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         int ret;
1039
1040         assert_spin_locked(pmd_lockptr(mm, pmd));
1041
1042         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1043         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1044                          (FOLL_PIN | FOLL_GET)))
1045                 return NULL;
1046
1047         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1048                 return NULL;
1049
1050         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1051                 /* pass */;
1052         else
1053                 return NULL;
1054
1055         if (flags & FOLL_TOUCH)
1056                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1057
1058         /*
1059          * device mapped pages can only be returned if the
1060          * caller will manage the page reference count.
1061          */
1062         if (!(flags & (FOLL_GET | FOLL_PIN)))
1063                 return ERR_PTR(-EEXIST);
1064
1065         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1066         *pgmap = get_dev_pagemap(pfn, *pgmap);
1067         if (!*pgmap)
1068                 return ERR_PTR(-EFAULT);
1069         page = pfn_to_page(pfn);
1070         ret = try_grab_page(page, flags);
1071         if (ret)
1072                 page = ERR_PTR(ret);
1073
1074         return page;
1075 }
1076
1077 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1078                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1079                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1080 {
1081         spinlock_t *dst_ptl, *src_ptl;
1082         struct page *src_page;
1083         pmd_t pmd;
1084         pgtable_t pgtable = NULL;
1085         int ret = -ENOMEM;
1086
1087         /* Skip if can be re-fill on fault */
1088         if (!vma_is_anonymous(dst_vma))
1089                 return 0;
1090
1091         pgtable = pte_alloc_one(dst_mm);
1092         if (unlikely(!pgtable))
1093                 goto out;
1094
1095         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1096         src_ptl = pmd_lockptr(src_mm, src_pmd);
1097         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1098
1099         ret = -EAGAIN;
1100         pmd = *src_pmd;
1101
1102 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1103         if (unlikely(is_swap_pmd(pmd))) {
1104                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1105
1106                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1107                 if (!is_readable_migration_entry(entry)) {
1108                         entry = make_readable_migration_entry(
1109                                                         swp_offset(entry));
1110                         pmd = swp_entry_to_pmd(entry);
1111                         if (pmd_swp_soft_dirty(*src_pmd))
1112                                 pmd = pmd_swp_mksoft_dirty(pmd);
1113                         if (pmd_swp_uffd_wp(*src_pmd))
1114                                 pmd = pmd_swp_mkuffd_wp(pmd);
1115                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1116                 }
1117                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118                 mm_inc_nr_ptes(dst_mm);
1119                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1120                 if (!userfaultfd_wp(dst_vma))
1121                         pmd = pmd_swp_clear_uffd_wp(pmd);
1122                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1123                 ret = 0;
1124                 goto out_unlock;
1125         }
1126 #endif
1127
1128         if (unlikely(!pmd_trans_huge(pmd))) {
1129                 pte_free(dst_mm, pgtable);
1130                 goto out_unlock;
1131         }
1132         /*
1133          * When page table lock is held, the huge zero pmd should not be
1134          * under splitting since we don't split the page itself, only pmd to
1135          * a page table.
1136          */
1137         if (is_huge_zero_pmd(pmd)) {
1138                 /*
1139                  * get_huge_zero_page() will never allocate a new page here,
1140                  * since we already have a zero page to copy. It just takes a
1141                  * reference.
1142                  */
1143                 mm_get_huge_zero_page(dst_mm);
1144                 goto out_zero_page;
1145         }
1146
1147         src_page = pmd_page(pmd);
1148         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1149
1150         get_page(src_page);
1151         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1152                 /* Page maybe pinned: split and retry the fault on PTEs. */
1153                 put_page(src_page);
1154                 pte_free(dst_mm, pgtable);
1155                 spin_unlock(src_ptl);
1156                 spin_unlock(dst_ptl);
1157                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1158                 return -EAGAIN;
1159         }
1160         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1161 out_zero_page:
1162         mm_inc_nr_ptes(dst_mm);
1163         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1164         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1165         if (!userfaultfd_wp(dst_vma))
1166                 pmd = pmd_clear_uffd_wp(pmd);
1167         pmd = pmd_mkold(pmd_wrprotect(pmd));
1168         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1169
1170         ret = 0;
1171 out_unlock:
1172         spin_unlock(src_ptl);
1173         spin_unlock(dst_ptl);
1174 out:
1175         return ret;
1176 }
1177
1178 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1179 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1180                       pud_t *pud, bool write)
1181 {
1182         pud_t _pud;
1183
1184         _pud = pud_mkyoung(*pud);
1185         if (write)
1186                 _pud = pud_mkdirty(_pud);
1187         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1188                                   pud, _pud, write))
1189                 update_mmu_cache_pud(vma, addr, pud);
1190 }
1191
1192 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1193                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1194 {
1195         unsigned long pfn = pud_pfn(*pud);
1196         struct mm_struct *mm = vma->vm_mm;
1197         struct page *page;
1198         int ret;
1199
1200         assert_spin_locked(pud_lockptr(mm, pud));
1201
1202         if (flags & FOLL_WRITE && !pud_write(*pud))
1203                 return NULL;
1204
1205         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1206         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1207                          (FOLL_PIN | FOLL_GET)))
1208                 return NULL;
1209
1210         if (pud_present(*pud) && pud_devmap(*pud))
1211                 /* pass */;
1212         else
1213                 return NULL;
1214
1215         if (flags & FOLL_TOUCH)
1216                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1217
1218         /*
1219          * device mapped pages can only be returned if the
1220          * caller will manage the page reference count.
1221          *
1222          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1223          */
1224         if (!(flags & (FOLL_GET | FOLL_PIN)))
1225                 return ERR_PTR(-EEXIST);
1226
1227         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1228         *pgmap = get_dev_pagemap(pfn, *pgmap);
1229         if (!*pgmap)
1230                 return ERR_PTR(-EFAULT);
1231         page = pfn_to_page(pfn);
1232
1233         ret = try_grab_page(page, flags);
1234         if (ret)
1235                 page = ERR_PTR(ret);
1236
1237         return page;
1238 }
1239
1240 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1241                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1242                   struct vm_area_struct *vma)
1243 {
1244         spinlock_t *dst_ptl, *src_ptl;
1245         pud_t pud;
1246         int ret;
1247
1248         dst_ptl = pud_lock(dst_mm, dst_pud);
1249         src_ptl = pud_lockptr(src_mm, src_pud);
1250         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1251
1252         ret = -EAGAIN;
1253         pud = *src_pud;
1254         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1255                 goto out_unlock;
1256
1257         /*
1258          * When page table lock is held, the huge zero pud should not be
1259          * under splitting since we don't split the page itself, only pud to
1260          * a page table.
1261          */
1262         if (is_huge_zero_pud(pud)) {
1263                 /* No huge zero pud yet */
1264         }
1265
1266         /*
1267          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1268          * and split if duplicating fails.
1269          */
1270         pudp_set_wrprotect(src_mm, addr, src_pud);
1271         pud = pud_mkold(pud_wrprotect(pud));
1272         set_pud_at(dst_mm, addr, dst_pud, pud);
1273
1274         ret = 0;
1275 out_unlock:
1276         spin_unlock(src_ptl);
1277         spin_unlock(dst_ptl);
1278         return ret;
1279 }
1280
1281 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1282 {
1283         bool write = vmf->flags & FAULT_FLAG_WRITE;
1284
1285         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1286         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1287                 goto unlock;
1288
1289         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1290 unlock:
1291         spin_unlock(vmf->ptl);
1292 }
1293 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1294
1295 void huge_pmd_set_accessed(struct vm_fault *vmf)
1296 {
1297         bool write = vmf->flags & FAULT_FLAG_WRITE;
1298
1299         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1300         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1301                 goto unlock;
1302
1303         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1304
1305 unlock:
1306         spin_unlock(vmf->ptl);
1307 }
1308
1309 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1310 {
1311         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1312         struct vm_area_struct *vma = vmf->vma;
1313         struct folio *folio;
1314         struct page *page;
1315         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1316         pmd_t orig_pmd = vmf->orig_pmd;
1317
1318         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1319         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1320
1321         VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE));
1322         VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE));
1323
1324         if (is_huge_zero_pmd(orig_pmd))
1325                 goto fallback;
1326
1327         spin_lock(vmf->ptl);
1328
1329         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1330                 spin_unlock(vmf->ptl);
1331                 return 0;
1332         }
1333
1334         page = pmd_page(orig_pmd);
1335         folio = page_folio(page);
1336         VM_BUG_ON_PAGE(!PageHead(page), page);
1337
1338         /* Early check when only holding the PT lock. */
1339         if (PageAnonExclusive(page))
1340                 goto reuse;
1341
1342         if (!folio_trylock(folio)) {
1343                 folio_get(folio);
1344                 spin_unlock(vmf->ptl);
1345                 folio_lock(folio);
1346                 spin_lock(vmf->ptl);
1347                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1348                         spin_unlock(vmf->ptl);
1349                         folio_unlock(folio);
1350                         folio_put(folio);
1351                         return 0;
1352                 }
1353                 folio_put(folio);
1354         }
1355
1356         /* Recheck after temporarily dropping the PT lock. */
1357         if (PageAnonExclusive(page)) {
1358                 folio_unlock(folio);
1359                 goto reuse;
1360         }
1361
1362         /*
1363          * See do_wp_page(): we can only reuse the folio exclusively if
1364          * there are no additional references. Note that we always drain
1365          * the LRU pagevecs immediately after adding a THP.
1366          */
1367         if (folio_ref_count(folio) >
1368                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1369                 goto unlock_fallback;
1370         if (folio_test_swapcache(folio))
1371                 folio_free_swap(folio);
1372         if (folio_ref_count(folio) == 1) {
1373                 pmd_t entry;
1374
1375                 page_move_anon_rmap(page, vma);
1376                 folio_unlock(folio);
1377 reuse:
1378                 if (unlikely(unshare)) {
1379                         spin_unlock(vmf->ptl);
1380                         return 0;
1381                 }
1382                 entry = pmd_mkyoung(orig_pmd);
1383                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1384                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1385                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1386                 spin_unlock(vmf->ptl);
1387                 return VM_FAULT_WRITE;
1388         }
1389
1390 unlock_fallback:
1391         folio_unlock(folio);
1392         spin_unlock(vmf->ptl);
1393 fallback:
1394         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1395         return VM_FAULT_FALLBACK;
1396 }
1397
1398 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1399 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1400                                         struct vm_area_struct *vma,
1401                                         unsigned int flags)
1402 {
1403         /* If the pmd is writable, we can write to the page. */
1404         if (pmd_write(pmd))
1405                 return true;
1406
1407         /* Maybe FOLL_FORCE is set to override it? */
1408         if (!(flags & FOLL_FORCE))
1409                 return false;
1410
1411         /* But FOLL_FORCE has no effect on shared mappings */
1412         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1413                 return false;
1414
1415         /* ... or read-only private ones */
1416         if (!(vma->vm_flags & VM_MAYWRITE))
1417                 return false;
1418
1419         /* ... or already writable ones that just need to take a write fault */
1420         if (vma->vm_flags & VM_WRITE)
1421                 return false;
1422
1423         /*
1424          * See can_change_pte_writable(): we broke COW and could map the page
1425          * writable if we have an exclusive anonymous page ...
1426          */
1427         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1428                 return false;
1429
1430         /* ... and a write-fault isn't required for other reasons. */
1431         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1432                 return false;
1433         return !userfaultfd_huge_pmd_wp(vma, pmd);
1434 }
1435
1436 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1437                                    unsigned long addr,
1438                                    pmd_t *pmd,
1439                                    unsigned int flags)
1440 {
1441         struct mm_struct *mm = vma->vm_mm;
1442         struct page *page;
1443         int ret;
1444
1445         assert_spin_locked(pmd_lockptr(mm, pmd));
1446
1447         page = pmd_page(*pmd);
1448         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1449
1450         if ((flags & FOLL_WRITE) &&
1451             !can_follow_write_pmd(*pmd, page, vma, flags))
1452                 return NULL;
1453
1454         /* Avoid dumping huge zero page */
1455         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1456                 return ERR_PTR(-EFAULT);
1457
1458         /* Full NUMA hinting faults to serialise migration in fault paths */
1459         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1460                 return NULL;
1461
1462         if (!pmd_write(*pmd) && gup_must_unshare(flags, page))
1463                 return ERR_PTR(-EMLINK);
1464
1465         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1466                         !PageAnonExclusive(page), page);
1467
1468         ret = try_grab_page(page, flags);
1469         if (ret)
1470                 return ERR_PTR(ret);
1471
1472         if (flags & FOLL_TOUCH)
1473                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1474
1475         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1476         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1477
1478         return page;
1479 }
1480
1481 /* NUMA hinting page fault entry point for trans huge pmds */
1482 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1483 {
1484         struct vm_area_struct *vma = vmf->vma;
1485         pmd_t oldpmd = vmf->orig_pmd;
1486         pmd_t pmd;
1487         struct page *page;
1488         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1489         int page_nid = NUMA_NO_NODE;
1490         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1491         bool migrated = false;
1492         bool was_writable = pmd_savedwrite(oldpmd);
1493         int flags = 0;
1494
1495         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1496         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1497                 spin_unlock(vmf->ptl);
1498                 goto out;
1499         }
1500
1501         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1502         page = vm_normal_page_pmd(vma, haddr, pmd);
1503         if (!page)
1504                 goto out_map;
1505
1506         /* See similar comment in do_numa_page for explanation */
1507         if (!was_writable)
1508                 flags |= TNF_NO_GROUP;
1509
1510         page_nid = page_to_nid(page);
1511         /*
1512          * For memory tiering mode, cpupid of slow memory page is used
1513          * to record page access time.  So use default value.
1514          */
1515         if (node_is_toptier(page_nid))
1516                 last_cpupid = page_cpupid_last(page);
1517         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1518                                        &flags);
1519
1520         if (target_nid == NUMA_NO_NODE) {
1521                 put_page(page);
1522                 goto out_map;
1523         }
1524
1525         spin_unlock(vmf->ptl);
1526
1527         migrated = migrate_misplaced_page(page, vma, target_nid);
1528         if (migrated) {
1529                 flags |= TNF_MIGRATED;
1530                 page_nid = target_nid;
1531         } else {
1532                 flags |= TNF_MIGRATE_FAIL;
1533                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1534                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1535                         spin_unlock(vmf->ptl);
1536                         goto out;
1537                 }
1538                 goto out_map;
1539         }
1540
1541 out:
1542         if (page_nid != NUMA_NO_NODE)
1543                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1544                                 flags);
1545
1546         return 0;
1547
1548 out_map:
1549         /* Restore the PMD */
1550         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1551         pmd = pmd_mkyoung(pmd);
1552         if (was_writable)
1553                 pmd = pmd_mkwrite(pmd);
1554         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1555         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1556         spin_unlock(vmf->ptl);
1557         goto out;
1558 }
1559
1560 /*
1561  * Return true if we do MADV_FREE successfully on entire pmd page.
1562  * Otherwise, return false.
1563  */
1564 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1565                 pmd_t *pmd, unsigned long addr, unsigned long next)
1566 {
1567         spinlock_t *ptl;
1568         pmd_t orig_pmd;
1569         struct page *page;
1570         struct mm_struct *mm = tlb->mm;
1571         bool ret = false;
1572
1573         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1574
1575         ptl = pmd_trans_huge_lock(pmd, vma);
1576         if (!ptl)
1577                 goto out_unlocked;
1578
1579         orig_pmd = *pmd;
1580         if (is_huge_zero_pmd(orig_pmd))
1581                 goto out;
1582
1583         if (unlikely(!pmd_present(orig_pmd))) {
1584                 VM_BUG_ON(thp_migration_supported() &&
1585                                   !is_pmd_migration_entry(orig_pmd));
1586                 goto out;
1587         }
1588
1589         page = pmd_page(orig_pmd);
1590         /*
1591          * If other processes are mapping this page, we couldn't discard
1592          * the page unless they all do MADV_FREE so let's skip the page.
1593          */
1594         if (total_mapcount(page) != 1)
1595                 goto out;
1596
1597         if (!trylock_page(page))
1598                 goto out;
1599
1600         /*
1601          * If user want to discard part-pages of THP, split it so MADV_FREE
1602          * will deactivate only them.
1603          */
1604         if (next - addr != HPAGE_PMD_SIZE) {
1605                 get_page(page);
1606                 spin_unlock(ptl);
1607                 split_huge_page(page);
1608                 unlock_page(page);
1609                 put_page(page);
1610                 goto out_unlocked;
1611         }
1612
1613         if (PageDirty(page))
1614                 ClearPageDirty(page);
1615         unlock_page(page);
1616
1617         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1618                 pmdp_invalidate(vma, addr, pmd);
1619                 orig_pmd = pmd_mkold(orig_pmd);
1620                 orig_pmd = pmd_mkclean(orig_pmd);
1621
1622                 set_pmd_at(mm, addr, pmd, orig_pmd);
1623                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1624         }
1625
1626         mark_page_lazyfree(page);
1627         ret = true;
1628 out:
1629         spin_unlock(ptl);
1630 out_unlocked:
1631         return ret;
1632 }
1633
1634 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1635 {
1636         pgtable_t pgtable;
1637
1638         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1639         pte_free(mm, pgtable);
1640         mm_dec_nr_ptes(mm);
1641 }
1642
1643 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1644                  pmd_t *pmd, unsigned long addr)
1645 {
1646         pmd_t orig_pmd;
1647         spinlock_t *ptl;
1648
1649         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1650
1651         ptl = __pmd_trans_huge_lock(pmd, vma);
1652         if (!ptl)
1653                 return 0;
1654         /*
1655          * For architectures like ppc64 we look at deposited pgtable
1656          * when calling pmdp_huge_get_and_clear. So do the
1657          * pgtable_trans_huge_withdraw after finishing pmdp related
1658          * operations.
1659          */
1660         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1661                                                 tlb->fullmm);
1662         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1663         if (vma_is_special_huge(vma)) {
1664                 if (arch_needs_pgtable_deposit())
1665                         zap_deposited_table(tlb->mm, pmd);
1666                 spin_unlock(ptl);
1667         } else if (is_huge_zero_pmd(orig_pmd)) {
1668                 zap_deposited_table(tlb->mm, pmd);
1669                 spin_unlock(ptl);
1670         } else {
1671                 struct page *page = NULL;
1672                 int flush_needed = 1;
1673
1674                 if (pmd_present(orig_pmd)) {
1675                         page = pmd_page(orig_pmd);
1676                         page_remove_rmap(page, vma, true);
1677                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1678                         VM_BUG_ON_PAGE(!PageHead(page), page);
1679                 } else if (thp_migration_supported()) {
1680                         swp_entry_t entry;
1681
1682                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1683                         entry = pmd_to_swp_entry(orig_pmd);
1684                         page = pfn_swap_entry_to_page(entry);
1685                         flush_needed = 0;
1686                 } else
1687                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1688
1689                 if (PageAnon(page)) {
1690                         zap_deposited_table(tlb->mm, pmd);
1691                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1692                 } else {
1693                         if (arch_needs_pgtable_deposit())
1694                                 zap_deposited_table(tlb->mm, pmd);
1695                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1696                 }
1697
1698                 spin_unlock(ptl);
1699                 if (flush_needed)
1700                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1701         }
1702         return 1;
1703 }
1704
1705 #ifndef pmd_move_must_withdraw
1706 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1707                                          spinlock_t *old_pmd_ptl,
1708                                          struct vm_area_struct *vma)
1709 {
1710         /*
1711          * With split pmd lock we also need to move preallocated
1712          * PTE page table if new_pmd is on different PMD page table.
1713          *
1714          * We also don't deposit and withdraw tables for file pages.
1715          */
1716         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1717 }
1718 #endif
1719
1720 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1721 {
1722 #ifdef CONFIG_MEM_SOFT_DIRTY
1723         if (unlikely(is_pmd_migration_entry(pmd)))
1724                 pmd = pmd_swp_mksoft_dirty(pmd);
1725         else if (pmd_present(pmd))
1726                 pmd = pmd_mksoft_dirty(pmd);
1727 #endif
1728         return pmd;
1729 }
1730
1731 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1732                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1733 {
1734         spinlock_t *old_ptl, *new_ptl;
1735         pmd_t pmd;
1736         struct mm_struct *mm = vma->vm_mm;
1737         bool force_flush = false;
1738
1739         /*
1740          * The destination pmd shouldn't be established, free_pgtables()
1741          * should have release it.
1742          */
1743         if (WARN_ON(!pmd_none(*new_pmd))) {
1744                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1745                 return false;
1746         }
1747
1748         /*
1749          * We don't have to worry about the ordering of src and dst
1750          * ptlocks because exclusive mmap_lock prevents deadlock.
1751          */
1752         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1753         if (old_ptl) {
1754                 new_ptl = pmd_lockptr(mm, new_pmd);
1755                 if (new_ptl != old_ptl)
1756                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1757                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1758                 if (pmd_present(pmd))
1759                         force_flush = true;
1760                 VM_BUG_ON(!pmd_none(*new_pmd));
1761
1762                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1763                         pgtable_t pgtable;
1764                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1765                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1766                 }
1767                 pmd = move_soft_dirty_pmd(pmd);
1768                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1769                 if (force_flush)
1770                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1771                 if (new_ptl != old_ptl)
1772                         spin_unlock(new_ptl);
1773                 spin_unlock(old_ptl);
1774                 return true;
1775         }
1776         return false;
1777 }
1778
1779 /*
1780  * Returns
1781  *  - 0 if PMD could not be locked
1782  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1783  *      or if prot_numa but THP migration is not supported
1784  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1785  */
1786 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1787                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1788                     unsigned long cp_flags)
1789 {
1790         struct mm_struct *mm = vma->vm_mm;
1791         spinlock_t *ptl;
1792         pmd_t oldpmd, entry;
1793         bool preserve_write;
1794         int ret;
1795         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1796         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1797         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1798
1799         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1800
1801         if (prot_numa && !thp_migration_supported())
1802                 return 1;
1803
1804         ptl = __pmd_trans_huge_lock(pmd, vma);
1805         if (!ptl)
1806                 return 0;
1807
1808         preserve_write = prot_numa && pmd_write(*pmd);
1809         ret = 1;
1810
1811 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1812         if (is_swap_pmd(*pmd)) {
1813                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1814                 struct page *page = pfn_swap_entry_to_page(entry);
1815
1816                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1817                 if (is_writable_migration_entry(entry)) {
1818                         pmd_t newpmd;
1819                         /*
1820                          * A protection check is difficult so
1821                          * just be safe and disable write
1822                          */
1823                         if (PageAnon(page))
1824                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1825                         else
1826                                 entry = make_readable_migration_entry(swp_offset(entry));
1827                         newpmd = swp_entry_to_pmd(entry);
1828                         if (pmd_swp_soft_dirty(*pmd))
1829                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1830                         if (pmd_swp_uffd_wp(*pmd))
1831                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1832                         set_pmd_at(mm, addr, pmd, newpmd);
1833                 }
1834                 goto unlock;
1835         }
1836 #endif
1837
1838         if (prot_numa) {
1839                 struct page *page;
1840                 bool toptier;
1841                 /*
1842                  * Avoid trapping faults against the zero page. The read-only
1843                  * data is likely to be read-cached on the local CPU and
1844                  * local/remote hits to the zero page are not interesting.
1845                  */
1846                 if (is_huge_zero_pmd(*pmd))
1847                         goto unlock;
1848
1849                 if (pmd_protnone(*pmd))
1850                         goto unlock;
1851
1852                 page = pmd_page(*pmd);
1853                 toptier = node_is_toptier(page_to_nid(page));
1854                 /*
1855                  * Skip scanning top tier node if normal numa
1856                  * balancing is disabled
1857                  */
1858                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1859                     toptier)
1860                         goto unlock;
1861
1862                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1863                     !toptier)
1864                         xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1865         }
1866         /*
1867          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1868          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1869          * which is also under mmap_read_lock(mm):
1870          *
1871          *      CPU0:                           CPU1:
1872          *                              change_huge_pmd(prot_numa=1)
1873          *                               pmdp_huge_get_and_clear_notify()
1874          * madvise_dontneed()
1875          *  zap_pmd_range()
1876          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1877          *   // skip the pmd
1878          *                               set_pmd_at();
1879          *                               // pmd is re-established
1880          *
1881          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1882          * which may break userspace.
1883          *
1884          * pmdp_invalidate_ad() is required to make sure we don't miss
1885          * dirty/young flags set by hardware.
1886          */
1887         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1888
1889         entry = pmd_modify(oldpmd, newprot);
1890         if (preserve_write)
1891                 entry = pmd_mk_savedwrite(entry);
1892         if (uffd_wp) {
1893                 entry = pmd_wrprotect(entry);
1894                 entry = pmd_mkuffd_wp(entry);
1895         } else if (uffd_wp_resolve) {
1896                 /*
1897                  * Leave the write bit to be handled by PF interrupt
1898                  * handler, then things like COW could be properly
1899                  * handled.
1900                  */
1901                 entry = pmd_clear_uffd_wp(entry);
1902         }
1903         ret = HPAGE_PMD_NR;
1904         set_pmd_at(mm, addr, pmd, entry);
1905
1906         if (huge_pmd_needs_flush(oldpmd, entry))
1907                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1908
1909         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1910 unlock:
1911         spin_unlock(ptl);
1912         return ret;
1913 }
1914
1915 /*
1916  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1917  *
1918  * Note that if it returns page table lock pointer, this routine returns without
1919  * unlocking page table lock. So callers must unlock it.
1920  */
1921 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1922 {
1923         spinlock_t *ptl;
1924         ptl = pmd_lock(vma->vm_mm, pmd);
1925         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1926                         pmd_devmap(*pmd)))
1927                 return ptl;
1928         spin_unlock(ptl);
1929         return NULL;
1930 }
1931
1932 /*
1933  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1934  *
1935  * Note that if it returns page table lock pointer, this routine returns without
1936  * unlocking page table lock. So callers must unlock it.
1937  */
1938 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1939 {
1940         spinlock_t *ptl;
1941
1942         ptl = pud_lock(vma->vm_mm, pud);
1943         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1944                 return ptl;
1945         spin_unlock(ptl);
1946         return NULL;
1947 }
1948
1949 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1950 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1951                  pud_t *pud, unsigned long addr)
1952 {
1953         spinlock_t *ptl;
1954
1955         ptl = __pud_trans_huge_lock(pud, vma);
1956         if (!ptl)
1957                 return 0;
1958
1959         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1960         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1961         if (vma_is_special_huge(vma)) {
1962                 spin_unlock(ptl);
1963                 /* No zero page support yet */
1964         } else {
1965                 /* No support for anonymous PUD pages yet */
1966                 BUG();
1967         }
1968         return 1;
1969 }
1970
1971 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1972                 unsigned long haddr)
1973 {
1974         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1975         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1976         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1977         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1978
1979         count_vm_event(THP_SPLIT_PUD);
1980
1981         pudp_huge_clear_flush_notify(vma, haddr, pud);
1982 }
1983
1984 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1985                 unsigned long address)
1986 {
1987         spinlock_t *ptl;
1988         struct mmu_notifier_range range;
1989
1990         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1991                                 address & HPAGE_PUD_MASK,
1992                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1993         mmu_notifier_invalidate_range_start(&range);
1994         ptl = pud_lock(vma->vm_mm, pud);
1995         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1996                 goto out;
1997         __split_huge_pud_locked(vma, pud, range.start);
1998
1999 out:
2000         spin_unlock(ptl);
2001         /*
2002          * No need to double call mmu_notifier->invalidate_range() callback as
2003          * the above pudp_huge_clear_flush_notify() did already call it.
2004          */
2005         mmu_notifier_invalidate_range_only_end(&range);
2006 }
2007 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2008
2009 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2010                 unsigned long haddr, pmd_t *pmd)
2011 {
2012         struct mm_struct *mm = vma->vm_mm;
2013         pgtable_t pgtable;
2014         pmd_t _pmd;
2015         int i;
2016
2017         /*
2018          * Leave pmd empty until pte is filled note that it is fine to delay
2019          * notification until mmu_notifier_invalidate_range_end() as we are
2020          * replacing a zero pmd write protected page with a zero pte write
2021          * protected page.
2022          *
2023          * See Documentation/mm/mmu_notifier.rst
2024          */
2025         pmdp_huge_clear_flush(vma, haddr, pmd);
2026
2027         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2028         pmd_populate(mm, &_pmd, pgtable);
2029
2030         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2031                 pte_t *pte, entry;
2032                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2033                 entry = pte_mkspecial(entry);
2034                 pte = pte_offset_map(&_pmd, haddr);
2035                 VM_BUG_ON(!pte_none(*pte));
2036                 set_pte_at(mm, haddr, pte, entry);
2037                 pte_unmap(pte);
2038         }
2039         smp_wmb(); /* make pte visible before pmd */
2040         pmd_populate(mm, pmd, pgtable);
2041 }
2042
2043 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2044                 unsigned long haddr, bool freeze)
2045 {
2046         struct mm_struct *mm = vma->vm_mm;
2047         struct page *page;
2048         pgtable_t pgtable;
2049         pmd_t old_pmd, _pmd;
2050         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2051         bool anon_exclusive = false, dirty = false;
2052         unsigned long addr;
2053         int i;
2054
2055         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2056         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2057         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2058         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2059                                 && !pmd_devmap(*pmd));
2060
2061         count_vm_event(THP_SPLIT_PMD);
2062
2063         if (!vma_is_anonymous(vma)) {
2064                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2065                 /*
2066                  * We are going to unmap this huge page. So
2067                  * just go ahead and zap it
2068                  */
2069                 if (arch_needs_pgtable_deposit())
2070                         zap_deposited_table(mm, pmd);
2071                 if (vma_is_special_huge(vma))
2072                         return;
2073                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2074                         swp_entry_t entry;
2075
2076                         entry = pmd_to_swp_entry(old_pmd);
2077                         page = pfn_swap_entry_to_page(entry);
2078                 } else {
2079                         page = pmd_page(old_pmd);
2080                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2081                                 set_page_dirty(page);
2082                         if (!PageReferenced(page) && pmd_young(old_pmd))
2083                                 SetPageReferenced(page);
2084                         page_remove_rmap(page, vma, true);
2085                         put_page(page);
2086                 }
2087                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2088                 return;
2089         }
2090
2091         if (is_huge_zero_pmd(*pmd)) {
2092                 /*
2093                  * FIXME: Do we want to invalidate secondary mmu by calling
2094                  * mmu_notifier_invalidate_range() see comments below inside
2095                  * __split_huge_pmd() ?
2096                  *
2097                  * We are going from a zero huge page write protected to zero
2098                  * small page also write protected so it does not seems useful
2099                  * to invalidate secondary mmu at this time.
2100                  */
2101                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2102         }
2103
2104         /*
2105          * Up to this point the pmd is present and huge and userland has the
2106          * whole access to the hugepage during the split (which happens in
2107          * place). If we overwrite the pmd with the not-huge version pointing
2108          * to the pte here (which of course we could if all CPUs were bug
2109          * free), userland could trigger a small page size TLB miss on the
2110          * small sized TLB while the hugepage TLB entry is still established in
2111          * the huge TLB. Some CPU doesn't like that.
2112          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2113          * 383 on page 105. Intel should be safe but is also warns that it's
2114          * only safe if the permission and cache attributes of the two entries
2115          * loaded in the two TLB is identical (which should be the case here).
2116          * But it is generally safer to never allow small and huge TLB entries
2117          * for the same virtual address to be loaded simultaneously. So instead
2118          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2119          * current pmd notpresent (atomically because here the pmd_trans_huge
2120          * must remain set at all times on the pmd until the split is complete
2121          * for this pmd), then we flush the SMP TLB and finally we write the
2122          * non-huge version of the pmd entry with pmd_populate.
2123          */
2124         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2125
2126         pmd_migration = is_pmd_migration_entry(old_pmd);
2127         if (unlikely(pmd_migration)) {
2128                 swp_entry_t entry;
2129
2130                 entry = pmd_to_swp_entry(old_pmd);
2131                 page = pfn_swap_entry_to_page(entry);
2132                 write = is_writable_migration_entry(entry);
2133                 if (PageAnon(page))
2134                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2135                 young = is_migration_entry_young(entry);
2136                 dirty = is_migration_entry_dirty(entry);
2137                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2138                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2139         } else {
2140                 page = pmd_page(old_pmd);
2141                 if (pmd_dirty(old_pmd)) {
2142                         dirty = true;
2143                         SetPageDirty(page);
2144                 }
2145                 write = pmd_write(old_pmd);
2146                 young = pmd_young(old_pmd);
2147                 soft_dirty = pmd_soft_dirty(old_pmd);
2148                 uffd_wp = pmd_uffd_wp(old_pmd);
2149
2150                 VM_BUG_ON_PAGE(!page_count(page), page);
2151                 page_ref_add(page, HPAGE_PMD_NR - 1);
2152
2153                 /*
2154                  * Without "freeze", we'll simply split the PMD, propagating the
2155                  * PageAnonExclusive() flag for each PTE by setting it for
2156                  * each subpage -- no need to (temporarily) clear.
2157                  *
2158                  * With "freeze" we want to replace mapped pages by
2159                  * migration entries right away. This is only possible if we
2160                  * managed to clear PageAnonExclusive() -- see
2161                  * set_pmd_migration_entry().
2162                  *
2163                  * In case we cannot clear PageAnonExclusive(), split the PMD
2164                  * only and let try_to_migrate_one() fail later.
2165                  *
2166                  * See page_try_share_anon_rmap(): invalidate PMD first.
2167                  */
2168                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2169                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2170                         freeze = false;
2171         }
2172
2173         /*
2174          * Withdraw the table only after we mark the pmd entry invalid.
2175          * This's critical for some architectures (Power).
2176          */
2177         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2178         pmd_populate(mm, &_pmd, pgtable);
2179
2180         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2181                 pte_t entry, *pte;
2182                 /*
2183                  * Note that NUMA hinting access restrictions are not
2184                  * transferred to avoid any possibility of altering
2185                  * permissions across VMAs.
2186                  */
2187                 if (freeze || pmd_migration) {
2188                         swp_entry_t swp_entry;
2189                         if (write)
2190                                 swp_entry = make_writable_migration_entry(
2191                                                         page_to_pfn(page + i));
2192                         else if (anon_exclusive)
2193                                 swp_entry = make_readable_exclusive_migration_entry(
2194                                                         page_to_pfn(page + i));
2195                         else
2196                                 swp_entry = make_readable_migration_entry(
2197                                                         page_to_pfn(page + i));
2198                         if (young)
2199                                 swp_entry = make_migration_entry_young(swp_entry);
2200                         if (dirty)
2201                                 swp_entry = make_migration_entry_dirty(swp_entry);
2202                         entry = swp_entry_to_pte(swp_entry);
2203                         if (soft_dirty)
2204                                 entry = pte_swp_mksoft_dirty(entry);
2205                         if (uffd_wp)
2206                                 entry = pte_swp_mkuffd_wp(entry);
2207                 } else {
2208                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2209                         entry = maybe_mkwrite(entry, vma);
2210                         if (anon_exclusive)
2211                                 SetPageAnonExclusive(page + i);
2212                         if (!write)
2213                                 entry = pte_wrprotect(entry);
2214                         if (!young)
2215                                 entry = pte_mkold(entry);
2216                         /*
2217                          * NOTE: we don't do pte_mkdirty when dirty==true
2218                          * because it breaks sparc64 which can sigsegv
2219                          * random process.  Need to revisit when we figure
2220                          * out what is special with sparc64.
2221                          */
2222                         if (soft_dirty)
2223                                 entry = pte_mksoft_dirty(entry);
2224                         if (uffd_wp)
2225                                 entry = pte_mkuffd_wp(entry);
2226                 }
2227                 pte = pte_offset_map(&_pmd, addr);
2228                 BUG_ON(!pte_none(*pte));
2229                 set_pte_at(mm, addr, pte, entry);
2230                 if (!pmd_migration)
2231                         atomic_inc(&page[i]._mapcount);
2232                 pte_unmap(pte);
2233         }
2234
2235         if (!pmd_migration) {
2236                 /*
2237                  * Set PG_double_map before dropping compound_mapcount to avoid
2238                  * false-negative page_mapped().
2239                  */
2240                 if (compound_mapcount(page) > 1 &&
2241                     !TestSetPageDoubleMap(page)) {
2242                         for (i = 0; i < HPAGE_PMD_NR; i++)
2243                                 atomic_inc(&page[i]._mapcount);
2244                 }
2245
2246                 lock_page_memcg(page);
2247                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2248                         /* Last compound_mapcount is gone. */
2249                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2250                                                 -HPAGE_PMD_NR);
2251                         if (TestClearPageDoubleMap(page)) {
2252                                 /* No need in mapcount reference anymore */
2253                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2254                                         atomic_dec(&page[i]._mapcount);
2255                         }
2256                 }
2257                 unlock_page_memcg(page);
2258
2259                 /* Above is effectively page_remove_rmap(page, vma, true) */
2260                 munlock_vma_page(page, vma, true);
2261         }
2262
2263         smp_wmb(); /* make pte visible before pmd */
2264         pmd_populate(mm, pmd, pgtable);
2265
2266         if (freeze) {
2267                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2268                         page_remove_rmap(page + i, vma, false);
2269                         put_page(page + i);
2270                 }
2271         }
2272 }
2273
2274 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2275                 unsigned long address, bool freeze, struct folio *folio)
2276 {
2277         spinlock_t *ptl;
2278         struct mmu_notifier_range range;
2279
2280         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2281                                 address & HPAGE_PMD_MASK,
2282                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2283         mmu_notifier_invalidate_range_start(&range);
2284         ptl = pmd_lock(vma->vm_mm, pmd);
2285
2286         /*
2287          * If caller asks to setup a migration entry, we need a folio to check
2288          * pmd against. Otherwise we can end up replacing wrong folio.
2289          */
2290         VM_BUG_ON(freeze && !folio);
2291         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2292
2293         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2294             is_pmd_migration_entry(*pmd)) {
2295                 /*
2296                  * It's safe to call pmd_page when folio is set because it's
2297                  * guaranteed that pmd is present.
2298                  */
2299                 if (folio && folio != page_folio(pmd_page(*pmd)))
2300                         goto out;
2301                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2302         }
2303
2304 out:
2305         spin_unlock(ptl);
2306         /*
2307          * No need to double call mmu_notifier->invalidate_range() callback.
2308          * They are 3 cases to consider inside __split_huge_pmd_locked():
2309          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2310          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2311          *    fault will trigger a flush_notify before pointing to a new page
2312          *    (it is fine if the secondary mmu keeps pointing to the old zero
2313          *    page in the meantime)
2314          *  3) Split a huge pmd into pte pointing to the same page. No need
2315          *     to invalidate secondary tlb entry they are all still valid.
2316          *     any further changes to individual pte will notify. So no need
2317          *     to call mmu_notifier->invalidate_range()
2318          */
2319         mmu_notifier_invalidate_range_only_end(&range);
2320 }
2321
2322 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2323                 bool freeze, struct folio *folio)
2324 {
2325         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2326
2327         if (!pmd)
2328                 return;
2329
2330         __split_huge_pmd(vma, pmd, address, freeze, folio);
2331 }
2332
2333 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2334 {
2335         /*
2336          * If the new address isn't hpage aligned and it could previously
2337          * contain an hugepage: check if we need to split an huge pmd.
2338          */
2339         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2340             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2341                          ALIGN(address, HPAGE_PMD_SIZE)))
2342                 split_huge_pmd_address(vma, address, false, NULL);
2343 }
2344
2345 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2346                              unsigned long start,
2347                              unsigned long end,
2348                              long adjust_next)
2349 {
2350         /* Check if we need to split start first. */
2351         split_huge_pmd_if_needed(vma, start);
2352
2353         /* Check if we need to split end next. */
2354         split_huge_pmd_if_needed(vma, end);
2355
2356         /*
2357          * If we're also updating the next vma vm_start,
2358          * check if we need to split it.
2359          */
2360         if (adjust_next > 0) {
2361                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2362                 unsigned long nstart = next->vm_start;
2363                 nstart += adjust_next;
2364                 split_huge_pmd_if_needed(next, nstart);
2365         }
2366 }
2367
2368 static void unmap_folio(struct folio *folio)
2369 {
2370         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2371                 TTU_SYNC;
2372
2373         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2374
2375         /*
2376          * Anon pages need migration entries to preserve them, but file
2377          * pages can simply be left unmapped, then faulted back on demand.
2378          * If that is ever changed (perhaps for mlock), update remap_page().
2379          */
2380         if (folio_test_anon(folio))
2381                 try_to_migrate(folio, ttu_flags);
2382         else
2383                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2384 }
2385
2386 static void remap_page(struct folio *folio, unsigned long nr)
2387 {
2388         int i = 0;
2389
2390         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2391         if (!folio_test_anon(folio))
2392                 return;
2393         for (;;) {
2394                 remove_migration_ptes(folio, folio, true);
2395                 i += folio_nr_pages(folio);
2396                 if (i >= nr)
2397                         break;
2398                 folio = folio_next(folio);
2399         }
2400 }
2401
2402 static void lru_add_page_tail(struct page *head, struct page *tail,
2403                 struct lruvec *lruvec, struct list_head *list)
2404 {
2405         VM_BUG_ON_PAGE(!PageHead(head), head);
2406         VM_BUG_ON_PAGE(PageCompound(tail), head);
2407         VM_BUG_ON_PAGE(PageLRU(tail), head);
2408         lockdep_assert_held(&lruvec->lru_lock);
2409
2410         if (list) {
2411                 /* page reclaim is reclaiming a huge page */
2412                 VM_WARN_ON(PageLRU(head));
2413                 get_page(tail);
2414                 list_add_tail(&tail->lru, list);
2415         } else {
2416                 /* head is still on lru (and we have it frozen) */
2417                 VM_WARN_ON(!PageLRU(head));
2418                 if (PageUnevictable(tail))
2419                         tail->mlock_count = 0;
2420                 else
2421                         list_add_tail(&tail->lru, &head->lru);
2422                 SetPageLRU(tail);
2423         }
2424 }
2425
2426 static void __split_huge_page_tail(struct page *head, int tail,
2427                 struct lruvec *lruvec, struct list_head *list)
2428 {
2429         struct page *page_tail = head + tail;
2430
2431         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2432
2433         /*
2434          * Clone page flags before unfreezing refcount.
2435          *
2436          * After successful get_page_unless_zero() might follow flags change,
2437          * for example lock_page() which set PG_waiters.
2438          *
2439          * Note that for mapped sub-pages of an anonymous THP,
2440          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2441          * the migration entry instead from where remap_page() will restore it.
2442          * We can still have PG_anon_exclusive set on effectively unmapped and
2443          * unreferenced sub-pages of an anonymous THP: we can simply drop
2444          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2445          */
2446         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2447         page_tail->flags |= (head->flags &
2448                         ((1L << PG_referenced) |
2449                          (1L << PG_swapbacked) |
2450                          (1L << PG_swapcache) |
2451                          (1L << PG_mlocked) |
2452                          (1L << PG_uptodate) |
2453                          (1L << PG_active) |
2454                          (1L << PG_workingset) |
2455                          (1L << PG_locked) |
2456                          (1L << PG_unevictable) |
2457 #ifdef CONFIG_64BIT
2458                          (1L << PG_arch_2) |
2459 #endif
2460                          (1L << PG_dirty) |
2461                          LRU_GEN_MASK | LRU_REFS_MASK));
2462
2463         /* ->mapping in first tail page is compound_mapcount */
2464         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2465                         page_tail);
2466         page_tail->mapping = head->mapping;
2467         page_tail->index = head->index + tail;
2468
2469         /*
2470          * page->private should not be set in tail pages with the exception
2471          * of swap cache pages that store the swp_entry_t in tail pages.
2472          * Fix up and warn once if private is unexpectedly set.
2473          */
2474         if (!folio_test_swapcache(page_folio(head))) {
2475                 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2476                 page_tail->private = 0;
2477         }
2478
2479         /* Page flags must be visible before we make the page non-compound. */
2480         smp_wmb();
2481
2482         /*
2483          * Clear PageTail before unfreezing page refcount.
2484          *
2485          * After successful get_page_unless_zero() might follow put_page()
2486          * which needs correct compound_head().
2487          */
2488         clear_compound_head(page_tail);
2489
2490         /* Finally unfreeze refcount. Additional reference from page cache. */
2491         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2492                                           PageSwapCache(head)));
2493
2494         if (page_is_young(head))
2495                 set_page_young(page_tail);
2496         if (page_is_idle(head))
2497                 set_page_idle(page_tail);
2498
2499         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2500
2501         /*
2502          * always add to the tail because some iterators expect new
2503          * pages to show after the currently processed elements - e.g.
2504          * migrate_pages
2505          */
2506         lru_add_page_tail(head, page_tail, lruvec, list);
2507 }
2508
2509 static void __split_huge_page(struct page *page, struct list_head *list,
2510                 pgoff_t end)
2511 {
2512         struct folio *folio = page_folio(page);
2513         struct page *head = &folio->page;
2514         struct lruvec *lruvec;
2515         struct address_space *swap_cache = NULL;
2516         unsigned long offset = 0;
2517         unsigned int nr = thp_nr_pages(head);
2518         int i;
2519
2520         /* complete memcg works before add pages to LRU */
2521         split_page_memcg(head, nr);
2522
2523         if (PageAnon(head) && PageSwapCache(head)) {
2524                 swp_entry_t entry = { .val = page_private(head) };
2525
2526                 offset = swp_offset(entry);
2527                 swap_cache = swap_address_space(entry);
2528                 xa_lock(&swap_cache->i_pages);
2529         }
2530
2531         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2532         lruvec = folio_lruvec_lock(folio);
2533
2534         ClearPageHasHWPoisoned(head);
2535
2536         for (i = nr - 1; i >= 1; i--) {
2537                 __split_huge_page_tail(head, i, lruvec, list);
2538                 /* Some pages can be beyond EOF: drop them from page cache */
2539                 if (head[i].index >= end) {
2540                         struct folio *tail = page_folio(head + i);
2541
2542                         if (shmem_mapping(head->mapping))
2543                                 shmem_uncharge(head->mapping->host, 1);
2544                         else if (folio_test_clear_dirty(tail))
2545                                 folio_account_cleaned(tail,
2546                                         inode_to_wb(folio->mapping->host));
2547                         __filemap_remove_folio(tail, NULL);
2548                         folio_put(tail);
2549                 } else if (!PageAnon(page)) {
2550                         __xa_store(&head->mapping->i_pages, head[i].index,
2551                                         head + i, 0);
2552                 } else if (swap_cache) {
2553                         __xa_store(&swap_cache->i_pages, offset + i,
2554                                         head + i, 0);
2555                 }
2556         }
2557
2558         ClearPageCompound(head);
2559         unlock_page_lruvec(lruvec);
2560         /* Caller disabled irqs, so they are still disabled here */
2561
2562         split_page_owner(head, nr);
2563
2564         /* See comment in __split_huge_page_tail() */
2565         if (PageAnon(head)) {
2566                 /* Additional pin to swap cache */
2567                 if (PageSwapCache(head)) {
2568                         page_ref_add(head, 2);
2569                         xa_unlock(&swap_cache->i_pages);
2570                 } else {
2571                         page_ref_inc(head);
2572                 }
2573         } else {
2574                 /* Additional pin to page cache */
2575                 page_ref_add(head, 2);
2576                 xa_unlock(&head->mapping->i_pages);
2577         }
2578         local_irq_enable();
2579
2580         remap_page(folio, nr);
2581
2582         if (PageSwapCache(head)) {
2583                 swp_entry_t entry = { .val = page_private(head) };
2584
2585                 split_swap_cluster(entry);
2586         }
2587
2588         for (i = 0; i < nr; i++) {
2589                 struct page *subpage = head + i;
2590                 if (subpage == page)
2591                         continue;
2592                 unlock_page(subpage);
2593
2594                 /*
2595                  * Subpages may be freed if there wasn't any mapping
2596                  * like if add_to_swap() is running on a lru page that
2597                  * had its mapping zapped. And freeing these pages
2598                  * requires taking the lru_lock so we do the put_page
2599                  * of the tail pages after the split is complete.
2600                  */
2601                 free_page_and_swap_cache(subpage);
2602         }
2603 }
2604
2605 /* Racy check whether the huge page can be split */
2606 bool can_split_folio(struct folio *folio, int *pextra_pins)
2607 {
2608         int extra_pins;
2609
2610         /* Additional pins from page cache */
2611         if (folio_test_anon(folio))
2612                 extra_pins = folio_test_swapcache(folio) ?
2613                                 folio_nr_pages(folio) : 0;
2614         else
2615                 extra_pins = folio_nr_pages(folio);
2616         if (pextra_pins)
2617                 *pextra_pins = extra_pins;
2618         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2619 }
2620
2621 /*
2622  * This function splits huge page into normal pages. @page can point to any
2623  * subpage of huge page to split. Split doesn't change the position of @page.
2624  *
2625  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2626  * The huge page must be locked.
2627  *
2628  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2629  *
2630  * Both head page and tail pages will inherit mapping, flags, and so on from
2631  * the hugepage.
2632  *
2633  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2634  * they are not mapped.
2635  *
2636  * Returns 0 if the hugepage is split successfully.
2637  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2638  * us.
2639  */
2640 int split_huge_page_to_list(struct page *page, struct list_head *list)
2641 {
2642         struct folio *folio = page_folio(page);
2643         struct deferred_split *ds_queue = get_deferred_split_queue(&folio->page);
2644         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2645         struct anon_vma *anon_vma = NULL;
2646         struct address_space *mapping = NULL;
2647         int extra_pins, ret;
2648         pgoff_t end;
2649         bool is_hzp;
2650
2651         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2652         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2653
2654         is_hzp = is_huge_zero_page(&folio->page);
2655         VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2656         if (is_hzp)
2657                 return -EBUSY;
2658
2659         if (folio_test_writeback(folio))
2660                 return -EBUSY;
2661
2662         if (folio_test_anon(folio)) {
2663                 /*
2664                  * The caller does not necessarily hold an mmap_lock that would
2665                  * prevent the anon_vma disappearing so we first we take a
2666                  * reference to it and then lock the anon_vma for write. This
2667                  * is similar to folio_lock_anon_vma_read except the write lock
2668                  * is taken to serialise against parallel split or collapse
2669                  * operations.
2670                  */
2671                 anon_vma = folio_get_anon_vma(folio);
2672                 if (!anon_vma) {
2673                         ret = -EBUSY;
2674                         goto out;
2675                 }
2676                 end = -1;
2677                 mapping = NULL;
2678                 anon_vma_lock_write(anon_vma);
2679         } else {
2680                 gfp_t gfp;
2681
2682                 mapping = folio->mapping;
2683
2684                 /* Truncated ? */
2685                 if (!mapping) {
2686                         ret = -EBUSY;
2687                         goto out;
2688                 }
2689
2690                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2691                                                         GFP_RECLAIM_MASK);
2692
2693                 if (folio_test_private(folio) &&
2694                                 !filemap_release_folio(folio, gfp)) {
2695                         ret = -EBUSY;
2696                         goto out;
2697                 }
2698
2699                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2700                 if (xas_error(&xas)) {
2701                         ret = xas_error(&xas);
2702                         goto out;
2703                 }
2704
2705                 anon_vma = NULL;
2706                 i_mmap_lock_read(mapping);
2707
2708                 /*
2709                  *__split_huge_page() may need to trim off pages beyond EOF:
2710                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2711                  * which cannot be nested inside the page tree lock. So note
2712                  * end now: i_size itself may be changed at any moment, but
2713                  * folio lock is good enough to serialize the trimming.
2714                  */
2715                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2716                 if (shmem_mapping(mapping))
2717                         end = shmem_fallocend(mapping->host, end);
2718         }
2719
2720         /*
2721          * Racy check if we can split the page, before unmap_folio() will
2722          * split PMDs
2723          */
2724         if (!can_split_folio(folio, &extra_pins)) {
2725                 ret = -EBUSY;
2726                 goto out_unlock;
2727         }
2728
2729         unmap_folio(folio);
2730
2731         /* block interrupt reentry in xa_lock and spinlock */
2732         local_irq_disable();
2733         if (mapping) {
2734                 /*
2735                  * Check if the folio is present in page cache.
2736                  * We assume all tail are present too, if folio is there.
2737                  */
2738                 xas_lock(&xas);
2739                 xas_reset(&xas);
2740                 if (xas_load(&xas) != folio)
2741                         goto fail;
2742         }
2743
2744         /* Prevent deferred_split_scan() touching ->_refcount */
2745         spin_lock(&ds_queue->split_queue_lock);
2746         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2747                 if (!list_empty(page_deferred_list(&folio->page))) {
2748                         ds_queue->split_queue_len--;
2749                         list_del(page_deferred_list(&folio->page));
2750                 }
2751                 spin_unlock(&ds_queue->split_queue_lock);
2752                 if (mapping) {
2753                         int nr = folio_nr_pages(folio);
2754
2755                         xas_split(&xas, folio, folio_order(folio));
2756                         if (folio_test_swapbacked(folio)) {
2757                                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2758                                                         -nr);
2759                         } else {
2760                                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2761                                                         -nr);
2762                                 filemap_nr_thps_dec(mapping);
2763                         }
2764                 }
2765
2766                 __split_huge_page(page, list, end);
2767                 ret = 0;
2768         } else {
2769                 spin_unlock(&ds_queue->split_queue_lock);
2770 fail:
2771                 if (mapping)
2772                         xas_unlock(&xas);
2773                 local_irq_enable();
2774                 remap_page(folio, folio_nr_pages(folio));
2775                 ret = -EBUSY;
2776         }
2777
2778 out_unlock:
2779         if (anon_vma) {
2780                 anon_vma_unlock_write(anon_vma);
2781                 put_anon_vma(anon_vma);
2782         }
2783         if (mapping)
2784                 i_mmap_unlock_read(mapping);
2785 out:
2786         xas_destroy(&xas);
2787         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2788         return ret;
2789 }
2790
2791 void free_transhuge_page(struct page *page)
2792 {
2793         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2794         unsigned long flags;
2795
2796         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2797         if (!list_empty(page_deferred_list(page))) {
2798                 ds_queue->split_queue_len--;
2799                 list_del(page_deferred_list(page));
2800         }
2801         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2802         free_compound_page(page);
2803 }
2804
2805 void deferred_split_huge_page(struct page *page)
2806 {
2807         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2808 #ifdef CONFIG_MEMCG
2809         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2810 #endif
2811         unsigned long flags;
2812
2813         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2814
2815         /*
2816          * The try_to_unmap() in page reclaim path might reach here too,
2817          * this may cause a race condition to corrupt deferred split queue.
2818          * And, if page reclaim is already handling the same page, it is
2819          * unnecessary to handle it again in shrinker.
2820          *
2821          * Check PageSwapCache to determine if the page is being
2822          * handled by page reclaim since THP swap would add the page into
2823          * swap cache before calling try_to_unmap().
2824          */
2825         if (PageSwapCache(page))
2826                 return;
2827
2828         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2829         if (list_empty(page_deferred_list(page))) {
2830                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2831                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2832                 ds_queue->split_queue_len++;
2833 #ifdef CONFIG_MEMCG
2834                 if (memcg)
2835                         set_shrinker_bit(memcg, page_to_nid(page),
2836                                          deferred_split_shrinker.id);
2837 #endif
2838         }
2839         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2840 }
2841
2842 static unsigned long deferred_split_count(struct shrinker *shrink,
2843                 struct shrink_control *sc)
2844 {
2845         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2846         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2847
2848 #ifdef CONFIG_MEMCG
2849         if (sc->memcg)
2850                 ds_queue = &sc->memcg->deferred_split_queue;
2851 #endif
2852         return READ_ONCE(ds_queue->split_queue_len);
2853 }
2854
2855 static unsigned long deferred_split_scan(struct shrinker *shrink,
2856                 struct shrink_control *sc)
2857 {
2858         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2859         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2860         unsigned long flags;
2861         LIST_HEAD(list), *pos, *next;
2862         struct page *page;
2863         int split = 0;
2864
2865 #ifdef CONFIG_MEMCG
2866         if (sc->memcg)
2867                 ds_queue = &sc->memcg->deferred_split_queue;
2868 #endif
2869
2870         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2871         /* Take pin on all head pages to avoid freeing them under us */
2872         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2873                 page = list_entry((void *)pos, struct page, deferred_list);
2874                 page = compound_head(page);
2875                 if (get_page_unless_zero(page)) {
2876                         list_move(page_deferred_list(page), &list);
2877                 } else {
2878                         /* We lost race with put_compound_page() */
2879                         list_del_init(page_deferred_list(page));
2880                         ds_queue->split_queue_len--;
2881                 }
2882                 if (!--sc->nr_to_scan)
2883                         break;
2884         }
2885         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2886
2887         list_for_each_safe(pos, next, &list) {
2888                 page = list_entry((void *)pos, struct page, deferred_list);
2889                 if (!trylock_page(page))
2890                         goto next;
2891                 /* split_huge_page() removes page from list on success */
2892                 if (!split_huge_page(page))
2893                         split++;
2894                 unlock_page(page);
2895 next:
2896                 put_page(page);
2897         }
2898
2899         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2900         list_splice_tail(&list, &ds_queue->split_queue);
2901         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2902
2903         /*
2904          * Stop shrinker if we didn't split any page, but the queue is empty.
2905          * This can happen if pages were freed under us.
2906          */
2907         if (!split && list_empty(&ds_queue->split_queue))
2908                 return SHRINK_STOP;
2909         return split;
2910 }
2911
2912 static struct shrinker deferred_split_shrinker = {
2913         .count_objects = deferred_split_count,
2914         .scan_objects = deferred_split_scan,
2915         .seeks = DEFAULT_SEEKS,
2916         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2917                  SHRINKER_NONSLAB,
2918 };
2919
2920 #ifdef CONFIG_DEBUG_FS
2921 static void split_huge_pages_all(void)
2922 {
2923         struct zone *zone;
2924         struct page *page;
2925         unsigned long pfn, max_zone_pfn;
2926         unsigned long total = 0, split = 0;
2927
2928         pr_debug("Split all THPs\n");
2929         for_each_zone(zone) {
2930                 if (!managed_zone(zone))
2931                         continue;
2932                 max_zone_pfn = zone_end_pfn(zone);
2933                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2934                         int nr_pages;
2935
2936                         page = pfn_to_online_page(pfn);
2937                         if (!page || !get_page_unless_zero(page))
2938                                 continue;
2939
2940                         if (zone != page_zone(page))
2941                                 goto next;
2942
2943                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2944                                 goto next;
2945
2946                         total++;
2947                         lock_page(page);
2948                         nr_pages = thp_nr_pages(page);
2949                         if (!split_huge_page(page))
2950                                 split++;
2951                         pfn += nr_pages - 1;
2952                         unlock_page(page);
2953 next:
2954                         put_page(page);
2955                         cond_resched();
2956                 }
2957         }
2958
2959         pr_debug("%lu of %lu THP split\n", split, total);
2960 }
2961
2962 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2963 {
2964         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2965                     is_vm_hugetlb_page(vma);
2966 }
2967
2968 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2969                                 unsigned long vaddr_end)
2970 {
2971         int ret = 0;
2972         struct task_struct *task;
2973         struct mm_struct *mm;
2974         unsigned long total = 0, split = 0;
2975         unsigned long addr;
2976
2977         vaddr_start &= PAGE_MASK;
2978         vaddr_end &= PAGE_MASK;
2979
2980         /* Find the task_struct from pid */
2981         rcu_read_lock();
2982         task = find_task_by_vpid(pid);
2983         if (!task) {
2984                 rcu_read_unlock();
2985                 ret = -ESRCH;
2986                 goto out;
2987         }
2988         get_task_struct(task);
2989         rcu_read_unlock();
2990
2991         /* Find the mm_struct */
2992         mm = get_task_mm(task);
2993         put_task_struct(task);
2994
2995         if (!mm) {
2996                 ret = -EINVAL;
2997                 goto out;
2998         }
2999
3000         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3001                  pid, vaddr_start, vaddr_end);
3002
3003         mmap_read_lock(mm);
3004         /*
3005          * always increase addr by PAGE_SIZE, since we could have a PTE page
3006          * table filled with PTE-mapped THPs, each of which is distinct.
3007          */
3008         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3009                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3010                 struct page *page;
3011
3012                 if (!vma)
3013                         break;
3014
3015                 /* skip special VMA and hugetlb VMA */
3016                 if (vma_not_suitable_for_thp_split(vma)) {
3017                         addr = vma->vm_end;
3018                         continue;
3019                 }
3020
3021                 /* FOLL_DUMP to ignore special (like zero) pages */
3022                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3023
3024                 if (IS_ERR_OR_NULL(page))
3025                         continue;
3026
3027                 if (!is_transparent_hugepage(page))
3028                         goto next;
3029
3030                 total++;
3031                 if (!can_split_folio(page_folio(page), NULL))
3032                         goto next;
3033
3034                 if (!trylock_page(page))
3035                         goto next;
3036
3037                 if (!split_huge_page(page))
3038                         split++;
3039
3040                 unlock_page(page);
3041 next:
3042                 put_page(page);
3043                 cond_resched();
3044         }
3045         mmap_read_unlock(mm);
3046         mmput(mm);
3047
3048         pr_debug("%lu of %lu THP split\n", split, total);
3049
3050 out:
3051         return ret;
3052 }
3053
3054 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3055                                 pgoff_t off_end)
3056 {
3057         struct filename *file;
3058         struct file *candidate;
3059         struct address_space *mapping;
3060         int ret = -EINVAL;
3061         pgoff_t index;
3062         int nr_pages = 1;
3063         unsigned long total = 0, split = 0;
3064
3065         file = getname_kernel(file_path);
3066         if (IS_ERR(file))
3067                 return ret;
3068
3069         candidate = file_open_name(file, O_RDONLY, 0);
3070         if (IS_ERR(candidate))
3071                 goto out;
3072
3073         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3074                  file_path, off_start, off_end);
3075
3076         mapping = candidate->f_mapping;
3077
3078         for (index = off_start; index < off_end; index += nr_pages) {
3079                 struct page *fpage = pagecache_get_page(mapping, index,
3080                                                 FGP_ENTRY | FGP_HEAD, 0);
3081
3082                 nr_pages = 1;
3083                 if (xa_is_value(fpage) || !fpage)
3084                         continue;
3085
3086                 if (!is_transparent_hugepage(fpage))
3087                         goto next;
3088
3089                 total++;
3090                 nr_pages = thp_nr_pages(fpage);
3091
3092                 if (!trylock_page(fpage))
3093                         goto next;
3094
3095                 if (!split_huge_page(fpage))
3096                         split++;
3097
3098                 unlock_page(fpage);
3099 next:
3100                 put_page(fpage);
3101                 cond_resched();
3102         }
3103
3104         filp_close(candidate, NULL);
3105         ret = 0;
3106
3107         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3108 out:
3109         putname(file);
3110         return ret;
3111 }
3112
3113 #define MAX_INPUT_BUF_SZ 255
3114
3115 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3116                                 size_t count, loff_t *ppops)
3117 {
3118         static DEFINE_MUTEX(split_debug_mutex);
3119         ssize_t ret;
3120         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3121         char input_buf[MAX_INPUT_BUF_SZ];
3122         int pid;
3123         unsigned long vaddr_start, vaddr_end;
3124
3125         ret = mutex_lock_interruptible(&split_debug_mutex);
3126         if (ret)
3127                 return ret;
3128
3129         ret = -EFAULT;
3130
3131         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3132         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3133                 goto out;
3134
3135         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3136
3137         if (input_buf[0] == '/') {
3138                 char *tok;
3139                 char *buf = input_buf;
3140                 char file_path[MAX_INPUT_BUF_SZ];
3141                 pgoff_t off_start = 0, off_end = 0;
3142                 size_t input_len = strlen(input_buf);
3143
3144                 tok = strsep(&buf, ",");
3145                 if (tok) {
3146                         strcpy(file_path, tok);
3147                 } else {
3148                         ret = -EINVAL;
3149                         goto out;
3150                 }
3151
3152                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3153                 if (ret != 2) {
3154                         ret = -EINVAL;
3155                         goto out;
3156                 }
3157                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3158                 if (!ret)
3159                         ret = input_len;
3160
3161                 goto out;
3162         }
3163
3164         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3165         if (ret == 1 && pid == 1) {
3166                 split_huge_pages_all();
3167                 ret = strlen(input_buf);
3168                 goto out;
3169         } else if (ret != 3) {
3170                 ret = -EINVAL;
3171                 goto out;
3172         }
3173
3174         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3175         if (!ret)
3176                 ret = strlen(input_buf);
3177 out:
3178         mutex_unlock(&split_debug_mutex);
3179         return ret;
3180
3181 }
3182
3183 static const struct file_operations split_huge_pages_fops = {
3184         .owner   = THIS_MODULE,
3185         .write   = split_huge_pages_write,
3186         .llseek  = no_llseek,
3187 };
3188
3189 static int __init split_huge_pages_debugfs(void)
3190 {
3191         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3192                             &split_huge_pages_fops);
3193         return 0;
3194 }
3195 late_initcall(split_huge_pages_debugfs);
3196 #endif
3197
3198 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3199 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3200                 struct page *page)
3201 {
3202         struct vm_area_struct *vma = pvmw->vma;
3203         struct mm_struct *mm = vma->vm_mm;
3204         unsigned long address = pvmw->address;
3205         bool anon_exclusive;
3206         pmd_t pmdval;
3207         swp_entry_t entry;
3208         pmd_t pmdswp;
3209
3210         if (!(pvmw->pmd && !pvmw->pte))
3211                 return 0;
3212
3213         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3214         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3215
3216         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3217         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3218         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3219                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3220                 return -EBUSY;
3221         }
3222
3223         if (pmd_dirty(pmdval))
3224                 set_page_dirty(page);
3225         if (pmd_write(pmdval))
3226                 entry = make_writable_migration_entry(page_to_pfn(page));
3227         else if (anon_exclusive)
3228                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3229         else
3230                 entry = make_readable_migration_entry(page_to_pfn(page));
3231         if (pmd_young(pmdval))
3232                 entry = make_migration_entry_young(entry);
3233         if (pmd_dirty(pmdval))
3234                 entry = make_migration_entry_dirty(entry);
3235         pmdswp = swp_entry_to_pmd(entry);
3236         if (pmd_soft_dirty(pmdval))
3237                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3238         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3239         page_remove_rmap(page, vma, true);
3240         put_page(page);
3241         trace_set_migration_pmd(address, pmd_val(pmdswp));
3242
3243         return 0;
3244 }
3245
3246 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3247 {
3248         struct vm_area_struct *vma = pvmw->vma;
3249         struct mm_struct *mm = vma->vm_mm;
3250         unsigned long address = pvmw->address;
3251         unsigned long haddr = address & HPAGE_PMD_MASK;
3252         pmd_t pmde;
3253         swp_entry_t entry;
3254
3255         if (!(pvmw->pmd && !pvmw->pte))
3256                 return;
3257
3258         entry = pmd_to_swp_entry(*pvmw->pmd);
3259         get_page(new);
3260         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3261         if (pmd_swp_soft_dirty(*pvmw->pmd))
3262                 pmde = pmd_mksoft_dirty(pmde);
3263         if (is_writable_migration_entry(entry))
3264                 pmde = maybe_pmd_mkwrite(pmde, vma);
3265         if (pmd_swp_uffd_wp(*pvmw->pmd))
3266                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3267         if (!is_migration_entry_young(entry))
3268                 pmde = pmd_mkold(pmde);
3269         /* NOTE: this may contain setting soft-dirty on some archs */
3270         if (PageDirty(new) && is_migration_entry_dirty(entry))
3271                 pmde = pmd_mkdirty(pmde);
3272
3273         if (PageAnon(new)) {
3274                 rmap_t rmap_flags = RMAP_COMPOUND;
3275
3276                 if (!is_readable_migration_entry(entry))
3277                         rmap_flags |= RMAP_EXCLUSIVE;
3278
3279                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3280         } else {
3281                 page_add_file_rmap(new, vma, true);
3282         }
3283         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3284         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3285
3286         /* No need to invalidate - it was non-present before */
3287         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3288         trace_remove_migration_pmd(address, pmd_val(pmde));
3289 }
3290 #endif