mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/backing-dev.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37 #include <linux/page_owner.h>
  38 #include <linux/sched/sysctl.h>
  39 #include <linux/memory-tiers.h>
  40
  41 #include <asm/tlb.h>
  42 #include <asm/pgalloc.h>
  43 #include "internal.h"
  44 #include "swap.h"
  45
  46 #define CREATE_TRACE_POINTS
  47 #include <trace/events/thp.h>
  48
  49 /*
  50  * By default, transparent hugepage support is disabled in order to avoid
  51  * risking an increased memory footprint for applications that are not
  52  * guaranteed to benefit from it. When transparent hugepage support is
  53  * enabled, it is for all mappings, and khugepaged scans all mappings.
  54  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  55  * for all hugepage allocations.
  56  */
  57 unsigned long transparent_hugepage_flags __read_mostly =
  58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  59         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  60 #endif
  61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  62         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  63 #endif
  64         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  65         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  66         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  67
  68 static struct shrinker deferred_split_shrinker;
  69
  70 static atomic_t huge_zero_refcount;
  71 struct page *huge_zero_page __read_mostly;
  72 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  73
  74 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
  75                         bool smaps, bool in_pf, bool enforce_sysfs)
  76 {
  77         if (!vma->vm_mm)                /* vdso */
  78                 return false;
  79
  80         /*
  81          * Explicitly disabled through madvise or prctl, or some
  82          * architectures may disable THP for some mappings, for
  83          * example, s390 kvm.
  84          * */
  85         if ((vm_flags & VM_NOHUGEPAGE) ||
  86             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  87                 return false;
  88         /*
  89          * If the hardware/firmware marked hugepage support disabled.
  90          */
  91         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
  92                 return false;
  93
  94         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  95         if (vma_is_dax(vma))
  96                 return in_pf;
  97
  98         /*
  99          * Special VMA and hugetlb VMA.
 100          * Must be checked after dax since some dax mappings may have
 101          * VM_MIXEDMAP set.
 102          */
 103         if (vm_flags & VM_NO_KHUGEPAGED)
 104                 return false;
 105
 106         /*
 107          * Check alignment for file vma and size for both file and anon vma.
 108          *
 109          * Skip the check for page fault. Huge fault does the check in fault
 110          * handlers. And this check is not suitable for huge PUD fault.
 111          */
 112         if (!in_pf &&
 113             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 114                 return false;
 115
 116         /*
 117          * Enabled via shmem mount options or sysfs settings.
 118          * Must be done before hugepage flags check since shmem has its
 119          * own flags.
 120          */
 121         if (!in_pf && shmem_file(vma->vm_file))
 122                 return shmem_huge_enabled(vma, !enforce_sysfs);
 123
 124         /* Enforce sysfs THP requirements as necessary */
 125         if (enforce_sysfs &&
 126             (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 127                                            !hugepage_flags_always())))
 128                 return false;
 129
 130         /* Only regular file is valid */
 131         if (!in_pf && file_thp_enabled(vma))
 132                 return true;
 133
 134         if (!vma_is_anonymous(vma))
 135                 return false;
 136
 137         if (vma_is_temporary_stack(vma))
 138                 return false;
 139
 140         /*
 141          * THPeligible bit of smaps should show 1 for proper VMAs even
 142          * though anon_vma is not initialized yet.
 143          *
 144          * Allow page fault since anon_vma may be not initialized until
 145          * the first page fault.
 146          */
 147         if (!vma->anon_vma)
 148                 return (smaps || in_pf);
 149
 150         return true;
 151 }
 152
 153 static bool get_huge_zero_page(void)
 154 {
 155         struct page *zero_page;
 156 retry:
 157         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 158                 return true;
 159
 160         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 161                         HPAGE_PMD_ORDER);
 162         if (!zero_page) {
 163                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 164                 return false;
 165         }
 166         preempt_disable();
 167         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 168                 preempt_enable();
 169                 __free_pages(zero_page, compound_order(zero_page));
 170                 goto retry;
 171         }
 172         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 173
 174         /* We take additional reference here. It will be put back by shrinker */
 175         atomic_set(&huge_zero_refcount, 2);
 176         preempt_enable();
 177         count_vm_event(THP_ZERO_PAGE_ALLOC);
 178         return true;
 179 }
 180
 181 static void put_huge_zero_page(void)
 182 {
 183         /*
 184          * Counter should never go to zero here. Only shrinker can put
 185          * last reference.
 186          */
 187         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 188 }
 189
 190 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 191 {
 192         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 193                 return READ_ONCE(huge_zero_page);
 194
 195         if (!get_huge_zero_page())
 196                 return NULL;
 197
 198         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 199                 put_huge_zero_page();
 200
 201         return READ_ONCE(huge_zero_page);
 202 }
 203
 204 void mm_put_huge_zero_page(struct mm_struct *mm)
 205 {
 206         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 207                 put_huge_zero_page();
 208 }
 209
 210 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 211                                         struct shrink_control *sc)
 212 {
 213         /* we can free zero page only if last reference remains */
 214         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 215 }
 216
 217 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 218                                        struct shrink_control *sc)
 219 {
 220         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 221                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 222                 BUG_ON(zero_page == NULL);
 223                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 224                 __free_pages(zero_page, compound_order(zero_page));
 225                 return HPAGE_PMD_NR;
 226         }
 227
 228         return 0;
 229 }
 230
 231 static struct shrinker huge_zero_page_shrinker = {
 232         .count_objects = shrink_huge_zero_page_count,
 233         .scan_objects = shrink_huge_zero_page_scan,
 234         .seeks = DEFAULT_SEEKS,
 235 };
 236
 237 #ifdef CONFIG_SYSFS
 238 static ssize_t enabled_show(struct kobject *kobj,
 239                             struct kobj_attribute *attr, char *buf)
 240 {
 241         const char *output;
 242
 243         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 244                 output = "[always] madvise never";
 245         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 246                           &transparent_hugepage_flags))
 247                 output = "always [madvise] never";
 248         else
 249                 output = "always madvise [never]";
 250
 251         return sysfs_emit(buf, "%s\n", output);
 252 }
 253
 254 static ssize_t enabled_store(struct kobject *kobj,
 255                              struct kobj_attribute *attr,
 256                              const char *buf, size_t count)
 257 {
 258         ssize_t ret = count;
 259
 260         if (sysfs_streq(buf, "always")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 263         } else if (sysfs_streq(buf, "madvise")) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 265                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 266         } else if (sysfs_streq(buf, "never")) {
 267                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 269         } else
 270                 ret = -EINVAL;
 271
 272         if (ret > 0) {
 273                 int err = start_stop_khugepaged();
 274                 if (err)
 275                         ret = err;
 276         }
 277         return ret;
 278 }
 279
 280 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 281
 282 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 283                                   struct kobj_attribute *attr, char *buf,
 284                                   enum transparent_hugepage_flag flag)
 285 {
 286         return sysfs_emit(buf, "%d\n",
 287                           !!test_bit(flag, &transparent_hugepage_flags));
 288 }
 289
 290 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 291                                  struct kobj_attribute *attr,
 292                                  const char *buf, size_t count,
 293                                  enum transparent_hugepage_flag flag)
 294 {
 295         unsigned long value;
 296         int ret;
 297
 298         ret = kstrtoul(buf, 10, &value);
 299         if (ret < 0)
 300                 return ret;
 301         if (value > 1)
 302                 return -EINVAL;
 303
 304         if (value)
 305                 set_bit(flag, &transparent_hugepage_flags);
 306         else
 307                 clear_bit(flag, &transparent_hugepage_flags);
 308
 309         return count;
 310 }
 311
 312 static ssize_t defrag_show(struct kobject *kobj,
 313                            struct kobj_attribute *attr, char *buf)
 314 {
 315         const char *output;
 316
 317         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 318                      &transparent_hugepage_flags))
 319                 output = "[always] defer defer+madvise madvise never";
 320         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 321                           &transparent_hugepage_flags))
 322                 output = "always [defer] defer+madvise madvise never";
 323         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 324                           &transparent_hugepage_flags))
 325                 output = "always defer [defer+madvise] madvise never";
 326         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 327                           &transparent_hugepage_flags))
 328                 output = "always defer defer+madvise [madvise] never";
 329         else
 330                 output = "always defer defer+madvise madvise [never]";
 331
 332         return sysfs_emit(buf, "%s\n", output);
 333 }
 334
 335 static ssize_t defrag_store(struct kobject *kobj,
 336                             struct kobj_attribute *attr,
 337                             const char *buf, size_t count)
 338 {
 339         if (sysfs_streq(buf, "always")) {
 340                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 341                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 342                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 343                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 344         } else if (sysfs_streq(buf, "defer+madvise")) {
 345                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 346                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 347                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 348                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 349         } else if (sysfs_streq(buf, "defer")) {
 350                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 351                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 352                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 353                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 354         } else if (sysfs_streq(buf, "madvise")) {
 355                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 356                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 357                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 358                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 359         } else if (sysfs_streq(buf, "never")) {
 360                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 361                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 362                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 363                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 364         } else
 365                 return -EINVAL;
 366
 367         return count;
 368 }
 369 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 370
 371 static ssize_t use_zero_page_show(struct kobject *kobj,
 372                                   struct kobj_attribute *attr, char *buf)
 373 {
 374         return single_hugepage_flag_show(kobj, attr, buf,
 375                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 376 }
 377 static ssize_t use_zero_page_store(struct kobject *kobj,
 378                 struct kobj_attribute *attr, const char *buf, size_t count)
 379 {
 380         return single_hugepage_flag_store(kobj, attr, buf, count,
 381                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 382 }
 383 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 384
 385 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 386                                    struct kobj_attribute *attr, char *buf)
 387 {
 388         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 389 }
 390 static struct kobj_attribute hpage_pmd_size_attr =
 391         __ATTR_RO(hpage_pmd_size);
 392
 393 static struct attribute *hugepage_attr[] = {
 394         &enabled_attr.attr,
 395         &defrag_attr.attr,
 396         &use_zero_page_attr.attr,
 397         &hpage_pmd_size_attr.attr,
 398 #ifdef CONFIG_SHMEM
 399         &shmem_enabled_attr.attr,
 400 #endif
 401         NULL,
 402 };
 403
 404 static const struct attribute_group hugepage_attr_group = {
 405         .attrs = hugepage_attr,
 406 };
 407
 408 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 409 {
 410         int err;
 411
 412         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 413         if (unlikely(!*hugepage_kobj)) {
 414                 pr_err("failed to create transparent hugepage kobject\n");
 415                 return -ENOMEM;
 416         }
 417
 418         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 419         if (err) {
 420                 pr_err("failed to register transparent hugepage group\n");
 421                 goto delete_obj;
 422         }
 423
 424         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 425         if (err) {
 426                 pr_err("failed to register transparent hugepage group\n");
 427                 goto remove_hp_group;
 428         }
 429
 430         return 0;
 431
 432 remove_hp_group:
 433         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 434 delete_obj:
 435         kobject_put(*hugepage_kobj);
 436         return err;
 437 }
 438
 439 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 440 {
 441         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 442         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 443         kobject_put(hugepage_kobj);
 444 }
 445 #else
 446 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 447 {
 448         return 0;
 449 }
 450
 451 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 452 {
 453 }
 454 #endif /* CONFIG_SYSFS */
 455
 456 static int __init hugepage_init(void)
 457 {
 458         int err;
 459         struct kobject *hugepage_kobj;
 460
 461         if (!has_transparent_hugepage()) {
 462                 /*
 463                  * Hardware doesn't support hugepages, hence disable
 464                  * DAX PMD support.
 465                  */
 466                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 467                 return -EINVAL;
 468         }
 469
 470         /*
 471          * hugepages can't be allocated by the buddy allocator
 472          */
 473         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 474         /*
 475          * we use page->mapping and page->index in second tail page
 476          * as list_head: assuming THP order >= 2
 477          */
 478         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 479
 480         err = hugepage_init_sysfs(&hugepage_kobj);
 481         if (err)
 482                 goto err_sysfs;
 483
 484         err = khugepaged_init();
 485         if (err)
 486                 goto err_slab;
 487
 488         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 489         if (err)
 490                 goto err_hzp_shrinker;
 491         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 492         if (err)
 493                 goto err_split_shrinker;
 494
 495         /*
 496          * By default disable transparent hugepages on smaller systems,
 497          * where the extra memory used could hurt more than TLB overhead
 498          * is likely to save.  The admin can still enable it through /sys.
 499          */
 500         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 501                 transparent_hugepage_flags = 0;
 502                 return 0;
 503         }
 504
 505         err = start_stop_khugepaged();
 506         if (err)
 507                 goto err_khugepaged;
 508
 509         return 0;
 510 err_khugepaged:
 511         unregister_shrinker(&deferred_split_shrinker);
 512 err_split_shrinker:
 513         unregister_shrinker(&huge_zero_page_shrinker);
 514 err_hzp_shrinker:
 515         khugepaged_destroy();
 516 err_slab:
 517         hugepage_exit_sysfs(hugepage_kobj);
 518 err_sysfs:
 519         return err;
 520 }
 521 subsys_initcall(hugepage_init);
 522
 523 static int __init setup_transparent_hugepage(char *str)
 524 {
 525         int ret = 0;
 526         if (!str)
 527                 goto out;
 528         if (!strcmp(str, "always")) {
 529                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 530                         &transparent_hugepage_flags);
 531                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 532                           &transparent_hugepage_flags);
 533                 ret = 1;
 534         } else if (!strcmp(str, "madvise")) {
 535                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 536                           &transparent_hugepage_flags);
 537                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 538                         &transparent_hugepage_flags);
 539                 ret = 1;
 540         } else if (!strcmp(str, "never")) {
 541                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 542                           &transparent_hugepage_flags);
 543                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 544                           &transparent_hugepage_flags);
 545                 ret = 1;
 546         }
 547 out:
 548         if (!ret)
 549                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 550         return ret;
 551 }
 552 __setup("transparent_hugepage=", setup_transparent_hugepage);
 553
 554 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 555 {
 556         if (likely(vma->vm_flags & VM_WRITE))
 557                 pmd = pmd_mkwrite(pmd);
 558         return pmd;
 559 }
 560
 561 #ifdef CONFIG_MEMCG
 562 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 563 {
 564         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 565         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 566
 567         if (memcg)
 568                 return &memcg->deferred_split_queue;
 569         else
 570                 return &pgdat->deferred_split_queue;
 571 }
 572 #else
 573 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 574 {
 575         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 576
 577         return &pgdat->deferred_split_queue;
 578 }
 579 #endif
 580
 581 void prep_transhuge_page(struct page *page)
 582 {
 583         /*
 584          * we use page->mapping and page->index in second tail page
 585          * as list_head: assuming THP order >= 2
 586          */
 587
 588         INIT_LIST_HEAD(page_deferred_list(page));
 589         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 590 }
 591
 592 static inline bool is_transparent_hugepage(struct page *page)
 593 {
 594         if (!PageCompound(page))
 595                 return false;
 596
 597         page = compound_head(page);
 598         return is_huge_zero_page(page) ||
 599                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 600 }
 601
 602 static unsigned long __thp_get_unmapped_area(struct file *filp,
 603                 unsigned long addr, unsigned long len,
 604                 loff_t off, unsigned long flags, unsigned long size)
 605 {
 606         loff_t off_end = off + len;
 607         loff_t off_align = round_up(off, size);
 608         unsigned long len_pad, ret;
 609
 610         if (off_end <= off_align || (off_end - off_align) < size)
 611                 return 0;
 612
 613         len_pad = len + size;
 614         if (len_pad < len || (off + len_pad) < off)
 615                 return 0;
 616
 617         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 618                                               off >> PAGE_SHIFT, flags);
 619
 620         /*
 621          * The failure might be due to length padding. The caller will retry
 622          * without the padding.
 623          */
 624         if (IS_ERR_VALUE(ret))
 625                 return 0;
 626
 627         /*
 628          * Do not try to align to THP boundary if allocation at the address
 629          * hint succeeds.
 630          */
 631         if (ret == addr)
 632                 return addr;
 633
 634         ret += (off - ret) & (size - 1);
 635         return ret;
 636 }
 637
 638 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 639                 unsigned long len, unsigned long pgoff, unsigned long flags)
 640 {
 641         unsigned long ret;
 642         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 643
 644         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 645         if (ret)
 646                 return ret;
 647
 648         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 649 }
 650 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 651
 652 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 653                         struct page *page, gfp_t gfp)
 654 {
 655         struct vm_area_struct *vma = vmf->vma;
 656         pgtable_t pgtable;
 657         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 658         vm_fault_t ret = 0;
 659
 660         VM_BUG_ON_PAGE(!PageCompound(page), page);
 661
 662         if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
 663                 put_page(page);
 664                 count_vm_event(THP_FAULT_FALLBACK);
 665                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 666                 return VM_FAULT_FALLBACK;
 667         }
 668         cgroup_throttle_swaprate(page, gfp);
 669
 670         pgtable = pte_alloc_one(vma->vm_mm);
 671         if (unlikely(!pgtable)) {
 672                 ret = VM_FAULT_OOM;
 673                 goto release;
 674         }
 675
 676         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 677         /*
 678          * The memory barrier inside __SetPageUptodate makes sure that
 679          * clear_huge_page writes become visible before the set_pmd_at()
 680          * write.
 681          */
 682         __SetPageUptodate(page);
 683
 684         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 685         if (unlikely(!pmd_none(*vmf->pmd))) {
 686                 goto unlock_release;
 687         } else {
 688                 pmd_t entry;
 689
 690                 ret = check_stable_address_space(vma->vm_mm);
 691                 if (ret)
 692                         goto unlock_release;
 693
 694                 /* Deliver the page fault to userland */
 695                 if (userfaultfd_missing(vma)) {
 696                         spin_unlock(vmf->ptl);
 697                         put_page(page);
 698                         pte_free(vma->vm_mm, pgtable);
 699                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 700                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 701                         return ret;
 702                 }
 703
 704                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 705                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 706                 page_add_new_anon_rmap(page, vma, haddr);
 707                 lru_cache_add_inactive_or_unevictable(page, vma);
 708                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 709                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 710                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 711                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 712                 mm_inc_nr_ptes(vma->vm_mm);
 713                 spin_unlock(vmf->ptl);
 714                 count_vm_event(THP_FAULT_ALLOC);
 715                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 716         }
 717
 718         return 0;
 719 unlock_release:
 720         spin_unlock(vmf->ptl);
 721 release:
 722         if (pgtable)
 723                 pte_free(vma->vm_mm, pgtable);
 724         put_page(page);
 725         return ret;
 726
 727 }
 728
 729 /*
 730  * always: directly stall for all thp allocations
 731  * defer: wake kswapd and fail if not immediately available
 732  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 733  *                fail if not immediately available
 734  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 735  *          available
 736  * never: never stall for any thp allocation
 737  */
 738 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 739 {
 740         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 741
 742         /* Always do synchronous compaction */
 743         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 744                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 745
 746         /* Kick kcompactd and fail quickly */
 747         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 748                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 749
 750         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 751         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 752                 return GFP_TRANSHUGE_LIGHT |
 753                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 754                                         __GFP_KSWAPD_RECLAIM);
 755
 756         /* Only do synchronous compaction if madvised */
 757         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 758                 return GFP_TRANSHUGE_LIGHT |
 759                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 760
 761         return GFP_TRANSHUGE_LIGHT;
 762 }
 763
 764 /* Caller must hold page table lock. */
 765 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 766                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 767                 struct page *zero_page)
 768 {
 769         pmd_t entry;
 770         if (!pmd_none(*pmd))
 771                 return;
 772         entry = mk_pmd(zero_page, vma->vm_page_prot);
 773         entry = pmd_mkhuge(entry);
 774         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 775         set_pmd_at(mm, haddr, pmd, entry);
 776         mm_inc_nr_ptes(mm);
 777 }
 778
 779 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 780 {
 781         struct vm_area_struct *vma = vmf->vma;
 782         gfp_t gfp;
 783         struct folio *folio;
 784         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 785
 786         if (!transhuge_vma_suitable(vma, haddr))
 787                 return VM_FAULT_FALLBACK;
 788         if (unlikely(anon_vma_prepare(vma)))
 789                 return VM_FAULT_OOM;
 790         khugepaged_enter_vma(vma, vma->vm_flags);
 791
 792         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 793                         !mm_forbids_zeropage(vma->vm_mm) &&
 794                         transparent_hugepage_use_zero_page()) {
 795                 pgtable_t pgtable;
 796                 struct page *zero_page;
 797                 vm_fault_t ret;
 798                 pgtable = pte_alloc_one(vma->vm_mm);
 799                 if (unlikely(!pgtable))
 800                         return VM_FAULT_OOM;
 801                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 802                 if (unlikely(!zero_page)) {
 803                         pte_free(vma->vm_mm, pgtable);
 804                         count_vm_event(THP_FAULT_FALLBACK);
 805                         return VM_FAULT_FALLBACK;
 806                 }
 807                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 808                 ret = 0;
 809                 if (pmd_none(*vmf->pmd)) {
 810                         ret = check_stable_address_space(vma->vm_mm);
 811                         if (ret) {
 812                                 spin_unlock(vmf->ptl);
 813                                 pte_free(vma->vm_mm, pgtable);
 814                         } else if (userfaultfd_missing(vma)) {
 815                                 spin_unlock(vmf->ptl);
 816                                 pte_free(vma->vm_mm, pgtable);
 817                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 818                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 819                         } else {
 820                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 821                                                    haddr, vmf->pmd, zero_page);
 822                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 823                                 spin_unlock(vmf->ptl);
 824                         }
 825                 } else {
 826                         spin_unlock(vmf->ptl);
 827                         pte_free(vma->vm_mm, pgtable);
 828                 }
 829                 return ret;
 830         }
 831         gfp = vma_thp_gfp_mask(vma);
 832         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 833         if (unlikely(!folio)) {
 834                 count_vm_event(THP_FAULT_FALLBACK);
 835                 return VM_FAULT_FALLBACK;
 836         }
 837         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 838 }
 839
 840 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 841                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 842                 pgtable_t pgtable)
 843 {
 844         struct mm_struct *mm = vma->vm_mm;
 845         pmd_t entry;
 846         spinlock_t *ptl;
 847
 848         ptl = pmd_lock(mm, pmd);
 849         if (!pmd_none(*pmd)) {
 850                 if (write) {
 851                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 852                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 853                                 goto out_unlock;
 854                         }
 855                         entry = pmd_mkyoung(*pmd);
 856                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 857                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 858                                 update_mmu_cache_pmd(vma, addr, pmd);
 859                 }
 860
 861                 goto out_unlock;
 862         }
 863
 864         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 865         if (pfn_t_devmap(pfn))
 866                 entry = pmd_mkdevmap(entry);
 867         if (write) {
 868                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 869                 entry = maybe_pmd_mkwrite(entry, vma);
 870         }
 871
 872         if (pgtable) {
 873                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 874                 mm_inc_nr_ptes(mm);
 875                 pgtable = NULL;
 876         }
 877
 878         set_pmd_at(mm, addr, pmd, entry);
 879         update_mmu_cache_pmd(vma, addr, pmd);
 880
 881 out_unlock:
 882         spin_unlock(ptl);
 883         if (pgtable)
 884                 pte_free(mm, pgtable);
 885 }
 886
 887 /**
 888  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 889  * @vmf: Structure describing the fault
 890  * @pfn: pfn to insert
 891  * @pgprot: page protection to use
 892  * @write: whether it's a write fault
 893  *
 894  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 895  * also consult the vmf_insert_mixed_prot() documentation when
 896  * @pgprot != @vmf->vma->vm_page_prot.
 897  *
 898  * Return: vm_fault_t value.
 899  */
 900 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 901                                    pgprot_t pgprot, bool write)
 902 {
 903         unsigned long addr = vmf->address & PMD_MASK;
 904         struct vm_area_struct *vma = vmf->vma;
 905         pgtable_t pgtable = NULL;
 906
 907         /*
 908          * If we had pmd_special, we could avoid all these restrictions,
 909          * but we need to be consistent with PTEs and architectures that
 910          * can't support a 'special' bit.
 911          */
 912         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 913                         !pfn_t_devmap(pfn));
 914         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 915                                                 (VM_PFNMAP|VM_MIXEDMAP));
 916         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 917
 918         if (addr < vma->vm_start || addr >= vma->vm_end)
 919                 return VM_FAULT_SIGBUS;
 920
 921         if (arch_needs_pgtable_deposit()) {
 922                 pgtable = pte_alloc_one(vma->vm_mm);
 923                 if (!pgtable)
 924                         return VM_FAULT_OOM;
 925         }
 926
 927         track_pfn_insert(vma, &pgprot, pfn);
 928
 929         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 930         return VM_FAULT_NOPAGE;
 931 }
 932 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 933
 934 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 935 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 936 {
 937         if (likely(vma->vm_flags & VM_WRITE))
 938                 pud = pud_mkwrite(pud);
 939         return pud;
 940 }
 941
 942 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 943                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 944 {
 945         struct mm_struct *mm = vma->vm_mm;
 946         pud_t entry;
 947         spinlock_t *ptl;
 948
 949         ptl = pud_lock(mm, pud);
 950         if (!pud_none(*pud)) {
 951                 if (write) {
 952                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 953                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 954                                 goto out_unlock;
 955                         }
 956                         entry = pud_mkyoung(*pud);
 957                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 958                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 959                                 update_mmu_cache_pud(vma, addr, pud);
 960                 }
 961                 goto out_unlock;
 962         }
 963
 964         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 965         if (pfn_t_devmap(pfn))
 966                 entry = pud_mkdevmap(entry);
 967         if (write) {
 968                 entry = pud_mkyoung(pud_mkdirty(entry));
 969                 entry = maybe_pud_mkwrite(entry, vma);
 970         }
 971         set_pud_at(mm, addr, pud, entry);
 972         update_mmu_cache_pud(vma, addr, pud);
 973
 974 out_unlock:
 975         spin_unlock(ptl);
 976 }
 977
 978 /**
 979  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 980  * @vmf: Structure describing the fault
 981  * @pfn: pfn to insert
 982  * @pgprot: page protection to use
 983  * @write: whether it's a write fault
 984  *
 985  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 986  * also consult the vmf_insert_mixed_prot() documentation when
 987  * @pgprot != @vmf->vma->vm_page_prot.
 988  *
 989  * Return: vm_fault_t value.
 990  */
 991 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 992                                    pgprot_t pgprot, bool write)
 993 {
 994         unsigned long addr = vmf->address & PUD_MASK;
 995         struct vm_area_struct *vma = vmf->vma;
 996
 997         /*
 998          * If we had pud_special, we could avoid all these restrictions,
 999          * but we need to be consistent with PTEs and architectures that
1000          * can't support a 'special' bit.
1001          */
1002         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1003                         !pfn_t_devmap(pfn));
1004         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005                                                 (VM_PFNMAP|VM_MIXEDMAP));
1006         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007
1008         if (addr < vma->vm_start || addr >= vma->vm_end)
1009                 return VM_FAULT_SIGBUS;
1010
1011         track_pfn_insert(vma, &pgprot, pfn);
1012
1013         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1014         return VM_FAULT_NOPAGE;
1015 }
1016 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1017 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1018
1019 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1020                       pmd_t *pmd, bool write)
1021 {
1022         pmd_t _pmd;
1023
1024         _pmd = pmd_mkyoung(*pmd);
1025         if (write)
1026                 _pmd = pmd_mkdirty(_pmd);
1027         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1028                                   pmd, _pmd, write))
1029                 update_mmu_cache_pmd(vma, addr, pmd);
1030 }
1031
1032 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1033                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1034 {
1035         unsigned long pfn = pmd_pfn(*pmd);
1036         struct mm_struct *mm = vma->vm_mm;
1037         struct page *page;
1038
1039         assert_spin_locked(pmd_lockptr(mm, pmd));
1040
1041         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1042         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1043                          (FOLL_PIN | FOLL_GET)))
1044                 return NULL;
1045
1046         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1047                 return NULL;
1048
1049         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1050                 /* pass */;
1051         else
1052                 return NULL;
1053
1054         if (flags & FOLL_TOUCH)
1055                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1056
1057         /*
1058          * device mapped pages can only be returned if the
1059          * caller will manage the page reference count.
1060          */
1061         if (!(flags & (FOLL_GET | FOLL_PIN)))
1062                 return ERR_PTR(-EEXIST);
1063
1064         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1065         *pgmap = get_dev_pagemap(pfn, *pgmap);
1066         if (!*pgmap)
1067                 return ERR_PTR(-EFAULT);
1068         page = pfn_to_page(pfn);
1069         if (!try_grab_page(page, flags))
1070                 page = ERR_PTR(-ENOMEM);
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
1197         assert_spin_locked(pud_lockptr(mm, pud));
1198
1199         if (flags & FOLL_WRITE && !pud_write(*pud))
1200                 return NULL;
1201
1202         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1203         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1204                          (FOLL_PIN | FOLL_GET)))
1205                 return NULL;
1206
1207         if (pud_present(*pud) && pud_devmap(*pud))
1208                 /* pass */;
1209         else
1210                 return NULL;
1211
1212         if (flags & FOLL_TOUCH)
1213                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1214
1215         /*
1216          * device mapped pages can only be returned if the
1217          * caller will manage the page reference count.
1218          *
1219          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1220          */
1221         if (!(flags & (FOLL_GET | FOLL_PIN)))
1222                 return ERR_PTR(-EEXIST);
1223
1224         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1225         *pgmap = get_dev_pagemap(pfn, *pgmap);
1226         if (!*pgmap)
1227                 return ERR_PTR(-EFAULT);
1228         page = pfn_to_page(pfn);
1229         if (!try_grab_page(page, flags))
1230                 page = ERR_PTR(-ENOMEM);
1231
1232         return page;
1233 }
1234
1235 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1236                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1237                   struct vm_area_struct *vma)
1238 {
1239         spinlock_t *dst_ptl, *src_ptl;
1240         pud_t pud;
1241         int ret;
1242
1243         dst_ptl = pud_lock(dst_mm, dst_pud);
1244         src_ptl = pud_lockptr(src_mm, src_pud);
1245         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1246
1247         ret = -EAGAIN;
1248         pud = *src_pud;
1249         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1250                 goto out_unlock;
1251
1252         /*
1253          * When page table lock is held, the huge zero pud should not be
1254          * under splitting since we don't split the page itself, only pud to
1255          * a page table.
1256          */
1257         if (is_huge_zero_pud(pud)) {
1258                 /* No huge zero pud yet */
1259         }
1260
1261         /*
1262          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1263          * and split if duplicating fails.
1264          */
1265         pudp_set_wrprotect(src_mm, addr, src_pud);
1266         pud = pud_mkold(pud_wrprotect(pud));
1267         set_pud_at(dst_mm, addr, dst_pud, pud);
1268
1269         ret = 0;
1270 out_unlock:
1271         spin_unlock(src_ptl);
1272         spin_unlock(dst_ptl);
1273         return ret;
1274 }
1275
1276 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1277 {
1278         bool write = vmf->flags & FAULT_FLAG_WRITE;
1279
1280         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1281         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1282                 goto unlock;
1283
1284         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1285 unlock:
1286         spin_unlock(vmf->ptl);
1287 }
1288 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1289
1290 void huge_pmd_set_accessed(struct vm_fault *vmf)
1291 {
1292         bool write = vmf->flags & FAULT_FLAG_WRITE;
1293
1294         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1295         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1296                 goto unlock;
1297
1298         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1299
1300 unlock:
1301         spin_unlock(vmf->ptl);
1302 }
1303
1304 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1305 {
1306         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1307         struct vm_area_struct *vma = vmf->vma;
1308         struct folio *folio;
1309         struct page *page;
1310         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1311         pmd_t orig_pmd = vmf->orig_pmd;
1312
1313         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1314         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1315
1316         VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE));
1317         VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE));
1318
1319         if (is_huge_zero_pmd(orig_pmd))
1320                 goto fallback;
1321
1322         spin_lock(vmf->ptl);
1323
1324         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1325                 spin_unlock(vmf->ptl);
1326                 return 0;
1327         }
1328
1329         page = pmd_page(orig_pmd);
1330         folio = page_folio(page);
1331         VM_BUG_ON_PAGE(!PageHead(page), page);
1332
1333         /* Early check when only holding the PT lock. */
1334         if (PageAnonExclusive(page))
1335                 goto reuse;
1336
1337         if (!folio_trylock(folio)) {
1338                 folio_get(folio);
1339                 spin_unlock(vmf->ptl);
1340                 folio_lock(folio);
1341                 spin_lock(vmf->ptl);
1342                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1343                         spin_unlock(vmf->ptl);
1344                         folio_unlock(folio);
1345                         folio_put(folio);
1346                         return 0;
1347                 }
1348                 folio_put(folio);
1349         }
1350
1351         /* Recheck after temporarily dropping the PT lock. */
1352         if (PageAnonExclusive(page)) {
1353                 folio_unlock(folio);
1354                 goto reuse;
1355         }
1356
1357         /*
1358          * See do_wp_page(): we can only reuse the folio exclusively if
1359          * there are no additional references. Note that we always drain
1360          * the LRU pagevecs immediately after adding a THP.
1361          */
1362         if (folio_ref_count(folio) >
1363                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1364                 goto unlock_fallback;
1365         if (folio_test_swapcache(folio))
1366                 folio_free_swap(folio);
1367         if (folio_ref_count(folio) == 1) {
1368                 pmd_t entry;
1369
1370                 page_move_anon_rmap(page, vma);
1371                 folio_unlock(folio);
1372 reuse:
1373                 if (unlikely(unshare)) {
1374                         spin_unlock(vmf->ptl);
1375                         return 0;
1376                 }
1377                 entry = pmd_mkyoung(orig_pmd);
1378                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1379                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1380                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1381                 spin_unlock(vmf->ptl);
1382                 return VM_FAULT_WRITE;
1383         }
1384
1385 unlock_fallback:
1386         folio_unlock(folio);
1387         spin_unlock(vmf->ptl);
1388 fallback:
1389         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1390         return VM_FAULT_FALLBACK;
1391 }
1392
1393 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1394 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1395                                         struct vm_area_struct *vma,
1396                                         unsigned int flags)
1397 {
1398         /* If the pmd is writable, we can write to the page. */
1399         if (pmd_write(pmd))
1400                 return true;
1401
1402         /* Maybe FOLL_FORCE is set to override it? */
1403         if (!(flags & FOLL_FORCE))
1404                 return false;
1405
1406         /* But FOLL_FORCE has no effect on shared mappings */
1407         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1408                 return false;
1409
1410         /* ... or read-only private ones */
1411         if (!(vma->vm_flags & VM_MAYWRITE))
1412                 return false;
1413
1414         /* ... or already writable ones that just need to take a write fault */
1415         if (vma->vm_flags & VM_WRITE)
1416                 return false;
1417
1418         /*
1419          * See can_change_pte_writable(): we broke COW and could map the page
1420          * writable if we have an exclusive anonymous page ...
1421          */
1422         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1423                 return false;
1424
1425         /* ... and a write-fault isn't required for other reasons. */
1426         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1427                 return false;
1428         return !userfaultfd_huge_pmd_wp(vma, pmd);
1429 }
1430
1431 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1432                                    unsigned long addr,
1433                                    pmd_t *pmd,
1434                                    unsigned int flags)
1435 {
1436         struct mm_struct *mm = vma->vm_mm;
1437         struct page *page;
1438
1439         assert_spin_locked(pmd_lockptr(mm, pmd));
1440
1441         page = pmd_page(*pmd);
1442         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1443
1444         if ((flags & FOLL_WRITE) &&
1445             !can_follow_write_pmd(*pmd, page, vma, flags))
1446                 return NULL;
1447
1448         /* Avoid dumping huge zero page */
1449         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1450                 return ERR_PTR(-EFAULT);
1451
1452         /* Full NUMA hinting faults to serialise migration in fault paths */
1453         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1454                 return NULL;
1455
1456         if (!pmd_write(*pmd) && gup_must_unshare(flags, page))
1457                 return ERR_PTR(-EMLINK);
1458
1459         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1460                         !PageAnonExclusive(page), page);
1461
1462         if (!try_grab_page(page, flags))
1463                 return ERR_PTR(-ENOMEM);
1464
1465         if (flags & FOLL_TOUCH)
1466                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1467
1468         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1469         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1470
1471         return page;
1472 }
1473
1474 /* NUMA hinting page fault entry point for trans huge pmds */
1475 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1476 {
1477         struct vm_area_struct *vma = vmf->vma;
1478         pmd_t oldpmd = vmf->orig_pmd;
1479         pmd_t pmd;
1480         struct page *page;
1481         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1482         int page_nid = NUMA_NO_NODE;
1483         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1484         bool migrated = false;
1485         bool was_writable = pmd_savedwrite(oldpmd);
1486         int flags = 0;
1487
1488         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1489         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1490                 spin_unlock(vmf->ptl);
1491                 goto out;
1492         }
1493
1494         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1495         page = vm_normal_page_pmd(vma, haddr, pmd);
1496         if (!page)
1497                 goto out_map;
1498
1499         /* See similar comment in do_numa_page for explanation */
1500         if (!was_writable)
1501                 flags |= TNF_NO_GROUP;
1502
1503         page_nid = page_to_nid(page);
1504         /*
1505          * For memory tiering mode, cpupid of slow memory page is used
1506          * to record page access time.  So use default value.
1507          */
1508         if (node_is_toptier(page_nid))
1509                 last_cpupid = page_cpupid_last(page);
1510         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1511                                        &flags);
1512
1513         if (target_nid == NUMA_NO_NODE) {
1514                 put_page(page);
1515                 goto out_map;
1516         }
1517
1518         spin_unlock(vmf->ptl);
1519
1520         migrated = migrate_misplaced_page(page, vma, target_nid);
1521         if (migrated) {
1522                 flags |= TNF_MIGRATED;
1523                 page_nid = target_nid;
1524         } else {
1525                 flags |= TNF_MIGRATE_FAIL;
1526                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1527                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1528                         spin_unlock(vmf->ptl);
1529                         goto out;
1530                 }
1531                 goto out_map;
1532         }
1533
1534 out:
1535         if (page_nid != NUMA_NO_NODE)
1536                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1537                                 flags);
1538
1539         return 0;
1540
1541 out_map:
1542         /* Restore the PMD */
1543         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1544         pmd = pmd_mkyoung(pmd);
1545         if (was_writable)
1546                 pmd = pmd_mkwrite(pmd);
1547         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1548         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1549         spin_unlock(vmf->ptl);
1550         goto out;
1551 }
1552
1553 /*
1554  * Return true if we do MADV_FREE successfully on entire pmd page.
1555  * Otherwise, return false.
1556  */
1557 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1558                 pmd_t *pmd, unsigned long addr, unsigned long next)
1559 {
1560         spinlock_t *ptl;
1561         pmd_t orig_pmd;
1562         struct page *page;
1563         struct mm_struct *mm = tlb->mm;
1564         bool ret = false;
1565
1566         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1567
1568         ptl = pmd_trans_huge_lock(pmd, vma);
1569         if (!ptl)
1570                 goto out_unlocked;
1571
1572         orig_pmd = *pmd;
1573         if (is_huge_zero_pmd(orig_pmd))
1574                 goto out;
1575
1576         if (unlikely(!pmd_present(orig_pmd))) {
1577                 VM_BUG_ON(thp_migration_supported() &&
1578                                   !is_pmd_migration_entry(orig_pmd));
1579                 goto out;
1580         }
1581
1582         page = pmd_page(orig_pmd);
1583         /*
1584          * If other processes are mapping this page, we couldn't discard
1585          * the page unless they all do MADV_FREE so let's skip the page.
1586          */
1587         if (total_mapcount(page) != 1)
1588                 goto out;
1589
1590         if (!trylock_page(page))
1591                 goto out;
1592
1593         /*
1594          * If user want to discard part-pages of THP, split it so MADV_FREE
1595          * will deactivate only them.
1596          */
1597         if (next - addr != HPAGE_PMD_SIZE) {
1598                 get_page(page);
1599                 spin_unlock(ptl);
1600                 split_huge_page(page);
1601                 unlock_page(page);
1602                 put_page(page);
1603                 goto out_unlocked;
1604         }
1605
1606         if (PageDirty(page))
1607                 ClearPageDirty(page);
1608         unlock_page(page);
1609
1610         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1611                 pmdp_invalidate(vma, addr, pmd);
1612                 orig_pmd = pmd_mkold(orig_pmd);
1613                 orig_pmd = pmd_mkclean(orig_pmd);
1614
1615                 set_pmd_at(mm, addr, pmd, orig_pmd);
1616                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1617         }
1618
1619         mark_page_lazyfree(page);
1620         ret = true;
1621 out:
1622         spin_unlock(ptl);
1623 out_unlocked:
1624         return ret;
1625 }
1626
1627 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1628 {
1629         pgtable_t pgtable;
1630
1631         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1632         pte_free(mm, pgtable);
1633         mm_dec_nr_ptes(mm);
1634 }
1635
1636 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1637                  pmd_t *pmd, unsigned long addr)
1638 {
1639         pmd_t orig_pmd;
1640         spinlock_t *ptl;
1641
1642         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1643
1644         ptl = __pmd_trans_huge_lock(pmd, vma);
1645         if (!ptl)
1646                 return 0;
1647         /*
1648          * For architectures like ppc64 we look at deposited pgtable
1649          * when calling pmdp_huge_get_and_clear. So do the
1650          * pgtable_trans_huge_withdraw after finishing pmdp related
1651          * operations.
1652          */
1653         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1654                                                 tlb->fullmm);
1655         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1656         if (vma_is_special_huge(vma)) {
1657                 if (arch_needs_pgtable_deposit())
1658                         zap_deposited_table(tlb->mm, pmd);
1659                 spin_unlock(ptl);
1660         } else if (is_huge_zero_pmd(orig_pmd)) {
1661                 zap_deposited_table(tlb->mm, pmd);
1662                 spin_unlock(ptl);
1663         } else {
1664                 struct page *page = NULL;
1665                 int flush_needed = 1;
1666
1667                 if (pmd_present(orig_pmd)) {
1668                         page = pmd_page(orig_pmd);
1669                         page_remove_rmap(page, vma, true);
1670                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1671                         VM_BUG_ON_PAGE(!PageHead(page), page);
1672                 } else if (thp_migration_supported()) {
1673                         swp_entry_t entry;
1674
1675                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1676                         entry = pmd_to_swp_entry(orig_pmd);
1677                         page = pfn_swap_entry_to_page(entry);
1678                         flush_needed = 0;
1679                 } else
1680                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1681
1682                 if (PageAnon(page)) {
1683                         zap_deposited_table(tlb->mm, pmd);
1684                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1685                 } else {
1686                         if (arch_needs_pgtable_deposit())
1687                                 zap_deposited_table(tlb->mm, pmd);
1688                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1689                 }
1690
1691                 spin_unlock(ptl);
1692                 if (flush_needed)
1693                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1694         }
1695         return 1;
1696 }
1697
1698 #ifndef pmd_move_must_withdraw
1699 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1700                                          spinlock_t *old_pmd_ptl,
1701                                          struct vm_area_struct *vma)
1702 {
1703         /*
1704          * With split pmd lock we also need to move preallocated
1705          * PTE page table if new_pmd is on different PMD page table.
1706          *
1707          * We also don't deposit and withdraw tables for file pages.
1708          */
1709         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1710 }
1711 #endif
1712
1713 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1714 {
1715 #ifdef CONFIG_MEM_SOFT_DIRTY
1716         if (unlikely(is_pmd_migration_entry(pmd)))
1717                 pmd = pmd_swp_mksoft_dirty(pmd);
1718         else if (pmd_present(pmd))
1719                 pmd = pmd_mksoft_dirty(pmd);
1720 #endif
1721         return pmd;
1722 }
1723
1724 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1725                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1726 {
1727         spinlock_t *old_ptl, *new_ptl;
1728         pmd_t pmd;
1729         struct mm_struct *mm = vma->vm_mm;
1730         bool force_flush = false;
1731
1732         /*
1733          * The destination pmd shouldn't be established, free_pgtables()
1734          * should have release it.
1735          */
1736         if (WARN_ON(!pmd_none(*new_pmd))) {
1737                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1738                 return false;
1739         }
1740
1741         /*
1742          * We don't have to worry about the ordering of src and dst
1743          * ptlocks because exclusive mmap_lock prevents deadlock.
1744          */
1745         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1746         if (old_ptl) {
1747                 new_ptl = pmd_lockptr(mm, new_pmd);
1748                 if (new_ptl != old_ptl)
1749                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1750                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1751                 if (pmd_present(pmd))
1752                         force_flush = true;
1753                 VM_BUG_ON(!pmd_none(*new_pmd));
1754
1755                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1756                         pgtable_t pgtable;
1757                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1758                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1759                 }
1760                 pmd = move_soft_dirty_pmd(pmd);
1761                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1762                 if (force_flush)
1763                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1764                 if (new_ptl != old_ptl)
1765                         spin_unlock(new_ptl);
1766                 spin_unlock(old_ptl);
1767                 return true;
1768         }
1769         return false;
1770 }
1771
1772 /*
1773  * Returns
1774  *  - 0 if PMD could not be locked
1775  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1776  *      or if prot_numa but THP migration is not supported
1777  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1778  */
1779 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1780                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1781                     unsigned long cp_flags)
1782 {
1783         struct mm_struct *mm = vma->vm_mm;
1784         spinlock_t *ptl;
1785         pmd_t oldpmd, entry;
1786         bool preserve_write;
1787         int ret;
1788         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1789         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1790         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1791
1792         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1793
1794         if (prot_numa && !thp_migration_supported())
1795                 return 1;
1796
1797         ptl = __pmd_trans_huge_lock(pmd, vma);
1798         if (!ptl)
1799                 return 0;
1800
1801         preserve_write = prot_numa && pmd_write(*pmd);
1802         ret = 1;
1803
1804 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1805         if (is_swap_pmd(*pmd)) {
1806                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1807                 struct page *page = pfn_swap_entry_to_page(entry);
1808
1809                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1810                 if (is_writable_migration_entry(entry)) {
1811                         pmd_t newpmd;
1812                         /*
1813                          * A protection check is difficult so
1814                          * just be safe and disable write
1815                          */
1816                         if (PageAnon(page))
1817                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1818                         else
1819                                 entry = make_readable_migration_entry(swp_offset(entry));
1820                         newpmd = swp_entry_to_pmd(entry);
1821                         if (pmd_swp_soft_dirty(*pmd))
1822                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1823                         if (pmd_swp_uffd_wp(*pmd))
1824                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1825                         set_pmd_at(mm, addr, pmd, newpmd);
1826                 }
1827                 goto unlock;
1828         }
1829 #endif
1830
1831         if (prot_numa) {
1832                 struct page *page;
1833                 bool toptier;
1834                 /*
1835                  * Avoid trapping faults against the zero page. The read-only
1836                  * data is likely to be read-cached on the local CPU and
1837                  * local/remote hits to the zero page are not interesting.
1838                  */
1839                 if (is_huge_zero_pmd(*pmd))
1840                         goto unlock;
1841
1842                 if (pmd_protnone(*pmd))
1843                         goto unlock;
1844
1845                 page = pmd_page(*pmd);
1846                 toptier = node_is_toptier(page_to_nid(page));
1847                 /*
1848                  * Skip scanning top tier node if normal numa
1849                  * balancing is disabled
1850                  */
1851                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1852                     toptier)
1853                         goto unlock;
1854
1855                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1856                     !toptier)
1857                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1858         }
1859         /*
1860          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1861          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1862          * which is also under mmap_read_lock(mm):
1863          *
1864          *      CPU0:                           CPU1:
1865          *                              change_huge_pmd(prot_numa=1)
1866          *                               pmdp_huge_get_and_clear_notify()
1867          * madvise_dontneed()
1868          *  zap_pmd_range()
1869          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1870          *   // skip the pmd
1871          *                               set_pmd_at();
1872          *                               // pmd is re-established
1873          *
1874          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1875          * which may break userspace.
1876          *
1877          * pmdp_invalidate_ad() is required to make sure we don't miss
1878          * dirty/young flags set by hardware.
1879          */
1880         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1881
1882         entry = pmd_modify(oldpmd, newprot);
1883         if (preserve_write)
1884                 entry = pmd_mk_savedwrite(entry);
1885         if (uffd_wp) {
1886                 entry = pmd_wrprotect(entry);
1887                 entry = pmd_mkuffd_wp(entry);
1888         } else if (uffd_wp_resolve) {
1889                 /*
1890                  * Leave the write bit to be handled by PF interrupt
1891                  * handler, then things like COW could be properly
1892                  * handled.
1893                  */
1894                 entry = pmd_clear_uffd_wp(entry);
1895         }
1896         ret = HPAGE_PMD_NR;
1897         set_pmd_at(mm, addr, pmd, entry);
1898
1899         if (huge_pmd_needs_flush(oldpmd, entry))
1900                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1901
1902         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1903 unlock:
1904         spin_unlock(ptl);
1905         return ret;
1906 }
1907
1908 /*
1909  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1910  *
1911  * Note that if it returns page table lock pointer, this routine returns without
1912  * unlocking page table lock. So callers must unlock it.
1913  */
1914 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1915 {
1916         spinlock_t *ptl;
1917         ptl = pmd_lock(vma->vm_mm, pmd);
1918         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1919                         pmd_devmap(*pmd)))
1920                 return ptl;
1921         spin_unlock(ptl);
1922         return NULL;
1923 }
1924
1925 /*
1926  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1927  *
1928  * Note that if it returns page table lock pointer, this routine returns without
1929  * unlocking page table lock. So callers must unlock it.
1930  */
1931 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1932 {
1933         spinlock_t *ptl;
1934
1935         ptl = pud_lock(vma->vm_mm, pud);
1936         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1937                 return ptl;
1938         spin_unlock(ptl);
1939         return NULL;
1940 }
1941
1942 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1943 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1944                  pud_t *pud, unsigned long addr)
1945 {
1946         spinlock_t *ptl;
1947
1948         ptl = __pud_trans_huge_lock(pud, vma);
1949         if (!ptl)
1950                 return 0;
1951
1952         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1953         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1954         if (vma_is_special_huge(vma)) {
1955                 spin_unlock(ptl);
1956                 /* No zero page support yet */
1957         } else {
1958                 /* No support for anonymous PUD pages yet */
1959                 BUG();
1960         }
1961         return 1;
1962 }
1963
1964 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1965                 unsigned long haddr)
1966 {
1967         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1968         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1969         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1970         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1971
1972         count_vm_event(THP_SPLIT_PUD);
1973
1974         pudp_huge_clear_flush_notify(vma, haddr, pud);
1975 }
1976
1977 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1978                 unsigned long address)
1979 {
1980         spinlock_t *ptl;
1981         struct mmu_notifier_range range;
1982
1983         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1984                                 address & HPAGE_PUD_MASK,
1985                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1986         mmu_notifier_invalidate_range_start(&range);
1987         ptl = pud_lock(vma->vm_mm, pud);
1988         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1989                 goto out;
1990         __split_huge_pud_locked(vma, pud, range.start);
1991
1992 out:
1993         spin_unlock(ptl);
1994         /*
1995          * No need to double call mmu_notifier->invalidate_range() callback as
1996          * the above pudp_huge_clear_flush_notify() did already call it.
1997          */
1998         mmu_notifier_invalidate_range_only_end(&range);
1999 }
2000 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2001
2002 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2003                 unsigned long haddr, pmd_t *pmd)
2004 {
2005         struct mm_struct *mm = vma->vm_mm;
2006         pgtable_t pgtable;
2007         pmd_t _pmd;
2008         int i;
2009
2010         /*
2011          * Leave pmd empty until pte is filled note that it is fine to delay
2012          * notification until mmu_notifier_invalidate_range_end() as we are
2013          * replacing a zero pmd write protected page with a zero pte write
2014          * protected page.
2015          *
2016          * See Documentation/mm/mmu_notifier.rst
2017          */
2018         pmdp_huge_clear_flush(vma, haddr, pmd);
2019
2020         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2021         pmd_populate(mm, &_pmd, pgtable);
2022
2023         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2024                 pte_t *pte, entry;
2025                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2026                 entry = pte_mkspecial(entry);
2027                 pte = pte_offset_map(&_pmd, haddr);
2028                 VM_BUG_ON(!pte_none(*pte));
2029                 set_pte_at(mm, haddr, pte, entry);
2030                 pte_unmap(pte);
2031         }
2032         smp_wmb(); /* make pte visible before pmd */
2033         pmd_populate(mm, pmd, pgtable);
2034 }
2035
2036 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2037                 unsigned long haddr, bool freeze)
2038 {
2039         struct mm_struct *mm = vma->vm_mm;
2040         struct page *page;
2041         pgtable_t pgtable;
2042         pmd_t old_pmd, _pmd;
2043         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2044         bool anon_exclusive = false, dirty = false;
2045         unsigned long addr;
2046         int i;
2047
2048         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2049         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2050         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2051         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2052                                 && !pmd_devmap(*pmd));
2053
2054         count_vm_event(THP_SPLIT_PMD);
2055
2056         if (!vma_is_anonymous(vma)) {
2057                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2058                 /*
2059                  * We are going to unmap this huge page. So
2060                  * just go ahead and zap it
2061                  */
2062                 if (arch_needs_pgtable_deposit())
2063                         zap_deposited_table(mm, pmd);
2064                 if (vma_is_special_huge(vma))
2065                         return;
2066                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2067                         swp_entry_t entry;
2068
2069                         entry = pmd_to_swp_entry(old_pmd);
2070                         page = pfn_swap_entry_to_page(entry);
2071                 } else {
2072                         page = pmd_page(old_pmd);
2073                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2074                                 set_page_dirty(page);
2075                         if (!PageReferenced(page) && pmd_young(old_pmd))
2076                                 SetPageReferenced(page);
2077                         page_remove_rmap(page, vma, true);
2078                         put_page(page);
2079                 }
2080                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2081                 return;
2082         }
2083
2084         if (is_huge_zero_pmd(*pmd)) {
2085                 /*
2086                  * FIXME: Do we want to invalidate secondary mmu by calling
2087                  * mmu_notifier_invalidate_range() see comments below inside
2088                  * __split_huge_pmd() ?
2089                  *
2090                  * We are going from a zero huge page write protected to zero
2091                  * small page also write protected so it does not seems useful
2092                  * to invalidate secondary mmu at this time.
2093                  */
2094                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2095         }
2096
2097         /*
2098          * Up to this point the pmd is present and huge and userland has the
2099          * whole access to the hugepage during the split (which happens in
2100          * place). If we overwrite the pmd with the not-huge version pointing
2101          * to the pte here (which of course we could if all CPUs were bug
2102          * free), userland could trigger a small page size TLB miss on the
2103          * small sized TLB while the hugepage TLB entry is still established in
2104          * the huge TLB. Some CPU doesn't like that.
2105          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2106          * 383 on page 105. Intel should be safe but is also warns that it's
2107          * only safe if the permission and cache attributes of the two entries
2108          * loaded in the two TLB is identical (which should be the case here).
2109          * But it is generally safer to never allow small and huge TLB entries
2110          * for the same virtual address to be loaded simultaneously. So instead
2111          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2112          * current pmd notpresent (atomically because here the pmd_trans_huge
2113          * must remain set at all times on the pmd until the split is complete
2114          * for this pmd), then we flush the SMP TLB and finally we write the
2115          * non-huge version of the pmd entry with pmd_populate.
2116          */
2117         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2118
2119         pmd_migration = is_pmd_migration_entry(old_pmd);
2120         if (unlikely(pmd_migration)) {
2121                 swp_entry_t entry;
2122
2123                 entry = pmd_to_swp_entry(old_pmd);
2124                 page = pfn_swap_entry_to_page(entry);
2125                 write = is_writable_migration_entry(entry);
2126                 if (PageAnon(page))
2127                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2128                 young = is_migration_entry_young(entry);
2129                 dirty = is_migration_entry_dirty(entry);
2130                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2131                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2132         } else {
2133                 page = pmd_page(old_pmd);
2134                 if (pmd_dirty(old_pmd)) {
2135                         dirty = true;
2136                         SetPageDirty(page);
2137                 }
2138                 write = pmd_write(old_pmd);
2139                 young = pmd_young(old_pmd);
2140                 soft_dirty = pmd_soft_dirty(old_pmd);
2141                 uffd_wp = pmd_uffd_wp(old_pmd);
2142
2143                 VM_BUG_ON_PAGE(!page_count(page), page);
2144                 page_ref_add(page, HPAGE_PMD_NR - 1);
2145
2146                 /*
2147                  * Without "freeze", we'll simply split the PMD, propagating the
2148                  * PageAnonExclusive() flag for each PTE by setting it for
2149                  * each subpage -- no need to (temporarily) clear.
2150                  *
2151                  * With "freeze" we want to replace mapped pages by
2152                  * migration entries right away. This is only possible if we
2153                  * managed to clear PageAnonExclusive() -- see
2154                  * set_pmd_migration_entry().
2155                  *
2156                  * In case we cannot clear PageAnonExclusive(), split the PMD
2157                  * only and let try_to_migrate_one() fail later.
2158                  *
2159                  * See page_try_share_anon_rmap(): invalidate PMD first.
2160                  */
2161                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2162                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2163                         freeze = false;
2164         }
2165
2166         /*
2167          * Withdraw the table only after we mark the pmd entry invalid.
2168          * This's critical for some architectures (Power).
2169          */
2170         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2171         pmd_populate(mm, &_pmd, pgtable);
2172
2173         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2174                 pte_t entry, *pte;
2175                 /*
2176                  * Note that NUMA hinting access restrictions are not
2177                  * transferred to avoid any possibility of altering
2178                  * permissions across VMAs.
2179                  */
2180                 if (freeze || pmd_migration) {
2181                         swp_entry_t swp_entry;
2182                         if (write)
2183                                 swp_entry = make_writable_migration_entry(
2184                                                         page_to_pfn(page + i));
2185                         else if (anon_exclusive)
2186                                 swp_entry = make_readable_exclusive_migration_entry(
2187                                                         page_to_pfn(page + i));
2188                         else
2189                                 swp_entry = make_readable_migration_entry(
2190                                                         page_to_pfn(page + i));
2191                         if (young)
2192                                 swp_entry = make_migration_entry_young(swp_entry);
2193                         if (dirty)
2194                                 swp_entry = make_migration_entry_dirty(swp_entry);
2195                         entry = swp_entry_to_pte(swp_entry);
2196                         if (soft_dirty)
2197                                 entry = pte_swp_mksoft_dirty(entry);
2198                         if (uffd_wp)
2199                                 entry = pte_swp_mkuffd_wp(entry);
2200                 } else {
2201                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2202                         entry = maybe_mkwrite(entry, vma);
2203                         if (anon_exclusive)
2204                                 SetPageAnonExclusive(page + i);
2205                         if (!write)
2206                                 entry = pte_wrprotect(entry);
2207                         if (!young)
2208                                 entry = pte_mkold(entry);
2209                         /* NOTE: this may set soft-dirty too on some archs */
2210                         if (dirty)
2211                                 entry = pte_mkdirty(entry);
2212                         if (soft_dirty)
2213                                 entry = pte_mksoft_dirty(entry);
2214                         if (uffd_wp)
2215                                 entry = pte_mkuffd_wp(entry);
2216                 }
2217                 pte = pte_offset_map(&_pmd, addr);
2218                 BUG_ON(!pte_none(*pte));
2219                 set_pte_at(mm, addr, pte, entry);
2220                 if (!pmd_migration)
2221                         atomic_inc(&page[i]._mapcount);
2222                 pte_unmap(pte);
2223         }
2224
2225         if (!pmd_migration) {
2226                 /*
2227                  * Set PG_double_map before dropping compound_mapcount to avoid
2228                  * false-negative page_mapped().
2229                  */
2230                 if (compound_mapcount(page) > 1 &&
2231                     !TestSetPageDoubleMap(page)) {
2232                         for (i = 0; i < HPAGE_PMD_NR; i++)
2233                                 atomic_inc(&page[i]._mapcount);
2234                 }
2235
2236                 lock_page_memcg(page);
2237                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2238                         /* Last compound_mapcount is gone. */
2239                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2240                                                 -HPAGE_PMD_NR);
2241                         if (TestClearPageDoubleMap(page)) {
2242                                 /* No need in mapcount reference anymore */
2243                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2244                                         atomic_dec(&page[i]._mapcount);
2245                         }
2246                 }
2247                 unlock_page_memcg(page);
2248
2249                 /* Above is effectively page_remove_rmap(page, vma, true) */
2250                 munlock_vma_page(page, vma, true);
2251         }
2252
2253         smp_wmb(); /* make pte visible before pmd */
2254         pmd_populate(mm, pmd, pgtable);
2255
2256         if (freeze) {
2257                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2258                         page_remove_rmap(page + i, vma, false);
2259                         put_page(page + i);
2260                 }
2261         }
2262 }
2263
2264 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2265                 unsigned long address, bool freeze, struct folio *folio)
2266 {
2267         spinlock_t *ptl;
2268         struct mmu_notifier_range range;
2269
2270         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2271                                 address & HPAGE_PMD_MASK,
2272                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2273         mmu_notifier_invalidate_range_start(&range);
2274         ptl = pmd_lock(vma->vm_mm, pmd);
2275
2276         /*
2277          * If caller asks to setup a migration entry, we need a folio to check
2278          * pmd against. Otherwise we can end up replacing wrong folio.
2279          */
2280         VM_BUG_ON(freeze && !folio);
2281         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2282
2283         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2284             is_pmd_migration_entry(*pmd)) {
2285                 /*
2286                  * It's safe to call pmd_page when folio is set because it's
2287                  * guaranteed that pmd is present.
2288                  */
2289                 if (folio && folio != page_folio(pmd_page(*pmd)))
2290                         goto out;
2291                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2292         }
2293
2294 out:
2295         spin_unlock(ptl);
2296         /*
2297          * No need to double call mmu_notifier->invalidate_range() callback.
2298          * They are 3 cases to consider inside __split_huge_pmd_locked():
2299          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2300          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2301          *    fault will trigger a flush_notify before pointing to a new page
2302          *    (it is fine if the secondary mmu keeps pointing to the old zero
2303          *    page in the meantime)
2304          *  3) Split a huge pmd into pte pointing to the same page. No need
2305          *     to invalidate secondary tlb entry they are all still valid.
2306          *     any further changes to individual pte will notify. So no need
2307          *     to call mmu_notifier->invalidate_range()
2308          */
2309         mmu_notifier_invalidate_range_only_end(&range);
2310 }
2311
2312 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2313                 bool freeze, struct folio *folio)
2314 {
2315         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2316
2317         if (!pmd)
2318                 return;
2319
2320         __split_huge_pmd(vma, pmd, address, freeze, folio);
2321 }
2322
2323 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2324 {
2325         /*
2326          * If the new address isn't hpage aligned and it could previously
2327          * contain an hugepage: check if we need to split an huge pmd.
2328          */
2329         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2330             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2331                          ALIGN(address, HPAGE_PMD_SIZE)))
2332                 split_huge_pmd_address(vma, address, false, NULL);
2333 }
2334
2335 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2336                              unsigned long start,
2337                              unsigned long end,
2338                              long adjust_next)
2339 {
2340         /* Check if we need to split start first. */
2341         split_huge_pmd_if_needed(vma, start);
2342
2343         /* Check if we need to split end next. */
2344         split_huge_pmd_if_needed(vma, end);
2345
2346         /*
2347          * If we're also updating the next vma vm_start,
2348          * check if we need to split it.
2349          */
2350         if (adjust_next > 0) {
2351                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2352                 unsigned long nstart = next->vm_start;
2353                 nstart += adjust_next;
2354                 split_huge_pmd_if_needed(next, nstart);
2355         }
2356 }
2357
2358 static void unmap_folio(struct folio *folio)
2359 {
2360         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2361                 TTU_SYNC;
2362
2363         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2364
2365         /*
2366          * Anon pages need migration entries to preserve them, but file
2367          * pages can simply be left unmapped, then faulted back on demand.
2368          * If that is ever changed (perhaps for mlock), update remap_page().
2369          */
2370         if (folio_test_anon(folio))
2371                 try_to_migrate(folio, ttu_flags);
2372         else
2373                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2374 }
2375
2376 static void remap_page(struct folio *folio, unsigned long nr)
2377 {
2378         int i = 0;
2379
2380         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2381         if (!folio_test_anon(folio))
2382                 return;
2383         for (;;) {
2384                 remove_migration_ptes(folio, folio, true);
2385                 i += folio_nr_pages(folio);
2386                 if (i >= nr)
2387                         break;
2388                 folio = folio_next(folio);
2389         }
2390 }
2391
2392 static void lru_add_page_tail(struct page *head, struct page *tail,
2393                 struct lruvec *lruvec, struct list_head *list)
2394 {
2395         VM_BUG_ON_PAGE(!PageHead(head), head);
2396         VM_BUG_ON_PAGE(PageCompound(tail), head);
2397         VM_BUG_ON_PAGE(PageLRU(tail), head);
2398         lockdep_assert_held(&lruvec->lru_lock);
2399
2400         if (list) {
2401                 /* page reclaim is reclaiming a huge page */
2402                 VM_WARN_ON(PageLRU(head));
2403                 get_page(tail);
2404                 list_add_tail(&tail->lru, list);
2405         } else {
2406                 /* head is still on lru (and we have it frozen) */
2407                 VM_WARN_ON(!PageLRU(head));
2408                 if (PageUnevictable(tail))
2409                         tail->mlock_count = 0;
2410                 else
2411                         list_add_tail(&tail->lru, &head->lru);
2412                 SetPageLRU(tail);
2413         }
2414 }
2415
2416 static void __split_huge_page_tail(struct page *head, int tail,
2417                 struct lruvec *lruvec, struct list_head *list)
2418 {
2419         struct page *page_tail = head + tail;
2420
2421         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2422
2423         /*
2424          * Clone page flags before unfreezing refcount.
2425          *
2426          * After successful get_page_unless_zero() might follow flags change,
2427          * for example lock_page() which set PG_waiters.
2428          *
2429          * Note that for mapped sub-pages of an anonymous THP,
2430          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2431          * the migration entry instead from where remap_page() will restore it.
2432          * We can still have PG_anon_exclusive set on effectively unmapped and
2433          * unreferenced sub-pages of an anonymous THP: we can simply drop
2434          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2435          */
2436         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2437         page_tail->flags |= (head->flags &
2438                         ((1L << PG_referenced) |
2439                          (1L << PG_swapbacked) |
2440                          (1L << PG_swapcache) |
2441                          (1L << PG_mlocked) |
2442                          (1L << PG_uptodate) |
2443                          (1L << PG_active) |
2444                          (1L << PG_workingset) |
2445                          (1L << PG_locked) |
2446                          (1L << PG_unevictable) |
2447 #ifdef CONFIG_64BIT
2448                          (1L << PG_arch_2) |
2449 #endif
2450                          (1L << PG_dirty) |
2451                          LRU_GEN_MASK | LRU_REFS_MASK));
2452
2453         /* ->mapping in first tail page is compound_mapcount */
2454         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2455                         page_tail);
2456         page_tail->mapping = head->mapping;
2457         page_tail->index = head->index + tail;
2458         page_tail->private = 0;
2459
2460         /* Page flags must be visible before we make the page non-compound. */
2461         smp_wmb();
2462
2463         /*
2464          * Clear PageTail before unfreezing page refcount.
2465          *
2466          * After successful get_page_unless_zero() might follow put_page()
2467          * which needs correct compound_head().
2468          */
2469         clear_compound_head(page_tail);
2470
2471         /* Finally unfreeze refcount. Additional reference from page cache. */
2472         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2473                                           PageSwapCache(head)));
2474
2475         if (page_is_young(head))
2476                 set_page_young(page_tail);
2477         if (page_is_idle(head))
2478                 set_page_idle(page_tail);
2479
2480         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2481
2482         /*
2483          * always add to the tail because some iterators expect new
2484          * pages to show after the currently processed elements - e.g.
2485          * migrate_pages
2486          */
2487         lru_add_page_tail(head, page_tail, lruvec, list);
2488 }
2489
2490 static void __split_huge_page(struct page *page, struct list_head *list,
2491                 pgoff_t end)
2492 {
2493         struct folio *folio = page_folio(page);
2494         struct page *head = &folio->page;
2495         struct lruvec *lruvec;
2496         struct address_space *swap_cache = NULL;
2497         unsigned long offset = 0;
2498         unsigned int nr = thp_nr_pages(head);
2499         int i;
2500
2501         /* complete memcg works before add pages to LRU */
2502         split_page_memcg(head, nr);
2503
2504         if (PageAnon(head) && PageSwapCache(head)) {
2505                 swp_entry_t entry = { .val = page_private(head) };
2506
2507                 offset = swp_offset(entry);
2508                 swap_cache = swap_address_space(entry);
2509                 xa_lock(&swap_cache->i_pages);
2510         }
2511
2512         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2513         lruvec = folio_lruvec_lock(folio);
2514
2515         ClearPageHasHWPoisoned(head);
2516
2517         for (i = nr - 1; i >= 1; i--) {
2518                 __split_huge_page_tail(head, i, lruvec, list);
2519                 /* Some pages can be beyond EOF: drop them from page cache */
2520                 if (head[i].index >= end) {
2521                         struct folio *tail = page_folio(head + i);
2522
2523                         if (shmem_mapping(head->mapping))
2524                                 shmem_uncharge(head->mapping->host, 1);
2525                         else if (folio_test_clear_dirty(tail))
2526                                 folio_account_cleaned(tail,
2527                                         inode_to_wb(folio->mapping->host));
2528                         __filemap_remove_folio(tail, NULL);
2529                         folio_put(tail);
2530                 } else if (!PageAnon(page)) {
2531                         __xa_store(&head->mapping->i_pages, head[i].index,
2532                                         head + i, 0);
2533                 } else if (swap_cache) {
2534                         __xa_store(&swap_cache->i_pages, offset + i,
2535                                         head + i, 0);
2536                 }
2537         }
2538
2539         ClearPageCompound(head);
2540         unlock_page_lruvec(lruvec);
2541         /* Caller disabled irqs, so they are still disabled here */
2542
2543         split_page_owner(head, nr);
2544
2545         /* See comment in __split_huge_page_tail() */
2546         if (PageAnon(head)) {
2547                 /* Additional pin to swap cache */
2548                 if (PageSwapCache(head)) {
2549                         page_ref_add(head, 2);
2550                         xa_unlock(&swap_cache->i_pages);
2551                 } else {
2552                         page_ref_inc(head);
2553                 }
2554         } else {
2555                 /* Additional pin to page cache */
2556                 page_ref_add(head, 2);
2557                 xa_unlock(&head->mapping->i_pages);
2558         }
2559         local_irq_enable();
2560
2561         remap_page(folio, nr);
2562
2563         if (PageSwapCache(head)) {
2564                 swp_entry_t entry = { .val = page_private(head) };
2565
2566                 split_swap_cluster(entry);
2567         }
2568
2569         for (i = 0; i < nr; i++) {
2570                 struct page *subpage = head + i;
2571                 if (subpage == page)
2572                         continue;
2573                 unlock_page(subpage);
2574
2575                 /*
2576                  * Subpages may be freed if there wasn't any mapping
2577                  * like if add_to_swap() is running on a lru page that
2578                  * had its mapping zapped. And freeing these pages
2579                  * requires taking the lru_lock so we do the put_page
2580                  * of the tail pages after the split is complete.
2581                  */
2582                 free_page_and_swap_cache(subpage);
2583         }
2584 }
2585
2586 /* Racy check whether the huge page can be split */
2587 bool can_split_folio(struct folio *folio, int *pextra_pins)
2588 {
2589         int extra_pins;
2590
2591         /* Additional pins from page cache */
2592         if (folio_test_anon(folio))
2593                 extra_pins = folio_test_swapcache(folio) ?
2594                                 folio_nr_pages(folio) : 0;
2595         else
2596                 extra_pins = folio_nr_pages(folio);
2597         if (pextra_pins)
2598                 *pextra_pins = extra_pins;
2599         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2600 }
2601
2602 /*
2603  * This function splits huge page into normal pages. @page can point to any
2604  * subpage of huge page to split. Split doesn't change the position of @page.
2605  *
2606  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2607  * The huge page must be locked.
2608  *
2609  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2610  *
2611  * Both head page and tail pages will inherit mapping, flags, and so on from
2612  * the hugepage.
2613  *
2614  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2615  * they are not mapped.
2616  *
2617  * Returns 0 if the hugepage is split successfully.
2618  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2619  * us.
2620  */
2621 int split_huge_page_to_list(struct page *page, struct list_head *list)
2622 {
2623         struct folio *folio = page_folio(page);
2624         struct deferred_split *ds_queue = get_deferred_split_queue(&folio->page);
2625         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2626         struct anon_vma *anon_vma = NULL;
2627         struct address_space *mapping = NULL;
2628         int extra_pins, ret;
2629         pgoff_t end;
2630         bool is_hzp;
2631
2632         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2633         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2634
2635         is_hzp = is_huge_zero_page(&folio->page);
2636         VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2637         if (is_hzp)
2638                 return -EBUSY;
2639
2640         if (folio_test_writeback(folio))
2641                 return -EBUSY;
2642
2643         if (folio_test_anon(folio)) {
2644                 /*
2645                  * The caller does not necessarily hold an mmap_lock that would
2646                  * prevent the anon_vma disappearing so we first we take a
2647                  * reference to it and then lock the anon_vma for write. This
2648                  * is similar to folio_lock_anon_vma_read except the write lock
2649                  * is taken to serialise against parallel split or collapse
2650                  * operations.
2651                  */
2652                 anon_vma = folio_get_anon_vma(folio);
2653                 if (!anon_vma) {
2654                         ret = -EBUSY;
2655                         goto out;
2656                 }
2657                 end = -1;
2658                 mapping = NULL;
2659                 anon_vma_lock_write(anon_vma);
2660         } else {
2661                 gfp_t gfp;
2662
2663                 mapping = folio->mapping;
2664
2665                 /* Truncated ? */
2666                 if (!mapping) {
2667                         ret = -EBUSY;
2668                         goto out;
2669                 }
2670
2671                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2672                                                         GFP_RECLAIM_MASK);
2673
2674                 if (folio_test_private(folio) &&
2675                                 !filemap_release_folio(folio, gfp)) {
2676                         ret = -EBUSY;
2677                         goto out;
2678                 }
2679
2680                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2681                 if (xas_error(&xas)) {
2682                         ret = xas_error(&xas);
2683                         goto out;
2684                 }
2685
2686                 anon_vma = NULL;
2687                 i_mmap_lock_read(mapping);
2688
2689                 /*
2690                  *__split_huge_page() may need to trim off pages beyond EOF:
2691                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2692                  * which cannot be nested inside the page tree lock. So note
2693                  * end now: i_size itself may be changed at any moment, but
2694                  * folio lock is good enough to serialize the trimming.
2695                  */
2696                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2697                 if (shmem_mapping(mapping))
2698                         end = shmem_fallocend(mapping->host, end);
2699         }
2700
2701         /*
2702          * Racy check if we can split the page, before unmap_folio() will
2703          * split PMDs
2704          */
2705         if (!can_split_folio(folio, &extra_pins)) {
2706                 ret = -EBUSY;
2707                 goto out_unlock;
2708         }
2709
2710         unmap_folio(folio);
2711
2712         /* block interrupt reentry in xa_lock and spinlock */
2713         local_irq_disable();
2714         if (mapping) {
2715                 /*
2716                  * Check if the folio is present in page cache.
2717                  * We assume all tail are present too, if folio is there.
2718                  */
2719                 xas_lock(&xas);
2720                 xas_reset(&xas);
2721                 if (xas_load(&xas) != folio)
2722                         goto fail;
2723         }
2724
2725         /* Prevent deferred_split_scan() touching ->_refcount */
2726         spin_lock(&ds_queue->split_queue_lock);
2727         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2728                 if (!list_empty(page_deferred_list(&folio->page))) {
2729                         ds_queue->split_queue_len--;
2730                         list_del(page_deferred_list(&folio->page));
2731                 }
2732                 spin_unlock(&ds_queue->split_queue_lock);
2733                 if (mapping) {
2734                         int nr = folio_nr_pages(folio);
2735
2736                         xas_split(&xas, folio, folio_order(folio));
2737                         if (folio_test_swapbacked(folio)) {
2738                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2739                                                         -nr);
2740                         } else {
2741                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2742                                                         -nr);
2743                                 filemap_nr_thps_dec(mapping);
2744                         }
2745                 }
2746
2747                 __split_huge_page(page, list, end);
2748                 ret = 0;
2749         } else {
2750                 spin_unlock(&ds_queue->split_queue_lock);
2751 fail:
2752                 if (mapping)
2753                         xas_unlock(&xas);
2754                 local_irq_enable();
2755                 remap_page(folio, folio_nr_pages(folio));
2756                 ret = -EBUSY;
2757         }
2758
2759 out_unlock:
2760         if (anon_vma) {
2761                 anon_vma_unlock_write(anon_vma);
2762                 put_anon_vma(anon_vma);
2763         }
2764         if (mapping)
2765                 i_mmap_unlock_read(mapping);
2766 out:
2767         xas_destroy(&xas);
2768         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2769         return ret;
2770 }
2771
2772 void free_transhuge_page(struct page *page)
2773 {
2774         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2775         unsigned long flags;
2776
2777         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2778         if (!list_empty(page_deferred_list(page))) {
2779                 ds_queue->split_queue_len--;
2780                 list_del(page_deferred_list(page));
2781         }
2782         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2783         free_compound_page(page);
2784 }
2785
2786 void deferred_split_huge_page(struct page *page)
2787 {
2788         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2789 #ifdef CONFIG_MEMCG
2790         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2791 #endif
2792         unsigned long flags;
2793
2794         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2795
2796         /*
2797          * The try_to_unmap() in page reclaim path might reach here too,
2798          * this may cause a race condition to corrupt deferred split queue.
2799          * And, if page reclaim is already handling the same page, it is
2800          * unnecessary to handle it again in shrinker.
2801          *
2802          * Check PageSwapCache to determine if the page is being
2803          * handled by page reclaim since THP swap would add the page into
2804          * swap cache before calling try_to_unmap().
2805          */
2806         if (PageSwapCache(page))
2807                 return;
2808
2809         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2810         if (list_empty(page_deferred_list(page))) {
2811                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2812                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2813                 ds_queue->split_queue_len++;
2814 #ifdef CONFIG_MEMCG
2815                 if (memcg)
2816                         set_shrinker_bit(memcg, page_to_nid(page),
2817                                          deferred_split_shrinker.id);
2818 #endif
2819         }
2820         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2821 }
2822
2823 static unsigned long deferred_split_count(struct shrinker *shrink,
2824                 struct shrink_control *sc)
2825 {
2826         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2827         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2828
2829 #ifdef CONFIG_MEMCG
2830         if (sc->memcg)
2831                 ds_queue = &sc->memcg->deferred_split_queue;
2832 #endif
2833         return READ_ONCE(ds_queue->split_queue_len);
2834 }
2835
2836 static unsigned long deferred_split_scan(struct shrinker *shrink,
2837                 struct shrink_control *sc)
2838 {
2839         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2840         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2841         unsigned long flags;
2842         LIST_HEAD(list), *pos, *next;
2843         struct page *page;
2844         int split = 0;
2845
2846 #ifdef CONFIG_MEMCG
2847         if (sc->memcg)
2848                 ds_queue = &sc->memcg->deferred_split_queue;
2849 #endif
2850
2851         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2852         /* Take pin on all head pages to avoid freeing them under us */
2853         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2854                 page = list_entry((void *)pos, struct page, deferred_list);
2855                 page = compound_head(page);
2856                 if (get_page_unless_zero(page)) {
2857                         list_move(page_deferred_list(page), &list);
2858                 } else {
2859                         /* We lost race with put_compound_page() */
2860                         list_del_init(page_deferred_list(page));
2861                         ds_queue->split_queue_len--;
2862                 }
2863                 if (!--sc->nr_to_scan)
2864                         break;
2865         }
2866         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2867
2868         list_for_each_safe(pos, next, &list) {
2869                 page = list_entry((void *)pos, struct page, deferred_list);
2870                 if (!trylock_page(page))
2871                         goto next;
2872                 /* split_huge_page() removes page from list on success */
2873                 if (!split_huge_page(page))
2874                         split++;
2875                 unlock_page(page);
2876 next:
2877                 put_page(page);
2878         }
2879
2880         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2881         list_splice_tail(&list, &ds_queue->split_queue);
2882         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2883
2884         /*
2885          * Stop shrinker if we didn't split any page, but the queue is empty.
2886          * This can happen if pages were freed under us.
2887          */
2888         if (!split && list_empty(&ds_queue->split_queue))
2889                 return SHRINK_STOP;
2890         return split;
2891 }
2892
2893 static struct shrinker deferred_split_shrinker = {
2894         .count_objects = deferred_split_count,
2895         .scan_objects = deferred_split_scan,
2896         .seeks = DEFAULT_SEEKS,
2897         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2898                  SHRINKER_NONSLAB,
2899 };
2900
2901 #ifdef CONFIG_DEBUG_FS
2902 static void split_huge_pages_all(void)
2903 {
2904         struct zone *zone;
2905         struct page *page;
2906         unsigned long pfn, max_zone_pfn;
2907         unsigned long total = 0, split = 0;
2908
2909         pr_debug("Split all THPs\n");
2910         for_each_zone(zone) {
2911                 if (!managed_zone(zone))
2912                         continue;
2913                 max_zone_pfn = zone_end_pfn(zone);
2914                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2915                         int nr_pages;
2916
2917                         page = pfn_to_online_page(pfn);
2918                         if (!page || !get_page_unless_zero(page))
2919                                 continue;
2920
2921                         if (zone != page_zone(page))
2922                                 goto next;
2923
2924                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2925                                 goto next;
2926
2927                         total++;
2928                         lock_page(page);
2929                         nr_pages = thp_nr_pages(page);
2930                         if (!split_huge_page(page))
2931                                 split++;
2932                         pfn += nr_pages - 1;
2933                         unlock_page(page);
2934 next:
2935                         put_page(page);
2936                         cond_resched();
2937                 }
2938         }
2939
2940         pr_debug("%lu of %lu THP split\n", split, total);
2941 }
2942
2943 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2944 {
2945         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2946                     is_vm_hugetlb_page(vma);
2947 }
2948
2949 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2950                                 unsigned long vaddr_end)
2951 {
2952         int ret = 0;
2953         struct task_struct *task;
2954         struct mm_struct *mm;
2955         unsigned long total = 0, split = 0;
2956         unsigned long addr;
2957
2958         vaddr_start &= PAGE_MASK;
2959         vaddr_end &= PAGE_MASK;
2960
2961         /* Find the task_struct from pid */
2962         rcu_read_lock();
2963         task = find_task_by_vpid(pid);
2964         if (!task) {
2965                 rcu_read_unlock();
2966                 ret = -ESRCH;
2967                 goto out;
2968         }
2969         get_task_struct(task);
2970         rcu_read_unlock();
2971
2972         /* Find the mm_struct */
2973         mm = get_task_mm(task);
2974         put_task_struct(task);
2975
2976         if (!mm) {
2977                 ret = -EINVAL;
2978                 goto out;
2979         }
2980
2981         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2982                  pid, vaddr_start, vaddr_end);
2983
2984         mmap_read_lock(mm);
2985         /*
2986          * always increase addr by PAGE_SIZE, since we could have a PTE page
2987          * table filled with PTE-mapped THPs, each of which is distinct.
2988          */
2989         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2990                 struct vm_area_struct *vma = vma_lookup(mm, addr);
2991                 struct page *page;
2992
2993                 if (!vma)
2994                         break;
2995
2996                 /* skip special VMA and hugetlb VMA */
2997                 if (vma_not_suitable_for_thp_split(vma)) {
2998                         addr = vma->vm_end;
2999                         continue;
3000                 }
3001
3002                 /* FOLL_DUMP to ignore special (like zero) pages */
3003                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3004
3005                 if (IS_ERR_OR_NULL(page))
3006                         continue;
3007
3008                 if (!is_transparent_hugepage(page))
3009                         goto next;
3010
3011                 total++;
3012                 if (!can_split_folio(page_folio(page), NULL))
3013                         goto next;
3014
3015                 if (!trylock_page(page))
3016                         goto next;
3017
3018                 if (!split_huge_page(page))
3019                         split++;
3020
3021                 unlock_page(page);
3022 next:
3023                 put_page(page);
3024                 cond_resched();
3025         }
3026         mmap_read_unlock(mm);
3027         mmput(mm);
3028
3029         pr_debug("%lu of %lu THP split\n", split, total);
3030
3031 out:
3032         return ret;
3033 }
3034
3035 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3036                                 pgoff_t off_end)
3037 {
3038         struct filename *file;
3039         struct file *candidate;
3040         struct address_space *mapping;
3041         int ret = -EINVAL;
3042         pgoff_t index;
3043         int nr_pages = 1;
3044         unsigned long total = 0, split = 0;
3045
3046         file = getname_kernel(file_path);
3047         if (IS_ERR(file))
3048                 return ret;
3049
3050         candidate = file_open_name(file, O_RDONLY, 0);
3051         if (IS_ERR(candidate))
3052                 goto out;
3053
3054         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3055                  file_path, off_start, off_end);
3056
3057         mapping = candidate->f_mapping;
3058
3059         for (index = off_start; index < off_end; index += nr_pages) {
3060                 struct page *fpage = pagecache_get_page(mapping, index,
3061                                                 FGP_ENTRY | FGP_HEAD, 0);
3062
3063                 nr_pages = 1;
3064                 if (xa_is_value(fpage) || !fpage)
3065                         continue;
3066
3067                 if (!is_transparent_hugepage(fpage))
3068                         goto next;
3069
3070                 total++;
3071                 nr_pages = thp_nr_pages(fpage);
3072
3073                 if (!trylock_page(fpage))
3074                         goto next;
3075
3076                 if (!split_huge_page(fpage))
3077                         split++;
3078
3079                 unlock_page(fpage);
3080 next:
3081                 put_page(fpage);
3082                 cond_resched();
3083         }
3084
3085         filp_close(candidate, NULL);
3086         ret = 0;
3087
3088         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3089 out:
3090         putname(file);
3091         return ret;
3092 }
3093
3094 #define MAX_INPUT_BUF_SZ 255
3095
3096 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3097                                 size_t count, loff_t *ppops)
3098 {
3099         static DEFINE_MUTEX(split_debug_mutex);
3100         ssize_t ret;
3101         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3102         char input_buf[MAX_INPUT_BUF_SZ];
3103         int pid;
3104         unsigned long vaddr_start, vaddr_end;
3105
3106         ret = mutex_lock_interruptible(&split_debug_mutex);
3107         if (ret)
3108                 return ret;
3109
3110         ret = -EFAULT;
3111
3112         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3113         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3114                 goto out;
3115
3116         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3117
3118         if (input_buf[0] == '/') {
3119                 char *tok;
3120                 char *buf = input_buf;
3121                 char file_path[MAX_INPUT_BUF_SZ];
3122                 pgoff_t off_start = 0, off_end = 0;
3123                 size_t input_len = strlen(input_buf);
3124
3125                 tok = strsep(&buf, ",");
3126                 if (tok) {
3127                         strcpy(file_path, tok);
3128                 } else {
3129                         ret = -EINVAL;
3130                         goto out;
3131                 }
3132
3133                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3134                 if (ret != 2) {
3135                         ret = -EINVAL;
3136                         goto out;
3137                 }
3138                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3139                 if (!ret)
3140                         ret = input_len;
3141
3142                 goto out;
3143         }
3144
3145         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3146         if (ret == 1 && pid == 1) {
3147                 split_huge_pages_all();
3148                 ret = strlen(input_buf);
3149                 goto out;
3150         } else if (ret != 3) {
3151                 ret = -EINVAL;
3152                 goto out;
3153         }
3154
3155         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3156         if (!ret)
3157                 ret = strlen(input_buf);
3158 out:
3159         mutex_unlock(&split_debug_mutex);
3160         return ret;
3161
3162 }
3163
3164 static const struct file_operations split_huge_pages_fops = {
3165         .owner   = THIS_MODULE,
3166         .write   = split_huge_pages_write,
3167         .llseek  = no_llseek,
3168 };
3169
3170 static int __init split_huge_pages_debugfs(void)
3171 {
3172         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3173                             &split_huge_pages_fops);
3174         return 0;
3175 }
3176 late_initcall(split_huge_pages_debugfs);
3177 #endif
3178
3179 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3180 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3181                 struct page *page)
3182 {
3183         struct vm_area_struct *vma = pvmw->vma;
3184         struct mm_struct *mm = vma->vm_mm;
3185         unsigned long address = pvmw->address;
3186         bool anon_exclusive;
3187         pmd_t pmdval;
3188         swp_entry_t entry;
3189         pmd_t pmdswp;
3190
3191         if (!(pvmw->pmd && !pvmw->pte))
3192                 return 0;
3193
3194         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3195         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3196
3197         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3198         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3199         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3200                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3201                 return -EBUSY;
3202         }
3203
3204         if (pmd_dirty(pmdval))
3205                 set_page_dirty(page);
3206         if (pmd_write(pmdval))
3207                 entry = make_writable_migration_entry(page_to_pfn(page));
3208         else if (anon_exclusive)
3209                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3210         else
3211                 entry = make_readable_migration_entry(page_to_pfn(page));
3212         if (pmd_young(pmdval))
3213                 entry = make_migration_entry_young(entry);
3214         if (pmd_dirty(pmdval))
3215                 entry = make_migration_entry_dirty(entry);
3216         pmdswp = swp_entry_to_pmd(entry);
3217         if (pmd_soft_dirty(pmdval))
3218                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3219         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3220         page_remove_rmap(page, vma, true);
3221         put_page(page);
3222         trace_set_migration_pmd(address, pmd_val(pmdswp));
3223
3224         return 0;
3225 }
3226
3227 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3228 {
3229         struct vm_area_struct *vma = pvmw->vma;
3230         struct mm_struct *mm = vma->vm_mm;
3231         unsigned long address = pvmw->address;
3232         unsigned long haddr = address & HPAGE_PMD_MASK;
3233         pmd_t pmde;
3234         swp_entry_t entry;
3235
3236         if (!(pvmw->pmd && !pvmw->pte))
3237                 return;
3238
3239         entry = pmd_to_swp_entry(*pvmw->pmd);
3240         get_page(new);
3241         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3242         if (pmd_swp_soft_dirty(*pvmw->pmd))
3243                 pmde = pmd_mksoft_dirty(pmde);
3244         if (is_writable_migration_entry(entry))
3245                 pmde = maybe_pmd_mkwrite(pmde, vma);
3246         if (pmd_swp_uffd_wp(*pvmw->pmd))
3247                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3248         if (!is_migration_entry_young(entry))
3249                 pmde = pmd_mkold(pmde);
3250         /* NOTE: this may contain setting soft-dirty on some archs */
3251         if (PageDirty(new) && is_migration_entry_dirty(entry))
3252                 pmde = pmd_mkdirty(pmde);
3253
3254         if (PageAnon(new)) {
3255                 rmap_t rmap_flags = RMAP_COMPOUND;
3256
3257                 if (!is_readable_migration_entry(entry))
3258                         rmap_flags |= RMAP_EXCLUSIVE;
3259
3260                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3261         } else {
3262                 page_add_file_rmap(new, vma, true);
3263         }
3264         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3265         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3266
3267         /* No need to invalidate - it was non-present before */
3268         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3269         trace_remove_migration_pmd(address, pmd_val(pmde));
3270 }
3271 #endif