mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/sched/coredump.h>
  12 #include <linux/sched/numa_balancing.h>
  13 #include <linux/highmem.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/rmap.h>
  17 #include <linux/swap.h>
  18 #include <linux/shrinker.h>
  19 #include <linux/mm_inline.h>
  20 #include <linux/swapops.h>
  21 #include <linux/backing-dev.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37 #include <linux/page_owner.h>
  38 #include <linux/sched/sysctl.h>
  39 #include <linux/memory-tiers.h>
  40
  41 #include <asm/tlb.h>
  42 #include <asm/pgalloc.h>
  43 #include "internal.h"
  44 #include "swap.h"
  45
  46 #define CREATE_TRACE_POINTS
  47 #include <trace/events/thp.h>
  48
  49 /*
  50  * By default, transparent hugepage support is disabled in order to avoid
  51  * risking an increased memory footprint for applications that are not
  52  * guaranteed to benefit from it. When transparent hugepage support is
  53  * enabled, it is for all mappings, and khugepaged scans all mappings.
  54  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  55  * for all hugepage allocations.
  56  */
  57 unsigned long transparent_hugepage_flags __read_mostly =
  58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  59         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  60 #endif
  61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  62         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  63 #endif
  64         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  65         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  66         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  67
  68 static struct shrinker *deferred_split_shrinker;
  69 static unsigned long deferred_split_count(struct shrinker *shrink,
  70                                           struct shrink_control *sc);
  71 static unsigned long deferred_split_scan(struct shrinker *shrink,
  72                                          struct shrink_control *sc);
  73
  74 static atomic_t huge_zero_refcount;
  75 struct page *huge_zero_page __read_mostly;
  76 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  77
  78 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
  79                         bool smaps, bool in_pf, bool enforce_sysfs)
  80 {
  81         if (!vma->vm_mm)                /* vdso */
  82                 return false;
  83
  84         /*
  85          * Explicitly disabled through madvise or prctl, or some
  86          * architectures may disable THP for some mappings, for
  87          * example, s390 kvm.
  88          * */
  89         if ((vm_flags & VM_NOHUGEPAGE) ||
  90             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  91                 return false;
  92         /*
  93          * If the hardware/firmware marked hugepage support disabled.
  94          */
  95         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED))
  96                 return false;
  97
  98         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  99         if (vma_is_dax(vma))
 100                 return in_pf;
 101
 102         /*
 103          * khugepaged special VMA and hugetlb VMA.
 104          * Must be checked after dax since some dax mappings may have
 105          * VM_MIXEDMAP set.
 106          */
 107         if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
 108                 return false;
 109
 110         /*
 111          * Check alignment for file vma and size for both file and anon vma.
 112          *
 113          * Skip the check for page fault. Huge fault does the check in fault
 114          * handlers. And this check is not suitable for huge PUD fault.
 115          */
 116         if (!in_pf &&
 117             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 118                 return false;
 119
 120         /*
 121          * Enabled via shmem mount options or sysfs settings.
 122          * Must be done before hugepage flags check since shmem has its
 123          * own flags.
 124          */
 125         if (!in_pf && shmem_file(vma->vm_file))
 126                 return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
 127                                      !enforce_sysfs, vma->vm_mm, vm_flags);
 128
 129         /* Enforce sysfs THP requirements as necessary */
 130         if (enforce_sysfs &&
 131             (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
 132                                            !hugepage_flags_always())))
 133                 return false;
 134
 135         if (!vma_is_anonymous(vma)) {
 136                 /*
 137                  * Trust that ->huge_fault() handlers know what they are doing
 138                  * in fault path.
 139                  */
 140                 if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
 141                         return true;
 142                 /* Only regular file is valid in collapse path */
 143                 if (((!in_pf || smaps)) && file_thp_enabled(vma))
 144                         return true;
 145                 return false;
 146         }
 147
 148         if (vma_is_temporary_stack(vma))
 149                 return false;
 150
 151         /*
 152          * THPeligible bit of smaps should show 1 for proper VMAs even
 153          * though anon_vma is not initialized yet.
 154          *
 155          * Allow page fault since anon_vma may be not initialized until
 156          * the first page fault.
 157          */
 158         if (!vma->anon_vma)
 159                 return (smaps || in_pf);
 160
 161         return true;
 162 }
 163
 164 static bool get_huge_zero_page(void)
 165 {
 166         struct page *zero_page;
 167 retry:
 168         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 169                 return true;
 170
 171         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 172                         HPAGE_PMD_ORDER);
 173         if (!zero_page) {
 174                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 175                 return false;
 176         }
 177         preempt_disable();
 178         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 179                 preempt_enable();
 180                 __free_pages(zero_page, compound_order(zero_page));
 181                 goto retry;
 182         }
 183         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 184
 185         /* We take additional reference here. It will be put back by shrinker */
 186         atomic_set(&huge_zero_refcount, 2);
 187         preempt_enable();
 188         count_vm_event(THP_ZERO_PAGE_ALLOC);
 189         return true;
 190 }
 191
 192 static void put_huge_zero_page(void)
 193 {
 194         /*
 195          * Counter should never go to zero here. Only shrinker can put
 196          * last reference.
 197          */
 198         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 199 }
 200
 201 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 202 {
 203         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 204                 return READ_ONCE(huge_zero_page);
 205
 206         if (!get_huge_zero_page())
 207                 return NULL;
 208
 209         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 210                 put_huge_zero_page();
 211
 212         return READ_ONCE(huge_zero_page);
 213 }
 214
 215 void mm_put_huge_zero_page(struct mm_struct *mm)
 216 {
 217         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 218                 put_huge_zero_page();
 219 }
 220
 221 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 222                                         struct shrink_control *sc)
 223 {
 224         /* we can free zero page only if last reference remains */
 225         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 226 }
 227
 228 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 229                                        struct shrink_control *sc)
 230 {
 231         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 232                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 233                 BUG_ON(zero_page == NULL);
 234                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 235                 __free_pages(zero_page, compound_order(zero_page));
 236                 return HPAGE_PMD_NR;
 237         }
 238
 239         return 0;
 240 }
 241
 242 static struct shrinker *huge_zero_page_shrinker;
 243
 244 #ifdef CONFIG_SYSFS
 245 static ssize_t enabled_show(struct kobject *kobj,
 246                             struct kobj_attribute *attr, char *buf)
 247 {
 248         const char *output;
 249
 250         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 251                 output = "[always] madvise never";
 252         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 253                           &transparent_hugepage_flags))
 254                 output = "always [madvise] never";
 255         else
 256                 output = "always madvise [never]";
 257
 258         return sysfs_emit(buf, "%s\n", output);
 259 }
 260
 261 static ssize_t enabled_store(struct kobject *kobj,
 262                              struct kobj_attribute *attr,
 263                              const char *buf, size_t count)
 264 {
 265         ssize_t ret = count;
 266
 267         if (sysfs_streq(buf, "always")) {
 268                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "madvise")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 272                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 273         } else if (sysfs_streq(buf, "never")) {
 274                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 275                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 276         } else
 277                 ret = -EINVAL;
 278
 279         if (ret > 0) {
 280                 int err = start_stop_khugepaged();
 281                 if (err)
 282                         ret = err;
 283         }
 284         return ret;
 285 }
 286
 287 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 288
 289 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 290                                   struct kobj_attribute *attr, char *buf,
 291                                   enum transparent_hugepage_flag flag)
 292 {
 293         return sysfs_emit(buf, "%d\n",
 294                           !!test_bit(flag, &transparent_hugepage_flags));
 295 }
 296
 297 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 298                                  struct kobj_attribute *attr,
 299                                  const char *buf, size_t count,
 300                                  enum transparent_hugepage_flag flag)
 301 {
 302         unsigned long value;
 303         int ret;
 304
 305         ret = kstrtoul(buf, 10, &value);
 306         if (ret < 0)
 307                 return ret;
 308         if (value > 1)
 309                 return -EINVAL;
 310
 311         if (value)
 312                 set_bit(flag, &transparent_hugepage_flags);
 313         else
 314                 clear_bit(flag, &transparent_hugepage_flags);
 315
 316         return count;
 317 }
 318
 319 static ssize_t defrag_show(struct kobject *kobj,
 320                            struct kobj_attribute *attr, char *buf)
 321 {
 322         const char *output;
 323
 324         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 325                      &transparent_hugepage_flags))
 326                 output = "[always] defer defer+madvise madvise never";
 327         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 328                           &transparent_hugepage_flags))
 329                 output = "always [defer] defer+madvise madvise never";
 330         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 331                           &transparent_hugepage_flags))
 332                 output = "always defer [defer+madvise] madvise never";
 333         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 334                           &transparent_hugepage_flags))
 335                 output = "always defer defer+madvise [madvise] never";
 336         else
 337                 output = "always defer defer+madvise madvise [never]";
 338
 339         return sysfs_emit(buf, "%s\n", output);
 340 }
 341
 342 static ssize_t defrag_store(struct kobject *kobj,
 343                             struct kobj_attribute *attr,
 344                             const char *buf, size_t count)
 345 {
 346         if (sysfs_streq(buf, "always")) {
 347                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 348                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 349                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 350                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 351         } else if (sysfs_streq(buf, "defer+madvise")) {
 352                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 353                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 354                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 355                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 356         } else if (sysfs_streq(buf, "defer")) {
 357                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 358                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 359                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 360                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 361         } else if (sysfs_streq(buf, "madvise")) {
 362                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 363                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 364                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 365                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 366         } else if (sysfs_streq(buf, "never")) {
 367                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 368                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 369                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 370                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 371         } else
 372                 return -EINVAL;
 373
 374         return count;
 375 }
 376 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 377
 378 static ssize_t use_zero_page_show(struct kobject *kobj,
 379                                   struct kobj_attribute *attr, char *buf)
 380 {
 381         return single_hugepage_flag_show(kobj, attr, buf,
 382                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static ssize_t use_zero_page_store(struct kobject *kobj,
 385                 struct kobj_attribute *attr, const char *buf, size_t count)
 386 {
 387         return single_hugepage_flag_store(kobj, attr, buf, count,
 388                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 389 }
 390 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 391
 392 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 393                                    struct kobj_attribute *attr, char *buf)
 394 {
 395         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 396 }
 397 static struct kobj_attribute hpage_pmd_size_attr =
 398         __ATTR_RO(hpage_pmd_size);
 399
 400 static struct attribute *hugepage_attr[] = {
 401         &enabled_attr.attr,
 402         &defrag_attr.attr,
 403         &use_zero_page_attr.attr,
 404         &hpage_pmd_size_attr.attr,
 405 #ifdef CONFIG_SHMEM
 406         &shmem_enabled_attr.attr,
 407 #endif
 408         NULL,
 409 };
 410
 411 static const struct attribute_group hugepage_attr_group = {
 412         .attrs = hugepage_attr,
 413 };
 414
 415 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 416 {
 417         int err;
 418
 419         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 420         if (unlikely(!*hugepage_kobj)) {
 421                 pr_err("failed to create transparent hugepage kobject\n");
 422                 return -ENOMEM;
 423         }
 424
 425         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 426         if (err) {
 427                 pr_err("failed to register transparent hugepage group\n");
 428                 goto delete_obj;
 429         }
 430
 431         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 432         if (err) {
 433                 pr_err("failed to register transparent hugepage group\n");
 434                 goto remove_hp_group;
 435         }
 436
 437         return 0;
 438
 439 remove_hp_group:
 440         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 441 delete_obj:
 442         kobject_put(*hugepage_kobj);
 443         return err;
 444 }
 445
 446 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 447 {
 448         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 449         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 450         kobject_put(hugepage_kobj);
 451 }
 452 #else
 453 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 454 {
 455         return 0;
 456 }
 457
 458 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 459 {
 460 }
 461 #endif /* CONFIG_SYSFS */
 462
 463 static int __init thp_shrinker_init(void)
 464 {
 465         huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");
 466         if (!huge_zero_page_shrinker)
 467                 return -ENOMEM;
 468
 469         deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
 470                                                  SHRINKER_MEMCG_AWARE |
 471                                                  SHRINKER_NONSLAB,
 472                                                  "thp-deferred_split");
 473         if (!deferred_split_shrinker) {
 474                 shrinker_free(huge_zero_page_shrinker);
 475                 return -ENOMEM;
 476         }
 477
 478         huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;
 479         huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;
 480         shrinker_register(huge_zero_page_shrinker);
 481
 482         deferred_split_shrinker->count_objects = deferred_split_count;
 483         deferred_split_shrinker->scan_objects = deferred_split_scan;
 484         shrinker_register(deferred_split_shrinker);
 485
 486         return 0;
 487 }
 488
 489 static void __init thp_shrinker_exit(void)
 490 {
 491         shrinker_free(huge_zero_page_shrinker);
 492         shrinker_free(deferred_split_shrinker);
 493 }
 494
 495 static int __init hugepage_init(void)
 496 {
 497         int err;
 498         struct kobject *hugepage_kobj;
 499
 500         if (!has_transparent_hugepage()) {
 501                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
 502                 return -EINVAL;
 503         }
 504
 505         /*
 506          * hugepages can't be allocated by the buddy allocator
 507          */
 508         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_ORDER);
 509         /*
 510          * we use page->mapping and page->index in second tail page
 511          * as list_head: assuming THP order >= 2
 512          */
 513         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 514
 515         err = hugepage_init_sysfs(&hugepage_kobj);
 516         if (err)
 517                 goto err_sysfs;
 518
 519         err = khugepaged_init();
 520         if (err)
 521                 goto err_slab;
 522
 523         err = thp_shrinker_init();
 524         if (err)
 525                 goto err_shrinker;
 526
 527         /*
 528          * By default disable transparent hugepages on smaller systems,
 529          * where the extra memory used could hurt more than TLB overhead
 530          * is likely to save.  The admin can still enable it through /sys.
 531          */
 532         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 533                 transparent_hugepage_flags = 0;
 534                 return 0;
 535         }
 536
 537         err = start_stop_khugepaged();
 538         if (err)
 539                 goto err_khugepaged;
 540
 541         return 0;
 542 err_khugepaged:
 543         thp_shrinker_exit();
 544 err_shrinker:
 545         khugepaged_destroy();
 546 err_slab:
 547         hugepage_exit_sysfs(hugepage_kobj);
 548 err_sysfs:
 549         return err;
 550 }
 551 subsys_initcall(hugepage_init);
 552
 553 static int __init setup_transparent_hugepage(char *str)
 554 {
 555         int ret = 0;
 556         if (!str)
 557                 goto out;
 558         if (!strcmp(str, "always")) {
 559                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 560                         &transparent_hugepage_flags);
 561                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 562                           &transparent_hugepage_flags);
 563                 ret = 1;
 564         } else if (!strcmp(str, "madvise")) {
 565                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 566                           &transparent_hugepage_flags);
 567                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 568                         &transparent_hugepage_flags);
 569                 ret = 1;
 570         } else if (!strcmp(str, "never")) {
 571                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 572                           &transparent_hugepage_flags);
 573                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 574                           &transparent_hugepage_flags);
 575                 ret = 1;
 576         }
 577 out:
 578         if (!ret)
 579                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 580         return ret;
 581 }
 582 __setup("transparent_hugepage=", setup_transparent_hugepage);
 583
 584 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 585 {
 586         if (likely(vma->vm_flags & VM_WRITE))
 587                 pmd = pmd_mkwrite(pmd, vma);
 588         return pmd;
 589 }
 590
 591 #ifdef CONFIG_MEMCG
 592 static inline
 593 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 594 {
 595         struct mem_cgroup *memcg = folio_memcg(folio);
 596         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 597
 598         if (memcg)
 599                 return &memcg->deferred_split_queue;
 600         else
 601                 return &pgdat->deferred_split_queue;
 602 }
 603 #else
 604 static inline
 605 struct deferred_split *get_deferred_split_queue(struct folio *folio)
 606 {
 607         struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
 608
 609         return &pgdat->deferred_split_queue;
 610 }
 611 #endif
 612
 613 void folio_prep_large_rmappable(struct folio *folio)
 614 {
 615         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
 616         INIT_LIST_HEAD(&folio->_deferred_list);
 617         folio_set_large_rmappable(folio);
 618 }
 619
 620 static inline bool is_transparent_hugepage(struct folio *folio)
 621 {
 622         if (!folio_test_large(folio))
 623                 return false;
 624
 625         return is_huge_zero_page(&folio->page) ||
 626                 folio_test_large_rmappable(folio);
 627 }
 628
 629 static unsigned long __thp_get_unmapped_area(struct file *filp,
 630                 unsigned long addr, unsigned long len,
 631                 loff_t off, unsigned long flags, unsigned long size)
 632 {
 633         loff_t off_end = off + len;
 634         loff_t off_align = round_up(off, size);
 635         unsigned long len_pad, ret;
 636
 637         if (off_end <= off_align || (off_end - off_align) < size)
 638                 return 0;
 639
 640         len_pad = len + size;
 641         if (len_pad < len || (off + len_pad) < off)
 642                 return 0;
 643
 644         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 645                                               off >> PAGE_SHIFT, flags);
 646
 647         /*
 648          * The failure might be due to length padding. The caller will retry
 649          * without the padding.
 650          */
 651         if (IS_ERR_VALUE(ret))
 652                 return 0;
 653
 654         /*
 655          * Do not try to align to THP boundary if allocation at the address
 656          * hint succeeds.
 657          */
 658         if (ret == addr)
 659                 return addr;
 660
 661         ret += (off - ret) & (size - 1);
 662         return ret;
 663 }
 664
 665 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 666                 unsigned long len, unsigned long pgoff, unsigned long flags)
 667 {
 668         unsigned long ret;
 669         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 670
 671         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 672         if (ret)
 673                 return ret;
 674
 675         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 676 }
 677 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 678
 679 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 680                         struct page *page, gfp_t gfp)
 681 {
 682         struct vm_area_struct *vma = vmf->vma;
 683         struct folio *folio = page_folio(page);
 684         pgtable_t pgtable;
 685         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 686         vm_fault_t ret = 0;
 687
 688         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
 689
 690         if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
 691                 folio_put(folio);
 692                 count_vm_event(THP_FAULT_FALLBACK);
 693                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 694                 return VM_FAULT_FALLBACK;
 695         }
 696         folio_throttle_swaprate(folio, gfp);
 697
 698         pgtable = pte_alloc_one(vma->vm_mm);
 699         if (unlikely(!pgtable)) {
 700                 ret = VM_FAULT_OOM;
 701                 goto release;
 702         }
 703
 704         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 705         /*
 706          * The memory barrier inside __folio_mark_uptodate makes sure that
 707          * clear_huge_page writes become visible before the set_pmd_at()
 708          * write.
 709          */
 710         __folio_mark_uptodate(folio);
 711
 712         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 713         if (unlikely(!pmd_none(*vmf->pmd))) {
 714                 goto unlock_release;
 715         } else {
 716                 pmd_t entry;
 717
 718                 ret = check_stable_address_space(vma->vm_mm);
 719                 if (ret)
 720                         goto unlock_release;
 721
 722                 /* Deliver the page fault to userland */
 723                 if (userfaultfd_missing(vma)) {
 724                         spin_unlock(vmf->ptl);
 725                         folio_put(folio);
 726                         pte_free(vma->vm_mm, pgtable);
 727                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 728                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 729                         return ret;
 730                 }
 731
 732                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 733                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 734                 folio_add_new_anon_rmap(folio, vma, haddr);
 735                 folio_add_lru_vma(folio, vma);
 736                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 737                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 738                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 739                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 740                 mm_inc_nr_ptes(vma->vm_mm);
 741                 spin_unlock(vmf->ptl);
 742                 count_vm_event(THP_FAULT_ALLOC);
 743                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 744         }
 745
 746         return 0;
 747 unlock_release:
 748         spin_unlock(vmf->ptl);
 749 release:
 750         if (pgtable)
 751                 pte_free(vma->vm_mm, pgtable);
 752         folio_put(folio);
 753         return ret;
 754
 755 }
 756
 757 /*
 758  * always: directly stall for all thp allocations
 759  * defer: wake kswapd and fail if not immediately available
 760  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 761  *                fail if not immediately available
 762  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 763  *          available
 764  * never: never stall for any thp allocation
 765  */
 766 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 767 {
 768         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 769
 770         /* Always do synchronous compaction */
 771         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 772                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 773
 774         /* Kick kcompactd and fail quickly */
 775         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 776                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 777
 778         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 779         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 780                 return GFP_TRANSHUGE_LIGHT |
 781                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 782                                         __GFP_KSWAPD_RECLAIM);
 783
 784         /* Only do synchronous compaction if madvised */
 785         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 786                 return GFP_TRANSHUGE_LIGHT |
 787                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 788
 789         return GFP_TRANSHUGE_LIGHT;
 790 }
 791
 792 /* Caller must hold page table lock. */
 793 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 794                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 795                 struct page *zero_page)
 796 {
 797         pmd_t entry;
 798         if (!pmd_none(*pmd))
 799                 return;
 800         entry = mk_pmd(zero_page, vma->vm_page_prot);
 801         entry = pmd_mkhuge(entry);
 802         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 803         set_pmd_at(mm, haddr, pmd, entry);
 804         mm_inc_nr_ptes(mm);
 805 }
 806
 807 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 808 {
 809         struct vm_area_struct *vma = vmf->vma;
 810         gfp_t gfp;
 811         struct folio *folio;
 812         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 813
 814         if (!transhuge_vma_suitable(vma, haddr))
 815                 return VM_FAULT_FALLBACK;
 816         if (unlikely(anon_vma_prepare(vma)))
 817                 return VM_FAULT_OOM;
 818         khugepaged_enter_vma(vma, vma->vm_flags);
 819
 820         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 821                         !mm_forbids_zeropage(vma->vm_mm) &&
 822                         transparent_hugepage_use_zero_page()) {
 823                 pgtable_t pgtable;
 824                 struct page *zero_page;
 825                 vm_fault_t ret;
 826                 pgtable = pte_alloc_one(vma->vm_mm);
 827                 if (unlikely(!pgtable))
 828                         return VM_FAULT_OOM;
 829                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 830                 if (unlikely(!zero_page)) {
 831                         pte_free(vma->vm_mm, pgtable);
 832                         count_vm_event(THP_FAULT_FALLBACK);
 833                         return VM_FAULT_FALLBACK;
 834                 }
 835                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 836                 ret = 0;
 837                 if (pmd_none(*vmf->pmd)) {
 838                         ret = check_stable_address_space(vma->vm_mm);
 839                         if (ret) {
 840                                 spin_unlock(vmf->ptl);
 841                                 pte_free(vma->vm_mm, pgtable);
 842                         } else if (userfaultfd_missing(vma)) {
 843                                 spin_unlock(vmf->ptl);
 844                                 pte_free(vma->vm_mm, pgtable);
 845                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 846                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 847                         } else {
 848                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 849                                                    haddr, vmf->pmd, zero_page);
 850                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 851                                 spin_unlock(vmf->ptl);
 852                         }
 853                 } else {
 854                         spin_unlock(vmf->ptl);
 855                         pte_free(vma->vm_mm, pgtable);
 856                 }
 857                 return ret;
 858         }
 859         gfp = vma_thp_gfp_mask(vma);
 860         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 861         if (unlikely(!folio)) {
 862                 count_vm_event(THP_FAULT_FALLBACK);
 863                 return VM_FAULT_FALLBACK;
 864         }
 865         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 866 }
 867
 868 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 869                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 870                 pgtable_t pgtable)
 871 {
 872         struct mm_struct *mm = vma->vm_mm;
 873         pmd_t entry;
 874         spinlock_t *ptl;
 875
 876         ptl = pmd_lock(mm, pmd);
 877         if (!pmd_none(*pmd)) {
 878                 if (write) {
 879                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 880                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 881                                 goto out_unlock;
 882                         }
 883                         entry = pmd_mkyoung(*pmd);
 884                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 885                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 886                                 update_mmu_cache_pmd(vma, addr, pmd);
 887                 }
 888
 889                 goto out_unlock;
 890         }
 891
 892         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 893         if (pfn_t_devmap(pfn))
 894                 entry = pmd_mkdevmap(entry);
 895         if (write) {
 896                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 897                 entry = maybe_pmd_mkwrite(entry, vma);
 898         }
 899
 900         if (pgtable) {
 901                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 902                 mm_inc_nr_ptes(mm);
 903                 pgtable = NULL;
 904         }
 905
 906         set_pmd_at(mm, addr, pmd, entry);
 907         update_mmu_cache_pmd(vma, addr, pmd);
 908
 909 out_unlock:
 910         spin_unlock(ptl);
 911         if (pgtable)
 912                 pte_free(mm, pgtable);
 913 }
 914
 915 /**
 916  * vmf_insert_pfn_pmd - insert a pmd size pfn
 917  * @vmf: Structure describing the fault
 918  * @pfn: pfn to insert
 919  * @write: whether it's a write fault
 920  *
 921  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
 922  *
 923  * Return: vm_fault_t value.
 924  */
 925 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 926 {
 927         unsigned long addr = vmf->address & PMD_MASK;
 928         struct vm_area_struct *vma = vmf->vma;
 929         pgprot_t pgprot = vma->vm_page_prot;
 930         pgtable_t pgtable = NULL;
 931
 932         /*
 933          * If we had pmd_special, we could avoid all these restrictions,
 934          * but we need to be consistent with PTEs and architectures that
 935          * can't support a 'special' bit.
 936          */
 937         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 938                         !pfn_t_devmap(pfn));
 939         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 940                                                 (VM_PFNMAP|VM_MIXEDMAP));
 941         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 942
 943         if (addr < vma->vm_start || addr >= vma->vm_end)
 944                 return VM_FAULT_SIGBUS;
 945
 946         if (arch_needs_pgtable_deposit()) {
 947                 pgtable = pte_alloc_one(vma->vm_mm);
 948                 if (!pgtable)
 949                         return VM_FAULT_OOM;
 950         }
 951
 952         track_pfn_insert(vma, &pgprot, pfn);
 953
 954         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 955         return VM_FAULT_NOPAGE;
 956 }
 957 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 958
 959 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 960 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 961 {
 962         if (likely(vma->vm_flags & VM_WRITE))
 963                 pud = pud_mkwrite(pud);
 964         return pud;
 965 }
 966
 967 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 968                 pud_t *pud, pfn_t pfn, bool write)
 969 {
 970         struct mm_struct *mm = vma->vm_mm;
 971         pgprot_t prot = vma->vm_page_prot;
 972         pud_t entry;
 973         spinlock_t *ptl;
 974
 975         ptl = pud_lock(mm, pud);
 976         if (!pud_none(*pud)) {
 977                 if (write) {
 978                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 979                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 980                                 goto out_unlock;
 981                         }
 982                         entry = pud_mkyoung(*pud);
 983                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 984                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 985                                 update_mmu_cache_pud(vma, addr, pud);
 986                 }
 987                 goto out_unlock;
 988         }
 989
 990         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 991         if (pfn_t_devmap(pfn))
 992                 entry = pud_mkdevmap(entry);
 993         if (write) {
 994                 entry = pud_mkyoung(pud_mkdirty(entry));
 995                 entry = maybe_pud_mkwrite(entry, vma);
 996         }
 997         set_pud_at(mm, addr, pud, entry);
 998         update_mmu_cache_pud(vma, addr, pud);
 999
1000 out_unlock:
1001         spin_unlock(ptl);
1002 }
1003
1004 /**
1005  * vmf_insert_pfn_pud - insert a pud size pfn
1006  * @vmf: Structure describing the fault
1007  * @pfn: pfn to insert
1008  * @write: whether it's a write fault
1009  *
1010  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1011  *
1012  * Return: vm_fault_t value.
1013  */
1014 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
1015 {
1016         unsigned long addr = vmf->address & PUD_MASK;
1017         struct vm_area_struct *vma = vmf->vma;
1018         pgprot_t pgprot = vma->vm_page_prot;
1019
1020         /*
1021          * If we had pud_special, we could avoid all these restrictions,
1022          * but we need to be consistent with PTEs and architectures that
1023          * can't support a 'special' bit.
1024          */
1025         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1026                         !pfn_t_devmap(pfn));
1027         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1028                                                 (VM_PFNMAP|VM_MIXEDMAP));
1029         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1030
1031         if (addr < vma->vm_start || addr >= vma->vm_end)
1032                 return VM_FAULT_SIGBUS;
1033
1034         track_pfn_insert(vma, &pgprot, pfn);
1035
1036         insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1037         return VM_FAULT_NOPAGE;
1038 }
1039 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1040 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1041
1042 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1043                       pmd_t *pmd, bool write)
1044 {
1045         pmd_t _pmd;
1046
1047         _pmd = pmd_mkyoung(*pmd);
1048         if (write)
1049                 _pmd = pmd_mkdirty(_pmd);
1050         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1051                                   pmd, _pmd, write))
1052                 update_mmu_cache_pmd(vma, addr, pmd);
1053 }
1054
1055 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1056                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1057 {
1058         unsigned long pfn = pmd_pfn(*pmd);
1059         struct mm_struct *mm = vma->vm_mm;
1060         struct page *page;
1061         int ret;
1062
1063         assert_spin_locked(pmd_lockptr(mm, pmd));
1064
1065         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1066                 return NULL;
1067
1068         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1069                 /* pass */;
1070         else
1071                 return NULL;
1072
1073         if (flags & FOLL_TOUCH)
1074                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1075
1076         /*
1077          * device mapped pages can only be returned if the
1078          * caller will manage the page reference count.
1079          */
1080         if (!(flags & (FOLL_GET | FOLL_PIN)))
1081                 return ERR_PTR(-EEXIST);
1082
1083         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1084         *pgmap = get_dev_pagemap(pfn, *pgmap);
1085         if (!*pgmap)
1086                 return ERR_PTR(-EFAULT);
1087         page = pfn_to_page(pfn);
1088         ret = try_grab_page(page, flags);
1089         if (ret)
1090                 page = ERR_PTR(ret);
1091
1092         return page;
1093 }
1094
1095 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1096                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1097                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1098 {
1099         spinlock_t *dst_ptl, *src_ptl;
1100         struct page *src_page;
1101         pmd_t pmd;
1102         pgtable_t pgtable = NULL;
1103         int ret = -ENOMEM;
1104
1105         /* Skip if can be re-fill on fault */
1106         if (!vma_is_anonymous(dst_vma))
1107                 return 0;
1108
1109         pgtable = pte_alloc_one(dst_mm);
1110         if (unlikely(!pgtable))
1111                 goto out;
1112
1113         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1114         src_ptl = pmd_lockptr(src_mm, src_pmd);
1115         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1116
1117         ret = -EAGAIN;
1118         pmd = *src_pmd;
1119
1120 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1121         if (unlikely(is_swap_pmd(pmd))) {
1122                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1123
1124                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1125                 if (!is_readable_migration_entry(entry)) {
1126                         entry = make_readable_migration_entry(
1127                                                         swp_offset(entry));
1128                         pmd = swp_entry_to_pmd(entry);
1129                         if (pmd_swp_soft_dirty(*src_pmd))
1130                                 pmd = pmd_swp_mksoft_dirty(pmd);
1131                         if (pmd_swp_uffd_wp(*src_pmd))
1132                                 pmd = pmd_swp_mkuffd_wp(pmd);
1133                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1134                 }
1135                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1136                 mm_inc_nr_ptes(dst_mm);
1137                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1138                 if (!userfaultfd_wp(dst_vma))
1139                         pmd = pmd_swp_clear_uffd_wp(pmd);
1140                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1141                 ret = 0;
1142                 goto out_unlock;
1143         }
1144 #endif
1145
1146         if (unlikely(!pmd_trans_huge(pmd))) {
1147                 pte_free(dst_mm, pgtable);
1148                 goto out_unlock;
1149         }
1150         /*
1151          * When page table lock is held, the huge zero pmd should not be
1152          * under splitting since we don't split the page itself, only pmd to
1153          * a page table.
1154          */
1155         if (is_huge_zero_pmd(pmd)) {
1156                 /*
1157                  * get_huge_zero_page() will never allocate a new page here,
1158                  * since we already have a zero page to copy. It just takes a
1159                  * reference.
1160                  */
1161                 mm_get_huge_zero_page(dst_mm);
1162                 goto out_zero_page;
1163         }
1164
1165         src_page = pmd_page(pmd);
1166         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1167
1168         get_page(src_page);
1169         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1170                 /* Page maybe pinned: split and retry the fault on PTEs. */
1171                 put_page(src_page);
1172                 pte_free(dst_mm, pgtable);
1173                 spin_unlock(src_ptl);
1174                 spin_unlock(dst_ptl);
1175                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1176                 return -EAGAIN;
1177         }
1178         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1179 out_zero_page:
1180         mm_inc_nr_ptes(dst_mm);
1181         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1182         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1183         if (!userfaultfd_wp(dst_vma))
1184                 pmd = pmd_clear_uffd_wp(pmd);
1185         pmd = pmd_mkold(pmd_wrprotect(pmd));
1186         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1187
1188         ret = 0;
1189 out_unlock:
1190         spin_unlock(src_ptl);
1191         spin_unlock(dst_ptl);
1192 out:
1193         return ret;
1194 }
1195
1196 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1197 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1198                       pud_t *pud, bool write)
1199 {
1200         pud_t _pud;
1201
1202         _pud = pud_mkyoung(*pud);
1203         if (write)
1204                 _pud = pud_mkdirty(_pud);
1205         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1206                                   pud, _pud, write))
1207                 update_mmu_cache_pud(vma, addr, pud);
1208 }
1209
1210 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1211                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1212 {
1213         unsigned long pfn = pud_pfn(*pud);
1214         struct mm_struct *mm = vma->vm_mm;
1215         struct page *page;
1216         int ret;
1217
1218         assert_spin_locked(pud_lockptr(mm, pud));
1219
1220         if (flags & FOLL_WRITE && !pud_write(*pud))
1221                 return NULL;
1222
1223         if (pud_present(*pud) && pud_devmap(*pud))
1224                 /* pass */;
1225         else
1226                 return NULL;
1227
1228         if (flags & FOLL_TOUCH)
1229                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1230
1231         /*
1232          * device mapped pages can only be returned if the
1233          * caller will manage the page reference count.
1234          *
1235          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1236          */
1237         if (!(flags & (FOLL_GET | FOLL_PIN)))
1238                 return ERR_PTR(-EEXIST);
1239
1240         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1241         *pgmap = get_dev_pagemap(pfn, *pgmap);
1242         if (!*pgmap)
1243                 return ERR_PTR(-EFAULT);
1244         page = pfn_to_page(pfn);
1245
1246         ret = try_grab_page(page, flags);
1247         if (ret)
1248                 page = ERR_PTR(ret);
1249
1250         return page;
1251 }
1252
1253 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1254                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1255                   struct vm_area_struct *vma)
1256 {
1257         spinlock_t *dst_ptl, *src_ptl;
1258         pud_t pud;
1259         int ret;
1260
1261         dst_ptl = pud_lock(dst_mm, dst_pud);
1262         src_ptl = pud_lockptr(src_mm, src_pud);
1263         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1264
1265         ret = -EAGAIN;
1266         pud = *src_pud;
1267         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1268                 goto out_unlock;
1269
1270         /*
1271          * When page table lock is held, the huge zero pud should not be
1272          * under splitting since we don't split the page itself, only pud to
1273          * a page table.
1274          */
1275         if (is_huge_zero_pud(pud)) {
1276                 /* No huge zero pud yet */
1277         }
1278
1279         /*
1280          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1281          * and split if duplicating fails.
1282          */
1283         pudp_set_wrprotect(src_mm, addr, src_pud);
1284         pud = pud_mkold(pud_wrprotect(pud));
1285         set_pud_at(dst_mm, addr, dst_pud, pud);
1286
1287         ret = 0;
1288 out_unlock:
1289         spin_unlock(src_ptl);
1290         spin_unlock(dst_ptl);
1291         return ret;
1292 }
1293
1294 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1295 {
1296         bool write = vmf->flags & FAULT_FLAG_WRITE;
1297
1298         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1299         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1300                 goto unlock;
1301
1302         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1303 unlock:
1304         spin_unlock(vmf->ptl);
1305 }
1306 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1307
1308 void huge_pmd_set_accessed(struct vm_fault *vmf)
1309 {
1310         bool write = vmf->flags & FAULT_FLAG_WRITE;
1311
1312         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1313         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1314                 goto unlock;
1315
1316         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1317
1318 unlock:
1319         spin_unlock(vmf->ptl);
1320 }
1321
1322 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1323 {
1324         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1325         struct vm_area_struct *vma = vmf->vma;
1326         struct folio *folio;
1327         struct page *page;
1328         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1329         pmd_t orig_pmd = vmf->orig_pmd;
1330
1331         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1332         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1333
1334         if (is_huge_zero_pmd(orig_pmd))
1335                 goto fallback;
1336
1337         spin_lock(vmf->ptl);
1338
1339         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1340                 spin_unlock(vmf->ptl);
1341                 return 0;
1342         }
1343
1344         page = pmd_page(orig_pmd);
1345         folio = page_folio(page);
1346         VM_BUG_ON_PAGE(!PageHead(page), page);
1347
1348         /* Early check when only holding the PT lock. */
1349         if (PageAnonExclusive(page))
1350                 goto reuse;
1351
1352         if (!folio_trylock(folio)) {
1353                 folio_get(folio);
1354                 spin_unlock(vmf->ptl);
1355                 folio_lock(folio);
1356                 spin_lock(vmf->ptl);
1357                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1358                         spin_unlock(vmf->ptl);
1359                         folio_unlock(folio);
1360                         folio_put(folio);
1361                         return 0;
1362                 }
1363                 folio_put(folio);
1364         }
1365
1366         /* Recheck after temporarily dropping the PT lock. */
1367         if (PageAnonExclusive(page)) {
1368                 folio_unlock(folio);
1369                 goto reuse;
1370         }
1371
1372         /*
1373          * See do_wp_page(): we can only reuse the folio exclusively if
1374          * there are no additional references. Note that we always drain
1375          * the LRU cache immediately after adding a THP.
1376          */
1377         if (folio_ref_count(folio) >
1378                         1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1379                 goto unlock_fallback;
1380         if (folio_test_swapcache(folio))
1381                 folio_free_swap(folio);
1382         if (folio_ref_count(folio) == 1) {
1383                 pmd_t entry;
1384
1385                 folio_move_anon_rmap(folio, vma);
1386                 SetPageAnonExclusive(page);
1387                 folio_unlock(folio);
1388 reuse:
1389                 if (unlikely(unshare)) {
1390                         spin_unlock(vmf->ptl);
1391                         return 0;
1392                 }
1393                 entry = pmd_mkyoung(orig_pmd);
1394                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1395                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1396                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1397                 spin_unlock(vmf->ptl);
1398                 return 0;
1399         }
1400
1401 unlock_fallback:
1402         folio_unlock(folio);
1403         spin_unlock(vmf->ptl);
1404 fallback:
1405         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1406         return VM_FAULT_FALLBACK;
1407 }
1408
1409 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1410                                            unsigned long addr, pmd_t pmd)
1411 {
1412         struct page *page;
1413
1414         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1415                 return false;
1416
1417         /* Don't touch entries that are not even readable (NUMA hinting). */
1418         if (pmd_protnone(pmd))
1419                 return false;
1420
1421         /* Do we need write faults for softdirty tracking? */
1422         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1423                 return false;
1424
1425         /* Do we need write faults for uffd-wp tracking? */
1426         if (userfaultfd_huge_pmd_wp(vma, pmd))
1427                 return false;
1428
1429         if (!(vma->vm_flags & VM_SHARED)) {
1430                 /* See can_change_pte_writable(). */
1431                 page = vm_normal_page_pmd(vma, addr, pmd);
1432                 return page && PageAnon(page) && PageAnonExclusive(page);
1433         }
1434
1435         /* See can_change_pte_writable(). */
1436         return pmd_dirty(pmd);
1437 }
1438
1439 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1440 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1441                                         struct vm_area_struct *vma,
1442                                         unsigned int flags)
1443 {
1444         /* If the pmd is writable, we can write to the page. */
1445         if (pmd_write(pmd))
1446                 return true;
1447
1448         /* Maybe FOLL_FORCE is set to override it? */
1449         if (!(flags & FOLL_FORCE))
1450                 return false;
1451
1452         /* But FOLL_FORCE has no effect on shared mappings */
1453         if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1454                 return false;
1455
1456         /* ... or read-only private ones */
1457         if (!(vma->vm_flags & VM_MAYWRITE))
1458                 return false;
1459
1460         /* ... or already writable ones that just need to take a write fault */
1461         if (vma->vm_flags & VM_WRITE)
1462                 return false;
1463
1464         /*
1465          * See can_change_pte_writable(): we broke COW and could map the page
1466          * writable if we have an exclusive anonymous page ...
1467          */
1468         if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1469                 return false;
1470
1471         /* ... and a write-fault isn't required for other reasons. */
1472         if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1473                 return false;
1474         return !userfaultfd_huge_pmd_wp(vma, pmd);
1475 }
1476
1477 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1478                                    unsigned long addr,
1479                                    pmd_t *pmd,
1480                                    unsigned int flags)
1481 {
1482         struct mm_struct *mm = vma->vm_mm;
1483         struct page *page;
1484         int ret;
1485
1486         assert_spin_locked(pmd_lockptr(mm, pmd));
1487
1488         page = pmd_page(*pmd);
1489         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1490
1491         if ((flags & FOLL_WRITE) &&
1492             !can_follow_write_pmd(*pmd, page, vma, flags))
1493                 return NULL;
1494
1495         /* Avoid dumping huge zero page */
1496         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1497                 return ERR_PTR(-EFAULT);
1498
1499         if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
1500                 return NULL;
1501
1502         if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1503                 return ERR_PTR(-EMLINK);
1504
1505         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1506                         !PageAnonExclusive(page), page);
1507
1508         ret = try_grab_page(page, flags);
1509         if (ret)
1510                 return ERR_PTR(ret);
1511
1512         if (flags & FOLL_TOUCH)
1513                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1514
1515         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1516         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1517
1518         return page;
1519 }
1520
1521 /* NUMA hinting page fault entry point for trans huge pmds */
1522 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1523 {
1524         struct vm_area_struct *vma = vmf->vma;
1525         pmd_t oldpmd = vmf->orig_pmd;
1526         pmd_t pmd;
1527         struct folio *folio;
1528         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1529         int nid = NUMA_NO_NODE;
1530         int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1531         bool migrated = false, writable = false;
1532         int flags = 0;
1533
1534         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1535         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1536                 spin_unlock(vmf->ptl);
1537                 goto out;
1538         }
1539
1540         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1541
1542         /*
1543          * Detect now whether the PMD could be writable; this information
1544          * is only valid while holding the PT lock.
1545          */
1546         writable = pmd_write(pmd);
1547         if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1548             can_change_pmd_writable(vma, vmf->address, pmd))
1549                 writable = true;
1550
1551         folio = vm_normal_folio_pmd(vma, haddr, pmd);
1552         if (!folio)
1553                 goto out_map;
1554
1555         /* See similar comment in do_numa_page for explanation */
1556         if (!writable)
1557                 flags |= TNF_NO_GROUP;
1558
1559         nid = folio_nid(folio);
1560         /*
1561          * For memory tiering mode, cpupid of slow memory page is used
1562          * to record page access time.  So use default value.
1563          */
1564         if (node_is_toptier(nid))
1565                 last_cpupid = folio_last_cpupid(folio);
1566         target_nid = numa_migrate_prep(folio, vma, haddr, nid, &flags);
1567         if (target_nid == NUMA_NO_NODE) {
1568                 folio_put(folio);
1569                 goto out_map;
1570         }
1571
1572         spin_unlock(vmf->ptl);
1573         writable = false;
1574
1575         migrated = migrate_misplaced_folio(folio, vma, target_nid);
1576         if (migrated) {
1577                 flags |= TNF_MIGRATED;
1578                 nid = target_nid;
1579         } else {
1580                 flags |= TNF_MIGRATE_FAIL;
1581                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1582                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1583                         spin_unlock(vmf->ptl);
1584                         goto out;
1585                 }
1586                 goto out_map;
1587         }
1588
1589 out:
1590         if (nid != NUMA_NO_NODE)
1591                 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
1592
1593         return 0;
1594
1595 out_map:
1596         /* Restore the PMD */
1597         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1598         pmd = pmd_mkyoung(pmd);
1599         if (writable)
1600                 pmd = pmd_mkwrite(pmd, vma);
1601         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1602         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1603         spin_unlock(vmf->ptl);
1604         goto out;
1605 }
1606
1607 /*
1608  * Return true if we do MADV_FREE successfully on entire pmd page.
1609  * Otherwise, return false.
1610  */
1611 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1612                 pmd_t *pmd, unsigned long addr, unsigned long next)
1613 {
1614         spinlock_t *ptl;
1615         pmd_t orig_pmd;
1616         struct folio *folio;
1617         struct mm_struct *mm = tlb->mm;
1618         bool ret = false;
1619
1620         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1621
1622         ptl = pmd_trans_huge_lock(pmd, vma);
1623         if (!ptl)
1624                 goto out_unlocked;
1625
1626         orig_pmd = *pmd;
1627         if (is_huge_zero_pmd(orig_pmd))
1628                 goto out;
1629
1630         if (unlikely(!pmd_present(orig_pmd))) {
1631                 VM_BUG_ON(thp_migration_supported() &&
1632                                   !is_pmd_migration_entry(orig_pmd));
1633                 goto out;
1634         }
1635
1636         folio = pfn_folio(pmd_pfn(orig_pmd));
1637         /*
1638          * If other processes are mapping this folio, we couldn't discard
1639          * the folio unless they all do MADV_FREE so let's skip the folio.
1640          */
1641         if (folio_estimated_sharers(folio) != 1)
1642                 goto out;
1643
1644         if (!folio_trylock(folio))
1645                 goto out;
1646
1647         /*
1648          * If user want to discard part-pages of THP, split it so MADV_FREE
1649          * will deactivate only them.
1650          */
1651         if (next - addr != HPAGE_PMD_SIZE) {
1652                 folio_get(folio);
1653                 spin_unlock(ptl);
1654                 split_folio(folio);
1655                 folio_unlock(folio);
1656                 folio_put(folio);
1657                 goto out_unlocked;
1658         }
1659
1660         if (folio_test_dirty(folio))
1661                 folio_clear_dirty(folio);
1662         folio_unlock(folio);
1663
1664         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1665                 pmdp_invalidate(vma, addr, pmd);
1666                 orig_pmd = pmd_mkold(orig_pmd);
1667                 orig_pmd = pmd_mkclean(orig_pmd);
1668
1669                 set_pmd_at(mm, addr, pmd, orig_pmd);
1670                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1671         }
1672
1673         folio_mark_lazyfree(folio);
1674         ret = true;
1675 out:
1676         spin_unlock(ptl);
1677 out_unlocked:
1678         return ret;
1679 }
1680
1681 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1682 {
1683         pgtable_t pgtable;
1684
1685         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1686         pte_free(mm, pgtable);
1687         mm_dec_nr_ptes(mm);
1688 }
1689
1690 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1691                  pmd_t *pmd, unsigned long addr)
1692 {
1693         pmd_t orig_pmd;
1694         spinlock_t *ptl;
1695
1696         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1697
1698         ptl = __pmd_trans_huge_lock(pmd, vma);
1699         if (!ptl)
1700                 return 0;
1701         /*
1702          * For architectures like ppc64 we look at deposited pgtable
1703          * when calling pmdp_huge_get_and_clear. So do the
1704          * pgtable_trans_huge_withdraw after finishing pmdp related
1705          * operations.
1706          */
1707         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1708                                                 tlb->fullmm);
1709         arch_check_zapped_pmd(vma, orig_pmd);
1710         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711         if (vma_is_special_huge(vma)) {
1712                 if (arch_needs_pgtable_deposit())
1713                         zap_deposited_table(tlb->mm, pmd);
1714                 spin_unlock(ptl);
1715         } else if (is_huge_zero_pmd(orig_pmd)) {
1716                 zap_deposited_table(tlb->mm, pmd);
1717                 spin_unlock(ptl);
1718         } else {
1719                 struct page *page = NULL;
1720                 int flush_needed = 1;
1721
1722                 if (pmd_present(orig_pmd)) {
1723                         page = pmd_page(orig_pmd);
1724                         page_remove_rmap(page, vma, true);
1725                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1726                         VM_BUG_ON_PAGE(!PageHead(page), page);
1727                 } else if (thp_migration_supported()) {
1728                         swp_entry_t entry;
1729
1730                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1731                         entry = pmd_to_swp_entry(orig_pmd);
1732                         page = pfn_swap_entry_to_page(entry);
1733                         flush_needed = 0;
1734                 } else
1735                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1736
1737                 if (PageAnon(page)) {
1738                         zap_deposited_table(tlb->mm, pmd);
1739                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1740                 } else {
1741                         if (arch_needs_pgtable_deposit())
1742                                 zap_deposited_table(tlb->mm, pmd);
1743                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1744                 }
1745
1746                 spin_unlock(ptl);
1747                 if (flush_needed)
1748                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1749         }
1750         return 1;
1751 }
1752
1753 #ifndef pmd_move_must_withdraw
1754 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1755                                          spinlock_t *old_pmd_ptl,
1756                                          struct vm_area_struct *vma)
1757 {
1758         /*
1759          * With split pmd lock we also need to move preallocated
1760          * PTE page table if new_pmd is on different PMD page table.
1761          *
1762          * We also don't deposit and withdraw tables for file pages.
1763          */
1764         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1765 }
1766 #endif
1767
1768 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1769 {
1770 #ifdef CONFIG_MEM_SOFT_DIRTY
1771         if (unlikely(is_pmd_migration_entry(pmd)))
1772                 pmd = pmd_swp_mksoft_dirty(pmd);
1773         else if (pmd_present(pmd))
1774                 pmd = pmd_mksoft_dirty(pmd);
1775 #endif
1776         return pmd;
1777 }
1778
1779 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1780                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1781 {
1782         spinlock_t *old_ptl, *new_ptl;
1783         pmd_t pmd;
1784         struct mm_struct *mm = vma->vm_mm;
1785         bool force_flush = false;
1786
1787         /*
1788          * The destination pmd shouldn't be established, free_pgtables()
1789          * should have released it; but move_page_tables() might have already
1790          * inserted a page table, if racing against shmem/file collapse.
1791          */
1792         if (!pmd_none(*new_pmd)) {
1793                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1794                 return false;
1795         }
1796
1797         /*
1798          * We don't have to worry about the ordering of src and dst
1799          * ptlocks because exclusive mmap_lock prevents deadlock.
1800          */
1801         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1802         if (old_ptl) {
1803                 new_ptl = pmd_lockptr(mm, new_pmd);
1804                 if (new_ptl != old_ptl)
1805                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1806                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1807                 if (pmd_present(pmd))
1808                         force_flush = true;
1809                 VM_BUG_ON(!pmd_none(*new_pmd));
1810
1811                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1812                         pgtable_t pgtable;
1813                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1814                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1815                 }
1816                 pmd = move_soft_dirty_pmd(pmd);
1817                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1818                 if (force_flush)
1819                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1820                 if (new_ptl != old_ptl)
1821                         spin_unlock(new_ptl);
1822                 spin_unlock(old_ptl);
1823                 return true;
1824         }
1825         return false;
1826 }
1827
1828 /*
1829  * Returns
1830  *  - 0 if PMD could not be locked
1831  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1832  *      or if prot_numa but THP migration is not supported
1833  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1834  */
1835 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1836                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1837                     unsigned long cp_flags)
1838 {
1839         struct mm_struct *mm = vma->vm_mm;
1840         spinlock_t *ptl;
1841         pmd_t oldpmd, entry;
1842         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1843         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1844         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1845         int ret = 1;
1846
1847         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1848
1849         if (prot_numa && !thp_migration_supported())
1850                 return 1;
1851
1852         ptl = __pmd_trans_huge_lock(pmd, vma);
1853         if (!ptl)
1854                 return 0;
1855
1856 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1857         if (is_swap_pmd(*pmd)) {
1858                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1859                 struct folio *folio = page_folio(pfn_swap_entry_to_page(entry));
1860                 pmd_t newpmd;
1861
1862                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1863                 if (is_writable_migration_entry(entry)) {
1864                         /*
1865                          * A protection check is difficult so
1866                          * just be safe and disable write
1867                          */
1868                         if (folio_test_anon(folio))
1869                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1870                         else
1871                                 entry = make_readable_migration_entry(swp_offset(entry));
1872                         newpmd = swp_entry_to_pmd(entry);
1873                         if (pmd_swp_soft_dirty(*pmd))
1874                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1875                 } else {
1876                         newpmd = *pmd;
1877                 }
1878
1879                 if (uffd_wp)
1880                         newpmd = pmd_swp_mkuffd_wp(newpmd);
1881                 else if (uffd_wp_resolve)
1882                         newpmd = pmd_swp_clear_uffd_wp(newpmd);
1883                 if (!pmd_same(*pmd, newpmd))
1884                         set_pmd_at(mm, addr, pmd, newpmd);
1885                 goto unlock;
1886         }
1887 #endif
1888
1889         if (prot_numa) {
1890                 struct folio *folio;
1891                 bool toptier;
1892                 /*
1893                  * Avoid trapping faults against the zero page. The read-only
1894                  * data is likely to be read-cached on the local CPU and
1895                  * local/remote hits to the zero page are not interesting.
1896                  */
1897                 if (is_huge_zero_pmd(*pmd))
1898                         goto unlock;
1899
1900                 if (pmd_protnone(*pmd))
1901                         goto unlock;
1902
1903                 folio = page_folio(pmd_page(*pmd));
1904                 toptier = node_is_toptier(folio_nid(folio));
1905                 /*
1906                  * Skip scanning top tier node if normal numa
1907                  * balancing is disabled
1908                  */
1909                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1910                     toptier)
1911                         goto unlock;
1912
1913                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1914                     !toptier)
1915                         folio_xchg_access_time(folio,
1916                                                jiffies_to_msecs(jiffies));
1917         }
1918         /*
1919          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1920          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1921          * which is also under mmap_read_lock(mm):
1922          *
1923          *      CPU0:                           CPU1:
1924          *                              change_huge_pmd(prot_numa=1)
1925          *                               pmdp_huge_get_and_clear_notify()
1926          * madvise_dontneed()
1927          *  zap_pmd_range()
1928          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1929          *   // skip the pmd
1930          *                               set_pmd_at();
1931          *                               // pmd is re-established
1932          *
1933          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1934          * which may break userspace.
1935          *
1936          * pmdp_invalidate_ad() is required to make sure we don't miss
1937          * dirty/young flags set by hardware.
1938          */
1939         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1940
1941         entry = pmd_modify(oldpmd, newprot);
1942         if (uffd_wp)
1943                 entry = pmd_mkuffd_wp(entry);
1944         else if (uffd_wp_resolve)
1945                 /*
1946                  * Leave the write bit to be handled by PF interrupt
1947                  * handler, then things like COW could be properly
1948                  * handled.
1949                  */
1950                 entry = pmd_clear_uffd_wp(entry);
1951
1952         /* See change_pte_range(). */
1953         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1954             can_change_pmd_writable(vma, addr, entry))
1955                 entry = pmd_mkwrite(entry, vma);
1956
1957         ret = HPAGE_PMD_NR;
1958         set_pmd_at(mm, addr, pmd, entry);
1959
1960         if (huge_pmd_needs_flush(oldpmd, entry))
1961                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1962 unlock:
1963         spin_unlock(ptl);
1964         return ret;
1965 }
1966
1967 /*
1968  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1969  *
1970  * Note that if it returns page table lock pointer, this routine returns without
1971  * unlocking page table lock. So callers must unlock it.
1972  */
1973 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1974 {
1975         spinlock_t *ptl;
1976         ptl = pmd_lock(vma->vm_mm, pmd);
1977         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1978                         pmd_devmap(*pmd)))
1979                 return ptl;
1980         spin_unlock(ptl);
1981         return NULL;
1982 }
1983
1984 /*
1985  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1986  *
1987  * Note that if it returns page table lock pointer, this routine returns without
1988  * unlocking page table lock. So callers must unlock it.
1989  */
1990 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1991 {
1992         spinlock_t *ptl;
1993
1994         ptl = pud_lock(vma->vm_mm, pud);
1995         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1996                 return ptl;
1997         spin_unlock(ptl);
1998         return NULL;
1999 }
2000
2001 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2002 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2003                  pud_t *pud, unsigned long addr)
2004 {
2005         spinlock_t *ptl;
2006
2007         ptl = __pud_trans_huge_lock(pud, vma);
2008         if (!ptl)
2009                 return 0;
2010
2011         pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2012         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2013         if (vma_is_special_huge(vma)) {
2014                 spin_unlock(ptl);
2015                 /* No zero page support yet */
2016         } else {
2017                 /* No support for anonymous PUD pages yet */
2018                 BUG();
2019         }
2020         return 1;
2021 }
2022
2023 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2024                 unsigned long haddr)
2025 {
2026         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2027         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2028         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2029         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2030
2031         count_vm_event(THP_SPLIT_PUD);
2032
2033         pudp_huge_clear_flush(vma, haddr, pud);
2034 }
2035
2036 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2037                 unsigned long address)
2038 {
2039         spinlock_t *ptl;
2040         struct mmu_notifier_range range;
2041
2042         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2043                                 address & HPAGE_PUD_MASK,
2044                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2045         mmu_notifier_invalidate_range_start(&range);
2046         ptl = pud_lock(vma->vm_mm, pud);
2047         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2048                 goto out;
2049         __split_huge_pud_locked(vma, pud, range.start);
2050
2051 out:
2052         spin_unlock(ptl);
2053         mmu_notifier_invalidate_range_end(&range);
2054 }
2055 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2056
2057 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2058                 unsigned long haddr, pmd_t *pmd)
2059 {
2060         struct mm_struct *mm = vma->vm_mm;
2061         pgtable_t pgtable;
2062         pmd_t _pmd, old_pmd;
2063         unsigned long addr;
2064         pte_t *pte;
2065         int i;
2066
2067         /*
2068          * Leave pmd empty until pte is filled note that it is fine to delay
2069          * notification until mmu_notifier_invalidate_range_end() as we are
2070          * replacing a zero pmd write protected page with a zero pte write
2071          * protected page.
2072          *
2073          * See Documentation/mm/mmu_notifier.rst
2074          */
2075         old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2076
2077         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2078         pmd_populate(mm, &_pmd, pgtable);
2079
2080         pte = pte_offset_map(&_pmd, haddr);
2081         VM_BUG_ON(!pte);
2082         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2083                 pte_t entry;
2084
2085                 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2086                 entry = pte_mkspecial(entry);
2087                 if (pmd_uffd_wp(old_pmd))
2088                         entry = pte_mkuffd_wp(entry);
2089                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2090                 set_pte_at(mm, addr, pte, entry);
2091                 pte++;
2092         }
2093         pte_unmap(pte - 1);
2094         smp_wmb(); /* make pte visible before pmd */
2095         pmd_populate(mm, pmd, pgtable);
2096 }
2097
2098 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2099                 unsigned long haddr, bool freeze)
2100 {
2101         struct mm_struct *mm = vma->vm_mm;
2102         struct page *page;
2103         pgtable_t pgtable;
2104         pmd_t old_pmd, _pmd;
2105         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2106         bool anon_exclusive = false, dirty = false;
2107         unsigned long addr;
2108         pte_t *pte;
2109         int i;
2110
2111         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2112         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2113         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2114         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2115                                 && !pmd_devmap(*pmd));
2116
2117         count_vm_event(THP_SPLIT_PMD);
2118
2119         if (!vma_is_anonymous(vma)) {
2120                 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2121                 /*
2122                  * We are going to unmap this huge page. So
2123                  * just go ahead and zap it
2124                  */
2125                 if (arch_needs_pgtable_deposit())
2126                         zap_deposited_table(mm, pmd);
2127                 if (vma_is_special_huge(vma))
2128                         return;
2129                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2130                         swp_entry_t entry;
2131
2132                         entry = pmd_to_swp_entry(old_pmd);
2133                         page = pfn_swap_entry_to_page(entry);
2134                 } else {
2135                         page = pmd_page(old_pmd);
2136                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2137                                 set_page_dirty(page);
2138                         if (!PageReferenced(page) && pmd_young(old_pmd))
2139                                 SetPageReferenced(page);
2140                         page_remove_rmap(page, vma, true);
2141                         put_page(page);
2142                 }
2143                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2144                 return;
2145         }
2146
2147         if (is_huge_zero_pmd(*pmd)) {
2148                 /*
2149                  * FIXME: Do we want to invalidate secondary mmu by calling
2150                  * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2151                  * inside __split_huge_pmd() ?
2152                  *
2153                  * We are going from a zero huge page write protected to zero
2154                  * small page also write protected so it does not seems useful
2155                  * to invalidate secondary mmu at this time.
2156                  */
2157                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2158         }
2159
2160         /*
2161          * Up to this point the pmd is present and huge and userland has the
2162          * whole access to the hugepage during the split (which happens in
2163          * place). If we overwrite the pmd with the not-huge version pointing
2164          * to the pte here (which of course we could if all CPUs were bug
2165          * free), userland could trigger a small page size TLB miss on the
2166          * small sized TLB while the hugepage TLB entry is still established in
2167          * the huge TLB. Some CPU doesn't like that.
2168          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2169          * 383 on page 105. Intel should be safe but is also warns that it's
2170          * only safe if the permission and cache attributes of the two entries
2171          * loaded in the two TLB is identical (which should be the case here).
2172          * But it is generally safer to never allow small and huge TLB entries
2173          * for the same virtual address to be loaded simultaneously. So instead
2174          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2175          * current pmd notpresent (atomically because here the pmd_trans_huge
2176          * must remain set at all times on the pmd until the split is complete
2177          * for this pmd), then we flush the SMP TLB and finally we write the
2178          * non-huge version of the pmd entry with pmd_populate.
2179          */
2180         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2181
2182         pmd_migration = is_pmd_migration_entry(old_pmd);
2183         if (unlikely(pmd_migration)) {
2184                 swp_entry_t entry;
2185
2186                 entry = pmd_to_swp_entry(old_pmd);
2187                 page = pfn_swap_entry_to_page(entry);
2188                 write = is_writable_migration_entry(entry);
2189                 if (PageAnon(page))
2190                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2191                 young = is_migration_entry_young(entry);
2192                 dirty = is_migration_entry_dirty(entry);
2193                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2194                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2195         } else {
2196                 page = pmd_page(old_pmd);
2197                 if (pmd_dirty(old_pmd)) {
2198                         dirty = true;
2199                         SetPageDirty(page);
2200                 }
2201                 write = pmd_write(old_pmd);
2202                 young = pmd_young(old_pmd);
2203                 soft_dirty = pmd_soft_dirty(old_pmd);
2204                 uffd_wp = pmd_uffd_wp(old_pmd);
2205
2206                 VM_BUG_ON_PAGE(!page_count(page), page);
2207
2208                 /*
2209                  * Without "freeze", we'll simply split the PMD, propagating the
2210                  * PageAnonExclusive() flag for each PTE by setting it for
2211                  * each subpage -- no need to (temporarily) clear.
2212                  *
2213                  * With "freeze" we want to replace mapped pages by
2214                  * migration entries right away. This is only possible if we
2215                  * managed to clear PageAnonExclusive() -- see
2216                  * set_pmd_migration_entry().
2217                  *
2218                  * In case we cannot clear PageAnonExclusive(), split the PMD
2219                  * only and let try_to_migrate_one() fail later.
2220                  *
2221                  * See page_try_share_anon_rmap(): invalidate PMD first.
2222                  */
2223                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2224                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2225                         freeze = false;
2226                 if (!freeze)
2227                         page_ref_add(page, HPAGE_PMD_NR - 1);
2228         }
2229
2230         /*
2231          * Withdraw the table only after we mark the pmd entry invalid.
2232          * This's critical for some architectures (Power).
2233          */
2234         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2235         pmd_populate(mm, &_pmd, pgtable);
2236
2237         pte = pte_offset_map(&_pmd, haddr);
2238         VM_BUG_ON(!pte);
2239         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2240                 pte_t entry;
2241                 /*
2242                  * Note that NUMA hinting access restrictions are not
2243                  * transferred to avoid any possibility of altering
2244                  * permissions across VMAs.
2245                  */
2246                 if (freeze || pmd_migration) {
2247                         swp_entry_t swp_entry;
2248                         if (write)
2249                                 swp_entry = make_writable_migration_entry(
2250                                                         page_to_pfn(page + i));
2251                         else if (anon_exclusive)
2252                                 swp_entry = make_readable_exclusive_migration_entry(
2253                                                         page_to_pfn(page + i));
2254                         else
2255                                 swp_entry = make_readable_migration_entry(
2256                                                         page_to_pfn(page + i));
2257                         if (young)
2258                                 swp_entry = make_migration_entry_young(swp_entry);
2259                         if (dirty)
2260                                 swp_entry = make_migration_entry_dirty(swp_entry);
2261                         entry = swp_entry_to_pte(swp_entry);
2262                         if (soft_dirty)
2263                                 entry = pte_swp_mksoft_dirty(entry);
2264                         if (uffd_wp)
2265                                 entry = pte_swp_mkuffd_wp(entry);
2266                 } else {
2267                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2268                         if (write)
2269                                 entry = pte_mkwrite(entry, vma);
2270                         if (anon_exclusive)
2271                                 SetPageAnonExclusive(page + i);
2272                         if (!young)
2273                                 entry = pte_mkold(entry);
2274                         /* NOTE: this may set soft-dirty too on some archs */
2275                         if (dirty)
2276                                 entry = pte_mkdirty(entry);
2277                         if (soft_dirty)
2278                                 entry = pte_mksoft_dirty(entry);
2279                         if (uffd_wp)
2280                                 entry = pte_mkuffd_wp(entry);
2281                         page_add_anon_rmap(page + i, vma, addr, RMAP_NONE);
2282                 }
2283                 VM_BUG_ON(!pte_none(ptep_get(pte)));
2284                 set_pte_at(mm, addr, pte, entry);
2285                 pte++;
2286         }
2287         pte_unmap(pte - 1);
2288
2289         if (!pmd_migration)
2290                 page_remove_rmap(page, vma, true);
2291         if (freeze)
2292                 put_page(page);
2293
2294         smp_wmb(); /* make pte visible before pmd */
2295         pmd_populate(mm, pmd, pgtable);
2296 }
2297
2298 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2299                 unsigned long address, bool freeze, struct folio *folio)
2300 {
2301         spinlock_t *ptl;
2302         struct mmu_notifier_range range;
2303
2304         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2305                                 address & HPAGE_PMD_MASK,
2306                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2307         mmu_notifier_invalidate_range_start(&range);
2308         ptl = pmd_lock(vma->vm_mm, pmd);
2309
2310         /*
2311          * If caller asks to setup a migration entry, we need a folio to check
2312          * pmd against. Otherwise we can end up replacing wrong folio.
2313          */
2314         VM_BUG_ON(freeze && !folio);
2315         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2316
2317         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2318             is_pmd_migration_entry(*pmd)) {
2319                 /*
2320                  * It's safe to call pmd_page when folio is set because it's
2321                  * guaranteed that pmd is present.
2322                  */
2323                 if (folio && folio != page_folio(pmd_page(*pmd)))
2324                         goto out;
2325                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2326         }
2327
2328 out:
2329         spin_unlock(ptl);
2330         mmu_notifier_invalidate_range_end(&range);
2331 }
2332
2333 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2334                 bool freeze, struct folio *folio)
2335 {
2336         pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2337
2338         if (!pmd)
2339                 return;
2340
2341         __split_huge_pmd(vma, pmd, address, freeze, folio);
2342 }
2343
2344 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2345 {
2346         /*
2347          * If the new address isn't hpage aligned and it could previously
2348          * contain an hugepage: check if we need to split an huge pmd.
2349          */
2350         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2351             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2352                          ALIGN(address, HPAGE_PMD_SIZE)))
2353                 split_huge_pmd_address(vma, address, false, NULL);
2354 }
2355
2356 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2357                              unsigned long start,
2358                              unsigned long end,
2359                              long adjust_next)
2360 {
2361         /* Check if we need to split start first. */
2362         split_huge_pmd_if_needed(vma, start);
2363
2364         /* Check if we need to split end next. */
2365         split_huge_pmd_if_needed(vma, end);
2366
2367         /*
2368          * If we're also updating the next vma vm_start,
2369          * check if we need to split it.
2370          */
2371         if (adjust_next > 0) {
2372                 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2373                 unsigned long nstart = next->vm_start;
2374                 nstart += adjust_next;
2375                 split_huge_pmd_if_needed(next, nstart);
2376         }
2377 }
2378
2379 static void unmap_folio(struct folio *folio)
2380 {
2381         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2382                 TTU_SYNC;
2383
2384         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2385
2386         /*
2387          * Anon pages need migration entries to preserve them, but file
2388          * pages can simply be left unmapped, then faulted back on demand.
2389          * If that is ever changed (perhaps for mlock), update remap_page().
2390          */
2391         if (folio_test_anon(folio))
2392                 try_to_migrate(folio, ttu_flags);
2393         else
2394                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2395 }
2396
2397 static void remap_page(struct folio *folio, unsigned long nr)
2398 {
2399         int i = 0;
2400
2401         /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2402         if (!folio_test_anon(folio))
2403                 return;
2404         for (;;) {
2405                 remove_migration_ptes(folio, folio, true);
2406                 i += folio_nr_pages(folio);
2407                 if (i >= nr)
2408                         break;
2409                 folio = folio_next(folio);
2410         }
2411 }
2412
2413 static void lru_add_page_tail(struct page *head, struct page *tail,
2414                 struct lruvec *lruvec, struct list_head *list)
2415 {
2416         VM_BUG_ON_PAGE(!PageHead(head), head);
2417         VM_BUG_ON_PAGE(PageCompound(tail), head);
2418         VM_BUG_ON_PAGE(PageLRU(tail), head);
2419         lockdep_assert_held(&lruvec->lru_lock);
2420
2421         if (list) {
2422                 /* page reclaim is reclaiming a huge page */
2423                 VM_WARN_ON(PageLRU(head));
2424                 get_page(tail);
2425                 list_add_tail(&tail->lru, list);
2426         } else {
2427                 /* head is still on lru (and we have it frozen) */
2428                 VM_WARN_ON(!PageLRU(head));
2429                 if (PageUnevictable(tail))
2430                         tail->mlock_count = 0;
2431                 else
2432                         list_add_tail(&tail->lru, &head->lru);
2433                 SetPageLRU(tail);
2434         }
2435 }
2436
2437 static void __split_huge_page_tail(struct folio *folio, int tail,
2438                 struct lruvec *lruvec, struct list_head *list)
2439 {
2440         struct page *head = &folio->page;
2441         struct page *page_tail = head + tail;
2442         /*
2443          * Careful: new_folio is not a "real" folio before we cleared PageTail.
2444          * Don't pass it around before clear_compound_head().
2445          */
2446         struct folio *new_folio = (struct folio *)page_tail;
2447
2448         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2449
2450         /*
2451          * Clone page flags before unfreezing refcount.
2452          *
2453          * After successful get_page_unless_zero() might follow flags change,
2454          * for example lock_page() which set PG_waiters.
2455          *
2456          * Note that for mapped sub-pages of an anonymous THP,
2457          * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2458          * the migration entry instead from where remap_page() will restore it.
2459          * We can still have PG_anon_exclusive set on effectively unmapped and
2460          * unreferenced sub-pages of an anonymous THP: we can simply drop
2461          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2462          */
2463         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2464         page_tail->flags |= (head->flags &
2465                         ((1L << PG_referenced) |
2466                          (1L << PG_swapbacked) |
2467                          (1L << PG_swapcache) |
2468                          (1L << PG_mlocked) |
2469                          (1L << PG_uptodate) |
2470                          (1L << PG_active) |
2471                          (1L << PG_workingset) |
2472                          (1L << PG_locked) |
2473                          (1L << PG_unevictable) |
2474 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2475                          (1L << PG_arch_2) |
2476                          (1L << PG_arch_3) |
2477 #endif
2478                          (1L << PG_dirty) |
2479                          LRU_GEN_MASK | LRU_REFS_MASK));
2480
2481         /* ->mapping in first and second tail page is replaced by other uses */
2482         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2483                         page_tail);
2484         page_tail->mapping = head->mapping;
2485         page_tail->index = head->index + tail;
2486
2487         /*
2488          * page->private should not be set in tail pages. Fix up and warn once
2489          * if private is unexpectedly set.
2490          */
2491         if (unlikely(page_tail->private)) {
2492                 VM_WARN_ON_ONCE_PAGE(true, page_tail);
2493                 page_tail->private = 0;
2494         }
2495         if (folio_test_swapcache(folio))
2496                 new_folio->swap.val = folio->swap.val + tail;
2497
2498         /* Page flags must be visible before we make the page non-compound. */
2499         smp_wmb();
2500
2501         /*
2502          * Clear PageTail before unfreezing page refcount.
2503          *
2504          * After successful get_page_unless_zero() might follow put_page()
2505          * which needs correct compound_head().
2506          */
2507         clear_compound_head(page_tail);
2508
2509         /* Finally unfreeze refcount. Additional reference from page cache. */
2510         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2511                                           PageSwapCache(head)));
2512
2513         if (page_is_young(head))
2514                 set_page_young(page_tail);
2515         if (page_is_idle(head))
2516                 set_page_idle(page_tail);
2517
2518         folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
2519
2520         /*
2521          * always add to the tail because some iterators expect new
2522          * pages to show after the currently processed elements - e.g.
2523          * migrate_pages
2524          */
2525         lru_add_page_tail(head, page_tail, lruvec, list);
2526 }
2527
2528 static void __split_huge_page(struct page *page, struct list_head *list,
2529                 pgoff_t end)
2530 {
2531         struct folio *folio = page_folio(page);
2532         struct page *head = &folio->page;
2533         struct lruvec *lruvec;
2534         struct address_space *swap_cache = NULL;
2535         unsigned long offset = 0;
2536         unsigned int nr = thp_nr_pages(head);
2537         int i, nr_dropped = 0;
2538
2539         /* complete memcg works before add pages to LRU */
2540         split_page_memcg(head, nr);
2541
2542         if (folio_test_anon(folio) && folio_test_swapcache(folio)) {
2543                 offset = swp_offset(folio->swap);
2544                 swap_cache = swap_address_space(folio->swap);
2545                 xa_lock(&swap_cache->i_pages);
2546         }
2547
2548         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2549         lruvec = folio_lruvec_lock(folio);
2550
2551         ClearPageHasHWPoisoned(head);
2552
2553         for (i = nr - 1; i >= 1; i--) {
2554                 __split_huge_page_tail(folio, i, lruvec, list);
2555                 /* Some pages can be beyond EOF: drop them from page cache */
2556                 if (head[i].index >= end) {
2557                         struct folio *tail = page_folio(head + i);
2558
2559                         if (shmem_mapping(head->mapping))
2560                                 nr_dropped++;
2561                         else if (folio_test_clear_dirty(tail))
2562                                 folio_account_cleaned(tail,
2563                                         inode_to_wb(folio->mapping->host));
2564                         __filemap_remove_folio(tail, NULL);
2565                         folio_put(tail);
2566                 } else if (!PageAnon(page)) {
2567                         __xa_store(&head->mapping->i_pages, head[i].index,
2568                                         head + i, 0);
2569                 } else if (swap_cache) {
2570                         __xa_store(&swap_cache->i_pages, offset + i,
2571                                         head + i, 0);
2572                 }
2573         }
2574
2575         ClearPageCompound(head);
2576         unlock_page_lruvec(lruvec);
2577         /* Caller disabled irqs, so they are still disabled here */
2578
2579         split_page_owner(head, nr);
2580
2581         /* See comment in __split_huge_page_tail() */
2582         if (PageAnon(head)) {
2583                 /* Additional pin to swap cache */
2584                 if (PageSwapCache(head)) {
2585                         page_ref_add(head, 2);
2586                         xa_unlock(&swap_cache->i_pages);
2587                 } else {
2588                         page_ref_inc(head);
2589                 }
2590         } else {
2591                 /* Additional pin to page cache */
2592                 page_ref_add(head, 2);
2593                 xa_unlock(&head->mapping->i_pages);
2594         }
2595         local_irq_enable();
2596
2597         if (nr_dropped)
2598                 shmem_uncharge(head->mapping->host, nr_dropped);
2599         remap_page(folio, nr);
2600
2601         if (folio_test_swapcache(folio))
2602                 split_swap_cluster(folio->swap);
2603
2604         for (i = 0; i < nr; i++) {
2605                 struct page *subpage = head + i;
2606                 if (subpage == page)
2607                         continue;
2608                 unlock_page(subpage);
2609
2610                 /*
2611                  * Subpages may be freed if there wasn't any mapping
2612                  * like if add_to_swap() is running on a lru page that
2613                  * had its mapping zapped. And freeing these pages
2614                  * requires taking the lru_lock so we do the put_page
2615                  * of the tail pages after the split is complete.
2616                  */
2617                 free_page_and_swap_cache(subpage);
2618         }
2619 }
2620
2621 /* Racy check whether the huge page can be split */
2622 bool can_split_folio(struct folio *folio, int *pextra_pins)
2623 {
2624         int extra_pins;
2625
2626         /* Additional pins from page cache */
2627         if (folio_test_anon(folio))
2628                 extra_pins = folio_test_swapcache(folio) ?
2629                                 folio_nr_pages(folio) : 0;
2630         else
2631                 extra_pins = folio_nr_pages(folio);
2632         if (pextra_pins)
2633                 *pextra_pins = extra_pins;
2634         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2635 }
2636
2637 /*
2638  * This function splits huge page into normal pages. @page can point to any
2639  * subpage of huge page to split. Split doesn't change the position of @page.
2640  *
2641  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2642  * The huge page must be locked.
2643  *
2644  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2645  *
2646  * Both head page and tail pages will inherit mapping, flags, and so on from
2647  * the hugepage.
2648  *
2649  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2650  * they are not mapped.
2651  *
2652  * Returns 0 if the hugepage is split successfully.
2653  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2654  * us.
2655  */
2656 int split_huge_page_to_list(struct page *page, struct list_head *list)
2657 {
2658         struct folio *folio = page_folio(page);
2659         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2660         XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2661         struct anon_vma *anon_vma = NULL;
2662         struct address_space *mapping = NULL;
2663         int extra_pins, ret;
2664         pgoff_t end;
2665         bool is_hzp;
2666
2667         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2668         VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2669
2670         is_hzp = is_huge_zero_page(&folio->page);
2671         if (is_hzp) {
2672                 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2673                 return -EBUSY;
2674         }
2675
2676         if (folio_test_writeback(folio))
2677                 return -EBUSY;
2678
2679         if (folio_test_anon(folio)) {
2680                 /*
2681                  * The caller does not necessarily hold an mmap_lock that would
2682                  * prevent the anon_vma disappearing so we first we take a
2683                  * reference to it and then lock the anon_vma for write. This
2684                  * is similar to folio_lock_anon_vma_read except the write lock
2685                  * is taken to serialise against parallel split or collapse
2686                  * operations.
2687                  */
2688                 anon_vma = folio_get_anon_vma(folio);
2689                 if (!anon_vma) {
2690                         ret = -EBUSY;
2691                         goto out;
2692                 }
2693                 end = -1;
2694                 mapping = NULL;
2695                 anon_vma_lock_write(anon_vma);
2696         } else {
2697                 gfp_t gfp;
2698
2699                 mapping = folio->mapping;
2700
2701                 /* Truncated ? */
2702                 if (!mapping) {
2703                         ret = -EBUSY;
2704                         goto out;
2705                 }
2706
2707                 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2708                                                         GFP_RECLAIM_MASK);
2709
2710                 if (!filemap_release_folio(folio, gfp)) {
2711                         ret = -EBUSY;
2712                         goto out;
2713                 }
2714
2715                 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2716                 if (xas_error(&xas)) {
2717                         ret = xas_error(&xas);
2718                         goto out;
2719                 }
2720
2721                 anon_vma = NULL;
2722                 i_mmap_lock_read(mapping);
2723
2724                 /*
2725                  *__split_huge_page() may need to trim off pages beyond EOF:
2726                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2727                  * which cannot be nested inside the page tree lock. So note
2728                  * end now: i_size itself may be changed at any moment, but
2729                  * folio lock is good enough to serialize the trimming.
2730                  */
2731                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2732                 if (shmem_mapping(mapping))
2733                         end = shmem_fallocend(mapping->host, end);
2734         }
2735
2736         /*
2737          * Racy check if we can split the page, before unmap_folio() will
2738          * split PMDs
2739          */
2740         if (!can_split_folio(folio, &extra_pins)) {
2741                 ret = -EAGAIN;
2742                 goto out_unlock;
2743         }
2744
2745         unmap_folio(folio);
2746
2747         /* block interrupt reentry in xa_lock and spinlock */
2748         local_irq_disable();
2749         if (mapping) {
2750                 /*
2751                  * Check if the folio is present in page cache.
2752                  * We assume all tail are present too, if folio is there.
2753                  */
2754                 xas_lock(&xas);
2755                 xas_reset(&xas);
2756                 if (xas_load(&xas) != folio)
2757                         goto fail;
2758         }
2759
2760         /* Prevent deferred_split_scan() touching ->_refcount */
2761         spin_lock(&ds_queue->split_queue_lock);
2762         if (folio_ref_freeze(folio, 1 + extra_pins)) {
2763                 if (!list_empty(&folio->_deferred_list)) {
2764                         ds_queue->split_queue_len--;
2765                         list_del(&folio->_deferred_list);
2766                 }
2767                 spin_unlock(&ds_queue->split_queue_lock);
2768                 if (mapping) {
2769                         int nr = folio_nr_pages(folio);
2770
2771                         xas_split(&xas, folio, folio_order(folio));
2772                         if (folio_test_pmd_mappable(folio)) {
2773                                 if (folio_test_swapbacked(folio)) {
2774                                         __lruvec_stat_mod_folio(folio,
2775                                                         NR_SHMEM_THPS, -nr);
2776                                 } else {
2777                                         __lruvec_stat_mod_folio(folio,
2778                                                         NR_FILE_THPS, -nr);
2779                                         filemap_nr_thps_dec(mapping);
2780                                 }
2781                         }
2782                 }
2783
2784                 __split_huge_page(page, list, end);
2785                 ret = 0;
2786         } else {
2787                 spin_unlock(&ds_queue->split_queue_lock);
2788 fail:
2789                 if (mapping)
2790                         xas_unlock(&xas);
2791                 local_irq_enable();
2792                 remap_page(folio, folio_nr_pages(folio));
2793                 ret = -EAGAIN;
2794         }
2795
2796 out_unlock:
2797         if (anon_vma) {
2798                 anon_vma_unlock_write(anon_vma);
2799                 put_anon_vma(anon_vma);
2800         }
2801         if (mapping)
2802                 i_mmap_unlock_read(mapping);
2803 out:
2804         xas_destroy(&xas);
2805         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2806         return ret;
2807 }
2808
2809 void folio_undo_large_rmappable(struct folio *folio)
2810 {
2811         struct deferred_split *ds_queue;
2812         unsigned long flags;
2813
2814         /*
2815          * At this point, there is no one trying to add the folio to
2816          * deferred_list. If folio is not in deferred_list, it's safe
2817          * to check without acquiring the split_queue_lock.
2818          */
2819         if (data_race(list_empty(&folio->_deferred_list)))
2820                 return;
2821
2822         ds_queue = get_deferred_split_queue(folio);
2823         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2824         if (!list_empty(&folio->_deferred_list)) {
2825                 ds_queue->split_queue_len--;
2826                 list_del(&folio->_deferred_list);
2827         }
2828         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2829 }
2830
2831 void deferred_split_folio(struct folio *folio)
2832 {
2833         struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2834 #ifdef CONFIG_MEMCG
2835         struct mem_cgroup *memcg = folio_memcg(folio);
2836 #endif
2837         unsigned long flags;
2838
2839         VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2840
2841         /*
2842          * The try_to_unmap() in page reclaim path might reach here too,
2843          * this may cause a race condition to corrupt deferred split queue.
2844          * And, if page reclaim is already handling the same folio, it is
2845          * unnecessary to handle it again in shrinker.
2846          *
2847          * Check the swapcache flag to determine if the folio is being
2848          * handled by page reclaim since THP swap would add the folio into
2849          * swap cache before calling try_to_unmap().
2850          */
2851         if (folio_test_swapcache(folio))
2852                 return;
2853
2854         if (!list_empty(&folio->_deferred_list))
2855                 return;
2856
2857         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2858         if (list_empty(&folio->_deferred_list)) {
2859                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2860                 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2861                 ds_queue->split_queue_len++;
2862 #ifdef CONFIG_MEMCG
2863                 if (memcg)
2864                         set_shrinker_bit(memcg, folio_nid(folio),
2865                                          deferred_split_shrinker->id);
2866 #endif
2867         }
2868         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2869 }
2870
2871 static unsigned long deferred_split_count(struct shrinker *shrink,
2872                 struct shrink_control *sc)
2873 {
2874         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2875         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2876
2877 #ifdef CONFIG_MEMCG
2878         if (sc->memcg)
2879                 ds_queue = &sc->memcg->deferred_split_queue;
2880 #endif
2881         return READ_ONCE(ds_queue->split_queue_len);
2882 }
2883
2884 static unsigned long deferred_split_scan(struct shrinker *shrink,
2885                 struct shrink_control *sc)
2886 {
2887         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2888         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2889         unsigned long flags;
2890         LIST_HEAD(list);
2891         struct folio *folio, *next;
2892         int split = 0;
2893
2894 #ifdef CONFIG_MEMCG
2895         if (sc->memcg)
2896                 ds_queue = &sc->memcg->deferred_split_queue;
2897 #endif
2898
2899         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2900         /* Take pin on all head pages to avoid freeing them under us */
2901         list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2902                                                         _deferred_list) {
2903                 if (folio_try_get(folio)) {
2904                         list_move(&folio->_deferred_list, &list);
2905                 } else {
2906                         /* We lost race with folio_put() */
2907                         list_del_init(&folio->_deferred_list);
2908                         ds_queue->split_queue_len--;
2909                 }
2910                 if (!--sc->nr_to_scan)
2911                         break;
2912         }
2913         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2914
2915         list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2916                 if (!folio_trylock(folio))
2917                         goto next;
2918                 /* split_huge_page() removes page from list on success */
2919                 if (!split_folio(folio))
2920                         split++;
2921                 folio_unlock(folio);
2922 next:
2923                 folio_put(folio);
2924         }
2925
2926         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2927         list_splice_tail(&list, &ds_queue->split_queue);
2928         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2929
2930         /*
2931          * Stop shrinker if we didn't split any page, but the queue is empty.
2932          * This can happen if pages were freed under us.
2933          */
2934         if (!split && list_empty(&ds_queue->split_queue))
2935                 return SHRINK_STOP;
2936         return split;
2937 }
2938
2939 #ifdef CONFIG_DEBUG_FS
2940 static void split_huge_pages_all(void)
2941 {
2942         struct zone *zone;
2943         struct page *page;
2944         struct folio *folio;
2945         unsigned long pfn, max_zone_pfn;
2946         unsigned long total = 0, split = 0;
2947
2948         pr_debug("Split all THPs\n");
2949         for_each_zone(zone) {
2950                 if (!managed_zone(zone))
2951                         continue;
2952                 max_zone_pfn = zone_end_pfn(zone);
2953                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2954                         int nr_pages;
2955
2956                         page = pfn_to_online_page(pfn);
2957                         if (!page || PageTail(page))
2958                                 continue;
2959                         folio = page_folio(page);
2960                         if (!folio_try_get(folio))
2961                                 continue;
2962
2963                         if (unlikely(page_folio(page) != folio))
2964                                 goto next;
2965
2966                         if (zone != folio_zone(folio))
2967                                 goto next;
2968
2969                         if (!folio_test_large(folio)
2970                                 || folio_test_hugetlb(folio)
2971                                 || !folio_test_lru(folio))
2972                                 goto next;
2973
2974                         total++;
2975                         folio_lock(folio);
2976                         nr_pages = folio_nr_pages(folio);
2977                         if (!split_folio(folio))
2978                                 split++;
2979                         pfn += nr_pages - 1;
2980                         folio_unlock(folio);
2981 next:
2982                         folio_put(folio);
2983                         cond_resched();
2984                 }
2985         }
2986
2987         pr_debug("%lu of %lu THP split\n", split, total);
2988 }
2989
2990 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2991 {
2992         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2993                     is_vm_hugetlb_page(vma);
2994 }
2995
2996 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2997                                 unsigned long vaddr_end)
2998 {
2999         int ret = 0;
3000         struct task_struct *task;
3001         struct mm_struct *mm;
3002         unsigned long total = 0, split = 0;
3003         unsigned long addr;
3004
3005         vaddr_start &= PAGE_MASK;
3006         vaddr_end &= PAGE_MASK;
3007
3008         /* Find the task_struct from pid */
3009         rcu_read_lock();
3010         task = find_task_by_vpid(pid);
3011         if (!task) {
3012                 rcu_read_unlock();
3013                 ret = -ESRCH;
3014                 goto out;
3015         }
3016         get_task_struct(task);
3017         rcu_read_unlock();
3018
3019         /* Find the mm_struct */
3020         mm = get_task_mm(task);
3021         put_task_struct(task);
3022
3023         if (!mm) {
3024                 ret = -EINVAL;
3025                 goto out;
3026         }
3027
3028         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3029                  pid, vaddr_start, vaddr_end);
3030
3031         mmap_read_lock(mm);
3032         /*
3033          * always increase addr by PAGE_SIZE, since we could have a PTE page
3034          * table filled with PTE-mapped THPs, each of which is distinct.
3035          */
3036         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3037                 struct vm_area_struct *vma = vma_lookup(mm, addr);
3038                 struct page *page;
3039                 struct folio *folio;
3040
3041                 if (!vma)
3042                         break;
3043
3044                 /* skip special VMA and hugetlb VMA */
3045                 if (vma_not_suitable_for_thp_split(vma)) {
3046                         addr = vma->vm_end;
3047                         continue;
3048                 }
3049
3050                 /* FOLL_DUMP to ignore special (like zero) pages */
3051                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3052
3053                 if (IS_ERR_OR_NULL(page))
3054                         continue;
3055
3056                 folio = page_folio(page);
3057                 if (!is_transparent_hugepage(folio))
3058                         goto next;
3059
3060                 total++;
3061                 if (!can_split_folio(folio, NULL))
3062                         goto next;
3063
3064                 if (!folio_trylock(folio))
3065                         goto next;
3066
3067                 if (!split_folio(folio))
3068                         split++;
3069
3070                 folio_unlock(folio);
3071 next:
3072                 folio_put(folio);
3073                 cond_resched();
3074         }
3075         mmap_read_unlock(mm);
3076         mmput(mm);
3077
3078         pr_debug("%lu of %lu THP split\n", split, total);
3079
3080 out:
3081         return ret;
3082 }
3083
3084 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3085                                 pgoff_t off_end)
3086 {
3087         struct filename *file;
3088         struct file *candidate;
3089         struct address_space *mapping;
3090         int ret = -EINVAL;
3091         pgoff_t index;
3092         int nr_pages = 1;
3093         unsigned long total = 0, split = 0;
3094
3095         file = getname_kernel(file_path);
3096         if (IS_ERR(file))
3097                 return ret;
3098
3099         candidate = file_open_name(file, O_RDONLY, 0);
3100         if (IS_ERR(candidate))
3101                 goto out;
3102
3103         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3104                  file_path, off_start, off_end);
3105
3106         mapping = candidate->f_mapping;
3107
3108         for (index = off_start; index < off_end; index += nr_pages) {
3109                 struct folio *folio = filemap_get_folio(mapping, index);
3110
3111                 nr_pages = 1;
3112                 if (IS_ERR(folio))
3113                         continue;
3114
3115                 if (!folio_test_large(folio))
3116                         goto next;
3117
3118                 total++;
3119                 nr_pages = folio_nr_pages(folio);
3120
3121                 if (!folio_trylock(folio))
3122                         goto next;
3123
3124                 if (!split_folio(folio))
3125                         split++;
3126
3127                 folio_unlock(folio);
3128 next:
3129                 folio_put(folio);
3130                 cond_resched();
3131         }
3132
3133         filp_close(candidate, NULL);
3134         ret = 0;
3135
3136         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3137 out:
3138         putname(file);
3139         return ret;
3140 }
3141
3142 #define MAX_INPUT_BUF_SZ 255
3143
3144 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3145                                 size_t count, loff_t *ppops)
3146 {
3147         static DEFINE_MUTEX(split_debug_mutex);
3148         ssize_t ret;
3149         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3150         char input_buf[MAX_INPUT_BUF_SZ];
3151         int pid;
3152         unsigned long vaddr_start, vaddr_end;
3153
3154         ret = mutex_lock_interruptible(&split_debug_mutex);
3155         if (ret)
3156                 return ret;
3157
3158         ret = -EFAULT;
3159
3160         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3161         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3162                 goto out;
3163
3164         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3165
3166         if (input_buf[0] == '/') {
3167                 char *tok;
3168                 char *buf = input_buf;
3169                 char file_path[MAX_INPUT_BUF_SZ];
3170                 pgoff_t off_start = 0, off_end = 0;
3171                 size_t input_len = strlen(input_buf);
3172
3173                 tok = strsep(&buf, ",");
3174                 if (tok) {
3175                         strcpy(file_path, tok);
3176                 } else {
3177                         ret = -EINVAL;
3178                         goto out;
3179                 }
3180
3181                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3182                 if (ret != 2) {
3183                         ret = -EINVAL;
3184                         goto out;
3185                 }
3186                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3187                 if (!ret)
3188                         ret = input_len;
3189
3190                 goto out;
3191         }
3192
3193         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3194         if (ret == 1 && pid == 1) {
3195                 split_huge_pages_all();
3196                 ret = strlen(input_buf);
3197                 goto out;
3198         } else if (ret != 3) {
3199                 ret = -EINVAL;
3200                 goto out;
3201         }
3202
3203         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3204         if (!ret)
3205                 ret = strlen(input_buf);
3206 out:
3207         mutex_unlock(&split_debug_mutex);
3208         return ret;
3209
3210 }
3211
3212 static const struct file_operations split_huge_pages_fops = {
3213         .owner   = THIS_MODULE,
3214         .write   = split_huge_pages_write,
3215         .llseek  = no_llseek,
3216 };
3217
3218 static int __init split_huge_pages_debugfs(void)
3219 {
3220         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3221                             &split_huge_pages_fops);
3222         return 0;
3223 }
3224 late_initcall(split_huge_pages_debugfs);
3225 #endif
3226
3227 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3228 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3229                 struct page *page)
3230 {
3231         struct vm_area_struct *vma = pvmw->vma;
3232         struct mm_struct *mm = vma->vm_mm;
3233         unsigned long address = pvmw->address;
3234         bool anon_exclusive;
3235         pmd_t pmdval;
3236         swp_entry_t entry;
3237         pmd_t pmdswp;
3238
3239         if (!(pvmw->pmd && !pvmw->pte))
3240                 return 0;
3241
3242         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3243         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3244
3245         /* See page_try_share_anon_rmap(): invalidate PMD first. */
3246         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3247         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3248                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3249                 return -EBUSY;
3250         }
3251
3252         if (pmd_dirty(pmdval))
3253                 set_page_dirty(page);
3254         if (pmd_write(pmdval))
3255                 entry = make_writable_migration_entry(page_to_pfn(page));
3256         else if (anon_exclusive)
3257                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3258         else
3259                 entry = make_readable_migration_entry(page_to_pfn(page));
3260         if (pmd_young(pmdval))
3261                 entry = make_migration_entry_young(entry);
3262         if (pmd_dirty(pmdval))
3263                 entry = make_migration_entry_dirty(entry);
3264         pmdswp = swp_entry_to_pmd(entry);
3265         if (pmd_soft_dirty(pmdval))
3266                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3267         if (pmd_uffd_wp(pmdval))
3268                 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3269         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3270         page_remove_rmap(page, vma, true);
3271         put_page(page);
3272         trace_set_migration_pmd(address, pmd_val(pmdswp));
3273
3274         return 0;
3275 }
3276
3277 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3278 {
3279         struct vm_area_struct *vma = pvmw->vma;
3280         struct mm_struct *mm = vma->vm_mm;
3281         unsigned long address = pvmw->address;
3282         unsigned long haddr = address & HPAGE_PMD_MASK;
3283         pmd_t pmde;
3284         swp_entry_t entry;
3285
3286         if (!(pvmw->pmd && !pvmw->pte))
3287                 return;
3288
3289         entry = pmd_to_swp_entry(*pvmw->pmd);
3290         get_page(new);
3291         pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3292         if (pmd_swp_soft_dirty(*pvmw->pmd))
3293                 pmde = pmd_mksoft_dirty(pmde);
3294         if (is_writable_migration_entry(entry))
3295                 pmde = pmd_mkwrite(pmde, vma);
3296         if (pmd_swp_uffd_wp(*pvmw->pmd))
3297                 pmde = pmd_mkuffd_wp(pmde);
3298         if (!is_migration_entry_young(entry))
3299                 pmde = pmd_mkold(pmde);
3300         /* NOTE: this may contain setting soft-dirty on some archs */
3301         if (PageDirty(new) && is_migration_entry_dirty(entry))
3302                 pmde = pmd_mkdirty(pmde);
3303
3304         if (PageAnon(new)) {
3305                 rmap_t rmap_flags = RMAP_COMPOUND;
3306
3307                 if (!is_readable_migration_entry(entry))
3308                         rmap_flags |= RMAP_EXCLUSIVE;
3309
3310                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3311         } else {
3312                 page_add_file_rmap(new, vma, true);
3313         }
3314         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3315         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3316
3317         /* No need to invalidate - it was non-present before */
3318         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3319         trace_remove_migration_pmd(address, pmd_val(pmde));
3320 }
3321 #endif