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         set_compound_page_dtor(page, 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_prot - insert a pmd size pfn
 889  * @vmf: Structure describing the fault
 890  * @pfn: pfn to insert
 891  * @pgprot: page protection to use
 892  * @write: whether it's a write fault
 893  *
 894  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 895  * also consult the vmf_insert_mixed_prot() documentation when
 896  * @pgprot != @vmf->vma->vm_page_prot.
 897  *
 898  * Return: vm_fault_t value.
 899  */
 900 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 901                                    pgprot_t pgprot, bool write)
 902 {
 903         unsigned long addr = vmf->address & PMD_MASK;
 904         struct vm_area_struct *vma = vmf->vma;
 905         pgtable_t pgtable = NULL;
 906
 907         /*
 908          * If we had pmd_special, we could avoid all these restrictions,
 909          * but we need to be consistent with PTEs and architectures that
 910          * can't support a 'special' bit.
 911          */
 912         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 913                         !pfn_t_devmap(pfn));
 914         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 915                                                 (VM_PFNMAP|VM_MIXEDMAP));
 916         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 917
 918         if (addr < vma->vm_start || addr >= vma->vm_end)
 919                 return VM_FAULT_SIGBUS;
 920
 921         if (arch_needs_pgtable_deposit()) {
 922                 pgtable = pte_alloc_one(vma->vm_mm);
 923                 if (!pgtable)
 924                         return VM_FAULT_OOM;
 925         }
 926
 927         track_pfn_insert(vma, &pgprot, pfn);
 928
 929         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 930         return VM_FAULT_NOPAGE;
 931 }
 932 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 933
 934 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 935 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 936 {
 937         if (likely(vma->vm_flags & VM_WRITE))
 938                 pud = pud_mkwrite(pud);
 939         return pud;
 940 }
 941
 942 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 943                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 944 {
 945         struct mm_struct *mm = vma->vm_mm;
 946         pud_t entry;
 947         spinlock_t *ptl;
 948
 949         ptl = pud_lock(mm, pud);
 950         if (!pud_none(*pud)) {
 951                 if (write) {
 952                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 953                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 954                                 goto out_unlock;
 955                         }
 956                         entry = pud_mkyoung(*pud);
 957                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 958                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 959                                 update_mmu_cache_pud(vma, addr, pud);
 960                 }
 961                 goto out_unlock;
 962         }
 963
 964         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 965         if (pfn_t_devmap(pfn))
 966                 entry = pud_mkdevmap(entry);
 967         if (write) {
 968                 entry = pud_mkyoung(pud_mkdirty(entry));
 969                 entry = maybe_pud_mkwrite(entry, vma);
 970         }
 971         set_pud_at(mm, addr, pud, entry);
 972         update_mmu_cache_pud(vma, addr, pud);
 973
 974 out_unlock:
 975         spin_unlock(ptl);
 976 }
 977
 978 /**
 979  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 980  * @vmf: Structure describing the fault
 981  * @pfn: pfn to insert
 982  * @pgprot: page protection to use
 983  * @write: whether it's a write fault
 984  *
 985  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 986  * also consult the vmf_insert_mixed_prot() documentation when
 987  * @pgprot != @vmf->vma->vm_page_prot.
 988  *
 989  * Return: vm_fault_t value.
 990  */
 991 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 992                                    pgprot_t pgprot, bool write)
 993 {
 994         unsigned long addr = vmf->address & PUD_MASK;
 995         struct vm_area_struct *vma = vmf->vma;
 996
 997         /*
 998          * If we had pud_special, we could avoid all these restrictions,
 999          * but we need to be consistent with PTEs and architectures that
1000          * can't support a 'special' bit.
1001          */
1002         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1003                         !pfn_t_devmap(pfn));
1004         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005                                                 (VM_PFNMAP|VM_MIXEDMAP));
1006         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007
1008         if (addr < vma->vm_start || addr >= vma->vm_end)
1009                 return VM_FAULT_SIGBUS;
1010
1011         track_pfn_insert(vma, &pgprot, pfn);
1012
1013         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1014         return VM_FAULT_NOPAGE;
1015 }
1016 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1017 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1018
1019 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1020                       pmd_t *pmd, bool write)
1021 {
1022         pmd_t _pmd;
1023
1024         _pmd = pmd_mkyoung(*pmd);
1025         if (write)
1026                 _pmd = pmd_mkdirty(_pmd);
1027         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1028                                   pmd, _pmd, write))
1029                 update_mmu_cache_pmd(vma, addr, pmd);
1030 }
1031
1032 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1033                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1034 {
1035         unsigned long pfn = pmd_pfn(*pmd);
1036         struct mm_struct *mm = vma->vm_mm;
1037         struct page *page;
1038         int ret;
1039
1040         assert_spin_locked(pmd_lockptr(mm, pmd));
1041
1042         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1043                 return NULL;
1044
1045         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1046                 /* pass */;
1047         else
1048                 return NULL;
1049
1050         if (flags & FOLL_TOUCH)
1051                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1052
1053         /*
1054          * device mapped pages can only be returned if the
1055          * caller will manage the page reference count.
1056          */
1057         if (!(flags & (FOLL_GET | FOLL_PIN)))
1058                 return ERR_PTR(-EEXIST);
1059
1060         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1061         *pgmap = get_dev_pagemap(pfn, *pgmap);
1062         if (!*pgmap)
1063                 return ERR_PTR(-EFAULT);
1064         page = pfn_to_page(pfn);
1065         ret = try_grab_page(page, flags);
1066         if (ret)
1067                 page = ERR_PTR(ret);
1068
1069         return page;
1070 }
1071
1072 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1073                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1074                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1075 {
1076         spinlock_t *dst_ptl, *src_ptl;
1077         struct page *src_page;
1078         pmd_t pmd;
1079         pgtable_t pgtable = NULL;
1080         int ret = -ENOMEM;
1081
1082         /* Skip if can be re-fill on fault */
1083         if (!vma_is_anonymous(dst_vma))
1084                 return 0;
1085
1086         pgtable = pte_alloc_one(dst_mm);
1087         if (unlikely(!pgtable))
1088                 goto out;
1089
1090         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1091         src_ptl = pmd_lockptr(src_mm, src_pmd);
1092         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1093
1094         ret = -EAGAIN;
1095         pmd = *src_pmd;
1096
1097 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1098         if (unlikely(is_swap_pmd(pmd))) {
1099                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1100
1101                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1102                 if (!is_readable_migration_entry(entry)) {
1103                         entry = make_readable_migration_entry(
1104                                                         swp_offset(entry));
1105                         pmd = swp_entry_to_pmd(entry);
1106                         if (pmd_swp_soft_dirty(*src_pmd))
1107                                 pmd = pmd_swp_mksoft_dirty(pmd);
1108                         if (pmd_swp_uffd_wp(*src_pmd))
1109                                 pmd = pmd_swp_mkuffd_wp(pmd);
1110                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1111                 }
1112                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1113                 mm_inc_nr_ptes(dst_mm);
1114                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1115                 if (!userfaultfd_wp(dst_vma))
1116                         pmd = pmd_swp_clear_uffd_wp(pmd);
1117                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1118                 ret = 0;
1119                 goto out_unlock;
1120         }
1121 #endif
1122
1123         if (unlikely(!pmd_trans_huge(pmd))) {
1124                 pte_free(dst_mm, pgtable);
1125                 goto out_unlock;
1126         }
1127         /*
1128          * When page table lock is held, the huge zero pmd should not be
1129          * under splitting since we don't split the page itself, only pmd to
1130          * a page table.
1131          */
1132         if (is_huge_zero_pmd(pmd)) {
1133                 /*
1134                  * get_huge_zero_page() will never allocate a new page here,
1135                  * since we already have a zero page to copy. It just takes a
1136                  * reference.
1137                  */
1138                 mm_get_huge_zero_page(dst_mm);
1139                 goto out_zero_page;
1140         }
1141
1142         src_page = pmd_page(pmd);
1143         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1144
1145         get_page(src_page);
1146         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1147                 /* Page maybe pinned: split and retry the fault on PTEs. */
1148                 put_page(src_page);
1149                 pte_free(dst_mm, pgtable);
1150                 spin_unlock(src_ptl);
1151                 spin_unlock(dst_ptl);
1152                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1153                 return -EAGAIN;
1154         }
1155         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1156 out_zero_page:
1157         mm_inc_nr_ptes(dst_mm);
1158         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1159         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1160         if (!userfaultfd_wp(dst_vma))
1161                 pmd = pmd_clear_uffd_wp(pmd);
1162         pmd = pmd_mkold(pmd_wrprotect(pmd));
1163         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1164
1165         ret = 0;
1166 out_unlock:
1167         spin_unlock(src_ptl);
1168         spin_unlock(dst_ptl);
1169 out:
1170         return ret;
1171 }
1172
1173 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1174 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1175                       pud_t *pud, bool write)
1176 {
1177         pud_t _pud;
1178
1179         _pud = pud_mkyoung(*pud);
1180         if (write)
1181                 _pud = pud_mkdirty(_pud);
1182         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1183                                   pud, _pud, write))
1184                 update_mmu_cache_pud(vma, addr, pud);
1185 }
1186
1187 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1188                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1189 {
1190         unsigned long pfn = pud_pfn(*pud);
1191         struct mm_struct *mm = vma->vm_mm;
1192         struct page *page;
1193         int ret;
1194
1195         assert_spin_locked(pud_lockptr(mm, pud));
1196
1197         if (flags & FOLL_WRITE && !pud_write(*pud))
1198                 return NULL;
1199
1200         if (pud_present(*pud) && pud_devmap(*pud))
1201                 /* pass */;
1202         else
1203                 return NULL;
1204
1205         if (flags & FOLL_TOUCH)
1206                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1207
1208         /*
1209          * device mapped pages can only be returned if the
1210          * caller will manage the page reference count.
1211          *
1212          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1213          */
1214         if (!(flags & (FOLL_GET | FOLL_PIN)))
1215                 return ERR_PTR(-EEXIST);
1216
1217         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1218         *pgmap = get_dev_pagemap(pfn, *pgmap);
1219         if (!*pgmap)
1220                 return ERR_PTR(-EFAULT);
1221         page = pfn_to_page(pfn);
1222
1223         ret = try_grab_page(page, flags);
1224         if (ret)
1225                 page = ERR_PTR(ret);
1226
1227         return page;
1228 }
1229
1230 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1231                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1232                   struct vm_area_struct *vma)
1233 {
1234         spinlock_t *dst_ptl, *src_ptl;
1235         pud_t pud;
1236         int ret;
1237
1238         dst_ptl = pud_lock(dst_mm, dst_pud);
1239         src_ptl = pud_lockptr(src_mm, src_pud);
1240         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1241
1242         ret = -EAGAIN;
1243         pud = *src_pud;
1244         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1245                 goto out_unlock;
1246
1247         /*
1248          * When page table lock is held, the huge zero pud should not be
1249          * under splitting since we don't split the page itself, only pud to
1250          * a page table.
1251          */
1252         if (is_huge_zero_pud(pud)) {
1253                 /* No huge zero pud yet */
1254         }
1255
1256         /*
1257          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1258          * and split if duplicating fails.
1259          */
1260         pudp_set_wrprotect(src_mm, addr, src_pud);
1261         pud = pud_mkold(pud_wrprotect(pud));
1262         set_pud_at(dst_mm, addr, dst_pud, pud);
1263
1264         ret = 0;
1265 out_unlock:
1266         spin_unlock(src_ptl);
1267         spin_unlock(dst_ptl);
1268         return ret;
1269 }
1270
1271 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1272 {
1273         bool write = vmf->flags & FAULT_FLAG_WRITE;
1274
1275         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1276         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1277                 goto unlock;
1278
1279         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1280 unlock:
1281         spin_unlock(vmf->ptl);
1282 }
1283 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1284
1285 void huge_pmd_set_accessed(struct vm_fault *vmf)
1286 {
1287         bool write = vmf->flags & FAULT_FLAG_WRITE;
1288
1289         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1290         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1291                 goto unlock;
1292
1293         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1294
1295 unlock:
1296         spin_unlock(vmf->ptl);
1297 }
1298
1299 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1300 {
1301         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1302         struct vm_area_struct *vma = vmf->vma;
1303         struct folio *folio;
1304         struct page *page;
1305         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1306         pmd_t orig_pmd = vmf->orig_pmd;
1307
1308         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1309         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1310
1311         if (is_huge_zero_pmd(orig_pmd))
1312                 goto fallback;
1313
1314         spin_lock(vmf->ptl);
1315
1316         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1317                 spin_unlock(vmf->ptl);
1318                 return 0;
1319         }
1320
1321         page = pmd_page(orig_pmd);
1322         folio = page_folio(page);
1323         VM_BUG_ON_PAGE(!PageHead(page), page);
1324
1325         /* Early check when only holding the PT lock. */
1326         if (PageAnonExclusive(page))
1327                 goto reuse;
1328
1329         if (!folio_trylock(folio)) {
1330                 folio_get(folio);
1331                 spin_unlock(vmf->ptl);
1332                 folio_lock(folio);
1333                 spin_lock(vmf->ptl);
1334                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1335                         spin_unlock(vmf->ptl);
1336                         folio_unlock(folio);
1337                         folio_put(folio);
1338                         return 0;
1339                 }
1340                 folio_put(folio);
1341         }
1342
1343         /* Recheck after temporarily dropping the PT lock. */
1344         if (PageAnonExclusive(page)) {
1345                 folio_unlock(folio);
1346                 goto reuse;
1347         }
1348
1349         /*
1350          * See do_wp_page(): we can only reuse the folio exclusively if
1351          * there are no additional references. Note that we always drain
1352          * the LRU pagevecs immediately after adding a THP.
1353          */
1354         if (folio_ref_count(folio) >
1355                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1356                 goto unlock_fallback;
1357         if (folio_test_swapcache(folio))
1358                 folio_free_swap(folio);
1359         if (folio_ref_count(folio) == 1) {
1360                 pmd_t entry;
1361
1362                 page_move_anon_rmap(page, vma);
1363                 folio_unlock(folio);
1364 reuse:
1365                 if (unlikely(unshare)) {
1366                         spin_unlock(vmf->ptl);
1367                         return 0;
1368                 }
1369                 entry = pmd_mkyoung(orig_pmd);
1370                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1371                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1372                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1373                 spin_unlock(vmf->ptl);
1374                 return 0;
1375         }
1376
1377 unlock_fallback:
1378         folio_unlock(folio);
1379         spin_unlock(vmf->ptl);
1380 fallback:
1381         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1382         return VM_FAULT_FALLBACK;
1383 }
1384
1385 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1386                                            unsigned long addr, pmd_t pmd)
1387 {
1388         struct page *page;
1389
1390         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1391                 return false;
1392
1393         /* Don't touch entries that are not even readable (NUMA hinting). */
1394         if (pmd_protnone(pmd))
1395                 return false;
1396
1397         /* Do we need write faults for softdirty tracking? */
1398         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1399                 return false;
1400
1401         /* Do we need write faults for uffd-wp tracking? */
1402         if (userfaultfd_huge_pmd_wp(vma, pmd))
1403                 return false;
1404
1405         if (!(vma->vm_flags & VM_SHARED)) {
1406                 /* See can_change_pte_writable(). */
1407                 page = vm_normal_page_pmd(vma, addr, pmd);
1408                 return page && PageAnon(page) && PageAnonExclusive(page);
1409         }
1410
1411         /* See can_change_pte_writable(). */
1412         return pmd_dirty(pmd);
1413 }
1414
1415 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1416 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1417                                         struct vm_area_struct *vma,
1418                                         unsigned int flags)
1419 {
1420         /* If the pmd is writable, we can write to the page. */
1421         if (pmd_write(pmd))
1422                 return true;
1423
1424         /* Maybe FOLL_FORCE is set to override it? */
1425         if (!(flags & FOLL_FORCE))
1426                 return false;
1427
1428         /* But FOLL_FORCE has no effect on shared mappings */
1429         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1430                 return false;
1431
1432         /* ... or read-only private ones */
1433         if (!(vma->vm_flags & VM_MAYWRITE))
1434                 return false;
1435
1436         /* ... or already writable ones that just need to take a write fault */
1437         if (vma->vm_flags & VM_WRITE)
1438                 return false;
1439
1440         /*
1441          * See can_change_pte_writable(): we broke COW and could map the page
1442          * writable if we have an exclusive anonymous page ...
1443          */
1444         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1445                 return false;
1446
1447         /* ... and a write-fault isn't required for other reasons. */
1448         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1449                 return false;
1450         return !userfaultfd_huge_pmd_wp(vma, pmd);
1451 }
1452
1453 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1454                                    unsigned long addr,
1455                                    pmd_t *pmd,
1456                                    unsigned int flags)
1457 {
1458         struct mm_struct *mm = vma->vm_mm;
1459         struct page *page;
1460         int ret;
1461
1462         assert_spin_locked(pmd_lockptr(mm, pmd));
1463
1464         page = pmd_page(*pmd);
1465         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1466
1467         if ((flags & FOLL_WRITE) &&
1468             !can_follow_write_pmd(*pmd, page, vma, flags))
1469                 return NULL;
1470
1471         /* Avoid dumping huge zero page */
1472         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1473                 return ERR_PTR(-EFAULT);
1474
1475         /* Full NUMA hinting faults to serialise migration in fault paths */
1476         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1477                 return NULL;
1478
1479         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1480                 return ERR_PTR(-EMLINK);
1481
1482         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1483                         !PageAnonExclusive(page), page);
1484
1485         ret = try_grab_page(page, flags);
1486         if (ret)
1487                 return ERR_PTR(ret);
1488
1489         if (flags & FOLL_TOUCH)
1490                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1491
1492         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1493         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1494
1495         return page;
1496 }
1497
1498 /* NUMA hinting page fault entry point for trans huge pmds */
1499 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1500 {
1501         struct vm_area_struct *vma = vmf->vma;
1502         pmd_t oldpmd = vmf->orig_pmd;
1503         pmd_t pmd;
1504         struct page *page;
1505         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1506         int page_nid = NUMA_NO_NODE;
1507         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1508         bool migrated = false, writable = false;
1509         int flags = 0;
1510
1511         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1512         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1513                 spin_unlock(vmf->ptl);
1514                 goto out;
1515         }
1516
1517         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1518
1519         /*
1520          * Detect now whether the PMD could be writable; this information
1521          * is only valid while holding the PT lock.
1522          */
1523         writable = pmd_write(pmd);
1524         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1525             can_change_pmd_writable(vma, vmf->address, pmd))
1526                 writable = true;
1527
1528         page = vm_normal_page_pmd(vma, haddr, pmd);
1529         if (!page)
1530                 goto out_map;
1531
1532         /* See similar comment in do_numa_page for explanation */
1533         if (!writable)
1534                 flags |= TNF_NO_GROUP;
1535
1536         page_nid = page_to_nid(page);
1537         /*
1538          * For memory tiering mode, cpupid of slow memory page is used
1539          * to record page access time.  So use default value.
1540          */
1541         if (node_is_toptier(page_nid))
1542                 last_cpupid = page_cpupid_last(page);
1543         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1544                                        &flags);
1545
1546         if (target_nid == NUMA_NO_NODE) {
1547                 put_page(page);
1548                 goto out_map;
1549         }
1550
1551         spin_unlock(vmf->ptl);
1552         writable = false;
1553
1554         migrated = migrate_misplaced_page(page, vma, target_nid);
1555         if (migrated) {
1556                 flags |= TNF_MIGRATED;
1557                 page_nid = target_nid;
1558         } else {
1559                 flags |= TNF_MIGRATE_FAIL;
1560                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1561                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1562                         spin_unlock(vmf->ptl);
1563                         goto out;
1564                 }
1565                 goto out_map;
1566         }
1567
1568 out:
1569         if (page_nid != NUMA_NO_NODE)
1570                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1571                                 flags);
1572
1573         return 0;
1574
1575 out_map:
1576         /* Restore the PMD */
1577         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1578         pmd = pmd_mkyoung(pmd);
1579         if (writable)
1580                 pmd = pmd_mkwrite(pmd);
1581         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1582         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1583         spin_unlock(vmf->ptl);
1584         goto out;
1585 }
1586
1587 /*
1588  * Return true if we do MADV_FREE successfully on entire pmd page.
1589  * Otherwise, return false.
1590  */
1591 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1592                 pmd_t *pmd, unsigned long addr, unsigned long next)
1593 {
1594         spinlock_t *ptl;
1595         pmd_t orig_pmd;
1596         struct folio *folio;
1597         struct mm_struct *mm = tlb->mm;
1598         bool ret = false;
1599
1600         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1601
1602         ptl = pmd_trans_huge_lock(pmd, vma);
1603         if (!ptl)
1604                 goto out_unlocked;
1605
1606         orig_pmd = *pmd;
1607         if (is_huge_zero_pmd(orig_pmd))
1608                 goto out;
1609
1610         if (unlikely(!pmd_present(orig_pmd))) {
1611                 VM_BUG_ON(thp_migration_supported() &&
1612                                   !is_pmd_migration_entry(orig_pmd));
1613                 goto out;
1614         }
1615
1616         folio = pfn_folio(pmd_pfn(orig_pmd));
1617         /*
1618          * If other processes are mapping this folio, we couldn't discard
1619          * the folio unless they all do MADV_FREE so let's skip the folio.
1620          */
1621         if (folio_mapcount(folio) != 1)
1622                 goto out;
1623
1624         if (!folio_trylock(folio))
1625                 goto out;
1626
1627         /*
1628          * If user want to discard part-pages of THP, split it so MADV_FREE
1629          * will deactivate only them.
1630          */
1631         if (next - addr != HPAGE_PMD_SIZE) {
1632                 folio_get(folio);
1633                 spin_unlock(ptl);
1634                 split_folio(folio);
1635                 folio_unlock(folio);
1636                 folio_put(folio);
1637                 goto out_unlocked;
1638         }
1639
1640         if (folio_test_dirty(folio))
1641                 folio_clear_dirty(folio);
1642         folio_unlock(folio);
1643
1644         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1645                 pmdp_invalidate(vma, addr, pmd);
1646                 orig_pmd = pmd_mkold(orig_pmd);
1647                 orig_pmd = pmd_mkclean(orig_pmd);
1648
1649                 set_pmd_at(mm, addr, pmd, orig_pmd);
1650                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1651         }
1652
1653         folio_mark_lazyfree(folio);
1654         ret = true;
1655 out:
1656         spin_unlock(ptl);
1657 out_unlocked:
1658         return ret;
1659 }
1660
1661 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1662 {
1663         pgtable_t pgtable;
1664
1665         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1666         pte_free(mm, pgtable);
1667         mm_dec_nr_ptes(mm);
1668 }
1669
1670 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1671                  pmd_t *pmd, unsigned long addr)
1672 {
1673         pmd_t orig_pmd;
1674         spinlock_t *ptl;
1675
1676         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1677
1678         ptl = __pmd_trans_huge_lock(pmd, vma);
1679         if (!ptl)
1680                 return 0;
1681         /*
1682          * For architectures like ppc64 we look at deposited pgtable
1683          * when calling pmdp_huge_get_and_clear. So do the
1684          * pgtable_trans_huge_withdraw after finishing pmdp related
1685          * operations.
1686          */
1687         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1688                                                 tlb->fullmm);
1689         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1690         if (vma_is_special_huge(vma)) {
1691                 if (arch_needs_pgtable_deposit())
1692                         zap_deposited_table(tlb->mm, pmd);
1693                 spin_unlock(ptl);
1694         } else if (is_huge_zero_pmd(orig_pmd)) {
1695                 zap_deposited_table(tlb->mm, pmd);
1696                 spin_unlock(ptl);
1697         } else {
1698                 struct page *page = NULL;
1699                 int flush_needed = 1;
1700
1701                 if (pmd_present(orig_pmd)) {
1702                         page = pmd_page(orig_pmd);
1703                         page_remove_rmap(page, vma, true);
1704                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1705                         VM_BUG_ON_PAGE(!PageHead(page), page);
1706                 } else if (thp_migration_supported()) {
1707                         swp_entry_t entry;
1708
1709                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1710                         entry = pmd_to_swp_entry(orig_pmd);
1711                         page = pfn_swap_entry_to_page(entry);
1712                         flush_needed = 0;
1713                 } else
1714                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1715
1716                 if (PageAnon(page)) {
1717                         zap_deposited_table(tlb->mm, pmd);
1718                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1719                 } else {
1720                         if (arch_needs_pgtable_deposit())
1721                                 zap_deposited_table(tlb->mm, pmd);
1722                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1723                 }
1724
1725                 spin_unlock(ptl);
1726                 if (flush_needed)
1727                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1728         }
1729         return 1;
1730 }
1731
1732 #ifndef pmd_move_must_withdraw
1733 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1734                                          spinlock_t *old_pmd_ptl,
1735                                          struct vm_area_struct *vma)
1736 {
1737         /*
1738          * With split pmd lock we also need to move preallocated
1739          * PTE page table if new_pmd is on different PMD page table.
1740          *
1741          * We also don't deposit and withdraw tables for file pages.
1742          */
1743         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1744 }
1745 #endif
1746
1747 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1748 {
1749 #ifdef CONFIG_MEM_SOFT_DIRTY
1750         if (unlikely(is_pmd_migration_entry(pmd)))
1751                 pmd = pmd_swp_mksoft_dirty(pmd);
1752         else if (pmd_present(pmd))
1753                 pmd = pmd_mksoft_dirty(pmd);
1754 #endif
1755         return pmd;
1756 }
1757
1758 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1759                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1760 {
1761         spinlock_t *old_ptl, *new_ptl;
1762         pmd_t pmd;
1763         struct mm_struct *mm = vma->vm_mm;
1764         bool force_flush = false;
1765
1766         /*
1767          * The destination pmd shouldn't be established, free_pgtables()
1768          * should have release it.
1769          */
1770         if (WARN_ON(!pmd_none(*new_pmd))) {
1771                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1772                 return false;
1773         }
1774
1775         /*
1776          * We don't have to worry about the ordering of src and dst
1777          * ptlocks because exclusive mmap_lock prevents deadlock.
1778          */
1779         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1780         if (old_ptl) {
1781                 new_ptl = pmd_lockptr(mm, new_pmd);
1782                 if (new_ptl != old_ptl)
1783                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1784                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1785                 if (pmd_present(pmd))
1786                         force_flush = true;
1787                 VM_BUG_ON(!pmd_none(*new_pmd));
1788
1789                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1790                         pgtable_t pgtable;
1791                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1792                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1793                 }
1794                 pmd = move_soft_dirty_pmd(pmd);
1795                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1796                 if (force_flush)
1797                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1798                 if (new_ptl != old_ptl)
1799                         spin_unlock(new_ptl);
1800                 spin_unlock(old_ptl);
1801                 return true;
1802         }
1803         return false;
1804 }
1805
1806 /*
1807  * Returns
1808  *  - 0 if PMD could not be locked
1809  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1810  *      or if prot_numa but THP migration is not supported
1811  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1812  */
1813 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1814                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1815                     unsigned long cp_flags)
1816 {
1817         struct mm_struct *mm = vma->vm_mm;
1818         spinlock_t *ptl;
1819         pmd_t oldpmd, entry;
1820         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1821         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1822         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1823         int ret = 1;
1824
1825         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1826
1827         if (prot_numa && !thp_migration_supported())
1828                 return 1;
1829
1830         ptl = __pmd_trans_huge_lock(pmd, vma);
1831         if (!ptl)
1832                 return 0;
1833
1834 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1835         if (is_swap_pmd(*pmd)) {
1836                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1837                 struct page *page = pfn_swap_entry_to_page(entry);
1838
1839                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1840                 if (is_writable_migration_entry(entry)) {
1841                         pmd_t newpmd;
1842                         /*
1843                          * A protection check is difficult so
1844                          * just be safe and disable write
1845                          */
1846                         if (PageAnon(page))
1847                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1848                         else
1849                                 entry = make_readable_migration_entry(swp_offset(entry));
1850                         newpmd = swp_entry_to_pmd(entry);
1851                         if (pmd_swp_soft_dirty(*pmd))
1852                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1853                         if (pmd_swp_uffd_wp(*pmd))
1854                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1855                         set_pmd_at(mm, addr, pmd, newpmd);
1856                 }
1857                 goto unlock;
1858         }
1859 #endif
1860
1861         if (prot_numa) {
1862                 struct page *page;
1863                 bool toptier;
1864                 /*
1865                  * Avoid trapping faults against the zero page. The read-only
1866                  * data is likely to be read-cached on the local CPU and
1867                  * local/remote hits to the zero page are not interesting.
1868                  */
1869                 if (is_huge_zero_pmd(*pmd))
1870                         goto unlock;
1871
1872                 if (pmd_protnone(*pmd))
1873                         goto unlock;
1874
1875                 page = pmd_page(*pmd);
1876                 toptier = node_is_toptier(page_to_nid(page));
1877                 /*
1878                  * Skip scanning top tier node if normal numa
1879                  * balancing is disabled
1880                  */
1881                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1882                     toptier)
1883                         goto unlock;
1884
1885                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1886                     !toptier)
1887                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1888         }
1889         /*
1890          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1891          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1892          * which is also under mmap_read_lock(mm):
1893          *
1894          *      CPU0:                           CPU1:
1895          *                              change_huge_pmd(prot_numa=1)
1896          *                               pmdp_huge_get_and_clear_notify()
1897          * madvise_dontneed()
1898          *  zap_pmd_range()
1899          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1900          *   // skip the pmd
1901          *                               set_pmd_at();
1902          *                               // pmd is re-established
1903          *
1904          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1905          * which may break userspace.
1906          *
1907          * pmdp_invalidate_ad() is required to make sure we don't miss
1908          * dirty/young flags set by hardware.
1909          */
1910         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1911
1912         entry = pmd_modify(oldpmd, newprot);
1913         if (uffd_wp)
1914                 entry = pmd_mkuffd_wp(entry);
1915         else if (uffd_wp_resolve)
1916                 /*
1917                  * Leave the write bit to be handled by PF interrupt
1918                  * handler, then things like COW could be properly
1919                  * handled.
1920                  */
1921                 entry = pmd_clear_uffd_wp(entry);
1922
1923         /* See change_pte_range(). */
1924         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1925             can_change_pmd_writable(vma, addr, entry))
1926                 entry = pmd_mkwrite(entry);
1927
1928         ret = HPAGE_PMD_NR;
1929         set_pmd_at(mm, addr, pmd, entry);
1930
1931         if (huge_pmd_needs_flush(oldpmd, entry))
1932                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1933 unlock:
1934         spin_unlock(ptl);
1935         return ret;
1936 }
1937
1938 /*
1939  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1940  *
1941  * Note that if it returns page table lock pointer, this routine returns without
1942  * unlocking page table lock. So callers must unlock it.
1943  */
1944 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1945 {
1946         spinlock_t *ptl;
1947         ptl = pmd_lock(vma->vm_mm, pmd);
1948         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1949                         pmd_devmap(*pmd)))
1950                 return ptl;
1951         spin_unlock(ptl);
1952         return NULL;
1953 }
1954
1955 /*
1956  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1957  *
1958  * Note that if it returns page table lock pointer, this routine returns without
1959  * unlocking page table lock. So callers must unlock it.
1960  */
1961 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1962 {
1963         spinlock_t *ptl;
1964
1965         ptl = pud_lock(vma->vm_mm, pud);
1966         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1967                 return ptl;
1968         spin_unlock(ptl);
1969         return NULL;
1970 }
1971
1972 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1973 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1974                  pud_t *pud, unsigned long addr)
1975 {
1976         spinlock_t *ptl;
1977
1978         ptl = __pud_trans_huge_lock(pud, vma);
1979         if (!ptl)
1980                 return 0;
1981
1982         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1983         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1984         if (vma_is_special_huge(vma)) {
1985                 spin_unlock(ptl);
1986                 /* No zero page support yet */
1987         } else {
1988                 /* No support for anonymous PUD pages yet */
1989                 BUG();
1990         }
1991         return 1;
1992 }
1993
1994 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1995                 unsigned long haddr)
1996 {
1997         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1998         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1999         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2000         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2001
2002         count_vm_event(THP_SPLIT_PUD);
2003
2004         pudp_huge_clear_flush_notify(vma, haddr, pud);
2005 }
2006
2007 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2008                 unsigned long address)
2009 {
2010         spinlock_t *ptl;
2011         struct mmu_notifier_range range;
2012
2013         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2014                                 address & HPAGE_PUD_MASK,
2015                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2016         mmu_notifier_invalidate_range_start(&range);
2017         ptl = pud_lock(vma->vm_mm, pud);
2018         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2019                 goto out;
2020         __split_huge_pud_locked(vma, pud, range.start);
2021
2022 out:
2023         spin_unlock(ptl);
2024         /*
2025          * No need to double call mmu_notifier->invalidate_range() callback as
2026          * the above pudp_huge_clear_flush_notify() did already call it.
2027          */
2028         mmu_notifier_invalidate_range_only_end(&range);
2029 }
2030 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2031
2032 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2033                 unsigned long haddr, pmd_t *pmd)
2034 {
2035         struct mm_struct *mm = vma->vm_mm;
2036         pgtable_t pgtable;
2037         pmd_t _pmd, old_pmd;
2038         int i;
2039
2040         /*
2041          * Leave pmd empty until pte is filled note that it is fine to delay
2042          * notification until mmu_notifier_invalidate_range_end() as we are
2043          * replacing a zero pmd write protected page with a zero pte write
2044          * protected page.
2045          *
2046          * See Documentation/mm/mmu_notifier.rst
2047          */
2048         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2049
2050         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2051         pmd_populate(mm, &_pmd, pgtable);
2052
2053         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2054                 pte_t *pte, entry;
2055                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2056                 entry = pte_mkspecial(entry);
2057                 if (pmd_uffd_wp(old_pmd))
2058                         entry = pte_mkuffd_wp(entry);
2059                 pte = pte_offset_map(&_pmd, haddr);
2060                 VM_BUG_ON(!pte_none(*pte));
2061                 set_pte_at(mm, haddr, pte, entry);
2062                 pte_unmap(pte);
2063         }
2064         smp_wmb(); /* make pte visible before pmd */
2065         pmd_populate(mm, pmd, pgtable);
2066 }
2067
2068 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2069                 unsigned long haddr, bool freeze)
2070 {
2071         struct mm_struct *mm = vma->vm_mm;
2072         struct page *page;
2073         pgtable_t pgtable;
2074         pmd_t old_pmd, _pmd;
2075         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2076         bool anon_exclusive = false, dirty = false;
2077         unsigned long addr;
2078         int i;
2079
2080         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2081         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2082         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2083         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2084                                 && !pmd_devmap(*pmd));
2085
2086         count_vm_event(THP_SPLIT_PMD);
2087
2088         if (!vma_is_anonymous(vma)) {
2089                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2090                 /*
2091                  * We are going to unmap this huge page. So
2092                  * just go ahead and zap it
2093                  */
2094                 if (arch_needs_pgtable_deposit())
2095                         zap_deposited_table(mm, pmd);
2096                 if (vma_is_special_huge(vma))
2097                         return;
2098                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2099                         swp_entry_t entry;
2100
2101                         entry = pmd_to_swp_entry(old_pmd);
2102                         page = pfn_swap_entry_to_page(entry);
2103                 } else {
2104                         page = pmd_page(old_pmd);
2105                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2106                                 set_page_dirty(page);
2107                         if (!PageReferenced(page) && pmd_young(old_pmd))
2108                                 SetPageReferenced(page);
2109                         page_remove_rmap(page, vma, true);
2110                         put_page(page);
2111                 }
2112                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2113                 return;
2114         }
2115
2116         if (is_huge_zero_pmd(*pmd)) {
2117                 /*
2118                  * FIXME: Do we want to invalidate secondary mmu by calling
2119                  * mmu_notifier_invalidate_range() see comments below inside
2120                  * __split_huge_pmd() ?
2121                  *
2122                  * We are going from a zero huge page write protected to zero
2123                  * small page also write protected so it does not seems useful
2124                  * to invalidate secondary mmu at this time.
2125                  */
2126                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2127         }
2128
2129         /*
2130          * Up to this point the pmd is present and huge and userland has the
2131          * whole access to the hugepage during the split (which happens in
2132          * place). If we overwrite the pmd with the not-huge version pointing
2133          * to the pte here (which of course we could if all CPUs were bug
2134          * free), userland could trigger a small page size TLB miss on the
2135          * small sized TLB while the hugepage TLB entry is still established in
2136          * the huge TLB. Some CPU doesn't like that.
2137          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2138          * 383 on page 105. Intel should be safe but is also warns that it's
2139          * only safe if the permission and cache attributes of the two entries
2140          * loaded in the two TLB is identical (which should be the case here).
2141          * But it is generally safer to never allow small and huge TLB entries
2142          * for the same virtual address to be loaded simultaneously. So instead
2143          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2144          * current pmd notpresent (atomically because here the pmd_trans_huge
2145          * must remain set at all times on the pmd until the split is complete
2146          * for this pmd), then we flush the SMP TLB and finally we write the
2147          * non-huge version of the pmd entry with pmd_populate.
2148          */
2149         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2150
2151         pmd_migration = is_pmd_migration_entry(old_pmd);
2152         if (unlikely(pmd_migration)) {
2153                 swp_entry_t entry;
2154
2155                 entry = pmd_to_swp_entry(old_pmd);
2156                 page = pfn_swap_entry_to_page(entry);
2157                 write = is_writable_migration_entry(entry);
2158                 if (PageAnon(page))
2159                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2160                 young = is_migration_entry_young(entry);
2161                 dirty = is_migration_entry_dirty(entry);
2162                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2163                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2164         } else {
2165                 page = pmd_page(old_pmd);
2166                 if (pmd_dirty(old_pmd)) {
2167                         dirty = true;
2168                         SetPageDirty(page);
2169                 }
2170                 write = pmd_write(old_pmd);
2171                 young = pmd_young(old_pmd);
2172                 soft_dirty = pmd_soft_dirty(old_pmd);
2173                 uffd_wp = pmd_uffd_wp(old_pmd);
2174
2175                 VM_BUG_ON_PAGE(!page_count(page), page);
2176
2177                 /*
2178                  * Without "freeze", we'll simply split the PMD, propagating the
2179                  * PageAnonExclusive() flag for each PTE by setting it for
2180                  * each subpage -- no need to (temporarily) clear.
2181                  *
2182                  * With "freeze" we want to replace mapped pages by
2183                  * migration entries right away. This is only possible if we
2184                  * managed to clear PageAnonExclusive() -- see
2185                  * set_pmd_migration_entry().
2186                  *
2187                  * In case we cannot clear PageAnonExclusive(), split the PMD
2188                  * only and let try_to_migrate_one() fail later.
2189                  *
2190                  * See page_try_share_anon_rmap(): invalidate PMD first.
2191                  */
2192                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2193                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2194                         freeze = false;
2195                 if (!freeze)
2196                         page_ref_add(page, HPAGE_PMD_NR - 1);
2197         }
2198
2199         /*
2200          * Withdraw the table only after we mark the pmd entry invalid.
2201          * This's critical for some architectures (Power).
2202          */
2203         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2204         pmd_populate(mm, &_pmd, pgtable);
2205
2206         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2207                 pte_t entry, *pte;
2208                 /*
2209                  * Note that NUMA hinting access restrictions are not
2210                  * transferred to avoid any possibility of altering
2211                  * permissions across VMAs.
2212                  */
2213                 if (freeze || pmd_migration) {
2214                         swp_entry_t swp_entry;
2215                         if (write)
2216                                 swp_entry = make_writable_migration_entry(
2217                                                         page_to_pfn(page + i));
2218                         else if (anon_exclusive)
2219                                 swp_entry = make_readable_exclusive_migration_entry(
2220                                                         page_to_pfn(page + i));
2221                         else
2222                                 swp_entry = make_readable_migration_entry(
2223                                                         page_to_pfn(page + i));
2224                         if (young)
2225                                 swp_entry = make_migration_entry_young(swp_entry);
2226                         if (dirty)
2227                                 swp_entry = make_migration_entry_dirty(swp_entry);
2228                         entry = swp_entry_to_pte(swp_entry);
2229                         if (soft_dirty)
2230                                 entry = pte_swp_mksoft_dirty(entry);
2231                         if (uffd_wp)
2232                                 entry = pte_swp_mkuffd_wp(entry);
2233                 } else {
2234                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2235                         entry = maybe_mkwrite(entry, vma);
2236                         if (anon_exclusive)
2237                                 SetPageAnonExclusive(page + i);
2238                         if (!young)
2239                                 entry = pte_mkold(entry);
2240                         /* NOTE: this may set soft-dirty too on some archs */
2241                         if (dirty)
2242                                 entry = pte_mkdirty(entry);
2243                         /*
2244                          * NOTE: this needs to happen after pte_mkdirty,
2245                          * because some archs (sparc64, loongarch) could
2246                          * set hw write bit when mkdirty.
2247                          */
2248                         if (!write)
2249                                 entry = pte_wrprotect(entry);
2250                         if (soft_dirty)
2251                                 entry = pte_mksoft_dirty(entry);
2252                         if (uffd_wp)
2253                                 entry = pte_mkuffd_wp(entry);
2254                         page_add_anon_rmap(page + i, vma, addr, false);
2255                 }
2256                 pte = pte_offset_map(&_pmd, addr);
2257                 BUG_ON(!pte_none(*pte));
2258                 set_pte_at(mm, addr, pte, entry);
2259                 pte_unmap(pte);
2260         }
2261
2262         if (!pmd_migration)
2263                 page_remove_rmap(page, vma, true);
2264         if (freeze)
2265                 put_page(page);
2266
2267         smp_wmb(); /* make pte visible before pmd */
2268         pmd_populate(mm, pmd, pgtable);
2269 }
2270
2271 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2272                 unsigned long address, bool freeze, struct folio *folio)
2273 {
2274         spinlock_t *ptl;
2275         struct mmu_notifier_range range;
2276
2277         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2278                                 address & HPAGE_PMD_MASK,
2279                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2280         mmu_notifier_invalidate_range_start(&range);
2281         ptl = pmd_lock(vma->vm_mm, pmd);
2282
2283         /*
2284          * If caller asks to setup a migration entry, we need a folio to check
2285          * pmd against. Otherwise we can end up replacing wrong folio.
2286          */
2287         VM_BUG_ON(freeze && !folio);
2288         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2289
2290         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2291             is_pmd_migration_entry(*pmd)) {
2292                 /*
2293                  * It's safe to call pmd_page when folio is set because it's
2294                  * guaranteed that pmd is present.
2295                  */
2296                 if (folio && folio != page_folio(pmd_page(*pmd)))
2297                         goto out;
2298                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2299         }
2300
2301 out:
2302         spin_unlock(ptl);
2303         /*
2304          * No need to double call mmu_notifier->invalidate_range() callback.
2305          * They are 3 cases to consider inside __split_huge_pmd_locked():
2306          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2307          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2308          *    fault will trigger a flush_notify before pointing to a new page
2309          *    (it is fine if the secondary mmu keeps pointing to the old zero
2310          *    page in the meantime)
2311          *  3) Split a huge pmd into pte pointing to the same page. No need
2312          *     to invalidate secondary tlb entry they are all still valid.
2313          *     any further changes to individual pte will notify. So no need
2314          *     to call mmu_notifier->invalidate_range()
2315          */
2316         mmu_notifier_invalidate_range_only_end(&range);
2317 }
2318
2319 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2320                 bool freeze, struct folio *folio)
2321 {
2322         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2323
2324         if (!pmd)
2325                 return;
2326
2327         __split_huge_pmd(vma, pmd, address, freeze, folio);
2328 }
2329
2330 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2331 {
2332         /*
2333          * If the new address isn't hpage aligned and it could previously
2334          * contain an hugepage: check if we need to split an huge pmd.
2335          */
2336         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2337             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2338                          ALIGN(address, HPAGE_PMD_SIZE)))
2339                 split_huge_pmd_address(vma, address, false, NULL);
2340 }
2341
2342 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2343                              unsigned long start,
2344                              unsigned long end,
2345                              long adjust_next)
2346 {
2347         /* Check if we need to split start first. */
2348         split_huge_pmd_if_needed(vma, start);
2349
2350         /* Check if we need to split end next. */
2351         split_huge_pmd_if_needed(vma, end);
2352
2353         /*
2354          * If we're also updating the next vma vm_start,
2355          * check if we need to split it.
2356          */
2357         if (adjust_next > 0) {
2358                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2359                 unsigned long nstart = next->vm_start;
2360                 nstart += adjust_next;
2361                 split_huge_pmd_if_needed(next, nstart);
2362         }
2363 }
2364
2365 static void unmap_folio(struct folio *folio)
2366 {
2367         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2368                 TTU_SYNC;
2369
2370         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2371
2372         /*
2373          * Anon pages need migration entries to preserve them, but file
2374          * pages can simply be left unmapped, then faulted back on demand.
2375          * If that is ever changed (perhaps for mlock), update remap_page().
2376          */
2377         if (folio_test_anon(folio))
2378                 try_to_migrate(folio, ttu_flags);
2379         else
2380                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2381 }
2382
2383 static void remap_page(struct folio *folio, unsigned long nr)
2384 {
2385         int i = 0;
2386
2387         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2388         if (!folio_test_anon(folio))
2389                 return;
2390         for (;;) {
2391                 remove_migration_ptes(folio, folio, true);
2392                 i += folio_nr_pages(folio);
2393                 if (i >= nr)
2394                         break;
2395                 folio = folio_next(folio);
2396         }
2397 }
2398
2399 static void lru_add_page_tail(struct page *head, struct page *tail,
2400                 struct lruvec *lruvec, struct list_head *list)
2401 {
2402         VM_BUG_ON_PAGE(!PageHead(head), head);
2403         VM_BUG_ON_PAGE(PageCompound(tail), head);
2404         VM_BUG_ON_PAGE(PageLRU(tail), head);
2405         lockdep_assert_held(&lruvec->lru_lock);
2406
2407         if (list) {
2408                 /* page reclaim is reclaiming a huge page */
2409                 VM_WARN_ON(PageLRU(head));
2410                 get_page(tail);
2411                 list_add_tail(&tail->lru, list);
2412         } else {
2413                 /* head is still on lru (and we have it frozen) */
2414                 VM_WARN_ON(!PageLRU(head));
2415                 if (PageUnevictable(tail))
2416                         tail->mlock_count = 0;
2417                 else
2418                         list_add_tail(&tail->lru, &head->lru);
2419                 SetPageLRU(tail);
2420         }
2421 }
2422
2423 static void __split_huge_page_tail(struct page *head, int tail,
2424                 struct lruvec *lruvec, struct list_head *list)
2425 {
2426         struct page *page_tail = head + tail;
2427
2428         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2429
2430         /*
2431          * Clone page flags before unfreezing refcount.
2432          *
2433          * After successful get_page_unless_zero() might follow flags change,
2434          * for example lock_page() which set PG_waiters.
2435          *
2436          * Note that for mapped sub-pages of an anonymous THP,
2437          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2438          * the migration entry instead from where remap_page() will restore it.
2439          * We can still have PG_anon_exclusive set on effectively unmapped and
2440          * unreferenced sub-pages of an anonymous THP: we can simply drop
2441          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2442          */
2443         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2444         page_tail->flags |= (head->flags &
2445                         ((1L << PG_referenced) |
2446                          (1L << PG_swapbacked) |
2447                          (1L << PG_swapcache) |
2448                          (1L << PG_mlocked) |
2449                          (1L << PG_uptodate) |
2450                          (1L << PG_active) |
2451                          (1L << PG_workingset) |
2452                          (1L << PG_locked) |
2453                          (1L << PG_unevictable) |
2454 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2455                          (1L << PG_arch_2) |
2456                          (1L << PG_arch_3) |
2457 #endif
2458                          (1L << PG_dirty) |
2459                          LRU_GEN_MASK | LRU_REFS_MASK));
2460
2461         /* ->mapping in first and second tail page is replaced by other uses */
2462         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2463                         page_tail);
2464         page_tail->mapping = head->mapping;
2465         page_tail->index = head->index + tail;
2466
2467         /*
2468          * page->private should not be set in tail pages with the exception
2469          * of swap cache pages that store the swp_entry_t in tail pages.
2470          * Fix up and warn once if private is unexpectedly set.
2471          *
2472          * What of 32-bit systems, on which folio->_pincount overlays
2473          * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2474          * pincount must be 0 for folio_ref_freeze() to have succeeded.
2475          */
2476         if (!folio_test_swapcache(page_folio(head))) {
2477                 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2478                 page_tail->private = 0;
2479         }
2480
2481         /* Page flags must be visible before we make the page non-compound. */
2482         smp_wmb();
2483
2484         /*
2485          * Clear PageTail before unfreezing page refcount.
2486          *
2487          * After successful get_page_unless_zero() might follow put_page()
2488          * which needs correct compound_head().
2489          */
2490         clear_compound_head(page_tail);
2491
2492         /* Finally unfreeze refcount. Additional reference from page cache. */
2493         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2494                                           PageSwapCache(head)));
2495
2496         if (page_is_young(head))
2497                 set_page_young(page_tail);
2498         if (page_is_idle(head))
2499                 set_page_idle(page_tail);
2500
2501         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2502
2503         /*
2504          * always add to the tail because some iterators expect new
2505          * pages to show after the currently processed elements - e.g.
2506          * migrate_pages
2507          */
2508         lru_add_page_tail(head, page_tail, lruvec, list);
2509 }
2510
2511 static void __split_huge_page(struct page *page, struct list_head *list,
2512                 pgoff_t end)
2513 {
2514         struct folio *folio = page_folio(page);
2515         struct page *head = &folio->page;
2516         struct lruvec *lruvec;
2517         struct address_space *swap_cache = NULL;
2518         unsigned long offset = 0;
2519         unsigned int nr = thp_nr_pages(head);
2520         int i;
2521
2522         /* complete memcg works before add pages to LRU */
2523         split_page_memcg(head, nr);
2524
2525         if (PageAnon(head) && PageSwapCache(head)) {
2526                 swp_entry_t entry = { .val = page_private(head) };
2527
2528                 offset = swp_offset(entry);
2529                 swap_cache = swap_address_space(entry);
2530                 xa_lock(&swap_cache->i_pages);
2531         }
2532
2533         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2534         lruvec = folio_lruvec_lock(folio);
2535
2536         ClearPageHasHWPoisoned(head);
2537
2538         for (i = nr - 1; i >= 1; i--) {
2539                 __split_huge_page_tail(head, i, lruvec, list);
2540                 /* Some pages can be beyond EOF: drop them from page cache */
2541                 if (head[i].index >= end) {
2542                         struct folio *tail = page_folio(head + i);
2543
2544                         if (shmem_mapping(head->mapping))
2545                                 shmem_uncharge(head->mapping->host, 1);
2546                         else if (folio_test_clear_dirty(tail))
2547                                 folio_account_cleaned(tail,
2548                                         inode_to_wb(folio->mapping->host));
2549                         __filemap_remove_folio(tail, NULL);
2550                         folio_put(tail);
2551                 } else if (!PageAnon(page)) {
2552                         __xa_store(&head->mapping->i_pages, head[i].index,
2553                                         head + i, 0);
2554                 } else if (swap_cache) {
2555                         __xa_store(&swap_cache->i_pages, offset + i,
2556                                         head + i, 0);
2557                 }
2558         }
2559
2560         ClearPageCompound(head);
2561         unlock_page_lruvec(lruvec);
2562         /* Caller disabled irqs, so they are still disabled here */
2563
2564         split_page_owner(head, nr);
2565
2566         /* See comment in __split_huge_page_tail() */
2567         if (PageAnon(head)) {
2568                 /* Additional pin to swap cache */
2569                 if (PageSwapCache(head)) {
2570                         page_ref_add(head, 2);
2571                         xa_unlock(&swap_cache->i_pages);
2572                 } else {
2573                         page_ref_inc(head);
2574                 }
2575         } else {
2576                 /* Additional pin to page cache */
2577                 page_ref_add(head, 2);
2578                 xa_unlock(&head->mapping->i_pages);
2579         }
2580         local_irq_enable();
2581
2582         remap_page(folio, nr);
2583
2584         if (PageSwapCache(head)) {
2585                 swp_entry_t entry = { .val = page_private(head) };
2586
2587                 split_swap_cluster(entry);
2588         }
2589
2590         for (i = 0; i < nr; i++) {
2591                 struct page *subpage = head + i;
2592                 if (subpage == page)
2593                         continue;
2594                 unlock_page(subpage);
2595
2596                 /*
2597                  * Subpages may be freed if there wasn't any mapping
2598                  * like if add_to_swap() is running on a lru page that
2599                  * had its mapping zapped. And freeing these pages
2600                  * requires taking the lru_lock so we do the put_page
2601                  * of the tail pages after the split is complete.
2602                  */
2603                 free_page_and_swap_cache(subpage);
2604         }
2605 }
2606
2607 /* Racy check whether the huge page can be split */
2608 bool can_split_folio(struct folio *folio, int *pextra_pins)
2609 {
2610         int extra_pins;
2611
2612         /* Additional pins from page cache */
2613         if (folio_test_anon(folio))
2614                 extra_pins = folio_test_swapcache(folio) ?
2615                                 folio_nr_pages(folio) : 0;
2616         else
2617                 extra_pins = folio_nr_pages(folio);
2618         if (pextra_pins)
2619                 *pextra_pins = extra_pins;
2620         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2621 }
2622
2623 /*
2624  * This function splits huge page into normal pages. @page can point to any
2625  * subpage of huge page to split. Split doesn't change the position of @page.
2626  *
2627  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2628  * The huge page must be locked.
2629  *
2630  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2631  *
2632  * Both head page and tail pages will inherit mapping, flags, and so on from
2633  * the hugepage.
2634  *
2635  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2636  * they are not mapped.
2637  *
2638  * Returns 0 if the hugepage is split successfully.
2639  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2640  * us.
2641  */
2642 int split_huge_page_to_list(struct page *page, struct list_head *list)
2643 {
2644         struct folio *folio = page_folio(page);
2645         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2646         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2647         struct anon_vma *anon_vma = NULL;
2648         struct address_space *mapping = NULL;
2649         int extra_pins, ret;
2650         pgoff_t end;
2651         bool is_hzp;
2652
2653         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2654         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2655
2656         is_hzp = is_huge_zero_page(&folio->page);
2657         VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2658         if (is_hzp)
2659                 return -EBUSY;
2660
2661         if (folio_test_writeback(folio))
2662                 return -EBUSY;
2663
2664         if (folio_test_anon(folio)) {
2665                 /*
2666                  * The caller does not necessarily hold an mmap_lock that would
2667                  * prevent the anon_vma disappearing so we first we take a
2668                  * reference to it and then lock the anon_vma for write. This
2669                  * is similar to folio_lock_anon_vma_read except the write lock
2670                  * is taken to serialise against parallel split or collapse
2671                  * operations.
2672                  */
2673                 anon_vma = folio_get_anon_vma(folio);
2674                 if (!anon_vma) {
2675                         ret = -EBUSY;
2676                         goto out;
2677                 }
2678                 end = -1;
2679                 mapping = NULL;
2680                 anon_vma_lock_write(anon_vma);
2681         } else {
2682                 gfp_t gfp;
2683
2684                 mapping = folio->mapping;
2685
2686                 /* Truncated ? */
2687                 if (!mapping) {
2688                         ret = -EBUSY;
2689                         goto out;
2690                 }
2691
2692                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2693                                                         GFP_RECLAIM_MASK);
2694
2695                 if (folio_test_private(folio) &&
2696                                 !filemap_release_folio(folio, gfp)) {
2697                         ret = -EBUSY;
2698                         goto out;
2699                 }
2700
2701                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2702                 if (xas_error(&xas)) {
2703                         ret = xas_error(&xas);
2704                         goto out;
2705                 }
2706
2707                 anon_vma = NULL;
2708                 i_mmap_lock_read(mapping);
2709
2710                 /*
2711                  *__split_huge_page() may need to trim off pages beyond EOF:
2712                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2713                  * which cannot be nested inside the page tree lock. So note
2714                  * end now: i_size itself may be changed at any moment, but
2715                  * folio lock is good enough to serialize the trimming.
2716                  */
2717                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2718                 if (shmem_mapping(mapping))
2719                         end = shmem_fallocend(mapping->host, end);
2720         }
2721
2722         /*
2723          * Racy check if we can split the page, before unmap_folio() will
2724          * split PMDs
2725          */
2726         if (!can_split_folio(folio, &extra_pins)) {
2727                 ret = -EAGAIN;
2728                 goto out_unlock;
2729         }
2730
2731         unmap_folio(folio);
2732
2733         /* block interrupt reentry in xa_lock and spinlock */
2734         local_irq_disable();
2735         if (mapping) {
2736                 /*
2737                  * Check if the folio is present in page cache.
2738                  * We assume all tail are present too, if folio is there.
2739                  */
2740                 xas_lock(&xas);
2741                 xas_reset(&xas);
2742                 if (xas_load(&xas) != folio)
2743                         goto fail;
2744         }
2745
2746         /* Prevent deferred_split_scan() touching ->_refcount */
2747         spin_lock(&ds_queue->split_queue_lock);
2748         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2749                 if (!list_empty(&folio->_deferred_list)) {
2750                         ds_queue->split_queue_len--;
2751                         list_del(&folio->_deferred_list);
2752                 }
2753                 spin_unlock(&ds_queue->split_queue_lock);
2754                 if (mapping) {
2755                         int nr = folio_nr_pages(folio);
2756
2757                         xas_split(&xas, folio, folio_order(folio));
2758                         if (folio_test_swapbacked(folio)) {
2759                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2760                                                         -nr);
2761                         } else {
2762                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2763                                                         -nr);
2764                                 filemap_nr_thps_dec(mapping);
2765                         }
2766                 }
2767
2768                 __split_huge_page(page, list, end);
2769                 ret = 0;
2770         } else {
2771                 spin_unlock(&ds_queue->split_queue_lock);
2772 fail:
2773                 if (mapping)
2774                         xas_unlock(&xas);
2775                 local_irq_enable();
2776                 remap_page(folio, folio_nr_pages(folio));
2777                 ret = -EAGAIN;
2778         }
2779
2780 out_unlock:
2781         if (anon_vma) {
2782                 anon_vma_unlock_write(anon_vma);
2783                 put_anon_vma(anon_vma);
2784         }
2785         if (mapping)
2786                 i_mmap_unlock_read(mapping);
2787 out:
2788         xas_destroy(&xas);
2789         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2790         return ret;
2791 }
2792
2793 void free_transhuge_page(struct page *page)
2794 {
2795         struct folio *folio = (struct folio *)page;
2796         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2797         unsigned long flags;
2798
2799         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2800         if (!list_empty(&folio->_deferred_list)) {
2801                 ds_queue->split_queue_len--;
2802                 list_del(&folio->_deferred_list);
2803         }
2804         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2805         free_compound_page(page);
2806 }
2807
2808 void deferred_split_folio(struct folio *folio)
2809 {
2810         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2811 #ifdef CONFIG_MEMCG
2812         struct mem_cgroup *memcg = folio_memcg(folio);
2813 #endif
2814         unsigned long flags;
2815
2816         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2817
2818         /*
2819          * The try_to_unmap() in page reclaim path might reach here too,
2820          * this may cause a race condition to corrupt deferred split queue.
2821          * And, if page reclaim is already handling the same folio, it is
2822          * unnecessary to handle it again in shrinker.
2823          *
2824          * Check the swapcache flag to determine if the folio is being
2825          * handled by page reclaim since THP swap would add the folio into
2826          * swap cache before calling try_to_unmap().
2827          */
2828         if (folio_test_swapcache(folio))
2829                 return;
2830
2831         if (!list_empty(&folio->_deferred_list))
2832                 return;
2833
2834         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2835         if (list_empty(&folio->_deferred_list)) {
2836                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2837                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2838                 ds_queue->split_queue_len++;
2839 #ifdef CONFIG_MEMCG
2840                 if (memcg)
2841                         set_shrinker_bit(memcg, folio_nid(folio),
2842                                          deferred_split_shrinker.id);
2843 #endif
2844         }
2845         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2846 }
2847
2848 static unsigned long deferred_split_count(struct shrinker *shrink,
2849                 struct shrink_control *sc)
2850 {
2851         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2852         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2853
2854 #ifdef CONFIG_MEMCG
2855         if (sc->memcg)
2856                 ds_queue = &sc->memcg->deferred_split_queue;
2857 #endif
2858         return READ_ONCE(ds_queue->split_queue_len);
2859 }
2860
2861 static unsigned long deferred_split_scan(struct shrinker *shrink,
2862                 struct shrink_control *sc)
2863 {
2864         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2865         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2866         unsigned long flags;
2867         LIST_HEAD(list);
2868         struct folio *folio, *next;
2869         int split = 0;
2870
2871 #ifdef CONFIG_MEMCG
2872         if (sc->memcg)
2873                 ds_queue = &sc->memcg->deferred_split_queue;
2874 #endif
2875
2876         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2877         /* Take pin on all head pages to avoid freeing them under us */
2878         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2879                                                         _deferred_list) {
2880                 if (folio_try_get(folio)) {
2881                         list_move(&folio->_deferred_list, &list);
2882                 } else {
2883                         /* We lost race with folio_put() */
2884                         list_del_init(&folio->_deferred_list);
2885                         ds_queue->split_queue_len--;
2886                 }
2887                 if (!--sc->nr_to_scan)
2888                         break;
2889         }
2890         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2891
2892         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2893                 if (!folio_trylock(folio))
2894                         goto next;
2895                 /* split_huge_page() removes page from list on success */
2896                 if (!split_folio(folio))
2897                         split++;
2898                 folio_unlock(folio);
2899 next:
2900                 folio_put(folio);
2901         }
2902
2903         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2904         list_splice_tail(&list, &ds_queue->split_queue);
2905         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2906
2907         /*
2908          * Stop shrinker if we didn't split any page, but the queue is empty.
2909          * This can happen if pages were freed under us.
2910          */
2911         if (!split && list_empty(&ds_queue->split_queue))
2912                 return SHRINK_STOP;
2913         return split;
2914 }
2915
2916 static struct shrinker deferred_split_shrinker = {
2917         .count_objects = deferred_split_count,
2918         .scan_objects = deferred_split_scan,
2919         .seeks = DEFAULT_SEEKS,
2920         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2921                  SHRINKER_NONSLAB,
2922 };
2923
2924 #ifdef CONFIG_DEBUG_FS
2925 static void split_huge_pages_all(void)
2926 {
2927         struct zone *zone;
2928         struct page *page;
2929         struct folio *folio;
2930         unsigned long pfn, max_zone_pfn;
2931         unsigned long total = 0, split = 0;
2932
2933         pr_debug("Split all THPs\n");
2934         for_each_zone(zone) {
2935                 if (!managed_zone(zone))
2936                         continue;
2937                 max_zone_pfn = zone_end_pfn(zone);
2938                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2939                         int nr_pages;
2940
2941                         page = pfn_to_online_page(pfn);
2942                         if (!page || PageTail(page))
2943                                 continue;
2944                         folio = page_folio(page);
2945                         if (!folio_try_get(folio))
2946                                 continue;
2947
2948                         if (unlikely(page_folio(page) != folio))
2949                                 goto next;
2950
2951                         if (zone != folio_zone(folio))
2952                                 goto next;
2953
2954                         if (!folio_test_large(folio)
2955                                 || folio_test_hugetlb(folio)
2956                                 || !folio_test_lru(folio))
2957                                 goto next;
2958
2959                         total++;
2960                         folio_lock(folio);
2961                         nr_pages = folio_nr_pages(folio);
2962                         if (!split_folio(folio))
2963                                 split++;
2964                         pfn += nr_pages - 1;
2965                         folio_unlock(folio);
2966 next:
2967                         folio_put(folio);
2968                         cond_resched();
2969                 }
2970         }
2971
2972         pr_debug("%lu of %lu THP split\n", split, total);
2973 }
2974
2975 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2976 {
2977         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2978                     is_vm_hugetlb_page(vma);
2979 }
2980
2981 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2982                                 unsigned long vaddr_end)
2983 {
2984         int ret = 0;
2985         struct task_struct *task;
2986         struct mm_struct *mm;
2987         unsigned long total = 0, split = 0;
2988         unsigned long addr;
2989
2990         vaddr_start &= PAGE_MASK;
2991         vaddr_end &= PAGE_MASK;
2992
2993         /* Find the task_struct from pid */
2994         rcu_read_lock();
2995         task = find_task_by_vpid(pid);
2996         if (!task) {
2997                 rcu_read_unlock();
2998                 ret = -ESRCH;
2999                 goto out;
3000         }
3001         get_task_struct(task);
3002         rcu_read_unlock();
3003
3004         /* Find the mm_struct */
3005         mm = get_task_mm(task);
3006         put_task_struct(task);
3007
3008         if (!mm) {
3009                 ret = -EINVAL;
3010                 goto out;
3011         }
3012
3013         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3014                  pid, vaddr_start, vaddr_end);
3015
3016         mmap_read_lock(mm);
3017         /*
3018          * always increase addr by PAGE_SIZE, since we could have a PTE page
3019          * table filled with PTE-mapped THPs, each of which is distinct.
3020          */
3021         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3022                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3023                 struct page *page;
3024
3025                 if (!vma)
3026                         break;
3027
3028                 /* skip special VMA and hugetlb VMA */
3029                 if (vma_not_suitable_for_thp_split(vma)) {
3030                         addr = vma->vm_end;
3031                         continue;
3032                 }
3033
3034                 /* FOLL_DUMP to ignore special (like zero) pages */
3035                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3036
3037                 if (IS_ERR_OR_NULL(page))
3038                         continue;
3039
3040                 if (!is_transparent_hugepage(page))
3041                         goto next;
3042
3043                 total++;
3044                 if (!can_split_folio(page_folio(page), NULL))
3045                         goto next;
3046
3047                 if (!trylock_page(page))
3048                         goto next;
3049
3050                 if (!split_huge_page(page))
3051                         split++;
3052
3053                 unlock_page(page);
3054 next:
3055                 put_page(page);
3056                 cond_resched();
3057         }
3058         mmap_read_unlock(mm);
3059         mmput(mm);
3060
3061         pr_debug("%lu of %lu THP split\n", split, total);
3062
3063 out:
3064         return ret;
3065 }
3066
3067 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3068                                 pgoff_t off_end)
3069 {
3070         struct filename *file;
3071         struct file *candidate;
3072         struct address_space *mapping;
3073         int ret = -EINVAL;
3074         pgoff_t index;
3075         int nr_pages = 1;
3076         unsigned long total = 0, split = 0;
3077
3078         file = getname_kernel(file_path);
3079         if (IS_ERR(file))
3080                 return ret;
3081
3082         candidate = file_open_name(file, O_RDONLY, 0);
3083         if (IS_ERR(candidate))
3084                 goto out;
3085
3086         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3087                  file_path, off_start, off_end);
3088
3089         mapping = candidate->f_mapping;
3090
3091         for (index = off_start; index < off_end; index += nr_pages) {
3092                 struct folio *folio = filemap_get_folio(mapping, index);
3093
3094                 nr_pages = 1;
3095                 if (IS_ERR(folio))
3096                         continue;
3097
3098                 if (!folio_test_large(folio))
3099                         goto next;
3100
3101                 total++;
3102                 nr_pages = folio_nr_pages(folio);
3103
3104                 if (!folio_trylock(folio))
3105                         goto next;
3106
3107                 if (!split_folio(folio))
3108                         split++;
3109
3110                 folio_unlock(folio);
3111 next:
3112                 folio_put(folio);
3113                 cond_resched();
3114         }
3115
3116         filp_close(candidate, NULL);
3117         ret = 0;
3118
3119         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3120 out:
3121         putname(file);
3122         return ret;
3123 }
3124
3125 #define MAX_INPUT_BUF_SZ 255
3126
3127 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3128                                 size_t count, loff_t *ppops)
3129 {
3130         static DEFINE_MUTEX(split_debug_mutex);
3131         ssize_t ret;
3132         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3133         char input_buf[MAX_INPUT_BUF_SZ];
3134         int pid;
3135         unsigned long vaddr_start, vaddr_end;
3136
3137         ret = mutex_lock_interruptible(&split_debug_mutex);
3138         if (ret)
3139                 return ret;
3140
3141         ret = -EFAULT;
3142
3143         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3144         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3145                 goto out;
3146
3147         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3148
3149         if (input_buf[0] == '/') {
3150                 char *tok;
3151                 char *buf = input_buf;
3152                 char file_path[MAX_INPUT_BUF_SZ];
3153                 pgoff_t off_start = 0, off_end = 0;
3154                 size_t input_len = strlen(input_buf);
3155
3156                 tok = strsep(&buf, ",");
3157                 if (tok) {
3158                         strcpy(file_path, tok);
3159                 } else {
3160                         ret = -EINVAL;
3161                         goto out;
3162                 }
3163
3164                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3165                 if (ret != 2) {
3166                         ret = -EINVAL;
3167                         goto out;
3168                 }
3169                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3170                 if (!ret)
3171                         ret = input_len;
3172
3173                 goto out;
3174         }
3175
3176         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3177         if (ret == 1 && pid == 1) {
3178                 split_huge_pages_all();
3179                 ret = strlen(input_buf);
3180                 goto out;
3181         } else if (ret != 3) {
3182                 ret = -EINVAL;
3183                 goto out;
3184         }
3185
3186         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3187         if (!ret)
3188                 ret = strlen(input_buf);
3189 out:
3190         mutex_unlock(&split_debug_mutex);
3191         return ret;
3192
3193 }
3194
3195 static const struct file_operations split_huge_pages_fops = {
3196         .owner   = THIS_MODULE,
3197         .write   = split_huge_pages_write,
3198         .llseek  = no_llseek,
3199 };
3200
3201 static int __init split_huge_pages_debugfs(void)
3202 {
3203         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3204                             &split_huge_pages_fops);
3205         return 0;
3206 }
3207 late_initcall(split_huge_pages_debugfs);
3208 #endif
3209
3210 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3211 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3212                 struct page *page)
3213 {
3214         struct vm_area_struct *vma = pvmw->vma;
3215         struct mm_struct *mm = vma->vm_mm;
3216         unsigned long address = pvmw->address;
3217         bool anon_exclusive;
3218         pmd_t pmdval;
3219         swp_entry_t entry;
3220         pmd_t pmdswp;
3221
3222         if (!(pvmw->pmd && !pvmw->pte))
3223                 return 0;
3224
3225         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3226         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3227
3228         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3229         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3230         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3231                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3232                 return -EBUSY;
3233         }
3234
3235         if (pmd_dirty(pmdval))
3236                 set_page_dirty(page);
3237         if (pmd_write(pmdval))
3238                 entry = make_writable_migration_entry(page_to_pfn(page));
3239         else if (anon_exclusive)
3240                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3241         else
3242                 entry = make_readable_migration_entry(page_to_pfn(page));
3243         if (pmd_young(pmdval))
3244                 entry = make_migration_entry_young(entry);
3245         if (pmd_dirty(pmdval))
3246                 entry = make_migration_entry_dirty(entry);
3247         pmdswp = swp_entry_to_pmd(entry);
3248         if (pmd_soft_dirty(pmdval))
3249                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3250         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3251         page_remove_rmap(page, vma, true);
3252         put_page(page);
3253         trace_set_migration_pmd(address, pmd_val(pmdswp));
3254
3255         return 0;
3256 }
3257
3258 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3259 {
3260         struct vm_area_struct *vma = pvmw->vma;
3261         struct mm_struct *mm = vma->vm_mm;
3262         unsigned long address = pvmw->address;
3263         unsigned long haddr = address & HPAGE_PMD_MASK;
3264         pmd_t pmde;
3265         swp_entry_t entry;
3266
3267         if (!(pvmw->pmd && !pvmw->pte))
3268                 return;
3269
3270         entry = pmd_to_swp_entry(*pvmw->pmd);
3271         get_page(new);
3272         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3273         if (pmd_swp_soft_dirty(*pvmw->pmd))
3274                 pmde = pmd_mksoft_dirty(pmde);
3275         if (pmd_swp_uffd_wp(*pvmw->pmd))
3276                 pmde = pmd_mkuffd_wp(pmde);
3277         if (!is_migration_entry_young(entry))
3278                 pmde = pmd_mkold(pmde);
3279         /* NOTE: this may contain setting soft-dirty on some archs */
3280         if (PageDirty(new) && is_migration_entry_dirty(entry))
3281                 pmde = pmd_mkdirty(pmde);
3282         if (is_writable_migration_entry(entry))
3283                 pmde = maybe_pmd_mkwrite(pmde, vma);
3284         else
3285                 pmde = pmd_wrprotect(pmde);
3286
3287         if (PageAnon(new)) {
3288                 rmap_t rmap_flags = RMAP_COMPOUND;
3289
3290                 if (!is_readable_migration_entry(entry))
3291                         rmap_flags |= RMAP_EXCLUSIVE;
3292
3293                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3294         } else {
3295                 page_add_file_rmap(new, vma, true);
3296         }
3297         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3298         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3299
3300         /* No need to invalidate - it was non-present before */
3301         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3302         trace_remove_migration_pmd(address, pmd_val(pmde));
3303 }
3304 #endif