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
  40 #include <asm/tlb.h>
  41 #include <asm/pgalloc.h>
  42 #include "internal.h"
  43 #include "swap.h"
  44
  45 #define CREATE_TRACE_POINTS
  46 #include <trace/events/thp.h>
  47
  48 /*
  49  * By default, transparent hugepage support is disabled in order to avoid
  50  * risking an increased memory footprint for applications that are not
  51  * guaranteed to benefit from it. When transparent hugepage support is
  52  * enabled, it is for all mappings, and khugepaged scans all mappings.
  53  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  54  * for all hugepage allocations.
  55  */
  56 unsigned long transparent_hugepage_flags __read_mostly =
  57 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  58         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  59 #endif
  60 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  61         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  62 #endif
  63         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  64         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  65         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  66
  67 static struct shrinker deferred_split_shrinker;
  68
  69 static atomic_t huge_zero_refcount;
  70 struct page *huge_zero_page __read_mostly;
  71 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  72
  73 bool hugepage_vma_check(struct vm_area_struct *vma,
  74                         unsigned long vm_flags,
  75                         bool smaps, bool in_pf)
  76 {
  77         if (!vma->vm_mm)                /* vdso */
  78                 return false;
  79
  80         /*
  81          * Explicitly disabled through madvise or prctl, or some
  82          * architectures may disable THP for some mappings, for
  83          * example, s390 kvm.
  84          * */
  85         if ((vm_flags & VM_NOHUGEPAGE) ||
  86             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
  87                 return false;
  88         /*
  89          * If the hardware/firmware marked hugepage support disabled.
  90          */
  91         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
  92                 return false;
  93
  94         /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
  95         if (vma_is_dax(vma))
  96                 return in_pf;
  97
  98         /*
  99          * Special VMA and hugetlb VMA.
 100          * Must be checked after dax since some dax mappings may have
 101          * VM_MIXEDMAP set.
 102          */
 103         if (vm_flags & VM_NO_KHUGEPAGED)
 104                 return false;
 105
 106         /*
 107          * Check alignment for file vma and size for both file and anon vma.
 108          *
 109          * Skip the check for page fault. Huge fault does the check in fault
 110          * handlers. And this check is not suitable for huge PUD fault.
 111          */
 112         if (!in_pf &&
 113             !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
 114                 return false;
 115
 116         /*
 117          * Enabled via shmem mount options or sysfs settings.
 118          * Must be done before hugepage flags check since shmem has its
 119          * own flags.
 120          */
 121         if (!in_pf && shmem_file(vma->vm_file))
 122                 return shmem_huge_enabled(vma);
 123
 124         if (!hugepage_flags_enabled())
 125                 return false;
 126
 127         /* THP settings require madvise. */
 128         if (!(vm_flags & VM_HUGEPAGE) && !hugepage_flags_always())
 129                 return false;
 130
 131         /* Only regular file is valid */
 132         if (!in_pf && file_thp_enabled(vma))
 133                 return true;
 134
 135         if (!vma_is_anonymous(vma))
 136                 return false;
 137
 138         if (vma_is_temporary_stack(vma))
 139                 return false;
 140
 141         /*
 142          * THPeligible bit of smaps should show 1 for proper VMAs even
 143          * though anon_vma is not initialized yet.
 144          *
 145          * Allow page fault since anon_vma may be not initialized until
 146          * the first page fault.
 147          */
 148         if (!vma->anon_vma)
 149                 return (smaps || in_pf);
 150
 151         return true;
 152 }
 153
 154 static bool get_huge_zero_page(void)
 155 {
 156         struct page *zero_page;
 157 retry:
 158         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 159                 return true;
 160
 161         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 162                         HPAGE_PMD_ORDER);
 163         if (!zero_page) {
 164                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 165                 return false;
 166         }
 167         count_vm_event(THP_ZERO_PAGE_ALLOC);
 168         preempt_disable();
 169         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 170                 preempt_enable();
 171                 __free_pages(zero_page, compound_order(zero_page));
 172                 goto retry;
 173         }
 174         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 175
 176         /* We take additional reference here. It will be put back by shrinker */
 177         atomic_set(&huge_zero_refcount, 2);
 178         preempt_enable();
 179         return true;
 180 }
 181
 182 static void put_huge_zero_page(void)
 183 {
 184         /*
 185          * Counter should never go to zero here. Only shrinker can put
 186          * last reference.
 187          */
 188         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 189 }
 190
 191 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 192 {
 193         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 194                 return READ_ONCE(huge_zero_page);
 195
 196         if (!get_huge_zero_page())
 197                 return NULL;
 198
 199         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 200                 put_huge_zero_page();
 201
 202         return READ_ONCE(huge_zero_page);
 203 }
 204
 205 void mm_put_huge_zero_page(struct mm_struct *mm)
 206 {
 207         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 208                 put_huge_zero_page();
 209 }
 210
 211 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 212                                         struct shrink_control *sc)
 213 {
 214         /* we can free zero page only if last reference remains */
 215         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 216 }
 217
 218 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 219                                        struct shrink_control *sc)
 220 {
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 return HPAGE_PMD_NR;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .count_objects = shrink_huge_zero_page_count,
 234         .scan_objects = shrink_huge_zero_page_scan,
 235         .seeks = DEFAULT_SEEKS,
 236 };
 237
 238 #ifdef CONFIG_SYSFS
 239 static ssize_t enabled_show(struct kobject *kobj,
 240                             struct kobj_attribute *attr, char *buf)
 241 {
 242         const char *output;
 243
 244         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 245                 output = "[always] madvise never";
 246         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 247                           &transparent_hugepage_flags))
 248                 output = "always [madvise] never";
 249         else
 250                 output = "always madvise [never]";
 251
 252         return sysfs_emit(buf, "%s\n", output);
 253 }
 254
 255 static ssize_t enabled_store(struct kobject *kobj,
 256                              struct kobj_attribute *attr,
 257                              const char *buf, size_t count)
 258 {
 259         ssize_t ret = count;
 260
 261         if (sysfs_streq(buf, "always")) {
 262                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 263                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 264         } else if (sysfs_streq(buf, "madvise")) {
 265                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 266                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 267         } else if (sysfs_streq(buf, "never")) {
 268                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 269                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else
 271                 ret = -EINVAL;
 272
 273         if (ret > 0) {
 274                 int err = start_stop_khugepaged();
 275                 if (err)
 276                         ret = err;
 277         }
 278         return ret;
 279 }
 280
 281 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
 282
 283 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 284                                   struct kobj_attribute *attr, char *buf,
 285                                   enum transparent_hugepage_flag flag)
 286 {
 287         return sysfs_emit(buf, "%d\n",
 288                           !!test_bit(flag, &transparent_hugepage_flags));
 289 }
 290
 291 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 292                                  struct kobj_attribute *attr,
 293                                  const char *buf, size_t count,
 294                                  enum transparent_hugepage_flag flag)
 295 {
 296         unsigned long value;
 297         int ret;
 298
 299         ret = kstrtoul(buf, 10, &value);
 300         if (ret < 0)
 301                 return ret;
 302         if (value > 1)
 303                 return -EINVAL;
 304
 305         if (value)
 306                 set_bit(flag, &transparent_hugepage_flags);
 307         else
 308                 clear_bit(flag, &transparent_hugepage_flags);
 309
 310         return count;
 311 }
 312
 313 static ssize_t defrag_show(struct kobject *kobj,
 314                            struct kobj_attribute *attr, char *buf)
 315 {
 316         const char *output;
 317
 318         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 319                      &transparent_hugepage_flags))
 320                 output = "[always] defer defer+madvise madvise never";
 321         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 322                           &transparent_hugepage_flags))
 323                 output = "always [defer] defer+madvise madvise never";
 324         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 325                           &transparent_hugepage_flags))
 326                 output = "always defer [defer+madvise] madvise never";
 327         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 328                           &transparent_hugepage_flags))
 329                 output = "always defer defer+madvise [madvise] never";
 330         else
 331                 output = "always defer defer+madvise madvise [never]";
 332
 333         return sysfs_emit(buf, "%s\n", output);
 334 }
 335
 336 static ssize_t defrag_store(struct kobject *kobj,
 337                             struct kobj_attribute *attr,
 338                             const char *buf, size_t count)
 339 {
 340         if (sysfs_streq(buf, "always")) {
 341                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 342                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 343                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 344                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 345         } else if (sysfs_streq(buf, "defer+madvise")) {
 346                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 347                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 348                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 349                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 350         } else if (sysfs_streq(buf, "defer")) {
 351                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 352                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 353                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 354                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 355         } else if (sysfs_streq(buf, "madvise")) {
 356                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 357                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 358                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 359                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 360         } else if (sysfs_streq(buf, "never")) {
 361                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 362                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 363                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 364                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 365         } else
 366                 return -EINVAL;
 367
 368         return count;
 369 }
 370 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
 371
 372 static ssize_t use_zero_page_show(struct kobject *kobj,
 373                                   struct kobj_attribute *attr, char *buf)
 374 {
 375         return single_hugepage_flag_show(kobj, attr, buf,
 376                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static ssize_t use_zero_page_store(struct kobject *kobj,
 379                 struct kobj_attribute *attr, const char *buf, size_t count)
 380 {
 381         return single_hugepage_flag_store(kobj, attr, buf, count,
 382                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 383 }
 384 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
 385
 386 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 387                                    struct kobj_attribute *attr, char *buf)
 388 {
 389         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 390 }
 391 static struct kobj_attribute hpage_pmd_size_attr =
 392         __ATTR_RO(hpage_pmd_size);
 393
 394 static struct attribute *hugepage_attr[] = {
 395         &enabled_attr.attr,
 396         &defrag_attr.attr,
 397         &use_zero_page_attr.attr,
 398         &hpage_pmd_size_attr.attr,
 399 #ifdef CONFIG_SHMEM
 400         &shmem_enabled_attr.attr,
 401 #endif
 402         NULL,
 403 };
 404
 405 static const struct attribute_group hugepage_attr_group = {
 406         .attrs = hugepage_attr,
 407 };
 408
 409 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 410 {
 411         int err;
 412
 413         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 414         if (unlikely(!*hugepage_kobj)) {
 415                 pr_err("failed to create transparent hugepage kobject\n");
 416                 return -ENOMEM;
 417         }
 418
 419         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 420         if (err) {
 421                 pr_err("failed to register transparent hugepage group\n");
 422                 goto delete_obj;
 423         }
 424
 425         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 426         if (err) {
 427                 pr_err("failed to register transparent hugepage group\n");
 428                 goto remove_hp_group;
 429         }
 430
 431         return 0;
 432
 433 remove_hp_group:
 434         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 435 delete_obj:
 436         kobject_put(*hugepage_kobj);
 437         return err;
 438 }
 439
 440 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 441 {
 442         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 443         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 444         kobject_put(hugepage_kobj);
 445 }
 446 #else
 447 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 448 {
 449         return 0;
 450 }
 451
 452 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 453 {
 454 }
 455 #endif /* CONFIG_SYSFS */
 456
 457 static int __init hugepage_init(void)
 458 {
 459         int err;
 460         struct kobject *hugepage_kobj;
 461
 462         if (!has_transparent_hugepage()) {
 463                 /*
 464                  * Hardware doesn't support hugepages, hence disable
 465                  * DAX PMD support.
 466                  */
 467                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 468                 return -EINVAL;
 469         }
 470
 471         /*
 472          * hugepages can't be allocated by the buddy allocator
 473          */
 474         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 475         /*
 476          * we use page->mapping and page->index in second tail page
 477          * as list_head: assuming THP order >= 2
 478          */
 479         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 480
 481         err = hugepage_init_sysfs(&hugepage_kobj);
 482         if (err)
 483                 goto err_sysfs;
 484
 485         err = khugepaged_init();
 486         if (err)
 487                 goto err_slab;
 488
 489         err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 490         if (err)
 491                 goto err_hzp_shrinker;
 492         err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 493         if (err)
 494                 goto err_split_shrinker;
 495
 496         /*
 497          * By default disable transparent hugepages on smaller systems,
 498          * where the extra memory used could hurt more than TLB overhead
 499          * is likely to save.  The admin can still enable it through /sys.
 500          */
 501         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 502                 transparent_hugepage_flags = 0;
 503                 return 0;
 504         }
 505
 506         err = start_stop_khugepaged();
 507         if (err)
 508                 goto err_khugepaged;
 509
 510         return 0;
 511 err_khugepaged:
 512         unregister_shrinker(&deferred_split_shrinker);
 513 err_split_shrinker:
 514         unregister_shrinker(&huge_zero_page_shrinker);
 515 err_hzp_shrinker:
 516         khugepaged_destroy();
 517 err_slab:
 518         hugepage_exit_sysfs(hugepage_kobj);
 519 err_sysfs:
 520         return err;
 521 }
 522 subsys_initcall(hugepage_init);
 523
 524 static int __init setup_transparent_hugepage(char *str)
 525 {
 526         int ret = 0;
 527         if (!str)
 528                 goto out;
 529         if (!strcmp(str, "always")) {
 530                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 531                         &transparent_hugepage_flags);
 532                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 533                           &transparent_hugepage_flags);
 534                 ret = 1;
 535         } else if (!strcmp(str, "madvise")) {
 536                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 537                           &transparent_hugepage_flags);
 538                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 539                         &transparent_hugepage_flags);
 540                 ret = 1;
 541         } else if (!strcmp(str, "never")) {
 542                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 543                           &transparent_hugepage_flags);
 544                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 545                           &transparent_hugepage_flags);
 546                 ret = 1;
 547         }
 548 out:
 549         if (!ret)
 550                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 551         return ret;
 552 }
 553 __setup("transparent_hugepage=", setup_transparent_hugepage);
 554
 555 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 556 {
 557         if (likely(vma->vm_flags & VM_WRITE))
 558                 pmd = pmd_mkwrite(pmd);
 559         return pmd;
 560 }
 561
 562 #ifdef CONFIG_MEMCG
 563 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 564 {
 565         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 566         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 567
 568         if (memcg)
 569                 return &memcg->deferred_split_queue;
 570         else
 571                 return &pgdat->deferred_split_queue;
 572 }
 573 #else
 574 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 575 {
 576         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 577
 578         return &pgdat->deferred_split_queue;
 579 }
 580 #endif
 581
 582 void prep_transhuge_page(struct page *page)
 583 {
 584         /*
 585          * we use page->mapping and page->index in second tail page
 586          * as list_head: assuming THP order >= 2
 587          */
 588
 589         INIT_LIST_HEAD(page_deferred_list(page));
 590         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 591 }
 592
 593 static inline bool is_transparent_hugepage(struct page *page)
 594 {
 595         if (!PageCompound(page))
 596                 return false;
 597
 598         page = compound_head(page);
 599         return is_huge_zero_page(page) ||
 600                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 601 }
 602
 603 static unsigned long __thp_get_unmapped_area(struct file *filp,
 604                 unsigned long addr, unsigned long len,
 605                 loff_t off, unsigned long flags, unsigned long size)
 606 {
 607         loff_t off_end = off + len;
 608         loff_t off_align = round_up(off, size);
 609         unsigned long len_pad, ret;
 610
 611         if (off_end <= off_align || (off_end - off_align) < size)
 612                 return 0;
 613
 614         len_pad = len + size;
 615         if (len_pad < len || (off + len_pad) < off)
 616                 return 0;
 617
 618         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 619                                               off >> PAGE_SHIFT, flags);
 620
 621         /*
 622          * The failure might be due to length padding. The caller will retry
 623          * without the padding.
 624          */
 625         if (IS_ERR_VALUE(ret))
 626                 return 0;
 627
 628         /*
 629          * Do not try to align to THP boundary if allocation at the address
 630          * hint succeeds.
 631          */
 632         if (ret == addr)
 633                 return addr;
 634
 635         ret += (off - ret) & (size - 1);
 636         return ret;
 637 }
 638
 639 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 640                 unsigned long len, unsigned long pgoff, unsigned long flags)
 641 {
 642         unsigned long ret;
 643         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 644
 645         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 646         if (ret)
 647                 return ret;
 648
 649         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 650 }
 651 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 652
 653 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 654                         struct page *page, gfp_t gfp)
 655 {
 656         struct vm_area_struct *vma = vmf->vma;
 657         pgtable_t pgtable;
 658         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 659         vm_fault_t ret = 0;
 660
 661         VM_BUG_ON_PAGE(!PageCompound(page), page);
 662
 663         if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
 664                 put_page(page);
 665                 count_vm_event(THP_FAULT_FALLBACK);
 666                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 667                 return VM_FAULT_FALLBACK;
 668         }
 669         cgroup_throttle_swaprate(page, gfp);
 670
 671         pgtable = pte_alloc_one(vma->vm_mm);
 672         if (unlikely(!pgtable)) {
 673                 ret = VM_FAULT_OOM;
 674                 goto release;
 675         }
 676
 677         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 678         /*
 679          * The memory barrier inside __SetPageUptodate makes sure that
 680          * clear_huge_page writes become visible before the set_pmd_at()
 681          * write.
 682          */
 683         __SetPageUptodate(page);
 684
 685         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 686         if (unlikely(!pmd_none(*vmf->pmd))) {
 687                 goto unlock_release;
 688         } else {
 689                 pmd_t entry;
 690
 691                 ret = check_stable_address_space(vma->vm_mm);
 692                 if (ret)
 693                         goto unlock_release;
 694
 695                 /* Deliver the page fault to userland */
 696                 if (userfaultfd_missing(vma)) {
 697                         spin_unlock(vmf->ptl);
 698                         put_page(page);
 699                         pte_free(vma->vm_mm, pgtable);
 700                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 701                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 702                         return ret;
 703                 }
 704
 705                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 706                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 707                 page_add_new_anon_rmap(page, vma, haddr);
 708                 lru_cache_add_inactive_or_unevictable(page, vma);
 709                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 710                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 711                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 712                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 713                 mm_inc_nr_ptes(vma->vm_mm);
 714                 spin_unlock(vmf->ptl);
 715                 count_vm_event(THP_FAULT_ALLOC);
 716                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 717         }
 718
 719         return 0;
 720 unlock_release:
 721         spin_unlock(vmf->ptl);
 722 release:
 723         if (pgtable)
 724                 pte_free(vma->vm_mm, pgtable);
 725         put_page(page);
 726         return ret;
 727
 728 }
 729
 730 /*
 731  * always: directly stall for all thp allocations
 732  * defer: wake kswapd and fail if not immediately available
 733  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 734  *                fail if not immediately available
 735  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 736  *          available
 737  * never: never stall for any thp allocation
 738  */
 739 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 740 {
 741         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 742
 743         /* Always do synchronous compaction */
 744         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 745                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 746
 747         /* Kick kcompactd and fail quickly */
 748         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 749                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 750
 751         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 752         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 753                 return GFP_TRANSHUGE_LIGHT |
 754                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 755                                         __GFP_KSWAPD_RECLAIM);
 756
 757         /* Only do synchronous compaction if madvised */
 758         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 759                 return GFP_TRANSHUGE_LIGHT |
 760                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 761
 762         return GFP_TRANSHUGE_LIGHT;
 763 }
 764
 765 /* Caller must hold page table lock. */
 766 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 767                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 768                 struct page *zero_page)
 769 {
 770         pmd_t entry;
 771         if (!pmd_none(*pmd))
 772                 return;
 773         entry = mk_pmd(zero_page, vma->vm_page_prot);
 774         entry = pmd_mkhuge(entry);
 775         if (pgtable)
 776                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 777         set_pmd_at(mm, haddr, pmd, entry);
 778         mm_inc_nr_ptes(mm);
 779 }
 780
 781 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 782 {
 783         struct vm_area_struct *vma = vmf->vma;
 784         gfp_t gfp;
 785         struct folio *folio;
 786         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 787
 788         if (!transhuge_vma_suitable(vma, haddr))
 789                 return VM_FAULT_FALLBACK;
 790         if (unlikely(anon_vma_prepare(vma)))
 791                 return VM_FAULT_OOM;
 792         khugepaged_enter_vma(vma, vma->vm_flags);
 793
 794         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 795                         !mm_forbids_zeropage(vma->vm_mm) &&
 796                         transparent_hugepage_use_zero_page()) {
 797                 pgtable_t pgtable;
 798                 struct page *zero_page;
 799                 vm_fault_t ret;
 800                 pgtable = pte_alloc_one(vma->vm_mm);
 801                 if (unlikely(!pgtable))
 802                         return VM_FAULT_OOM;
 803                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 804                 if (unlikely(!zero_page)) {
 805                         pte_free(vma->vm_mm, pgtable);
 806                         count_vm_event(THP_FAULT_FALLBACK);
 807                         return VM_FAULT_FALLBACK;
 808                 }
 809                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 810                 ret = 0;
 811                 if (pmd_none(*vmf->pmd)) {
 812                         ret = check_stable_address_space(vma->vm_mm);
 813                         if (ret) {
 814                                 spin_unlock(vmf->ptl);
 815                                 pte_free(vma->vm_mm, pgtable);
 816                         } else if (userfaultfd_missing(vma)) {
 817                                 spin_unlock(vmf->ptl);
 818                                 pte_free(vma->vm_mm, pgtable);
 819                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 820                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 821                         } else {
 822                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 823                                                    haddr, vmf->pmd, zero_page);
 824                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 825                                 spin_unlock(vmf->ptl);
 826                         }
 827                 } else {
 828                         spin_unlock(vmf->ptl);
 829                         pte_free(vma->vm_mm, pgtable);
 830                 }
 831                 return ret;
 832         }
 833         gfp = vma_thp_gfp_mask(vma);
 834         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
 835         if (unlikely(!folio)) {
 836                 count_vm_event(THP_FAULT_FALLBACK);
 837                 return VM_FAULT_FALLBACK;
 838         }
 839         return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
 840 }
 841
 842 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 843                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 844                 pgtable_t pgtable)
 845 {
 846         struct mm_struct *mm = vma->vm_mm;
 847         pmd_t entry;
 848         spinlock_t *ptl;
 849
 850         ptl = pmd_lock(mm, pmd);
 851         if (!pmd_none(*pmd)) {
 852                 if (write) {
 853                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 854                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 855                                 goto out_unlock;
 856                         }
 857                         entry = pmd_mkyoung(*pmd);
 858                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 859                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 860                                 update_mmu_cache_pmd(vma, addr, pmd);
 861                 }
 862
 863                 goto out_unlock;
 864         }
 865
 866         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 867         if (pfn_t_devmap(pfn))
 868                 entry = pmd_mkdevmap(entry);
 869         if (write) {
 870                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 871                 entry = maybe_pmd_mkwrite(entry, vma);
 872         }
 873
 874         if (pgtable) {
 875                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 876                 mm_inc_nr_ptes(mm);
 877                 pgtable = NULL;
 878         }
 879
 880         set_pmd_at(mm, addr, pmd, entry);
 881         update_mmu_cache_pmd(vma, addr, pmd);
 882
 883 out_unlock:
 884         spin_unlock(ptl);
 885         if (pgtable)
 886                 pte_free(mm, pgtable);
 887 }
 888
 889 /**
 890  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 891  * @vmf: Structure describing the fault
 892  * @pfn: pfn to insert
 893  * @pgprot: page protection to use
 894  * @write: whether it's a write fault
 895  *
 896  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 897  * also consult the vmf_insert_mixed_prot() documentation when
 898  * @pgprot != @vmf->vma->vm_page_prot.
 899  *
 900  * Return: vm_fault_t value.
 901  */
 902 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 903                                    pgprot_t pgprot, bool write)
 904 {
 905         unsigned long addr = vmf->address & PMD_MASK;
 906         struct vm_area_struct *vma = vmf->vma;
 907         pgtable_t pgtable = NULL;
 908
 909         /*
 910          * If we had pmd_special, we could avoid all these restrictions,
 911          * but we need to be consistent with PTEs and architectures that
 912          * can't support a 'special' bit.
 913          */
 914         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 915                         !pfn_t_devmap(pfn));
 916         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 917                                                 (VM_PFNMAP|VM_MIXEDMAP));
 918         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 919
 920         if (addr < vma->vm_start || addr >= vma->vm_end)
 921                 return VM_FAULT_SIGBUS;
 922
 923         if (arch_needs_pgtable_deposit()) {
 924                 pgtable = pte_alloc_one(vma->vm_mm);
 925                 if (!pgtable)
 926                         return VM_FAULT_OOM;
 927         }
 928
 929         track_pfn_insert(vma, &pgprot, pfn);
 930
 931         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 932         return VM_FAULT_NOPAGE;
 933 }
 934 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 935
 936 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 937 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 938 {
 939         if (likely(vma->vm_flags & VM_WRITE))
 940                 pud = pud_mkwrite(pud);
 941         return pud;
 942 }
 943
 944 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 945                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 946 {
 947         struct mm_struct *mm = vma->vm_mm;
 948         pud_t entry;
 949         spinlock_t *ptl;
 950
 951         ptl = pud_lock(mm, pud);
 952         if (!pud_none(*pud)) {
 953                 if (write) {
 954                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 955                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 956                                 goto out_unlock;
 957                         }
 958                         entry = pud_mkyoung(*pud);
 959                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 960                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 961                                 update_mmu_cache_pud(vma, addr, pud);
 962                 }
 963                 goto out_unlock;
 964         }
 965
 966         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 967         if (pfn_t_devmap(pfn))
 968                 entry = pud_mkdevmap(entry);
 969         if (write) {
 970                 entry = pud_mkyoung(pud_mkdirty(entry));
 971                 entry = maybe_pud_mkwrite(entry, vma);
 972         }
 973         set_pud_at(mm, addr, pud, entry);
 974         update_mmu_cache_pud(vma, addr, pud);
 975
 976 out_unlock:
 977         spin_unlock(ptl);
 978 }
 979
 980 /**
 981  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 982  * @vmf: Structure describing the fault
 983  * @pfn: pfn to insert
 984  * @pgprot: page protection to use
 985  * @write: whether it's a write fault
 986  *
 987  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 988  * also consult the vmf_insert_mixed_prot() documentation when
 989  * @pgprot != @vmf->vma->vm_page_prot.
 990  *
 991  * Return: vm_fault_t value.
 992  */
 993 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 994                                    pgprot_t pgprot, bool write)
 995 {
 996         unsigned long addr = vmf->address & PUD_MASK;
 997         struct vm_area_struct *vma = vmf->vma;
 998
 999         /*
1000          * If we had pud_special, we could avoid all these restrictions,
1001          * but we need to be consistent with PTEs and architectures that
1002          * can't support a 'special' bit.
1003          */
1004         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1005                         !pfn_t_devmap(pfn));
1006         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1007                                                 (VM_PFNMAP|VM_MIXEDMAP));
1008         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1009
1010         if (addr < vma->vm_start || addr >= vma->vm_end)
1011                 return VM_FAULT_SIGBUS;
1012
1013         track_pfn_insert(vma, &pgprot, pfn);
1014
1015         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1016         return VM_FAULT_NOPAGE;
1017 }
1018 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1019 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1020
1021 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1022                       pmd_t *pmd, bool write)
1023 {
1024         pmd_t _pmd;
1025
1026         _pmd = pmd_mkyoung(*pmd);
1027         if (write)
1028                 _pmd = pmd_mkdirty(_pmd);
1029         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1030                                   pmd, _pmd, write))
1031                 update_mmu_cache_pmd(vma, addr, pmd);
1032 }
1033
1034 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1035                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1036 {
1037         unsigned long pfn = pmd_pfn(*pmd);
1038         struct mm_struct *mm = vma->vm_mm;
1039         struct page *page;
1040
1041         assert_spin_locked(pmd_lockptr(mm, pmd));
1042
1043         /*
1044          * When we COW a devmap PMD entry, we split it into PTEs, so we should
1045          * not be in this function with `flags & FOLL_COW` set.
1046          */
1047         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
1048
1049         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1050         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1051                          (FOLL_PIN | FOLL_GET)))
1052                 return NULL;
1053
1054         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1055                 return NULL;
1056
1057         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1058                 /* pass */;
1059         else
1060                 return NULL;
1061
1062         if (flags & FOLL_TOUCH)
1063                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1064
1065         /*
1066          * device mapped pages can only be returned if the
1067          * caller will manage the page reference count.
1068          */
1069         if (!(flags & (FOLL_GET | FOLL_PIN)))
1070                 return ERR_PTR(-EEXIST);
1071
1072         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1073         *pgmap = get_dev_pagemap(pfn, *pgmap);
1074         if (!*pgmap)
1075                 return ERR_PTR(-EFAULT);
1076         page = pfn_to_page(pfn);
1077         if (!try_grab_page(page, flags))
1078                 page = ERR_PTR(-ENOMEM);
1079
1080         return page;
1081 }
1082
1083 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1084                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1085                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1086 {
1087         spinlock_t *dst_ptl, *src_ptl;
1088         struct page *src_page;
1089         pmd_t pmd;
1090         pgtable_t pgtable = NULL;
1091         int ret = -ENOMEM;
1092
1093         /* Skip if can be re-fill on fault */
1094         if (!vma_is_anonymous(dst_vma))
1095                 return 0;
1096
1097         pgtable = pte_alloc_one(dst_mm);
1098         if (unlikely(!pgtable))
1099                 goto out;
1100
1101         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1102         src_ptl = pmd_lockptr(src_mm, src_pmd);
1103         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1104
1105         ret = -EAGAIN;
1106         pmd = *src_pmd;
1107
1108 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1109         if (unlikely(is_swap_pmd(pmd))) {
1110                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1111
1112                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1113                 if (!is_readable_migration_entry(entry)) {
1114                         entry = make_readable_migration_entry(
1115                                                         swp_offset(entry));
1116                         pmd = swp_entry_to_pmd(entry);
1117                         if (pmd_swp_soft_dirty(*src_pmd))
1118                                 pmd = pmd_swp_mksoft_dirty(pmd);
1119                         if (pmd_swp_uffd_wp(*src_pmd))
1120                                 pmd = pmd_swp_mkuffd_wp(pmd);
1121                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1122                 }
1123                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1124                 mm_inc_nr_ptes(dst_mm);
1125                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1126                 if (!userfaultfd_wp(dst_vma))
1127                         pmd = pmd_swp_clear_uffd_wp(pmd);
1128                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1129                 ret = 0;
1130                 goto out_unlock;
1131         }
1132 #endif
1133
1134         if (unlikely(!pmd_trans_huge(pmd))) {
1135                 pte_free(dst_mm, pgtable);
1136                 goto out_unlock;
1137         }
1138         /*
1139          * When page table lock is held, the huge zero pmd should not be
1140          * under splitting since we don't split the page itself, only pmd to
1141          * a page table.
1142          */
1143         if (is_huge_zero_pmd(pmd)) {
1144                 /*
1145                  * get_huge_zero_page() will never allocate a new page here,
1146                  * since we already have a zero page to copy. It just takes a
1147                  * reference.
1148                  */
1149                 mm_get_huge_zero_page(dst_mm);
1150                 goto out_zero_page;
1151         }
1152
1153         src_page = pmd_page(pmd);
1154         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1155
1156         get_page(src_page);
1157         if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1158                 /* Page maybe pinned: split and retry the fault on PTEs. */
1159                 put_page(src_page);
1160                 pte_free(dst_mm, pgtable);
1161                 spin_unlock(src_ptl);
1162                 spin_unlock(dst_ptl);
1163                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1164                 return -EAGAIN;
1165         }
1166         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1167 out_zero_page:
1168         mm_inc_nr_ptes(dst_mm);
1169         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1170         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1171         if (!userfaultfd_wp(dst_vma))
1172                 pmd = pmd_clear_uffd_wp(pmd);
1173         pmd = pmd_mkold(pmd_wrprotect(pmd));
1174         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1175
1176         ret = 0;
1177 out_unlock:
1178         spin_unlock(src_ptl);
1179         spin_unlock(dst_ptl);
1180 out:
1181         return ret;
1182 }
1183
1184 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1185 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1186                       pud_t *pud, bool write)
1187 {
1188         pud_t _pud;
1189
1190         _pud = pud_mkyoung(*pud);
1191         if (write)
1192                 _pud = pud_mkdirty(_pud);
1193         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1194                                   pud, _pud, write))
1195                 update_mmu_cache_pud(vma, addr, pud);
1196 }
1197
1198 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1199                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1200 {
1201         unsigned long pfn = pud_pfn(*pud);
1202         struct mm_struct *mm = vma->vm_mm;
1203         struct page *page;
1204
1205         assert_spin_locked(pud_lockptr(mm, pud));
1206
1207         if (flags & FOLL_WRITE && !pud_write(*pud))
1208                 return NULL;
1209
1210         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1211         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1212                          (FOLL_PIN | FOLL_GET)))
1213                 return NULL;
1214
1215         if (pud_present(*pud) && pud_devmap(*pud))
1216                 /* pass */;
1217         else
1218                 return NULL;
1219
1220         if (flags & FOLL_TOUCH)
1221                 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1222
1223         /*
1224          * device mapped pages can only be returned if the
1225          * caller will manage the page reference count.
1226          *
1227          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1228          */
1229         if (!(flags & (FOLL_GET | FOLL_PIN)))
1230                 return ERR_PTR(-EEXIST);
1231
1232         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1233         *pgmap = get_dev_pagemap(pfn, *pgmap);
1234         if (!*pgmap)
1235                 return ERR_PTR(-EFAULT);
1236         page = pfn_to_page(pfn);
1237         if (!try_grab_page(page, flags))
1238                 page = ERR_PTR(-ENOMEM);
1239
1240         return page;
1241 }
1242
1243 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1244                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1245                   struct vm_area_struct *vma)
1246 {
1247         spinlock_t *dst_ptl, *src_ptl;
1248         pud_t pud;
1249         int ret;
1250
1251         dst_ptl = pud_lock(dst_mm, dst_pud);
1252         src_ptl = pud_lockptr(src_mm, src_pud);
1253         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1254
1255         ret = -EAGAIN;
1256         pud = *src_pud;
1257         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1258                 goto out_unlock;
1259
1260         /*
1261          * When page table lock is held, the huge zero pud should not be
1262          * under splitting since we don't split the page itself, only pud to
1263          * a page table.
1264          */
1265         if (is_huge_zero_pud(pud)) {
1266                 /* No huge zero pud yet */
1267         }
1268
1269         /*
1270          * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1271          * and split if duplicating fails.
1272          */
1273         pudp_set_wrprotect(src_mm, addr, src_pud);
1274         pud = pud_mkold(pud_wrprotect(pud));
1275         set_pud_at(dst_mm, addr, dst_pud, pud);
1276
1277         ret = 0;
1278 out_unlock:
1279         spin_unlock(src_ptl);
1280         spin_unlock(dst_ptl);
1281         return ret;
1282 }
1283
1284 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1285 {
1286         bool write = vmf->flags & FAULT_FLAG_WRITE;
1287
1288         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1289         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1290                 goto unlock;
1291
1292         touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1293 unlock:
1294         spin_unlock(vmf->ptl);
1295 }
1296 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1297
1298 void huge_pmd_set_accessed(struct vm_fault *vmf)
1299 {
1300         bool write = vmf->flags & FAULT_FLAG_WRITE;
1301
1302         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1303         if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1304                 goto unlock;
1305
1306         touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1307
1308 unlock:
1309         spin_unlock(vmf->ptl);
1310 }
1311
1312 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1313 {
1314         const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1315         struct vm_area_struct *vma = vmf->vma;
1316         struct page *page;
1317         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1318         pmd_t orig_pmd = vmf->orig_pmd;
1319
1320         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1321         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1322
1323         VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE));
1324         VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE));
1325
1326         if (is_huge_zero_pmd(orig_pmd))
1327                 goto fallback;
1328
1329         spin_lock(vmf->ptl);
1330
1331         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1332                 spin_unlock(vmf->ptl);
1333                 return 0;
1334         }
1335
1336         page = pmd_page(orig_pmd);
1337         VM_BUG_ON_PAGE(!PageHead(page), page);
1338
1339         /* Early check when only holding the PT lock. */
1340         if (PageAnonExclusive(page))
1341                 goto reuse;
1342
1343         if (!trylock_page(page)) {
1344                 get_page(page);
1345                 spin_unlock(vmf->ptl);
1346                 lock_page(page);
1347                 spin_lock(vmf->ptl);
1348                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1349                         spin_unlock(vmf->ptl);
1350                         unlock_page(page);
1351                         put_page(page);
1352                         return 0;
1353                 }
1354                 put_page(page);
1355         }
1356
1357         /* Recheck after temporarily dropping the PT lock. */
1358         if (PageAnonExclusive(page)) {
1359                 unlock_page(page);
1360                 goto reuse;
1361         }
1362
1363         /*
1364          * See do_wp_page(): we can only reuse the page exclusively if there are
1365          * no additional references. Note that we always drain the LRU
1366          * pagevecs immediately after adding a THP.
1367          */
1368         if (page_count(page) > 1 + PageSwapCache(page) * thp_nr_pages(page))
1369                 goto unlock_fallback;
1370         if (PageSwapCache(page))
1371                 try_to_free_swap(page);
1372         if (page_count(page) == 1) {
1373                 pmd_t entry;
1374
1375                 page_move_anon_rmap(page, vma);
1376                 unlock_page(page);
1377 reuse:
1378                 if (unlikely(unshare)) {
1379                         spin_unlock(vmf->ptl);
1380                         return 0;
1381                 }
1382                 entry = pmd_mkyoung(orig_pmd);
1383                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1384                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1385                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1386                 spin_unlock(vmf->ptl);
1387                 return VM_FAULT_WRITE;
1388         }
1389
1390 unlock_fallback:
1391         unlock_page(page);
1392         spin_unlock(vmf->ptl);
1393 fallback:
1394         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1395         return VM_FAULT_FALLBACK;
1396 }
1397
1398 /*
1399  * FOLL_FORCE can write to even unwritable pmd's, but only
1400  * after we've gone through a COW cycle and they are dirty.
1401  */
1402 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1403 {
1404         return pmd_write(pmd) ||
1405                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1406 }
1407
1408 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1409                                    unsigned long addr,
1410                                    pmd_t *pmd,
1411                                    unsigned int flags)
1412 {
1413         struct mm_struct *mm = vma->vm_mm;
1414         struct page *page = NULL;
1415
1416         assert_spin_locked(pmd_lockptr(mm, pmd));
1417
1418         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1419                 goto out;
1420
1421         /* Avoid dumping huge zero page */
1422         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1423                 return ERR_PTR(-EFAULT);
1424
1425         /* Full NUMA hinting faults to serialise migration in fault paths */
1426         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1427                 goto out;
1428
1429         page = pmd_page(*pmd);
1430         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1431
1432         if (!pmd_write(*pmd) && gup_must_unshare(flags, page))
1433                 return ERR_PTR(-EMLINK);
1434
1435         VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1436                         !PageAnonExclusive(page), page);
1437
1438         if (!try_grab_page(page, flags))
1439                 return ERR_PTR(-ENOMEM);
1440
1441         if (flags & FOLL_TOUCH)
1442                 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1443
1444         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1445         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1446
1447 out:
1448         return page;
1449 }
1450
1451 /* NUMA hinting page fault entry point for trans huge pmds */
1452 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1453 {
1454         struct vm_area_struct *vma = vmf->vma;
1455         pmd_t oldpmd = vmf->orig_pmd;
1456         pmd_t pmd;
1457         struct page *page;
1458         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1459         int page_nid = NUMA_NO_NODE;
1460         int target_nid, last_cpupid = -1;
1461         bool migrated = false;
1462         bool was_writable = pmd_savedwrite(oldpmd);
1463         int flags = 0;
1464
1465         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1466         if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1467                 spin_unlock(vmf->ptl);
1468                 goto out;
1469         }
1470
1471         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1472         page = vm_normal_page_pmd(vma, haddr, pmd);
1473         if (!page)
1474                 goto out_map;
1475
1476         /* See similar comment in do_numa_page for explanation */
1477         if (!was_writable)
1478                 flags |= TNF_NO_GROUP;
1479
1480         page_nid = page_to_nid(page);
1481         last_cpupid = page_cpupid_last(page);
1482         target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1483                                        &flags);
1484
1485         if (target_nid == NUMA_NO_NODE) {
1486                 put_page(page);
1487                 goto out_map;
1488         }
1489
1490         spin_unlock(vmf->ptl);
1491
1492         migrated = migrate_misplaced_page(page, vma, target_nid);
1493         if (migrated) {
1494                 flags |= TNF_MIGRATED;
1495                 page_nid = target_nid;
1496         } else {
1497                 flags |= TNF_MIGRATE_FAIL;
1498                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1499                 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1500                         spin_unlock(vmf->ptl);
1501                         goto out;
1502                 }
1503                 goto out_map;
1504         }
1505
1506 out:
1507         if (page_nid != NUMA_NO_NODE)
1508                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1509                                 flags);
1510
1511         return 0;
1512
1513 out_map:
1514         /* Restore the PMD */
1515         pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1516         pmd = pmd_mkyoung(pmd);
1517         if (was_writable)
1518                 pmd = pmd_mkwrite(pmd);
1519         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1520         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1521         spin_unlock(vmf->ptl);
1522         goto out;
1523 }
1524
1525 /*
1526  * Return true if we do MADV_FREE successfully on entire pmd page.
1527  * Otherwise, return false.
1528  */
1529 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1530                 pmd_t *pmd, unsigned long addr, unsigned long next)
1531 {
1532         spinlock_t *ptl;
1533         pmd_t orig_pmd;
1534         struct page *page;
1535         struct mm_struct *mm = tlb->mm;
1536         bool ret = false;
1537
1538         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1539
1540         ptl = pmd_trans_huge_lock(pmd, vma);
1541         if (!ptl)
1542                 goto out_unlocked;
1543
1544         orig_pmd = *pmd;
1545         if (is_huge_zero_pmd(orig_pmd))
1546                 goto out;
1547
1548         if (unlikely(!pmd_present(orig_pmd))) {
1549                 VM_BUG_ON(thp_migration_supported() &&
1550                                   !is_pmd_migration_entry(orig_pmd));
1551                 goto out;
1552         }
1553
1554         page = pmd_page(orig_pmd);
1555         /*
1556          * If other processes are mapping this page, we couldn't discard
1557          * the page unless they all do MADV_FREE so let's skip the page.
1558          */
1559         if (total_mapcount(page) != 1)
1560                 goto out;
1561
1562         if (!trylock_page(page))
1563                 goto out;
1564
1565         /*
1566          * If user want to discard part-pages of THP, split it so MADV_FREE
1567          * will deactivate only them.
1568          */
1569         if (next - addr != HPAGE_PMD_SIZE) {
1570                 get_page(page);
1571                 spin_unlock(ptl);
1572                 split_huge_page(page);
1573                 unlock_page(page);
1574                 put_page(page);
1575                 goto out_unlocked;
1576         }
1577
1578         if (PageDirty(page))
1579                 ClearPageDirty(page);
1580         unlock_page(page);
1581
1582         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1583                 pmdp_invalidate(vma, addr, pmd);
1584                 orig_pmd = pmd_mkold(orig_pmd);
1585                 orig_pmd = pmd_mkclean(orig_pmd);
1586
1587                 set_pmd_at(mm, addr, pmd, orig_pmd);
1588                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1589         }
1590
1591         mark_page_lazyfree(page);
1592         ret = true;
1593 out:
1594         spin_unlock(ptl);
1595 out_unlocked:
1596         return ret;
1597 }
1598
1599 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1600 {
1601         pgtable_t pgtable;
1602
1603         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1604         pte_free(mm, pgtable);
1605         mm_dec_nr_ptes(mm);
1606 }
1607
1608 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1609                  pmd_t *pmd, unsigned long addr)
1610 {
1611         pmd_t orig_pmd;
1612         spinlock_t *ptl;
1613
1614         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1615
1616         ptl = __pmd_trans_huge_lock(pmd, vma);
1617         if (!ptl)
1618                 return 0;
1619         /*
1620          * For architectures like ppc64 we look at deposited pgtable
1621          * when calling pmdp_huge_get_and_clear. So do the
1622          * pgtable_trans_huge_withdraw after finishing pmdp related
1623          * operations.
1624          */
1625         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1626                                                 tlb->fullmm);
1627         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1628         if (vma_is_special_huge(vma)) {
1629                 if (arch_needs_pgtable_deposit())
1630                         zap_deposited_table(tlb->mm, pmd);
1631                 spin_unlock(ptl);
1632         } else if (is_huge_zero_pmd(orig_pmd)) {
1633                 zap_deposited_table(tlb->mm, pmd);
1634                 spin_unlock(ptl);
1635         } else {
1636                 struct page *page = NULL;
1637                 int flush_needed = 1;
1638
1639                 if (pmd_present(orig_pmd)) {
1640                         page = pmd_page(orig_pmd);
1641                         page_remove_rmap(page, vma, true);
1642                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1643                         VM_BUG_ON_PAGE(!PageHead(page), page);
1644                 } else if (thp_migration_supported()) {
1645                         swp_entry_t entry;
1646
1647                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1648                         entry = pmd_to_swp_entry(orig_pmd);
1649                         page = pfn_swap_entry_to_page(entry);
1650                         flush_needed = 0;
1651                 } else
1652                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1653
1654                 if (PageAnon(page)) {
1655                         zap_deposited_table(tlb->mm, pmd);
1656                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1657                 } else {
1658                         if (arch_needs_pgtable_deposit())
1659                                 zap_deposited_table(tlb->mm, pmd);
1660                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1661                 }
1662
1663                 spin_unlock(ptl);
1664                 if (flush_needed)
1665                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1666         }
1667         return 1;
1668 }
1669
1670 #ifndef pmd_move_must_withdraw
1671 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1672                                          spinlock_t *old_pmd_ptl,
1673                                          struct vm_area_struct *vma)
1674 {
1675         /*
1676          * With split pmd lock we also need to move preallocated
1677          * PTE page table if new_pmd is on different PMD page table.
1678          *
1679          * We also don't deposit and withdraw tables for file pages.
1680          */
1681         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1682 }
1683 #endif
1684
1685 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1686 {
1687 #ifdef CONFIG_MEM_SOFT_DIRTY
1688         if (unlikely(is_pmd_migration_entry(pmd)))
1689                 pmd = pmd_swp_mksoft_dirty(pmd);
1690         else if (pmd_present(pmd))
1691                 pmd = pmd_mksoft_dirty(pmd);
1692 #endif
1693         return pmd;
1694 }
1695
1696 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1697                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1698 {
1699         spinlock_t *old_ptl, *new_ptl;
1700         pmd_t pmd;
1701         struct mm_struct *mm = vma->vm_mm;
1702         bool force_flush = false;
1703
1704         /*
1705          * The destination pmd shouldn't be established, free_pgtables()
1706          * should have release it.
1707          */
1708         if (WARN_ON(!pmd_none(*new_pmd))) {
1709                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1710                 return false;
1711         }
1712
1713         /*
1714          * We don't have to worry about the ordering of src and dst
1715          * ptlocks because exclusive mmap_lock prevents deadlock.
1716          */
1717         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1718         if (old_ptl) {
1719                 new_ptl = pmd_lockptr(mm, new_pmd);
1720                 if (new_ptl != old_ptl)
1721                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1722                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1723                 if (pmd_present(pmd))
1724                         force_flush = true;
1725                 VM_BUG_ON(!pmd_none(*new_pmd));
1726
1727                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1728                         pgtable_t pgtable;
1729                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1730                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1731                 }
1732                 pmd = move_soft_dirty_pmd(pmd);
1733                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1734                 if (force_flush)
1735                         flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1736                 if (new_ptl != old_ptl)
1737                         spin_unlock(new_ptl);
1738                 spin_unlock(old_ptl);
1739                 return true;
1740         }
1741         return false;
1742 }
1743
1744 /*
1745  * Returns
1746  *  - 0 if PMD could not be locked
1747  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1748  *      or if prot_numa but THP migration is not supported
1749  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1750  */
1751 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1752                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1753                     unsigned long cp_flags)
1754 {
1755         struct mm_struct *mm = vma->vm_mm;
1756         spinlock_t *ptl;
1757         pmd_t oldpmd, entry;
1758         bool preserve_write;
1759         int ret;
1760         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1761         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1762         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1763
1764         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1765
1766         if (prot_numa && !thp_migration_supported())
1767                 return 1;
1768
1769         ptl = __pmd_trans_huge_lock(pmd, vma);
1770         if (!ptl)
1771                 return 0;
1772
1773         preserve_write = prot_numa && pmd_write(*pmd);
1774         ret = 1;
1775
1776 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1777         if (is_swap_pmd(*pmd)) {
1778                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1779                 struct page *page = pfn_swap_entry_to_page(entry);
1780
1781                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1782                 if (is_writable_migration_entry(entry)) {
1783                         pmd_t newpmd;
1784                         /*
1785                          * A protection check is difficult so
1786                          * just be safe and disable write
1787                          */
1788                         if (PageAnon(page))
1789                                 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1790                         else
1791                                 entry = make_readable_migration_entry(swp_offset(entry));
1792                         newpmd = swp_entry_to_pmd(entry);
1793                         if (pmd_swp_soft_dirty(*pmd))
1794                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1795                         if (pmd_swp_uffd_wp(*pmd))
1796                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1797                         set_pmd_at(mm, addr, pmd, newpmd);
1798                 }
1799                 goto unlock;
1800         }
1801 #endif
1802
1803         if (prot_numa) {
1804                 struct page *page;
1805                 /*
1806                  * Avoid trapping faults against the zero page. The read-only
1807                  * data is likely to be read-cached on the local CPU and
1808                  * local/remote hits to the zero page are not interesting.
1809                  */
1810                 if (is_huge_zero_pmd(*pmd))
1811                         goto unlock;
1812
1813                 if (pmd_protnone(*pmd))
1814                         goto unlock;
1815
1816                 page = pmd_page(*pmd);
1817                 /*
1818                  * Skip scanning top tier node if normal numa
1819                  * balancing is disabled
1820                  */
1821                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1822                     node_is_toptier(page_to_nid(page)))
1823                         goto unlock;
1824         }
1825         /*
1826          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1827          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1828          * which is also under mmap_read_lock(mm):
1829          *
1830          *      CPU0:                           CPU1:
1831          *                              change_huge_pmd(prot_numa=1)
1832          *                               pmdp_huge_get_and_clear_notify()
1833          * madvise_dontneed()
1834          *  zap_pmd_range()
1835          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1836          *   // skip the pmd
1837          *                               set_pmd_at();
1838          *                               // pmd is re-established
1839          *
1840          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1841          * which may break userspace.
1842          *
1843          * pmdp_invalidate_ad() is required to make sure we don't miss
1844          * dirty/young flags set by hardware.
1845          */
1846         oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1847
1848         entry = pmd_modify(oldpmd, newprot);
1849         if (preserve_write)
1850                 entry = pmd_mk_savedwrite(entry);
1851         if (uffd_wp) {
1852                 entry = pmd_wrprotect(entry);
1853                 entry = pmd_mkuffd_wp(entry);
1854         } else if (uffd_wp_resolve) {
1855                 /*
1856                  * Leave the write bit to be handled by PF interrupt
1857                  * handler, then things like COW could be properly
1858                  * handled.
1859                  */
1860                 entry = pmd_clear_uffd_wp(entry);
1861         }
1862         ret = HPAGE_PMD_NR;
1863         set_pmd_at(mm, addr, pmd, entry);
1864
1865         if (huge_pmd_needs_flush(oldpmd, entry))
1866                 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1867
1868         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1869 unlock:
1870         spin_unlock(ptl);
1871         return ret;
1872 }
1873
1874 /*
1875  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1876  *
1877  * Note that if it returns page table lock pointer, this routine returns without
1878  * unlocking page table lock. So callers must unlock it.
1879  */
1880 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1881 {
1882         spinlock_t *ptl;
1883         ptl = pmd_lock(vma->vm_mm, pmd);
1884         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1885                         pmd_devmap(*pmd)))
1886                 return ptl;
1887         spin_unlock(ptl);
1888         return NULL;
1889 }
1890
1891 /*
1892  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1893  *
1894  * Note that if it returns page table lock pointer, this routine returns without
1895  * unlocking page table lock. So callers must unlock it.
1896  */
1897 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1898 {
1899         spinlock_t *ptl;
1900
1901         ptl = pud_lock(vma->vm_mm, pud);
1902         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1903                 return ptl;
1904         spin_unlock(ptl);
1905         return NULL;
1906 }
1907
1908 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1909 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1910                  pud_t *pud, unsigned long addr)
1911 {
1912         spinlock_t *ptl;
1913
1914         ptl = __pud_trans_huge_lock(pud, vma);
1915         if (!ptl)
1916                 return 0;
1917
1918         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1919         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1920         if (vma_is_special_huge(vma)) {
1921                 spin_unlock(ptl);
1922                 /* No zero page support yet */
1923         } else {
1924                 /* No support for anonymous PUD pages yet */
1925                 BUG();
1926         }
1927         return 1;
1928 }
1929
1930 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1931                 unsigned long haddr)
1932 {
1933         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1934         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1935         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1936         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1937
1938         count_vm_event(THP_SPLIT_PUD);
1939
1940         pudp_huge_clear_flush_notify(vma, haddr, pud);
1941 }
1942
1943 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1944                 unsigned long address)
1945 {
1946         spinlock_t *ptl;
1947         struct mmu_notifier_range range;
1948
1949         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1950                                 address & HPAGE_PUD_MASK,
1951                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1952         mmu_notifier_invalidate_range_start(&range);
1953         ptl = pud_lock(vma->vm_mm, pud);
1954         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1955                 goto out;
1956         __split_huge_pud_locked(vma, pud, range.start);
1957
1958 out:
1959         spin_unlock(ptl);
1960         /*
1961          * No need to double call mmu_notifier->invalidate_range() callback as
1962          * the above pudp_huge_clear_flush_notify() did already call it.
1963          */
1964         mmu_notifier_invalidate_range_only_end(&range);
1965 }
1966 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1967
1968 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1969                 unsigned long haddr, pmd_t *pmd)
1970 {
1971         struct mm_struct *mm = vma->vm_mm;
1972         pgtable_t pgtable;
1973         pmd_t _pmd;
1974         int i;
1975
1976         /*
1977          * Leave pmd empty until pte is filled note that it is fine to delay
1978          * notification until mmu_notifier_invalidate_range_end() as we are
1979          * replacing a zero pmd write protected page with a zero pte write
1980          * protected page.
1981          *
1982          * See Documentation/mm/mmu_notifier.rst
1983          */
1984         pmdp_huge_clear_flush(vma, haddr, pmd);
1985
1986         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1987         pmd_populate(mm, &_pmd, pgtable);
1988
1989         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1990                 pte_t *pte, entry;
1991                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1992                 entry = pte_mkspecial(entry);
1993                 pte = pte_offset_map(&_pmd, haddr);
1994                 VM_BUG_ON(!pte_none(*pte));
1995                 set_pte_at(mm, haddr, pte, entry);
1996                 pte_unmap(pte);
1997         }
1998         smp_wmb(); /* make pte visible before pmd */
1999         pmd_populate(mm, pmd, pgtable);
2000 }
2001
2002 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2003                 unsigned long haddr, bool freeze)
2004 {
2005         struct mm_struct *mm = vma->vm_mm;
2006         struct page *page;
2007         pgtable_t pgtable;
2008         pmd_t old_pmd, _pmd;
2009         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2010         bool anon_exclusive = false;
2011         unsigned long addr;
2012         int i;
2013
2014         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2015         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2016         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2017         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2018                                 && !pmd_devmap(*pmd));
2019
2020         count_vm_event(THP_SPLIT_PMD);
2021
2022         if (!vma_is_anonymous(vma)) {
2023                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2024                 /*
2025                  * We are going to unmap this huge page. So
2026                  * just go ahead and zap it
2027                  */
2028                 if (arch_needs_pgtable_deposit())
2029                         zap_deposited_table(mm, pmd);
2030                 if (vma_is_special_huge(vma))
2031                         return;
2032                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2033                         swp_entry_t entry;
2034
2035                         entry = pmd_to_swp_entry(old_pmd);
2036                         page = pfn_swap_entry_to_page(entry);
2037                 } else {
2038                         page = pmd_page(old_pmd);
2039                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2040                                 set_page_dirty(page);
2041                         if (!PageReferenced(page) && pmd_young(old_pmd))
2042                                 SetPageReferenced(page);
2043                         page_remove_rmap(page, vma, true);
2044                         put_page(page);
2045                 }
2046                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2047                 return;
2048         }
2049
2050         if (is_huge_zero_pmd(*pmd)) {
2051                 /*
2052                  * FIXME: Do we want to invalidate secondary mmu by calling
2053                  * mmu_notifier_invalidate_range() see comments below inside
2054                  * __split_huge_pmd() ?
2055                  *
2056                  * We are going from a zero huge page write protected to zero
2057                  * small page also write protected so it does not seems useful
2058                  * to invalidate secondary mmu at this time.
2059                  */
2060                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2061         }
2062
2063         /*
2064          * Up to this point the pmd is present and huge and userland has the
2065          * whole access to the hugepage during the split (which happens in
2066          * place). If we overwrite the pmd with the not-huge version pointing
2067          * to the pte here (which of course we could if all CPUs were bug
2068          * free), userland could trigger a small page size TLB miss on the
2069          * small sized TLB while the hugepage TLB entry is still established in
2070          * the huge TLB. Some CPU doesn't like that.
2071          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2072          * 383 on page 105. Intel should be safe but is also warns that it's
2073          * only safe if the permission and cache attributes of the two entries
2074          * loaded in the two TLB is identical (which should be the case here).
2075          * But it is generally safer to never allow small and huge TLB entries
2076          * for the same virtual address to be loaded simultaneously. So instead
2077          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2078          * current pmd notpresent (atomically because here the pmd_trans_huge
2079          * must remain set at all times on the pmd until the split is complete
2080          * for this pmd), then we flush the SMP TLB and finally we write the
2081          * non-huge version of the pmd entry with pmd_populate.
2082          */
2083         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2084
2085         pmd_migration = is_pmd_migration_entry(old_pmd);
2086         if (unlikely(pmd_migration)) {
2087                 swp_entry_t entry;
2088
2089                 entry = pmd_to_swp_entry(old_pmd);
2090                 page = pfn_swap_entry_to_page(entry);
2091                 write = is_writable_migration_entry(entry);
2092                 if (PageAnon(page))
2093                         anon_exclusive = is_readable_exclusive_migration_entry(entry);
2094                 young = false;
2095                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2096                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2097         } else {
2098                 page = pmd_page(old_pmd);
2099                 if (pmd_dirty(old_pmd))
2100                         SetPageDirty(page);
2101                 write = pmd_write(old_pmd);
2102                 young = pmd_young(old_pmd);
2103                 soft_dirty = pmd_soft_dirty(old_pmd);
2104                 uffd_wp = pmd_uffd_wp(old_pmd);
2105
2106                 VM_BUG_ON_PAGE(!page_count(page), page);
2107                 page_ref_add(page, HPAGE_PMD_NR - 1);
2108
2109                 /*
2110                  * Without "freeze", we'll simply split the PMD, propagating the
2111                  * PageAnonExclusive() flag for each PTE by setting it for
2112                  * each subpage -- no need to (temporarily) clear.
2113                  *
2114                  * With "freeze" we want to replace mapped pages by
2115                  * migration entries right away. This is only possible if we
2116                  * managed to clear PageAnonExclusive() -- see
2117                  * set_pmd_migration_entry().
2118                  *
2119                  * In case we cannot clear PageAnonExclusive(), split the PMD
2120                  * only and let try_to_migrate_one() fail later.
2121                  */
2122                 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2123                 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2124                         freeze = false;
2125         }
2126
2127         /*
2128          * Withdraw the table only after we mark the pmd entry invalid.
2129          * This's critical for some architectures (Power).
2130          */
2131         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2132         pmd_populate(mm, &_pmd, pgtable);
2133
2134         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2135                 pte_t entry, *pte;
2136                 /*
2137                  * Note that NUMA hinting access restrictions are not
2138                  * transferred to avoid any possibility of altering
2139                  * permissions across VMAs.
2140                  */
2141                 if (freeze || pmd_migration) {
2142                         swp_entry_t swp_entry;
2143                         if (write)
2144                                 swp_entry = make_writable_migration_entry(
2145                                                         page_to_pfn(page + i));
2146                         else if (anon_exclusive)
2147                                 swp_entry = make_readable_exclusive_migration_entry(
2148                                                         page_to_pfn(page + i));
2149                         else
2150                                 swp_entry = make_readable_migration_entry(
2151                                                         page_to_pfn(page + i));
2152                         entry = swp_entry_to_pte(swp_entry);
2153                         if (soft_dirty)
2154                                 entry = pte_swp_mksoft_dirty(entry);
2155                         if (uffd_wp)
2156                                 entry = pte_swp_mkuffd_wp(entry);
2157                 } else {
2158                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2159                         entry = maybe_mkwrite(entry, vma);
2160                         if (anon_exclusive)
2161                                 SetPageAnonExclusive(page + i);
2162                         if (!write)
2163                                 entry = pte_wrprotect(entry);
2164                         if (!young)
2165                                 entry = pte_mkold(entry);
2166                         if (soft_dirty)
2167                                 entry = pte_mksoft_dirty(entry);
2168                         if (uffd_wp)
2169                                 entry = pte_mkuffd_wp(entry);
2170                 }
2171                 pte = pte_offset_map(&_pmd, addr);
2172                 BUG_ON(!pte_none(*pte));
2173                 set_pte_at(mm, addr, pte, entry);
2174                 if (!pmd_migration)
2175                         atomic_inc(&page[i]._mapcount);
2176                 pte_unmap(pte);
2177         }
2178
2179         if (!pmd_migration) {
2180                 /*
2181                  * Set PG_double_map before dropping compound_mapcount to avoid
2182                  * false-negative page_mapped().
2183                  */
2184                 if (compound_mapcount(page) > 1 &&
2185                     !TestSetPageDoubleMap(page)) {
2186                         for (i = 0; i < HPAGE_PMD_NR; i++)
2187                                 atomic_inc(&page[i]._mapcount);
2188                 }
2189
2190                 lock_page_memcg(page);
2191                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2192                         /* Last compound_mapcount is gone. */
2193                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2194                                                 -HPAGE_PMD_NR);
2195                         if (TestClearPageDoubleMap(page)) {
2196                                 /* No need in mapcount reference anymore */
2197                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2198                                         atomic_dec(&page[i]._mapcount);
2199                         }
2200                 }
2201                 unlock_page_memcg(page);
2202
2203                 /* Above is effectively page_remove_rmap(page, vma, true) */
2204                 munlock_vma_page(page, vma, true);
2205         }
2206
2207         smp_wmb(); /* make pte visible before pmd */
2208         pmd_populate(mm, pmd, pgtable);
2209
2210         if (freeze) {
2211                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2212                         page_remove_rmap(page + i, vma, false);
2213                         put_page(page + i);
2214                 }
2215         }
2216 }
2217
2218 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2219                 unsigned long address, bool freeze, struct folio *folio)
2220 {
2221         spinlock_t *ptl;
2222         struct mmu_notifier_range range;
2223
2224         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2225                                 address & HPAGE_PMD_MASK,
2226                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2227         mmu_notifier_invalidate_range_start(&range);
2228         ptl = pmd_lock(vma->vm_mm, pmd);
2229
2230         /*
2231          * If caller asks to setup a migration entry, we need a folio to check
2232          * pmd against. Otherwise we can end up replacing wrong folio.
2233          */
2234         VM_BUG_ON(freeze && !folio);
2235         VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2236
2237         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2238             is_pmd_migration_entry(*pmd)) {
2239                 /*
2240                  * It's safe to call pmd_page when folio is set because it's
2241                  * guaranteed that pmd is present.
2242                  */
2243                 if (folio && folio != page_folio(pmd_page(*pmd)))
2244                         goto out;
2245                 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2246         }
2247
2248 out:
2249         spin_unlock(ptl);
2250         /*
2251          * No need to double call mmu_notifier->invalidate_range() callback.
2252          * They are 3 cases to consider inside __split_huge_pmd_locked():
2253          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2254          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2255          *    fault will trigger a flush_notify before pointing to a new page
2256          *    (it is fine if the secondary mmu keeps pointing to the old zero
2257          *    page in the meantime)
2258          *  3) Split a huge pmd into pte pointing to the same page. No need
2259          *     to invalidate secondary tlb entry they are all still valid.
2260          *     any further changes to individual pte will notify. So no need
2261          *     to call mmu_notifier->invalidate_range()
2262          */
2263         mmu_notifier_invalidate_range_only_end(&range);
2264 }
2265
2266 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2267                 bool freeze, struct folio *folio)
2268 {
2269         pgd_t *pgd;
2270         p4d_t *p4d;
2271         pud_t *pud;
2272         pmd_t *pmd;
2273
2274         pgd = pgd_offset(vma->vm_mm, address);
2275         if (!pgd_present(*pgd))
2276                 return;
2277
2278         p4d = p4d_offset(pgd, address);
2279         if (!p4d_present(*p4d))
2280                 return;
2281
2282         pud = pud_offset(p4d, address);
2283         if (!pud_present(*pud))
2284                 return;
2285
2286         pmd = pmd_offset(pud, address);
2287
2288         __split_huge_pmd(vma, pmd, address, freeze, folio);
2289 }
2290
2291 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2292 {
2293         /*
2294          * If the new address isn't hpage aligned and it could previously
2295          * contain an hugepage: check if we need to split an huge pmd.
2296          */
2297         if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2298             range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2299                          ALIGN(address, HPAGE_PMD_SIZE)))
2300                 split_huge_pmd_address(vma, address, false, NULL);
2301 }
2302
2303 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2304                              unsigned long start,
2305                              unsigned long end,
2306                              long adjust_next)
2307 {
2308         /* Check if we need to split start first. */
2309         split_huge_pmd_if_needed(vma, start);
2310
2311         /* Check if we need to split end next. */
2312         split_huge_pmd_if_needed(vma, end);
2313
2314         /*
2315          * If we're also updating the vma->vm_next->vm_start,
2316          * check if we need to split it.
2317          */
2318         if (adjust_next > 0) {
2319                 struct vm_area_struct *next = vma->vm_next;
2320                 unsigned long nstart = next->vm_start;
2321                 nstart += adjust_next;
2322                 split_huge_pmd_if_needed(next, nstart);
2323         }
2324 }
2325
2326 static void unmap_page(struct page *page)
2327 {
2328         struct folio *folio = page_folio(page);
2329         enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2330                 TTU_SYNC;
2331
2332         VM_BUG_ON_PAGE(!PageHead(page), page);
2333
2334         /*
2335          * Anon pages need migration entries to preserve them, but file
2336          * pages can simply be left unmapped, then faulted back on demand.
2337          * If that is ever changed (perhaps for mlock), update remap_page().
2338          */
2339         if (folio_test_anon(folio))
2340                 try_to_migrate(folio, ttu_flags);
2341         else
2342                 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2343 }
2344
2345 static void remap_page(struct folio *folio, unsigned long nr)
2346 {
2347         int i = 0;
2348
2349         /* If unmap_page() uses try_to_migrate() on file, remove this check */
2350         if (!folio_test_anon(folio))
2351                 return;
2352         for (;;) {
2353                 remove_migration_ptes(folio, folio, true);
2354                 i += folio_nr_pages(folio);
2355                 if (i >= nr)
2356                         break;
2357                 folio = folio_next(folio);
2358         }
2359 }
2360
2361 static void lru_add_page_tail(struct page *head, struct page *tail,
2362                 struct lruvec *lruvec, struct list_head *list)
2363 {
2364         VM_BUG_ON_PAGE(!PageHead(head), head);
2365         VM_BUG_ON_PAGE(PageCompound(tail), head);
2366         VM_BUG_ON_PAGE(PageLRU(tail), head);
2367         lockdep_assert_held(&lruvec->lru_lock);
2368
2369         if (list) {
2370                 /* page reclaim is reclaiming a huge page */
2371                 VM_WARN_ON(PageLRU(head));
2372                 get_page(tail);
2373                 list_add_tail(&tail->lru, list);
2374         } else {
2375                 /* head is still on lru (and we have it frozen) */
2376                 VM_WARN_ON(!PageLRU(head));
2377                 if (PageUnevictable(tail))
2378                         tail->mlock_count = 0;
2379                 else
2380                         list_add_tail(&tail->lru, &head->lru);
2381                 SetPageLRU(tail);
2382         }
2383 }
2384
2385 static void __split_huge_page_tail(struct page *head, int tail,
2386                 struct lruvec *lruvec, struct list_head *list)
2387 {
2388         struct page *page_tail = head + tail;
2389
2390         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2391
2392         /*
2393          * Clone page flags before unfreezing refcount.
2394          *
2395          * After successful get_page_unless_zero() might follow flags change,
2396          * for example lock_page() which set PG_waiters.
2397          *
2398          * Note that for mapped sub-pages of an anonymous THP,
2399          * PG_anon_exclusive has been cleared in unmap_page() and is stored in
2400          * the migration entry instead from where remap_page() will restore it.
2401          * We can still have PG_anon_exclusive set on effectively unmapped and
2402          * unreferenced sub-pages of an anonymous THP: we can simply drop
2403          * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2404          */
2405         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2406         page_tail->flags |= (head->flags &
2407                         ((1L << PG_referenced) |
2408                          (1L << PG_swapbacked) |
2409                          (1L << PG_swapcache) |
2410                          (1L << PG_mlocked) |
2411                          (1L << PG_uptodate) |
2412                          (1L << PG_active) |
2413                          (1L << PG_workingset) |
2414                          (1L << PG_locked) |
2415                          (1L << PG_unevictable) |
2416 #ifdef CONFIG_64BIT
2417                          (1L << PG_arch_2) |
2418 #endif
2419                          (1L << PG_dirty)));
2420
2421         /* ->mapping in first tail page is compound_mapcount */
2422         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2423                         page_tail);
2424         page_tail->mapping = head->mapping;
2425         page_tail->index = head->index + tail;
2426         page_tail->private = 0;
2427
2428         /* Page flags must be visible before we make the page non-compound. */
2429         smp_wmb();
2430
2431         /*
2432          * Clear PageTail before unfreezing page refcount.
2433          *
2434          * After successful get_page_unless_zero() might follow put_page()
2435          * which needs correct compound_head().
2436          */
2437         clear_compound_head(page_tail);
2438
2439         /* Finally unfreeze refcount. Additional reference from page cache. */
2440         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2441                                           PageSwapCache(head)));
2442
2443         if (page_is_young(head))
2444                 set_page_young(page_tail);
2445         if (page_is_idle(head))
2446                 set_page_idle(page_tail);
2447
2448         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2449
2450         /*
2451          * always add to the tail because some iterators expect new
2452          * pages to show after the currently processed elements - e.g.
2453          * migrate_pages
2454          */
2455         lru_add_page_tail(head, page_tail, lruvec, list);
2456 }
2457
2458 static void __split_huge_page(struct page *page, struct list_head *list,
2459                 pgoff_t end)
2460 {
2461         struct folio *folio = page_folio(page);
2462         struct page *head = &folio->page;
2463         struct lruvec *lruvec;
2464         struct address_space *swap_cache = NULL;
2465         unsigned long offset = 0;
2466         unsigned int nr = thp_nr_pages(head);
2467         int i;
2468
2469         /* complete memcg works before add pages to LRU */
2470         split_page_memcg(head, nr);
2471
2472         if (PageAnon(head) && PageSwapCache(head)) {
2473                 swp_entry_t entry = { .val = page_private(head) };
2474
2475                 offset = swp_offset(entry);
2476                 swap_cache = swap_address_space(entry);
2477                 xa_lock(&swap_cache->i_pages);
2478         }
2479
2480         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2481         lruvec = folio_lruvec_lock(folio);
2482
2483         ClearPageHasHWPoisoned(head);
2484
2485         for (i = nr - 1; i >= 1; i--) {
2486                 __split_huge_page_tail(head, i, lruvec, list);
2487                 /* Some pages can be beyond EOF: drop them from page cache */
2488                 if (head[i].index >= end) {
2489                         struct folio *tail = page_folio(head + i);
2490
2491                         if (shmem_mapping(head->mapping))
2492                                 shmem_uncharge(head->mapping->host, 1);
2493                         else if (folio_test_clear_dirty(tail))
2494                                 folio_account_cleaned(tail,
2495                                         inode_to_wb(folio->mapping->host));
2496                         __filemap_remove_folio(tail, NULL);
2497                         folio_put(tail);
2498                 } else if (!PageAnon(page)) {
2499                         __xa_store(&head->mapping->i_pages, head[i].index,
2500                                         head + i, 0);
2501                 } else if (swap_cache) {
2502                         __xa_store(&swap_cache->i_pages, offset + i,
2503                                         head + i, 0);
2504                 }
2505         }
2506
2507         ClearPageCompound(head);
2508         unlock_page_lruvec(lruvec);
2509         /* Caller disabled irqs, so they are still disabled here */
2510
2511         split_page_owner(head, nr);
2512
2513         /* See comment in __split_huge_page_tail() */
2514         if (PageAnon(head)) {
2515                 /* Additional pin to swap cache */
2516                 if (PageSwapCache(head)) {
2517                         page_ref_add(head, 2);
2518                         xa_unlock(&swap_cache->i_pages);
2519                 } else {
2520                         page_ref_inc(head);
2521                 }
2522         } else {
2523                 /* Additional pin to page cache */
2524                 page_ref_add(head, 2);
2525                 xa_unlock(&head->mapping->i_pages);
2526         }
2527         local_irq_enable();
2528
2529         remap_page(folio, nr);
2530
2531         if (PageSwapCache(head)) {
2532                 swp_entry_t entry = { .val = page_private(head) };
2533
2534                 split_swap_cluster(entry);
2535         }
2536
2537         for (i = 0; i < nr; i++) {
2538                 struct page *subpage = head + i;
2539                 if (subpage == page)
2540                         continue;
2541                 unlock_page(subpage);
2542
2543                 /*
2544                  * Subpages may be freed if there wasn't any mapping
2545                  * like if add_to_swap() is running on a lru page that
2546                  * had its mapping zapped. And freeing these pages
2547                  * requires taking the lru_lock so we do the put_page
2548                  * of the tail pages after the split is complete.
2549                  */
2550                 free_page_and_swap_cache(subpage);
2551         }
2552 }
2553
2554 /* Racy check whether the huge page can be split */
2555 bool can_split_folio(struct folio *folio, int *pextra_pins)
2556 {
2557         int extra_pins;
2558
2559         /* Additional pins from page cache */
2560         if (folio_test_anon(folio))
2561                 extra_pins = folio_test_swapcache(folio) ?
2562                                 folio_nr_pages(folio) : 0;
2563         else
2564                 extra_pins = folio_nr_pages(folio);
2565         if (pextra_pins)
2566                 *pextra_pins = extra_pins;
2567         return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2568 }
2569
2570 /*
2571  * This function splits huge page into normal pages. @page can point to any
2572  * subpage of huge page to split. Split doesn't change the position of @page.
2573  *
2574  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2575  * The huge page must be locked.
2576  *
2577  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2578  *
2579  * Both head page and tail pages will inherit mapping, flags, and so on from
2580  * the hugepage.
2581  *
2582  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2583  * they are not mapped.
2584  *
2585  * Returns 0 if the hugepage is split successfully.
2586  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2587  * us.
2588  */
2589 int split_huge_page_to_list(struct page *page, struct list_head *list)
2590 {
2591         struct folio *folio = page_folio(page);
2592         struct page *head = &folio->page;
2593         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2594         XA_STATE(xas, &head->mapping->i_pages, head->index);
2595         struct anon_vma *anon_vma = NULL;
2596         struct address_space *mapping = NULL;
2597         int extra_pins, ret;
2598         pgoff_t end;
2599         bool is_hzp;
2600
2601         VM_BUG_ON_PAGE(!PageLocked(head), head);
2602         VM_BUG_ON_PAGE(!PageCompound(head), head);
2603
2604         is_hzp = is_huge_zero_page(head);
2605         VM_WARN_ON_ONCE_PAGE(is_hzp, head);
2606         if (is_hzp)
2607                 return -EBUSY;
2608
2609         if (PageWriteback(head))
2610                 return -EBUSY;
2611
2612         if (PageAnon(head)) {
2613                 /*
2614                  * The caller does not necessarily hold an mmap_lock that would
2615                  * prevent the anon_vma disappearing so we first we take a
2616                  * reference to it and then lock the anon_vma for write. This
2617                  * is similar to folio_lock_anon_vma_read except the write lock
2618                  * is taken to serialise against parallel split or collapse
2619                  * operations.
2620                  */
2621                 anon_vma = page_get_anon_vma(head);
2622                 if (!anon_vma) {
2623                         ret = -EBUSY;
2624                         goto out;
2625                 }
2626                 end = -1;
2627                 mapping = NULL;
2628                 anon_vma_lock_write(anon_vma);
2629         } else {
2630                 mapping = head->mapping;
2631
2632                 /* Truncated ? */
2633                 if (!mapping) {
2634                         ret = -EBUSY;
2635                         goto out;
2636                 }
2637
2638                 xas_split_alloc(&xas, head, compound_order(head),
2639                                 mapping_gfp_mask(mapping) & GFP_RECLAIM_MASK);
2640                 if (xas_error(&xas)) {
2641                         ret = xas_error(&xas);
2642                         goto out;
2643                 }
2644
2645                 anon_vma = NULL;
2646                 i_mmap_lock_read(mapping);
2647
2648                 /*
2649                  *__split_huge_page() may need to trim off pages beyond EOF:
2650                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2651                  * which cannot be nested inside the page tree lock. So note
2652                  * end now: i_size itself may be changed at any moment, but
2653                  * head page lock is good enough to serialize the trimming.
2654                  */
2655                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2656                 if (shmem_mapping(mapping))
2657                         end = shmem_fallocend(mapping->host, end);
2658         }
2659
2660         /*
2661          * Racy check if we can split the page, before unmap_page() will
2662          * split PMDs
2663          */
2664         if (!can_split_folio(folio, &extra_pins)) {
2665                 ret = -EBUSY;
2666                 goto out_unlock;
2667         }
2668
2669         unmap_page(head);
2670
2671         /* block interrupt reentry in xa_lock and spinlock */
2672         local_irq_disable();
2673         if (mapping) {
2674                 /*
2675                  * Check if the head page is present in page cache.
2676                  * We assume all tail are present too, if head is there.
2677                  */
2678                 xas_lock(&xas);
2679                 xas_reset(&xas);
2680                 if (xas_load(&xas) != head)
2681                         goto fail;
2682         }
2683
2684         /* Prevent deferred_split_scan() touching ->_refcount */
2685         spin_lock(&ds_queue->split_queue_lock);
2686         if (page_ref_freeze(head, 1 + extra_pins)) {
2687                 if (!list_empty(page_deferred_list(head))) {
2688                         ds_queue->split_queue_len--;
2689                         list_del(page_deferred_list(head));
2690                 }
2691                 spin_unlock(&ds_queue->split_queue_lock);
2692                 if (mapping) {
2693                         int nr = thp_nr_pages(head);
2694
2695                         xas_split(&xas, head, thp_order(head));
2696                         if (PageSwapBacked(head)) {
2697                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2698                                                         -nr);
2699                         } else {
2700                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2701                                                         -nr);
2702                                 filemap_nr_thps_dec(mapping);
2703                         }
2704                 }
2705
2706                 __split_huge_page(page, list, end);
2707                 ret = 0;
2708         } else {
2709                 spin_unlock(&ds_queue->split_queue_lock);
2710 fail:
2711                 if (mapping)
2712                         xas_unlock(&xas);
2713                 local_irq_enable();
2714                 remap_page(folio, folio_nr_pages(folio));
2715                 ret = -EBUSY;
2716         }
2717
2718 out_unlock:
2719         if (anon_vma) {
2720                 anon_vma_unlock_write(anon_vma);
2721                 put_anon_vma(anon_vma);
2722         }
2723         if (mapping)
2724                 i_mmap_unlock_read(mapping);
2725 out:
2726         xas_destroy(&xas);
2727         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2728         return ret;
2729 }
2730
2731 void free_transhuge_page(struct page *page)
2732 {
2733         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2734         unsigned long flags;
2735
2736         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2737         if (!list_empty(page_deferred_list(page))) {
2738                 ds_queue->split_queue_len--;
2739                 list_del(page_deferred_list(page));
2740         }
2741         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2742         free_compound_page(page);
2743 }
2744
2745 void deferred_split_huge_page(struct page *page)
2746 {
2747         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2748 #ifdef CONFIG_MEMCG
2749         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2750 #endif
2751         unsigned long flags;
2752
2753         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2754
2755         /*
2756          * The try_to_unmap() in page reclaim path might reach here too,
2757          * this may cause a race condition to corrupt deferred split queue.
2758          * And, if page reclaim is already handling the same page, it is
2759          * unnecessary to handle it again in shrinker.
2760          *
2761          * Check PageSwapCache to determine if the page is being
2762          * handled by page reclaim since THP swap would add the page into
2763          * swap cache before calling try_to_unmap().
2764          */
2765         if (PageSwapCache(page))
2766                 return;
2767
2768         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2769         if (list_empty(page_deferred_list(page))) {
2770                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2771                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2772                 ds_queue->split_queue_len++;
2773 #ifdef CONFIG_MEMCG
2774                 if (memcg)
2775                         set_shrinker_bit(memcg, page_to_nid(page),
2776                                          deferred_split_shrinker.id);
2777 #endif
2778         }
2779         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2780 }
2781
2782 static unsigned long deferred_split_count(struct shrinker *shrink,
2783                 struct shrink_control *sc)
2784 {
2785         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2786         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2787
2788 #ifdef CONFIG_MEMCG
2789         if (sc->memcg)
2790                 ds_queue = &sc->memcg->deferred_split_queue;
2791 #endif
2792         return READ_ONCE(ds_queue->split_queue_len);
2793 }
2794
2795 static unsigned long deferred_split_scan(struct shrinker *shrink,
2796                 struct shrink_control *sc)
2797 {
2798         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2799         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2800         unsigned long flags;
2801         LIST_HEAD(list), *pos, *next;
2802         struct page *page;
2803         int split = 0;
2804
2805 #ifdef CONFIG_MEMCG
2806         if (sc->memcg)
2807                 ds_queue = &sc->memcg->deferred_split_queue;
2808 #endif
2809
2810         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2811         /* Take pin on all head pages to avoid freeing them under us */
2812         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2813                 page = list_entry((void *)pos, struct page, deferred_list);
2814                 page = compound_head(page);
2815                 if (get_page_unless_zero(page)) {
2816                         list_move(page_deferred_list(page), &list);
2817                 } else {
2818                         /* We lost race with put_compound_page() */
2819                         list_del_init(page_deferred_list(page));
2820                         ds_queue->split_queue_len--;
2821                 }
2822                 if (!--sc->nr_to_scan)
2823                         break;
2824         }
2825         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2826
2827         list_for_each_safe(pos, next, &list) {
2828                 page = list_entry((void *)pos, struct page, deferred_list);
2829                 if (!trylock_page(page))
2830                         goto next;
2831                 /* split_huge_page() removes page from list on success */
2832                 if (!split_huge_page(page))
2833                         split++;
2834                 unlock_page(page);
2835 next:
2836                 put_page(page);
2837         }
2838
2839         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2840         list_splice_tail(&list, &ds_queue->split_queue);
2841         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2842
2843         /*
2844          * Stop shrinker if we didn't split any page, but the queue is empty.
2845          * This can happen if pages were freed under us.
2846          */
2847         if (!split && list_empty(&ds_queue->split_queue))
2848                 return SHRINK_STOP;
2849         return split;
2850 }
2851
2852 static struct shrinker deferred_split_shrinker = {
2853         .count_objects = deferred_split_count,
2854         .scan_objects = deferred_split_scan,
2855         .seeks = DEFAULT_SEEKS,
2856         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2857                  SHRINKER_NONSLAB,
2858 };
2859
2860 #ifdef CONFIG_DEBUG_FS
2861 static void split_huge_pages_all(void)
2862 {
2863         struct zone *zone;
2864         struct page *page;
2865         unsigned long pfn, max_zone_pfn;
2866         unsigned long total = 0, split = 0;
2867
2868         pr_debug("Split all THPs\n");
2869         for_each_zone(zone) {
2870                 if (!managed_zone(zone))
2871                         continue;
2872                 max_zone_pfn = zone_end_pfn(zone);
2873                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2874                         int nr_pages;
2875                         if (!pfn_valid(pfn))
2876                                 continue;
2877
2878                         page = pfn_to_page(pfn);
2879                         if (!get_page_unless_zero(page))
2880                                 continue;
2881
2882                         if (zone != page_zone(page))
2883                                 goto next;
2884
2885                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2886                                 goto next;
2887
2888                         total++;
2889                         lock_page(page);
2890                         nr_pages = thp_nr_pages(page);
2891                         if (!split_huge_page(page))
2892                                 split++;
2893                         pfn += nr_pages - 1;
2894                         unlock_page(page);
2895 next:
2896                         put_page(page);
2897                         cond_resched();
2898                 }
2899         }
2900
2901         pr_debug("%lu of %lu THP split\n", split, total);
2902 }
2903
2904 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2905 {
2906         return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2907                     is_vm_hugetlb_page(vma);
2908 }
2909
2910 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2911                                 unsigned long vaddr_end)
2912 {
2913         int ret = 0;
2914         struct task_struct *task;
2915         struct mm_struct *mm;
2916         unsigned long total = 0, split = 0;
2917         unsigned long addr;
2918
2919         vaddr_start &= PAGE_MASK;
2920         vaddr_end &= PAGE_MASK;
2921
2922         /* Find the task_struct from pid */
2923         rcu_read_lock();
2924         task = find_task_by_vpid(pid);
2925         if (!task) {
2926                 rcu_read_unlock();
2927                 ret = -ESRCH;
2928                 goto out;
2929         }
2930         get_task_struct(task);
2931         rcu_read_unlock();
2932
2933         /* Find the mm_struct */
2934         mm = get_task_mm(task);
2935         put_task_struct(task);
2936
2937         if (!mm) {
2938                 ret = -EINVAL;
2939                 goto out;
2940         }
2941
2942         pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2943                  pid, vaddr_start, vaddr_end);
2944
2945         mmap_read_lock(mm);
2946         /*
2947          * always increase addr by PAGE_SIZE, since we could have a PTE page
2948          * table filled with PTE-mapped THPs, each of which is distinct.
2949          */
2950         for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2951                 struct vm_area_struct *vma = vma_lookup(mm, addr);
2952                 struct page *page;
2953
2954                 if (!vma)
2955                         break;
2956
2957                 /* skip special VMA and hugetlb VMA */
2958                 if (vma_not_suitable_for_thp_split(vma)) {
2959                         addr = vma->vm_end;
2960                         continue;
2961                 }
2962
2963                 /* FOLL_DUMP to ignore special (like zero) pages */
2964                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2965
2966                 if (IS_ERR_OR_NULL(page) || is_zone_device_page(page))
2967                         continue;
2968
2969                 if (!is_transparent_hugepage(page))
2970                         goto next;
2971
2972                 total++;
2973                 if (!can_split_folio(page_folio(page), NULL))
2974                         goto next;
2975
2976                 if (!trylock_page(page))
2977                         goto next;
2978
2979                 if (!split_huge_page(page))
2980                         split++;
2981
2982                 unlock_page(page);
2983 next:
2984                 put_page(page);
2985                 cond_resched();
2986         }
2987         mmap_read_unlock(mm);
2988         mmput(mm);
2989
2990         pr_debug("%lu of %lu THP split\n", split, total);
2991
2992 out:
2993         return ret;
2994 }
2995
2996 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
2997                                 pgoff_t off_end)
2998 {
2999         struct filename *file;
3000         struct file *candidate;
3001         struct address_space *mapping;
3002         int ret = -EINVAL;
3003         pgoff_t index;
3004         int nr_pages = 1;
3005         unsigned long total = 0, split = 0;
3006
3007         file = getname_kernel(file_path);
3008         if (IS_ERR(file))
3009                 return ret;
3010
3011         candidate = file_open_name(file, O_RDONLY, 0);
3012         if (IS_ERR(candidate))
3013                 goto out;
3014
3015         pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3016                  file_path, off_start, off_end);
3017
3018         mapping = candidate->f_mapping;
3019
3020         for (index = off_start; index < off_end; index += nr_pages) {
3021                 struct page *fpage = pagecache_get_page(mapping, index,
3022                                                 FGP_ENTRY | FGP_HEAD, 0);
3023
3024                 nr_pages = 1;
3025                 if (xa_is_value(fpage) || !fpage)
3026                         continue;
3027
3028                 if (!is_transparent_hugepage(fpage))
3029                         goto next;
3030
3031                 total++;
3032                 nr_pages = thp_nr_pages(fpage);
3033
3034                 if (!trylock_page(fpage))
3035                         goto next;
3036
3037                 if (!split_huge_page(fpage))
3038                         split++;
3039
3040                 unlock_page(fpage);
3041 next:
3042                 put_page(fpage);
3043                 cond_resched();
3044         }
3045
3046         filp_close(candidate, NULL);
3047         ret = 0;
3048
3049         pr_debug("%lu of %lu file-backed THP split\n", split, total);
3050 out:
3051         putname(file);
3052         return ret;
3053 }
3054
3055 #define MAX_INPUT_BUF_SZ 255
3056
3057 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3058                                 size_t count, loff_t *ppops)
3059 {
3060         static DEFINE_MUTEX(split_debug_mutex);
3061         ssize_t ret;
3062         /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3063         char input_buf[MAX_INPUT_BUF_SZ];
3064         int pid;
3065         unsigned long vaddr_start, vaddr_end;
3066
3067         ret = mutex_lock_interruptible(&split_debug_mutex);
3068         if (ret)
3069                 return ret;
3070
3071         ret = -EFAULT;
3072
3073         memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3074         if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3075                 goto out;
3076
3077         input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3078
3079         if (input_buf[0] == '/') {
3080                 char *tok;
3081                 char *buf = input_buf;
3082                 char file_path[MAX_INPUT_BUF_SZ];
3083                 pgoff_t off_start = 0, off_end = 0;
3084                 size_t input_len = strlen(input_buf);
3085
3086                 tok = strsep(&buf, ",");
3087                 if (tok) {
3088                         strcpy(file_path, tok);
3089                 } else {
3090                         ret = -EINVAL;
3091                         goto out;
3092                 }
3093
3094                 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3095                 if (ret != 2) {
3096                         ret = -EINVAL;
3097                         goto out;
3098                 }
3099                 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3100                 if (!ret)
3101                         ret = input_len;
3102
3103                 goto out;
3104         }
3105
3106         ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3107         if (ret == 1 && pid == 1) {
3108                 split_huge_pages_all();
3109                 ret = strlen(input_buf);
3110                 goto out;
3111         } else if (ret != 3) {
3112                 ret = -EINVAL;
3113                 goto out;
3114         }
3115
3116         ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3117         if (!ret)
3118                 ret = strlen(input_buf);
3119 out:
3120         mutex_unlock(&split_debug_mutex);
3121         return ret;
3122
3123 }
3124
3125 static const struct file_operations split_huge_pages_fops = {
3126         .owner   = THIS_MODULE,
3127         .write   = split_huge_pages_write,
3128         .llseek  = no_llseek,
3129 };
3130
3131 static int __init split_huge_pages_debugfs(void)
3132 {
3133         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3134                             &split_huge_pages_fops);
3135         return 0;
3136 }
3137 late_initcall(split_huge_pages_debugfs);
3138 #endif
3139
3140 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3141 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3142                 struct page *page)
3143 {
3144         struct vm_area_struct *vma = pvmw->vma;
3145         struct mm_struct *mm = vma->vm_mm;
3146         unsigned long address = pvmw->address;
3147         bool anon_exclusive;
3148         pmd_t pmdval;
3149         swp_entry_t entry;
3150         pmd_t pmdswp;
3151
3152         if (!(pvmw->pmd && !pvmw->pte))
3153                 return 0;
3154
3155         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3156         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3157
3158         anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3159         if (anon_exclusive && page_try_share_anon_rmap(page)) {
3160                 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3161                 return -EBUSY;
3162         }
3163
3164         if (pmd_dirty(pmdval))
3165                 set_page_dirty(page);
3166         if (pmd_write(pmdval))
3167                 entry = make_writable_migration_entry(page_to_pfn(page));
3168         else if (anon_exclusive)
3169                 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3170         else
3171                 entry = make_readable_migration_entry(page_to_pfn(page));
3172         pmdswp = swp_entry_to_pmd(entry);
3173         if (pmd_soft_dirty(pmdval))
3174                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3175         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3176         page_remove_rmap(page, vma, true);
3177         put_page(page);
3178         trace_set_migration_pmd(address, pmd_val(pmdswp));
3179
3180         return 0;
3181 }
3182
3183 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3184 {
3185         struct vm_area_struct *vma = pvmw->vma;
3186         struct mm_struct *mm = vma->vm_mm;
3187         unsigned long address = pvmw->address;
3188         unsigned long haddr = address & HPAGE_PMD_MASK;
3189         pmd_t pmde;
3190         swp_entry_t entry;
3191
3192         if (!(pvmw->pmd && !pvmw->pte))
3193                 return;
3194
3195         entry = pmd_to_swp_entry(*pvmw->pmd);
3196         get_page(new);
3197         pmde = pmd_mkold(mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)));
3198         if (pmd_swp_soft_dirty(*pvmw->pmd))
3199                 pmde = pmd_mksoft_dirty(pmde);
3200         if (is_writable_migration_entry(entry))
3201                 pmde = maybe_pmd_mkwrite(pmde, vma);
3202         if (pmd_swp_uffd_wp(*pvmw->pmd))
3203                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3204
3205         if (PageAnon(new)) {
3206                 rmap_t rmap_flags = RMAP_COMPOUND;
3207
3208                 if (!is_readable_migration_entry(entry))
3209                         rmap_flags |= RMAP_EXCLUSIVE;
3210
3211                 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3212         } else {
3213                 page_add_file_rmap(new, vma, true);
3214         }
3215         VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3216         set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3217
3218         /* No need to invalidate - it was non-present before */
3219         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3220         trace_remove_migration_pmd(address, pmd_val(pmde));
3221 }
3222 #endif