirty(pmd);
1115                         if (pmd_swp_uffd_wp(*src_pmd))
1116                                 pmd = pmd_swp_mkuffd_wp(pmd);
1117                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1118                 }
1119                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1120                 mm_inc_nr_ptes(dst_mm);
1121                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1122                 if (!userfaultfd_wp(dst_vma))
1123                         pmd = pmd_swp_clear_uffd_wp(pmd);
1124                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1125                 ret = 0;
1126                 goto out_unlock;
1127         }
1128 #endif
1129
1130         if (unlikely(!pmd_trans_huge(pmd))) {
1131                 pte_free(dst_mm, pgtable);
1132                 goto out_unlock;
1133         }
1134         /*
1135          * When page table lock is held, the huge zero pmd should not be
1136          * under splitting since we don't split the page itself, only pmd to
1137          * a page table.
1138          */
1139         if (is_huge_zero_pmd(pmd)) {
1140                 /*
1141                  * get_huge_zero_page() will never allocate a new page here,
1142                  * since we already have a zero page to copy. It just takes a
1143                  * reference.
1144                  */
1145                 mm_get_huge_zero_page(dst_mm);
1146                 goto out_zero_page;
1147         }
1148
1149         src_page = pmd_page(pmd);
1150         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1151
1152         get_page(src_page);
1153         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1154                 /* Page maybe pinned: split and retry the fault on PTEs. */
1155                 put_page(src_page);
1156                 pte_free(dst_mm, pgtable);
1157                 spin_unlock(src_ptl);
1158                 spin_unlock(dst_ptl);
1159                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1160                 return -EAGAIN;
1161         }
1162         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1163 out_zero_page:
1164         mm_inc_nr_ptes(dst_mm);
1165         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1166         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1167         if (!userfaultfd_wp(dst_vma))
1168                 pmd = pmd_clear_uffd_wp(pmd);
1169         pmd = pmd_mkold(pmd_wrprotect(pmd));
1170         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1171
1172         ret = 0;
1173 out_unlock:
1174         spin_unlock(src_ptl);
1175         spin_unlock(dst_ptl);
1176 out:
1177         return ret;
1178 }
1179
1180 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1181 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1182                       pud_t *pud, bool write)
1183 {
1184         pud_t _pud;
1185
1186         _pud = pud_mkyoung(*pud);
1187         if (write)
1188                 _pud = pud_mkdirty(_pud);
1189         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1190                                   pud, _pud, write))
1191                 update_mmu_cache_pud(vma, addr, pud);
1192 }
1193
1194 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1195                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1196 {
1197         unsigned long pfn = pud_pfn(*pud);
1198         struct mm_struct *mm = vma->vm_mm;
1199         struct page *page;
1200         int ret;
1201
1202         assert_spin_locked(pud_lockptr(mm, pud));
1203
1204         if (flags & FOLL_WRITE && !pud_write(*pud))
1205                 return NULL;
1206
1207         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1208         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1209                          (FOLL_PIN | FOLL_GET)))
1210                 return NULL;
1211
1212         if (pud_present(*pud) && pud_devmap(*pud))
1213                 /* pass */;
1214         else
1215                 return NULL;
1216
1217         if (flags & FOLL_TOUCH)
1218                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1219
1220         /*
1221          * device mapped pages can only be returned if the
1222          * caller will manage the page reference count.
1223          *
1224          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1225          */
1226         if (!(flags & (FOLL_GET | FOLL_PIN)))
1227                 return ERR_PTR(-EEXIST);
1228
1229         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1230         *pgmap = get_dev_pagemap(pfn, *pgmap);
1231         if (!*pgmap)
1232                 return ERR_PTR(-EFAULT);
1233         page = pfn_to_page(pfn);
1234
1235         ret = try_grab_page(page, flags);
1236         if (ret)
1237                 page = ERR_PTR(ret);
1238
1239         return page;
1240 }
1241
1242 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1243                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1244                   struct vm_area_struct *vma)
1245 {
1246         spinlock_t *dst_ptl, *src_ptl;
1247         pud_t pud;
1248         int ret;
1249
1250         dst_ptl = pud_lock(dst_mm, dst_pud);
1251         src_ptl = pud_lockptr(src_mm, src_pud);
1252         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1253
1254         ret = -EAGAIN;
1255         pud = *src_pud;
1256         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1257                 goto out_unlock;
1258
1259         /*
1260          * When page table lock is held, the huge zero pud should not be
1261          * under splitting since we don't split the page itself, only pud to
1262          * a page table.
1263          */
1264         if (is_huge_zero_pud(pud)) {
1265                 /* No huge zero pud yet */
1266         }
1267
1268         /*
1269          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1270          * and split if duplicating fails.
1271          */
1272         pudp_set_wrprotect(src_mm, addr, src_pud);
1273         pud = pud_mkold(pud_wrprotect(pud));
1274         set_pud_at(dst_mm, addr, dst_pud, pud);
1275
1276         ret = 0;
1277 out_unlock:
1278         spin_unlock(src_ptl);
1279         spin_unlock(dst_ptl);
1280         return ret;
1281 }
1282
1283 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1284 {
1285         bool write = vmf->flags & FAULT_FLAG_WRITE;
1286
1287         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1288         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1289                 goto unlock;
1290
1291         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1292 unlock:
1293         spin_unlock(vmf->ptl);
1294 }
1295 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1296
1297 void huge_pmd_set_accessed(struct vm_fault *vmf)
1298 {
1299         bool write = vmf->flags & FAULT_FLAG_WRITE;
1300
1301         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1302         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1303                 goto unlock;
1304
1305         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1306
1307 unlock:
1308         spin_unlock(vmf->ptl);
1309 }
1310
1311 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1312 {
1313         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1314         struct vm_area_struct *vma = vmf->vma;
1315         struct folio *folio;
1316         struct page *page;
1317         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1318         pmd_t orig_pmd = vmf->orig_pmd;
1319
1320         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1321         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1322
1323         if (is_huge_zero_pmd(orig_pmd))
1324                 goto fallback;
1325
1326         spin_lock(vmf->ptl);
1327
1328         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1329                 spin_unlock(vmf->ptl);
1330                 return 0;
1331         }
1332
1333         page = pmd_page(orig_pmd);
1334         folio = page_folio(page);
1335         VM_BUG_ON_PAGE(!PageHead(page), page);
1336
1337         /* Early check when only holding the PT lock. */
1338         if (PageAnonExclusive(page))
1339                 goto reuse;
1340
1341         if (!folio_trylock(folio)) {
1342                 folio_get(folio);
1343                 spin_unlock(vmf->ptl);
1344                 folio_lock(folio);
1345                 spin_lock(vmf->ptl);
1346                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1347                         spin_unlock(vmf->ptl);
1348                         folio_unlock(folio);
1349                         folio_put(folio);
1350                         return 0;
1351                 }
1352                 folio_put(folio);
1353         }
1354
1355         /* Recheck after temporarily dropping the PT lock. */
1356         if (PageAnonExclusive(page)) {
1357                 folio_unlock(folio);
1358                 goto reuse;
1359         }
1360
1361         /*
1362          * See do_wp_page(): we can only reuse the folio exclusively if
1363          * there are no additional references. Note that we always drain
1364          * the LRU pagevecs immediately after adding a THP.
1365          */
1366         if (folio_ref_count(folio) >
1367                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1368                 goto unlock_fallback;
1369         if (folio_test_swapcache(folio))
1370                 folio_free_swap(folio);
1371         if (folio_ref_count(folio) == 1) {
1372                 pmd_t entry;
1373
1374                 page_move_anon_rmap(page, vma);
1375                 folio_unlock(folio);
1376 reuse:
1377                 if (unlikely(unshare)) {
1378                         spin_unlock(vmf->ptl);
1379                         return 0;
1380                 }
1381                 entry = pmd_mkyoung(orig_pmd);
1382                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1383                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1384                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1385                 spin_unlock(vmf->ptl);
1386                 return 0;
1387         }
1388
1389 unlock_fallback:
1390         folio_unlock(folio);
1391         spin_unlock(vmf->ptl);
1392 fallback:
1393         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1394         return VM_FAULT_FALLBACK;
1395 }
1396
1397 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1398                                            unsigned long addr, pmd_t pmd)
1399 {
1400         struct page *page;
1401
1402         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1403                 return false;
1404
1405         /* Don't touch entries that are not even readable (NUMA hinting). */
1406         if (pmd_protnone(pmd))
1407                 return false;
1408
1409         /* Do we need write faults for softdirty tracking? */
1410         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1411                 return false;
1412
1413         /* Do we need write faults for uffd-wp tracking? */
1414         if (userfaultfd_huge_pmd_wp(vma, pmd))
1415                 return false;
1416
1417         if (!(vma->vm_flags & VM_SHARED)) {
1418                 /* See can_change_pte_writable(). */
1419                 page = vm_normal_page_pmd(vma, addr, pmd);
1420                 return page && PageAnon(page) && PageAnonExclusive(page);
1421         }
1422
1423         /* See can_change_pte_writable(). */
1424         return pmd_dirty(pmd);
1425 }
1426
1427 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1428 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1429                                         struct vm_area_struct *vma,
1430                                         unsigned int flags)
1431 {
1432         /* If the pmd is writable, we can write to the page. */
1433         if (pmd_write(pmd))
1434                 return true;
1435
1436         /* Maybe FOLL_FORCE is set to override it? */
1437         if (!(flags & FOLL_FORCE))
1438                 return false;
1439
1440         /* But FOLL_FORCE has no effect on shared mappings */
1441         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1442                 return false;
1443
1444         /* ... or read-only private ones */
1445         if (!(vma->vm_flags & VM_MAYWRITE))
1446                 return false;
1447
1448         /* ... or already writable ones that just need to take a write fault */
1449         if (vma->vm_flags & VM_WRITE)
1450                 return false;
1451
1452         /*
1453          * See can_change_pte_writable(): we broke COW and could map the page
1454          * writable if we have an exclusive anonymous page ...
1455          */
1456         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1457                 return false;
1458
1459         /* ... and a write-fault isn't required for other reasons. */
1460         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1461                 return false;
1462         return !userfaultfd_huge_pmd_wp(vma, pmd);
1463 }
1464
1465 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1466                                    unsigned long addr,
1467                                    pmd_t *pmd,
1468                                    unsigned int flags)
1469 {
1470         struct mm_struct *mm = vma->vm_mm;
1471         struct page *page;
1472         int ret;
1473
1474         assert_spin_locked(pmd_lockptr(mm, pmd));
1475
1476         page = pmd_page(*pmd);
1477         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1478
1479         if ((flags & FOLL_WRITE) &&
1480             !can_follow_write_pmd(*pmd, page, vma, flags))
1481                 return NULL;
1482
1483         /* Avoid dumping huge zero page */
1484         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1485                 return ERR_PTR(-EFAULT);
1486
1487         /* Full NUMA hinting faults to serialise migration in fault paths */
1488         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1489                 return NULL;
1490
1491         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1492                 return ERR_PTR(-EMLINK);
1493
1494         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1495                         !PageAnonExclusive(page), page);
1496
1497         ret = try_grab_page(page, flags);
1498         if (ret)
1499                 return ERR_PTR(ret);
1500
1501         if (flags & FOLL_TOUCH)
1502                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1503
1504         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1505         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1506
1507         return page;
1508 }
1509
1510 /* NUMA hinting page fault entry point for trans huge pmds */
1511 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1512 {
1513         struct vm_area_struct *vma = vmf->vma;
1514         pmd_t oldpmd = vmf->orig_pmd;
1515         pmd_t pmd;
1516         struct page *page;
1517         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1518         int page_nid = NUMA_NO_NODE;
1519         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1520         bool migrated = false, writable = false;
1521         int flags = 0;
1522
1523         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1524         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1525                 spin_unlock(vmf->ptl);
1526                 goto out;
1527         }
1528
1529         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1530
1531         /*
1532          * Detect now whether the PMD could be writable; this information
1533          * is only valid while holding the PT lock.
1534          */
1535         writable = pmd_write(pmd);
1536         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1537             can_change_pmd_writable(vma, vmf->address, pmd))
1538                 writable = true;
1539
1540         page = vm_normal_page_pmd(vma, haddr, pmd);
1541         if (!page)
1542                 goto out_map;
1543
1544         /* See similar comment in do_numa_page for explanation */
1545         if (!writable)
1546                 flags |= TNF_NO_GROUP;
1547
1548         page_nid = page_to_nid(page);
1549         /*
1550          * For memory tiering mode, cpupid of slow memory page is used
1551          * to record page access time.  So use default value.
1552          */
1553         if (node_is_toptier(page_nid))
1554                 last_cpupid = page_cpupid_last(page);
1555         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1556                                        &flags);
1557
1558         if (target_nid == NUMA_NO_NODE) {
1559                 put_page(page);
1560                 goto out_map;
1561         }
1562
1563         spin_unlock(vmf->ptl);
1564         writable = false;
1565
1566         migrated = migrate_misplaced_page(page, vma, target_nid);
1567         if (migrated) {
1568                 flags |= TNF_MIGRATED;
1569                 page_nid = target_nid;
1570         } else {
1571                 flags |= TNF_MIGRATE_FAIL;
1572                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1573                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1574                         spin_unlock(vmf->ptl);
1575                         goto out;
1576                 }
1577                 goto out_map;
1578         }
1579
1580 out:
1581         if (page_nid != NUMA_NO_NODE)
1582                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1583                                 flags);
1584
1585         return 0;
1586
1587 out_map:
1588         /* Restore the PMD */
1589         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1590         pmd = pmd_mkyoung(pmd);
1591         if (writable)
1592                 pmd = pmd_mkwrite(pmd);
1593         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1594         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1595         spin_unlock(vmf->ptl);
1596         goto out;
1597 }
1598
1599 /*
1600  * Return true if we do MADV_FREE successfully on entire pmd page.
1601  * Otherwise, return false.
1602  */
1603 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1604                 pmd_t *pmd, unsigned long addr, unsigned long next)
1605 {
1606         spinlock_t *ptl;
1607         pmd_t orig_pmd;
1608         struct folio *folio;
1609         struct mm_struct *mm = tlb->mm;
1610         bool ret = false;
1611
1612         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1613
1614         ptl = pmd_trans_huge_lock(pmd, vma);
1615         if (!ptl)
1616                 goto out_unlocked;
1617
1618         orig_pmd = *pmd;
1619         if (is_huge_zero_pmd(orig_pmd))
1620                 goto out;
1621
1622         if (unlikely(!pmd_present(orig_pmd))) {
1623                 VM_BUG_ON(thp_migration_supported() &&
1624                                   !is_pmd_migration_entry(orig_pmd));
1625                 goto out;
1626         }
1627
1628         folio = pfn_folio(pmd_pfn(orig_pmd));
1629         /*
1630          * If other processes are mapping this folio, we couldn't discard
1631          * the folio unless they all do MADV_FREE so let's skip the folio.
1632          */
1633         if (folio_mapcount(folio) != 1)
1634                 goto out;
1635
1636         if (!folio_trylock(folio))
1637                 goto out;
1638
1639         /*
1640          * If user want to discard part-pages of THP, split it so MADV_FREE
1641          * will deactivate only them.
1642          */
1643         if (next - addr != HPAGE_PMD_SIZE) {
1644                 folio_get(folio);
1645                 spin_unlock(ptl);
1646                 split_folio(folio);
1647                 folio_unlock(folio);
1648                 folio_put(folio);
1649                 goto out_unlocked;
1650         }
1651
1652         if (folio_test_dirty(folio))
1653                 folio_clear_dirty(folio);
1654         folio_unlock(folio);
1655
1656         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1657                 pmdp_invalidate(vma, addr, pmd);
1658                 orig_pmd = pmd_mkold(orig_pmd);
1659                 orig_pmd = pmd_mkclean(orig_pmd);
1660
1661                 set_pmd_at(mm, addr, pmd, orig_pmd);
1662                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1663         }
1664
1665         folio_mark_lazyfree(folio);
1666         ret = true;
1667 out:
1668         spin_unlock(ptl);
1669 out_unlocked:
1670         return ret;
1671 }
1672
1673 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1674 {
1675         pgtable_t pgtable;
1676
1677         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678         pte_free(mm, pgtable);
1679         mm_dec_nr_ptes(mm);
1680 }
1681
1682 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1683                  pmd_t *pmd, unsigned long addr)
1684 {
1685         pmd_t orig_pmd;
1686         spinlock_t *ptl;
1687
1688         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1689
1690         ptl = __pmd_trans_huge_lock(pmd, vma);
1691         if (!ptl)
1692                 return 0;
1693         /*
1694          * For architectures like ppc64 we look at deposited pgtable
1695          * when calling pmdp_huge_get_and_clear. So do the
1696          * pgtable_trans_huge_withdraw after finishing pmdp related
1697          * operations.
1698          */
1699         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1700                                                 tlb->fullmm);
1701         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1702         if (vma_is_special_huge(vma)) {
1703                 if (arch_needs_pgtable_deposit())
1704                         zap_deposited_table(tlb->mm, pmd);
1705                 spin_unlock(ptl);
1706         } else if (is_huge_zero_pmd(orig_pmd)) {
1707                 zap_deposited_table(tlb->mm, pmd);
1708                 spin_unlock(ptl);
1709         } else {
1710                 struct page *page = NULL;
1711                 int flush_needed = 1;
1712
1713                 if (pmd_present(orig_pmd)) {
1714                         page = pmd_page(orig_pmd);
1715                         page_remove_rmap(page, vma, true);
1716                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1717                         VM_BUG_ON_PAGE(!PageHead(page), page);
1718                 } else if (thp_migration_supported()) {
1719                         swp_entry_t entry;
1720
1721                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1722                         entry = pmd_to_swp_entry(orig_pmd);
1723                         page = pfn_swap_entry_to_page(entry);
1724                         flush_needed = 0;
1725                 } else
1726                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1727
1728                 if (PageAnon(page)) {
1729                         zap_deposited_table(tlb->mm, pmd);
1730                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1731                 } else {
1732                         if (arch_needs_pgtable_deposit())
1733                                 zap_deposited_table(tlb->mm, pmd);
1734                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1735                 }
1736
1737                 spin_unlock(ptl);
1738                 if (flush_needed)
1739                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1740         }
1741         return 1;
1742 }
1743
1744 #ifndef pmd_move_must_withdraw
1745 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1746                                          spinlock_t *old_pmd_ptl,
1747                                          struct vm_area_struct *vma)
1748 {
1749         /*
1750          * With split pmd lock we also need to move preallocated
1751          * PTE page table if new_pmd is on different PMD page table.
1752          *
1753          * We also don't deposit and withdraw tables for file pages.
1754          */
1755         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1756 }
1757 #endif
1758
1759 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1760 {
1761 #ifdef CONFIG_MEM_SOFT_DIRTY
1762         if (unlikely(is_pmd_migration_entry(pmd)))
1763                 pmd = pmd_swp_mksoft_dirty(pmd);
1764         else if (pmd_present(pmd))
1765                 pmd = pmd_mksoft_dirty(pmd);
1766 #endif
1767         return pmd;
1768 }
1769
1770 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1771                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1772 {
1773         spinlock_t *old_ptl, *new_ptl;
1774         pmd_t pmd;
1775         struct mm_struct *mm = vma->vm_mm;
1776         bool force_flush = false;
1777
1778         /*
1779          * The destination pmd shouldn't be established, free_pgtables()
1780          * should have release it.
1781          */
1782         if (WARN_ON(!pmd_none(*new_pmd))) {
1783                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1784                 return false;
1785         }
1786
1787         /*
1788          * We don't have to worry about the ordering of src and dst
1789          * ptlocks because exclusive mmap_lock prevents deadlock.
1790          */
1791         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1792         if (old_ptl) {
1793                 new_ptl = pmd_lockptr(mm, new_pmd);
1794                 if (new_ptl != old_ptl)
1795                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1796                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1797                 if (pmd_present(pmd))
1798                         force_flush = true;
1799                 VM_BUG_ON(!pmd_none(*new_pmd));
1800
1801                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1802                         pgtable_t pgtable;
1803                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1804                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1805                 }
1806                 pmd = move_soft_dirty_pmd(pmd);
1807                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1808                 if (force_flush)
1809                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1810                 if (new_ptl != old_ptl)
1811                         spin_unlock(new_ptl);
1812                 spin_unlock(old_ptl);
1813                 return true;
1814         }
1815         return false;
1816 }
1817
1818 /*
1819  * Returns
1820  *  - 0 if PMD could not be locked
1821  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1822  *      or if prot_numa but THP migration is not supported
1823  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1824  */
1825 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1826                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1827                     unsigned long cp_flags)
1828 {
1829         struct mm_struct *mm = vma->vm_mm;
1830         spinlock_t *ptl;
1831         pmd_t oldpmd, entry;
1832         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1833         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1834         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1835         int ret = 1;
1836
1837         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1838
1839         if (prot_numa && !thp_migration_supported())
1840                 return 1;
1841
1842         ptl = __pmd_trans_huge_lock(pmd, vma);
1843         if (!ptl)
1844                 return 0;
1845
1846 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1847         if (is_swap_pmd(*pmd)) {
1848                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1849                 struct page *page = pfn_swap_entry_to_page(entry);
1850
1851                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1852                 if (is_writable_migration_entry(entry)) {
1853                         pmd_t newpmd;
1854                         /*
1855                          * A protection check is difficult so
1856                          * just be safe and disable write
1857                          */
1858                         if (PageAnon(page))
1859                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1860                         else
1861                                 entry = make_readable_migration_entry(swp_offset(entry));
1862                         newpmd = swp_entry_to_pmd(entry);
1863                         if (pmd_swp_soft_dirty(*pmd))
1864                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1865                         if (pmd_swp_uffd_wp(*pmd))
1866                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1867                         set_pmd_at(mm, addr, pmd, newpmd);
1868                 }
1869                 goto unlock;
1870         }
1871 #endif
1872
1873         if (prot_numa) {
1874                 struct page *page;
1875                 bool toptier;
1876                 /*
1877                  * Avoid trapping faults against the zero page. The read-only
1878                  * data is likely to be read-cached on the local CPU and
1879                  * local/remote hits to the zero page are not interesting.
1880                  */
1881                 if (is_huge_zero_pmd(*pmd))
1882                         goto unlock;
1883
1884                 if (pmd_protnone(*pmd))
1885                         goto unlock;
1886
1887                 page = pmd_page(*pmd);
1888                 toptier = node_is_toptier(page_to_nid(page));
1889                 /*
1890                  * Skip scanning top tier node if normal numa
1891                  * balancing is disabled
1892                  */
1893                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1894                     toptier)
1895                         goto unlock;
1896
1897                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1898                     !toptier)
1899                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1900         }
1901         /*
1902          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1903          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1904          * which is also under mmap_read_lock(mm):
1905          *
1906          *      CPU0:                           CPU1:
1907          *                              change_huge_pmd(prot_numa=1)
1908          *                               pmdp_huge_get_and_clear_notify()
1909          * madvise_dontneed()
1910          *  zap_pmd_range()
1911          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1912          *   // skip the pmd
1913          *                               set_pmd_at();
1914          *                               // pmd is re-established
1915          *
1916          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1917          * which may break userspace.
1918          *
1919          * pmdp_invalidate_ad() is required to make sure we don't miss
1920          * dirty/young flags set by hardware.
1921          */
1922         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1923
1924         entry = pmd_modify(oldpmd, newprot);
1925         if (uffd_wp)
1926                 entry = pmd_mkuffd_wp(entry);
1927         else if (uffd_wp_resolve)
1928                 /*
1929                  * Leave the write bit to be handled by PF interrupt
1930                  * handler, then things like COW could be properly
1931                  * handled.
1932                  */
1933                 entry = pmd_clear_uffd_wp(entry);
1934
1935         /* See change_pte_range(). */
1936         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1937             can_change_pmd_writable(vma, addr, entry))
1938                 entry = pmd_mkwrite(entry);
1939
1940         ret = HPAGE_PMD_NR;
1941         set_pmd_at(mm, addr, pmd, entry);
1942
1943         if (huge_pmd_needs_flush(oldpmd, entry))
1944                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1945 unlock:
1946         spin_unlock(ptl);
1947         return ret;
1948 }
1949
1950 /*
1951  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1952  *
1953  * Note that if it returns page table lock pointer, this routine returns without
1954  * unlocking page table lock. So callers must unlock it.
1955  */
1956 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1957 {
1958         spinlock_t *ptl;
1959         ptl = pmd_lock(vma->vm_mm, pmd);
1960         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1961                         pmd_devmap(*pmd)))
1962                 return ptl;
1963         spin_unlock(ptl);
1964         return NULL;
1965 }
1966
1967 /*
1968  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1969  *
1970  * Note that if it returns page table lock pointer, this routine returns without
1971  * unlocking page table lock. So callers must unlock it.
1972  */
1973 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1974 {
1975         spinlock_t *ptl;
1976
1977         ptl = pud_lock(vma->vm_mm, pud);
1978         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1979                 return ptl;
1980         spin_unlock(ptl);
1981         return NULL;
1982 }
1983
1984 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1985 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1986                  pud_t *pud, unsigned long addr)
1987 {
1988         spinlock_t *ptl;
1989
1990         ptl = __pud_trans_huge_lock(pud, vma);
1991         if (!ptl)
1992                 return 0;
1993
1994         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1995         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1996         if (vma_is_special_huge(vma)) {
1997                 spin_unlock(ptl);
1998                 /* No zero page support yet */
1999         } else {
2000                 /* No support for anonymous PUD pages yet */
2001                 BUG();
2002         }
2003         return 1;
2004 }
2005
2006 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2007                 unsigned long haddr)
2008 {
2009         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2010         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2011         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2012         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2013
2014         count_vm_event(THP_SPLIT_PUD);
2015
2016         pudp_huge_clear_flush_notify(vma, haddr, pud);
2017 }
2018
2019 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2020                 unsigned long address)
2021 {
2022         spinlock_t *ptl;
2023         struct mmu_notifier_range range;
2024
2025         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2026                                 address & HPAGE_PUD_MASK,
2027                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2028         mmu_notifier_invalidate_range_start(&range);
2029         ptl = pud_lock(vma->vm_mm, pud);
2030         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2031                 goto out;
2032         __split_huge_pud_locked(vma, pud, range.start);
2033
2034 out:
2035         spin_unlock(ptl);
2036         /*
2037          * No need to double call mmu_notifier->invalidate_range() callback as
2038          * the above pudp_huge_clear_flush_notify() did already call it.
2039          */
2040         mmu_notifier_invalidate_range_only_end(&range);
2041 }
2042 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2043
2044 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2045                 unsigned long haddr, pmd_t *pmd)
2046 {
2047         struct mm_struct *mm = vma->vm_mm;
2048         pgtable_t pgtable;
2049         pmd_t _pmd;
2050         int i;
2051
2052         /*
2053          * Leave pmd empty until pte is filled note that it is fine to delay
2054          * notification until mmu_notifier_invalidate_range_end() as we are
2055          * replacing a zero pmd write protected page with a zero pte write
2056          * protected page.
2057          *
2058          * See Documentation/mm/mmu_notifier.rst
2059          */
2060         pmdp_huge_clear_flush(vma, haddr, pmd);
2061
2062         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2063         pmd_populate(mm, &_pmd, pgtable);
2064
2065         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2066                 pte_t *pte, entry;
2067                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2068                 entry = pte_mkspecial(entry);
2069                 pte = pte_offset_map(&_pmd, haddr);
2070                 VM_BUG_ON(!pte_none(*pte));
2071                 set_pte_at(mm, haddr, pte, entry);
2072                 pte_unmap(pte);
2073         }
2074         smp_wmb(); /* make pte visible before pmd */
2075         pmd_populate(mm, pmd, pgtable);
2076 }
2077
2078 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2079                 unsigned long haddr, bool freeze)
2080 {
2081         struct mm_struct *mm = vma->vm_mm;
2082         struct page *page;
2083         pgtable_t pgtable;
2084         pmd_t old_pmd, _pmd;
2085         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2086         bool anon_exclusive = false, dirty = false;
2087         unsigned long addr;
2088         int i;
2089
2090         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2091         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2092         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2093         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2094                                 && !pmd_devmap(*pmd));
2095
2096         count_vm_event(THP_SPLIT_PMD);
2097
2098         if (!vma_is_anonymous(vma)) {
2099                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2100                 /*
2101                  * We are going to unmap this huge page. So
2102                  * just go ahead and zap it
2103                  */
2104                 if (arch_needs_pgtable_deposit())
2105                         zap_deposited_table(mm, pmd);
2106                 if (vma_is_special_huge(vma))
2107                         return;
2108                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2109                         swp_entry_t entry;
2110
2111                         entry = pmd_to_swp_entry(old_pmd);
2112                         page = pfn_swap_entry_to_page(entry);
2113                 } else {
2114                         page = pmd_page(old_pmd);
2115                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2116                                 set_page_dirty(page);
2117                         if (!PageReferenced(page) && pmd_young(old_pmd))
2118                                 SetPageReferenced(page);
2119                         page_remove_rmap(page, vma, true);
2120                         put_page(page);
2121                 }
2122                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2123                 return;
2124         }
2125
2126         if (is_huge_zero_pmd(*pmd)) {
2127                 /*
2128                  * FIXME: Do we want to invalidate secondary mmu by calling
2129                  * mmu_notifier_invalidate_range() see comments below inside
2130                  * __split_huge_pmd() ?
2131                  *
2132                  * We are going from a zero huge page write protected to zero
2133                  * small page also write protected so it does not seems useful
2134                  * to invalidate secondary mmu at this time.
2135                  */
2136                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2137         }
2138
2139         /*
2140          * Up to this point the pmd is present and huge and userland has the
2141          * whole access to the hugepage during the split (which happens in
2142          * place). If we overwrite the pmd with the not-huge version pointing
2143          * to the pte here (which of course we could if all CPUs were bug
2144          * free), userland could trigger a small page size TLB miss on the
2145          * small sized TLB while the hugepage TLB entry is still established in
2146          * the huge TLB. Some CPU doesn't like that.
2147          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2148          * 383 on page 105. Intel should be safe but is also warns that it's
2149          * only safe if the permission and cache attributes of the two entries
2150          * loaded in the two TLB is identical (which should be the case here).
2151          * But it is generally safer to never allow small and huge TLB entries
2152          * for the same virtual address to be loaded simultaneously. So instead
2153          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2154          * current pmd notpresent (atomically because here the pmd_trans_huge
2155          * must remain set at all times on the pmd until the split is complete
2156          * for this pmd), then we flush the SMP TLB and finally we write the
2157          * non-huge version of the pmd entry with pmd_populate.
2158          */
2159         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2160
2161         pmd_migration = is_pmd_migration_entry(old_pmd);
2162         if (unlikely(pmd_migration)) {
2163                 swp_entry_t entry;
2164
2165                 entry = pmd_to_swp_entry(old_pmd);
2166                 page = pfn_swap_entry_to_page(entry);
2167                 write = is_writable_migration_entry(entry);
2168                 if (PageAnon(page))
2169                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2170                 young = is_migration_entry_young(entry);
2171                 dirty = is_migration_entry_dirty(entry);
2172                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2173                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2174         } else {
2175                 page = pmd_page(old_pmd);
2176                 if (pmd_dirty(old_pmd)) {
2177                         dirty = true;
2178                         SetPageDirty(page);
2179                 }
2180                 write = pmd_write(old_pmd);
2181                 young = pmd_young(old_pmd);
2182                 soft_dirty = pmd_soft_dirty(old_pmd);
2183                 uffd_wp = pmd_uffd_wp(old_pmd);
2184
2185                 VM_BUG_ON_PAGE(!page_count(page), page);
2186
2187                 /*
2188                  * Without "freeze", we'll simply split the PMD, propagating the
2189                  * PageAnonExclusive() flag for each PTE by setting it for
2190                  * each subpage -- no need to (temporarily) clear.
2191                  *
2192                  * With "freeze" we want to replace mapped pages by
2193                  * migration entries right away. This is only possible if we
2194                  * managed to clear PageAnonExclusive() -- see
2195                  * set_pmd_migration_entry().
2196                  *
2197                  * In case we cannot clear PageAnonExclusive(), split the PMD
2198                  * only and let try_to_migrate_one() fail later.
2199                  *
2200                  * See page_try_share_anon_rmap(): invalidate PMD first.
2201                  */
2202                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2203                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2204                         freeze = false;
2205                 if (!freeze)
2206                         page_ref_add(page, HPAGE_PMD_NR - 1);
2207         }
2208
2209         /*
2210          * Withdraw the table only after we mark the pmd entry invalid.
2211          * This's critical for some architectures (Power).
2212          */
2213         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2214         pmd_populate(mm, &_pmd, pgtable);
2215
2216         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2217                 pte_t entry, *pte;
2218                 /*
2219                  * Note that NUMA hinting access restrictions are not
2220                  * transferred to avoid any possibility of altering
2221                  * permissions across VMAs.
2222                  */
2223                 if (freeze || pmd_migration) {
2224                         swp_entry_t swp_entry;
2225                         if (write)
2226                                 swp_entry = make_writable_migration_entry(
2227                                                         page_to_pfn(page + i));
2228                         else if (anon_exclusive)
2229                                 swp_entry = make_readable_exclusive_migration_entry(
2230                                                         page_to_pfn(page + i));
2231                         else
2232                                 swp_entry = make_readable_migration_entry(
2233                                                         page_to_pfn(page + i));
2234                         if (young)
2235                                 swp_entry = make_migration_entry_young(swp_entry);
2236                         if (dirty)
2237                                 swp_entry = make_migration_entry_dirty(swp_entry);
2238                         entry = swp_entry_to_pte(swp_entry);
2239                         if (soft_dirty)
2240                                 entry = pte_swp_mksoft_dirty(entry);
2241                         if (uffd_wp)
2242                                 entry = pte_swp_mkuffd_wp(entry);
2243                 } else {
2244                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2245                         entry = maybe_mkwrite(entry, vma);
2246                         if (anon_exclusive)
2247                                 SetPageAnonExclusive(page + i);
2248                         if (!young)
2249                                 entry = pte_mkold(entry);
2250                         /* NOTE: this may set soft-dirty too on some archs */
2251                         if (dirty)
2252                                 entry = pte_mkdirty(entry);
2253                         /*
2254                          * NOTE: this needs to happen after pte_mkdirty,
2255                          * because some archs (sparc64, loongarch) could
2256                          * set hw write bit when mkdirty.
2257                          */
2258                         if (!write)
2259                                 entry = pte_wrprotect(entry);
2260                         if (soft_dirty)
2261                                 entry = pte_mksoft_dirty(entry);
2262                         if (uffd_wp)
2263                                 entry = pte_mkuffd_wp(entry);
2264                         page_add_anon_rmap(page + i, vma, addr, false);
2265                 }
2266                 pte = pte_offset_map(&_pmd, addr);
2267                 BUG_ON(!pte_none(*pte));
2268                 set_pte_at(mm, addr, pte, entry);
2269                 pte_unmap(pte);
2270         }
2271
2272         if (!pmd_migration)
2273                 page_remove_rmap(page, vma, true);
2274         if (freeze)
2275                 put_page(page);
2276
2277         smp_wmb(); /* make pte visible before pmd */
2278         pmd_populate(mm, pmd, pgtable);
2279 }
2280
2281 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2282                 unsigned long address, bool freeze, struct folio *folio)
2283 {
2284         spinlock_t *ptl;
2285         struct mmu_notifier_range range;
2286
2287         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2288                                 address & HPAGE_PMD_MASK,
2289                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2290         mmu_notifier_invalidate_range_start(&range);
2291         ptl = pmd_lock(vma->vm_mm, pmd);
2292
2293         /*
2294          * If caller asks to setup a migration entry, we need a folio to check
2295          * pmd against. Otherwise we can end up replacing wrong folio.
2296          */
2297         VM_BUG_ON(freeze && !folio);
2298         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2299
2300         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2301             is_pmd_migration_entry(*pmd)) {
2302                 /*
2303                  * It's safe to call pmd_page when folio is set because it's
2304                  * guaranteed that pmd is present.
2305                  */
2306                 if (folio && folio != page_folio(pmd_page(*pmd)))
2307                         goto out;
2308                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2309         }
2310
2311 out:
2312         spin_unlock(ptl);
2313         /*
2314          * No need to double call mmu_notifier->invalidate_range() callback.
2315          * They are 3 cases to consider inside __split_huge_pmd_locked():
2316          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2317          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2318          *    fault will trigger a flush_notify before pointing to a new page
2319          *    (it is fine if the secondary mmu keeps pointing to the old zero
2320          *    page in the meantime)
2321          *  3) Split a huge pmd into pte pointing to the same page. No need
2322          *     to invalidate secondary tlb entry they are all still valid.
2323          *     any further changes to individual pte will notify. So no need
2324          *     to call mmu_notifier->invalidate_range()
2325          */
2326         mmu_notifier_invalidate_range_only_end(&range);
2327 }
2328
2329 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2330                 bool freeze, struct folio *folio)
2331 {
2332         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2333
2334         if (!pmd)
2335                 return;
2336
2337         __split_huge_pmd(vma, pmd, address, freeze, folio);
2338 }
2339
2340 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2341 {
2342         /*
2343          * If the new address isn't hpage aligned and it could previously
2344          * contain an hugepage: check if we need to split an huge pmd.
2345          */
2346         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2347             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2348                          ALIGN(address, HPAGE_PMD_SIZE)))
2349                 split_huge_pmd_address(vma, address, false, NULL);
2350 }
2351
2352 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2353                              unsigned long start,
2354                              unsigned long end,
2355                              long adjust_next)
2356 {
2357         /* Check if we need to split start first. */
2358         split_huge_pmd_if_needed(vma, start);
2359
2360         /* Check if we need to split end next. */
2361         split_huge_pmd_if_needed(vma, end);
2362
2363         /*
2364          * If we're also updating the next vma vm_start,
2365          * check if we need to split it.
2366          */
2367         if (adjust_next > 0) {
2368                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2369                 unsigned long nstart = next->vm_start;
2370                 nstart += adjust_next;
2371                 split_huge_pmd_if_needed(next, nstart);
2372         }
2373 }
2374
2375 static void unmap_folio(struct folio *folio)
2376 {
2377         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2378                 TTU_SYNC;
2379
2380         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2381
2382         /*
2383          * Anon pages need migration entries to preserve them, but file
2384          * pages can simply be left unmapped, then faulted back on demand.
2385          * If that is ever changed (perhaps for mlock), update remap_page().
2386          */
2387         if (folio_test_anon(folio))
2388                 try_to_migrate(folio, ttu_flags);
2389         else
2390                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2391 }
2392
2393 static void remap_page(struct folio *folio, unsigned long nr)
2394 {
2395         int i = 0;
2396
2397         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2398         if (!folio_test_anon(folio))
2399                 return;
2400         for (;;) {
2401                 remove_migration_ptes(folio, folio, true);
2402                 i += folio_nr_pages(folio);
2403                 if (i >= nr)
2404                         break;
2405                 folio = folio_next(folio);
2406         }
2407 }
2408
2409 static void lru_add_page_tail(struct page *head, struct page *tail,
2410                 struct lruvec *lruvec, struct list_head *list)
2411 {
2412         VM_BUG_ON_PAGE(!PageHead(head), head);
2413         VM_BUG_ON_PAGE(PageCompound(tail), head);
2414         VM_BUG_ON_PAGE(PageLRU(tail), head);
2415         lockdep_assert_held(&lruvec->lru_lock);
2416
2417         if (list) {
2418                 /* page reclaim is reclaiming a huge page */
2419                 VM_WARN_ON(PageLRU(head));
2420                 get_page(tail);
2421                 list_add_tail(&tail->lru, list);
2422         } else {
2423                 /* head is still on lru (and we have it frozen) */
2424                 VM_WARN_ON(!PageLRU(head));
2425                 if (PageUnevictable(tail))
2426                         tail->mlock_count = 0;
2427                 else
2428                         list_add_tail(&tail->lru, &head->lru);
2429                 SetPageLRU(tail);
2430         }
2431 }
2432
2433 static void __split_huge_page_tail(struct page *head, int tail,
2434                 struct lruvec *lruvec, struct list_head *list)
2435 {
2436         struct page *page_tail = head + tail;
2437
2438         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2439
2440         /*
2441          * Clone page flags before unfreezing refcount.
2442          *
2443          * After successful get_page_unless_zero() might follow flags change,
2444          * for example lock_page() which set PG_waiters.
2445          *
2446          * Note that for mapped sub-pages of an anonymous THP,
2447          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2448          * the migration entry instead from where remap_page() will restore it.
2449          * We can still have PG_anon_exclusive set on effectively unmapped and
2450          * unreferenced sub-pages of an anonymous THP: we can simply drop
2451          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2452          */
2453         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2454         page_tail->flags |= (head->flags &
2455                         ((1L << PG_referenced) |
2456                          (1L << PG_swapbacked) |
2457                          (1L << PG_swapcache) |
2458                          (1L << PG_mlocked) |
2459                          (1L << PG_uptodate) |
2460                          (1L << PG_active) |
2461                          (1L << PG_workingset) |
2462                          (1L << PG_locked) |
2463                          (1L << PG_unevictable) |
2464 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2465                          (1L << PG_arch_2) |
2466                          (1L << PG_arch_3) |
2467 #endif
2468                          (1L << PG_dirty) |
2469                          LRU_GEN_MASK | LRU_REFS_MASK));
2470
2471         /* ->mapping in first and second tail page is replaced by other uses */
2472         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2473                         page_tail);
2474         page_tail->mapping = head->mapping;
2475         page_tail->index = head->index + tail;
2476
2477         /*
2478          * page->private should not be set in tail pages with the exception
2479          * of swap cache pages that store the swp_entry_t in tail pages.
2480          * Fix up and warn once if private is unexpectedly set.
2481          *
2482          * What of 32-bit systems, on which folio->_pincount overlays
2483          * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2484          * pincount must be 0 for folio_ref_freeze() to have succeeded.
2485          */
2486         if (!folio_test_swapcache(page_folio(head))) {
2487                 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2488                 page_tail->private = 0;
2489         }
2490
2491         /* Page flags must be visible before we make the page non-compound. */
2492         smp_wmb();
2493
2494         /*
2495          * Clear PageTail before unfreezing page refcount.
2496          *
2497          * After successful get_page_unless_zero() might follow put_page()
2498          * which needs correct compound_head().
2499          */
2500         clear_compound_head(page_tail);
2501
2502         /* Finally unfreeze refcount. Additional reference from page cache. */
2503         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2504                                           PageSwapCache(head)));
2505
2506         if (page_is_young(head))
2507                 set_page_young(page_tail);
2508         if (page_is_idle(head))
2509                 set_page_idle(page_tail);
2510
2511         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2512
2513         /*
2514          * always add to the tail because some iterators expect new
2515          * pages to show after the currently processed elements - e.g.
2516          * migrate_pages
2517          */
2518         lru_add_page_tail(head, page_tail, lruvec, list);
2519 }
2520
2521 static void __split_huge_page(struct page *page, struct list_head *list,
2522                 pgoff_t end)
2523 {
2524         struct folio *folio = page_folio(page);
2525         struct page *head = &folio->page;
2526         struct lruvec *lruvec;
2527         struct address_space *swap_cache = NULL;
2528         unsigned long offset = 0;
2529         unsigned int nr = thp_nr_pages(head);
2530         int i;
2531
2532         /* complete memcg works before add pages to LRU */
2533         split_page_memcg(head, nr);
2534
2535         if (PageAnon(head) && PageSwapCache(head)) {
2536                 swp_entry_t entry = { .val = page_private(head) };
2537
2538                 offset = swp_offset(entry);
2539                 swap_cache = swap_address_space(entry);
2540                 xa_lock(&swap_cache->i_pages);
2541         }
2542
2543         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2544         lruvec = folio_lruvec_lock(folio);
2545
2546         ClearPageHasHWPoisoned(head);
2547
2548         for (i = nr - 1; i >= 1; i--) {
2549                 __split_huge_page_tail(head, i, lruvec, list);
2550                 /* Some pages can be beyond EOF: drop them from page cache */
2551                 if (head[i].index >= end) {
2552                         struct folio *tail = page_folio(head + i);
2553
2554                         if (shmem_mapping(head->mapping))
2555                                 shmem_uncharge(head->mapping->host, 1);
2556                         else if (folio_test_clear_dirty(tail))
2557                                 folio_account_cleaned(tail,
2558                                         inode_to_wb(folio->mapping->host));
2559                         __filemap_remove_folio(tail, NULL);
2560                         folio_put(tail);
2561                 } else if (!PageAnon(page)) {
2562                         __xa_store(&head->mapping->i_pages, head[i].index,
2563                                         head + i, 0);
2564                 } else if (swap_cache) {
2565                         __xa_store(&swap_cache->i_pages, offset + i,
2566                                         head + i, 0);
2567                 }
2568         }
2569
2570         ClearPageCompound(head);
2571         unlock_page_lruvec(lruvec);
2572         /* Caller disabled irqs, so they are still disabled here */
2573
2574         split_page_owner(head, nr);
2575
2576         /* See comment in __split_huge_page_tail() */
2577         if (PageAnon(head)) {
2578                 /* Additional pin to swap cache */
2579                 if (PageSwapCache(head)) {
2580                         page_ref_add(head, 2);
2581                         xa_unlock(&swap_cache->i_pages);
2582                 } else {
2583                         page_ref_inc(head);
2584                 }
2585         } else {
2586                 /* Additional pin to page cache */
2587                 page_ref_add(head, 2);
2588                 xa_unlock(&head->mapping->i_pages);
2589         }
2590         local_irq_enable();
2591
2592         remap_page(folio, nr);
2593
2594         if (PageSwapCache(head)) {
2595                 swp_entry_t entry = { .val = page_private(head) };
2596
2597                 split_swap_cluster(entry);
2598         }
2599
2600         for (i = 0; i < nr; i++) {
2601                 struct page *subpage = head + i;
2602                 if (subpage == page)
2603                         continue;
2604                 unlock_page(subpage);
2605
2606                 /*
2607                  * Subpages may be freed if there wasn't any mapping
2608                  * like if add_to_swap() is running on a lru page that
2609                  * had its mapping zapped. And freeing these pages
2610                  * requires taking the lru_lock so we do the put_page
2611                  * of the tail pages after the split is complete.
2612                  */
2613                 free_page_and_swap_cache(subpage);
2614         }
2615 }
2616
2617 /* Racy check whether the huge page can be split */
2618 bool can_split_folio(struct folio *folio, int *pextra_pins)
2619 {
2620         int extra_pins;
2621
2622         /* Additional pins from page cache */
2623         if (folio_test_anon(folio))
2624                 extra_pins = folio_test_swapcache(folio) ?
2625                                 folio_nr_pages(folio) : 0;
2626         else
2627                 extra_pins = folio_nr_pages(folio);
2628         if (pextra_pins)
2629                 *pextra_pins = extra_pins;
2630         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2631 }
2632
2633 /*
2634  * This function splits huge page into normal pages. @page can point to any
2635  * subpage of huge page to split. Split doesn't change the position of @page.
2636  *
2637  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2638  * The huge page must be locked.
2639  *
2640  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2641  *
2642  * Both head page and tail pages will inherit mapping, flags, and so on from
2643  * the hugepage.
2644  *
2645  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2646  * they are not mapped.
2647  *
2648  * Returns 0 if the hugepage is split successfully.
2649  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2650  * us.
2651  */
2652 int split_huge_page_to_list(struct page *page, struct list_head *list)
2653 {
2654         struct folio *folio = page_folio(page);
2655         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2656         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2657         struct anon_vma *anon_vma = NULL;
2658         struct address_space *mapping = NULL;
2659         int extra_pins, ret;
2660         pgoff_t end;
2661         bool is_hzp;
2662
2663         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2664         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2665
2666         is_hzp = is_huge_zero_page(&folio->page);
2667         VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2668         if (is_hzp)
2669                 return -EBUSY;
2670
2671         if (folio_test_writeback(folio))
2672                 return -EBUSY;
2673
2674         if (folio_test_anon(folio)) {
2675                 /*
2676                  * The caller does not necessarily hold an mmap_lock that would
2677                  * prevent the anon_vma disappearing so we first we take a
2678                  * reference to it and then lock the anon_vma for write. This
2679                  * is similar to folio_lock_anon_vma_read except the write lock
2680                  * is taken to serialise against parallel split or collapse
2681                  * operations.
2682                  */
2683                 anon_vma = folio_get_anon_vma(folio);
2684                 if (!anon_vma) {
2685                         ret = -EBUSY;
2686                         goto out;
2687                 }
2688                 end = -1;
2689                 mapping = NULL;
2690                 anon_vma_lock_write(anon_vma);
2691         } else {
2692                 gfp_t gfp;
2693
2694                 mapping = folio->mapping;
2695
2696                 /* Truncated ? */
2697                 if (!mapping) {
2698                         ret = -EBUSY;
2699                         goto out;
2700                 }
2701
2702                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2703                                                         GFP_RECLAIM_MASK);
2704
2705                 if (folio_test_private(folio) &&
2706                                 !filemap_release_folio(folio, gfp)) {
2707                         ret = -EBUSY;
2708                         goto out;
2709                 }
2710
2711                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2712                 if (xas_error(&xas)) {
2713                         ret = xas_error(&xas);
2714                         goto out;
2715                 }
2716
2717                 anon_vma = NULL;
2718                 i_mmap_lock_read(mapping);
2719
2720                 /*
2721                  *__split_huge_page() may need to trim off pages beyond EOF:
2722                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2723                  * which cannot be nested inside the page tree lock. So note
2724                  * end now: i_size itself may be changed at any moment, but
2725                  * folio lock is good enough to serialize the trimming.
2726                  */
2727                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2728                 if (shmem_mapping(mapping))
2729                         end = shmem_fallocend(mapping->host, end);
2730         }
2731
2732         /*
2733          * Racy check if we can split the page, before unmap_folio() will
2734          * split PMDs
2735          */
2736         if (!can_split_folio(folio, &extra_pins)) {
2737                 ret = -EAGAIN;
2738                 goto out_unlock;
2739         }
2740
2741         unmap_folio(folio);
2742
2743         /* block interrupt reentry in xa_lock and spinlock */
2744         local_irq_disable();
2745         if (mapping) {
2746                 /*
2747                  * Check if the folio is present in page cache.
2748                  * We assume all tail are present too, if folio is there.
2749                  */
2750                 xas_lock(&xas);
2751                 xas_reset(&xas);
2752                 if (xas_load(&xas) != folio)
2753                         goto fail;
2754         }
2755
2756         /* Prevent deferred_split_scan() touching ->_refcount */
2757         spin_lock(&ds_queue->split_queue_lock);
2758         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2759                 if (!list_empty(&folio->_deferred_list)) {
2760                         ds_queue->split_queue_len--;
2761                         list_del(&folio->_deferred_list);
2762                 }
2763                 spin_unlock(&ds_queue->split_queue_lock);
2764                 if (mapping) {
2765                         int nr = folio_nr_pages(folio);
2766
2767                         xas_split(&xas, folio, folio_order(folio));
2768                         if (folio_test_swapbacked(folio)) {
2769                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2770                                                         -nr);
2771                         } else {
2772                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2773                                                         -nr);
2774                                 filemap_nr_thps_dec(mapping);
2775                         }
2776                 }
2777
2778                 __split_huge_page(page, list, end);
2779                 ret = 0;
2780         } else {
2781                 spin_unlock(&ds_queue->split_queue_lock);
2782 fail:
2783                 if (mapping)
2784                         xas_unlock(&xas);
2785                 local_irq_enable();
2786                 remap_page(folio, folio_nr_pages(folio));
2787                 ret = -EAGAIN;
2788         }
2789
2790 out_unlock:
2791         if (anon_vma) {
2792                 anon_vma_unlock_write(anon_vma);
2793                 put_anon_vma(anon_vma);
2794         }
2795         if (mapping)
2796                 i_mmap_unlock_read(mapping);
2797 out:
2798         xas_destroy(&xas);
2799         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2800         return ret;
2801 }
2802
2803 void free_transhuge_page(struct page *page)
2804 {
2805         struct folio *folio = (struct folio *)page;
2806         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2807         unsigned long flags;
2808
2809         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2810         if (!list_empty(&folio->_deferred_list)) {
2811                 ds_queue->split_queue_len--;
2812                 list_del(&folio->_deferred_list);
2813         }
2814         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2815         free_compound_page(page);
2816 }
2817
2818 void deferred_split_folio(struct folio *folio)
2819 {
2820         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2821 #ifdef CONFIG_MEMCG
2822         struct mem_cgroup *memcg = folio_memcg(folio);
2823 #endif
2824         unsigned long flags;
2825
2826         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2827
2828         /*
2829          * The try_to_unmap() in page reclaim path might reach here too,
2830          * this may cause a race condition to corrupt deferred split queue.
2831          * And, if page reclaim is already handling the same folio, it is
2832          * unnecessary to handle it again in shrinker.
2833          *
2834          * Check the swapcache flag to determine if the folio is being
2835          * handled by page reclaim since THP swap would add the folio into
2836          * swap cache before calling try_to_unmap().
2837          */
2838         if (folio_test_swapcache(folio))
2839                 return;
2840
2841         if (!list_empty(&folio->_deferred_list))
2842                 return;
2843
2844         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2845         if (list_empty(&folio->_deferred_list)) {
2846                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2847                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2848                 ds_queue->split_queue_len++;
2849 #ifdef CONFIG_MEMCG
2850                 if (memcg)
2851                         set_shrinker_bit(memcg, folio_nid(folio),
2852                                          deferred_split_shrinker.id);
2853 #endif
2854         }
2855         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2856 }
2857
2858 static unsigned long deferred_split_count(struct shrinker *shrink,
2859                 struct shrink_control *sc)
2860 {
2861         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2862         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2863
2864 #ifdef CONFIG_MEMCG
2865         if (sc->memcg)
2866                 ds_queue = &sc->memcg->deferred_split_queue;
2867 #endif
2868         return READ_ONCE(ds_queue->split_queue_len);
2869 }
2870
2871 static unsigned long deferred_split_scan(struct shrinker *shrink,
2872                 struct shrink_control *sc)
2873 {
2874         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2875         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2876         unsigned long flags;
2877         LIST_HEAD(list);
2878         struct folio *folio, *next;
2879         int split = 0;
2880
2881 #ifdef CONFIG_MEMCG
2882         if (sc->memcg)
2883                 ds_queue = &sc->memcg->deferred_split_queue;
2884 #endif
2885
2886         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2887         /* Take pin on all head pages to avoid freeing them under us */
2888         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2889                                                         _deferred_list) {
2890                 if (folio_try_get(folio)) {
2891                         list_move(&folio->_deferred_list, &list);
2892                 } else {
2893                         /* We lost race with folio_put() */
2894                         list_del_init(&folio->_deferred_list);
2895                         ds_queue->split_queue_len--;
2896                 }
2897                 if (!--sc->nr_to_scan)
2898                         break;
2899         }
2900         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2901
2902         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2903                 if (!folio_trylock(folio))
2904                         goto next;
2905                 /* split_huge_page() removes page from list on success */
2906                 if (!split_folio(folio))
2907                         split++;
2908                 folio_unlock(folio);
2909 next:
2910                 folio_put(folio);
2911         }
2912
2913         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2914         list_splice_tail(&list, &ds_queue->split_queue);
2915         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2916
2917         /*
2918          * Stop shrinker if we didn't split any page, but the queue is empty.
2919          * This can happen if pages were freed under us.
2920          */
2921         if (!split && list_empty(&ds_queue->split_queue))
2922                 return SHRINK_STOP;
2923         return split;
2924 }
2925
2926 static struct shrinker deferred_split_shrinker = {
2927         .count_objects = deferred_split_count,
2928         .scan_objects = deferred_split_scan,
2929         .seeks = DEFAULT_SEEKS,
2930         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2931                  SHRINKER_NONSLAB,
2932 };
2933
2934 #ifdef CONFIG_DEBUG_FS
2935 static void split_huge_pages_all(void)
2936 {
2937         struct zone *zone;
2938         struct page *page;
2939         struct folio *folio;
2940         unsigned long pfn, max_zone_pfn;
2941         unsigned long total = 0, split = 0;
2942
2943         pr_debug("Split all THPs\n");
2944         for_each_zone(zone) {
2945                 if (!managed_zone(zone))
2946                         continue;
2947                 max_zone_pfn = zone_end_pfn(zone);
2948                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2949                         int nr_pages;
2950
2951                         page = pfn_to_online_page(pfn);
2952                         if (!page || PageTail(page))
2953                                 continue;
2954                         folio = page_folio(page);
2955                         if (!folio_try_get(folio))
2956                                 continue;
2957
2958                         if (unlikely(page_folio(page) != folio))
2959                                 goto next;
2960
2961                         if (zone != folio_zone(folio))
2962                                 goto next;
2963
2964                         if (!folio_test_large(folio)
2965                                 || folio_test_hugetlb(folio)
2966                                 || !folio_test_lru(folio))
2967                                 goto next;
2968
2969                         total++;
2970                         folio_lock(folio);
2971                         nr_pages = folio_nr_pages(folio);
2972                         if (!split_folio(folio))
2973                                 split++;
2974                         pfn += nr_pages - 1;
2975                         folio_unlock(folio);
2976 next:
2977                         folio_put(folio);
2978                         cond_resched();
2979                 }
2980         }
2981
2982         pr_debug("%lu of %lu THP split\n", split, total);
2983 }
2984
2985 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2986 {
2987         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2988                     is_vm_hugetlb_page(vma);
2989 }
2990
2991 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2992                                 unsigned long vaddr_end)
2993 {
2994         int ret = 0;
2995         struct task_struct *task;
2996         struct mm_struct *mm;
2997         unsigned long total = 0, split = 0;
2998         unsigned long addr;
2999
3000         vaddr_start &= PAGE_MASK;
3001         vaddr_end &= PAGE_MASK;
3002
3003         /* Find the task_struct from pid */
3004         rcu_read_lock();
3005         task = find_task_by_vpid(pid);
3006         if (!task) {
3007                 rcu_read_unlock();
3008                 ret = -ESRCH;
3009                 goto out;
3010         }
3011         get_task_struct(task);
3012         rcu_read_unlock();
3013
3014         /* Find the mm_struct */
3015         mm = get_task_mm(task);
3016         put_task_struct(task);
3017
3018         if (!mm) {
3019                 ret = -EINVAL;
3020                 goto out;
3021         }
3022
3023         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3024                  pid, vaddr_start, vaddr_end);
3025
3026         mmap_read_lock(mm);
3027         /*
3028          * always increase addr by PAGE_SIZE, since we could have a PTE page
3029          * table filled with PTE-mapped THPs, each of which is distinct.
3030          */
3031         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3032                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3033                 struct page *page;
3034
3035                 if (!vma)
3036                         break;
3037
3038                 /* skip special VMA and hugetlb VMA */
3039                 if (vma_not_suitable_for_thp_split(vma)) {
3040                         addr = vma->vm_end;
3041                         continue;
3042                 }
3043
3044                 /* FOLL_DUMP to ignore special (like zero) pages */
3045                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3046
3047                 if (IS_ERR_OR_NULL(page))
3048                         continue;
3049
3050                 if (!is_transparent_hugepage(page))
3051                         goto next;
3052
3053                 total++;
3054                 if (!can_split_folio(page_folio(page), NULL))
3055                         goto next;
3056
3057                 if (!trylock_page(page))
3058                         goto next;
3059
3060                 if (!split_huge_page(page))
3061                         split++;
3062
3063                 unlock_page(page);
3064 next:
3065                 put_page(page);
3066                 cond_resched();
3067         }
3068         mmap_read_unlock(mm);
3069         mmput(mm);
3070
3071         pr_debug("%lu of %lu THP split\n", split, total);
3072
3073 out:
3074         return ret;
3075 }
3076
3077 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3078                                 pgoff_t off_end)
3079 {
3080         struct filename *file;
3081         struct file *candidate;
3082         struct address_space *mapping;
3083         int ret = -EINVAL;
3084         pgoff_t index;
3085         int nr_pages = 1;
3086         unsigned long total = 0, split = 0;
3087
3088         file = getname_kernel(file_path);
3089         if (IS_ERR(file))
3090                 return ret;
3091
3092         candidate = file_open_name(file, O_RDONLY, 0);
3093         if (IS_ERR(candidate))
3094                 goto out;
3095
3096         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3097                  file_path, off_start, off_end);
3098
3099         mapping = candidate->f_mapping;
3100
3101         for (index = off_start; index < off_end; index += nr_pages) {
3102                 struct folio *folio = __filemap_get_folio(mapping, index,
3103                                                 FGP_ENTRY, 0);
3104
3105                 nr_pages = 1;
3106                 if (xa_is_value(folio) || !folio)
3107                         continue;
3108
3109                 if (!folio_test_large(folio))
3110                         goto next;
3111
3112                 total++;
3113                 nr_pages = folio_nr_pages(folio);
3114
3115                 if (!folio_trylock(folio))
3116                         goto next;
3117
3118                 if (!split_folio(folio))
3119                         split++;
3120
3121                 folio_unlock(folio);
3122 next:
3123                 folio_put(folio);
3124                 cond_resched();
3125         }
3126
3127         filp_close(candidate, NULL);
3128         ret = 0;
3129
3130         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3131 out:
3132         putname(file);
3133         return ret;
3134 }
3135
3136 #define MAX_INPUT_BUF_SZ 255
3137
3138 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3139                                 size_t count, loff_t *ppops)
3140 {
3141         static DEFINE_MUTEX(split_debug_mutex);
3142         ssize_t ret;
3143         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3144         char input_buf[MAX_INPUT_BUF_SZ];
3145         int pid;
3146         unsigned long vaddr_start, vaddr_end;
3147
3148         ret = mutex_lock_interruptible(&split_debug_mutex);
3149         if (ret)
3150                 return ret;
3151
3152         ret = -EFAULT;
3153
3154         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3155         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3156                 goto out;
3157
3158         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3159
3160         if (input_buf[0] == '/') {
3161                 char *tok;
3162                 char *buf = input_buf;
3163                 char file_path[MAX_INPUT_BUF_SZ];
3164                 pgoff_t off_start = 0, off_end = 0;
3165                 size_t input_len = strlen(input_buf);
3166
3167                 tok = strsep(&buf, ",");
3168                 if (tok) {
3169                         strcpy(file_path, tok);
3170                 } else {
3171                         ret = -EINVAL;
3172                         goto out;
3173                 }
3174
3175                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3176                 if (ret != 2) {
3177                         ret = -EINVAL;
3178                         goto out;
3179                 }
3180                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3181                 if (!ret)
3182                         ret = input_len;
3183
3184                 goto out;
3185         }
3186
3187         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3188         if (ret == 1 && pid == 1) {
3189                 split_huge_pages_all();
3190                 ret = strlen(input_buf);
3191                 goto out;
3192         } else if (ret != 3) {
3193                 ret = -EINVAL;
3194                 goto out;
3195         }
3196
3197         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3198         if (!ret)
3199                 ret = strlen(input_buf);
3200 out:
3201         mutex_unlock(&split_debug_mutex);
3202         return ret;
3203
3204 }
3205
3206 static const struct file_operations split_huge_pages_fops = {
3207         .owner   = THIS_MODULE,
3208         .write   = split_huge_pages_write,
3209         .llseek  = no_llseek,
3210 };
3211
3212 static int __init split_huge_pages_debugfs(void)
3213 {
3214         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3215                             &split_huge_pages_fops);
3216         return 0;
3217 }
3218 late_initcall(split_huge_pages_debugfs);
3219 #endif
3220
3221 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3222 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3223                 struct page *page)
3224 {
3225         struct vm_area_struct *vma = pvmw->vma;
3226         struct mm_struct *mm = vma->vm_mm;
3227         unsigned long address = pvmw->address;
3228         bool anon_exclusive;
3229         pmd_t pmdval;
3230         swp_entry_t entry;
3231         pmd_t pmdswp;
3232
3233         if (!(pvmw->pmd && !pvmw->pte))
3234                 return 0;
3235
3236         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3237         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3238
3239         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3240         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3241         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3242                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3243                 return -EBUSY;
3244         }
3245
3246         if (pmd_dirty(pmdval))
3247                 set_page_dirty(page);
3248         if (pmd_write(pmdval))
3249                 entry = make_writable_migration_entry(page_to_pfn(page));
3250         else if (anon_exclusive)
3251                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3252         else
3253                 entry = make_readable_migration_entry(page_to_pfn(page));
3254         if (pmd_young(pmdval))
3255                 entry = make_migration_entry_young(entry);
3256         if (pmd_dirty(pmdval))
3257                 entry = make_migration_entry_dirty(entry);
3258         pmdswp = swp_entry_to_pmd(entry);
3259         if (pmd_soft_dirty(pmdval))
3260                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3261         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3262         page_remove_rmap(page, vma, true);
3263         put_page(page);
3264         trace_set_migration_pmd(address, pmd_val(pmdswp));
3265
3266         return 0;
3267 }
3268
3269 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3270 {
3271         struct vm_area_struct *vma = pvmw->vma;
3272         struct mm_struct *mm = vma->vm_mm;
3273         unsigned long address = pvmw->address;
3274         unsigned long haddr = address & HPAGE_PMD_MASK;
3275         pmd_t pmde;
3276         swp_entry_t entry;
3277
3278         if (!(pvmw->pmd && !pvmw->pte))
3279                 return;
3280
3281         entry = pmd_to_swp_entry(*pvmw->pmd);
3282         get_page(new);
3283         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3284         if (pmd_swp_soft_dirty(*pvmw->pmd))
3285                 pmde = pmd_mksoft_dirty(pmde);
3286         if (is_writable_migration_entry(entry))
3287                 pmde = maybe_pmd_mkwrite(pmde, vma);
3288         if (pmd_swp_uffd_wp(*pvmw->pmd))
3289                 pmde = pmd_mkuffd_wp(pmde);
3290         if (!is_migration_entry_young(entry))
3291                 pmde = pmd_mkold(pmde);
3292         /* NOTE: this may contain setting soft-dirty on some archs */
3293         if (PageDirty(new) && is_migration_entry_dirty(entry))
3294                 pmde = pmd_mkdirty(pmde);
3295
3296         if (PageAnon(new)) {
3297                 rmap_t rmap_flags = RMAP_COMPOUND;
3298
3299                 if (!is_readable_migration_entry(entry))
3300                         rmap_flags |= RMAP_EXCLUSIVE;
3301
3302                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3303         } else {
3304                 page_add_file_rmap(new, vma, true);
3305         }
3306         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3307         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3308
3309         /* No need to invalidate - it was non-present before */
3310         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3311         trace_remove_migration_pmd(address, pmd_val(pmde));
3312 }
3313 #endif