mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/backing-dev.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37 #include <linux/page_owner.h>
  38 #include <linux/sched/sysctl.h>
  39 #include <linux/memory-tiers.h>
  40
  41 #include <asm/tlb.h>
  42 #include <asm/pgalloc.h>
  43 #include "internal.h"
  44 #include "swap.h"
  45
  46 #define CREATE_TRACE_POINTS
  47 #include <trace/events/thp.h>
  48
  49 /*
  50  * By default, transparent hugepage support is disabled in order to avoid
  51  * risking an increased memory footprint for applications that are not
  52  * guaranteed to benefit from it. When transparent hugepage support is
  53  * enabled, it is for all mappings, and khugepaged scans all mappings.
  54  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  55  * for all hugepage allocations.
  56  */
  57 unsigned long transparent_hugepage_flags __read_mostly =
  58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  59         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  60 #endif
  61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  62         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  63 #endif
  64         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  65         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  66         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  67
  68 static struct shrinker deferred_split_shrinker;
  69
  70 static atomic_t huge_zero_refcount;
  71 struct page *huge_zero_page __read_mostly;
  72 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  73
  74 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
  75                         bool smaps, bool in_pf, bool enforce_sysfs)
  76 {
  77         if (!vma->vm_mm)                /* vdso */
  78                 return false;
  79
  80         /*
  81          * Explicitly disabled through madvise or prctl, or some
  82          * architectures may disable THP for some mappings, for
  83          * example, s390 kvm.
  84          * */
  85         if ((vm_flags & VM_NOHUGEPAGE) ||
  86             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  87                 return false;
  88         /*
  89          * If the hardware/firmware marked hugepage support disabled.
  90          */
  91         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED))
  92                 return false;
  93
  94         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  95         if (vma_is_dax(vma))
  96                 return in_pf;
  97
  98         /*
  99          * Special VMA and hugetlb VMA.
 100          * Must be checked after dax since some dax mappings may have
 101          * VM_MIXEDMAP set.
 102          */
 103         if (vm_flags & VM_NO_KHUGEPAGED)
 104                 return false;
 105
 106         /*
 107          * Check alignment for file vma and size for both file and anon vma.
 108          *
 109          * Skip the check for page fault. Huge fault does the check in fault
 110          * handlers. And this check is not suitable for huge PUD fault.
 111          */
 112         if (!in_pf &&
 113             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 114                 return false;
 115
 116         /*
 117          * Enabled via shmem mount options or sysfs settings.
 118          * Must be done before hugepage flags check since shmem has its
 119          * own flags.
 120          */
 121         if (!in_pf && shmem_file(vma->vm_file))
 122                 return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
 123                                      !enforce_sysfs, vma->vm_mm, vm_flags);
 124
 125         /* Enforce sysfs THP requirements as necessary */
 126         if (enforce_sysfs &&
 127             (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 128                                            !hugepage_flags_always())))
 129                 return false;
 130
 131         /* Only regular file is valid */
 132         if (!in_pf && file_thp_enabled(vma))
 133                 return true;
 134
 135         if (!vma_is_anonymous(vma))
 136                 return false;
 137
 138         if (vma_is_temporary_stack(vma))
 139                 return false;
 140
 141         /*
 142          * THPeligible bit of smaps should show 1 for proper VMAs even
 143          * though anon_vma is not initialized yet.
 144          *
 145          * Allow page fault since anon_vma may be not initialized until
 146          * the first page fault.
 147          */
 148         if (!vma->anon_vma)
 149                 return (smaps || in_pf);
 150
 151         return true;
 152 }
 153
 154 static bool get_huge_zero_page(void)
 155 {
 156         struct page *zero_page;
 157 retry:
 158         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 159                 return true;
 160
 161         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 162                         HPAGE_PMD_ORDER);
 163         if (!zero_page) {
 164                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 165                 return false;
 166         }
 167         preempt_disable();
 168         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 169                 preempt_enable();
 170                 __free_pages(zero_page, compound_order(zero_page));
 171                 goto retry;
 172         }
 173         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 174
 175         /* We take additional reference here. It will be put back by shrinker */
 176         atomic_set(&huge_zero_refcount, 2);
 177         preempt_enable();
 178         count_vm_event(THP_ZERO_PAGE_ALLOC);
 179         return true;
 180 }
 181
 182 static void put_huge_zero_page(void)
 183 {
 184         /*
 185          * Counter should never go to zero here. Only shrinker can put
 186          * last reference.
 187          */
 188         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 189 }
 190
 191 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 192 {
 193         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 194                 return READ_ONCE(huge_zero_page);
 195
 196         if (!get_huge_zero_page())
 197                 return NULL;
 198
 199         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 200                 put_huge_zero_page();
 201
 202         return READ_ONCE(huge_zero_page);
 203 }
 204
 205 void mm_put_huge_zero_page(struct mm_struct *mm)
 206 {
 207         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 208                 put_huge_zero_page();
 209 }
 210
 211 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 212                                         struct shrink_control *sc)
 213 {
 214         /* we can free zero page only if last reference remains */
 215         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 216 }
 217
 218 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 219                                        struct shrink_control *sc)
 220 {
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 return HPAGE_PMD_NR;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .count_objects = shrink_huge_zero_page_count,
 234         .scan_objects = shrink_huge_zero_page_scan,
 235         .seeks = DEFAULT_SEEKS,
 236 };
 237
 238 #ifdef CONFIG_SYSFS
 239 static ssize_t enabled_show(struct kobject *kobj,
 240                             struct kobj_attribute *attr, char *buf)
 241 {
 242         const char *output;
 243
 244         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 245                 output = "[always] madvise never";
 246         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 247                           &transparent_hugepage_flags))
 248                 output = "always [madvise] never";
 249         else
 250                 output = "always madvise [never]";
 251
 252         return sysfs_emit(buf, "%s\n", output);
 253 }
 254
 255 static ssize_t enabled_store(struct kobject *kobj,
 256                              struct kobj_attribute *attr,
 257                              const char *buf, size_t count)
 258 {
 259         ssize_t ret = count;
 260
 261         if (sysfs_streq(buf, "always")) {
 262                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 263                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 264         } else if (sysfs_streq(buf, "madvise")) {
 265                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 266                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 267         } else if (sysfs_streq(buf, "never")) {
 268                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 269                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else
 271                 ret = -EINVAL;
 272
 273         if (ret > 0) {
 274                 int err = start_stop_khugepaged();
 275                 if (err)
 276                         ret = err;
 277         }
 278         return ret;
 279 }
 280
 281 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 282
 283 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 284                                   struct kobj_attribute *attr, char *buf,
 285                                   enum transparent_hugepage_flag flag)
 286 {
 287         return sysfs_emit(buf, "%d\n",
 288                           !!test_bit(flag, &transparent_hugepage_flags));
 289 }
 290
 291 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 292                                  struct kobj_attribute *attr,
 293                                  const char *buf, size_t count,
 294                                  enum transparent_hugepage_flag flag)
 295 {
 296         unsigned long value;
 297         int ret;
 298
 299         ret = kstrtoul(buf, 10, &value);
 300         if (ret < 0)
 301                 return ret;
 302         if (value > 1)
 303                 return -EINVAL;
 304
 305         if (value)
 306                 set_bit(flag, &transparent_hugepage_flags);
 307         else
 308                 clear_bit(flag, &transparent_hugepage_flags);
 309
 310         return count;
 311 }
 312
 313 static ssize_t defrag_show(struct kobject *kobj,
 314                            struct kobj_attribute *attr, char *buf)
 315 {
 316         const char *output;
 317
 318         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 319                      &transparent_hugepage_flags))
 320                 output = "[always] defer defer+madvise madvise never";
 321         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 322                           &transparent_hugepage_flags))
 323                 output = "always [defer] defer+madvise madvise never";
 324         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 325                           &transparent_hugepage_flags))
 326                 output = "always defer [defer+madvise] madvise never";
 327         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 328                           &transparent_hugepage_flags))
 329                 output = "always defer defer+madvise [madvise] never";
 330         else
 331                 output = "always defer defer+madvise madvise [never]";
 332
 333         return sysfs_emit(buf, "%s\n", output);
 334 }
 335
 336 static ssize_t defrag_store(struct kobject *kobj,
 337                             struct kobj_attribute *attr,
 338                             const char *buf, size_t count)
 339 {
 340         if (sysfs_streq(buf, "always")) {
 341                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 342                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 343                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 344                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 345         } else if (sysfs_streq(buf, "defer+madvise")) {
 346                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 347                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 348                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 349                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 350         } else if (sysfs_streq(buf, "defer")) {
 351                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 352                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 353                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 354                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 355         } else if (sysfs_streq(buf, "madvise")) {
 356                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 357                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 358                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 359                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 360         } else if (sysfs_streq(buf, "never")) {
 361                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 362                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 363                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 364                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 365         } else
 366                 return -EINVAL;
 367
 368         return count;
 369 }
 370 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 371
 372 static ssize_t use_zero_page_show(struct kobject *kobj,
 373                                   struct kobj_attribute *attr, char *buf)
 374 {
 375         return single_hugepage_flag_show(kobj, attr, buf,
 376                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static ssize_t use_zero_page_store(struct kobject *kobj,
 379                 struct kobj_attribute *attr, const char *buf, size_t count)
 380 {
 381         return single_hugepage_flag_store(kobj, attr, buf, count,
 382                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 385
 386 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 387                                    struct kobj_attribute *attr, char *buf)
 388 {
 389         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 390 }
 391 static struct kobj_attribute hpage_pmd_size_attr =
 392         __ATTR_RO(hpage_pmd_size);
 393
 394 static struct attribute *hugepage_attr[] = {
 395         &enabled_attr.attr,
 396         &defrag_attr.attr,
 397         &use_zero_page_attr.attr,
 398         &hpage_pmd_size_attr.attr,
 399 #ifdef CONFIG_SHMEM
 400         &shmem_enabled_attr.attr,
 401 #endif
 402         NULL,
 403 };
 404
 405 static const struct attribute_group hugepage_attr_group = {
 406         .attrs = hugepage_attr,
 407 };
 408
 409 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 410 {
 411         int err;
 412
 413         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 414         if (unlikely(!*hugepage_kobj)) {
 415                 pr_err("failed to create transparent hugepage kobject\n");
 416                 return -ENOMEM;
 417         }
 418
 419         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 420         if (err) {
 421                 pr_err("failed to register transparent hugepage group\n");
 422                 goto delete_obj;
 423         }
 424
 425         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 426         if (err) {
 427                 pr_err("failed to register transparent hugepage group\n");
 428                 goto remove_hp_group;
 429         }
 430
 431         return 0;
 432
 433 remove_hp_group:
 434         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 435 delete_obj:
 436         kobject_put(*hugepage_kobj);
 437         return err;
 438 }
 439
 440 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 441 {
 442         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 443         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 444         kobject_put(hugepage_kobj);
 445 }
 446 #else
 447 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 448 {
 449         return 0;
 450 }
 451
 452 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 453 {
 454 }
 455 #endif /* CONFIG_SYSFS */
 456
 457 static int __init hugepage_init(void)
 458 {
 459         int err;
 460         struct kobject *hugepage_kobj;
 461
 462         if (!has_transparent_hugepage()) {
 463                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
 464                 return -EINVAL;
 465         }
 466
 467         /*
 468          * hugepages can't be allocated by the buddy allocator
 469          */
 470         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_ORDER);
 471         /*
 472          * we use page->mapping and page->index in second tail page
 473          * as list_head: assuming THP order >= 2
 474          */
 475         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 476
 477         err = hugepage_init_sysfs(&hugepage_kobj);
 478         if (err)
 479                 goto err_sysfs;
 480
 481         err = khugepaged_init();
 482         if (err)
 483                 goto err_slab;
 484
 485         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 486         if (err)
 487                 goto err_hzp_shrinker;
 488         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 489         if (err)
 490                 goto err_split_shrinker;
 491
 492         /*
 493          * By default disable transparent hugepages on smaller systems,
 494          * where the extra memory used could hurt more than TLB overhead
 495          * is likely to save.  The admin can still enable it through /sys.
 496          */
 497         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 498                 transparent_hugepage_flags = 0;
 499                 return 0;
 500         }
 501
 502         err = start_stop_khugepaged();
 503         if (err)
 504                 goto err_khugepaged;
 505
 506         return 0;
 507 err_khugepaged:
 508         unregister_shrinker(&deferred_split_shrinker);
 509 err_split_shrinker:
 510         unregister_shrinker(&huge_zero_page_shrinker);
 511 err_hzp_shrinker:
 512         khugepaged_destroy();
 513 err_slab:
 514         hugepage_exit_sysfs(hugepage_kobj);
 515 err_sysfs:
 516         return err;
 517 }
 518 subsys_initcall(hugepage_init);
 519
 520 static int __init setup_transparent_hugepage(char *str)
 521 {
 522         int ret = 0;
 523         if (!str)
 524                 goto out;
 525         if (!strcmp(str, "always")) {
 526                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 527                         &transparent_hugepage_flags);
 528                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 529                           &transparent_hugepage_flags);
 530                 ret = 1;
 531         } else if (!strcmp(str, "madvise")) {
 532                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 533                           &transparent_hugepage_flags);
 534                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 535                         &transparent_hugepage_flags);
 536                 ret = 1;
 537         } else if (!strcmp(str, "never")) {
 538                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 539                           &transparent_hugepage_flags);
 540                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 541                           &transparent_hugepage_flags);
 542                 ret = 1;
 543         }
 544 out:
 545         if (!ret)
 546                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 547         return ret;
 548 }
 549 __setup("transparent_hugepage=", setup_transparent_hugepage);
 550
 551 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 552 {
 553         if (likely(vma->vm_flags & VM_WRITE))
 554                 pmd = pmd_mkwrite(pmd);
 555         return pmd;
 556 }
 557
 558 #ifdef CONFIG_MEMCG
 559 static inline
 560 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 561 {
 562         struct mem_cgroup *memcg = folio_memcg(folio);
 563         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 564
 565         if (memcg)
 566                 return &memcg->deferred_split_queue;
 567         else
 568                 return &pgdat->deferred_split_queue;
 569 }
 570 #else
 571 static inline
 572 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 573 {
 574         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 575
 576         return &pgdat->deferred_split_queue;
 577 }
 578 #endif
 579
 580 void prep_transhuge_page(struct page *page)
 581 {
 582         struct folio *folio = (struct folio *)page;
 583
 584         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
 585         INIT_LIST_HEAD(&folio->_deferred_list);
 586         folio_set_compound_dtor(folio, TRANSHUGE_PAGE_DTOR);
 587 }
 588
 589 static inline bool is_transparent_hugepage(struct page *page)
 590 {
 591         struct folio *folio;
 592
 593         if (!PageCompound(page))
 594                 return false;
 595
 596         folio = page_folio(page);
 597         return is_huge_zero_page(&folio->page) ||
 598                folio->_folio_dtor == TRANSHUGE_PAGE_DTOR;
 599 }
 600
 601 static unsigned long __thp_get_unmapped_area(struct file *filp,
 602                 unsigned long addr, unsigned long len,
 603                 loff_t off, unsigned long flags, unsigned long size)
 604 {
 605         loff_t off_end = off + len;
 606         loff_t off_align = round_up(off, size);
 607         unsigned long len_pad, ret;
 608
 609         if (off_end <= off_align || (off_end - off_align) < size)
 610                 return 0;
 611
 612         len_pad = len + size;
 613         if (len_pad < len || (off + len_pad) < off)
 614                 return 0;
 615
 616         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 617                                               off >> PAGE_SHIFT, flags);
 618
 619         /*
 620          * The failure might be due to length padding. The caller will retry
 621          * without the padding.
 622          */
 623         if (IS_ERR_VALUE(ret))
 624                 return 0;
 625
 626         /*
 627          * Do not try to align to THP boundary if allocation at the address
 628          * hint succeeds.
 629          */
 630         if (ret == addr)
 631                 return addr;
 632
 633         ret += (off - ret) & (size - 1);
 634         return ret;
 635 }
 636
 637 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 638                 unsigned long len, unsigned long pgoff, unsigned long flags)
 639 {
 640         unsigned long ret;
 641         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 642
 643         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 644         if (ret)
 645                 return ret;
 646
 647         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 648 }
 649 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 650
 651 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 652                         struct page *page, gfp_t gfp)
 653 {
 654         struct vm_area_struct *vma = vmf->vma;
 655         struct folio *folio = page_folio(page);
 656         pgtable_t pgtable;
 657         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 658         vm_fault_t ret = 0;
 659
 660         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
 661
 662         if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
 663                 folio_put(folio);
 664                 count_vm_event(THP_FAULT_FALLBACK);
 665                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 666                 return VM_FAULT_FALLBACK;
 667         }
 668         folio_throttle_swaprate(folio, gfp);
 669
 670         pgtable = pte_alloc_one(vma->vm_mm);
 671         if (unlikely(!pgtable)) {
 672                 ret = VM_FAULT_OOM;
 673                 goto release;
 674         }
 675
 676         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 677         /*
 678          * The memory barrier inside __folio_mark_uptodate makes sure that
 679          * clear_huge_page writes become visible before the set_pmd_at()
 680          * write.
 681          */
 682         __folio_mark_uptodate(folio);
 683
 684         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 685         if (unlikely(!pmd_none(*vmf->pmd))) {
 686                 goto unlock_release;
 687         } else {
 688                 pmd_t entry;
 689
 690                 ret = check_stable_address_space(vma->vm_mm);
 691                 if (ret)
 692                         goto unlock_release;
 693
 694                 /* Deliver the page fault to userland */
 695                 if (userfaultfd_missing(vma)) {
 696                         spin_unlock(vmf->ptl);
 697                         folio_put(folio);
 698                         pte_free(vma->vm_mm, pgtable);
 699                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 700                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 701                         return ret;
 702                 }
 703
 704                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 705                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 706                 folio_add_new_anon_rmap(folio, vma, haddr);
 707                 folio_add_lru_vma(folio, vma);
 708                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 709                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 710                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 711                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 712                 mm_inc_nr_ptes(vma->vm_mm);
 713                 spin_unlock(vmf->ptl);
 714                 count_vm_event(THP_FAULT_ALLOC);
 715                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 716         }
 717
 718         return 0;
 719 unlock_release:
 720         spin_unlock(vmf->ptl);
 721 release:
 722         if (pgtable)
 723                 pte_free(vma->vm_mm, pgtable);
 724         folio_put(folio);
 725         return ret;
 726
 727 }
 728
 729 /*
 730  * always: directly stall for all thp allocations
 731  * defer: wake kswapd and fail if not immediately available
 732  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 733  *                fail if not immediately available
 734  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 735  *          available
 736  * never: never stall for any thp allocation
 737  */
 738 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 739 {
 740         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 741
 742         /* Always do synchronous compaction */
 743         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 744                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 745
 746         /* Kick kcompactd and fail quickly */
 747         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 748                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 749
 750         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 751         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 752                 return GFP_TRANSHUGE_LIGHT |
 753                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 754                                         __GFP_KSWAPD_RECLAIM);
 755
 756         /* Only do synchronous compaction if madvised */
 757         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 758                 return GFP_TRANSHUGE_LIGHT |
 759                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 760
 761         return GFP_TRANSHUGE_LIGHT;
 762 }
 763
 764 /* Caller must hold page table lock. */
 765 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 766                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 767                 struct page *zero_page)
 768 {
 769         pmd_t entry;
 770         if (!pmd_none(*pmd))
 771                 return;
 772         entry = mk_pmd(zero_page, vma->vm_page_prot);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         mm_inc_nr_ptes(mm);
 777 }
 778
 779 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 780 {
 781         struct vm_area_struct *vma = vmf->vma;
 782         gfp_t gfp;
 783         struct folio *folio;
 784         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 785
 786         if (!transhuge_vma_suitable(vma, haddr))
 787                 return VM_FAULT_FALLBACK;
 788         if (unlikely(anon_vma_prepare(vma)))
 789                 return VM_FAULT_OOM;
 790         khugepaged_enter_vma(vma, vma->vm_flags);
 791
 792         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 793                         !mm_forbids_zeropage(vma->vm_mm) &&
 794                         transparent_hugepage_use_zero_page()) {
 795                 pgtable_t pgtable;
 796                 struct page *zero_page;
 797                 vm_fault_t ret;
 798                 pgtable = pte_alloc_one(vma->vm_mm);
 799                 if (unlikely(!pgtable))
 800                         return VM_FAULT_OOM;
 801                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 802                 if (unlikely(!zero_page)) {
 803                         pte_free(vma->vm_mm, pgtable);
 804                         count_vm_event(THP_FAULT_FALLBACK);
 805                         return VM_FAULT_FALLBACK;
 806                 }
 807                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 808                 ret = 0;
 809                 if (pmd_none(*vmf->pmd)) {
 810                         ret = check_stable_address_space(vma->vm_mm);
 811                         if (ret) {
 812                                 spin_unlock(vmf->ptl);
 813                                 pte_free(vma->vm_mm, pgtable);
 814                         } else if (userfaultfd_missing(vma)) {
 815                                 spin_unlock(vmf->ptl);
 816                                 pte_free(vma->vm_mm, pgtable);
 817                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 818                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 819                         } else {
 820                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 821                                                    haddr, vmf->pmd, zero_page);
 822                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 823                                 spin_unlock(vmf->ptl);
 824                         }
 825                 } else {
 826                         spin_unlock(vmf->ptl);
 827                         pte_free(vma->vm_mm, pgtable);
 828                 }
 829                 return ret;
 830         }
 831         gfp = vma_thp_gfp_mask(vma);
 832         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 833         if (unlikely(!folio)) {
 834                 count_vm_event(THP_FAULT_FALLBACK);
 835                 return VM_FAULT_FALLBACK;
 836         }
 837         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 838 }
 839
 840 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 841                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 842                 pgtable_t pgtable)
 843 {
 844         struct mm_struct *mm = vma->vm_mm;
 845         pmd_t entry;
 846         spinlock_t *ptl;
 847
 848         ptl = pmd_lock(mm, pmd);
 849         if (!pmd_none(*pmd)) {
 850                 if (write) {
 851                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 852                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 853                                 goto out_unlock;
 854                         }
 855                         entry = pmd_mkyoung(*pmd);
 856                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 857                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 858                                 update_mmu_cache_pmd(vma, addr, pmd);
 859                 }
 860
 861                 goto out_unlock;
 862         }
 863
 864         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 865         if (pfn_t_devmap(pfn))
 866                 entry = pmd_mkdevmap(entry);
 867         if (write) {
 868                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 869                 entry = maybe_pmd_mkwrite(entry, vma);
 870         }
 871
 872         if (pgtable) {
 873                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 874                 mm_inc_nr_ptes(mm);
 875                 pgtable = NULL;
 876         }
 877
 878         set_pmd_at(mm, addr, pmd, entry);
 879         update_mmu_cache_pmd(vma, addr, pmd);
 880
 881 out_unlock:
 882         spin_unlock(ptl);
 883         if (pgtable)
 884                 pte_free(mm, pgtable);
 885 }
 886
 887 /**
 888  * vmf_insert_pfn_pmd - insert a pmd size pfn
 889  * @vmf: Structure describing the fault
 890  * @pfn: pfn to insert
 891  * @write: whether it's a write fault
 892  *
 893  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
 894  *
 895  * Return: vm_fault_t value.
 896  */
 897 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 898 {
 899         unsigned long addr = vmf->address & PMD_MASK;
 900         struct vm_area_struct *vma = vmf->vma;
 901         pgprot_t pgprot = vma->vm_page_prot;
 902         pgtable_t pgtable = NULL;
 903
 904         /*
 905          * If we had pmd_special, we could avoid all these restrictions,
 906          * but we need to be consistent with PTEs and architectures that
 907          * can't support a 'special' bit.
 908          */
 909         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 910                         !pfn_t_devmap(pfn));
 911         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 912                                                 (VM_PFNMAP|VM_MIXEDMAP));
 913         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 914
 915         if (addr < vma->vm_start || addr >= vma->vm_end)
 916                 return VM_FAULT_SIGBUS;
 917
 918         if (arch_needs_pgtable_deposit()) {
 919                 pgtable = pte_alloc_one(vma->vm_mm);
 920                 if (!pgtable)
 921                         return VM_FAULT_OOM;
 922         }
 923
 924         track_pfn_insert(vma, &pgprot, pfn);
 925
 926         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 927         return VM_FAULT_NOPAGE;
 928 }
 929 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 930
 931 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 932 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 933 {
 934         if (likely(vma->vm_flags & VM_WRITE))
 935                 pud = pud_mkwrite(pud);
 936         return pud;
 937 }
 938
 939 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 940                 pud_t *pud, pfn_t pfn, bool write)
 941 {
 942         struct mm_struct *mm = vma->vm_mm;
 943         pgprot_t prot = vma->vm_page_prot;
 944         pud_t entry;
 945         spinlock_t *ptl;
 946
 947         ptl = pud_lock(mm, pud);
 948         if (!pud_none(*pud)) {
 949                 if (write) {
 950                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 951                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 952                                 goto out_unlock;
 953                         }
 954                         entry = pud_mkyoung(*pud);
 955                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 956                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 957                                 update_mmu_cache_pud(vma, addr, pud);
 958                 }
 959                 goto out_unlock;
 960         }
 961
 962         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 963         if (pfn_t_devmap(pfn))
 964                 entry = pud_mkdevmap(entry);
 965         if (write) {
 966                 entry = pud_mkyoung(pud_mkdirty(entry));
 967                 entry = maybe_pud_mkwrite(entry, vma);
 968         }
 969         set_pud_at(mm, addr, pud, entry);
 970         update_mmu_cache_pud(vma, addr, pud);
 971
 972 out_unlock:
 973         spin_unlock(ptl);
 974 }
 975
 976 /**
 977  * vmf_insert_pfn_pud - insert a pud size pfn
 978  * @vmf: Structure describing the fault
 979  * @pfn: pfn to insert
 980  * @write: whether it's a write fault
 981  *
 982  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
 983  *
 984  * Return: vm_fault_t value.
 985  */
 986 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 987 {
 988         unsigned long addr = vmf->address & PUD_MASK;
 989         struct vm_area_struct *vma = vmf->vma;
 990         pgprot_t pgprot = vma->vm_page_prot;
 991
 992         /*
 993          * If we had pud_special, we could avoid all these restrictions,
 994          * but we need to be consistent with PTEs and architectures that
 995          * can't support a 'special' bit.
 996          */
 997         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 998                         !pfn_t_devmap(pfn));
 999         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1000                                                 (VM_PFNMAP|VM_MIXEDMAP));
1001         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1002
1003         if (addr < vma->vm_start || addr >= vma->vm_end)
1004                 return VM_FAULT_SIGBUS;
1005
1006         track_pfn_insert(vma, &pgprot, pfn);
1007
1008         insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1009         return VM_FAULT_NOPAGE;
1010 }
1011 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1013
1014 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1015                       pmd_t *pmd, bool write)
1016 {
1017         pmd_t _pmd;
1018
1019         _pmd = pmd_mkyoung(*pmd);
1020         if (write)
1021                 _pmd = pmd_mkdirty(_pmd);
1022         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1023                                   pmd, _pmd, write))
1024                 update_mmu_cache_pmd(vma, addr, pmd);
1025 }
1026
1027 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1028                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1029 {
1030         unsigned long pfn = pmd_pfn(*pmd);
1031         struct mm_struct *mm = vma->vm_mm;
1032         struct page *page;
1033         int ret;
1034
1035         assert_spin_locked(pmd_lockptr(mm, pmd));
1036
1037         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1038                 return NULL;
1039
1040         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1041                 /* pass */;
1042         else
1043                 return NULL;
1044
1045         if (flags & FOLL_TOUCH)
1046                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1047
1048         /*
1049          * device mapped pages can only be returned if the
1050          * caller will manage the page reference count.
1051          */
1052         if (!(flags & (FOLL_GET | FOLL_PIN)))
1053                 return ERR_PTR(-EEXIST);
1054
1055         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1056         *pgmap = get_dev_pagemap(pfn, *pgmap);
1057         if (!*pgmap)
1058                 return ERR_PTR(-EFAULT);
1059         page = pfn_to_page(pfn);
1060         ret = try_grab_page(page, flags);
1061         if (ret)
1062                 page = ERR_PTR(ret);
1063
1064         return page;
1065 }
1066
1067 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1068                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1069                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1070 {
1071         spinlock_t *dst_ptl, *src_ptl;
1072         struct page *src_page;
1073         pmd_t pmd;
1074         pgtable_t pgtable = NULL;
1075         int ret = -ENOMEM;
1076
1077         /* Skip if can be re-fill on fault */
1078         if (!vma_is_anonymous(dst_vma))
1079                 return 0;
1080
1081         pgtable = pte_alloc_one(dst_mm);
1082         if (unlikely(!pgtable))
1083                 goto out;
1084
1085         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1086         src_ptl = pmd_lockptr(src_mm, src_pmd);
1087         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1088
1089         ret = -EAGAIN;
1090         pmd = *src_pmd;
1091
1092 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1093         if (unlikely(is_swap_pmd(pmd))) {
1094                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1095
1096                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1097                 if (!is_readable_migration_entry(entry)) {
1098                         entry = make_readable_migration_entry(
1099                                                         swp_offset(entry));
1100                         pmd = swp_entry_to_pmd(entry);
1101                         if (pmd_swp_soft_dirty(*src_pmd))
1102                                 pmd = pmd_swp_mksoft_dirty(pmd);
1103                         if (pmd_swp_uffd_wp(*src_pmd))
1104                                 pmd = pmd_swp_mkuffd_wp(pmd);
1105                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1106                 }
1107                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1108                 mm_inc_nr_ptes(dst_mm);
1109                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1110                 if (!userfaultfd_wp(dst_vma))
1111                         pmd = pmd_swp_clear_uffd_wp(pmd);
1112                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1113                 ret = 0;
1114                 goto out_unlock;
1115         }
1116 #endif
1117
1118         if (unlikely(!pmd_trans_huge(pmd))) {
1119                 pte_free(dst_mm, pgtable);
1120                 goto out_unlock;
1121         }
1122         /*
1123          * When page table lock is held, the huge zero pmd should not be
1124          * under splitting since we don't split the page itself, only pmd to
1125          * a page table.
1126          */
1127         if (is_huge_zero_pmd(pmd)) {
1128                 /*
1129                  * get_huge_zero_page() will never allocate a new page here,
1130                  * since we already have a zero page to copy. It just takes a
1131                  * reference.
1132                  */
1133                 mm_get_huge_zero_page(dst_mm);
1134                 goto out_zero_page;
1135         }
1136
1137         src_page = pmd_page(pmd);
1138         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1139
1140         get_page(src_page);
1141         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1142                 /* Page maybe pinned: split and retry the fault on PTEs. */
1143                 put_page(src_page);
1144                 pte_free(dst_mm, pgtable);
1145                 spin_unlock(src_ptl);
1146                 spin_unlock(dst_ptl);
1147                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1148                 return -EAGAIN;
1149         }
1150         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1151 out_zero_page:
1152         mm_inc_nr_ptes(dst_mm);
1153         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1154         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1155         if (!userfaultfd_wp(dst_vma))
1156                 pmd = pmd_clear_uffd_wp(pmd);
1157         pmd = pmd_mkold(pmd_wrprotect(pmd));
1158         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1159
1160         ret = 0;
1161 out_unlock:
1162         spin_unlock(src_ptl);
1163         spin_unlock(dst_ptl);
1164 out:
1165         return ret;
1166 }
1167
1168 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1169 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1170                       pud_t *pud, bool write)
1171 {
1172         pud_t _pud;
1173
1174         _pud = pud_mkyoung(*pud);
1175         if (write)
1176                 _pud = pud_mkdirty(_pud);
1177         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1178                                   pud, _pud, write))
1179                 update_mmu_cache_pud(vma, addr, pud);
1180 }
1181
1182 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1183                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1184 {
1185         unsigned long pfn = pud_pfn(*pud);
1186         struct mm_struct *mm = vma->vm_mm;
1187         struct page *page;
1188         int ret;
1189
1190         assert_spin_locked(pud_lockptr(mm, pud));
1191
1192         if (flags & FOLL_WRITE && !pud_write(*pud))
1193                 return NULL;
1194
1195         if (pud_present(*pud) && pud_devmap(*pud))
1196                 /* pass */;
1197         else
1198                 return NULL;
1199
1200         if (flags & FOLL_TOUCH)
1201                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1202
1203         /*
1204          * device mapped pages can only be returned if the
1205          * caller will manage the page reference count.
1206          *
1207          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1208          */
1209         if (!(flags & (FOLL_GET | FOLL_PIN)))
1210                 return ERR_PTR(-EEXIST);
1211
1212         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1213         *pgmap = get_dev_pagemap(pfn, *pgmap);
1214         if (!*pgmap)
1215                 return ERR_PTR(-EFAULT);
1216         page = pfn_to_page(pfn);
1217
1218         ret = try_grab_page(page, flags);
1219         if (ret)
1220                 page = ERR_PTR(ret);
1221
1222         return page;
1223 }
1224
1225 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1226                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1227                   struct vm_area_struct *vma)
1228 {
1229         spinlock_t *dst_ptl, *src_ptl;
1230         pud_t pud;
1231         int ret;
1232
1233         dst_ptl = pud_lock(dst_mm, dst_pud);
1234         src_ptl = pud_lockptr(src_mm, src_pud);
1235         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1236
1237         ret = -EAGAIN;
1238         pud = *src_pud;
1239         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1240                 goto out_unlock;
1241
1242         /*
1243          * When page table lock is held, the huge zero pud should not be
1244          * under splitting since we don't split the page itself, only pud to
1245          * a page table.
1246          */
1247         if (is_huge_zero_pud(pud)) {
1248                 /* No huge zero pud yet */
1249         }
1250
1251         /*
1252          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1253          * and split if duplicating fails.
1254          */
1255         pudp_set_wrprotect(src_mm, addr, src_pud);
1256         pud = pud_mkold(pud_wrprotect(pud));
1257         set_pud_at(dst_mm, addr, dst_pud, pud);
1258
1259         ret = 0;
1260 out_unlock:
1261         spin_unlock(src_ptl);
1262         spin_unlock(dst_ptl);
1263         return ret;
1264 }
1265
1266 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1267 {
1268         bool write = vmf->flags & FAULT_FLAG_WRITE;
1269
1270         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1271         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1272                 goto unlock;
1273
1274         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1275 unlock:
1276         spin_unlock(vmf->ptl);
1277 }
1278 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1279
1280 void huge_pmd_set_accessed(struct vm_fault *vmf)
1281 {
1282         bool write = vmf->flags & FAULT_FLAG_WRITE;
1283
1284         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1285         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1286                 goto unlock;
1287
1288         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1289
1290 unlock:
1291         spin_unlock(vmf->ptl);
1292 }
1293
1294 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1295 {
1296         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1297         struct vm_area_struct *vma = vmf->vma;
1298         struct folio *folio;
1299         struct page *page;
1300         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1301         pmd_t orig_pmd = vmf->orig_pmd;
1302
1303         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1304         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1305
1306         if (is_huge_zero_pmd(orig_pmd))
1307                 goto fallback;
1308
1309         spin_lock(vmf->ptl);
1310
1311         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1312                 spin_unlock(vmf->ptl);
1313                 return 0;
1314         }
1315
1316         page = pmd_page(orig_pmd);
1317         folio = page_folio(page);
1318         VM_BUG_ON_PAGE(!PageHead(page), page);
1319
1320         /* Early check when only holding the PT lock. */
1321         if (PageAnonExclusive(page))
1322                 goto reuse;
1323
1324         if (!folio_trylock(folio)) {
1325                 folio_get(folio);
1326                 spin_unlock(vmf->ptl);
1327                 folio_lock(folio);
1328                 spin_lock(vmf->ptl);
1329                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1330                         spin_unlock(vmf->ptl);
1331                         folio_unlock(folio);
1332                         folio_put(folio);
1333                         return 0;
1334                 }
1335                 folio_put(folio);
1336         }
1337
1338         /* Recheck after temporarily dropping the PT lock. */
1339         if (PageAnonExclusive(page)) {
1340                 folio_unlock(folio);
1341                 goto reuse;
1342         }
1343
1344         /*
1345          * See do_wp_page(): we can only reuse the folio exclusively if
1346          * there are no additional references. Note that we always drain
1347          * the LRU cache immediately after adding a THP.
1348          */
1349         if (folio_ref_count(folio) >
1350                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1351                 goto unlock_fallback;
1352         if (folio_test_swapcache(folio))
1353                 folio_free_swap(folio);
1354         if (folio_ref_count(folio) == 1) {
1355                 pmd_t entry;
1356
1357                 page_move_anon_rmap(page, vma);
1358                 folio_unlock(folio);
1359 reuse:
1360                 if (unlikely(unshare)) {
1361                         spin_unlock(vmf->ptl);
1362                         return 0;
1363                 }
1364                 entry = pmd_mkyoung(orig_pmd);
1365                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1366                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1367                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1368                 spin_unlock(vmf->ptl);
1369                 return 0;
1370         }
1371
1372 unlock_fallback:
1373         folio_unlock(folio);
1374         spin_unlock(vmf->ptl);
1375 fallback:
1376         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1377         return VM_FAULT_FALLBACK;
1378 }
1379
1380 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1381                                            unsigned long addr, pmd_t pmd)
1382 {
1383         struct page *page;
1384
1385         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1386                 return false;
1387
1388         /* Don't touch entries that are not even readable (NUMA hinting). */
1389         if (pmd_protnone(pmd))
1390                 return false;
1391
1392         /* Do we need write faults for softdirty tracking? */
1393         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1394                 return false;
1395
1396         /* Do we need write faults for uffd-wp tracking? */
1397         if (userfaultfd_huge_pmd_wp(vma, pmd))
1398                 return false;
1399
1400         if (!(vma->vm_flags & VM_SHARED)) {
1401                 /* See can_change_pte_writable(). */
1402                 page = vm_normal_page_pmd(vma, addr, pmd);
1403                 return page && PageAnon(page) && PageAnonExclusive(page);
1404         }
1405
1406         /* See can_change_pte_writable(). */
1407         return pmd_dirty(pmd);
1408 }
1409
1410 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1411 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1412                                         struct vm_area_struct *vma,
1413                                         unsigned int flags)
1414 {
1415         /* If the pmd is writable, we can write to the page. */
1416         if (pmd_write(pmd))
1417                 return true;
1418
1419         /* Maybe FOLL_FORCE is set to override it? */
1420         if (!(flags & FOLL_FORCE))
1421                 return false;
1422
1423         /* But FOLL_FORCE has no effect on shared mappings */
1424         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1425                 return false;
1426
1427         /* ... or read-only private ones */
1428         if (!(vma->vm_flags & VM_MAYWRITE))
1429                 return false;
1430
1431         /* ... or already writable ones that just need to take a write fault */
1432         if (vma->vm_flags & VM_WRITE)
1433                 return false;
1434
1435         /*
1436          * See can_change_pte_writable(): we broke COW and could map the page
1437          * writable if we have an exclusive anonymous page ...
1438          */
1439         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1440                 return false;
1441
1442         /* ... and a write-fault isn't required for other reasons. */
1443         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1444                 return false;
1445         return !userfaultfd_huge_pmd_wp(vma, pmd);
1446 }
1447
1448 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1449                                    unsigned long addr,
1450                                    pmd_t *pmd,
1451                                    unsigned int flags)
1452 {
1453         struct mm_struct *mm = vma->vm_mm;
1454         struct page *page;
1455         int ret;
1456
1457         assert_spin_locked(pmd_lockptr(mm, pmd));
1458
1459         page = pmd_page(*pmd);
1460         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1461
1462         if ((flags & FOLL_WRITE) &&
1463             !can_follow_write_pmd(*pmd, page, vma, flags))
1464                 return NULL;
1465
1466         /* Avoid dumping huge zero page */
1467         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1468                 return ERR_PTR(-EFAULT);
1469
1470         /* Full NUMA hinting faults to serialise migration in fault paths */
1471         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1472                 return NULL;
1473
1474         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1475                 return ERR_PTR(-EMLINK);
1476
1477         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1478                         !PageAnonExclusive(page), page);
1479
1480         ret = try_grab_page(page, flags);
1481         if (ret)
1482                 return ERR_PTR(ret);
1483
1484         if (flags & FOLL_TOUCH)
1485                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1486
1487         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1488         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1489
1490         return page;
1491 }
1492
1493 /* NUMA hinting page fault entry point for trans huge pmds */
1494 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1495 {
1496         struct vm_area_struct *vma = vmf->vma;
1497         pmd_t oldpmd = vmf->orig_pmd;
1498         pmd_t pmd;
1499         struct page *page;
1500         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1501         int page_nid = NUMA_NO_NODE;
1502         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1503         bool migrated = false, writable = false;
1504         int flags = 0;
1505
1506         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1507         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1508                 spin_unlock(vmf->ptl);
1509                 goto out;
1510         }
1511
1512         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1513
1514         /*
1515          * Detect now whether the PMD could be writable; this information
1516          * is only valid while holding the PT lock.
1517          */
1518         writable = pmd_write(pmd);
1519         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1520             can_change_pmd_writable(vma, vmf->address, pmd))
1521                 writable = true;
1522
1523         page = vm_normal_page_pmd(vma, haddr, pmd);
1524         if (!page)
1525                 goto out_map;
1526
1527         /* See similar comment in do_numa_page for explanation */
1528         if (!writable)
1529                 flags |= TNF_NO_GROUP;
1530
1531         page_nid = page_to_nid(page);
1532         /*
1533          * For memory tiering mode, cpupid of slow memory page is used
1534          * to record page access time.  So use default value.
1535          */
1536         if (node_is_toptier(page_nid))
1537                 last_cpupid = page_cpupid_last(page);
1538         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1539                                        &flags);
1540
1541         if (target_nid == NUMA_NO_NODE) {
1542                 put_page(page);
1543                 goto out_map;
1544         }
1545
1546         spin_unlock(vmf->ptl);
1547         writable = false;
1548
1549         migrated = migrate_misplaced_page(page, vma, target_nid);
1550         if (migrated) {
1551                 flags |= TNF_MIGRATED;
1552                 page_nid = target_nid;
1553         } else {
1554                 flags |= TNF_MIGRATE_FAIL;
1555                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1556                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1557                         spin_unlock(vmf->ptl);
1558                         goto out;
1559                 }
1560                 goto out_map;
1561         }
1562
1563 out:
1564         if (page_nid != NUMA_NO_NODE)
1565                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1566                                 flags);
1567
1568         return 0;
1569
1570 out_map:
1571         /* Restore the PMD */
1572         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1573         pmd = pmd_mkyoung(pmd);
1574         if (writable)
1575                 pmd = pmd_mkwrite(pmd);
1576         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1577         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1578         spin_unlock(vmf->ptl);
1579         goto out;
1580 }
1581
1582 /*
1583  * Return true if we do MADV_FREE successfully on entire pmd page.
1584  * Otherwise, return false.
1585  */
1586 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1587                 pmd_t *pmd, unsigned long addr, unsigned long next)
1588 {
1589         spinlock_t *ptl;
1590         pmd_t orig_pmd;
1591         struct folio *folio;
1592         struct mm_struct *mm = tlb->mm;
1593         bool ret = false;
1594
1595         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1596
1597         ptl = pmd_trans_huge_lock(pmd, vma);
1598         if (!ptl)
1599                 goto out_unlocked;
1600
1601         orig_pmd = *pmd;
1602         if (is_huge_zero_pmd(orig_pmd))
1603                 goto out;
1604
1605         if (unlikely(!pmd_present(orig_pmd))) {
1606                 VM_BUG_ON(thp_migration_supported() &&
1607                                   !is_pmd_migration_entry(orig_pmd));
1608                 goto out;
1609         }
1610
1611         folio = pfn_folio(pmd_pfn(orig_pmd));
1612         /*
1613          * If other processes are mapping this folio, we couldn't discard
1614          * the folio unless they all do MADV_FREE so let's skip the folio.
1615          */
1616         if (folio_mapcount(folio) != 1)
1617                 goto out;
1618
1619         if (!folio_trylock(folio))
1620                 goto out;
1621
1622         /*
1623          * If user want to discard part-pages of THP, split it so MADV_FREE
1624          * will deactivate only them.
1625          */
1626         if (next - addr != HPAGE_PMD_SIZE) {
1627                 folio_get(folio);
1628                 spin_unlock(ptl);
1629                 split_folio(folio);
1630                 folio_unlock(folio);
1631                 folio_put(folio);
1632                 goto out_unlocked;
1633         }
1634
1635         if (folio_test_dirty(folio))
1636                 folio_clear_dirty(folio);
1637         folio_unlock(folio);
1638
1639         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1640                 pmdp_invalidate(vma, addr, pmd);
1641                 orig_pmd = pmd_mkold(orig_pmd);
1642                 orig_pmd = pmd_mkclean(orig_pmd);
1643
1644                 set_pmd_at(mm, addr, pmd, orig_pmd);
1645                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1646         }
1647
1648         folio_mark_lazyfree(folio);
1649         ret = true;
1650 out:
1651         spin_unlock(ptl);
1652 out_unlocked:
1653         return ret;
1654 }
1655
1656 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1657 {
1658         pgtable_t pgtable;
1659
1660         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1661         pte_free(mm, pgtable);
1662         mm_dec_nr_ptes(mm);
1663 }
1664
1665 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1666                  pmd_t *pmd, unsigned long addr)
1667 {
1668         pmd_t orig_pmd;
1669         spinlock_t *ptl;
1670
1671         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1672
1673         ptl = __pmd_trans_huge_lock(pmd, vma);
1674         if (!ptl)
1675                 return 0;
1676         /*
1677          * For architectures like ppc64 we look at deposited pgtable
1678          * when calling pmdp_huge_get_and_clear. So do the
1679          * pgtable_trans_huge_withdraw after finishing pmdp related
1680          * operations.
1681          */
1682         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1683                                                 tlb->fullmm);
1684         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1685         if (vma_is_special_huge(vma)) {
1686                 if (arch_needs_pgtable_deposit())
1687                         zap_deposited_table(tlb->mm, pmd);
1688                 spin_unlock(ptl);
1689         } else if (is_huge_zero_pmd(orig_pmd)) {
1690                 zap_deposited_table(tlb->mm, pmd);
1691                 spin_unlock(ptl);
1692         } else {
1693                 struct page *page = NULL;
1694                 int flush_needed = 1;
1695
1696                 if (pmd_present(orig_pmd)) {
1697                         page = pmd_page(orig_pmd);
1698                         page_remove_rmap(page, vma, true);
1699                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1700                         VM_BUG_ON_PAGE(!PageHead(page), page);
1701                 } else if (thp_migration_supported()) {
1702                         swp_entry_t entry;
1703
1704                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1705                         entry = pmd_to_swp_entry(orig_pmd);
1706                         page = pfn_swap_entry_to_page(entry);
1707                         flush_needed = 0;
1708                 } else
1709                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1710
1711                 if (PageAnon(page)) {
1712                         zap_deposited_table(tlb->mm, pmd);
1713                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1714                 } else {
1715                         if (arch_needs_pgtable_deposit())
1716                                 zap_deposited_table(tlb->mm, pmd);
1717                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1718                 }
1719
1720                 spin_unlock(ptl);
1721                 if (flush_needed)
1722                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1723         }
1724         return 1;
1725 }
1726
1727 #ifndef pmd_move_must_withdraw
1728 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1729                                          spinlock_t *old_pmd_ptl,
1730                                          struct vm_area_struct *vma)
1731 {
1732         /*
1733          * With split pmd lock we also need to move preallocated
1734          * PTE page table if new_pmd is on different PMD page table.
1735          *
1736          * We also don't deposit and withdraw tables for file pages.
1737          */
1738         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1739 }
1740 #endif
1741
1742 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1743 {
1744 #ifdef CONFIG_MEM_SOFT_DIRTY
1745         if (unlikely(is_pmd_migration_entry(pmd)))
1746                 pmd = pmd_swp_mksoft_dirty(pmd);
1747         else if (pmd_present(pmd))
1748                 pmd = pmd_mksoft_dirty(pmd);
1749 #endif
1750         return pmd;
1751 }
1752
1753 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1754                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1755 {
1756         spinlock_t *old_ptl, *new_ptl;
1757         pmd_t pmd;
1758         struct mm_struct *mm = vma->vm_mm;
1759         bool force_flush = false;
1760
1761         /*
1762          * The destination pmd shouldn't be established, free_pgtables()
1763          * should have released it; but move_page_tables() might have already
1764          * inserted a page table, if racing against shmem/file collapse.
1765          */
1766         if (!pmd_none(*new_pmd)) {
1767                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1768                 return false;
1769         }
1770
1771         /*
1772          * We don't have to worry about the ordering of src and dst
1773          * ptlocks because exclusive mmap_lock prevents deadlock.
1774          */
1775         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1776         if (old_ptl) {
1777                 new_ptl = pmd_lockptr(mm, new_pmd);
1778                 if (new_ptl != old_ptl)
1779                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1780                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1781                 if (pmd_present(pmd))
1782                         force_flush = true;
1783                 VM_BUG_ON(!pmd_none(*new_pmd));
1784
1785                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1786                         pgtable_t pgtable;
1787                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1788                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1789                 }
1790                 pmd = move_soft_dirty_pmd(pmd);
1791                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1792                 if (force_flush)
1793                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1794                 if (new_ptl != old_ptl)
1795                         spin_unlock(new_ptl);
1796                 spin_unlock(old_ptl);
1797                 return true;
1798         }
1799         return false;
1800 }
1801
1802 /*
1803  * Returns
1804  *  - 0 if PMD could not be locked
1805  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1806  *      or if prot_numa but THP migration is not supported
1807  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1808  */
1809 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1810                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1811                     unsigned long cp_flags)
1812 {
1813         struct mm_struct *mm = vma->vm_mm;
1814         spinlock_t *ptl;
1815         pmd_t oldpmd, entry;
1816         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1817         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1818         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1819         int ret = 1;
1820
1821         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1822
1823         if (prot_numa && !thp_migration_supported())
1824                 return 1;
1825
1826         ptl = __pmd_trans_huge_lock(pmd, vma);
1827         if (!ptl)
1828                 return 0;
1829
1830 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1831         if (is_swap_pmd(*pmd)) {
1832                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1833                 struct page *page = pfn_swap_entry_to_page(entry);
1834                 pmd_t newpmd;
1835
1836                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1837                 if (is_writable_migration_entry(entry)) {
1838                         /*
1839                          * A protection check is difficult so
1840                          * just be safe and disable write
1841                          */
1842                         if (PageAnon(page))
1843                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1844                         else
1845                                 entry = make_readable_migration_entry(swp_offset(entry));
1846                         newpmd = swp_entry_to_pmd(entry);
1847                         if (pmd_swp_soft_dirty(*pmd))
1848                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1849                 } else {
1850                         newpmd = *pmd;
1851                 }
1852
1853                 if (uffd_wp)
1854                         newpmd = pmd_swp_mkuffd_wp(newpmd);
1855                 else if (uffd_wp_resolve)
1856                         newpmd = pmd_swp_clear_uffd_wp(newpmd);
1857                 if (!pmd_same(*pmd, newpmd))
1858                         set_pmd_at(mm, addr, pmd, newpmd);
1859                 goto unlock;
1860         }
1861 #endif
1862
1863         if (prot_numa) {
1864                 struct page *page;
1865                 bool toptier;
1866                 /*
1867                  * Avoid trapping faults against the zero page. The read-only
1868                  * data is likely to be read-cached on the local CPU and
1869                  * local/remote hits to the zero page are not interesting.
1870                  */
1871                 if (is_huge_zero_pmd(*pmd))
1872                         goto unlock;
1873
1874                 if (pmd_protnone(*pmd))
1875                         goto unlock;
1876
1877                 page = pmd_page(*pmd);
1878                 toptier = node_is_toptier(page_to_nid(page));
1879                 /*
1880                  * Skip scanning top tier node if normal numa
1881                  * balancing is disabled
1882                  */
1883                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1884                     toptier)
1885                         goto unlock;
1886
1887                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1888                     !toptier)
1889                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1890         }
1891         /*
1892          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1893          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1894          * which is also under mmap_read_lock(mm):
1895          *
1896          *      CPU0:                           CPU1:
1897          *                              change_huge_pmd(prot_numa=1)
1898          *                               pmdp_huge_get_and_clear_notify()
1899          * madvise_dontneed()
1900          *  zap_pmd_range()
1901          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1902          *   // skip the pmd
1903          *                               set_pmd_at();
1904          *                               // pmd is re-established
1905          *
1906          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1907          * which may break userspace.
1908          *
1909          * pmdp_invalidate_ad() is required to make sure we don't miss
1910          * dirty/young flags set by hardware.
1911          */
1912         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1913
1914         entry = pmd_modify(oldpmd, newprot);
1915         if (uffd_wp)
1916                 entry = pmd_mkuffd_wp(entry);
1917         else if (uffd_wp_resolve)
1918                 /*
1919                  * Leave the write bit to be handled by PF interrupt
1920                  * handler, then things like COW could be properly
1921                  * handled.
1922                  */
1923                 entry = pmd_clear_uffd_wp(entry);
1924
1925         /* See change_pte_range(). */
1926         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1927             can_change_pmd_writable(vma, addr, entry))
1928                 entry = pmd_mkwrite(entry);
1929
1930         ret = HPAGE_PMD_NR;
1931         set_pmd_at(mm, addr, pmd, entry);
1932
1933         if (huge_pmd_needs_flush(oldpmd, entry))
1934                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1935 unlock:
1936         spin_unlock(ptl);
1937         return ret;
1938 }
1939
1940 /*
1941  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1942  *
1943  * Note that if it returns page table lock pointer, this routine returns without
1944  * unlocking page table lock. So callers must unlock it.
1945  */
1946 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1947 {
1948         spinlock_t *ptl;
1949         ptl = pmd_lock(vma->vm_mm, pmd);
1950         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1951                         pmd_devmap(*pmd)))
1952                 return ptl;
1953         spin_unlock(ptl);
1954         return NULL;
1955 }
1956
1957 /*
1958  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1959  *
1960  * Note that if it returns page table lock pointer, this routine returns without
1961  * unlocking page table lock. So callers must unlock it.
1962  */
1963 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1964 {
1965         spinlock_t *ptl;
1966
1967         ptl = pud_lock(vma->vm_mm, pud);
1968         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1969                 return ptl;
1970         spin_unlock(ptl);
1971         return NULL;
1972 }
1973
1974 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1975 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1976                  pud_t *pud, unsigned long addr)
1977 {
1978         spinlock_t *ptl;
1979
1980         ptl = __pud_trans_huge_lock(pud, vma);
1981         if (!ptl)
1982                 return 0;
1983
1984         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1985         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1986         if (vma_is_special_huge(vma)) {
1987                 spin_unlock(ptl);
1988                 /* No zero page support yet */
1989         } else {
1990                 /* No support for anonymous PUD pages yet */
1991                 BUG();
1992         }
1993         return 1;
1994 }
1995
1996 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1997                 unsigned long haddr)
1998 {
1999         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2000         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2001         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2002         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2003
2004         count_vm_event(THP_SPLIT_PUD);
2005
2006         pudp_huge_clear_flush_notify(vma, haddr, pud);
2007 }
2008
2009 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2010                 unsigned long address)
2011 {
2012         spinlock_t *ptl;
2013         struct mmu_notifier_range range;
2014
2015         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2016                                 address & HPAGE_PUD_MASK,
2017                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2018         mmu_notifier_invalidate_range_start(&range);
2019         ptl = pud_lock(vma->vm_mm, pud);
2020         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2021                 goto out;
2022         __split_huge_pud_locked(vma, pud, range.start);
2023
2024 out:
2025         spin_unlock(ptl);
2026         /*
2027          * No need to double call mmu_notifier->invalidate_range() callback as
2028          * the above pudp_huge_clear_flush_notify() did already call it.
2029          */
2030         mmu_notifier_invalidate_range_only_end(&range);
2031 }
2032 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2033
2034 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2035                 unsigned long haddr, pmd_t *pmd)
2036 {
2037         struct mm_struct *mm = vma->vm_mm;
2038         pgtable_t pgtable;
2039         pmd_t _pmd, old_pmd;
2040         unsigned long addr;
2041         pte_t *pte;
2042         int i;
2043
2044         /*
2045          * Leave pmd empty until pte is filled note that it is fine to delay
2046          * notification until mmu_notifier_invalidate_range_end() as we are
2047          * replacing a zero pmd write protected page with a zero pte write
2048          * protected page.
2049          *
2050          * See Documentation/mm/mmu_notifier.rst
2051          */
2052         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2053
2054         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2055         pmd_populate(mm, &_pmd, pgtable);
2056
2057         pte = pte_offset_map(&_pmd, haddr);
2058         VM_BUG_ON(!pte);
2059         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2060                 pte_t entry;
2061
2062                 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2063                 entry = pte_mkspecial(entry);
2064                 if (pmd_uffd_wp(old_pmd))
2065                         entry = pte_mkuffd_wp(entry);
2066                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2067                 set_pte_at(mm, addr, pte, entry);
2068                 pte++;
2069         }
2070         pte_unmap(pte - 1);
2071         smp_wmb(); /* make pte visible before pmd */
2072         pmd_populate(mm, pmd, pgtable);
2073 }
2074
2075 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2076                 unsigned long haddr, bool freeze)
2077 {
2078         struct mm_struct *mm = vma->vm_mm;
2079         struct page *page;
2080         pgtable_t pgtable;
2081         pmd_t old_pmd, _pmd;
2082         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2083         bool anon_exclusive = false, dirty = false;
2084         unsigned long addr;
2085         pte_t *pte;
2086         int i;
2087
2088         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2089         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2090         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2091         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2092                                 && !pmd_devmap(*pmd));
2093
2094         count_vm_event(THP_SPLIT_PMD);
2095
2096         if (!vma_is_anonymous(vma)) {
2097                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2098                 /*
2099                  * We are going to unmap this huge page. So
2100                  * just go ahead and zap it
2101                  */
2102                 if (arch_needs_pgtable_deposit())
2103                         zap_deposited_table(mm, pmd);
2104                 if (vma_is_special_huge(vma))
2105                         return;
2106                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2107                         swp_entry_t entry;
2108
2109                         entry = pmd_to_swp_entry(old_pmd);
2110                         page = pfn_swap_entry_to_page(entry);
2111                 } else {
2112                         page = pmd_page(old_pmd);
2113                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2114                                 set_page_dirty(page);
2115                         if (!PageReferenced(page) && pmd_young(old_pmd))
2116                                 SetPageReferenced(page);
2117                         page_remove_rmap(page, vma, true);
2118                         put_page(page);
2119                 }
2120                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2121                 return;
2122         }
2123
2124         if (is_huge_zero_pmd(*pmd)) {
2125                 /*
2126                  * FIXME: Do we want to invalidate secondary mmu by calling
2127                  * mmu_notifier_invalidate_range() see comments below inside
2128                  * __split_huge_pmd() ?
2129                  *
2130                  * We are going from a zero huge page write protected to zero
2131                  * small page also write protected so it does not seems useful
2132                  * to invalidate secondary mmu at this time.
2133                  */
2134                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2135         }
2136
2137         /*
2138          * Up to this point the pmd is present and huge and userland has the
2139          * whole access to the hugepage during the split (which happens in
2140          * place). If we overwrite the pmd with the not-huge version pointing
2141          * to the pte here (which of course we could if all CPUs were bug
2142          * free), userland could trigger a small page size TLB miss on the
2143          * small sized TLB while the hugepage TLB entry is still established in
2144          * the huge TLB. Some CPU doesn't like that.
2145          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2146          * 383 on page 105. Intel should be safe but is also warns that it's
2147          * only safe if the permission and cache attributes of the two entries
2148          * loaded in the two TLB is identical (which should be the case here).
2149          * But it is generally safer to never allow small and huge TLB entries
2150          * for the same virtual address to be loaded simultaneously. So instead
2151          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2152          * current pmd notpresent (atomically because here the pmd_trans_huge
2153          * must remain set at all times on the pmd until the split is complete
2154          * for this pmd), then we flush the SMP TLB and finally we write the
2155          * non-huge version of the pmd entry with pmd_populate.
2156          */
2157         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2158
2159         pmd_migration = is_pmd_migration_entry(old_pmd);
2160         if (unlikely(pmd_migration)) {
2161                 swp_entry_t entry;
2162
2163                 entry = pmd_to_swp_entry(old_pmd);
2164                 page = pfn_swap_entry_to_page(entry);
2165                 write = is_writable_migration_entry(entry);
2166                 if (PageAnon(page))
2167                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2168                 young = is_migration_entry_young(entry);
2169                 dirty = is_migration_entry_dirty(entry);
2170                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2171                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2172         } else {
2173                 page = pmd_page(old_pmd);
2174                 if (pmd_dirty(old_pmd)) {
2175                         dirty = true;
2176                         SetPageDirty(page);
2177                 }
2178                 write = pmd_write(old_pmd);
2179                 young = pmd_young(old_pmd);
2180                 soft_dirty = pmd_soft_dirty(old_pmd);
2181                 uffd_wp = pmd_uffd_wp(old_pmd);
2182
2183                 VM_BUG_ON_PAGE(!page_count(page), page);
2184
2185                 /*
2186                  * Without "freeze", we'll simply split the PMD, propagating the
2187                  * PageAnonExclusive() flag for each PTE by setting it for
2188                  * each subpage -- no need to (temporarily) clear.
2189                  *
2190                  * With "freeze" we want to replace mapped pages by
2191                  * migration entries right away. This is only possible if we
2192                  * managed to clear PageAnonExclusive() -- see
2193                  * set_pmd_migration_entry().
2194                  *
2195                  * In case we cannot clear PageAnonExclusive(), split the PMD
2196                  * only and let try_to_migrate_one() fail later.
2197                  *
2198                  * See page_try_share_anon_rmap(): invalidate PMD first.
2199                  */
2200                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2201                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2202                         freeze = false;
2203                 if (!freeze)
2204                         page_ref_add(page, HPAGE_PMD_NR - 1);
2205         }
2206
2207         /*
2208          * Withdraw the table only after we mark the pmd entry invalid.
2209          * This's critical for some architectures (Power).
2210          */
2211         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2212         pmd_populate(mm, &_pmd, pgtable);
2213
2214         pte = pte_offset_map(&_pmd, haddr);
2215         VM_BUG_ON(!pte);
2216         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2217                 pte_t entry;
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                         if (write)
2246                                 entry = pte_mkwrite(entry);
2247                         if (anon_exclusive)
2248                                 SetPageAnonExclusive(page + i);
2249                         if (!young)
2250                                 entry = pte_mkold(entry);
2251                         /* NOTE: this may set soft-dirty too on some archs */
2252                         if (dirty)
2253                                 entry = pte_mkdirty(entry);
2254                         if (soft_dirty)
2255                                 entry = pte_mksoft_dirty(entry);
2256                         if (uffd_wp)
2257                                 entry = pte_mkuffd_wp(entry);
2258                         page_add_anon_rmap(page + i, vma, addr, false);
2259                 }
2260                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2261                 set_pte_at(mm, addr, pte, entry);
2262                 pte++;
2263         }
2264         pte_unmap(pte - 1);
2265
2266         if (!pmd_migration)
2267                 page_remove_rmap(page, vma, true);
2268         if (freeze)
2269                 put_page(page);
2270
2271         smp_wmb(); /* make pte visible before pmd */
2272         pmd_populate(mm, pmd, pgtable);
2273 }
2274
2275 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2276                 unsigned long address, bool freeze, struct folio *folio)
2277 {
2278         spinlock_t *ptl;
2279         struct mmu_notifier_range range;
2280
2281         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2282                                 address & HPAGE_PMD_MASK,
2283                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2284         mmu_notifier_invalidate_range_start(&range);
2285         ptl = pmd_lock(vma->vm_mm, pmd);
2286
2287         /*
2288          * If caller asks to setup a migration entry, we need a folio to check
2289          * pmd against. Otherwise we can end up replacing wrong folio.
2290          */
2291         VM_BUG_ON(freeze && !folio);
2292         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2293
2294         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2295             is_pmd_migration_entry(*pmd)) {
2296                 /*
2297                  * It's safe to call pmd_page when folio is set because it's
2298                  * guaranteed that pmd is present.
2299                  */
2300                 if (folio && folio != page_folio(pmd_page(*pmd)))
2301                         goto out;
2302                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2303         }
2304
2305 out:
2306         spin_unlock(ptl);
2307         /*
2308          * No need to double call mmu_notifier->invalidate_range() callback.
2309          * They are 3 cases to consider inside __split_huge_pmd_locked():
2310          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2311          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2312          *    fault will trigger a flush_notify before pointing to a new page
2313          *    (it is fine if the secondary mmu keeps pointing to the old zero
2314          *    page in the meantime)
2315          *  3) Split a huge pmd into pte pointing to the same page. No need
2316          *     to invalidate secondary tlb entry they are all still valid.
2317          *     any further changes to individual pte will notify. So no need
2318          *     to call mmu_notifier->invalidate_range()
2319          */
2320         mmu_notifier_invalidate_range_only_end(&range);
2321 }
2322
2323 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2324                 bool freeze, struct folio *folio)
2325 {
2326         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2327
2328         if (!pmd)
2329                 return;
2330
2331         __split_huge_pmd(vma, pmd, address, freeze, folio);
2332 }
2333
2334 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2335 {
2336         /*
2337          * If the new address isn't hpage aligned and it could previously
2338          * contain an hugepage: check if we need to split an huge pmd.
2339          */
2340         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2341             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2342                          ALIGN(address, HPAGE_PMD_SIZE)))
2343                 split_huge_pmd_address(vma, address, false, NULL);
2344 }
2345
2346 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2347                              unsigned long start,
2348                              unsigned long end,
2349                              long adjust_next)
2350 {
2351         /* Check if we need to split start first. */
2352         split_huge_pmd_if_needed(vma, start);
2353
2354         /* Check if we need to split end next. */
2355         split_huge_pmd_if_needed(vma, end);
2356
2357         /*
2358          * If we're also updating the next vma vm_start,
2359          * check if we need to split it.
2360          */
2361         if (adjust_next > 0) {
2362                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2363                 unsigned long nstart = next->vm_start;
2364                 nstart += adjust_next;
2365                 split_huge_pmd_if_needed(next, nstart);
2366         }
2367 }
2368
2369 static void unmap_folio(struct folio *folio)
2370 {
2371         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2372                 TTU_SYNC;
2373
2374         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2375
2376         /*
2377          * Anon pages need migration entries to preserve them, but file
2378          * pages can simply be left unmapped, then faulted back on demand.
2379          * If that is ever changed (perhaps for mlock), update remap_page().
2380          */
2381         if (folio_test_anon(folio))
2382                 try_to_migrate(folio, ttu_flags);
2383         else
2384                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2385 }
2386
2387 static void remap_page(struct folio *folio, unsigned long nr)
2388 {
2389         int i = 0;
2390
2391         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2392         if (!folio_test_anon(folio))
2393                 return;
2394         for (;;) {
2395                 remove_migration_ptes(folio, folio, true);
2396                 i += folio_nr_pages(folio);
2397                 if (i >= nr)
2398                         break;
2399                 folio = folio_next(folio);
2400         }
2401 }
2402
2403 static void lru_add_page_tail(struct page *head, struct page *tail,
2404                 struct lruvec *lruvec, struct list_head *list)
2405 {
2406         VM_BUG_ON_PAGE(!PageHead(head), head);
2407         VM_BUG_ON_PAGE(PageCompound(tail), head);
2408         VM_BUG_ON_PAGE(PageLRU(tail), head);
2409         lockdep_assert_held(&lruvec->lru_lock);
2410
2411         if (list) {
2412                 /* page reclaim is reclaiming a huge page */
2413                 VM_WARN_ON(PageLRU(head));
2414                 get_page(tail);
2415                 list_add_tail(&tail->lru, list);
2416         } else {
2417                 /* head is still on lru (and we have it frozen) */
2418                 VM_WARN_ON(!PageLRU(head));
2419                 if (PageUnevictable(tail))
2420                         tail->mlock_count = 0;
2421                 else
2422                         list_add_tail(&tail->lru, &head->lru);
2423                 SetPageLRU(tail);
2424         }
2425 }
2426
2427 static void __split_huge_page_tail(struct page *head, int tail,
2428                 struct lruvec *lruvec, struct list_head *list)
2429 {
2430         struct page *page_tail = head + tail;
2431
2432         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2433
2434         /*
2435          * Clone page flags before unfreezing refcount.
2436          *
2437          * After successful get_page_unless_zero() might follow flags change,
2438          * for example lock_page() which set PG_waiters.
2439          *
2440          * Note that for mapped sub-pages of an anonymous THP,
2441          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2442          * the migration entry instead from where remap_page() will restore it.
2443          * We can still have PG_anon_exclusive set on effectively unmapped and
2444          * unreferenced sub-pages of an anonymous THP: we can simply drop
2445          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2446          */
2447         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2448         page_tail->flags |= (head->flags &
2449                         ((1L << PG_referenced) |
2450                          (1L << PG_swapbacked) |
2451                          (1L << PG_swapcache) |
2452                          (1L << PG_mlocked) |
2453                          (1L << PG_uptodate) |
2454                          (1L << PG_active) |
2455                          (1L << PG_workingset) |
2456                          (1L << PG_locked) |
2457                          (1L << PG_unevictable) |
2458 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2459                          (1L << PG_arch_2) |
2460                          (1L << PG_arch_3) |
2461 #endif
2462                          (1L << PG_dirty) |
2463                          LRU_GEN_MASK | LRU_REFS_MASK));
2464
2465         /* ->mapping in first and second tail page is replaced by other uses */
2466         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2467                         page_tail);
2468         page_tail->mapping = head->mapping;
2469         page_tail->index = head->index + tail;
2470
2471         /*
2472          * page->private should not be set in tail pages with the exception
2473          * of swap cache pages that store the swp_entry_t in tail pages.
2474          * Fix up and warn once if private is unexpectedly set.
2475          *
2476          * What of 32-bit systems, on which folio->_pincount overlays
2477          * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2478          * pincount must be 0 for folio_ref_freeze() to have succeeded.
2479          */
2480         if (!folio_test_swapcache(page_folio(head))) {
2481                 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2482                 page_tail->private = 0;
2483         }
2484
2485         /* Page flags must be visible before we make the page non-compound. */
2486         smp_wmb();
2487
2488         /*
2489          * Clear PageTail before unfreezing page refcount.
2490          *
2491          * After successful get_page_unless_zero() might follow put_page()
2492          * which needs correct compound_head().
2493          */
2494         clear_compound_head(page_tail);
2495
2496         /* Finally unfreeze refcount. Additional reference from page cache. */
2497         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2498                                           PageSwapCache(head)));
2499
2500         if (page_is_young(head))
2501                 set_page_young(page_tail);
2502         if (page_is_idle(head))
2503                 set_page_idle(page_tail);
2504
2505         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2506
2507         /*
2508          * always add to the tail because some iterators expect new
2509          * pages to show after the currently processed elements - e.g.
2510          * migrate_pages
2511          */
2512         lru_add_page_tail(head, page_tail, lruvec, list);
2513 }
2514
2515 static void __split_huge_page(struct page *page, struct list_head *list,
2516                 pgoff_t end)
2517 {
2518         struct folio *folio = page_folio(page);
2519         struct page *head = &folio->page;
2520         struct lruvec *lruvec;
2521         struct address_space *swap_cache = NULL;
2522         unsigned long offset = 0;
2523         unsigned int nr = thp_nr_pages(head);
2524         int i;
2525
2526         /* complete memcg works before add pages to LRU */
2527         split_page_memcg(head, nr);
2528
2529         if (PageAnon(head) && PageSwapCache(head)) {
2530                 swp_entry_t entry = { .val = page_private(head) };
2531
2532                 offset = swp_offset(entry);
2533                 swap_cache = swap_address_space(entry);
2534                 xa_lock(&swap_cache->i_pages);
2535         }
2536
2537         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2538         lruvec = folio_lruvec_lock(folio);
2539
2540         ClearPageHasHWPoisoned(head);
2541
2542         for (i = nr - 1; i >= 1; i--) {
2543                 __split_huge_page_tail(head, i, lruvec, list);
2544                 /* Some pages can be beyond EOF: drop them from page cache */
2545                 if (head[i].index >= end) {
2546                         struct folio *tail = page_folio(head + i);
2547
2548                         if (shmem_mapping(head->mapping))
2549                                 shmem_uncharge(head->mapping->host, 1);
2550                         else if (folio_test_clear_dirty(tail))
2551                                 folio_account_cleaned(tail,
2552                                         inode_to_wb(folio->mapping->host));
2553                         __filemap_remove_folio(tail, NULL);
2554                         folio_put(tail);
2555                 } else if (!PageAnon(page)) {
2556                         __xa_store(&head->mapping->i_pages, head[i].index,
2557                                         head + i, 0);
2558                 } else if (swap_cache) {
2559                         __xa_store(&swap_cache->i_pages, offset + i,
2560                                         head + i, 0);
2561                 }
2562         }
2563
2564         ClearPageCompound(head);
2565         unlock_page_lruvec(lruvec);
2566         /* Caller disabled irqs, so they are still disabled here */
2567
2568         split_page_owner(head, nr);
2569
2570         /* See comment in __split_huge_page_tail() */
2571         if (PageAnon(head)) {
2572                 /* Additional pin to swap cache */
2573                 if (PageSwapCache(head)) {
2574                         page_ref_add(head, 2);
2575                         xa_unlock(&swap_cache->i_pages);
2576                 } else {
2577                         page_ref_inc(head);
2578                 }
2579         } else {
2580                 /* Additional pin to page cache */
2581                 page_ref_add(head, 2);
2582                 xa_unlock(&head->mapping->i_pages);
2583         }
2584         local_irq_enable();
2585
2586         remap_page(folio, nr);
2587
2588         if (PageSwapCache(head)) {
2589                 swp_entry_t entry = { .val = page_private(head) };
2590
2591                 split_swap_cluster(entry);
2592         }
2593
2594         for (i = 0; i < nr; i++) {
2595                 struct page *subpage = head + i;
2596                 if (subpage == page)
2597                         continue;
2598                 unlock_page(subpage);
2599
2600                 /*
2601                  * Subpages may be freed if there wasn't any mapping
2602                  * like if add_to_swap() is running on a lru page that
2603                  * had its mapping zapped. And freeing these pages
2604                  * requires taking the lru_lock so we do the put_page
2605                  * of the tail pages after the split is complete.
2606                  */
2607                 free_page_and_swap_cache(subpage);
2608         }
2609 }
2610
2611 /* Racy check whether the huge page can be split */
2612 bool can_split_folio(struct folio *folio, int *pextra_pins)
2613 {
2614         int extra_pins;
2615
2616         /* Additional pins from page cache */
2617         if (folio_test_anon(folio))
2618                 extra_pins = folio_test_swapcache(folio) ?
2619                                 folio_nr_pages(folio) : 0;
2620         else
2621                 extra_pins = folio_nr_pages(folio);
2622         if (pextra_pins)
2623                 *pextra_pins = extra_pins;
2624         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2625 }
2626
2627 /*
2628  * This function splits huge page into normal pages. @page can point to any
2629  * subpage of huge page to split. Split doesn't change the position of @page.
2630  *
2631  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2632  * The huge page must be locked.
2633  *
2634  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2635  *
2636  * Both head page and tail pages will inherit mapping, flags, and so on from
2637  * the hugepage.
2638  *
2639  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2640  * they are not mapped.
2641  *
2642  * Returns 0 if the hugepage is split successfully.
2643  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2644  * us.
2645  */
2646 int split_huge_page_to_list(struct page *page, struct list_head *list)
2647 {
2648         struct folio *folio = page_folio(page);
2649         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2650         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2651         struct anon_vma *anon_vma = NULL;
2652         struct address_space *mapping = NULL;
2653         int extra_pins, ret;
2654         pgoff_t end;
2655         bool is_hzp;
2656
2657         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2658         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2659
2660         is_hzp = is_huge_zero_page(&folio->page);
2661         if (is_hzp) {
2662                 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2663                 return -EBUSY;
2664         }
2665
2666         if (folio_test_writeback(folio))
2667                 return -EBUSY;
2668
2669         if (folio_test_anon(folio)) {
2670                 /*
2671                  * The caller does not necessarily hold an mmap_lock that would
2672                  * prevent the anon_vma disappearing so we first we take a
2673                  * reference to it and then lock the anon_vma for write. This
2674                  * is similar to folio_lock_anon_vma_read except the write lock
2675                  * is taken to serialise against parallel split or collapse
2676                  * operations.
2677                  */
2678                 anon_vma = folio_get_anon_vma(folio);
2679                 if (!anon_vma) {
2680                         ret = -EBUSY;
2681                         goto out;
2682                 }
2683                 end = -1;
2684                 mapping = NULL;
2685                 anon_vma_lock_write(anon_vma);
2686         } else {
2687                 gfp_t gfp;
2688
2689                 mapping = folio->mapping;
2690
2691                 /* Truncated ? */
2692                 if (!mapping) {
2693                         ret = -EBUSY;
2694                         goto out;
2695                 }
2696
2697                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2698                                                         GFP_RECLAIM_MASK);
2699
2700                 if (folio_test_private(folio) &&
2701                                 !filemap_release_folio(folio, gfp)) {
2702                         ret = -EBUSY;
2703                         goto out;
2704                 }
2705
2706                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2707                 if (xas_error(&xas)) {
2708                         ret = xas_error(&xas);
2709                         goto out;
2710                 }
2711
2712                 anon_vma = NULL;
2713                 i_mmap_lock_read(mapping);
2714
2715                 /*
2716                  *__split_huge_page() may need to trim off pages beyond EOF:
2717                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2718                  * which cannot be nested inside the page tree lock. So note
2719                  * end now: i_size itself may be changed at any moment, but
2720                  * folio lock is good enough to serialize the trimming.
2721                  */
2722                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2723                 if (shmem_mapping(mapping))
2724                         end = shmem_fallocend(mapping->host, end);
2725         }
2726
2727         /*
2728          * Racy check if we can split the page, before unmap_folio() will
2729          * split PMDs
2730          */
2731         if (!can_split_folio(folio, &extra_pins)) {
2732                 ret = -EAGAIN;
2733                 goto out_unlock;
2734         }
2735
2736         unmap_folio(folio);
2737
2738         /* block interrupt reentry in xa_lock and spinlock */
2739         local_irq_disable();
2740         if (mapping) {
2741                 /*
2742                  * Check if the folio is present in page cache.
2743                  * We assume all tail are present too, if folio is there.
2744                  */
2745                 xas_lock(&xas);
2746                 xas_reset(&xas);
2747                 if (xas_load(&xas) != folio)
2748                         goto fail;
2749         }
2750
2751         /* Prevent deferred_split_scan() touching ->_refcount */
2752         spin_lock(&ds_queue->split_queue_lock);
2753         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2754                 if (!list_empty(&folio->_deferred_list)) {
2755                         ds_queue->split_queue_len--;
2756                         list_del(&folio->_deferred_list);
2757                 }
2758                 spin_unlock(&ds_queue->split_queue_lock);
2759                 if (mapping) {
2760                         int nr = folio_nr_pages(folio);
2761
2762                         xas_split(&xas, folio, folio_order(folio));
2763                         if (folio_test_swapbacked(folio)) {
2764                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2765                                                         -nr);
2766                         } else {
2767                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2768                                                         -nr);
2769                                 filemap_nr_thps_dec(mapping);
2770                         }
2771                 }
2772
2773                 __split_huge_page(page, list, end);
2774                 ret = 0;
2775         } else {
2776                 spin_unlock(&ds_queue->split_queue_lock);
2777 fail:
2778                 if (mapping)
2779                         xas_unlock(&xas);
2780                 local_irq_enable();
2781                 remap_page(folio, folio_nr_pages(folio));
2782                 ret = -EAGAIN;
2783         }
2784
2785 out_unlock:
2786         if (anon_vma) {
2787                 anon_vma_unlock_write(anon_vma);
2788                 put_anon_vma(anon_vma);
2789         }
2790         if (mapping)
2791                 i_mmap_unlock_read(mapping);
2792 out:
2793         xas_destroy(&xas);
2794         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2795         return ret;
2796 }
2797
2798 void free_transhuge_page(struct page *page)
2799 {
2800         struct folio *folio = (struct folio *)page;
2801         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2802         unsigned long flags;
2803
2804         /*
2805          * At this point, there is no one trying to add the folio to
2806          * deferred_list. If folio is not in deferred_list, it's safe
2807          * to check without acquiring the split_queue_lock.
2808          */
2809         if (data_race(!list_empty(&folio->_deferred_list))) {
2810                 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2811                 if (!list_empty(&folio->_deferred_list)) {
2812                         ds_queue->split_queue_len--;
2813                         list_del(&folio->_deferred_list);
2814                 }
2815                 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2816         }
2817         free_compound_page(page);
2818 }
2819
2820 void deferred_split_folio(struct folio *folio)
2821 {
2822         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2823 #ifdef CONFIG_MEMCG
2824         struct mem_cgroup *memcg = folio_memcg(folio);
2825 #endif
2826         unsigned long flags;
2827
2828         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2829
2830         /*
2831          * The try_to_unmap() in page reclaim path might reach here too,
2832          * this may cause a race condition to corrupt deferred split queue.
2833          * And, if page reclaim is already handling the same folio, it is
2834          * unnecessary to handle it again in shrinker.
2835          *
2836          * Check the swapcache flag to determine if the folio is being
2837          * handled by page reclaim since THP swap would add the folio into
2838          * swap cache before calling try_to_unmap().
2839          */
2840         if (folio_test_swapcache(folio))
2841                 return;
2842
2843         if (!list_empty(&folio->_deferred_list))
2844                 return;
2845
2846         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2847         if (list_empty(&folio->_deferred_list)) {
2848                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2849                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2850                 ds_queue->split_queue_len++;
2851 #ifdef CONFIG_MEMCG
2852                 if (memcg)
2853                         set_shrinker_bit(memcg, folio_nid(folio),
2854                                          deferred_split_shrinker.id);
2855 #endif
2856         }
2857         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2858 }
2859
2860 static unsigned long deferred_split_count(struct shrinker *shrink,
2861                 struct shrink_control *sc)
2862 {
2863         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2864         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2865
2866 #ifdef CONFIG_MEMCG
2867         if (sc->memcg)
2868                 ds_queue = &sc->memcg->deferred_split_queue;
2869 #endif
2870         return READ_ONCE(ds_queue->split_queue_len);
2871 }
2872
2873 static unsigned long deferred_split_scan(struct shrinker *shrink,
2874                 struct shrink_control *sc)
2875 {
2876         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2877         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2878         unsigned long flags;
2879         LIST_HEAD(list);
2880         struct folio *folio, *next;
2881         int split = 0;
2882
2883 #ifdef CONFIG_MEMCG
2884         if (sc->memcg)
2885                 ds_queue = &sc->memcg->deferred_split_queue;
2886 #endif
2887
2888         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2889         /* Take pin on all head pages to avoid freeing them under us */
2890         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2891                                                         _deferred_list) {
2892                 if (folio_try_get(folio)) {
2893                         list_move(&folio->_deferred_list, &list);
2894                 } else {
2895                         /* We lost race with folio_put() */
2896                         list_del_init(&folio->_deferred_list);
2897                         ds_queue->split_queue_len--;
2898                 }
2899                 if (!--sc->nr_to_scan)
2900                         break;
2901         }
2902         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2903
2904         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2905                 if (!folio_trylock(folio))
2906                         goto next;
2907                 /* split_huge_page() removes page from list on success */
2908                 if (!split_folio(folio))
2909                         split++;
2910                 folio_unlock(folio);
2911 next:
2912                 folio_put(folio);
2913         }
2914
2915         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2916         list_splice_tail(&list, &ds_queue->split_queue);
2917         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2918
2919         /*
2920          * Stop shrinker if we didn't split any page, but the queue is empty.
2921          * This can happen if pages were freed under us.
2922          */
2923         if (!split && list_empty(&ds_queue->split_queue))
2924                 return SHRINK_STOP;
2925         return split;
2926 }
2927
2928 static struct shrinker deferred_split_shrinker = {
2929         .count_objects = deferred_split_count,
2930         .scan_objects = deferred_split_scan,
2931         .seeks = DEFAULT_SEEKS,
2932         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2933                  SHRINKER_NONSLAB,
2934 };
2935
2936 #ifdef CONFIG_DEBUG_FS
2937 static void split_huge_pages_all(void)
2938 {
2939         struct zone *zone;
2940         struct page *page;
2941         struct folio *folio;
2942         unsigned long pfn, max_zone_pfn;
2943         unsigned long total = 0, split = 0;
2944
2945         pr_debug("Split all THPs\n");
2946         for_each_zone(zone) {
2947                 if (!managed_zone(zone))
2948                         continue;
2949                 max_zone_pfn = zone_end_pfn(zone);
2950                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2951                         int nr_pages;
2952
2953                         page = pfn_to_online_page(pfn);
2954                         if (!page || PageTail(page))
2955                                 continue;
2956                         folio = page_folio(page);
2957                         if (!folio_try_get(folio))
2958                                 continue;
2959
2960                         if (unlikely(page_folio(page) != folio))
2961                                 goto next;
2962
2963                         if (zone != folio_zone(folio))
2964                                 goto next;
2965
2966                         if (!folio_test_large(folio)
2967                                 || folio_test_hugetlb(folio)
2968                                 || !folio_test_lru(folio))
2969                                 goto next;
2970
2971                         total++;
2972                         folio_lock(folio);
2973                         nr_pages = folio_nr_pages(folio);
2974                         if (!split_folio(folio))
2975                                 split++;
2976                         pfn += nr_pages - 1;
2977                         folio_unlock(folio);
2978 next:
2979                         folio_put(folio);
2980                         cond_resched();
2981                 }
2982         }
2983
2984         pr_debug("%lu of %lu THP split\n", split, total);
2985 }
2986
2987 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2988 {
2989         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2990                     is_vm_hugetlb_page(vma);
2991 }
2992
2993 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2994                                 unsigned long vaddr_end)
2995 {
2996         int ret = 0;
2997         struct task_struct *task;
2998         struct mm_struct *mm;
2999         unsigned long total = 0, split = 0;
3000         unsigned long addr;
3001
3002         vaddr_start &= PAGE_MASK;
3003         vaddr_end &= PAGE_MASK;
3004
3005         /* Find the task_struct from pid */
3006         rcu_read_lock();
3007         task = find_task_by_vpid(pid);
3008         if (!task) {
3009                 rcu_read_unlock();
3010                 ret = -ESRCH;
3011                 goto out;
3012         }
3013         get_task_struct(task);
3014         rcu_read_unlock();
3015
3016         /* Find the mm_struct */
3017         mm = get_task_mm(task);
3018         put_task_struct(task);
3019
3020         if (!mm) {
3021                 ret = -EINVAL;
3022                 goto out;
3023         }
3024
3025         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3026                  pid, vaddr_start, vaddr_end);
3027
3028         mmap_read_lock(mm);
3029         /*
3030          * always increase addr by PAGE_SIZE, since we could have a PTE page
3031          * table filled with PTE-mapped THPs, each of which is distinct.
3032          */
3033         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3034                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3035                 struct page *page;
3036
3037                 if (!vma)
3038                         break;
3039
3040                 /* skip special VMA and hugetlb VMA */
3041                 if (vma_not_suitable_for_thp_split(vma)) {
3042                         addr = vma->vm_end;
3043                         continue;
3044                 }
3045
3046                 /* FOLL_DUMP to ignore special (like zero) pages */
3047                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3048
3049                 if (IS_ERR_OR_NULL(page))
3050                         continue;
3051
3052                 if (!is_transparent_hugepage(page))
3053                         goto next;
3054
3055                 total++;
3056                 if (!can_split_folio(page_folio(page), NULL))
3057                         goto next;
3058
3059                 if (!trylock_page(page))
3060                         goto next;
3061
3062                 if (!split_huge_page(page))
3063                         split++;
3064
3065                 unlock_page(page);
3066 next:
3067                 put_page(page);
3068                 cond_resched();
3069         }
3070         mmap_read_unlock(mm);
3071         mmput(mm);
3072
3073         pr_debug("%lu of %lu THP split\n", split, total);
3074
3075 out:
3076         return ret;
3077 }
3078
3079 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3080                                 pgoff_t off_end)
3081 {
3082         struct filename *file;
3083         struct file *candidate;
3084         struct address_space *mapping;
3085         int ret = -EINVAL;
3086         pgoff_t index;
3087         int nr_pages = 1;
3088         unsigned long total = 0, split = 0;
3089
3090         file = getname_kernel(file_path);
3091         if (IS_ERR(file))
3092                 return ret;
3093
3094         candidate = file_open_name(file, O_RDONLY, 0);
3095         if (IS_ERR(candidate))
3096                 goto out;
3097
3098         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3099                  file_path, off_start, off_end);
3100
3101         mapping = candidate->f_mapping;
3102
3103         for (index = off_start; index < off_end; index += nr_pages) {
3104                 struct folio *folio = filemap_get_folio(mapping, index);
3105
3106                 nr_pages = 1;
3107                 if (IS_ERR(folio))
3108                         continue;
3109
3110                 if (!folio_test_large(folio))
3111                         goto next;
3112
3113                 total++;
3114                 nr_pages = folio_nr_pages(folio);
3115
3116                 if (!folio_trylock(folio))
3117                         goto next;
3118
3119                 if (!split_folio(folio))
3120                         split++;
3121
3122                 folio_unlock(folio);
3123 next:
3124                 folio_put(folio);
3125                 cond_resched();
3126         }
3127
3128         filp_close(candidate, NULL);
3129         ret = 0;
3130
3131         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3132 out:
3133         putname(file);
3134         return ret;
3135 }
3136
3137 #define MAX_INPUT_BUF_SZ 255
3138
3139 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3140                                 size_t count, loff_t *ppops)
3141 {
3142         static DEFINE_MUTEX(split_debug_mutex);
3143         ssize_t ret;
3144         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3145         char input_buf[MAX_INPUT_BUF_SZ];
3146         int pid;
3147         unsigned long vaddr_start, vaddr_end;
3148
3149         ret = mutex_lock_interruptible(&split_debug_mutex);
3150         if (ret)
3151                 return ret;
3152
3153         ret = -EFAULT;
3154
3155         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3156         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3157                 goto out;
3158
3159         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3160
3161         if (input_buf[0] == '/') {
3162                 char *tok;
3163                 char *buf = input_buf;
3164                 char file_path[MAX_INPUT_BUF_SZ];
3165                 pgoff_t off_start = 0, off_end = 0;
3166                 size_t input_len = strlen(input_buf);
3167
3168                 tok = strsep(&buf, ",");
3169                 if (tok) {
3170                         strcpy(file_path, tok);
3171                 } else {
3172                         ret = -EINVAL;
3173                         goto out;
3174                 }
3175
3176                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3177                 if (ret != 2) {
3178                         ret = -EINVAL;
3179                         goto out;
3180                 }
3181                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3182                 if (!ret)
3183                         ret = input_len;
3184
3185                 goto out;
3186         }
3187
3188         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3189         if (ret == 1 && pid == 1) {
3190                 split_huge_pages_all();
3191                 ret = strlen(input_buf);
3192                 goto out;
3193         } else if (ret != 3) {
3194                 ret = -EINVAL;
3195                 goto out;
3196         }
3197
3198         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3199         if (!ret)
3200                 ret = strlen(input_buf);
3201 out:
3202         mutex_unlock(&split_debug_mutex);
3203         return ret;
3204
3205 }
3206
3207 static const struct file_operations split_huge_pages_fops = {
3208         .owner   = THIS_MODULE,
3209         .write   = split_huge_pages_write,
3210         .llseek  = no_llseek,
3211 };
3212
3213 static int __init split_huge_pages_debugfs(void)
3214 {
3215         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3216                             &split_huge_pages_fops);
3217         return 0;
3218 }
3219 late_initcall(split_huge_pages_debugfs);
3220 #endif
3221
3222 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3223 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3224                 struct page *page)
3225 {
3226         struct vm_area_struct *vma = pvmw->vma;
3227         struct mm_struct *mm = vma->vm_mm;
3228         unsigned long address = pvmw->address;
3229         bool anon_exclusive;
3230         pmd_t pmdval;
3231         swp_entry_t entry;
3232         pmd_t pmdswp;
3233
3234         if (!(pvmw->pmd && !pvmw->pte))
3235                 return 0;
3236
3237         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3238         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3239
3240         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3241         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3242         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3243                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3244                 return -EBUSY;
3245         }
3246
3247         if (pmd_dirty(pmdval))
3248                 set_page_dirty(page);
3249         if (pmd_write(pmdval))
3250                 entry = make_writable_migration_entry(page_to_pfn(page));
3251         else if (anon_exclusive)
3252                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3253         else
3254                 entry = make_readable_migration_entry(page_to_pfn(page));
3255         if (pmd_young(pmdval))
3256                 entry = make_migration_entry_young(entry);
3257         if (pmd_dirty(pmdval))
3258                 entry = make_migration_entry_dirty(entry);
3259         pmdswp = swp_entry_to_pmd(entry);
3260         if (pmd_soft_dirty(pmdval))
3261                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3262         if (pmd_uffd_wp(pmdval))
3263                 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3264         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3265         page_remove_rmap(page, vma, true);
3266         put_page(page);
3267         trace_set_migration_pmd(address, pmd_val(pmdswp));
3268
3269         return 0;
3270 }
3271
3272 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3273 {
3274         struct vm_area_struct *vma = pvmw->vma;
3275         struct mm_struct *mm = vma->vm_mm;
3276         unsigned long address = pvmw->address;
3277         unsigned long haddr = address & HPAGE_PMD_MASK;
3278         pmd_t pmde;
3279         swp_entry_t entry;
3280
3281         if (!(pvmw->pmd && !pvmw->pte))
3282                 return;
3283
3284         entry = pmd_to_swp_entry(*pvmw->pmd);
3285         get_page(new);
3286         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3287         if (pmd_swp_soft_dirty(*pvmw->pmd))
3288                 pmde = pmd_mksoft_dirty(pmde);
3289         if (is_writable_migration_entry(entry))
3290                 pmde = pmd_mkwrite(pmde);
3291         if (pmd_swp_uffd_wp(*pvmw->pmd))
3292                 pmde = pmd_mkuffd_wp(pmde);
3293         if (!is_migration_entry_young(entry))
3294                 pmde = pmd_mkold(pmde);
3295         /* NOTE: this may contain setting soft-dirty on some archs */
3296         if (PageDirty(new) && is_migration_entry_dirty(entry))
3297                 pmde = pmd_mkdirty(pmde);
3298
3299         if (PageAnon(new)) {
3300                 rmap_t rmap_flags = RMAP_COMPOUND;
3301
3302                 if (!is_readable_migration_entry(entry))
3303                         rmap_flags |= RMAP_EXCLUSIVE;
3304
3305                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3306         } else {
3307                 page_add_file_rmap(new, vma, true);
3308         }
3309         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3310         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3311
3312         /* No need to invalidate - it was non-present before */
3313         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3314         trace_remove_migration_pmd(address, pmd_val(pmde));
3315 }
3316 #endif