mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/backing-dev.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37 #include <linux/page_owner.h>
  38 #include <linux/sched/sysctl.h>
  39 #include <linux/memory-tiers.h>
  40
  41 #include <asm/tlb.h>
  42 #include <asm/pgalloc.h>
  43 #include "internal.h"
  44 #include "swap.h"
  45
  46 #define CREATE_TRACE_POINTS
  47 #include <trace/events/thp.h>
  48
  49 /*
  50  * By default, transparent hugepage support is disabled in order to avoid
  51  * risking an increased memory footprint for applications that are not
  52  * guaranteed to benefit from it. When transparent hugepage support is
  53  * enabled, it is for all mappings, and khugepaged scans all mappings.
  54  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  55  * for all hugepage allocations.
  56  */
  57 unsigned long transparent_hugepage_flags __read_mostly =
  58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  59         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  60 #endif
  61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  62         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  63 #endif
  64         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  65         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  66         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  67
  68 static struct shrinker deferred_split_shrinker;
  69
  70 static atomic_t huge_zero_refcount;
  71 struct page *huge_zero_page __read_mostly;
  72 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  73
  74 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
  75                         bool smaps, bool in_pf, bool enforce_sysfs)
  76 {
  77         if (!vma->vm_mm)                /* vdso */
  78                 return false;
  79
  80         /*
  81          * Explicitly disabled through madvise or prctl, or some
  82          * architectures may disable THP for some mappings, for
  83          * example, s390 kvm.
  84          * */
  85         if ((vm_flags & VM_NOHUGEPAGE) ||
  86             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  87                 return false;
  88         /*
  89          * If the hardware/firmware marked hugepage support disabled.
  90          */
  91         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
  92                 return false;
  93
  94         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  95         if (vma_is_dax(vma))
  96                 return in_pf;
  97
  98         /*
  99          * Special VMA and hugetlb VMA.
 100          * Must be checked after dax since some dax mappings may have
 101          * VM_MIXEDMAP set.
 102          */
 103         if (vm_flags & VM_NO_KHUGEPAGED)
 104                 return false;
 105
 106         /*
 107          * Check alignment for file vma and size for both file and anon vma.
 108          *
 109          * Skip the check for page fault. Huge fault does the check in fault
 110          * handlers. And this check is not suitable for huge PUD fault.
 111          */
 112         if (!in_pf &&
 113             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 114                 return false;
 115
 116         /*
 117          * Enabled via shmem mount options or sysfs settings.
 118          * Must be done before hugepage flags check since shmem has its
 119          * own flags.
 120          */
 121         if (!in_pf && shmem_file(vma->vm_file))
 122                 return shmem_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                 /*
 464                  * Hardware doesn't support hugepages, hence disable
 465                  * DAX PMD support.
 466                  */
 467                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 468                 return -EINVAL;
 469         }
 470
 471         /*
 472          * hugepages can't be allocated by the buddy allocator
 473          */
 474         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 475         /*
 476          * we use page->mapping and page->index in second tail page
 477          * as list_head: assuming THP order >= 2
 478          */
 479         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 480
 481         err = hugepage_init_sysfs(&hugepage_kobj);
 482         if (err)
 483                 goto err_sysfs;
 484
 485         err = khugepaged_init();
 486         if (err)
 487                 goto err_slab;
 488
 489         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 490         if (err)
 491                 goto err_hzp_shrinker;
 492         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 493         if (err)
 494                 goto err_split_shrinker;
 495
 496         /*
 497          * By default disable transparent hugepages on smaller systems,
 498          * where the extra memory used could hurt more than TLB overhead
 499          * is likely to save.  The admin can still enable it through /sys.
 500          */
 501         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 502                 transparent_hugepage_flags = 0;
 503                 return 0;
 504         }
 505
 506         err = start_stop_khugepaged();
 507         if (err)
 508                 goto err_khugepaged;
 509
 510         return 0;
 511 err_khugepaged:
 512         unregister_shrinker(&deferred_split_shrinker);
 513 err_split_shrinker:
 514         unregister_shrinker(&huge_zero_page_shrinker);
 515 err_hzp_shrinker:
 516         khugepaged_destroy();
 517 err_slab:
 518         hugepage_exit_sysfs(hugepage_kobj);
 519 err_sysfs:
 520         return err;
 521 }
 522 subsys_initcall(hugepage_init);
 523
 524 static int __init setup_transparent_hugepage(char *str)
 525 {
 526         int ret = 0;
 527         if (!str)
 528                 goto out;
 529         if (!strcmp(str, "always")) {
 530                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 531                         &transparent_hugepage_flags);
 532                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 533                           &transparent_hugepage_flags);
 534                 ret = 1;
 535         } else if (!strcmp(str, "madvise")) {
 536                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 537                           &transparent_hugepage_flags);
 538                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 539                         &transparent_hugepage_flags);
 540                 ret = 1;
 541         } else if (!strcmp(str, "never")) {
 542                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 543                           &transparent_hugepage_flags);
 544                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 545                           &transparent_hugepage_flags);
 546                 ret = 1;
 547         }
 548 out:
 549         if (!ret)
 550                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 551         return ret;
 552 }
 553 __setup("transparent_hugepage=", setup_transparent_hugepage);
 554
 555 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 556 {
 557         if (likely(vma->vm_flags & VM_WRITE))
 558                 pmd = pmd_mkwrite(pmd);
 559         return pmd;
 560 }
 561
 562 #ifdef CONFIG_MEMCG
 563 static inline
 564 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 565 {
 566         struct mem_cgroup *memcg = folio_memcg(folio);
 567         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 568
 569         if (memcg)
 570                 return &memcg->deferred_split_queue;
 571         else
 572                 return &pgdat->deferred_split_queue;
 573 }
 574 #else
 575 static inline
 576 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 577 {
 578         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 579
 580         return &pgdat->deferred_split_queue;
 581 }
 582 #endif
 583
 584 void prep_transhuge_page(struct page *page)
 585 {
 586         struct folio *folio = (struct folio *)page;
 587
 588         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
 589         INIT_LIST_HEAD(&folio->_deferred_list);
 590         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 591 }
 592
 593 static inline bool is_transparent_hugepage(struct page *page)
 594 {
 595         struct folio *folio;
 596
 597         if (!PageCompound(page))
 598                 return false;
 599
 600         folio = page_folio(page);
 601         return is_huge_zero_page(&folio->page) ||
 602                folio->_folio_dtor == TRANSHUGE_PAGE_DTOR;
 603 }
 604
 605 static unsigned long __thp_get_unmapped_area(struct file *filp,
 606                 unsigned long addr, unsigned long len,
 607                 loff_t off, unsigned long flags, unsigned long size)
 608 {
 609         loff_t off_end = off + len;
 610         loff_t off_align = round_up(off, size);
 611         unsigned long len_pad, ret;
 612
 613         if (off_end <= off_align || (off_end - off_align) < size)
 614                 return 0;
 615
 616         len_pad = len + size;
 617         if (len_pad < len || (off + len_pad) < off)
 618                 return 0;
 619
 620         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 621                                               off >> PAGE_SHIFT, flags);
 622
 623         /*
 624          * The failure might be due to length padding. The caller will retry
 625          * without the padding.
 626          */
 627         if (IS_ERR_VALUE(ret))
 628                 return 0;
 629
 630         /*
 631          * Do not try to align to THP boundary if allocation at the address
 632          * hint succeeds.
 633          */
 634         if (ret == addr)
 635                 return addr;
 636
 637         ret += (off - ret) & (size - 1);
 638         return ret;
 639 }
 640
 641 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 642                 unsigned long len, unsigned long pgoff, unsigned long flags)
 643 {
 644         unsigned long ret;
 645         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 646
 647         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 648         if (ret)
 649                 return ret;
 650
 651         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 652 }
 653 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 654
 655 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 656                         struct page *page, gfp_t gfp)
 657 {
 658         struct vm_area_struct *vma = vmf->vma;
 659         pgtable_t pgtable;
 660         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 661         vm_fault_t ret = 0;
 662
 663         VM_BUG_ON_PAGE(!PageCompound(page), page);
 664
 665         if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
 666                 put_page(page);
 667                 count_vm_event(THP_FAULT_FALLBACK);
 668                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 669                 return VM_FAULT_FALLBACK;
 670         }
 671         cgroup_throttle_swaprate(page, gfp);
 672
 673         pgtable = pte_alloc_one(vma->vm_mm);
 674         if (unlikely(!pgtable)) {
 675                 ret = VM_FAULT_OOM;
 676                 goto release;
 677         }
 678
 679         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 680         /*
 681          * The memory barrier inside __SetPageUptodate makes sure that
 682          * clear_huge_page writes become visible before the set_pmd_at()
 683          * write.
 684          */
 685         __SetPageUptodate(page);
 686
 687         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 688         if (unlikely(!pmd_none(*vmf->pmd))) {
 689                 goto unlock_release;
 690         } else {
 691                 pmd_t entry;
 692
 693                 ret = check_stable_address_space(vma->vm_mm);
 694                 if (ret)
 695                         goto unlock_release;
 696
 697                 /* Deliver the page fault to userland */
 698                 if (userfaultfd_missing(vma)) {
 699                         spin_unlock(vmf->ptl);
 700                         put_page(page);
 701                         pte_free(vma->vm_mm, pgtable);
 702                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 703                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 704                         return ret;
 705                 }
 706
 707                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 708                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 709                 page_add_new_anon_rmap(page, vma, haddr);
 710                 lru_cache_add_inactive_or_unevictable(page, vma);
 711                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 712                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 713                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 714                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 715                 mm_inc_nr_ptes(vma->vm_mm);
 716                 spin_unlock(vmf->ptl);
 717                 count_vm_event(THP_FAULT_ALLOC);
 718                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 719         }
 720
 721         return 0;
 722 unlock_release:
 723         spin_unlock(vmf->ptl);
 724 release:
 725         if (pgtable)
 726                 pte_free(vma->vm_mm, pgtable);
 727         put_page(page);
 728         return ret;
 729
 730 }
 731
 732 /*
 733  * always: directly stall for all thp allocations
 734  * defer: wake kswapd and fail if not immediately available
 735  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 736  *                fail if not immediately available
 737  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 738  *          available
 739  * never: never stall for any thp allocation
 740  */
 741 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 742 {
 743         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 744
 745         /* Always do synchronous compaction */
 746         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 747                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 748
 749         /* Kick kcompactd and fail quickly */
 750         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 751                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 752
 753         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 754         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 755                 return GFP_TRANSHUGE_LIGHT |
 756                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 757                                         __GFP_KSWAPD_RECLAIM);
 758
 759         /* Only do synchronous compaction if madvised */
 760         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 761                 return GFP_TRANSHUGE_LIGHT |
 762                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 763
 764         return GFP_TRANSHUGE_LIGHT;
 765 }
 766
 767 /* Caller must hold page table lock. */
 768 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 769                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 770                 struct page *zero_page)
 771 {
 772         pmd_t entry;
 773         if (!pmd_none(*pmd))
 774                 return;
 775         entry = mk_pmd(zero_page, vma->vm_page_prot);
 776         entry = pmd_mkhuge(entry);
 777         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 778         set_pmd_at(mm, haddr, pmd, entry);
 779         mm_inc_nr_ptes(mm);
 780 }
 781
 782 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 783 {
 784         struct vm_area_struct *vma = vmf->vma;
 785         gfp_t gfp;
 786         struct folio *folio;
 787         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 788
 789         if (!transhuge_vma_suitable(vma, haddr))
 790                 return VM_FAULT_FALLBACK;
 791         if (unlikely(anon_vma_prepare(vma)))
 792                 return VM_FAULT_OOM;
 793         khugepaged_enter_vma(vma, vma->vm_flags);
 794
 795         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 796                         !mm_forbids_zeropage(vma->vm_mm) &&
 797                         transparent_hugepage_use_zero_page()) {
 798                 pgtable_t pgtable;
 799                 struct page *zero_page;
 800                 vm_fault_t ret;
 801                 pgtable = pte_alloc_one(vma->vm_mm);
 802                 if (unlikely(!pgtable))
 803                         return VM_FAULT_OOM;
 804                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 805                 if (unlikely(!zero_page)) {
 806                         pte_free(vma->vm_mm, pgtable);
 807                         count_vm_event(THP_FAULT_FALLBACK);
 808                         return VM_FAULT_FALLBACK;
 809                 }
 810                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 811                 ret = 0;
 812                 if (pmd_none(*vmf->pmd)) {
 813                         ret = check_stable_address_space(vma->vm_mm);
 814                         if (ret) {
 815                                 spin_unlock(vmf->ptl);
 816                                 pte_free(vma->vm_mm, pgtable);
 817                         } else if (userfaultfd_missing(vma)) {
 818                                 spin_unlock(vmf->ptl);
 819                                 pte_free(vma->vm_mm, pgtable);
 820                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 821                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 822                         } else {
 823                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 824                                                    haddr, vmf->pmd, zero_page);
 825                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 826                                 spin_unlock(vmf->ptl);
 827                         }
 828                 } else {
 829                         spin_unlock(vmf->ptl);
 830                         pte_free(vma->vm_mm, pgtable);
 831                 }
 832                 return ret;
 833         }
 834         gfp = vma_thp_gfp_mask(vma);
 835         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 836         if (unlikely(!folio)) {
 837                 count_vm_event(THP_FAULT_FALLBACK);
 838                 return VM_FAULT_FALLBACK;
 839         }
 840         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 841 }
 842
 843 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 844                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 845                 pgtable_t pgtable)
 846 {
 847         struct mm_struct *mm = vma->vm_mm;
 848         pmd_t entry;
 849         spinlock_t *ptl;
 850
 851         ptl = pmd_lock(mm, pmd);
 852         if (!pmd_none(*pmd)) {
 853                 if (write) {
 854                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 855                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 856                                 goto out_unlock;
 857                         }
 858                         entry = pmd_mkyoung(*pmd);
 859                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 860                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 861                                 update_mmu_cache_pmd(vma, addr, pmd);
 862                 }
 863
 864                 goto out_unlock;
 865         }
 866
 867         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 868         if (pfn_t_devmap(pfn))
 869                 entry = pmd_mkdevmap(entry);
 870         if (write) {
 871                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 872                 entry = maybe_pmd_mkwrite(entry, vma);
 873         }
 874
 875         if (pgtable) {
 876                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 877                 mm_inc_nr_ptes(mm);
 878                 pgtable = NULL;
 879         }
 880
 881         set_pmd_at(mm, addr, pmd, entry);
 882         update_mmu_cache_pmd(vma, addr, pmd);
 883
 884 out_unlock:
 885         spin_unlock(ptl);
 886         if (pgtable)
 887                 pte_free(mm, pgtable);
 888 }
 889
 890 /**
 891  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 892  * @vmf: Structure describing the fault
 893  * @pfn: pfn to insert
 894  * @pgprot: page protection to use
 895  * @write: whether it's a write fault
 896  *
 897  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 898  * also consult the vmf_insert_mixed_prot() documentation when
 899  * @pgprot != @vmf->vma->vm_page_prot.
 900  *
 901  * Return: vm_fault_t value.
 902  */
 903 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 904                                    pgprot_t pgprot, bool write)
 905 {
 906         unsigned long addr = vmf->address & PMD_MASK;
 907         struct vm_area_struct *vma = vmf->vma;
 908         pgtable_t pgtable = NULL;
 909
 910         /*
 911          * If we had pmd_special, we could avoid all these restrictions,
 912          * but we need to be consistent with PTEs and architectures that
 913          * can't support a 'special' bit.
 914          */
 915         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 916                         !pfn_t_devmap(pfn));
 917         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 918                                                 (VM_PFNMAP|VM_MIXEDMAP));
 919         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 920
 921         if (addr < vma->vm_start || addr >= vma->vm_end)
 922                 return VM_FAULT_SIGBUS;
 923
 924         if (arch_needs_pgtable_deposit()) {
 925                 pgtable = pte_alloc_one(vma->vm_mm);
 926                 if (!pgtable)
 927                         return VM_FAULT_OOM;
 928         }
 929
 930         track_pfn_insert(vma, &pgprot, pfn);
 931
 932         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 933         return VM_FAULT_NOPAGE;
 934 }
 935 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 936
 937 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 938 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 939 {
 940         if (likely(vma->vm_flags & VM_WRITE))
 941                 pud = pud_mkwrite(pud);
 942         return pud;
 943 }
 944
 945 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 946                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 947 {
 948         struct mm_struct *mm = vma->vm_mm;
 949         pud_t entry;
 950         spinlock_t *ptl;
 951
 952         ptl = pud_lock(mm, pud);
 953         if (!pud_none(*pud)) {
 954                 if (write) {
 955                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 956                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 957                                 goto out_unlock;
 958                         }
 959                         entry = pud_mkyoung(*pud);
 960                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 961                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 962                                 update_mmu_cache_pud(vma, addr, pud);
 963                 }
 964                 goto out_unlock;
 965         }
 966
 967         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 968         if (pfn_t_devmap(pfn))
 969                 entry = pud_mkdevmap(entry);
 970         if (write) {
 971                 entry = pud_mkyoung(pud_mkdirty(entry));
 972                 entry = maybe_pud_mkwrite(entry, vma);
 973         }
 974         set_pud_at(mm, addr, pud, entry);
 975         update_mmu_cache_pud(vma, addr, pud);
 976
 977 out_unlock:
 978         spin_unlock(ptl);
 979 }
 980
 981 /**
 982  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 983  * @vmf: Structure describing the fault
 984  * @pfn: pfn to insert
 985  * @pgprot: page protection to use
 986  * @write: whether it's a write fault
 987  *
 988  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 989  * also consult the vmf_insert_mixed_prot() documentation when
 990  * @pgprot != @vmf->vma->vm_page_prot.
 991  *
 992  * Return: vm_fault_t value.
 993  */
 994 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 995                                    pgprot_t pgprot, bool write)
 996 {
 997         unsigned long addr = vmf->address & PUD_MASK;
 998         struct vm_area_struct *vma = vmf->vma;
 999
1000         /*
1001          * If we had pud_special, we could avoid all these restrictions,
1002          * but we need to be consistent with PTEs and architectures that
1003          * can't support a 'special' bit.
1004          */
1005         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1006                         !pfn_t_devmap(pfn));
1007         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1008                                                 (VM_PFNMAP|VM_MIXEDMAP));
1009         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1010
1011         if (addr < vma->vm_start || addr >= vma->vm_end)
1012                 return VM_FAULT_SIGBUS;
1013
1014         track_pfn_insert(vma, &pgprot, pfn);
1015
1016         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1017         return VM_FAULT_NOPAGE;
1018 }
1019 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1020 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1021
1022 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1023                       pmd_t *pmd, bool write)
1024 {
1025         pmd_t _pmd;
1026
1027         _pmd = pmd_mkyoung(*pmd);
1028         if (write)
1029                 _pmd = pmd_mkdirty(_pmd);
1030         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031                                   pmd, _pmd, write))
1032                 update_mmu_cache_pmd(vma, addr, pmd);
1033 }
1034
1035 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1037 {
1038         unsigned long pfn = pmd_pfn(*pmd);
1039         struct mm_struct *mm = vma->vm_mm;
1040         struct page *page;
1041         int ret;
1042
1043         assert_spin_locked(pmd_lockptr(mm, pmd));
1044
1045         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1046                 return NULL;
1047
1048         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1049                 /* pass */;
1050         else
1051                 return NULL;
1052
1053         if (flags & FOLL_TOUCH)
1054                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1055
1056         /*
1057          * device mapped pages can only be returned if the
1058          * caller will manage the page reference count.
1059          */
1060         if (!(flags & (FOLL_GET | FOLL_PIN)))
1061                 return ERR_PTR(-EEXIST);
1062
1063         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1064         *pgmap = get_dev_pagemap(pfn, *pgmap);
1065         if (!*pgmap)
1066                 return ERR_PTR(-EFAULT);
1067         page = pfn_to_page(pfn);
1068         ret = try_grab_page(page, flags);
1069         if (ret)
1070                 page = ERR_PTR(ret);
1071
1072         return page;
1073 }
1074
1075 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1076                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1077                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1078 {
1079         spinlock_t *dst_ptl, *src_ptl;
1080         struct page *src_page;
1081         pmd_t pmd;
1082         pgtable_t pgtable = NULL;
1083         int ret = -ENOMEM;
1084
1085         /* Skip if can be re-fill on fault */
1086         if (!vma_is_anonymous(dst_vma))
1087                 return 0;
1088
1089         pgtable = pte_alloc_one(dst_mm);
1090         if (unlikely(!pgtable))
1091                 goto out;
1092
1093         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1094         src_ptl = pmd_lockptr(src_mm, src_pmd);
1095         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1096
1097         ret = -EAGAIN;
1098         pmd = *src_pmd;
1099
1100 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1101         if (unlikely(is_swap_pmd(pmd))) {
1102                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1103
1104                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1105                 if (!is_readable_migration_entry(entry)) {
1106                         entry = make_readable_migration_entry(
1107                                                         swp_offset(entry));
1108                         pmd = swp_entry_to_pmd(entry);
1109                         if (pmd_swp_soft_dirty(*src_pmd))
1110                                 pmd = pmd_swp_mksoft_dirty(pmd);
1111                         if (pmd_swp_uffd_wp(*src_pmd))
1112                                 pmd = pmd_swp_mkuffd_wp(pmd);
1113                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1114                 }
1115                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1116                 mm_inc_nr_ptes(dst_mm);
1117                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1118                 if (!userfaultfd_wp(dst_vma))
1119                         pmd = pmd_swp_clear_uffd_wp(pmd);
1120                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1121                 ret = 0;
1122                 goto out_unlock;
1123         }
1124 #endif
1125
1126         if (unlikely(!pmd_trans_huge(pmd))) {
1127                 pte_free(dst_mm, pgtable);
1128                 goto out_unlock;
1129         }
1130         /*
1131          * When page table lock is held, the huge zero pmd should not be
1132          * under splitting since we don't split the page itself, only pmd to
1133          * a page table.
1134          */
1135         if (is_huge_zero_pmd(pmd)) {
1136                 /*
1137                  * get_huge_zero_page() will never allocate a new page here,
1138                  * since we already have a zero page to copy. It just takes a
1139                  * reference.
1140                  */
1141                 mm_get_huge_zero_page(dst_mm);
1142                 goto out_zero_page;
1143         }
1144
1145         src_page = pmd_page(pmd);
1146         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1147
1148         get_page(src_page);
1149         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1150                 /* Page maybe pinned: split and retry the fault on PTEs. */
1151                 put_page(src_page);
1152                 pte_free(dst_mm, pgtable);
1153                 spin_unlock(src_ptl);
1154                 spin_unlock(dst_ptl);
1155                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1156                 return -EAGAIN;
1157         }
1158         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1159 out_zero_page:
1160         mm_inc_nr_ptes(dst_mm);
1161         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1162         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1163         if (!userfaultfd_wp(dst_vma))
1164                 pmd = pmd_clear_uffd_wp(pmd);
1165         pmd = pmd_mkold(pmd_wrprotect(pmd));
1166         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1167
1168         ret = 0;
1169 out_unlock:
1170         spin_unlock(src_ptl);
1171         spin_unlock(dst_ptl);
1172 out:
1173         return ret;
1174 }
1175
1176 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1177 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1178                       pud_t *pud, bool write)
1179 {
1180         pud_t _pud;
1181
1182         _pud = pud_mkyoung(*pud);
1183         if (write)
1184                 _pud = pud_mkdirty(_pud);
1185         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1186                                   pud, _pud, write))
1187                 update_mmu_cache_pud(vma, addr, pud);
1188 }
1189
1190 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1191                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1192 {
1193         unsigned long pfn = pud_pfn(*pud);
1194         struct mm_struct *mm = vma->vm_mm;
1195         struct page *page;
1196         int ret;
1197
1198         assert_spin_locked(pud_lockptr(mm, pud));
1199
1200         if (flags & FOLL_WRITE && !pud_write(*pud))
1201                 return NULL;
1202
1203         if (pud_present(*pud) && pud_devmap(*pud))
1204                 /* pass */;
1205         else
1206                 return NULL;
1207
1208         if (flags & FOLL_TOUCH)
1209                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1210
1211         /*
1212          * device mapped pages can only be returned if the
1213          * caller will manage the page reference count.
1214          *
1215          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1216          */
1217         if (!(flags & (FOLL_GET | FOLL_PIN)))
1218                 return ERR_PTR(-EEXIST);
1219
1220         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1221         *pgmap = get_dev_pagemap(pfn, *pgmap);
1222         if (!*pgmap)
1223                 return ERR_PTR(-EFAULT);
1224         page = pfn_to_page(pfn);
1225
1226         ret = try_grab_page(page, flags);
1227         if (ret)
1228                 page = ERR_PTR(ret);
1229
1230         return page;
1231 }
1232
1233 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1234                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1235                   struct vm_area_struct *vma)
1236 {
1237         spinlock_t *dst_ptl, *src_ptl;
1238         pud_t pud;
1239         int ret;
1240
1241         dst_ptl = pud_lock(dst_mm, dst_pud);
1242         src_ptl = pud_lockptr(src_mm, src_pud);
1243         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1244
1245         ret = -EAGAIN;
1246         pud = *src_pud;
1247         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1248                 goto out_unlock;
1249
1250         /*
1251          * When page table lock is held, the huge zero pud should not be
1252          * under splitting since we don't split the page itself, only pud to
1253          * a page table.
1254          */
1255         if (is_huge_zero_pud(pud)) {
1256                 /* No huge zero pud yet */
1257         }
1258
1259         /*
1260          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1261          * and split if duplicating fails.
1262          */
1263         pudp_set_wrprotect(src_mm, addr, src_pud);
1264         pud = pud_mkold(pud_wrprotect(pud));
1265         set_pud_at(dst_mm, addr, dst_pud, pud);
1266
1267         ret = 0;
1268 out_unlock:
1269         spin_unlock(src_ptl);
1270         spin_unlock(dst_ptl);
1271         return ret;
1272 }
1273
1274 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1275 {
1276         bool write = vmf->flags & FAULT_FLAG_WRITE;
1277
1278         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1279         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1280                 goto unlock;
1281
1282         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1283 unlock:
1284         spin_unlock(vmf->ptl);
1285 }
1286 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1287
1288 void huge_pmd_set_accessed(struct vm_fault *vmf)
1289 {
1290         bool write = vmf->flags & FAULT_FLAG_WRITE;
1291
1292         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1293         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1294                 goto unlock;
1295
1296         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1297
1298 unlock:
1299         spin_unlock(vmf->ptl);
1300 }
1301
1302 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1303 {
1304         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1305         struct vm_area_struct *vma = vmf->vma;
1306         struct folio *folio;
1307         struct page *page;
1308         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1309         pmd_t orig_pmd = vmf->orig_pmd;
1310
1311         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1312         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1313
1314         if (is_huge_zero_pmd(orig_pmd))
1315                 goto fallback;
1316
1317         spin_lock(vmf->ptl);
1318
1319         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1320                 spin_unlock(vmf->ptl);
1321                 return 0;
1322         }
1323
1324         page = pmd_page(orig_pmd);
1325         folio = page_folio(page);
1326         VM_BUG_ON_PAGE(!PageHead(page), page);
1327
1328         /* Early check when only holding the PT lock. */
1329         if (PageAnonExclusive(page))
1330                 goto reuse;
1331
1332         if (!folio_trylock(folio)) {
1333                 folio_get(folio);
1334                 spin_unlock(vmf->ptl);
1335                 folio_lock(folio);
1336                 spin_lock(vmf->ptl);
1337                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1338                         spin_unlock(vmf->ptl);
1339                         folio_unlock(folio);
1340                         folio_put(folio);
1341                         return 0;
1342                 }
1343                 folio_put(folio);
1344         }
1345
1346         /* Recheck after temporarily dropping the PT lock. */
1347         if (PageAnonExclusive(page)) {
1348                 folio_unlock(folio);
1349                 goto reuse;
1350         }
1351
1352         /*
1353          * See do_wp_page(): we can only reuse the folio exclusively if
1354          * there are no additional references. Note that we always drain
1355          * the LRU pagevecs immediately after adding a THP.
1356          */
1357         if (folio_ref_count(folio) >
1358                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1359                 goto unlock_fallback;
1360         if (folio_test_swapcache(folio))
1361                 folio_free_swap(folio);
1362         if (folio_ref_count(folio) == 1) {
1363                 pmd_t entry;
1364
1365                 page_move_anon_rmap(page, vma);
1366                 folio_unlock(folio);
1367 reuse:
1368                 if (unlikely(unshare)) {
1369                         spin_unlock(vmf->ptl);
1370                         return 0;
1371                 }
1372                 entry = pmd_mkyoung(orig_pmd);
1373                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1374                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1375                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1376                 spin_unlock(vmf->ptl);
1377                 return 0;
1378         }
1379
1380 unlock_fallback:
1381         folio_unlock(folio);
1382         spin_unlock(vmf->ptl);
1383 fallback:
1384         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1385         return VM_FAULT_FALLBACK;
1386 }
1387
1388 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1389                                            unsigned long addr, pmd_t pmd)
1390 {
1391         struct page *page;
1392
1393         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1394                 return false;
1395
1396         /* Don't touch entries that are not even readable (NUMA hinting). */
1397         if (pmd_protnone(pmd))
1398                 return false;
1399
1400         /* Do we need write faults for softdirty tracking? */
1401         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1402                 return false;
1403
1404         /* Do we need write faults for uffd-wp tracking? */
1405         if (userfaultfd_huge_pmd_wp(vma, pmd))
1406                 return false;
1407
1408         if (!(vma->vm_flags & VM_SHARED)) {
1409                 /* See can_change_pte_writable(). */
1410                 page = vm_normal_page_pmd(vma, addr, pmd);
1411                 return page && PageAnon(page) && PageAnonExclusive(page);
1412         }
1413
1414         /* See can_change_pte_writable(). */
1415         return pmd_dirty(pmd);
1416 }
1417
1418 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1419 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1420                                         struct vm_area_struct *vma,
1421                                         unsigned int flags)
1422 {
1423         /* If the pmd is writable, we can write to the page. */
1424         if (pmd_write(pmd))
1425                 return true;
1426
1427         /* Maybe FOLL_FORCE is set to override it? */
1428         if (!(flags & FOLL_FORCE))
1429                 return false;
1430
1431         /* But FOLL_FORCE has no effect on shared mappings */
1432         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1433                 return false;
1434
1435         /* ... or read-only private ones */
1436         if (!(vma->vm_flags & VM_MAYWRITE))
1437                 return false;
1438
1439         /* ... or already writable ones that just need to take a write fault */
1440         if (vma->vm_flags & VM_WRITE)
1441                 return false;
1442
1443         /*
1444          * See can_change_pte_writable(): we broke COW and could map the page
1445          * writable if we have an exclusive anonymous page ...
1446          */
1447         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1448                 return false;
1449
1450         /* ... and a write-fault isn't required for other reasons. */
1451         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1452                 return false;
1453         return !userfaultfd_huge_pmd_wp(vma, pmd);
1454 }
1455
1456 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1457                                    unsigned long addr,
1458                                    pmd_t *pmd,
1459                                    unsigned int flags)
1460 {
1461         struct mm_struct *mm = vma->vm_mm;
1462         struct page *page;
1463         int ret;
1464
1465         assert_spin_locked(pmd_lockptr(mm, pmd));
1466
1467         page = pmd_page(*pmd);
1468         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1469
1470         if ((flags & FOLL_WRITE) &&
1471             !can_follow_write_pmd(*pmd, page, vma, flags))
1472                 return NULL;
1473
1474         /* Avoid dumping huge zero page */
1475         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1476                 return ERR_PTR(-EFAULT);
1477
1478         /* Full NUMA hinting faults to serialise migration in fault paths */
1479         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1480                 return NULL;
1481
1482         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1483                 return ERR_PTR(-EMLINK);
1484
1485         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1486                         !PageAnonExclusive(page), page);
1487
1488         ret = try_grab_page(page, flags);
1489         if (ret)
1490                 return ERR_PTR(ret);
1491
1492         if (flags & FOLL_TOUCH)
1493                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1494
1495         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1496         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1497
1498         return page;
1499 }
1500
1501 /* NUMA hinting page fault entry point for trans huge pmds */
1502 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1503 {
1504         struct vm_area_struct *vma = vmf->vma;
1505         pmd_t oldpmd = vmf->orig_pmd;
1506         pmd_t pmd;
1507         struct page *page;
1508         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1509         int page_nid = NUMA_NO_NODE;
1510         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1511         bool migrated = false, writable = false;
1512         int flags = 0;
1513
1514         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1515         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1516                 spin_unlock(vmf->ptl);
1517                 goto out;
1518         }
1519
1520         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1521
1522         /*
1523          * Detect now whether the PMD could be writable; this information
1524          * is only valid while holding the PT lock.
1525          */
1526         writable = pmd_write(pmd);
1527         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1528             can_change_pmd_writable(vma, vmf->address, pmd))
1529                 writable = true;
1530
1531         page = vm_normal_page_pmd(vma, haddr, pmd);
1532         if (!page)
1533                 goto out_map;
1534
1535         /* See similar comment in do_numa_page for explanation */
1536         if (!writable)
1537                 flags |= TNF_NO_GROUP;
1538
1539         page_nid = page_to_nid(page);
1540         /*
1541          * For memory tiering mode, cpupid of slow memory page is used
1542          * to record page access time.  So use default value.
1543          */
1544         if (node_is_toptier(page_nid))
1545                 last_cpupid = page_cpupid_last(page);
1546         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1547                                        &flags);
1548
1549         if (target_nid == NUMA_NO_NODE) {
1550                 put_page(page);
1551                 goto out_map;
1552         }
1553
1554         spin_unlock(vmf->ptl);
1555         writable = false;
1556
1557         migrated = migrate_misplaced_page(page, vma, target_nid);
1558         if (migrated) {
1559                 flags |= TNF_MIGRATED;
1560                 page_nid = target_nid;
1561         } else {
1562                 flags |= TNF_MIGRATE_FAIL;
1563                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1564                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1565                         spin_unlock(vmf->ptl);
1566                         goto out;
1567                 }
1568                 goto out_map;
1569         }
1570
1571 out:
1572         if (page_nid != NUMA_NO_NODE)
1573                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1574                                 flags);
1575
1576         return 0;
1577
1578 out_map:
1579         /* Restore the PMD */
1580         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1581         pmd = pmd_mkyoung(pmd);
1582         if (writable)
1583                 pmd = pmd_mkwrite(pmd);
1584         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1585         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1586         spin_unlock(vmf->ptl);
1587         goto out;
1588 }
1589
1590 /*
1591  * Return true if we do MADV_FREE successfully on entire pmd page.
1592  * Otherwise, return false.
1593  */
1594 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1595                 pmd_t *pmd, unsigned long addr, unsigned long next)
1596 {
1597         spinlock_t *ptl;
1598         pmd_t orig_pmd;
1599         struct folio *folio;
1600         struct mm_struct *mm = tlb->mm;
1601         bool ret = false;
1602
1603         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1604
1605         ptl = pmd_trans_huge_lock(pmd, vma);
1606         if (!ptl)
1607                 goto out_unlocked;
1608
1609         orig_pmd = *pmd;
1610         if (is_huge_zero_pmd(orig_pmd))
1611                 goto out;
1612
1613         if (unlikely(!pmd_present(orig_pmd))) {
1614                 VM_BUG_ON(thp_migration_supported() &&
1615                                   !is_pmd_migration_entry(orig_pmd));
1616                 goto out;
1617         }
1618
1619         folio = pfn_folio(pmd_pfn(orig_pmd));
1620         /*
1621          * If other processes are mapping this folio, we couldn't discard
1622          * the folio unless they all do MADV_FREE so let's skip the folio.
1623          */
1624         if (folio_mapcount(folio) != 1)
1625                 goto out;
1626
1627         if (!folio_trylock(folio))
1628                 goto out;
1629
1630         /*
1631          * If user want to discard part-pages of THP, split it so MADV_FREE
1632          * will deactivate only them.
1633          */
1634         if (next - addr != HPAGE_PMD_SIZE) {
1635                 folio_get(folio);
1636                 spin_unlock(ptl);
1637                 split_folio(folio);
1638                 folio_unlock(folio);
1639                 folio_put(folio);
1640                 goto out_unlocked;
1641         }
1642
1643         if (folio_test_dirty(folio))
1644                 folio_clear_dirty(folio);
1645         folio_unlock(folio);
1646
1647         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1648                 pmdp_invalidate(vma, addr, pmd);
1649                 orig_pmd = pmd_mkold(orig_pmd);
1650                 orig_pmd = pmd_mkclean(orig_pmd);
1651
1652                 set_pmd_at(mm, addr, pmd, orig_pmd);
1653                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1654         }
1655
1656         folio_mark_lazyfree(folio);
1657         ret = true;
1658 out:
1659         spin_unlock(ptl);
1660 out_unlocked:
1661         return ret;
1662 }
1663
1664 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1665 {
1666         pgtable_t pgtable;
1667
1668         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1669         pte_free(mm, pgtable);
1670         mm_dec_nr_ptes(mm);
1671 }
1672
1673 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1674                  pmd_t *pmd, unsigned long addr)
1675 {
1676         pmd_t orig_pmd;
1677         spinlock_t *ptl;
1678
1679         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1680
1681         ptl = __pmd_trans_huge_lock(pmd, vma);
1682         if (!ptl)
1683                 return 0;
1684         /*
1685          * For architectures like ppc64 we look at deposited pgtable
1686          * when calling pmdp_huge_get_and_clear. So do the
1687          * pgtable_trans_huge_withdraw after finishing pmdp related
1688          * operations.
1689          */
1690         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1691                                                 tlb->fullmm);
1692         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1693         if (vma_is_special_huge(vma)) {
1694                 if (arch_needs_pgtable_deposit())
1695                         zap_deposited_table(tlb->mm, pmd);
1696                 spin_unlock(ptl);
1697         } else if (is_huge_zero_pmd(orig_pmd)) {
1698                 zap_deposited_table(tlb->mm, pmd);
1699                 spin_unlock(ptl);
1700         } else {
1701                 struct page *page = NULL;
1702                 int flush_needed = 1;
1703
1704                 if (pmd_present(orig_pmd)) {
1705                         page = pmd_page(orig_pmd);
1706                         page_remove_rmap(page, vma, true);
1707                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1708                         VM_BUG_ON_PAGE(!PageHead(page), page);
1709                 } else if (thp_migration_supported()) {
1710                         swp_entry_t entry;
1711
1712                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1713                         entry = pmd_to_swp_entry(orig_pmd);
1714                         page = pfn_swap_entry_to_page(entry);
1715                         flush_needed = 0;
1716                 } else
1717                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1718
1719                 if (PageAnon(page)) {
1720                         zap_deposited_table(tlb->mm, pmd);
1721                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1722                 } else {
1723                         if (arch_needs_pgtable_deposit())
1724                                 zap_deposited_table(tlb->mm, pmd);
1725                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1726                 }
1727
1728                 spin_unlock(ptl);
1729                 if (flush_needed)
1730                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1731         }
1732         return 1;
1733 }
1734
1735 #ifndef pmd_move_must_withdraw
1736 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1737                                          spinlock_t *old_pmd_ptl,
1738                                          struct vm_area_struct *vma)
1739 {
1740         /*
1741          * With split pmd lock we also need to move preallocated
1742          * PTE page table if new_pmd is on different PMD page table.
1743          *
1744          * We also don't deposit and withdraw tables for file pages.
1745          */
1746         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1747 }
1748 #endif
1749
1750 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1751 {
1752 #ifdef CONFIG_MEM_SOFT_DIRTY
1753         if (unlikely(is_pmd_migration_entry(pmd)))
1754                 pmd = pmd_swp_mksoft_dirty(pmd);
1755         else if (pmd_present(pmd))
1756                 pmd = pmd_mksoft_dirty(pmd);
1757 #endif
1758         return pmd;
1759 }
1760
1761 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1762                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1763 {
1764         spinlock_t *old_ptl, *new_ptl;
1765         pmd_t pmd;
1766         struct mm_struct *mm = vma->vm_mm;
1767         bool force_flush = false;
1768
1769         /*
1770          * The destination pmd shouldn't be established, free_pgtables()
1771          * should have release it.
1772          */
1773         if (WARN_ON(!pmd_none(*new_pmd))) {
1774                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1775                 return false;
1776         }
1777
1778         /*
1779          * We don't have to worry about the ordering of src and dst
1780          * ptlocks because exclusive mmap_lock prevents deadlock.
1781          */
1782         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1783         if (old_ptl) {
1784                 new_ptl = pmd_lockptr(mm, new_pmd);
1785                 if (new_ptl != old_ptl)
1786                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1787                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1788                 if (pmd_present(pmd))
1789                         force_flush = true;
1790                 VM_BUG_ON(!pmd_none(*new_pmd));
1791
1792                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1793                         pgtable_t pgtable;
1794                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1795                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1796                 }
1797                 pmd = move_soft_dirty_pmd(pmd);
1798                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1799                 if (force_flush)
1800                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1801                 if (new_ptl != old_ptl)
1802                         spin_unlock(new_ptl);
1803                 spin_unlock(old_ptl);
1804                 return true;
1805         }
1806         return false;
1807 }
1808
1809 /*
1810  * Returns
1811  *  - 0 if PMD could not be locked
1812  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1813  *      or if prot_numa but THP migration is not supported
1814  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1815  */
1816 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1817                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1818                     unsigned long cp_flags)
1819 {
1820         struct mm_struct *mm = vma->vm_mm;
1821         spinlock_t *ptl;
1822         pmd_t oldpmd, entry;
1823         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1824         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1825         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1826         int ret = 1;
1827
1828         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1829
1830         if (prot_numa && !thp_migration_supported())
1831                 return 1;
1832
1833         ptl = __pmd_trans_huge_lock(pmd, vma);
1834         if (!ptl)
1835                 return 0;
1836
1837 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1838         if (is_swap_pmd(*pmd)) {
1839                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1840                 struct page *page = pfn_swap_entry_to_page(entry);
1841
1842                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1843                 if (is_writable_migration_entry(entry)) {
1844                         pmd_t newpmd;
1845                         /*
1846                          * A protection check is difficult so
1847                          * just be safe and disable write
1848                          */
1849                         if (PageAnon(page))
1850                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1851                         else
1852                                 entry = make_readable_migration_entry(swp_offset(entry));
1853                         newpmd = swp_entry_to_pmd(entry);
1854                         if (pmd_swp_soft_dirty(*pmd))
1855                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1856                         if (pmd_swp_uffd_wp(*pmd))
1857                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1858                         set_pmd_at(mm, addr, pmd, newpmd);
1859                 }
1860                 goto unlock;
1861         }
1862 #endif
1863
1864         if (prot_numa) {
1865                 struct page *page;
1866                 bool toptier;
1867                 /*
1868                  * Avoid trapping faults against the zero page. The read-only
1869                  * data is likely to be read-cached on the local CPU and
1870                  * local/remote hits to the zero page are not interesting.
1871                  */
1872                 if (is_huge_zero_pmd(*pmd))
1873                         goto unlock;
1874
1875                 if (pmd_protnone(*pmd))
1876                         goto unlock;
1877
1878                 page = pmd_page(*pmd);
1879                 toptier = node_is_toptier(page_to_nid(page));
1880                 /*
1881                  * Skip scanning top tier node if normal numa
1882                  * balancing is disabled
1883                  */
1884                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1885                     toptier)
1886                         goto unlock;
1887
1888                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1889                     !toptier)
1890                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1891         }
1892         /*
1893          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1894          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895          * which is also under mmap_read_lock(mm):
1896          *
1897          *      CPU0:                           CPU1:
1898          *                              change_huge_pmd(prot_numa=1)
1899          *                               pmdp_huge_get_and_clear_notify()
1900          * madvise_dontneed()
1901          *  zap_pmd_range()
1902          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1903          *   // skip the pmd
1904          *                               set_pmd_at();
1905          *                               // pmd is re-established
1906          *
1907          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908          * which may break userspace.
1909          *
1910          * pmdp_invalidate_ad() is required to make sure we don't miss
1911          * dirty/young flags set by hardware.
1912          */
1913         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1914
1915         entry = pmd_modify(oldpmd, newprot);
1916         if (uffd_wp)
1917                 entry = pmd_mkuffd_wp(entry);
1918         else if (uffd_wp_resolve)
1919                 /*
1920                  * Leave the write bit to be handled by PF interrupt
1921                  * handler, then things like COW could be properly
1922                  * handled.
1923                  */
1924                 entry = pmd_clear_uffd_wp(entry);
1925
1926         /* See change_pte_range(). */
1927         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1928             can_change_pmd_writable(vma, addr, entry))
1929                 entry = pmd_mkwrite(entry);
1930
1931         ret = HPAGE_PMD_NR;
1932         set_pmd_at(mm, addr, pmd, entry);
1933
1934         if (huge_pmd_needs_flush(oldpmd, entry))
1935                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1936 unlock:
1937         spin_unlock(ptl);
1938         return ret;
1939 }
1940
1941 /*
1942  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1943  *
1944  * Note that if it returns page table lock pointer, this routine returns without
1945  * unlocking page table lock. So callers must unlock it.
1946  */
1947 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1948 {
1949         spinlock_t *ptl;
1950         ptl = pmd_lock(vma->vm_mm, pmd);
1951         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1952                         pmd_devmap(*pmd)))
1953                 return ptl;
1954         spin_unlock(ptl);
1955         return NULL;
1956 }
1957
1958 /*
1959  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1960  *
1961  * Note that if it returns page table lock pointer, this routine returns without
1962  * unlocking page table lock. So callers must unlock it.
1963  */
1964 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1965 {
1966         spinlock_t *ptl;
1967
1968         ptl = pud_lock(vma->vm_mm, pud);
1969         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1970                 return ptl;
1971         spin_unlock(ptl);
1972         return NULL;
1973 }
1974
1975 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1976 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1977                  pud_t *pud, unsigned long addr)
1978 {
1979         spinlock_t *ptl;
1980
1981         ptl = __pud_trans_huge_lock(pud, vma);
1982         if (!ptl)
1983                 return 0;
1984
1985         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1986         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1987         if (vma_is_special_huge(vma)) {
1988                 spin_unlock(ptl);
1989                 /* No zero page support yet */
1990         } else {
1991                 /* No support for anonymous PUD pages yet */
1992                 BUG();
1993         }
1994         return 1;
1995 }
1996
1997 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1998                 unsigned long haddr)
1999 {
2000         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2001         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2002         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2003         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2004
2005         count_vm_event(THP_SPLIT_PUD);
2006
2007         pudp_huge_clear_flush_notify(vma, haddr, pud);
2008 }
2009
2010 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2011                 unsigned long address)
2012 {
2013         spinlock_t *ptl;
2014         struct mmu_notifier_range range;
2015
2016         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2017                                 address & HPAGE_PUD_MASK,
2018                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2019         mmu_notifier_invalidate_range_start(&range);
2020         ptl = pud_lock(vma->vm_mm, pud);
2021         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2022                 goto out;
2023         __split_huge_pud_locked(vma, pud, range.start);
2024
2025 out:
2026         spin_unlock(ptl);
2027         /*
2028          * No need to double call mmu_notifier->invalidate_range() callback as
2029          * the above pudp_huge_clear_flush_notify() did already call it.
2030          */
2031         mmu_notifier_invalidate_range_only_end(&range);
2032 }
2033 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2034
2035 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2036                 unsigned long haddr, pmd_t *pmd)
2037 {
2038         struct mm_struct *mm = vma->vm_mm;
2039         pgtable_t pgtable;
2040         pmd_t _pmd;
2041         int i;
2042
2043         /*
2044          * Leave pmd empty until pte is filled note that it is fine to delay
2045          * notification until mmu_notifier_invalidate_range_end() as we are
2046          * replacing a zero pmd write protected page with a zero pte write
2047          * protected page.
2048          *
2049          * See Documentation/mm/mmu_notifier.rst
2050          */
2051         pmdp_huge_clear_flush(vma, haddr, pmd);
2052
2053         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2054         pmd_populate(mm, &_pmd, pgtable);
2055
2056         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2057                 pte_t *pte, entry;
2058                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2059                 entry = pte_mkspecial(entry);
2060                 pte = pte_offset_map(&_pmd, haddr);
2061                 VM_BUG_ON(!pte_none(*pte));
2062                 set_pte_at(mm, haddr, pte, entry);
2063                 pte_unmap(pte);
2064         }
2065         smp_wmb(); /* make pte visible before pmd */
2066         pmd_populate(mm, pmd, pgtable);
2067 }
2068
2069 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2070                 unsigned long haddr, bool freeze)
2071 {
2072         struct mm_struct *mm = vma->vm_mm;
2073         struct page *page;
2074         pgtable_t pgtable;
2075         pmd_t old_pmd, _pmd;
2076         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2077         bool anon_exclusive = false, dirty = false;
2078         unsigned long addr;
2079         int i;
2080
2081         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2082         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2083         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2084         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2085                                 && !pmd_devmap(*pmd));
2086
2087         count_vm_event(THP_SPLIT_PMD);
2088
2089         if (!vma_is_anonymous(vma)) {
2090                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2091                 /*
2092                  * We are going to unmap this huge page. So
2093                  * just go ahead and zap it
2094                  */
2095                 if (arch_needs_pgtable_deposit())
2096                         zap_deposited_table(mm, pmd);
2097                 if (vma_is_special_huge(vma))
2098                         return;
2099                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2100                         swp_entry_t entry;
2101
2102                         entry = pmd_to_swp_entry(old_pmd);
2103                         page = pfn_swap_entry_to_page(entry);
2104                 } else {
2105                         page = pmd_page(old_pmd);
2106                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2107                                 set_page_dirty(page);
2108                         if (!PageReferenced(page) && pmd_young(old_pmd))
2109                                 SetPageReferenced(page);
2110                         page_remove_rmap(page, vma, true);
2111                         put_page(page);
2112                 }
2113                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2114                 return;
2115         }
2116
2117         if (is_huge_zero_pmd(*pmd)) {
2118                 /*
2119                  * FIXME: Do we want to invalidate secondary mmu by calling
2120                  * mmu_notifier_invalidate_range() see comments below inside
2121                  * __split_huge_pmd() ?
2122                  *
2123                  * We are going from a zero huge page write protected to zero
2124                  * small page also write protected so it does not seems useful
2125                  * to invalidate secondary mmu at this time.
2126                  */
2127                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2128         }
2129
2130         /*
2131          * Up to this point the pmd is present and huge and userland has the
2132          * whole access to the hugepage during the split (which happens in
2133          * place). If we overwrite the pmd with the not-huge version pointing
2134          * to the pte here (which of course we could if all CPUs were bug
2135          * free), userland could trigger a small page size TLB miss on the
2136          * small sized TLB while the hugepage TLB entry is still established in
2137          * the huge TLB. Some CPU doesn't like that.
2138          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2139          * 383 on page 105. Intel should be safe but is also warns that it's
2140          * only safe if the permission and cache attributes of the two entries
2141          * loaded in the two TLB is identical (which should be the case here).
2142          * But it is generally safer to never allow small and huge TLB entries
2143          * for the same virtual address to be loaded simultaneously. So instead
2144          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2145          * current pmd notpresent (atomically because here the pmd_trans_huge
2146          * must remain set at all times on the pmd until the split is complete
2147          * for this pmd), then we flush the SMP TLB and finally we write the
2148          * non-huge version of the pmd entry with pmd_populate.
2149          */
2150         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2151
2152         pmd_migration = is_pmd_migration_entry(old_pmd);
2153         if (unlikely(pmd_migration)) {
2154                 swp_entry_t entry;
2155
2156                 entry = pmd_to_swp_entry(old_pmd);
2157                 page = pfn_swap_entry_to_page(entry);
2158                 write = is_writable_migration_entry(entry);
2159                 if (PageAnon(page))
2160                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2161                 young = is_migration_entry_young(entry);
2162                 dirty = is_migration_entry_dirty(entry);
2163                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2164                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2165         } else {
2166                 page = pmd_page(old_pmd);
2167                 if (pmd_dirty(old_pmd)) {
2168                         dirty = true;
2169                         SetPageDirty(page);
2170                 }
2171                 write = pmd_write(old_pmd);
2172                 young = pmd_young(old_pmd);
2173                 soft_dirty = pmd_soft_dirty(old_pmd);
2174                 uffd_wp = pmd_uffd_wp(old_pmd);
2175
2176                 VM_BUG_ON_PAGE(!page_count(page), page);
2177
2178                 /*
2179                  * Without "freeze", we'll simply split the PMD, propagating the
2180                  * PageAnonExclusive() flag for each PTE by setting it for
2181                  * each subpage -- no need to (temporarily) clear.
2182                  *
2183                  * With "freeze" we want to replace mapped pages by
2184                  * migration entries right away. This is only possible if we
2185                  * managed to clear PageAnonExclusive() -- see
2186                  * set_pmd_migration_entry().
2187                  *
2188                  * In case we cannot clear PageAnonExclusive(), split the PMD
2189                  * only and let try_to_migrate_one() fail later.
2190                  *
2191                  * See page_try_share_anon_rmap(): invalidate PMD first.
2192                  */
2193                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2194                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2195                         freeze = false;
2196                 if (!freeze)
2197                         page_ref_add(page, HPAGE_PMD_NR - 1);
2198         }
2199
2200         /*
2201          * Withdraw the table only after we mark the pmd entry invalid.
2202          * This's critical for some architectures (Power).
2203          */
2204         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2205         pmd_populate(mm, &_pmd, pgtable);
2206
2207         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2208                 pte_t entry, *pte;
2209                 /*
2210                  * Note that NUMA hinting access restrictions are not
2211                  * transferred to avoid any possibility of altering
2212                  * permissions across VMAs.
2213                  */
2214                 if (freeze || pmd_migration) {
2215                         swp_entry_t swp_entry;
2216                         if (write)
2217                                 swp_entry = make_writable_migration_entry(
2218                                                         page_to_pfn(page + i));
2219                         else if (anon_exclusive)
2220                                 swp_entry = make_readable_exclusive_migration_entry(
2221                                                         page_to_pfn(page + i));
2222                         else
2223                                 swp_entry = make_readable_migration_entry(
2224                                                         page_to_pfn(page + i));
2225                         if (young)
2226                                 swp_entry = make_migration_entry_young(swp_entry);
2227                         if (dirty)
2228                                 swp_entry = make_migration_entry_dirty(swp_entry);
2229                         entry = swp_entry_to_pte(swp_entry);
2230                         if (soft_dirty)
2231                                 entry = pte_swp_mksoft_dirty(entry);
2232                         if (uffd_wp)
2233                                 entry = pte_swp_mkuffd_wp(entry);
2234                 } else {
2235                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2236                         entry = maybe_mkwrite(entry, vma);
2237                         if (anon_exclusive)
2238                                 SetPageAnonExclusive(page + i);
2239                         if (!young)
2240                                 entry = pte_mkold(entry);
2241                         /* NOTE: this may set soft-dirty too on some archs */
2242                         if (dirty)
2243                                 entry = pte_mkdirty(entry);
2244                         /*
2245                          * NOTE: this needs to happen after pte_mkdirty,
2246                          * because some archs (sparc64, loongarch) could
2247                          * set hw write bit when mkdirty.
2248                          */
2249                         if (!write)
2250                                 entry = pte_wrprotect(entry);
2251                         if (soft_dirty)
2252                                 entry = pte_mksoft_dirty(entry);
2253                         if (uffd_wp)
2254                                 entry = pte_mkuffd_wp(entry);
2255                         page_add_anon_rmap(page + i, vma, addr, false);
2256                 }
2257                 pte = pte_offset_map(&_pmd, addr);
2258                 BUG_ON(!pte_none(*pte));
2259                 set_pte_at(mm, addr, pte, entry);
2260                 pte_unmap(pte);
2261         }
2262
2263         if (!pmd_migration)
2264                 page_remove_rmap(page, vma, true);
2265         if (freeze)
2266                 put_page(page);
2267
2268         smp_wmb(); /* make pte visible before pmd */
2269         pmd_populate(mm, pmd, pgtable);
2270 }
2271
2272 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2273                 unsigned long address, bool freeze, struct folio *folio)
2274 {
2275         spinlock_t *ptl;
2276         struct mmu_notifier_range range;
2277
2278         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2279                                 address & HPAGE_PMD_MASK,
2280                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2281         mmu_notifier_invalidate_range_start(&range);
2282         ptl = pmd_lock(vma->vm_mm, pmd);
2283
2284         /*
2285          * If caller asks to setup a migration entry, we need a folio to check
2286          * pmd against. Otherwise we can end up replacing wrong folio.
2287          */
2288         VM_BUG_ON(freeze && !folio);
2289         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2290
2291         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2292             is_pmd_migration_entry(*pmd)) {
2293                 /*
2294                  * It's safe to call pmd_page when folio is set because it's
2295                  * guaranteed that pmd is present.
2296                  */
2297                 if (folio && folio != page_folio(pmd_page(*pmd)))
2298                         goto out;
2299                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2300         }
2301
2302 out:
2303         spin_unlock(ptl);
2304         /*
2305          * No need to double call mmu_notifier->invalidate_range() callback.
2306          * They are 3 cases to consider inside __split_huge_pmd_locked():
2307          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2308          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2309          *    fault will trigger a flush_notify before pointing to a new page
2310          *    (it is fine if the secondary mmu keeps pointing to the old zero
2311          *    page in the meantime)
2312          *  3) Split a huge pmd into pte pointing to the same page. No need
2313          *     to invalidate secondary tlb entry they are all still valid.
2314          *     any further changes to individual pte will notify. So no need
2315          *     to call mmu_notifier->invalidate_range()
2316          */
2317         mmu_notifier_invalidate_range_only_end(&range);
2318 }
2319
2320 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2321                 bool freeze, struct folio *folio)
2322 {
2323         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2324
2325         if (!pmd)
2326                 return;
2327
2328         __split_huge_pmd(vma, pmd, address, freeze, folio);
2329 }
2330
2331 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2332 {
2333         /*
2334          * If the new address isn't hpage aligned and it could previously
2335          * contain an hugepage: check if we need to split an huge pmd.
2336          */
2337         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2338             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2339                          ALIGN(address, HPAGE_PMD_SIZE)))
2340                 split_huge_pmd_address(vma, address, false, NULL);
2341 }
2342
2343 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2344                              unsigned long start,
2345                              unsigned long end,
2346                              long adjust_next)
2347 {
2348         /* Check if we need to split start first. */
2349         split_huge_pmd_if_needed(vma, start);
2350
2351         /* Check if we need to split end next. */
2352         split_huge_pmd_if_needed(vma, end);
2353
2354         /*
2355          * If we're also updating the next vma vm_start,
2356          * check if we need to split it.
2357          */
2358         if (adjust_next > 0) {
2359                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2360                 unsigned long nstart = next->vm_start;
2361                 nstart += adjust_next;
2362                 split_huge_pmd_if_needed(next, nstart);
2363         }
2364 }
2365
2366 static void unmap_folio(struct folio *folio)
2367 {
2368         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2369                 TTU_SYNC;
2370
2371         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2372
2373         /*
2374          * Anon pages need migration entries to preserve them, but file
2375          * pages can simply be left unmapped, then faulted back on demand.
2376          * If that is ever changed (perhaps for mlock), update remap_page().
2377          */
2378         if (folio_test_anon(folio))
2379                 try_to_migrate(folio, ttu_flags);
2380         else
2381                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2382 }
2383
2384 static void remap_page(struct folio *folio, unsigned long nr)
2385 {
2386         int i = 0;
2387
2388         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2389         if (!folio_test_anon(folio))
2390                 return;
2391         for (;;) {
2392                 remove_migration_ptes(folio, folio, true);
2393                 i += folio_nr_pages(folio);
2394                 if (i >= nr)
2395                         break;
2396                 folio = folio_next(folio);
2397         }
2398 }
2399
2400 static void lru_add_page_tail(struct page *head, struct page *tail,
2401                 struct lruvec *lruvec, struct list_head *list)
2402 {
2403         VM_BUG_ON_PAGE(!PageHead(head), head);
2404         VM_BUG_ON_PAGE(PageCompound(tail), head);
2405         VM_BUG_ON_PAGE(PageLRU(tail), head);
2406         lockdep_assert_held(&lruvec->lru_lock);
2407
2408         if (list) {
2409                 /* page reclaim is reclaiming a huge page */
2410                 VM_WARN_ON(PageLRU(head));
2411                 get_page(tail);
2412                 list_add_tail(&tail->lru, list);
2413         } else {
2414                 /* head is still on lru (and we have it frozen) */
2415                 VM_WARN_ON(!PageLRU(head));
2416                 if (PageUnevictable(tail))
2417                         tail->mlock_count = 0;
2418                 else
2419                         list_add_tail(&tail->lru, &head->lru);
2420                 SetPageLRU(tail);
2421         }
2422 }
2423
2424 static void __split_huge_page_tail(struct page *head, int tail,
2425                 struct lruvec *lruvec, struct list_head *list)
2426 {
2427         struct page *page_tail = head + tail;
2428
2429         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2430
2431         /*
2432          * Clone page flags before unfreezing refcount.
2433          *
2434          * After successful get_page_unless_zero() might follow flags change,
2435          * for example lock_page() which set PG_waiters.
2436          *
2437          * Note that for mapped sub-pages of an anonymous THP,
2438          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2439          * the migration entry instead from where remap_page() will restore it.
2440          * We can still have PG_anon_exclusive set on effectively unmapped and
2441          * unreferenced sub-pages of an anonymous THP: we can simply drop
2442          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2443          */
2444         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2445         page_tail->flags |= (head->flags &
2446                         ((1L << PG_referenced) |
2447                          (1L << PG_swapbacked) |
2448                          (1L << PG_swapcache) |
2449                          (1L << PG_mlocked) |
2450                          (1L << PG_uptodate) |
2451                          (1L << PG_active) |
2452                          (1L << PG_workingset) |
2453                          (1L << PG_locked) |
2454                          (1L << PG_unevictable) |
2455 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2456                          (1L << PG_arch_2) |
2457                          (1L << PG_arch_3) |
2458 #endif
2459                          (1L << PG_dirty) |
2460                          LRU_GEN_MASK | LRU_REFS_MASK));
2461
2462         /* ->mapping in first and second tail page is replaced by other uses */
2463         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2464                         page_tail);
2465         page_tail->mapping = head->mapping;
2466         page_tail->index = head->index + tail;
2467
2468         /*
2469          * page->private should not be set in tail pages with the exception
2470          * of swap cache pages that store the swp_entry_t in tail pages.
2471          * Fix up and warn once if private is unexpectedly set.
2472          *
2473          * What of 32-bit systems, on which folio->_pincount overlays
2474          * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2475          * pincount must be 0 for folio_ref_freeze() to have succeeded.
2476          */
2477         if (!folio_test_swapcache(page_folio(head))) {
2478                 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2479                 page_tail->private = 0;
2480         }
2481
2482         /* Page flags must be visible before we make the page non-compound. */
2483         smp_wmb();
2484
2485         /*
2486          * Clear PageTail before unfreezing page refcount.
2487          *
2488          * After successful get_page_unless_zero() might follow put_page()
2489          * which needs correct compound_head().
2490          */
2491         clear_compound_head(page_tail);
2492
2493         /* Finally unfreeze refcount. Additional reference from page cache. */
2494         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2495                                           PageSwapCache(head)));
2496
2497         if (page_is_young(head))
2498                 set_page_young(page_tail);
2499         if (page_is_idle(head))
2500                 set_page_idle(page_tail);
2501
2502         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2503
2504         /*
2505          * always add to the tail because some iterators expect new
2506          * pages to show after the currently processed elements - e.g.
2507          * migrate_pages
2508          */
2509         lru_add_page_tail(head, page_tail, lruvec, list);
2510 }
2511
2512 static void __split_huge_page(struct page *page, struct list_head *list,
2513                 pgoff_t end)
2514 {
2515         struct folio *folio = page_folio(page);
2516         struct page *head = &folio->page;
2517         struct lruvec *lruvec;
2518         struct address_space *swap_cache = NULL;
2519         unsigned long offset = 0;
2520         unsigned int nr = thp_nr_pages(head);
2521         int i;
2522
2523         /* complete memcg works before add pages to LRU */
2524         split_page_memcg(head, nr);
2525
2526         if (PageAnon(head) && PageSwapCache(head)) {
2527                 swp_entry_t entry = { .val = page_private(head) };
2528
2529                 offset = swp_offset(entry);
2530                 swap_cache = swap_address_space(entry);
2531                 xa_lock(&swap_cache->i_pages);
2532         }
2533
2534         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2535         lruvec = folio_lruvec_lock(folio);
2536
2537         ClearPageHasHWPoisoned(head);
2538
2539         for (i = nr - 1; i >= 1; i--) {
2540                 __split_huge_page_tail(head, i, lruvec, list);
2541                 /* Some pages can be beyond EOF: drop them from page cache */
2542                 if (head[i].index >= end) {
2543                         struct folio *tail = page_folio(head + i);
2544
2545                         if (shmem_mapping(head->mapping))
2546                                 shmem_uncharge(head->mapping->host, 1);
2547                         else if (folio_test_clear_dirty(tail))
2548                                 folio_account_cleaned(tail,
2549                                         inode_to_wb(folio->mapping->host));
2550                         __filemap_remove_folio(tail, NULL);
2551                         folio_put(tail);
2552                 } else if (!PageAnon(page)) {
2553                         __xa_store(&head->mapping->i_pages, head[i].index,
2554                                         head + i, 0);
2555                 } else if (swap_cache) {
2556                         __xa_store(&swap_cache->i_pages, offset + i,
2557                                         head + i, 0);
2558                 }
2559         }
2560
2561         ClearPageCompound(head);
2562         unlock_page_lruvec(lruvec);
2563         /* Caller disabled irqs, so they are still disabled here */
2564
2565         split_page_owner(head, nr);
2566
2567         /* See comment in __split_huge_page_tail() */
2568         if (PageAnon(head)) {
2569                 /* Additional pin to swap cache */
2570                 if (PageSwapCache(head)) {
2571                         page_ref_add(head, 2);
2572                         xa_unlock(&swap_cache->i_pages);
2573                 } else {
2574                         page_ref_inc(head);
2575                 }
2576         } else {
2577                 /* Additional pin to page cache */
2578                 page_ref_add(head, 2);
2579                 xa_unlock(&head->mapping->i_pages);
2580         }
2581         local_irq_enable();
2582
2583         remap_page(folio, nr);
2584
2585         if (PageSwapCache(head)) {
2586                 swp_entry_t entry = { .val = page_private(head) };
2587
2588                 split_swap_cluster(entry);
2589         }
2590
2591         for (i = 0; i < nr; i++) {
2592                 struct page *subpage = head + i;
2593                 if (subpage == page)
2594                         continue;
2595                 unlock_page(subpage);
2596
2597                 /*
2598                  * Subpages may be freed if there wasn't any mapping
2599                  * like if add_to_swap() is running on a lru page that
2600                  * had its mapping zapped. And freeing these pages
2601                  * requires taking the lru_lock so we do the put_page
2602                  * of the tail pages after the split is complete.
2603                  */
2604                 free_page_and_swap_cache(subpage);
2605         }
2606 }
2607
2608 /* Racy check whether the huge page can be split */
2609 bool can_split_folio(struct folio *folio, int *pextra_pins)
2610 {
2611         int extra_pins;
2612
2613         /* Additional pins from page cache */
2614         if (folio_test_anon(folio))
2615                 extra_pins = folio_test_swapcache(folio) ?
2616                                 folio_nr_pages(folio) : 0;
2617         else
2618                 extra_pins = folio_nr_pages(folio);
2619         if (pextra_pins)
2620                 *pextra_pins = extra_pins;
2621         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2622 }
2623
2624 /*
2625  * This function splits huge page into normal pages. @page can point to any
2626  * subpage of huge page to split. Split doesn't change the position of @page.
2627  *
2628  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2629  * The huge page must be locked.
2630  *
2631  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2632  *
2633  * Both head page and tail pages will inherit mapping, flags, and so on from
2634  * the hugepage.
2635  *
2636  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2637  * they are not mapped.
2638  *
2639  * Returns 0 if the hugepage is split successfully.
2640  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2641  * us.
2642  */
2643 int split_huge_page_to_list(struct page *page, struct list_head *list)
2644 {
2645         struct folio *folio = page_folio(page);
2646         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2647         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2648         struct anon_vma *anon_vma = NULL;
2649         struct address_space *mapping = NULL;
2650         int extra_pins, ret;
2651         pgoff_t end;
2652         bool is_hzp;
2653
2654         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2655         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2656
2657         is_hzp = is_huge_zero_page(&folio->page);
2658         VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2659         if (is_hzp)
2660                 return -EBUSY;
2661
2662         if (folio_test_writeback(folio))
2663                 return -EBUSY;
2664
2665         if (folio_test_anon(folio)) {
2666                 /*
2667                  * The caller does not necessarily hold an mmap_lock that would
2668                  * prevent the anon_vma disappearing so we first we take a
2669                  * reference to it and then lock the anon_vma for write. This
2670                  * is similar to folio_lock_anon_vma_read except the write lock
2671                  * is taken to serialise against parallel split or collapse
2672                  * operations.
2673                  */
2674                 anon_vma = folio_get_anon_vma(folio);
2675                 if (!anon_vma) {
2676                         ret = -EBUSY;
2677                         goto out;
2678                 }
2679                 end = -1;
2680                 mapping = NULL;
2681                 anon_vma_lock_write(anon_vma);
2682         } else {
2683                 gfp_t gfp;
2684
2685                 mapping = folio->mapping;
2686
2687                 /* Truncated ? */
2688                 if (!mapping) {
2689                         ret = -EBUSY;
2690                         goto out;
2691                 }
2692
2693                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2694                                                         GFP_RECLAIM_MASK);
2695
2696                 if (folio_test_private(folio) &&
2697                                 !filemap_release_folio(folio, gfp)) {
2698                         ret = -EBUSY;
2699                         goto out;
2700                 }
2701
2702                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2703                 if (xas_error(&xas)) {
2704                         ret = xas_error(&xas);
2705                         goto out;
2706                 }
2707
2708                 anon_vma = NULL;
2709                 i_mmap_lock_read(mapping);
2710
2711                 /*
2712                  *__split_huge_page() may need to trim off pages beyond EOF:
2713                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2714                  * which cannot be nested inside the page tree lock. So note
2715                  * end now: i_size itself may be changed at any moment, but
2716                  * folio lock is good enough to serialize the trimming.
2717                  */
2718                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2719                 if (shmem_mapping(mapping))
2720                         end = shmem_fallocend(mapping->host, end);
2721         }
2722
2723         /*
2724          * Racy check if we can split the page, before unmap_folio() will
2725          * split PMDs
2726          */
2727         if (!can_split_folio(folio, &extra_pins)) {
2728                 ret = -EAGAIN;
2729                 goto out_unlock;
2730         }
2731
2732         unmap_folio(folio);
2733
2734         /* block interrupt reentry in xa_lock and spinlock */
2735         local_irq_disable();
2736         if (mapping) {
2737                 /*
2738                  * Check if the folio is present in page cache.
2739                  * We assume all tail are present too, if folio is there.
2740                  */
2741                 xas_lock(&xas);
2742                 xas_reset(&xas);
2743                 if (xas_load(&xas) != folio)
2744                         goto fail;
2745         }
2746
2747         /* Prevent deferred_split_scan() touching ->_refcount */
2748         spin_lock(&ds_queue->split_queue_lock);
2749         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2750                 if (!list_empty(&folio->_deferred_list)) {
2751                         ds_queue->split_queue_len--;
2752                         list_del(&folio->_deferred_list);
2753                 }
2754                 spin_unlock(&ds_queue->split_queue_lock);
2755                 if (mapping) {
2756                         int nr = folio_nr_pages(folio);
2757
2758                         xas_split(&xas, folio, folio_order(folio));
2759                         if (folio_test_swapbacked(folio)) {
2760                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2761                                                         -nr);
2762                         } else {
2763                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2764                                                         -nr);
2765                                 filemap_nr_thps_dec(mapping);
2766                         }
2767                 }
2768
2769                 __split_huge_page(page, list, end);
2770                 ret = 0;
2771         } else {
2772                 spin_unlock(&ds_queue->split_queue_lock);
2773 fail:
2774                 if (mapping)
2775                         xas_unlock(&xas);
2776                 local_irq_enable();
2777                 remap_page(folio, folio_nr_pages(folio));
2778                 ret = -EAGAIN;
2779         }
2780
2781 out_unlock:
2782         if (anon_vma) {
2783                 anon_vma_unlock_write(anon_vma);
2784                 put_anon_vma(anon_vma);
2785         }
2786         if (mapping)
2787                 i_mmap_unlock_read(mapping);
2788 out:
2789         xas_destroy(&xas);
2790         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2791         return ret;
2792 }
2793
2794 void free_transhuge_page(struct page *page)
2795 {
2796         struct folio *folio = (struct folio *)page;
2797         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2798         unsigned long flags;
2799
2800         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2801         if (!list_empty(&folio->_deferred_list)) {
2802                 ds_queue->split_queue_len--;
2803                 list_del(&folio->_deferred_list);
2804         }
2805         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2806         free_compound_page(page);
2807 }
2808
2809 void deferred_split_folio(struct folio *folio)
2810 {
2811         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2812 #ifdef CONFIG_MEMCG
2813         struct mem_cgroup *memcg = folio_memcg(folio);
2814 #endif
2815         unsigned long flags;
2816
2817         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2818
2819         /*
2820          * The try_to_unmap() in page reclaim path might reach here too,
2821          * this may cause a race condition to corrupt deferred split queue.
2822          * And, if page reclaim is already handling the same folio, it is
2823          * unnecessary to handle it again in shrinker.
2824          *
2825          * Check the swapcache flag to determine if the folio is being
2826          * handled by page reclaim since THP swap would add the folio into
2827          * swap cache before calling try_to_unmap().
2828          */
2829         if (folio_test_swapcache(folio))
2830                 return;
2831
2832         if (!list_empty(&folio->_deferred_list))
2833                 return;
2834
2835         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2836         if (list_empty(&folio->_deferred_list)) {
2837                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2838                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2839                 ds_queue->split_queue_len++;
2840 #ifdef CONFIG_MEMCG
2841                 if (memcg)
2842                         set_shrinker_bit(memcg, folio_nid(folio),
2843                                          deferred_split_shrinker.id);
2844 #endif
2845         }
2846         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2847 }
2848
2849 static unsigned long deferred_split_count(struct shrinker *shrink,
2850                 struct shrink_control *sc)
2851 {
2852         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2853         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2854
2855 #ifdef CONFIG_MEMCG
2856         if (sc->memcg)
2857                 ds_queue = &sc->memcg->deferred_split_queue;
2858 #endif
2859         return READ_ONCE(ds_queue->split_queue_len);
2860 }
2861
2862 static unsigned long deferred_split_scan(struct shrinker *shrink,
2863                 struct shrink_control *sc)
2864 {
2865         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2866         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2867         unsigned long flags;
2868         LIST_HEAD(list);
2869         struct folio *folio, *next;
2870         int split = 0;
2871
2872 #ifdef CONFIG_MEMCG
2873         if (sc->memcg)
2874                 ds_queue = &sc->memcg->deferred_split_queue;
2875 #endif
2876
2877         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2878         /* Take pin on all head pages to avoid freeing them under us */
2879         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2880                                                         _deferred_list) {
2881                 if (folio_try_get(folio)) {
2882                         list_move(&folio->_deferred_list, &list);
2883                 } else {
2884                         /* We lost race with folio_put() */
2885                         list_del_init(&folio->_deferred_list);
2886                         ds_queue->split_queue_len--;
2887                 }
2888                 if (!--sc->nr_to_scan)
2889                         break;
2890         }
2891         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2892
2893         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2894                 if (!folio_trylock(folio))
2895                         goto next;
2896                 /* split_huge_page() removes page from list on success */
2897                 if (!split_folio(folio))
2898                         split++;
2899                 folio_unlock(folio);
2900 next:
2901                 folio_put(folio);
2902         }
2903
2904         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2905         list_splice_tail(&list, &ds_queue->split_queue);
2906         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2907
2908         /*
2909          * Stop shrinker if we didn't split any page, but the queue is empty.
2910          * This can happen if pages were freed under us.
2911          */
2912         if (!split && list_empty(&ds_queue->split_queue))
2913                 return SHRINK_STOP;
2914         return split;
2915 }
2916
2917 static struct shrinker deferred_split_shrinker = {
2918         .count_objects = deferred_split_count,
2919         .scan_objects = deferred_split_scan,
2920         .seeks = DEFAULT_SEEKS,
2921         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2922                  SHRINKER_NONSLAB,
2923 };
2924
2925 #ifdef CONFIG_DEBUG_FS
2926 static void split_huge_pages_all(void)
2927 {
2928         struct zone *zone;
2929         struct page *page;
2930         struct folio *folio;
2931         unsigned long pfn, max_zone_pfn;
2932         unsigned long total = 0, split = 0;
2933
2934         pr_debug("Split all THPs\n");
2935         for_each_zone(zone) {
2936                 if (!managed_zone(zone))
2937                         continue;
2938                 max_zone_pfn = zone_end_pfn(zone);
2939                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2940                         int nr_pages;
2941
2942                         page = pfn_to_online_page(pfn);
2943                         if (!page || PageTail(page))
2944                                 continue;
2945                         folio = page_folio(page);
2946                         if (!folio_try_get(folio))
2947                                 continue;
2948
2949                         if (unlikely(page_folio(page) != folio))
2950                                 goto next;
2951
2952                         if (zone != folio_zone(folio))
2953                                 goto next;
2954
2955                         if (!folio_test_large(folio)
2956                                 || folio_test_hugetlb(folio)
2957                                 || !folio_test_lru(folio))
2958                                 goto next;
2959
2960                         total++;
2961                         folio_lock(folio);
2962                         nr_pages = folio_nr_pages(folio);
2963                         if (!split_folio(folio))
2964                                 split++;
2965                         pfn += nr_pages - 1;
2966                         folio_unlock(folio);
2967 next:
2968                         folio_put(folio);
2969                         cond_resched();
2970                 }
2971         }
2972
2973         pr_debug("%lu of %lu THP split\n", split, total);
2974 }
2975
2976 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2977 {
2978         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2979                     is_vm_hugetlb_page(vma);
2980 }
2981
2982 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2983                                 unsigned long vaddr_end)
2984 {
2985         int ret = 0;
2986         struct task_struct *task;
2987         struct mm_struct *mm;
2988         unsigned long total = 0, split = 0;
2989         unsigned long addr;
2990
2991         vaddr_start &= PAGE_MASK;
2992         vaddr_end &= PAGE_MASK;
2993
2994         /* Find the task_struct from pid */
2995         rcu_read_lock();
2996         task = find_task_by_vpid(pid);
2997         if (!task) {
2998                 rcu_read_unlock();
2999                 ret = -ESRCH;
3000                 goto out;
3001         }
3002         get_task_struct(task);
3003         rcu_read_unlock();
3004
3005         /* Find the mm_struct */
3006         mm = get_task_mm(task);
3007         put_task_struct(task);
3008
3009         if (!mm) {
3010                 ret = -EINVAL;
3011                 goto out;
3012         }
3013
3014         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3015                  pid, vaddr_start, vaddr_end);
3016
3017         mmap_read_lock(mm);
3018         /*
3019          * always increase addr by PAGE_SIZE, since we could have a PTE page
3020          * table filled with PTE-mapped THPs, each of which is distinct.
3021          */
3022         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3023                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3024                 struct page *page;
3025
3026                 if (!vma)
3027                         break;
3028
3029                 /* skip special VMA and hugetlb VMA */
3030                 if (vma_not_suitable_for_thp_split(vma)) {
3031                         addr = vma->vm_end;
3032                         continue;
3033                 }
3034
3035                 /* FOLL_DUMP to ignore special (like zero) pages */
3036                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3037
3038                 if (IS_ERR_OR_NULL(page))
3039                         continue;
3040
3041                 if (!is_transparent_hugepage(page))
3042                         goto next;
3043
3044                 total++;
3045                 if (!can_split_folio(page_folio(page), NULL))
3046                         goto next;
3047
3048                 if (!trylock_page(page))
3049                         goto next;
3050
3051                 if (!split_huge_page(page))
3052                         split++;
3053
3054                 unlock_page(page);
3055 next:
3056                 put_page(page);
3057                 cond_resched();
3058         }
3059         mmap_read_unlock(mm);
3060         mmput(mm);
3061
3062         pr_debug("%lu of %lu THP split\n", split, total);
3063
3064 out:
3065         return ret;
3066 }
3067
3068 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3069                                 pgoff_t off_end)
3070 {
3071         struct filename *file;
3072         struct file *candidate;
3073         struct address_space *mapping;
3074         int ret = -EINVAL;
3075         pgoff_t index;
3076         int nr_pages = 1;
3077         unsigned long total = 0, split = 0;
3078
3079         file = getname_kernel(file_path);
3080         if (IS_ERR(file))
3081                 return ret;
3082
3083         candidate = file_open_name(file, O_RDONLY, 0);
3084         if (IS_ERR(candidate))
3085                 goto out;
3086
3087         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3088                  file_path, off_start, off_end);
3089
3090         mapping = candidate->f_mapping;
3091
3092         for (index = off_start; index < off_end; index += nr_pages) {
3093                 struct folio *folio = __filemap_get_folio(mapping, index,
3094                                                 FGP_ENTRY, 0);
3095
3096                 nr_pages = 1;
3097                 if (xa_is_value(folio) || !folio)
3098                         continue;
3099
3100                 if (!folio_test_large(folio))
3101                         goto next;
3102
3103                 total++;
3104                 nr_pages = folio_nr_pages(folio);
3105
3106                 if (!folio_trylock(folio))
3107                         goto next;
3108
3109                 if (!split_folio(folio))
3110                         split++;
3111
3112                 folio_unlock(folio);
3113 next:
3114                 folio_put(folio);
3115                 cond_resched();
3116         }
3117
3118         filp_close(candidate, NULL);
3119         ret = 0;
3120
3121         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3122 out:
3123         putname(file);
3124         return ret;
3125 }
3126
3127 #define MAX_INPUT_BUF_SZ 255
3128
3129 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3130                                 size_t count, loff_t *ppops)
3131 {
3132         static DEFINE_MUTEX(split_debug_mutex);
3133         ssize_t ret;
3134         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3135         char input_buf[MAX_INPUT_BUF_SZ];
3136         int pid;
3137         unsigned long vaddr_start, vaddr_end;
3138
3139         ret = mutex_lock_interruptible(&split_debug_mutex);
3140         if (ret)
3141                 return ret;
3142
3143         ret = -EFAULT;
3144
3145         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3146         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3147                 goto out;
3148
3149         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3150
3151         if (input_buf[0] == '/') {
3152                 char *tok;
3153                 char *buf = input_buf;
3154                 char file_path[MAX_INPUT_BUF_SZ];
3155                 pgoff_t off_start = 0, off_end = 0;
3156                 size_t input_len = strlen(input_buf);
3157
3158                 tok = strsep(&buf, ",");
3159                 if (tok) {
3160                         strcpy(file_path, tok);
3161                 } else {
3162                         ret = -EINVAL;
3163                         goto out;
3164                 }
3165
3166                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3167                 if (ret != 2) {
3168                         ret = -EINVAL;
3169                         goto out;
3170                 }
3171                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3172                 if (!ret)
3173                         ret = input_len;
3174
3175                 goto out;
3176         }
3177
3178         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3179         if (ret == 1 && pid == 1) {
3180                 split_huge_pages_all();
3181                 ret = strlen(input_buf);
3182                 goto out;
3183         } else if (ret != 3) {
3184                 ret = -EINVAL;
3185                 goto out;
3186         }
3187
3188         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3189         if (!ret)
3190                 ret = strlen(input_buf);
3191 out:
3192         mutex_unlock(&split_debug_mutex);
3193         return ret;
3194
3195 }
3196
3197 static const struct file_operations split_huge_pages_fops = {
3198         .owner   = THIS_MODULE,
3199         .write   = split_huge_pages_write,
3200         .llseek  = no_llseek,
3201 };
3202
3203 static int __init split_huge_pages_debugfs(void)
3204 {
3205         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3206                             &split_huge_pages_fops);
3207         return 0;
3208 }
3209 late_initcall(split_huge_pages_debugfs);
3210 #endif
3211
3212 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3213 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3214                 struct page *page)
3215 {
3216         struct vm_area_struct *vma = pvmw->vma;
3217         struct mm_struct *mm = vma->vm_mm;
3218         unsigned long address = pvmw->address;
3219         bool anon_exclusive;
3220         pmd_t pmdval;
3221         swp_entry_t entry;
3222         pmd_t pmdswp;
3223
3224         if (!(pvmw->pmd && !pvmw->pte))
3225                 return 0;
3226
3227         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3228         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3229
3230         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3231         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3232         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3233                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3234                 return -EBUSY;
3235         }
3236
3237         if (pmd_dirty(pmdval))
3238                 set_page_dirty(page);
3239         if (pmd_write(pmdval))
3240                 entry = make_writable_migration_entry(page_to_pfn(page));
3241         else if (anon_exclusive)
3242                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3243         else
3244                 entry = make_readable_migration_entry(page_to_pfn(page));
3245         if (pmd_young(pmdval))
3246                 entry = make_migration_entry_young(entry);
3247         if (pmd_dirty(pmdval))
3248                 entry = make_migration_entry_dirty(entry);
3249         pmdswp = swp_entry_to_pmd(entry);
3250         if (pmd_soft_dirty(pmdval))
3251                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3252         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3253         page_remove_rmap(page, vma, true);
3254         put_page(page);
3255         trace_set_migration_pmd(address, pmd_val(pmdswp));
3256
3257         return 0;
3258 }
3259
3260 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3261 {
3262         struct vm_area_struct *vma = pvmw->vma;
3263         struct mm_struct *mm = vma->vm_mm;
3264         unsigned long address = pvmw->address;
3265         unsigned long haddr = address & HPAGE_PMD_MASK;
3266         pmd_t pmde;
3267         swp_entry_t entry;
3268
3269         if (!(pvmw->pmd && !pvmw->pte))
3270                 return;
3271
3272         entry = pmd_to_swp_entry(*pvmw->pmd);
3273         get_page(new);
3274         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3275         if (pmd_swp_soft_dirty(*pvmw->pmd))
3276                 pmde = pmd_mksoft_dirty(pmde);
3277         if (pmd_swp_uffd_wp(*pvmw->pmd))
3278                 pmde = pmd_mkuffd_wp(pmde);
3279         if (!is_migration_entry_young(entry))
3280                 pmde = pmd_mkold(pmde);
3281         /* NOTE: this may contain setting soft-dirty on some archs */
3282         if (PageDirty(new) && is_migration_entry_dirty(entry))
3283                 pmde = pmd_mkdirty(pmde);
3284         if (is_writable_migration_entry(entry))
3285                 pmde = maybe_pmd_mkwrite(pmde, vma);
3286         else
3287                 pmde = pmd_wrprotect(pmde);
3288
3289         if (PageAnon(new)) {
3290                 rmap_t rmap_flags = RMAP_COMPOUND;
3291
3292                 if (!is_readable_migration_entry(entry))
3293                         rmap_flags |= RMAP_EXCLUSIVE;
3294
3295                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3296         } else {
3297                 page_add_file_rmap(new, vma, true);
3298         }
3299         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3300         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3301
3302         /* No need to invalidate - it was non-present before */
3303         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3304         trace_remove_migration_pmd(address, pmd_val(pmde));
3305 }
3306 #endif