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/dax.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33 #include <linux/shmem_fs.h>
  34 #include <linux/oom.h>
  35 #include <linux/numa.h>
  36 #include <linux/page_owner.h>
  37 #include <linux/sched/sysctl.h>
  38
  39 #include <asm/tlb.h>
  40 #include <asm/pgalloc.h>
  41 #include "internal.h"
  42 #include "swap.h"
  43
  44 #define CREATE_TRACE_POINTS
  45 #include <trace/events/thp.h>
  46
  47 /*
  48  * By default, transparent hugepage support is disabled in order to avoid
  49  * risking an increased memory footprint for applications that are not
  50  * guaranteed to benefit from it. When transparent hugepage support is
  51  * enabled, it is for all mappings, and khugepaged scans all mappings.
  52  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  53  * for all hugepage allocations.
  54  */
  55 unsigned long transparent_hugepage_flags __read_mostly =
  56 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  57         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  58 #endif
  59 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  60         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  61 #endif
  62         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  63         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  64         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  65
  66 static struct shrinker deferred_split_shrinker;
  67
  68 static atomic_t huge_zero_refcount;
  69 struct page *huge_zero_page __read_mostly;
  70 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  71
  72 bool hugepage_vma_check(struct vm_area_struct *vma,
  73                         unsigned long vm_flags,
  74                         bool smaps, bool in_pf)
  75 {
  76         if (!vma->vm_mm)                /* vdso */
  77                 return false;
  78
  79         /*
  80          * Explicitly disabled through madvise or prctl, or some
  81          * architectures may disable THP for some mappings, for
  82          * example, s390 kvm.
  83          * */
  84         if ((vm_flags & VM_NOHUGEPAGE) ||
  85             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  86                 return false;
  87         /*
  88          * If the hardware/firmware marked hugepage support disabled.
  89          */
  90         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
  91                 return false;
  92
  93         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  94         if (vma_is_dax(vma))
  95                 return in_pf;
  96
  97         /*
  98          * Special VMA and hugetlb VMA.
  99          * Must be checked after dax since some dax mappings may have
 100          * VM_MIXEDMAP set.
 101          */
 102         if (vm_flags & VM_NO_KHUGEPAGED)
 103                 return false;
 104
 105         /*
 106          * Check alignment for file vma and size for both file and anon vma.
 107          *
 108          * Skip the check for page fault. Huge fault does the check in fault
 109          * handlers. And this check is not suitable for huge PUD fault.
 110          */
 111         if (!in_pf &&
 112             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 113                 return false;
 114
 115         /*
 116          * Enabled via shmem mount options or sysfs settings.
 117          * Must be done before hugepage flags check since shmem has its
 118          * own flags.
 119          */
 120         if (!in_pf && shmem_file(vma->vm_file))
 121                 return shmem_huge_enabled(vma);
 122
 123         if (!hugepage_flags_enabled())
 124                 return false;
 125
 126         /* THP settings require madvise. */
 127         if (!(vm_flags & VM_HUGEPAGE) && !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         count_vm_event(THP_ZERO_PAGE_ALLOC);
 167         preempt_disable();
 168         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 169                 preempt_enable();
 170                 __free_pages(zero_page, compound_order(zero_page));
 171                 goto retry;
 172         }
 173         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 174
 175         /* We take additional reference here. It will be put back by shrinker */
 176         atomic_set(&huge_zero_refcount, 2);
 177         preempt_enable();
 178         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 static struct kobj_attribute enabled_attr =
 280         __ATTR(enabled, 0644, enabled_show, enabled_store);
 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 =
 370         __ATTR(defrag, 0644, defrag_show, defrag_store);
 371
 372 static ssize_t use_zero_page_show(struct kobject *kobj,
 373                                   struct kobj_attribute *attr, char *buf)
 374 {
 375         return single_hugepage_flag_show(kobj, attr, buf,
 376                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static ssize_t use_zero_page_store(struct kobject *kobj,
 379                 struct kobj_attribute *attr, const char *buf, size_t count)
 380 {
 381         return single_hugepage_flag_store(kobj, attr, buf, count,
 382                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static struct kobj_attribute use_zero_page_attr =
 385         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 386
 387 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 388                                    struct kobj_attribute *attr, char *buf)
 389 {
 390         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 391 }
 392 static struct kobj_attribute hpage_pmd_size_attr =
 393         __ATTR_RO(hpage_pmd_size);
 394
 395 static struct attribute *hugepage_attr[] = {
 396         &enabled_attr.attr,
 397         &defrag_attr.attr,
 398         &use_zero_page_attr.attr,
 399         &hpage_pmd_size_attr.attr,
 400 #ifdef CONFIG_SHMEM
 401         &shmem_enabled_attr.attr,
 402 #endif
 403         NULL,
 404 };
 405
 406 static const struct attribute_group hugepage_attr_group = {
 407         .attrs = hugepage_attr,
 408 };
 409
 410 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 411 {
 412         int err;
 413
 414         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 415         if (unlikely(!*hugepage_kobj)) {
 416                 pr_err("failed to create transparent hugepage kobject\n");
 417                 return -ENOMEM;
 418         }
 419
 420         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 421         if (err) {
 422                 pr_err("failed to register transparent hugepage group\n");
 423                 goto delete_obj;
 424         }
 425
 426         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 427         if (err) {
 428                 pr_err("failed to register transparent hugepage group\n");
 429                 goto remove_hp_group;
 430         }
 431
 432         return 0;
 433
 434 remove_hp_group:
 435         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 436 delete_obj:
 437         kobject_put(*hugepage_kobj);
 438         return err;
 439 }
 440
 441 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 442 {
 443         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 444         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 445         kobject_put(hugepage_kobj);
 446 }
 447 #else
 448 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 449 {
 450         return 0;
 451 }
 452
 453 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 454 {
 455 }
 456 #endif /* CONFIG_SYSFS */
 457
 458 static int __init hugepage_init(void)
 459 {
 460         int err;
 461         struct kobject *hugepage_kobj;
 462
 463         if (!has_transparent_hugepage()) {
 464                 /*
 465                  * Hardware doesn't support hugepages, hence disable
 466                  * DAX PMD support.
 467                  */
 468                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 469                 return -EINVAL;
 470         }
 471
 472         /*
 473          * hugepages can't be allocated by the buddy allocator
 474          */
 475         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 476         /*
 477          * we use page->mapping and page->index in second tail page
 478          * as list_head: assuming THP order >= 2
 479          */
 480         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 481
 482         err = hugepage_init_sysfs(&hugepage_kobj);
 483         if (err)
 484                 goto err_sysfs;
 485
 486         err = khugepaged_init();
 487         if (err)
 488                 goto err_slab;
 489
 490         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 491         if (err)
 492                 goto err_hzp_shrinker;
 493         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 494         if (err)
 495                 goto err_split_shrinker;
 496
 497         /*
 498          * By default disable transparent hugepages on smaller systems,
 499          * where the extra memory used could hurt more than TLB overhead
 500          * is likely to save.  The admin can still enable it through /sys.
 501          */
 502         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 503                 transparent_hugepage_flags = 0;
 504                 return 0;
 505         }
 506
 507         err = start_stop_khugepaged();
 508         if (err)
 509                 goto err_khugepaged;
 510
 511         return 0;
 512 err_khugepaged:
 513         unregister_shrinker(&deferred_split_shrinker);
 514 err_split_shrinker:
 515         unregister_shrinker(&huge_zero_page_shrinker);
 516 err_hzp_shrinker:
 517         khugepaged_destroy();
 518 err_slab:
 519         hugepage_exit_sysfs(hugepage_kobj);
 520 err_sysfs:
 521         return err;
 522 }
 523 subsys_initcall(hugepage_init);
 524
 525 static int __init setup_transparent_hugepage(char *str)
 526 {
 527         int ret = 0;
 528         if (!str)
 529                 goto out;
 530         if (!strcmp(str, "always")) {
 531                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 532                         &transparent_hugepage_flags);
 533                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 534                           &transparent_hugepage_flags);
 535                 ret = 1;
 536         } else if (!strcmp(str, "madvise")) {
 537                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 538                           &transparent_hugepage_flags);
 539                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 540                         &transparent_hugepage_flags);
 541                 ret = 1;
 542         } else if (!strcmp(str, "never")) {
 543                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 544                           &transparent_hugepage_flags);
 545                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 546                           &transparent_hugepage_flags);
 547                 ret = 1;
 548         }
 549 out:
 550         if (!ret)
 551                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 552         return ret;
 553 }
 554 __setup("transparent_hugepage=", setup_transparent_hugepage);
 555
 556 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 557 {
 558         if (likely(vma->vm_flags & VM_WRITE))
 559                 pmd = pmd_mkwrite(pmd);
 560         return pmd;
 561 }
 562
 563 #ifdef CONFIG_MEMCG
 564 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 565 {
 566         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 567         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 568
 569         if (memcg)
 570                 return &memcg->deferred_split_queue;
 571         else
 572                 return &pgdat->deferred_split_queue;
 573 }
 574 #else
 575 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 576 {
 577         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 578
 579         return &pgdat->deferred_split_queue;
 580 }
 581 #endif
 582
 583 void prep_transhuge_page(struct page *page)
 584 {
 585         /*
 586          * we use page->mapping and page->indexlru in second tail page
 587          * as list_head: assuming THP order >= 2
 588          */
 589
 590         INIT_LIST_HEAD(page_deferred_list(page));
 591         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 592 }
 593
 594 static inline bool is_transparent_hugepage(struct page *page)
 595 {
 596         if (!PageCompound(page))
 597                 return false;
 598
 599         page = compound_head(page);
 600         return is_huge_zero_page(page) ||
 601                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 602 }
 603
 604 static unsigned long __thp_get_unmapped_area(struct file *filp,
 605                 unsigned long addr, unsigned long len,
 606                 loff_t off, unsigned long flags, unsigned long size)
 607 {
 608         loff_t off_end = off + len;
 609         loff_t off_align = round_up(off, size);
 610         unsigned long len_pad, ret;
 611
 612         if (off_end <= off_align || (off_end - off_align) < size)
 613                 return 0;
 614
 615         len_pad = len + size;
 616         if (len_pad < len || (off + len_pad) < off)
 617                 return 0;
 618
 619         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 620                                               off >> PAGE_SHIFT, flags);
 621
 622         /*
 623          * The failure might be due to length padding. The caller will retry
 624          * without the padding.
 625          */
 626         if (IS_ERR_VALUE(ret))
 627                 return 0;
 628
 629         /*
 630          * Do not try to align to THP boundary if allocation at the address
 631          * hint succeeds.
 632          */
 633         if (ret == addr)
 634                 return addr;
 635
 636         ret += (off - ret) & (size - 1);
 637         return ret;
 638 }
 639
 640 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 641                 unsigned long len, unsigned long pgoff, unsigned long flags)
 642 {
 643         unsigned long ret;
 644         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 645
 646         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 647         if (ret)
 648                 return ret;
 649
 650         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 651 }
 652 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 653
 654 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 655                         struct page *page, gfp_t gfp)
 656 {
 657         struct vm_area_struct *vma = vmf->vma;
 658         pgtable_t pgtable;
 659         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 660         vm_fault_t ret = 0;
 661
 662         VM_BUG_ON_PAGE(!PageCompound(page), page);
 663
 664         if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
 665                 put_page(page);
 666                 count_vm_event(THP_FAULT_FALLBACK);
 667                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 668                 return VM_FAULT_FALLBACK;
 669         }
 670         cgroup_throttle_swaprate(page, gfp);
 671
 672         pgtable = pte_alloc_one(vma->vm_mm);
 673         if (unlikely(!pgtable)) {
 674                 ret = VM_FAULT_OOM;
 675                 goto release;
 676         }
 677
 678         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 679         /*
 680          * The memory barrier inside __SetPageUptodate makes sure that
 681          * clear_huge_page writes become visible before the set_pmd_at()
 682          * write.
 683          */
 684         __SetPageUptodate(page);
 685
 686         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 687         if (unlikely(!pmd_none(*vmf->pmd))) {
 688                 goto unlock_release;
 689         } else {
 690                 pmd_t entry;
 691
 692                 ret = check_stable_address_space(vma->vm_mm);
 693                 if (ret)
 694                         goto unlock_release;
 695
 696                 /* Deliver the page fault to userland */
 697                 if (userfaultfd_missing(vma)) {
 698                         spin_unlock(vmf->ptl);
 699                         put_page(page);
 700                         pte_free(vma->vm_mm, pgtable);
 701                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 702                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 703                         return ret;
 704                 }
 705
 706                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 707                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 708                 page_add_new_anon_rmap(page, vma, haddr);
 709                 lru_cache_add_inactive_or_unevictable(page, vma);
 710                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 711                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 712                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 713                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 714                 mm_inc_nr_ptes(vma->vm_mm);
 715                 spin_unlock(vmf->ptl);
 716                 count_vm_event(THP_FAULT_ALLOC);
 717                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 718         }
 719
 720         return 0;
 721 unlock_release:
 722         spin_unlock(vmf->ptl);
 723 release:
 724         if (pgtable)
 725                 pte_free(vma->vm_mm, pgtable);
 726         put_page(page);
 727         return ret;
 728
 729 }
 730
 731 /*
 732  * always: directly stall for all thp allocations
 733  * defer: wake kswapd and fail if not immediately available
 734  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 735  *                fail if not immediately available
 736  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 737  *          available
 738  * never: never stall for any thp allocation
 739  */
 740 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 741 {
 742         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 743
 744         /* Always do synchronous compaction */
 745         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 746                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 747
 748         /* Kick kcompactd and fail quickly */
 749         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 750                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 751
 752         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 753         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 754                 return GFP_TRANSHUGE_LIGHT |
 755                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 756                                         __GFP_KSWAPD_RECLAIM);
 757
 758         /* Only do synchronous compaction if madvised */
 759         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 760                 return GFP_TRANSHUGE_LIGHT |
 761                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 762
 763         return GFP_TRANSHUGE_LIGHT;
 764 }
 765
 766 /* Caller must hold page table lock. */
 767 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 768                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 769                 struct page *zero_page)
 770 {
 771         pmd_t entry;
 772         if (!pmd_none(*pmd))
 773                 return;
 774         entry = mk_pmd(zero_page, vma->vm_page_prot);
 775         entry = pmd_mkhuge(entry);
 776         if (pgtable)
 777                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 778         set_pmd_at(mm, haddr, pmd, entry);
 779         mm_inc_nr_ptes(mm);
 780 }
 781
 782 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 783 {
 784         struct vm_area_struct *vma = vmf->vma;
 785         gfp_t gfp;
 786         struct folio *folio;
 787         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 788
 789         if (!transhuge_vma_suitable(vma, haddr))
 790                 return VM_FAULT_FALLBACK;
 791         if (unlikely(anon_vma_prepare(vma)))
 792                 return VM_FAULT_OOM;
 793         khugepaged_enter_vma(vma, vma->vm_flags);
 794
 795         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 796                         !mm_forbids_zeropage(vma->vm_mm) &&
 797                         transparent_hugepage_use_zero_page()) {
 798                 pgtable_t pgtable;
 799                 struct page *zero_page;
 800                 vm_fault_t ret;
 801                 pgtable = pte_alloc_one(vma->vm_mm);
 802                 if (unlikely(!pgtable))
 803                         return VM_FAULT_OOM;
 804                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 805                 if (unlikely(!zero_page)) {
 806                         pte_free(vma->vm_mm, pgtable);
 807                         count_vm_event(THP_FAULT_FALLBACK);
 808                         return VM_FAULT_FALLBACK;
 809                 }
 810                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 811                 ret = 0;
 812                 if (pmd_none(*vmf->pmd)) {
 813                         ret = check_stable_address_space(vma->vm_mm);
 814                         if (ret) {
 815                                 spin_unlock(vmf->ptl);
 816                                 pte_free(vma->vm_mm, pgtable);
 817                         } else if (userfaultfd_missing(vma)) {
 818                                 spin_unlock(vmf->ptl);
 819                                 pte_free(vma->vm_mm, pgtable);
 820                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 821                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 822                         } else {
 823                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 824                                                    haddr, vmf->pmd, zero_page);
 825                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 826                                 spin_unlock(vmf->ptl);
 827                         }
 828                 } else {
 829                         spin_unlock(vmf->ptl);
 830                         pte_free(vma->vm_mm, pgtable);
 831                 }
 832                 return ret;
 833         }
 834         gfp = vma_thp_gfp_mask(vma);
 835         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 836         if (unlikely(!folio)) {
 837                 count_vm_event(THP_FAULT_FALLBACK);
 838                 return VM_FAULT_FALLBACK;
 839         }
 840         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 841 }
 842
 843 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 844                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 845                 pgtable_t pgtable)
 846 {
 847         struct mm_struct *mm = vma->vm_mm;
 848         pmd_t entry;
 849         spinlock_t *ptl;
 850
 851         ptl = pmd_lock(mm, pmd);
 852         if (!pmd_none(*pmd)) {
 853                 if (write) {
 854                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 855                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 856                                 goto out_unlock;
 857                         }
 858                         entry = pmd_mkyoung(*pmd);
 859                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 860                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 861                                 update_mmu_cache_pmd(vma, addr, pmd);
 862                 }
 863
 864                 goto out_unlock;
 865         }
 866
 867         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 868         if (pfn_t_devmap(pfn))
 869                 entry = pmd_mkdevmap(entry);
 870         if (write) {
 871                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 872                 entry = maybe_pmd_mkwrite(entry, vma);
 873         }
 874
 875         if (pgtable) {
 876                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 877                 mm_inc_nr_ptes(mm);
 878                 pgtable = NULL;
 879         }
 880
 881         set_pmd_at(mm, addr, pmd, entry);
 882         update_mmu_cache_pmd(vma, addr, pmd);
 883
 884 out_unlock:
 885         spin_unlock(ptl);
 886         if (pgtable)
 887                 pte_free(mm, pgtable);
 888 }
 889
 890 /**
 891  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 892  * @vmf: Structure describing the fault
 893  * @pfn: pfn to insert
 894  * @pgprot: page protection to use
 895  * @write: whether it's a write fault
 896  *
 897  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 898  * also consult the vmf_insert_mixed_prot() documentation when
 899  * @pgprot != @vmf->vma->vm_page_prot.
 900  *
 901  * Return: vm_fault_t value.
 902  */
 903 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 904                                    pgprot_t pgprot, bool write)
 905 {
 906         unsigned long addr = vmf->address & PMD_MASK;
 907         struct vm_area_struct *vma = vmf->vma;
 908         pgtable_t pgtable = NULL;
 909
 910         /*
 911          * If we had pmd_special, we could avoid all these restrictions,
 912          * but we need to be consistent with PTEs and architectures that
 913          * can't support a 'special' bit.
 914          */
 915         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 916                         !pfn_t_devmap(pfn));
 917         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 918                                                 (VM_PFNMAP|VM_MIXEDMAP));
 919         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 920
 921         if (addr < vma->vm_start || addr >= vma->vm_end)
 922                 return VM_FAULT_SIGBUS;
 923
 924         if (arch_needs_pgtable_deposit()) {
 925                 pgtable = pte_alloc_one(vma->vm_mm);
 926                 if (!pgtable)
 927                         return VM_FAULT_OOM;
 928         }
 929
 930         track_pfn_insert(vma, &pgprot, pfn);
 931
 932         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 933         return VM_FAULT_NOPAGE;
 934 }
 935 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 936
 937 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 938 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 939 {
 940         if (likely(vma->vm_flags & VM_WRITE))
 941                 pud = pud_mkwrite(pud);
 942         return pud;
 943 }
 944
 945 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 946                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 947 {
 948         struct mm_struct *mm = vma->vm_mm;
 949         pud_t entry;
 950         spinlock_t *ptl;
 951
 952         ptl = pud_lock(mm, pud);
 953         if (!pud_none(*pud)) {
 954                 if (write) {
 955                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 956                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 957                                 goto out_unlock;
 958                         }
 959                         entry = pud_mkyoung(*pud);
 960                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 961                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 962                                 update_mmu_cache_pud(vma, addr, pud);
 963                 }
 964                 goto out_unlock;
 965         }
 966
 967         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 968         if (pfn_t_devmap(pfn))
 969                 entry = pud_mkdevmap(entry);
 970         if (write) {
 971                 entry = pud_mkyoung(pud_mkdirty(entry));
 972                 entry = maybe_pud_mkwrite(entry, vma);
 973         }
 974         set_pud_at(mm, addr, pud, entry);
 975         update_mmu_cache_pud(vma, addr, pud);
 976
 977 out_unlock:
 978         spin_unlock(ptl);
 979 }
 980
 981 /**
 982  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 983  * @vmf: Structure describing the fault
 984  * @pfn: pfn to insert
 985  * @pgprot: page protection to use
 986  * @write: whether it's a write fault
 987  *
 988  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 989  * also consult the vmf_insert_mixed_prot() documentation when
 990  * @pgprot != @vmf->vma->vm_page_prot.
 991  *
 992  * Return: vm_fault_t value.
 993  */
 994 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 995                                    pgprot_t pgprot, bool write)
 996 {
 997         unsigned long addr = vmf->address & PUD_MASK;
 998         struct vm_area_struct *vma = vmf->vma;
 999
1000         /*
1001          * If we had pud_special, we could avoid all these restrictions,
1002          * but we need to be consistent with PTEs and architectures that
1003          * can't support a 'special' bit.
1004          */
1005         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1006                         !pfn_t_devmap(pfn));
1007         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1008                                                 (VM_PFNMAP|VM_MIXEDMAP));
1009         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1010
1011         if (addr < vma->vm_start || addr >= vma->vm_end)
1012                 return VM_FAULT_SIGBUS;
1013
1014         track_pfn_insert(vma, &pgprot, pfn);
1015
1016         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1017         return VM_FAULT_NOPAGE;
1018 }
1019 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1020 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1021
1022 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1023                       pmd_t *pmd, bool write)
1024 {
1025         pmd_t _pmd;
1026
1027         _pmd = pmd_mkyoung(*pmd);
1028         if (write)
1029                 _pmd = pmd_mkdirty(_pmd);
1030         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031                                   pmd, _pmd, write))
1032                 update_mmu_cache_pmd(vma, addr, pmd);
1033 }
1034
1035 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1037 {
1038         unsigned long pfn = pmd_pfn(*pmd);
1039         struct mm_struct *mm = vma->vm_mm;
1040         struct page *page;
1041
1042         assert_spin_locked(pmd_lockptr(mm, pmd));
1043
1044         /*
1045          * When we COW a devmap PMD entry, we split it into PTEs, so we should
1046          * not be in this function with `flags & FOLL_COW` set.
1047          */
1048         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
1049
1050         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1051         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1052                          (FOLL_PIN | FOLL_GET)))
1053                 return NULL;
1054
1055         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1056                 return NULL;
1057
1058         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1059                 /* pass */;
1060         else
1061                 return NULL;
1062
1063         if (flags & FOLL_TOUCH)
1064                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1065
1066         /*
1067          * device mapped pages can only be returned if the
1068          * caller will manage the page reference count.
1069          */
1070         if (!(flags & (FOLL_GET | FOLL_PIN)))
1071                 return ERR_PTR(-EEXIST);
1072
1073         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1074         *pgmap = get_dev_pagemap(pfn, *pgmap);
1075         if (!*pgmap)
1076                 return ERR_PTR(-EFAULT);
1077         page = pfn_to_page(pfn);
1078         if (!try_grab_page(page, flags))
1079                 page = ERR_PTR(-ENOMEM);
1080
1081         return page;
1082 }
1083
1084 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1085                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1086                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1087 {
1088         spinlock_t *dst_ptl, *src_ptl;
1089         struct page *src_page;
1090         pmd_t pmd;
1091         pgtable_t pgtable = NULL;
1092         int ret = -ENOMEM;
1093
1094         /* Skip if can be re-fill on fault */
1095         if (!vma_is_anonymous(dst_vma))
1096                 return 0;
1097
1098         pgtable = pte_alloc_one(dst_mm);
1099         if (unlikely(!pgtable))
1100                 goto out;
1101
1102         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1103         src_ptl = pmd_lockptr(src_mm, src_pmd);
1104         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1105
1106         ret = -EAGAIN;
1107         pmd = *src_pmd;
1108
1109 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1110         if (unlikely(is_swap_pmd(pmd))) {
1111                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1112
1113                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1114                 if (!is_readable_migration_entry(entry)) {
1115                         entry = make_readable_migration_entry(
1116                                                         swp_offset(entry));
1117                         pmd = swp_entry_to_pmd(entry);
1118                         if (pmd_swp_soft_dirty(*src_pmd))
1119                                 pmd = pmd_swp_mksoft_dirty(pmd);
1120                         if (pmd_swp_uffd_wp(*src_pmd))
1121                                 pmd = pmd_swp_mkuffd_wp(pmd);
1122                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1123                 }
1124                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1125                 mm_inc_nr_ptes(dst_mm);
1126                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1127                 if (!userfaultfd_wp(dst_vma))
1128                         pmd = pmd_swp_clear_uffd_wp(pmd);
1129                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1130                 ret = 0;
1131                 goto out_unlock;
1132         }
1133 #endif
1134
1135         if (unlikely(!pmd_trans_huge(pmd))) {
1136                 pte_free(dst_mm, pgtable);
1137                 goto out_unlock;
1138         }
1139         /*
1140          * When page table lock is held, the huge zero pmd should not be
1141          * under splitting since we don't split the page itself, only pmd to
1142          * a page table.
1143          */
1144         if (is_huge_zero_pmd(pmd)) {
1145                 /*
1146                  * get_huge_zero_page() will never allocate a new page here,
1147                  * since we already have a zero page to copy. It just takes a
1148                  * reference.
1149                  */
1150                 mm_get_huge_zero_page(dst_mm);
1151                 goto out_zero_page;
1152         }
1153
1154         src_page = pmd_page(pmd);
1155         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1156
1157         get_page(src_page);
1158         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1159                 /* Page maybe pinned: split and retry the fault on PTEs. */
1160                 put_page(src_page);
1161                 pte_free(dst_mm, pgtable);
1162                 spin_unlock(src_ptl);
1163                 spin_unlock(dst_ptl);
1164                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1165                 return -EAGAIN;
1166         }
1167         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1168 out_zero_page:
1169         mm_inc_nr_ptes(dst_mm);
1170         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1171         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1172         if (!userfaultfd_wp(dst_vma))
1173                 pmd = pmd_clear_uffd_wp(pmd);
1174         pmd = pmd_mkold(pmd_wrprotect(pmd));
1175         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1176
1177         ret = 0;
1178 out_unlock:
1179         spin_unlock(src_ptl);
1180         spin_unlock(dst_ptl);
1181 out:
1182         return ret;
1183 }
1184
1185 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1186 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1187                       pud_t *pud, bool write)
1188 {
1189         pud_t _pud;
1190
1191         _pud = pud_mkyoung(*pud);
1192         if (write)
1193                 _pud = pud_mkdirty(_pud);
1194         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1195                                   pud, _pud, write))
1196                 update_mmu_cache_pud(vma, addr, pud);
1197 }
1198
1199 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1200                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1201 {
1202         unsigned long pfn = pud_pfn(*pud);
1203         struct mm_struct *mm = vma->vm_mm;
1204         struct page *page;
1205
1206         assert_spin_locked(pud_lockptr(mm, pud));
1207
1208         if (flags & FOLL_WRITE && !pud_write(*pud))
1209                 return NULL;
1210
1211         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1212         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1213                          (FOLL_PIN | FOLL_GET)))
1214                 return NULL;
1215
1216         if (pud_present(*pud) && pud_devmap(*pud))
1217                 /* pass */;
1218         else
1219                 return NULL;
1220
1221         if (flags & FOLL_TOUCH)
1222                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1223
1224         /*
1225          * device mapped pages can only be returned if the
1226          * caller will manage the page reference count.
1227          *
1228          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1229          */
1230         if (!(flags & (FOLL_GET | FOLL_PIN)))
1231                 return ERR_PTR(-EEXIST);
1232
1233         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1234         *pgmap = get_dev_pagemap(pfn, *pgmap);
1235         if (!*pgmap)
1236                 return ERR_PTR(-EFAULT);
1237         page = pfn_to_page(pfn);
1238         if (!try_grab_page(page, flags))
1239                 page = ERR_PTR(-ENOMEM);
1240
1241         return page;
1242 }
1243
1244 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1245                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1246                   struct vm_area_struct *vma)
1247 {
1248         spinlock_t *dst_ptl, *src_ptl;
1249         pud_t pud;
1250         int ret;
1251
1252         dst_ptl = pud_lock(dst_mm, dst_pud);
1253         src_ptl = pud_lockptr(src_mm, src_pud);
1254         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1255
1256         ret = -EAGAIN;
1257         pud = *src_pud;
1258         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1259                 goto out_unlock;
1260
1261         /*
1262          * When page table lock is held, the huge zero pud should not be
1263          * under splitting since we don't split the page itself, only pud to
1264          * a page table.
1265          */
1266         if (is_huge_zero_pud(pud)) {
1267                 /* No huge zero pud yet */
1268         }
1269
1270         /*
1271          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1272          * and split if duplicating fails.
1273          */
1274         pudp_set_wrprotect(src_mm, addr, src_pud);
1275         pud = pud_mkold(pud_wrprotect(pud));
1276         set_pud_at(dst_mm, addr, dst_pud, pud);
1277
1278         ret = 0;
1279 out_unlock:
1280         spin_unlock(src_ptl);
1281         spin_unlock(dst_ptl);
1282         return ret;
1283 }
1284
1285 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1286 {
1287         bool write = vmf->flags & FAULT_FLAG_WRITE;
1288
1289         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1290         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1291                 goto unlock;
1292
1293         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1294 unlock:
1295         spin_unlock(vmf->ptl);
1296 }
1297 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1298
1299 void huge_pmd_set_accessed(struct vm_fault *vmf)
1300 {
1301         bool write = vmf->flags & FAULT_FLAG_WRITE;
1302
1303         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1304         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1305                 goto unlock;
1306
1307         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1308
1309 unlock:
1310         spin_unlock(vmf->ptl);
1311 }
1312
1313 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1314 {
1315         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1316         struct vm_area_struct *vma = vmf->vma;
1317         struct page *page;
1318         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1319         pmd_t orig_pmd = vmf->orig_pmd;
1320
1321         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1322         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1323
1324         VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE));
1325         VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE));
1326
1327         if (is_huge_zero_pmd(orig_pmd))
1328                 goto fallback;
1329
1330         spin_lock(vmf->ptl);
1331
1332         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1333                 spin_unlock(vmf->ptl);
1334                 return 0;
1335         }
1336
1337         page = pmd_page(orig_pmd);
1338         VM_BUG_ON_PAGE(!PageHead(page), page);
1339
1340         /* Early check when only holding the PT lock. */
1341         if (PageAnonExclusive(page))
1342                 goto reuse;
1343
1344         if (!trylock_page(page)) {
1345                 get_page(page);
1346                 spin_unlock(vmf->ptl);
1347                 lock_page(page);
1348                 spin_lock(vmf->ptl);
1349                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1350                         spin_unlock(vmf->ptl);
1351                         unlock_page(page);
1352                         put_page(page);
1353                         return 0;
1354                 }
1355                 put_page(page);
1356         }
1357
1358         /* Recheck after temporarily dropping the PT lock. */
1359         if (PageAnonExclusive(page)) {
1360                 unlock_page(page);
1361                 goto reuse;
1362         }
1363
1364         /*
1365          * See do_wp_page(): we can only reuse the page exclusively if there are
1366          * no additional references. Note that we always drain the LRU
1367          * pagevecs immediately after adding a THP.
1368          */
1369         if (page_count(page) > 1 + PageSwapCache(page) * thp_nr_pages(page))
1370                 goto unlock_fallback;
1371         if (PageSwapCache(page))
1372                 try_to_free_swap(page);
1373         if (page_count(page) == 1) {
1374                 pmd_t entry;
1375
1376                 page_move_anon_rmap(page, vma);
1377                 unlock_page(page);
1378 reuse:
1379                 if (unlikely(unshare)) {
1380                         spin_unlock(vmf->ptl);
1381                         return 0;
1382                 }
1383                 entry = pmd_mkyoung(orig_pmd);
1384                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1385                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1386                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1387                 spin_unlock(vmf->ptl);
1388                 return VM_FAULT_WRITE;
1389         }
1390
1391 unlock_fallback:
1392         unlock_page(page);
1393         spin_unlock(vmf->ptl);
1394 fallback:
1395         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1396         return VM_FAULT_FALLBACK;
1397 }
1398
1399 /*
1400  * FOLL_FORCE can write to even unwritable pmd's, but only
1401  * after we've gone through a COW cycle and they are dirty.
1402  */
1403 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1404 {
1405         return pmd_write(pmd) ||
1406                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1407 }
1408
1409 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1410                                    unsigned long addr,
1411                                    pmd_t *pmd,
1412                                    unsigned int flags)
1413 {
1414         struct mm_struct *mm = vma->vm_mm;
1415         struct page *page = NULL;
1416
1417         assert_spin_locked(pmd_lockptr(mm, pmd));
1418
1419         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1420                 goto out;
1421
1422         /* Avoid dumping huge zero page */
1423         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1424                 return ERR_PTR(-EFAULT);
1425
1426         /* Full NUMA hinting faults to serialise migration in fault paths */
1427         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1428                 goto out;
1429
1430         page = pmd_page(*pmd);
1431         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1432
1433         if (!pmd_write(*pmd) && gup_must_unshare(flags, page))
1434                 return ERR_PTR(-EMLINK);
1435
1436         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1437                         !PageAnonExclusive(page), page);
1438
1439         if (!try_grab_page(page, flags))
1440                 return ERR_PTR(-ENOMEM);
1441
1442         if (flags & FOLL_TOUCH)
1443                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1444
1445         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1446         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1447
1448 out:
1449         return page;
1450 }
1451
1452 /* NUMA hinting page fault entry point for trans huge pmds */
1453 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1454 {
1455         struct vm_area_struct *vma = vmf->vma;
1456         pmd_t oldpmd = vmf->orig_pmd;
1457         pmd_t pmd;
1458         struct page *page;
1459         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1460         int page_nid = NUMA_NO_NODE;
1461         int target_nid, last_cpupid = -1;
1462         bool migrated = false;
1463         bool was_writable = pmd_savedwrite(oldpmd);
1464         int flags = 0;
1465
1466         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1467         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1468                 spin_unlock(vmf->ptl);
1469                 goto out;
1470         }
1471
1472         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1473         page = vm_normal_page_pmd(vma, haddr, pmd);
1474         if (!page)
1475                 goto out_map;
1476
1477         /* See similar comment in do_numa_page for explanation */
1478         if (!was_writable)
1479                 flags |= TNF_NO_GROUP;
1480
1481         page_nid = page_to_nid(page);
1482         last_cpupid = page_cpupid_last(page);
1483         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1484                                        &flags);
1485
1486         if (target_nid == NUMA_NO_NODE) {
1487                 put_page(page);
1488                 goto out_map;
1489         }
1490
1491         spin_unlock(vmf->ptl);
1492
1493         migrated = migrate_misplaced_page(page, vma, target_nid);
1494         if (migrated) {
1495                 flags |= TNF_MIGRATED;
1496                 page_nid = target_nid;
1497         } else {
1498                 flags |= TNF_MIGRATE_FAIL;
1499                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1500                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1501                         spin_unlock(vmf->ptl);
1502                         goto out;
1503                 }
1504                 goto out_map;
1505         }
1506
1507 out:
1508         if (page_nid != NUMA_NO_NODE)
1509                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1510                                 flags);
1511
1512         return 0;
1513
1514 out_map:
1515         /* Restore the PMD */
1516         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1517         pmd = pmd_mkyoung(pmd);
1518         if (was_writable)
1519                 pmd = pmd_mkwrite(pmd);
1520         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1521         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1522         spin_unlock(vmf->ptl);
1523         goto out;
1524 }
1525
1526 /*
1527  * Return true if we do MADV_FREE successfully on entire pmd page.
1528  * Otherwise, return false.
1529  */
1530 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1531                 pmd_t *pmd, unsigned long addr, unsigned long next)
1532 {
1533         spinlock_t *ptl;
1534         pmd_t orig_pmd;
1535         struct page *page;
1536         struct mm_struct *mm = tlb->mm;
1537         bool ret = false;
1538
1539         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1540
1541         ptl = pmd_trans_huge_lock(pmd, vma);
1542         if (!ptl)
1543                 goto out_unlocked;
1544
1545         orig_pmd = *pmd;
1546         if (is_huge_zero_pmd(orig_pmd))
1547                 goto out;
1548
1549         if (unlikely(!pmd_present(orig_pmd))) {
1550                 VM_BUG_ON(thp_migration_supported() &&
1551                                   !is_pmd_migration_entry(orig_pmd));
1552                 goto out;
1553         }
1554
1555         page = pmd_page(orig_pmd);
1556         /*
1557          * If other processes are mapping this page, we couldn't discard
1558          * the page unless they all do MADV_FREE so let's skip the page.
1559          */
1560         if (total_mapcount(page) != 1)
1561                 goto out;
1562
1563         if (!trylock_page(page))
1564                 goto out;
1565
1566         /*
1567          * If user want to discard part-pages of THP, split it so MADV_FREE
1568          * will deactivate only them.
1569          */
1570         if (next - addr != HPAGE_PMD_SIZE) {
1571                 get_page(page);
1572                 spin_unlock(ptl);
1573                 split_huge_page(page);
1574                 unlock_page(page);
1575                 put_page(page);
1576                 goto out_unlocked;
1577         }
1578
1579         if (PageDirty(page))
1580                 ClearPageDirty(page);
1581         unlock_page(page);
1582
1583         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1584                 pmdp_invalidate(vma, addr, pmd);
1585                 orig_pmd = pmd_mkold(orig_pmd);
1586                 orig_pmd = pmd_mkclean(orig_pmd);
1587
1588                 set_pmd_at(mm, addr, pmd, orig_pmd);
1589                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1590         }
1591
1592         mark_page_lazyfree(page);
1593         ret = true;
1594 out:
1595         spin_unlock(ptl);
1596 out_unlocked:
1597         return ret;
1598 }
1599
1600 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1601 {
1602         pgtable_t pgtable;
1603
1604         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1605         pte_free(mm, pgtable);
1606         mm_dec_nr_ptes(mm);
1607 }
1608
1609 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1610                  pmd_t *pmd, unsigned long addr)
1611 {
1612         pmd_t orig_pmd;
1613         spinlock_t *ptl;
1614
1615         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1616
1617         ptl = __pmd_trans_huge_lock(pmd, vma);
1618         if (!ptl)
1619                 return 0;
1620         /*
1621          * For architectures like ppc64 we look at deposited pgtable
1622          * when calling pmdp_huge_get_and_clear. So do the
1623          * pgtable_trans_huge_withdraw after finishing pmdp related
1624          * operations.
1625          */
1626         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1627                                                 tlb->fullmm);
1628         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1629         if (vma_is_special_huge(vma)) {
1630                 if (arch_needs_pgtable_deposit())
1631                         zap_deposited_table(tlb->mm, pmd);
1632                 spin_unlock(ptl);
1633         } else if (is_huge_zero_pmd(orig_pmd)) {
1634                 zap_deposited_table(tlb->mm, pmd);
1635                 spin_unlock(ptl);
1636         } else {
1637                 struct page *page = NULL;
1638                 int flush_needed = 1;
1639
1640                 if (pmd_present(orig_pmd)) {
1641                         page = pmd_page(orig_pmd);
1642                         page_remove_rmap(page, vma, true);
1643                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1644                         VM_BUG_ON_PAGE(!PageHead(page), page);
1645                 } else if (thp_migration_supported()) {
1646                         swp_entry_t entry;
1647
1648                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1649                         entry = pmd_to_swp_entry(orig_pmd);
1650                         page = pfn_swap_entry_to_page(entry);
1651                         flush_needed = 0;
1652                 } else
1653                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1654
1655                 if (PageAnon(page)) {
1656                         zap_deposited_table(tlb->mm, pmd);
1657                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1658                 } else {
1659                         if (arch_needs_pgtable_deposit())
1660                                 zap_deposited_table(tlb->mm, pmd);
1661                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1662                 }
1663
1664                 spin_unlock(ptl);
1665                 if (flush_needed)
1666                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1667         }
1668         return 1;
1669 }
1670
1671 #ifndef pmd_move_must_withdraw
1672 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1673                                          spinlock_t *old_pmd_ptl,
1674                                          struct vm_area_struct *vma)
1675 {
1676         /*
1677          * With split pmd lock we also need to move preallocated
1678          * PTE page table if new_pmd is on different PMD page table.
1679          *
1680          * We also don't deposit and withdraw tables for file pages.
1681          */
1682         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1683 }
1684 #endif
1685
1686 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1687 {
1688 #ifdef CONFIG_MEM_SOFT_DIRTY
1689         if (unlikely(is_pmd_migration_entry(pmd)))
1690                 pmd = pmd_swp_mksoft_dirty(pmd);
1691         else if (pmd_present(pmd))
1692                 pmd = pmd_mksoft_dirty(pmd);
1693 #endif
1694         return pmd;
1695 }
1696
1697 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1698                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1699 {
1700         spinlock_t *old_ptl, *new_ptl;
1701         pmd_t pmd;
1702         struct mm_struct *mm = vma->vm_mm;
1703         bool force_flush = false;
1704
1705         /*
1706          * The destination pmd shouldn't be established, free_pgtables()
1707          * should have release it.
1708          */
1709         if (WARN_ON(!pmd_none(*new_pmd))) {
1710                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1711                 return false;
1712         }
1713
1714         /*
1715          * We don't have to worry about the ordering of src and dst
1716          * ptlocks because exclusive mmap_lock prevents deadlock.
1717          */
1718         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1719         if (old_ptl) {
1720                 new_ptl = pmd_lockptr(mm, new_pmd);
1721                 if (new_ptl != old_ptl)
1722                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1723                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1724                 if (pmd_present(pmd))
1725                         force_flush = true;
1726                 VM_BUG_ON(!pmd_none(*new_pmd));
1727
1728                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1729                         pgtable_t pgtable;
1730                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1731                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1732                 }
1733                 pmd = move_soft_dirty_pmd(pmd);
1734                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1735                 if (force_flush)
1736                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1737                 if (new_ptl != old_ptl)
1738                         spin_unlock(new_ptl);
1739                 spin_unlock(old_ptl);
1740                 return true;
1741         }
1742         return false;
1743 }
1744
1745 /*
1746  * Returns
1747  *  - 0 if PMD could not be locked
1748  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1749  *      or if prot_numa but THP migration is not supported
1750  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1751  */
1752 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1753                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1754                     unsigned long cp_flags)
1755 {
1756         struct mm_struct *mm = vma->vm_mm;
1757         spinlock_t *ptl;
1758         pmd_t oldpmd, entry;
1759         bool preserve_write;
1760         int ret;
1761         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1762         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1763         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1764
1765         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1766
1767         if (prot_numa && !thp_migration_supported())
1768                 return 1;
1769
1770         ptl = __pmd_trans_huge_lock(pmd, vma);
1771         if (!ptl)
1772                 return 0;
1773
1774         preserve_write = prot_numa && pmd_write(*pmd);
1775         ret = 1;
1776
1777 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1778         if (is_swap_pmd(*pmd)) {
1779                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1780                 struct page *page = pfn_swap_entry_to_page(entry);
1781
1782                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1783                 if (is_writable_migration_entry(entry)) {
1784                         pmd_t newpmd;
1785                         /*
1786                          * A protection check is difficult so
1787                          * just be safe and disable write
1788                          */
1789                         if (PageAnon(page))
1790                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1791                         else
1792                                 entry = make_readable_migration_entry(swp_offset(entry));
1793                         newpmd = swp_entry_to_pmd(entry);
1794                         if (pmd_swp_soft_dirty(*pmd))
1795                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1796                         if (pmd_swp_uffd_wp(*pmd))
1797                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1798                         set_pmd_at(mm, addr, pmd, newpmd);
1799                 }
1800                 goto unlock;
1801         }
1802 #endif
1803
1804         if (prot_numa) {
1805                 struct page *page;
1806                 /*
1807                  * Avoid trapping faults against the zero page. The read-only
1808                  * data is likely to be read-cached on the local CPU and
1809                  * local/remote hits to the zero page are not interesting.
1810                  */
1811                 if (is_huge_zero_pmd(*pmd))
1812                         goto unlock;
1813
1814                 if (pmd_protnone(*pmd))
1815                         goto unlock;
1816
1817                 page = pmd_page(*pmd);
1818                 /*
1819                  * Skip scanning top tier node if normal numa
1820                  * balancing is disabled
1821                  */
1822                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1823                     node_is_toptier(page_to_nid(page)))
1824                         goto unlock;
1825         }
1826         /*
1827          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1828          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1829          * which is also under mmap_read_lock(mm):
1830          *
1831          *      CPU0:                           CPU1:
1832          *                              change_huge_pmd(prot_numa=1)
1833          *                               pmdp_huge_get_and_clear_notify()
1834          * madvise_dontneed()
1835          *  zap_pmd_range()
1836          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1837          *   // skip the pmd
1838          *                               set_pmd_at();
1839          *                               // pmd is re-established
1840          *
1841          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1842          * which may break userspace.
1843          *
1844          * pmdp_invalidate_ad() is required to make sure we don't miss
1845          * dirty/young flags set by hardware.
1846          */
1847         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1848
1849         entry = pmd_modify(oldpmd, newprot);
1850         if (preserve_write)
1851                 entry = pmd_mk_savedwrite(entry);
1852         if (uffd_wp) {
1853                 entry = pmd_wrprotect(entry);
1854                 entry = pmd_mkuffd_wp(entry);
1855         } else if (uffd_wp_resolve) {
1856                 /*
1857                  * Leave the write bit to be handled by PF interrupt
1858                  * handler, then things like COW could be properly
1859                  * handled.
1860                  */
1861                 entry = pmd_clear_uffd_wp(entry);
1862         }
1863         ret = HPAGE_PMD_NR;
1864         set_pmd_at(mm, addr, pmd, entry);
1865
1866         if (huge_pmd_needs_flush(oldpmd, entry))
1867                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1868
1869         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1870 unlock:
1871         spin_unlock(ptl);
1872         return ret;
1873 }
1874
1875 /*
1876  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1877  *
1878  * Note that if it returns page table lock pointer, this routine returns without
1879  * unlocking page table lock. So callers must unlock it.
1880  */
1881 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1882 {
1883         spinlock_t *ptl;
1884         ptl = pmd_lock(vma->vm_mm, pmd);
1885         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1886                         pmd_devmap(*pmd)))
1887                 return ptl;
1888         spin_unlock(ptl);
1889         return NULL;
1890 }
1891
1892 /*
1893  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1894  *
1895  * Note that if it returns page table lock pointer, this routine returns without
1896  * unlocking page table lock. So callers must unlock it.
1897  */
1898 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1899 {
1900         spinlock_t *ptl;
1901
1902         ptl = pud_lock(vma->vm_mm, pud);
1903         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1904                 return ptl;
1905         spin_unlock(ptl);
1906         return NULL;
1907 }
1908
1909 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1910 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1911                  pud_t *pud, unsigned long addr)
1912 {
1913         spinlock_t *ptl;
1914
1915         ptl = __pud_trans_huge_lock(pud, vma);
1916         if (!ptl)
1917                 return 0;
1918         /*
1919          * For architectures like ppc64 we look at deposited pgtable
1920          * when calling pudp_huge_get_and_clear. So do the
1921          * pgtable_trans_huge_withdraw after finishing pudp related
1922          * operations.
1923          */
1924         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1925         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1926         if (vma_is_special_huge(vma)) {
1927                 spin_unlock(ptl);
1928                 /* No zero page support yet */
1929         } else {
1930                 /* No support for anonymous PUD pages yet */
1931                 BUG();
1932         }
1933         return 1;
1934 }
1935
1936 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1937                 unsigned long haddr)
1938 {
1939         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1940         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1941         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1942         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1943
1944         count_vm_event(THP_SPLIT_PUD);
1945
1946         pudp_huge_clear_flush_notify(vma, haddr, pud);
1947 }
1948
1949 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1950                 unsigned long address)
1951 {
1952         spinlock_t *ptl;
1953         struct mmu_notifier_range range;
1954
1955         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1956                                 address & HPAGE_PUD_MASK,
1957                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1958         mmu_notifier_invalidate_range_start(&range);
1959         ptl = pud_lock(vma->vm_mm, pud);
1960         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1961                 goto out;
1962         __split_huge_pud_locked(vma, pud, range.start);
1963
1964 out:
1965         spin_unlock(ptl);
1966         /*
1967          * No need to double call mmu_notifier->invalidate_range() callback as
1968          * the above pudp_huge_clear_flush_notify() did already call it.
1969          */
1970         mmu_notifier_invalidate_range_only_end(&range);
1971 }
1972 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1973
1974 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1975                 unsigned long haddr, pmd_t *pmd)
1976 {
1977         struct mm_struct *mm = vma->vm_mm;
1978         pgtable_t pgtable;
1979         pmd_t _pmd;
1980         int i;
1981
1982         /*
1983          * Leave pmd empty until pte is filled note that it is fine to delay
1984          * notification until mmu_notifier_invalidate_range_end() as we are
1985          * replacing a zero pmd write protected page with a zero pte write
1986          * protected page.
1987          *
1988          * See Documentation/mm/mmu_notifier.rst
1989          */
1990         pmdp_huge_clear_flush(vma, haddr, pmd);
1991
1992         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1993         pmd_populate(mm, &_pmd, pgtable);
1994
1995         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1996                 pte_t *pte, entry;
1997                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1998                 entry = pte_mkspecial(entry);
1999                 pte = pte_offset_map(&_pmd, haddr);
2000                 VM_BUG_ON(!pte_none(*pte));
2001                 set_pte_at(mm, haddr, pte, entry);
2002                 pte_unmap(pte);
2003         }
2004         smp_wmb(); /* make pte visible before pmd */
2005         pmd_populate(mm, pmd, pgtable);
2006 }
2007
2008 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2009                 unsigned long haddr, bool freeze)
2010 {
2011         struct mm_struct *mm = vma->vm_mm;
2012         struct page *page;
2013         pgtable_t pgtable;
2014         pmd_t old_pmd, _pmd;
2015         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2016         bool anon_exclusive = false;
2017         unsigned long addr;
2018         int i;
2019
2020         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2021         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2022         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2023         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2024                                 && !pmd_devmap(*pmd));
2025
2026         count_vm_event(THP_SPLIT_PMD);
2027
2028         if (!vma_is_anonymous(vma)) {
2029                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2030                 /*
2031                  * We are going to unmap this huge page. So
2032                  * just go ahead and zap it
2033                  */
2034                 if (arch_needs_pgtable_deposit())
2035                         zap_deposited_table(mm, pmd);
2036                 if (vma_is_special_huge(vma))
2037                         return;
2038                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2039                         swp_entry_t entry;
2040
2041                         entry = pmd_to_swp_entry(old_pmd);
2042                         page = pfn_swap_entry_to_page(entry);
2043                 } else {
2044                         page = pmd_page(old_pmd);
2045                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2046                                 set_page_dirty(page);
2047                         if (!PageReferenced(page) && pmd_young(old_pmd))
2048                                 SetPageReferenced(page);
2049                         page_remove_rmap(page, vma, true);
2050                         put_page(page);
2051                 }
2052                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2053                 return;
2054         }
2055
2056         if (is_huge_zero_pmd(*pmd)) {
2057                 /*
2058                  * FIXME: Do we want to invalidate secondary mmu by calling
2059                  * mmu_notifier_invalidate_range() see comments below inside
2060                  * __split_huge_pmd() ?
2061                  *
2062                  * We are going from a zero huge page write protected to zero
2063                  * small page also write protected so it does not seems useful
2064                  * to invalidate secondary mmu at this time.
2065                  */
2066                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2067         }
2068
2069         /*
2070          * Up to this point the pmd is present and huge and userland has the
2071          * whole access to the hugepage during the split (which happens in
2072          * place). If we overwrite the pmd with the not-huge version pointing
2073          * to the pte here (which of course we could if all CPUs were bug
2074          * free), userland could trigger a small page size TLB miss on the
2075          * small sized TLB while the hugepage TLB entry is still established in
2076          * the huge TLB. Some CPU doesn't like that.
2077          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2078          * 383 on page 105. Intel should be safe but is also warns that it's
2079          * only safe if the permission and cache attributes of the two entries
2080          * loaded in the two TLB is identical (which should be the case here).
2081          * But it is generally safer to never allow small and huge TLB entries
2082          * for the same virtual address to be loaded simultaneously. So instead
2083          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2084          * current pmd notpresent (atomically because here the pmd_trans_huge
2085          * must remain set at all times on the pmd until the split is complete
2086          * for this pmd), then we flush the SMP TLB and finally we write the
2087          * non-huge version of the pmd entry with pmd_populate.
2088          */
2089         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2090
2091         pmd_migration = is_pmd_migration_entry(old_pmd);
2092         if (unlikely(pmd_migration)) {
2093                 swp_entry_t entry;
2094
2095                 entry = pmd_to_swp_entry(old_pmd);
2096                 page = pfn_swap_entry_to_page(entry);
2097                 write = is_writable_migration_entry(entry);
2098                 if (PageAnon(page))
2099                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2100                 young = false;
2101                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2102                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2103         } else {
2104                 page = pmd_page(old_pmd);
2105                 if (pmd_dirty(old_pmd))
2106                         SetPageDirty(page);
2107                 write = pmd_write(old_pmd);
2108                 young = pmd_young(old_pmd);
2109                 soft_dirty = pmd_soft_dirty(old_pmd);
2110                 uffd_wp = pmd_uffd_wp(old_pmd);
2111
2112                 VM_BUG_ON_PAGE(!page_count(page), page);
2113                 page_ref_add(page, HPAGE_PMD_NR - 1);
2114
2115                 /*
2116                  * Without "freeze", we'll simply split the PMD, propagating the
2117                  * PageAnonExclusive() flag for each PTE by setting it for
2118                  * each subpage -- no need to (temporarily) clear.
2119                  *
2120                  * With "freeze" we want to replace mapped pages by
2121                  * migration entries right away. This is only possible if we
2122                  * managed to clear PageAnonExclusive() -- see
2123                  * set_pmd_migration_entry().
2124                  *
2125                  * In case we cannot clear PageAnonExclusive(), split the PMD
2126                  * only and let try_to_migrate_one() fail later.
2127                  */
2128                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2129                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2130                         freeze = false;
2131         }
2132
2133         /*
2134          * Withdraw the table only after we mark the pmd entry invalid.
2135          * This's critical for some architectures (Power).
2136          */
2137         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2138         pmd_populate(mm, &_pmd, pgtable);
2139
2140         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2141                 pte_t entry, *pte;
2142                 /*
2143                  * Note that NUMA hinting access restrictions are not
2144                  * transferred to avoid any possibility of altering
2145                  * permissions across VMAs.
2146                  */
2147                 if (freeze || pmd_migration) {
2148                         swp_entry_t swp_entry;
2149                         if (write)
2150                                 swp_entry = make_writable_migration_entry(
2151                                                         page_to_pfn(page + i));
2152                         else if (anon_exclusive)
2153                                 swp_entry = make_readable_exclusive_migration_entry(
2154                                                         page_to_pfn(page + i));
2155                         else
2156                                 swp_entry = make_readable_migration_entry(
2157                                                         page_to_pfn(page + i));
2158                         entry = swp_entry_to_pte(swp_entry);
2159                         if (soft_dirty)
2160                                 entry = pte_swp_mksoft_dirty(entry);
2161                         if (uffd_wp)
2162                                 entry = pte_swp_mkuffd_wp(entry);
2163                 } else {
2164                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2165                         entry = maybe_mkwrite(entry, vma);
2166                         if (anon_exclusive)
2167                                 SetPageAnonExclusive(page + i);
2168                         if (!write)
2169                                 entry = pte_wrprotect(entry);
2170                         if (!young)
2171                                 entry = pte_mkold(entry);
2172                         if (soft_dirty)
2173                                 entry = pte_mksoft_dirty(entry);
2174                         if (uffd_wp)
2175                                 entry = pte_mkuffd_wp(entry);
2176                 }
2177                 pte = pte_offset_map(&_pmd, addr);
2178                 BUG_ON(!pte_none(*pte));
2179                 set_pte_at(mm, addr, pte, entry);
2180                 if (!pmd_migration)
2181                         atomic_inc(&page[i]._mapcount);
2182                 pte_unmap(pte);
2183         }
2184
2185         if (!pmd_migration) {
2186                 /*
2187                  * Set PG_double_map before dropping compound_mapcount to avoid
2188                  * false-negative page_mapped().
2189                  */
2190                 if (compound_mapcount(page) > 1 &&
2191                     !TestSetPageDoubleMap(page)) {
2192                         for (i = 0; i < HPAGE_PMD_NR; i++)
2193                                 atomic_inc(&page[i]._mapcount);
2194                 }
2195
2196                 lock_page_memcg(page);
2197                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2198                         /* Last compound_mapcount is gone. */
2199                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2200                                                 -HPAGE_PMD_NR);
2201                         if (TestClearPageDoubleMap(page)) {
2202                                 /* No need in mapcount reference anymore */
2203                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2204                                         atomic_dec(&page[i]._mapcount);
2205                         }
2206                 }
2207                 unlock_page_memcg(page);
2208
2209                 /* Above is effectively page_remove_rmap(page, vma, true) */
2210                 munlock_vma_page(page, vma, true);
2211         }
2212
2213         smp_wmb(); /* make pte visible before pmd */
2214         pmd_populate(mm, pmd, pgtable);
2215
2216         if (freeze) {
2217                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2218                         page_remove_rmap(page + i, vma, false);
2219                         put_page(page + i);
2220                 }
2221         }
2222 }
2223
2224 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2225                 unsigned long address, bool freeze, struct folio *folio)
2226 {
2227         spinlock_t *ptl;
2228         struct mmu_notifier_range range;
2229
2230         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2231                                 address & HPAGE_PMD_MASK,
2232                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2233         mmu_notifier_invalidate_range_start(&range);
2234         ptl = pmd_lock(vma->vm_mm, pmd);
2235
2236         /*
2237          * If caller asks to setup a migration entry, we need a folio to check
2238          * pmd against. Otherwise we can end up replacing wrong folio.
2239          */
2240         VM_BUG_ON(freeze && !folio);
2241         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2242
2243         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2244             is_pmd_migration_entry(*pmd)) {
2245                 if (folio && folio != page_folio(pmd_page(*pmd)))
2246                         goto out;
2247                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2248         }
2249
2250 out:
2251         spin_unlock(ptl);
2252         /*
2253          * No need to double call mmu_notifier->invalidate_range() callback.
2254          * They are 3 cases to consider inside __split_huge_pmd_locked():
2255          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2256          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2257          *    fault will trigger a flush_notify before pointing to a new page
2258          *    (it is fine if the secondary mmu keeps pointing to the old zero
2259          *    page in the meantime)
2260          *  3) Split a huge pmd into pte pointing to the same page. No need
2261          *     to invalidate secondary tlb entry they are all still valid.
2262          *     any further changes to individual pte will notify. So no need
2263          *     to call mmu_notifier->invalidate_range()
2264          */
2265         mmu_notifier_invalidate_range_only_end(&range);
2266 }
2267
2268 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2269                 bool freeze, struct folio *folio)
2270 {
2271         pgd_t *pgd;
2272         p4d_t *p4d;
2273         pud_t *pud;
2274         pmd_t *pmd;
2275
2276         pgd = pgd_offset(vma->vm_mm, address);
2277         if (!pgd_present(*pgd))
2278                 return;
2279
2280         p4d = p4d_offset(pgd, address);
2281         if (!p4d_present(*p4d))
2282                 return;
2283
2284         pud = pud_offset(p4d, address);
2285         if (!pud_present(*pud))
2286                 return;
2287
2288         pmd = pmd_offset(pud, address);
2289
2290         __split_huge_pmd(vma, pmd, address, freeze, folio);
2291 }
2292
2293 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2294 {
2295         /*
2296          * If the new address isn't hpage aligned and it could previously
2297          * contain an hugepage: check if we need to split an huge pmd.
2298          */
2299         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2300             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2301                          ALIGN(address, HPAGE_PMD_SIZE)))
2302                 split_huge_pmd_address(vma, address, false, NULL);
2303 }
2304
2305 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2306                              unsigned long start,
2307                              unsigned long end,
2308                              long adjust_next)
2309 {
2310         /* Check if we need to split start first. */
2311         split_huge_pmd_if_needed(vma, start);
2312
2313         /* Check if we need to split end next. */
2314         split_huge_pmd_if_needed(vma, end);
2315
2316         /*
2317          * If we're also updating the vma->vm_next->vm_start,
2318          * check if we need to split it.
2319          */
2320         if (adjust_next > 0) {
2321                 struct vm_area_struct *next = vma->vm_next;
2322                 unsigned long nstart = next->vm_start;
2323                 nstart += adjust_next;
2324                 split_huge_pmd_if_needed(next, nstart);
2325         }
2326 }
2327
2328 static void unmap_page(struct page *page)
2329 {
2330         struct folio *folio = page_folio(page);
2331         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2332                 TTU_SYNC;
2333
2334         VM_BUG_ON_PAGE(!PageHead(page), page);
2335
2336         /*
2337          * Anon pages need migration entries to preserve them, but file
2338          * pages can simply be left unmapped, then faulted back on demand.
2339          * If that is ever changed (perhaps for mlock), update remap_page().
2340          */
2341         if (folio_test_anon(folio))
2342                 try_to_migrate(folio, ttu_flags);
2343         else
2344                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2345 }
2346
2347 static void remap_page(struct folio *folio, unsigned long nr)
2348 {
2349         int i = 0;
2350
2351         /* If unmap_page() uses try_to_migrate() on file, remove this check */
2352         if (!folio_test_anon(folio))
2353                 return;
2354         for (;;) {
2355                 remove_migration_ptes(folio, folio, true);
2356                 i += folio_nr_pages(folio);
2357                 if (i >= nr)
2358                         break;
2359                 folio = folio_next(folio);
2360         }
2361 }
2362
2363 static void lru_add_page_tail(struct page *head, struct page *tail,
2364                 struct lruvec *lruvec, struct list_head *list)
2365 {
2366         VM_BUG_ON_PAGE(!PageHead(head), head);
2367         VM_BUG_ON_PAGE(PageCompound(tail), head);
2368         VM_BUG_ON_PAGE(PageLRU(tail), head);
2369         lockdep_assert_held(&lruvec->lru_lock);
2370
2371         if (list) {
2372                 /* page reclaim is reclaiming a huge page */
2373                 VM_WARN_ON(PageLRU(head));
2374                 get_page(tail);
2375                 list_add_tail(&tail->lru, list);
2376         } else {
2377                 /* head is still on lru (and we have it frozen) */
2378                 VM_WARN_ON(!PageLRU(head));
2379                 if (PageUnevictable(tail))
2380                         tail->mlock_count = 0;
2381                 else
2382                         list_add_tail(&tail->lru, &head->lru);
2383                 SetPageLRU(tail);
2384         }
2385 }
2386
2387 static void __split_huge_page_tail(struct page *head, int tail,
2388                 struct lruvec *lruvec, struct list_head *list)
2389 {
2390         struct page *page_tail = head + tail;
2391
2392         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2393
2394         /*
2395          * Clone page flags before unfreezing refcount.
2396          *
2397          * After successful get_page_unless_zero() might follow flags change,
2398          * for example lock_page() which set PG_waiters.
2399          *
2400          * Note that for mapped sub-pages of an anonymous THP,
2401          * PG_anon_exclusive has been cleared in unmap_page() and is stored in
2402          * the migration entry instead from where remap_page() will restore it.
2403          * We can still have PG_anon_exclusive set on effectively unmapped and
2404          * unreferenced sub-pages of an anonymous THP: we can simply drop
2405          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2406          */
2407         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2408         page_tail->flags |= (head->flags &
2409                         ((1L << PG_referenced) |
2410                          (1L << PG_swapbacked) |
2411                          (1L << PG_swapcache) |
2412                          (1L << PG_mlocked) |
2413                          (1L << PG_uptodate) |
2414                          (1L << PG_active) |
2415                          (1L << PG_workingset) |
2416                          (1L << PG_locked) |
2417                          (1L << PG_unevictable) |
2418 #ifdef CONFIG_64BIT
2419                          (1L << PG_arch_2) |
2420 #endif
2421                          (1L << PG_dirty)));
2422
2423         /* ->mapping in first tail page is compound_mapcount */
2424         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2425                         page_tail);
2426         page_tail->mapping = head->mapping;
2427         page_tail->index = head->index + tail;
2428         page_tail->private = 0;
2429
2430         /* Page flags must be visible before we make the page non-compound. */
2431         smp_wmb();
2432
2433         /*
2434          * Clear PageTail before unfreezing page refcount.
2435          *
2436          * After successful get_page_unless_zero() might follow put_page()
2437          * which needs correct compound_head().
2438          */
2439         clear_compound_head(page_tail);
2440
2441         /* Finally unfreeze refcount. Additional reference from page cache. */
2442         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2443                                           PageSwapCache(head)));
2444
2445         if (page_is_young(head))
2446                 set_page_young(page_tail);
2447         if (page_is_idle(head))
2448                 set_page_idle(page_tail);
2449
2450         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2451
2452         /*
2453          * always add to the tail because some iterators expect new
2454          * pages to show after the currently processed elements - e.g.
2455          * migrate_pages
2456          */
2457         lru_add_page_tail(head, page_tail, lruvec, list);
2458 }
2459
2460 static void __split_huge_page(struct page *page, struct list_head *list,
2461                 pgoff_t end)
2462 {
2463         struct folio *folio = page_folio(page);
2464         struct page *head = &folio->page;
2465         struct lruvec *lruvec;
2466         struct address_space *swap_cache = NULL;
2467         unsigned long offset = 0;
2468         unsigned int nr = thp_nr_pages(head);
2469         int i;
2470
2471         /* complete memcg works before add pages to LRU */
2472         split_page_memcg(head, nr);
2473
2474         if (PageAnon(head) && PageSwapCache(head)) {
2475                 swp_entry_t entry = { .val = page_private(head) };
2476
2477                 offset = swp_offset(entry);
2478                 swap_cache = swap_address_space(entry);
2479                 xa_lock(&swap_cache->i_pages);
2480         }
2481
2482         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2483         lruvec = folio_lruvec_lock(folio);
2484
2485         ClearPageHasHWPoisoned(head);
2486
2487         for (i = nr - 1; i >= 1; i--) {
2488                 __split_huge_page_tail(head, i, lruvec, list);
2489                 /* Some pages can be beyond EOF: drop them from page cache */
2490                 if (head[i].index >= end) {
2491                         ClearPageDirty(head + i);
2492                         __delete_from_page_cache(head + i, NULL);
2493                         if (shmem_mapping(head->mapping))
2494                                 shmem_uncharge(head->mapping->host, 1);
2495                         put_page(head + i);
2496                 } else if (!PageAnon(page)) {
2497                         __xa_store(&head->mapping->i_pages, head[i].index,
2498                                         head + i, 0);
2499                 } else if (swap_cache) {
2500                         __xa_store(&swap_cache->i_pages, offset + i,
2501                                         head + i, 0);
2502                 }
2503         }
2504
2505         ClearPageCompound(head);
2506         unlock_page_lruvec(lruvec);
2507         /* Caller disabled irqs, so they are still disabled here */
2508
2509         split_page_owner(head, nr);
2510
2511         /* See comment in __split_huge_page_tail() */
2512         if (PageAnon(head)) {
2513                 /* Additional pin to swap cache */
2514                 if (PageSwapCache(head)) {
2515                         page_ref_add(head, 2);
2516                         xa_unlock(&swap_cache->i_pages);
2517                 } else {
2518                         page_ref_inc(head);
2519                 }
2520         } else {
2521                 /* Additional pin to page cache */
2522                 page_ref_add(head, 2);
2523                 xa_unlock(&head->mapping->i_pages);
2524         }
2525         local_irq_enable();
2526
2527         remap_page(folio, nr);
2528
2529         if (PageSwapCache(head)) {
2530                 swp_entry_t entry = { .val = page_private(head) };
2531
2532                 split_swap_cluster(entry);
2533         }
2534
2535         for (i = 0; i < nr; i++) {
2536                 struct page *subpage = head + i;
2537                 if (subpage == page)
2538                         continue;
2539                 unlock_page(subpage);
2540
2541                 /*
2542                  * Subpages may be freed if there wasn't any mapping
2543                  * like if add_to_swap() is running on a lru page that
2544                  * had its mapping zapped. And freeing these pages
2545                  * requires taking the lru_lock so we do the put_page
2546                  * of the tail pages after the split is complete.
2547                  */
2548                 put_page(subpage);
2549         }
2550 }
2551
2552 /* Racy check whether the huge page can be split */
2553 bool can_split_folio(struct folio *folio, int *pextra_pins)
2554 {
2555         int extra_pins;
2556
2557         /* Additional pins from page cache */
2558         if (folio_test_anon(folio))
2559                 extra_pins = folio_test_swapcache(folio) ?
2560                                 folio_nr_pages(folio) : 0;
2561         else
2562                 extra_pins = folio_nr_pages(folio);
2563         if (pextra_pins)
2564                 *pextra_pins = extra_pins;
2565         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2566 }
2567
2568 /*
2569  * This function splits huge page into normal pages. @page can point to any
2570  * subpage of huge page to split. Split doesn't change the position of @page.
2571  *
2572  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2573  * The huge page must be locked.
2574  *
2575  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2576  *
2577  * Both head page and tail pages will inherit mapping, flags, and so on from
2578  * the hugepage.
2579  *
2580  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2581  * they are not mapped.
2582  *
2583  * Returns 0 if the hugepage is split successfully.
2584  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2585  * us.
2586  */
2587 int split_huge_page_to_list(struct page *page, struct list_head *list)
2588 {
2589         struct folio *folio = page_folio(page);
2590         struct page *head = &folio->page;
2591         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2592         XA_STATE(xas, &head->mapping->i_pages, head->index);
2593         struct anon_vma *anon_vma = NULL;
2594         struct address_space *mapping = NULL;
2595         int extra_pins, ret;
2596         pgoff_t end;
2597         bool is_hzp;
2598
2599         VM_BUG_ON_PAGE(!PageLocked(head), head);
2600         VM_BUG_ON_PAGE(!PageCompound(head), head);
2601
2602         is_hzp = is_huge_zero_page(head);
2603         VM_WARN_ON_ONCE_PAGE(is_hzp, head);
2604         if (is_hzp)
2605                 return -EBUSY;
2606
2607         if (PageWriteback(head))
2608                 return -EBUSY;
2609
2610         if (PageAnon(head)) {
2611                 /*
2612                  * The caller does not necessarily hold an mmap_lock that would
2613                  * prevent the anon_vma disappearing so we first we take a
2614                  * reference to it and then lock the anon_vma for write. This
2615                  * is similar to folio_lock_anon_vma_read except the write lock
2616                  * is taken to serialise against parallel split or collapse
2617                  * operations.
2618                  */
2619                 anon_vma = page_get_anon_vma(head);
2620                 if (!anon_vma) {
2621                         ret = -EBUSY;
2622                         goto out;
2623                 }
2624                 end = -1;
2625                 mapping = NULL;
2626                 anon_vma_lock_write(anon_vma);
2627         } else {
2628                 mapping = head->mapping;
2629
2630                 /* Truncated ? */
2631                 if (!mapping) {
2632                         ret = -EBUSY;
2633                         goto out;
2634                 }
2635
2636                 xas_split_alloc(&xas, head, compound_order(head),
2637                                 mapping_gfp_mask(mapping) & GFP_RECLAIM_MASK);
2638                 if (xas_error(&xas)) {
2639                         ret = xas_error(&xas);
2640                         goto out;
2641                 }
2642
2643                 anon_vma = NULL;
2644                 i_mmap_lock_read(mapping);
2645
2646                 /*
2647                  *__split_huge_page() may need to trim off pages beyond EOF:
2648                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2649                  * which cannot be nested inside the page tree lock. So note
2650                  * end now: i_size itself may be changed at any moment, but
2651                  * head page lock is good enough to serialize the trimming.
2652                  */
2653                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2654                 if (shmem_mapping(mapping))
2655                         end = shmem_fallocend(mapping->host, end);
2656         }
2657
2658         /*
2659          * Racy check if we can split the page, before unmap_page() will
2660          * split PMDs
2661          */
2662         if (!can_split_folio(folio, &extra_pins)) {
2663                 ret = -EBUSY;
2664                 goto out_unlock;
2665         }
2666
2667         unmap_page(head);
2668
2669         /* block interrupt reentry in xa_lock and spinlock */
2670         local_irq_disable();
2671         if (mapping) {
2672                 /*
2673                  * Check if the head page is present in page cache.
2674                  * We assume all tail are present too, if head is there.
2675                  */
2676                 xas_lock(&xas);
2677                 xas_reset(&xas);
2678                 if (xas_load(&xas) != head)
2679                         goto fail;
2680         }
2681
2682         /* Prevent deferred_split_scan() touching ->_refcount */
2683         spin_lock(&ds_queue->split_queue_lock);
2684         if (page_ref_freeze(head, 1 + extra_pins)) {
2685                 if (!list_empty(page_deferred_list(head))) {
2686                         ds_queue->split_queue_len--;
2687                         list_del(page_deferred_list(head));
2688                 }
2689                 spin_unlock(&ds_queue->split_queue_lock);
2690                 if (mapping) {
2691                         int nr = thp_nr_pages(head);
2692
2693                         xas_split(&xas, head, thp_order(head));
2694                         if (PageSwapBacked(head)) {
2695                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2696                                                         -nr);
2697                         } else {
2698                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2699                                                         -nr);
2700                                 filemap_nr_thps_dec(mapping);
2701                         }
2702                 }
2703
2704                 __split_huge_page(page, list, end);
2705                 ret = 0;
2706         } else {
2707                 spin_unlock(&ds_queue->split_queue_lock);
2708 fail:
2709                 if (mapping)
2710                         xas_unlock(&xas);
2711                 local_irq_enable();
2712                 remap_page(folio, folio_nr_pages(folio));
2713                 ret = -EBUSY;
2714         }
2715
2716 out_unlock:
2717         if (anon_vma) {
2718                 anon_vma_unlock_write(anon_vma);
2719                 put_anon_vma(anon_vma);
2720         }
2721         if (mapping)
2722                 i_mmap_unlock_read(mapping);
2723 out:
2724         xas_destroy(&xas);
2725         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2726         return ret;
2727 }
2728
2729 void free_transhuge_page(struct page *page)
2730 {
2731         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2732         unsigned long flags;
2733
2734         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2735         if (!list_empty(page_deferred_list(page))) {
2736                 ds_queue->split_queue_len--;
2737                 list_del(page_deferred_list(page));
2738         }
2739         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2740         free_compound_page(page);
2741 }
2742
2743 void deferred_split_huge_page(struct page *page)
2744 {
2745         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2746 #ifdef CONFIG_MEMCG
2747         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2748 #endif
2749         unsigned long flags;
2750
2751         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2752
2753         /*
2754          * The try_to_unmap() in page reclaim path might reach here too,
2755          * this may cause a race condition to corrupt deferred split queue.
2756          * And, if page reclaim is already handling the same page, it is
2757          * unnecessary to handle it again in shrinker.
2758          *
2759          * Check PageSwapCache to determine if the page is being
2760          * handled by page reclaim since THP swap would add the page into
2761          * swap cache before calling try_to_unmap().
2762          */
2763         if (PageSwapCache(page))
2764                 return;
2765
2766         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2767         if (list_empty(page_deferred_list(page))) {
2768                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2769                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2770                 ds_queue->split_queue_len++;
2771 #ifdef CONFIG_MEMCG
2772                 if (memcg)
2773                         set_shrinker_bit(memcg, page_to_nid(page),
2774                                          deferred_split_shrinker.id);
2775 #endif
2776         }
2777         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2778 }
2779
2780 static unsigned long deferred_split_count(struct shrinker *shrink,
2781                 struct shrink_control *sc)
2782 {
2783         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2784         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2785
2786 #ifdef CONFIG_MEMCG
2787         if (sc->memcg)
2788                 ds_queue = &sc->memcg->deferred_split_queue;
2789 #endif
2790         return READ_ONCE(ds_queue->split_queue_len);
2791 }
2792
2793 static unsigned long deferred_split_scan(struct shrinker *shrink,
2794                 struct shrink_control *sc)
2795 {
2796         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2797         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2798         unsigned long flags;
2799         LIST_HEAD(list), *pos, *next;
2800         struct page *page;
2801         int split = 0;
2802
2803 #ifdef CONFIG_MEMCG
2804         if (sc->memcg)
2805                 ds_queue = &sc->memcg->deferred_split_queue;
2806 #endif
2807
2808         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2809         /* Take pin on all head pages to avoid freeing them under us */
2810         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2811                 page = list_entry((void *)pos, struct page, deferred_list);
2812                 page = compound_head(page);
2813                 if (get_page_unless_zero(page)) {
2814                         list_move(page_deferred_list(page), &list);
2815                 } else {
2816                         /* We lost race with put_compound_page() */
2817                         list_del_init(page_deferred_list(page));
2818                         ds_queue->split_queue_len--;
2819                 }
2820                 if (!--sc->nr_to_scan)
2821                         break;
2822         }
2823         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2824
2825         list_for_each_safe(pos, next, &list) {
2826                 page = list_entry((void *)pos, struct page, deferred_list);
2827                 if (!trylock_page(page))
2828                         goto next;
2829                 /* split_huge_page() removes page from list on success */
2830                 if (!split_huge_page(page))
2831                         split++;
2832                 unlock_page(page);
2833 next:
2834                 put_page(page);
2835         }
2836
2837         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2838         list_splice_tail(&list, &ds_queue->split_queue);
2839         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2840
2841         /*
2842          * Stop shrinker if we didn't split any page, but the queue is empty.
2843          * This can happen if pages were freed under us.
2844          */
2845         if (!split && list_empty(&ds_queue->split_queue))
2846                 return SHRINK_STOP;
2847         return split;
2848 }
2849
2850 static struct shrinker deferred_split_shrinker = {
2851         .count_objects = deferred_split_count,
2852         .scan_objects = deferred_split_scan,
2853         .seeks = DEFAULT_SEEKS,
2854         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2855                  SHRINKER_NONSLAB,
2856 };
2857
2858 #ifdef CONFIG_DEBUG_FS
2859 static void split_huge_pages_all(void)
2860 {
2861         struct zone *zone;
2862         struct page *page;
2863         unsigned long pfn, max_zone_pfn;
2864         unsigned long total = 0, split = 0;
2865
2866         pr_debug("Split all THPs\n");
2867         for_each_populated_zone(zone) {
2868                 max_zone_pfn = zone_end_pfn(zone);
2869                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2870                         if (!pfn_valid(pfn))
2871                                 continue;
2872
2873                         page = pfn_to_page(pfn);
2874                         if (!get_page_unless_zero(page))
2875                                 continue;
2876
2877                         if (zone != page_zone(page))
2878                                 goto next;
2879
2880                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2881                                 goto next;
2882
2883                         total++;
2884                         lock_page(page);
2885                         if (!split_huge_page(page))
2886                                 split++;
2887                         unlock_page(page);
2888 next:
2889                         put_page(page);
2890                         cond_resched();
2891                 }
2892         }
2893
2894         pr_debug("%lu of %lu THP split\n", split, total);
2895 }
2896
2897 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2898 {
2899         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2900                     is_vm_hugetlb_page(vma);
2901 }
2902
2903 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2904                                 unsigned long vaddr_end)
2905 {
2906         int ret = 0;
2907         struct task_struct *task;
2908         struct mm_struct *mm;
2909         unsigned long total = 0, split = 0;
2910         unsigned long addr;
2911
2912         vaddr_start &= PAGE_MASK;
2913         vaddr_end &= PAGE_MASK;
2914
2915         /* Find the task_struct from pid */
2916         rcu_read_lock();
2917         task = find_task_by_vpid(pid);
2918         if (!task) {
2919                 rcu_read_unlock();
2920                 ret = -ESRCH;
2921                 goto out;
2922         }
2923         get_task_struct(task);
2924         rcu_read_unlock();
2925
2926         /* Find the mm_struct */
2927         mm = get_task_mm(task);
2928         put_task_struct(task);
2929
2930         if (!mm) {
2931                 ret = -EINVAL;
2932                 goto out;
2933         }
2934
2935         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2936                  pid, vaddr_start, vaddr_end);
2937
2938         mmap_read_lock(mm);
2939         /*
2940          * always increase addr by PAGE_SIZE, since we could have a PTE page
2941          * table filled with PTE-mapped THPs, each of which is distinct.
2942          */
2943         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2944                 struct vm_area_struct *vma = find_vma(mm, addr);
2945                 struct page *page;
2946
2947                 if (!vma || addr < vma->vm_start)
2948                         break;
2949
2950                 /* skip special VMA and hugetlb VMA */
2951                 if (vma_not_suitable_for_thp_split(vma)) {
2952                         addr = vma->vm_end;
2953                         continue;
2954                 }
2955
2956                 /* FOLL_DUMP to ignore special (like zero) pages */
2957                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2958
2959                 if (IS_ERR(page))
2960                         continue;
2961                 if (!page || is_zone_device_page(page))
2962                         continue;
2963
2964                 if (!is_transparent_hugepage(page))
2965                         goto next;
2966
2967                 total++;
2968                 if (!can_split_folio(page_folio(page), NULL))
2969                         goto next;
2970
2971                 if (!trylock_page(page))
2972                         goto next;
2973
2974                 if (!split_huge_page(page))
2975                         split++;
2976
2977                 unlock_page(page);
2978 next:
2979                 put_page(page);
2980                 cond_resched();
2981         }
2982         mmap_read_unlock(mm);
2983         mmput(mm);
2984
2985         pr_debug("%lu of %lu THP split\n", split, total);
2986
2987 out:
2988         return ret;
2989 }
2990
2991 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
2992                                 pgoff_t off_end)
2993 {
2994         struct filename *file;
2995         struct file *candidate;
2996         struct address_space *mapping;
2997         int ret = -EINVAL;
2998         pgoff_t index;
2999         int nr_pages = 1;
3000         unsigned long total = 0, split = 0;
3001
3002         file = getname_kernel(file_path);
3003         if (IS_ERR(file))
3004                 return ret;
3005
3006         candidate = file_open_name(file, O_RDONLY, 0);
3007         if (IS_ERR(candidate))
3008                 goto out;
3009
3010         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3011                  file_path, off_start, off_end);
3012
3013         mapping = candidate->f_mapping;
3014
3015         for (index = off_start; index < off_end; index += nr_pages) {
3016                 struct page *fpage = pagecache_get_page(mapping, index,
3017                                                 FGP_ENTRY | FGP_HEAD, 0);
3018
3019                 nr_pages = 1;
3020                 if (xa_is_value(fpage) || !fpage)
3021                         continue;
3022
3023                 if (!is_transparent_hugepage(fpage))
3024                         goto next;
3025
3026                 total++;
3027                 nr_pages = thp_nr_pages(fpage);
3028
3029                 if (!trylock_page(fpage))
3030                         goto next;
3031
3032                 if (!split_huge_page(fpage))
3033                         split++;
3034
3035                 unlock_page(fpage);
3036 next:
3037                 put_page(fpage);
3038                 cond_resched();
3039         }
3040
3041         filp_close(candidate, NULL);
3042         ret = 0;
3043
3044         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3045 out:
3046         putname(file);
3047         return ret;
3048 }
3049
3050 #define MAX_INPUT_BUF_SZ 255
3051
3052 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3053                                 size_t count, loff_t *ppops)
3054 {
3055         static DEFINE_MUTEX(split_debug_mutex);
3056         ssize_t ret;
3057         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3058         char input_buf[MAX_INPUT_BUF_SZ];
3059         int pid;
3060         unsigned long vaddr_start, vaddr_end;
3061
3062         ret = mutex_lock_interruptible(&split_debug_mutex);
3063         if (ret)
3064                 return ret;
3065
3066         ret = -EFAULT;
3067
3068         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3069         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3070                 goto out;
3071
3072         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3073
3074         if (input_buf[0] == '/') {
3075                 char *tok;
3076                 char *buf = input_buf;
3077                 char file_path[MAX_INPUT_BUF_SZ];
3078                 pgoff_t off_start = 0, off_end = 0;
3079                 size_t input_len = strlen(input_buf);
3080
3081                 tok = strsep(&buf, ",");
3082                 if (tok) {
3083                         strcpy(file_path, tok);
3084                 } else {
3085                         ret = -EINVAL;
3086                         goto out;
3087                 }
3088
3089                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3090                 if (ret != 2) {
3091                         ret = -EINVAL;
3092                         goto out;
3093                 }
3094                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3095                 if (!ret)
3096                         ret = input_len;
3097
3098                 goto out;
3099         }
3100
3101         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3102         if (ret == 1 && pid == 1) {
3103                 split_huge_pages_all();
3104                 ret = strlen(input_buf);
3105                 goto out;
3106         } else if (ret != 3) {
3107                 ret = -EINVAL;
3108                 goto out;
3109         }
3110
3111         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3112         if (!ret)
3113                 ret = strlen(input_buf);
3114 out:
3115         mutex_unlock(&split_debug_mutex);
3116         return ret;
3117
3118 }
3119
3120 static const struct file_operations split_huge_pages_fops = {
3121         .owner   = THIS_MODULE,
3122         .write   = split_huge_pages_write,
3123         .llseek  = no_llseek,
3124 };
3125
3126 static int __init split_huge_pages_debugfs(void)
3127 {
3128         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3129                             &split_huge_pages_fops);
3130         return 0;
3131 }
3132 late_initcall(split_huge_pages_debugfs);
3133 #endif
3134
3135 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3136 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3137                 struct page *page)
3138 {
3139         struct vm_area_struct *vma = pvmw->vma;
3140         struct mm_struct *mm = vma->vm_mm;
3141         unsigned long address = pvmw->address;
3142         bool anon_exclusive;
3143         pmd_t pmdval;
3144         swp_entry_t entry;
3145         pmd_t pmdswp;
3146
3147         if (!(pvmw->pmd && !pvmw->pte))
3148                 return 0;
3149
3150         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3151         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3152
3153         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3154         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3155                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3156                 return -EBUSY;
3157         }
3158
3159         if (pmd_dirty(pmdval))
3160                 set_page_dirty(page);
3161         if (pmd_write(pmdval))
3162                 entry = make_writable_migration_entry(page_to_pfn(page));
3163         else if (anon_exclusive)
3164                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3165         else
3166                 entry = make_readable_migration_entry(page_to_pfn(page));
3167         pmdswp = swp_entry_to_pmd(entry);
3168         if (pmd_soft_dirty(pmdval))
3169                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3170         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3171         page_remove_rmap(page, vma, true);
3172         put_page(page);
3173         trace_set_migration_pmd(address, pmd_val(pmdswp));
3174
3175         return 0;
3176 }
3177
3178 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3179 {
3180         struct vm_area_struct *vma = pvmw->vma;
3181         struct mm_struct *mm = vma->vm_mm;
3182         unsigned long address = pvmw->address;
3183         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3184         pmd_t pmde;
3185         swp_entry_t entry;
3186
3187         if (!(pvmw->pmd && !pvmw->pte))
3188                 return;
3189
3190         entry = pmd_to_swp_entry(*pvmw->pmd);
3191         get_page(new);
3192         pmde = pmd_mkold(mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)));
3193         if (pmd_swp_soft_dirty(*pvmw->pmd))
3194                 pmde = pmd_mksoft_dirty(pmde);
3195         if (is_writable_migration_entry(entry))
3196                 pmde = maybe_pmd_mkwrite(pmde, vma);
3197         if (pmd_swp_uffd_wp(*pvmw->pmd))
3198                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3199
3200         if (PageAnon(new)) {
3201                 rmap_t rmap_flags = RMAP_COMPOUND;
3202
3203                 if (!is_readable_migration_entry(entry))
3204                         rmap_flags |= RMAP_EXCLUSIVE;
3205
3206                 page_add_anon_rmap(new, vma, mmun_start, rmap_flags);
3207         } else {
3208                 page_add_file_rmap(new, vma, true);
3209         }
3210         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3211         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3212
3213         /* No need to invalidate - it was non-present before */
3214         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3215         trace_remove_migration_pmd(address, pmd_val(pmde));
3216 }
3217 #endif