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, unsigned long vm_flags, 74 bool smaps, bool in_pf, bool enforce_sysfs) 75 { 76 if (!vma->vm_mm) /* vdso */ 77 return false; 78 79 /* 80 * Explicitly disabled through madvise or prctl, or some 81 * architectures may disable THP for some mappings, for 82 * example, s390 kvm. 83 * */ 84 if ((vm_flags & VM_NOHUGEPAGE) || 85 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) 86 return false; 87 /* 88 * If the hardware/firmware marked hugepage support disabled. 89 */ 90 if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX)) 91 return false; 92 93 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */ 94 if (vma_is_dax(vma)) 95 return in_pf; 96 97 /* 98 * Special VMA and hugetlb VMA. 99 * Must be checked after dax since some dax mappings may have 100 * VM_MIXEDMAP set. 101 */ 102 if (vm_flags & VM_NO_KHUGEPAGED) 103 return false; 104 105 /* 106 * Check alignment for file vma and size for both file and anon vma. 107 * 108 * Skip the check for page fault. Huge fault does the check in fault 109 * handlers. And this check is not suitable for huge PUD fault. 110 */ 111 if (!in_pf && 112 !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE))) 113 return false; 114 115 /* 116 * Enabled via shmem mount options or sysfs settings. 117 * Must be done before hugepage flags check since shmem has its 118 * own flags. 119 */ 120 if (!in_pf && shmem_file(vma->vm_file)) 121 return shmem_huge_enabled(vma); 122 123 /* Enforce sysfs THP requirements as necessary */ 124 if (enforce_sysfs && 125 (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) && 126 !hugepage_flags_always()))) 127 return false; 128 129 /* Only regular file is valid */ 130 if (!in_pf && file_thp_enabled(vma)) 131 return true; 132 133 if (!vma_is_anonymous(vma)) 134 return false; 135 136 if (vma_is_temporary_stack(vma)) 137 return false; 138 139 /* 140 * THPeligible bit of smaps should show 1 for proper VMAs even 141 * though anon_vma is not initialized yet. 142 * 143 * Allow page fault since anon_vma may be not initialized until 144 * the first page fault. 145 */ 146 if (!vma->anon_vma) 147 return (smaps || in_pf); 148 149 return true; 150 } 151 152 static bool get_huge_zero_page(void) 153 { 154 struct page *zero_page; 155 retry: 156 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 157 return true; 158 159 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 160 HPAGE_PMD_ORDER); 161 if (!zero_page) { 162 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 163 return false; 164 } 165 count_vm_event(THP_ZERO_PAGE_ALLOC); 166 preempt_disable(); 167 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 168 preempt_enable(); 169 __free_pages(zero_page, compound_order(zero_page)); 170 goto retry; 171 } 172 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); 173 174 /* We take additional reference here. It will be put back by shrinker */ 175 atomic_set(&huge_zero_refcount, 2); 176 preempt_enable(); 177 return true; 178 } 179 180 static void put_huge_zero_page(void) 181 { 182 /* 183 * Counter should never go to zero here. Only shrinker can put 184 * last reference. 185 */ 186 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 187 } 188 189 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 190 { 191 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 192 return READ_ONCE(huge_zero_page); 193 194 if (!get_huge_zero_page()) 195 return NULL; 196 197 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 198 put_huge_zero_page(); 199 200 return READ_ONCE(huge_zero_page); 201 } 202 203 void mm_put_huge_zero_page(struct mm_struct *mm) 204 { 205 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 206 put_huge_zero_page(); 207 } 208 209 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 210 struct shrink_control *sc) 211 { 212 /* we can free zero page only if last reference remains */ 213 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 214 } 215 216 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 217 struct shrink_control *sc) 218 { 219 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 220 struct page *zero_page = xchg(&huge_zero_page, NULL); 221 BUG_ON(zero_page == NULL); 222 WRITE_ONCE(huge_zero_pfn, ~0UL); 223 __free_pages(zero_page, compound_order(zero_page)); 224 return HPAGE_PMD_NR; 225 } 226 227 return 0; 228 } 229 230 static struct shrinker huge_zero_page_shrinker = { 231 .count_objects = shrink_huge_zero_page_count, 232 .scan_objects = shrink_huge_zero_page_scan, 233 .seeks = DEFAULT_SEEKS, 234 }; 235 236 #ifdef CONFIG_SYSFS 237 static ssize_t enabled_show(struct kobject *kobj, 238 struct kobj_attribute *attr, char *buf) 239 { 240 const char *output; 241 242 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 243 output = "[always] madvise never"; 244 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 245 &transparent_hugepage_flags)) 246 output = "always [madvise] never"; 247 else 248 output = "always madvise [never]"; 249 250 return sysfs_emit(buf, "%s\n", output); 251 } 252 253 static ssize_t enabled_store(struct kobject *kobj, 254 struct kobj_attribute *attr, 255 const char *buf, size_t count) 256 { 257 ssize_t ret = count; 258 259 if (sysfs_streq(buf, "always")) { 260 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 261 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 262 } else if (sysfs_streq(buf, "madvise")) { 263 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 264 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 265 } else if (sysfs_streq(buf, "never")) { 266 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 267 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 268 } else 269 ret = -EINVAL; 270 271 if (ret > 0) { 272 int err = start_stop_khugepaged(); 273 if (err) 274 ret = err; 275 } 276 return ret; 277 } 278 279 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled); 280 281 ssize_t single_hugepage_flag_show(struct kobject *kobj, 282 struct kobj_attribute *attr, char *buf, 283 enum transparent_hugepage_flag flag) 284 { 285 return sysfs_emit(buf, "%d\n", 286 !!test_bit(flag, &transparent_hugepage_flags)); 287 } 288 289 ssize_t single_hugepage_flag_store(struct kobject *kobj, 290 struct kobj_attribute *attr, 291 const char *buf, size_t count, 292 enum transparent_hugepage_flag flag) 293 { 294 unsigned long value; 295 int ret; 296 297 ret = kstrtoul(buf, 10, &value); 298 if (ret < 0) 299 return ret; 300 if (value > 1) 301 return -EINVAL; 302 303 if (value) 304 set_bit(flag, &transparent_hugepage_flags); 305 else 306 clear_bit(flag, &transparent_hugepage_flags); 307 308 return count; 309 } 310 311 static ssize_t defrag_show(struct kobject *kobj, 312 struct kobj_attribute *attr, char *buf) 313 { 314 const char *output; 315 316 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 317 &transparent_hugepage_flags)) 318 output = "[always] defer defer+madvise madvise never"; 319 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 320 &transparent_hugepage_flags)) 321 output = "always [defer] defer+madvise madvise never"; 322 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, 323 &transparent_hugepage_flags)) 324 output = "always defer [defer+madvise] madvise never"; 325 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, 326 &transparent_hugepage_flags)) 327 output = "always defer defer+madvise [madvise] never"; 328 else 329 output = "always defer defer+madvise madvise [never]"; 330 331 return sysfs_emit(buf, "%s\n", output); 332 } 333 334 static ssize_t defrag_store(struct kobject *kobj, 335 struct kobj_attribute *attr, 336 const char *buf, size_t count) 337 { 338 if (sysfs_streq(buf, "always")) { 339 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 340 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 341 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 342 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 343 } else if (sysfs_streq(buf, "defer+madvise")) { 344 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 345 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 346 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 347 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 348 } else if (sysfs_streq(buf, "defer")) { 349 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 350 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 351 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 352 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 353 } else if (sysfs_streq(buf, "madvise")) { 354 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 355 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 356 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 357 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 358 } else if (sysfs_streq(buf, "never")) { 359 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 360 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 361 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 362 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 363 } else 364 return -EINVAL; 365 366 return count; 367 } 368 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag); 369 370 static ssize_t use_zero_page_show(struct kobject *kobj, 371 struct kobj_attribute *attr, char *buf) 372 { 373 return single_hugepage_flag_show(kobj, attr, buf, 374 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 375 } 376 static ssize_t use_zero_page_store(struct kobject *kobj, 377 struct kobj_attribute *attr, const char *buf, size_t count) 378 { 379 return single_hugepage_flag_store(kobj, attr, buf, count, 380 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 381 } 382 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page); 383 384 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 385 struct kobj_attribute *attr, char *buf) 386 { 387 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); 388 } 389 static struct kobj_attribute hpage_pmd_size_attr = 390 __ATTR_RO(hpage_pmd_size); 391 392 static struct attribute *hugepage_attr[] = { 393 &enabled_attr.attr, 394 &defrag_attr.attr, 395 &use_zero_page_attr.attr, 396 &hpage_pmd_size_attr.attr, 397 #ifdef CONFIG_SHMEM 398 &shmem_enabled_attr.attr, 399 #endif 400 NULL, 401 }; 402 403 static const struct attribute_group hugepage_attr_group = { 404 .attrs = hugepage_attr, 405 }; 406 407 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 408 { 409 int err; 410 411 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 412 if (unlikely(!*hugepage_kobj)) { 413 pr_err("failed to create transparent hugepage kobject\n"); 414 return -ENOMEM; 415 } 416 417 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 418 if (err) { 419 pr_err("failed to register transparent hugepage group\n"); 420 goto delete_obj; 421 } 422 423 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 424 if (err) { 425 pr_err("failed to register transparent hugepage group\n"); 426 goto remove_hp_group; 427 } 428 429 return 0; 430 431 remove_hp_group: 432 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 433 delete_obj: 434 kobject_put(*hugepage_kobj); 435 return err; 436 } 437 438 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 439 { 440 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 441 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 442 kobject_put(hugepage_kobj); 443 } 444 #else 445 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 446 { 447 return 0; 448 } 449 450 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 451 { 452 } 453 #endif /* CONFIG_SYSFS */ 454 455 static int __init hugepage_init(void) 456 { 457 int err; 458 struct kobject *hugepage_kobj; 459 460 if (!has_transparent_hugepage()) { 461 /* 462 * Hardware doesn't support hugepages, hence disable 463 * DAX PMD support. 464 */ 465 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX; 466 return -EINVAL; 467 } 468 469 /* 470 * hugepages can't be allocated by the buddy allocator 471 */ 472 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 473 /* 474 * we use page->mapping and page->index in second tail page 475 * as list_head: assuming THP order >= 2 476 */ 477 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 478 479 err = hugepage_init_sysfs(&hugepage_kobj); 480 if (err) 481 goto err_sysfs; 482 483 err = khugepaged_init(); 484 if (err) 485 goto err_slab; 486 487 err = register_shrinker(&huge_zero_page_shrinker, "thp-zero"); 488 if (err) 489 goto err_hzp_shrinker; 490 err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split"); 491 if (err) 492 goto err_split_shrinker; 493 494 /* 495 * By default disable transparent hugepages on smaller systems, 496 * where the extra memory used could hurt more than TLB overhead 497 * is likely to save. The admin can still enable it through /sys. 498 */ 499 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 500 transparent_hugepage_flags = 0; 501 return 0; 502 } 503 504 err = start_stop_khugepaged(); 505 if (err) 506 goto err_khugepaged; 507 508 return 0; 509 err_khugepaged: 510 unregister_shrinker(&deferred_split_shrinker); 511 err_split_shrinker: 512 unregister_shrinker(&huge_zero_page_shrinker); 513 err_hzp_shrinker: 514 khugepaged_destroy(); 515 err_slab: 516 hugepage_exit_sysfs(hugepage_kobj); 517 err_sysfs: 518 return err; 519 } 520 subsys_initcall(hugepage_init); 521 522 static int __init setup_transparent_hugepage(char *str) 523 { 524 int ret = 0; 525 if (!str) 526 goto out; 527 if (!strcmp(str, "always")) { 528 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 529 &transparent_hugepage_flags); 530 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 531 &transparent_hugepage_flags); 532 ret = 1; 533 } else if (!strcmp(str, "madvise")) { 534 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 535 &transparent_hugepage_flags); 536 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 537 &transparent_hugepage_flags); 538 ret = 1; 539 } else if (!strcmp(str, "never")) { 540 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 541 &transparent_hugepage_flags); 542 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 543 &transparent_hugepage_flags); 544 ret = 1; 545 } 546 out: 547 if (!ret) 548 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 549 return ret; 550 } 551 __setup("transparent_hugepage=", setup_transparent_hugepage); 552 553 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 554 { 555 if (likely(vma->vm_flags & VM_WRITE)) 556 pmd = pmd_mkwrite(pmd); 557 return pmd; 558 } 559 560 #ifdef CONFIG_MEMCG 561 static inline struct deferred_split *get_deferred_split_queue(struct page *page) 562 { 563 struct mem_cgroup *memcg = page_memcg(compound_head(page)); 564 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 565 566 if (memcg) 567 return &memcg->deferred_split_queue; 568 else 569 return &pgdat->deferred_split_queue; 570 } 571 #else 572 static inline struct deferred_split *get_deferred_split_queue(struct page *page) 573 { 574 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 575 576 return &pgdat->deferred_split_queue; 577 } 578 #endif 579 580 void prep_transhuge_page(struct page *page) 581 { 582 /* 583 * we use page->mapping and page->index in second tail page 584 * as list_head: assuming THP order >= 2 585 */ 586 587 INIT_LIST_HEAD(page_deferred_list(page)); 588 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 589 } 590 591 static inline bool is_transparent_hugepage(struct page *page) 592 { 593 if (!PageCompound(page)) 594 return false; 595 596 page = compound_head(page); 597 return is_huge_zero_page(page) || 598 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; 599 } 600 601 static unsigned long __thp_get_unmapped_area(struct file *filp, 602 unsigned long addr, unsigned long len, 603 loff_t off, unsigned long flags, unsigned long size) 604 { 605 loff_t off_end = off + len; 606 loff_t off_align = round_up(off, size); 607 unsigned long len_pad, ret; 608 609 if (off_end <= off_align || (off_end - off_align) < size) 610 return 0; 611 612 len_pad = len + size; 613 if (len_pad < len || (off + len_pad) < off) 614 return 0; 615 616 ret = current->mm->get_unmapped_area(filp, addr, len_pad, 617 off >> PAGE_SHIFT, flags); 618 619 /* 620 * The failure might be due to length padding. The caller will retry 621 * without the padding. 622 */ 623 if (IS_ERR_VALUE(ret)) 624 return 0; 625 626 /* 627 * Do not try to align to THP boundary if allocation at the address 628 * hint succeeds. 629 */ 630 if (ret == addr) 631 return addr; 632 633 ret += (off - ret) & (size - 1); 634 return ret; 635 } 636 637 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 638 unsigned long len, unsigned long pgoff, unsigned long flags) 639 { 640 unsigned long ret; 641 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 642 643 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 644 if (ret) 645 return ret; 646 647 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 648 } 649 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 650 651 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 652 struct page *page, gfp_t gfp) 653 { 654 struct vm_area_struct *vma = vmf->vma; 655 pgtable_t pgtable; 656 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 657 vm_fault_t ret = 0; 658 659 VM_BUG_ON_PAGE(!PageCompound(page), page); 660 661 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) { 662 put_page(page); 663 count_vm_event(THP_FAULT_FALLBACK); 664 count_vm_event(THP_FAULT_FALLBACK_CHARGE); 665 return VM_FAULT_FALLBACK; 666 } 667 cgroup_throttle_swaprate(page, gfp); 668 669 pgtable = pte_alloc_one(vma->vm_mm); 670 if (unlikely(!pgtable)) { 671 ret = VM_FAULT_OOM; 672 goto release; 673 } 674 675 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 676 /* 677 * The memory barrier inside __SetPageUptodate makes sure that 678 * clear_huge_page writes become visible before the set_pmd_at() 679 * write. 680 */ 681 __SetPageUptodate(page); 682 683 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 684 if (unlikely(!pmd_none(*vmf->pmd))) { 685 goto unlock_release; 686 } else { 687 pmd_t entry; 688 689 ret = check_stable_address_space(vma->vm_mm); 690 if (ret) 691 goto unlock_release; 692 693 /* Deliver the page fault to userland */ 694 if (userfaultfd_missing(vma)) { 695 spin_unlock(vmf->ptl); 696 put_page(page); 697 pte_free(vma->vm_mm, pgtable); 698 ret = handle_userfault(vmf, VM_UFFD_MISSING); 699 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 700 return ret; 701 } 702 703 entry = mk_huge_pmd(page, vma->vm_page_prot); 704 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 705 page_add_new_anon_rmap(page, vma, haddr); 706 lru_cache_add_inactive_or_unevictable(page, vma); 707 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 708 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 709 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 710 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 711 mm_inc_nr_ptes(vma->vm_mm); 712 spin_unlock(vmf->ptl); 713 count_vm_event(THP_FAULT_ALLOC); 714 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 715 } 716 717 return 0; 718 unlock_release: 719 spin_unlock(vmf->ptl); 720 release: 721 if (pgtable) 722 pte_free(vma->vm_mm, pgtable); 723 put_page(page); 724 return ret; 725 726 } 727 728 /* 729 * always: directly stall for all thp allocations 730 * defer: wake kswapd and fail if not immediately available 731 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 732 * fail if not immediately available 733 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 734 * available 735 * never: never stall for any thp allocation 736 */ 737 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 738 { 739 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 740 741 /* Always do synchronous compaction */ 742 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 743 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 744 745 /* Kick kcompactd and fail quickly */ 746 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 747 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 748 749 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 750 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 751 return GFP_TRANSHUGE_LIGHT | 752 (vma_madvised ? __GFP_DIRECT_RECLAIM : 753 __GFP_KSWAPD_RECLAIM); 754 755 /* Only do synchronous compaction if madvised */ 756 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 757 return GFP_TRANSHUGE_LIGHT | 758 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 759 760 return GFP_TRANSHUGE_LIGHT; 761 } 762 763 /* Caller must hold page table lock. */ 764 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 766 struct page *zero_page) 767 { 768 pmd_t entry; 769 if (!pmd_none(*pmd)) 770 return; 771 entry = mk_pmd(zero_page, vma->vm_page_prot); 772 entry = pmd_mkhuge(entry); 773 pgtable_trans_huge_deposit(mm, pmd, pgtable); 774 set_pmd_at(mm, haddr, pmd, entry); 775 mm_inc_nr_ptes(mm); 776 } 777 778 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 779 { 780 struct vm_area_struct *vma = vmf->vma; 781 gfp_t gfp; 782 struct folio *folio; 783 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 784 785 if (!transhuge_vma_suitable(vma, haddr)) 786 return VM_FAULT_FALLBACK; 787 if (unlikely(anon_vma_prepare(vma))) 788 return VM_FAULT_OOM; 789 khugepaged_enter_vma(vma, vma->vm_flags); 790 791 if (!(vmf->flags & FAULT_FLAG_WRITE) && 792 !mm_forbids_zeropage(vma->vm_mm) && 793 transparent_hugepage_use_zero_page()) { 794 pgtable_t pgtable; 795 struct page *zero_page; 796 vm_fault_t ret; 797 pgtable = pte_alloc_one(vma->vm_mm); 798 if (unlikely(!pgtable)) 799 return VM_FAULT_OOM; 800 zero_page = mm_get_huge_zero_page(vma->vm_mm); 801 if (unlikely(!zero_page)) { 802 pte_free(vma->vm_mm, pgtable); 803 count_vm_event(THP_FAULT_FALLBACK); 804 return VM_FAULT_FALLBACK; 805 } 806 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 807 ret = 0; 808 if (pmd_none(*vmf->pmd)) { 809 ret = check_stable_address_space(vma->vm_mm); 810 if (ret) { 811 spin_unlock(vmf->ptl); 812 pte_free(vma->vm_mm, pgtable); 813 } else if (userfaultfd_missing(vma)) { 814 spin_unlock(vmf->ptl); 815 pte_free(vma->vm_mm, pgtable); 816 ret = handle_userfault(vmf, VM_UFFD_MISSING); 817 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 818 } else { 819 set_huge_zero_page(pgtable, vma->vm_mm, vma, 820 haddr, vmf->pmd, zero_page); 821 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 822 spin_unlock(vmf->ptl); 823 } 824 } else { 825 spin_unlock(vmf->ptl); 826 pte_free(vma->vm_mm, pgtable); 827 } 828 return ret; 829 } 830 gfp = vma_thp_gfp_mask(vma); 831 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true); 832 if (unlikely(!folio)) { 833 count_vm_event(THP_FAULT_FALLBACK); 834 return VM_FAULT_FALLBACK; 835 } 836 return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp); 837 } 838 839 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 840 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 841 pgtable_t pgtable) 842 { 843 struct mm_struct *mm = vma->vm_mm; 844 pmd_t entry; 845 spinlock_t *ptl; 846 847 ptl = pmd_lock(mm, pmd); 848 if (!pmd_none(*pmd)) { 849 if (write) { 850 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 851 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 852 goto out_unlock; 853 } 854 entry = pmd_mkyoung(*pmd); 855 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 856 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 857 update_mmu_cache_pmd(vma, addr, pmd); 858 } 859 860 goto out_unlock; 861 } 862 863 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 864 if (pfn_t_devmap(pfn)) 865 entry = pmd_mkdevmap(entry); 866 if (write) { 867 entry = pmd_mkyoung(pmd_mkdirty(entry)); 868 entry = maybe_pmd_mkwrite(entry, vma); 869 } 870 871 if (pgtable) { 872 pgtable_trans_huge_deposit(mm, pmd, pgtable); 873 mm_inc_nr_ptes(mm); 874 pgtable = NULL; 875 } 876 877 set_pmd_at(mm, addr, pmd, entry); 878 update_mmu_cache_pmd(vma, addr, pmd); 879 880 out_unlock: 881 spin_unlock(ptl); 882 if (pgtable) 883 pte_free(mm, pgtable); 884 } 885 886 /** 887 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn 888 * @vmf: Structure describing the fault 889 * @pfn: pfn to insert 890 * @pgprot: page protection to use 891 * @write: whether it's a write fault 892 * 893 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and 894 * also consult the vmf_insert_mixed_prot() documentation when 895 * @pgprot != @vmf->vma->vm_page_prot. 896 * 897 * Return: vm_fault_t value. 898 */ 899 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, 900 pgprot_t pgprot, bool write) 901 { 902 unsigned long addr = vmf->address & PMD_MASK; 903 struct vm_area_struct *vma = vmf->vma; 904 pgtable_t pgtable = NULL; 905 906 /* 907 * If we had pmd_special, we could avoid all these restrictions, 908 * but we need to be consistent with PTEs and architectures that 909 * can't support a 'special' bit. 910 */ 911 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 912 !pfn_t_devmap(pfn)); 913 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 914 (VM_PFNMAP|VM_MIXEDMAP)); 915 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 916 917 if (addr < vma->vm_start || addr >= vma->vm_end) 918 return VM_FAULT_SIGBUS; 919 920 if (arch_needs_pgtable_deposit()) { 921 pgtable = pte_alloc_one(vma->vm_mm); 922 if (!pgtable) 923 return VM_FAULT_OOM; 924 } 925 926 track_pfn_insert(vma, &pgprot, pfn); 927 928 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 929 return VM_FAULT_NOPAGE; 930 } 931 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); 932 933 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 934 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 935 { 936 if (likely(vma->vm_flags & VM_WRITE)) 937 pud = pud_mkwrite(pud); 938 return pud; 939 } 940 941 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 942 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 943 { 944 struct mm_struct *mm = vma->vm_mm; 945 pud_t entry; 946 spinlock_t *ptl; 947 948 ptl = pud_lock(mm, pud); 949 if (!pud_none(*pud)) { 950 if (write) { 951 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 952 WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 953 goto out_unlock; 954 } 955 entry = pud_mkyoung(*pud); 956 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 957 if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 958 update_mmu_cache_pud(vma, addr, pud); 959 } 960 goto out_unlock; 961 } 962 963 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 964 if (pfn_t_devmap(pfn)) 965 entry = pud_mkdevmap(entry); 966 if (write) { 967 entry = pud_mkyoung(pud_mkdirty(entry)); 968 entry = maybe_pud_mkwrite(entry, vma); 969 } 970 set_pud_at(mm, addr, pud, entry); 971 update_mmu_cache_pud(vma, addr, pud); 972 973 out_unlock: 974 spin_unlock(ptl); 975 } 976 977 /** 978 * vmf_insert_pfn_pud_prot - insert a pud size pfn 979 * @vmf: Structure describing the fault 980 * @pfn: pfn to insert 981 * @pgprot: page protection to use 982 * @write: whether it's a write fault 983 * 984 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and 985 * also consult the vmf_insert_mixed_prot() documentation when 986 * @pgprot != @vmf->vma->vm_page_prot. 987 * 988 * Return: vm_fault_t value. 989 */ 990 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, 991 pgprot_t pgprot, bool write) 992 { 993 unsigned long addr = vmf->address & PUD_MASK; 994 struct vm_area_struct *vma = vmf->vma; 995 996 /* 997 * If we had pud_special, we could avoid all these restrictions, 998 * but we need to be consistent with PTEs and architectures that 999 * can't support a 'special' bit. 1000 */ 1001 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 1002 !pfn_t_devmap(pfn)); 1003 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1004 (VM_PFNMAP|VM_MIXEDMAP)); 1005 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1006 1007 if (addr < vma->vm_start || addr >= vma->vm_end) 1008 return VM_FAULT_SIGBUS; 1009 1010 track_pfn_insert(vma, &pgprot, pfn); 1011 1012 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); 1013 return VM_FAULT_NOPAGE; 1014 } 1015 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); 1016 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1017 1018 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 1019 pmd_t *pmd, bool write) 1020 { 1021 pmd_t _pmd; 1022 1023 _pmd = pmd_mkyoung(*pmd); 1024 if (write) 1025 _pmd = pmd_mkdirty(_pmd); 1026 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1027 pmd, _pmd, write)) 1028 update_mmu_cache_pmd(vma, addr, pmd); 1029 } 1030 1031 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 1032 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 1033 { 1034 unsigned long pfn = pmd_pfn(*pmd); 1035 struct mm_struct *mm = vma->vm_mm; 1036 struct page *page; 1037 1038 assert_spin_locked(pmd_lockptr(mm, pmd)); 1039 1040 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 1041 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 1042 (FOLL_PIN | FOLL_GET))) 1043 return NULL; 1044 1045 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 1046 return NULL; 1047 1048 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 1049 /* pass */; 1050 else 1051 return NULL; 1052 1053 if (flags & FOLL_TOUCH) 1054 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 1055 1056 /* 1057 * device mapped pages can only be returned if the 1058 * caller will manage the page reference count. 1059 */ 1060 if (!(flags & (FOLL_GET | FOLL_PIN))) 1061 return ERR_PTR(-EEXIST); 1062 1063 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1064 *pgmap = get_dev_pagemap(pfn, *pgmap); 1065 if (!*pgmap) 1066 return ERR_PTR(-EFAULT); 1067 page = pfn_to_page(pfn); 1068 if (!try_grab_page(page, flags)) 1069 page = ERR_PTR(-ENOMEM); 1070 1071 return page; 1072 } 1073 1074 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1075 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1076 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1077 { 1078 spinlock_t *dst_ptl, *src_ptl; 1079 struct page *src_page; 1080 pmd_t pmd; 1081 pgtable_t pgtable = NULL; 1082 int ret = -ENOMEM; 1083 1084 /* Skip if can be re-fill on fault */ 1085 if (!vma_is_anonymous(dst_vma)) 1086 return 0; 1087 1088 pgtable = pte_alloc_one(dst_mm); 1089 if (unlikely(!pgtable)) 1090 goto out; 1091 1092 dst_ptl = pmd_lock(dst_mm, dst_pmd); 1093 src_ptl = pmd_lockptr(src_mm, src_pmd); 1094 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1095 1096 ret = -EAGAIN; 1097 pmd = *src_pmd; 1098 1099 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1100 if (unlikely(is_swap_pmd(pmd))) { 1101 swp_entry_t entry = pmd_to_swp_entry(pmd); 1102 1103 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 1104 if (!is_readable_migration_entry(entry)) { 1105 entry = make_readable_migration_entry( 1106 swp_offset(entry)); 1107 pmd = swp_entry_to_pmd(entry); 1108 if (pmd_swp_soft_dirty(*src_pmd)) 1109 pmd = pmd_swp_mksoft_dirty(pmd); 1110 if (pmd_swp_uffd_wp(*src_pmd)) 1111 pmd = pmd_swp_mkuffd_wp(pmd); 1112 set_pmd_at(src_mm, addr, src_pmd, pmd); 1113 } 1114 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1115 mm_inc_nr_ptes(dst_mm); 1116 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1117 if (!userfaultfd_wp(dst_vma)) 1118 pmd = pmd_swp_clear_uffd_wp(pmd); 1119 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1120 ret = 0; 1121 goto out_unlock; 1122 } 1123 #endif 1124 1125 if (unlikely(!pmd_trans_huge(pmd))) { 1126 pte_free(dst_mm, pgtable); 1127 goto out_unlock; 1128 } 1129 /* 1130 * When page table lock is held, the huge zero pmd should not be 1131 * under splitting since we don't split the page itself, only pmd to 1132 * a page table. 1133 */ 1134 if (is_huge_zero_pmd(pmd)) { 1135 /* 1136 * get_huge_zero_page() will never allocate a new page here, 1137 * since we already have a zero page to copy. It just takes a 1138 * reference. 1139 */ 1140 mm_get_huge_zero_page(dst_mm); 1141 goto out_zero_page; 1142 } 1143 1144 src_page = pmd_page(pmd); 1145 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 1146 1147 get_page(src_page); 1148 if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) { 1149 /* Page maybe pinned: split and retry the fault on PTEs. */ 1150 put_page(src_page); 1151 pte_free(dst_mm, pgtable); 1152 spin_unlock(src_ptl); 1153 spin_unlock(dst_ptl); 1154 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL); 1155 return -EAGAIN; 1156 } 1157 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1158 out_zero_page: 1159 mm_inc_nr_ptes(dst_mm); 1160 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1161 pmdp_set_wrprotect(src_mm, addr, src_pmd); 1162 if (!userfaultfd_wp(dst_vma)) 1163 pmd = pmd_clear_uffd_wp(pmd); 1164 pmd = pmd_mkold(pmd_wrprotect(pmd)); 1165 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1166 1167 ret = 0; 1168 out_unlock: 1169 spin_unlock(src_ptl); 1170 spin_unlock(dst_ptl); 1171 out: 1172 return ret; 1173 } 1174 1175 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1176 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1177 pud_t *pud, bool write) 1178 { 1179 pud_t _pud; 1180 1181 _pud = pud_mkyoung(*pud); 1182 if (write) 1183 _pud = pud_mkdirty(_pud); 1184 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1185 pud, _pud, write)) 1186 update_mmu_cache_pud(vma, addr, pud); 1187 } 1188 1189 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1190 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1191 { 1192 unsigned long pfn = pud_pfn(*pud); 1193 struct mm_struct *mm = vma->vm_mm; 1194 struct page *page; 1195 1196 assert_spin_locked(pud_lockptr(mm, pud)); 1197 1198 if (flags & FOLL_WRITE && !pud_write(*pud)) 1199 return NULL; 1200 1201 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 1202 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 1203 (FOLL_PIN | FOLL_GET))) 1204 return NULL; 1205 1206 if (pud_present(*pud) && pud_devmap(*pud)) 1207 /* pass */; 1208 else 1209 return NULL; 1210 1211 if (flags & FOLL_TOUCH) 1212 touch_pud(vma, addr, pud, flags & FOLL_WRITE); 1213 1214 /* 1215 * device mapped pages can only be returned if the 1216 * caller will manage the page reference count. 1217 * 1218 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1219 */ 1220 if (!(flags & (FOLL_GET | FOLL_PIN))) 1221 return ERR_PTR(-EEXIST); 1222 1223 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1224 *pgmap = get_dev_pagemap(pfn, *pgmap); 1225 if (!*pgmap) 1226 return ERR_PTR(-EFAULT); 1227 page = pfn_to_page(pfn); 1228 if (!try_grab_page(page, flags)) 1229 page = ERR_PTR(-ENOMEM); 1230 1231 return page; 1232 } 1233 1234 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1235 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1236 struct vm_area_struct *vma) 1237 { 1238 spinlock_t *dst_ptl, *src_ptl; 1239 pud_t pud; 1240 int ret; 1241 1242 dst_ptl = pud_lock(dst_mm, dst_pud); 1243 src_ptl = pud_lockptr(src_mm, src_pud); 1244 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1245 1246 ret = -EAGAIN; 1247 pud = *src_pud; 1248 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1249 goto out_unlock; 1250 1251 /* 1252 * When page table lock is held, the huge zero pud should not be 1253 * under splitting since we don't split the page itself, only pud to 1254 * a page table. 1255 */ 1256 if (is_huge_zero_pud(pud)) { 1257 /* No huge zero pud yet */ 1258 } 1259 1260 /* 1261 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap() 1262 * and split if duplicating fails. 1263 */ 1264 pudp_set_wrprotect(src_mm, addr, src_pud); 1265 pud = pud_mkold(pud_wrprotect(pud)); 1266 set_pud_at(dst_mm, addr, dst_pud, pud); 1267 1268 ret = 0; 1269 out_unlock: 1270 spin_unlock(src_ptl); 1271 spin_unlock(dst_ptl); 1272 return ret; 1273 } 1274 1275 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1276 { 1277 bool write = vmf->flags & FAULT_FLAG_WRITE; 1278 1279 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1280 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1281 goto unlock; 1282 1283 touch_pud(vmf->vma, vmf->address, vmf->pud, write); 1284 unlock: 1285 spin_unlock(vmf->ptl); 1286 } 1287 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1288 1289 void huge_pmd_set_accessed(struct vm_fault *vmf) 1290 { 1291 bool write = vmf->flags & FAULT_FLAG_WRITE; 1292 1293 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1294 if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd))) 1295 goto unlock; 1296 1297 touch_pmd(vmf->vma, vmf->address, vmf->pmd, write); 1298 1299 unlock: 1300 spin_unlock(vmf->ptl); 1301 } 1302 1303 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf) 1304 { 1305 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 1306 struct vm_area_struct *vma = vmf->vma; 1307 struct page *page; 1308 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1309 pmd_t orig_pmd = vmf->orig_pmd; 1310 1311 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1312 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1313 1314 VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE)); 1315 VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE)); 1316 1317 if (is_huge_zero_pmd(orig_pmd)) 1318 goto fallback; 1319 1320 spin_lock(vmf->ptl); 1321 1322 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1323 spin_unlock(vmf->ptl); 1324 return 0; 1325 } 1326 1327 page = pmd_page(orig_pmd); 1328 VM_BUG_ON_PAGE(!PageHead(page), page); 1329 1330 /* Early check when only holding the PT lock. */ 1331 if (PageAnonExclusive(page)) 1332 goto reuse; 1333 1334 if (!trylock_page(page)) { 1335 get_page(page); 1336 spin_unlock(vmf->ptl); 1337 lock_page(page); 1338 spin_lock(vmf->ptl); 1339 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1340 spin_unlock(vmf->ptl); 1341 unlock_page(page); 1342 put_page(page); 1343 return 0; 1344 } 1345 put_page(page); 1346 } 1347 1348 /* Recheck after temporarily dropping the PT lock. */ 1349 if (PageAnonExclusive(page)) { 1350 unlock_page(page); 1351 goto reuse; 1352 } 1353 1354 /* 1355 * See do_wp_page(): we can only reuse the page exclusively if there are 1356 * no additional references. Note that we always drain the LRU 1357 * pagevecs immediately after adding a THP. 1358 */ 1359 if (page_count(page) > 1 + PageSwapCache(page) * thp_nr_pages(page)) 1360 goto unlock_fallback; 1361 if (PageSwapCache(page)) 1362 try_to_free_swap(page); 1363 if (page_count(page) == 1) { 1364 pmd_t entry; 1365 1366 page_move_anon_rmap(page, vma); 1367 unlock_page(page); 1368 reuse: 1369 if (unlikely(unshare)) { 1370 spin_unlock(vmf->ptl); 1371 return 0; 1372 } 1373 entry = pmd_mkyoung(orig_pmd); 1374 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1375 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1376 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1377 spin_unlock(vmf->ptl); 1378 return VM_FAULT_WRITE; 1379 } 1380 1381 unlock_fallback: 1382 unlock_page(page); 1383 spin_unlock(vmf->ptl); 1384 fallback: 1385 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 1386 return VM_FAULT_FALLBACK; 1387 } 1388 1389 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ 1390 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, 1391 struct vm_area_struct *vma, 1392 unsigned int flags) 1393 { 1394 /* If the pmd is writable, we can write to the page. */ 1395 if (pmd_write(pmd)) 1396 return true; 1397 1398 /* Maybe FOLL_FORCE is set to override it? */ 1399 if (!(flags & FOLL_FORCE)) 1400 return false; 1401 1402 /* But FOLL_FORCE has no effect on shared mappings */ 1403 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) 1404 return false; 1405 1406 /* ... or read-only private ones */ 1407 if (!(vma->vm_flags & VM_MAYWRITE)) 1408 return false; 1409 1410 /* ... or already writable ones that just need to take a write fault */ 1411 if (vma->vm_flags & VM_WRITE) 1412 return false; 1413 1414 /* 1415 * See can_change_pte_writable(): we broke COW and could map the page 1416 * writable if we have an exclusive anonymous page ... 1417 */ 1418 if (!page || !PageAnon(page) || !PageAnonExclusive(page)) 1419 return false; 1420 1421 /* ... and a write-fault isn't required for other reasons. */ 1422 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 1423 return false; 1424 return !userfaultfd_huge_pmd_wp(vma, pmd); 1425 } 1426 1427 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1428 unsigned long addr, 1429 pmd_t *pmd, 1430 unsigned int flags) 1431 { 1432 struct mm_struct *mm = vma->vm_mm; 1433 struct page *page; 1434 1435 assert_spin_locked(pmd_lockptr(mm, pmd)); 1436 1437 page = pmd_page(*pmd); 1438 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1439 1440 if ((flags & FOLL_WRITE) && 1441 !can_follow_write_pmd(*pmd, page, vma, flags)) 1442 return NULL; 1443 1444 /* Avoid dumping huge zero page */ 1445 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1446 return ERR_PTR(-EFAULT); 1447 1448 /* Full NUMA hinting faults to serialise migration in fault paths */ 1449 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1450 return NULL; 1451 1452 if (!pmd_write(*pmd) && gup_must_unshare(flags, page)) 1453 return ERR_PTR(-EMLINK); 1454 1455 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 1456 !PageAnonExclusive(page), page); 1457 1458 if (!try_grab_page(page, flags)) 1459 return ERR_PTR(-ENOMEM); 1460 1461 if (flags & FOLL_TOUCH) 1462 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 1463 1464 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1465 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1466 1467 return page; 1468 } 1469 1470 /* NUMA hinting page fault entry point for trans huge pmds */ 1471 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf) 1472 { 1473 struct vm_area_struct *vma = vmf->vma; 1474 pmd_t oldpmd = vmf->orig_pmd; 1475 pmd_t pmd; 1476 struct page *page; 1477 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1478 int page_nid = NUMA_NO_NODE; 1479 int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK); 1480 bool migrated = false; 1481 bool was_writable = pmd_savedwrite(oldpmd); 1482 int flags = 0; 1483 1484 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1485 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 1486 spin_unlock(vmf->ptl); 1487 goto out; 1488 } 1489 1490 pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1491 page = vm_normal_page_pmd(vma, haddr, pmd); 1492 if (!page) 1493 goto out_map; 1494 1495 /* See similar comment in do_numa_page for explanation */ 1496 if (!was_writable) 1497 flags |= TNF_NO_GROUP; 1498 1499 page_nid = page_to_nid(page); 1500 /* 1501 * For memory tiering mode, cpupid of slow memory page is used 1502 * to record page access time. So use default value. 1503 */ 1504 if (node_is_toptier(page_nid)) 1505 last_cpupid = page_cpupid_last(page); 1506 target_nid = numa_migrate_prep(page, vma, haddr, page_nid, 1507 &flags); 1508 1509 if (target_nid == NUMA_NO_NODE) { 1510 put_page(page); 1511 goto out_map; 1512 } 1513 1514 spin_unlock(vmf->ptl); 1515 1516 migrated = migrate_misplaced_page(page, vma, target_nid); 1517 if (migrated) { 1518 flags |= TNF_MIGRATED; 1519 page_nid = target_nid; 1520 } else { 1521 flags |= TNF_MIGRATE_FAIL; 1522 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1523 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 1524 spin_unlock(vmf->ptl); 1525 goto out; 1526 } 1527 goto out_map; 1528 } 1529 1530 out: 1531 if (page_nid != NUMA_NO_NODE) 1532 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1533 flags); 1534 1535 return 0; 1536 1537 out_map: 1538 /* Restore the PMD */ 1539 pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1540 pmd = pmd_mkyoung(pmd); 1541 if (was_writable) 1542 pmd = pmd_mkwrite(pmd); 1543 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1544 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1545 spin_unlock(vmf->ptl); 1546 goto out; 1547 } 1548 1549 /* 1550 * Return true if we do MADV_FREE successfully on entire pmd page. 1551 * Otherwise, return false. 1552 */ 1553 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1554 pmd_t *pmd, unsigned long addr, unsigned long next) 1555 { 1556 spinlock_t *ptl; 1557 pmd_t orig_pmd; 1558 struct page *page; 1559 struct mm_struct *mm = tlb->mm; 1560 bool ret = false; 1561 1562 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1563 1564 ptl = pmd_trans_huge_lock(pmd, vma); 1565 if (!ptl) 1566 goto out_unlocked; 1567 1568 orig_pmd = *pmd; 1569 if (is_huge_zero_pmd(orig_pmd)) 1570 goto out; 1571 1572 if (unlikely(!pmd_present(orig_pmd))) { 1573 VM_BUG_ON(thp_migration_supported() && 1574 !is_pmd_migration_entry(orig_pmd)); 1575 goto out; 1576 } 1577 1578 page = pmd_page(orig_pmd); 1579 /* 1580 * If other processes are mapping this page, we couldn't discard 1581 * the page unless they all do MADV_FREE so let's skip the page. 1582 */ 1583 if (total_mapcount(page) != 1) 1584 goto out; 1585 1586 if (!trylock_page(page)) 1587 goto out; 1588 1589 /* 1590 * If user want to discard part-pages of THP, split it so MADV_FREE 1591 * will deactivate only them. 1592 */ 1593 if (next - addr != HPAGE_PMD_SIZE) { 1594 get_page(page); 1595 spin_unlock(ptl); 1596 split_huge_page(page); 1597 unlock_page(page); 1598 put_page(page); 1599 goto out_unlocked; 1600 } 1601 1602 if (PageDirty(page)) 1603 ClearPageDirty(page); 1604 unlock_page(page); 1605 1606 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1607 pmdp_invalidate(vma, addr, pmd); 1608 orig_pmd = pmd_mkold(orig_pmd); 1609 orig_pmd = pmd_mkclean(orig_pmd); 1610 1611 set_pmd_at(mm, addr, pmd, orig_pmd); 1612 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1613 } 1614 1615 mark_page_lazyfree(page); 1616 ret = true; 1617 out: 1618 spin_unlock(ptl); 1619 out_unlocked: 1620 return ret; 1621 } 1622 1623 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1624 { 1625 pgtable_t pgtable; 1626 1627 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1628 pte_free(mm, pgtable); 1629 mm_dec_nr_ptes(mm); 1630 } 1631 1632 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1633 pmd_t *pmd, unsigned long addr) 1634 { 1635 pmd_t orig_pmd; 1636 spinlock_t *ptl; 1637 1638 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1639 1640 ptl = __pmd_trans_huge_lock(pmd, vma); 1641 if (!ptl) 1642 return 0; 1643 /* 1644 * For architectures like ppc64 we look at deposited pgtable 1645 * when calling pmdp_huge_get_and_clear. So do the 1646 * pgtable_trans_huge_withdraw after finishing pmdp related 1647 * operations. 1648 */ 1649 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1650 tlb->fullmm); 1651 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1652 if (vma_is_special_huge(vma)) { 1653 if (arch_needs_pgtable_deposit()) 1654 zap_deposited_table(tlb->mm, pmd); 1655 spin_unlock(ptl); 1656 } else if (is_huge_zero_pmd(orig_pmd)) { 1657 zap_deposited_table(tlb->mm, pmd); 1658 spin_unlock(ptl); 1659 } else { 1660 struct page *page = NULL; 1661 int flush_needed = 1; 1662 1663 if (pmd_present(orig_pmd)) { 1664 page = pmd_page(orig_pmd); 1665 page_remove_rmap(page, vma, true); 1666 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1667 VM_BUG_ON_PAGE(!PageHead(page), page); 1668 } else if (thp_migration_supported()) { 1669 swp_entry_t entry; 1670 1671 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1672 entry = pmd_to_swp_entry(orig_pmd); 1673 page = pfn_swap_entry_to_page(entry); 1674 flush_needed = 0; 1675 } else 1676 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1677 1678 if (PageAnon(page)) { 1679 zap_deposited_table(tlb->mm, pmd); 1680 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1681 } else { 1682 if (arch_needs_pgtable_deposit()) 1683 zap_deposited_table(tlb->mm, pmd); 1684 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1685 } 1686 1687 spin_unlock(ptl); 1688 if (flush_needed) 1689 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1690 } 1691 return 1; 1692 } 1693 1694 #ifndef pmd_move_must_withdraw 1695 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1696 spinlock_t *old_pmd_ptl, 1697 struct vm_area_struct *vma) 1698 { 1699 /* 1700 * With split pmd lock we also need to move preallocated 1701 * PTE page table if new_pmd is on different PMD page table. 1702 * 1703 * We also don't deposit and withdraw tables for file pages. 1704 */ 1705 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1706 } 1707 #endif 1708 1709 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1710 { 1711 #ifdef CONFIG_MEM_SOFT_DIRTY 1712 if (unlikely(is_pmd_migration_entry(pmd))) 1713 pmd = pmd_swp_mksoft_dirty(pmd); 1714 else if (pmd_present(pmd)) 1715 pmd = pmd_mksoft_dirty(pmd); 1716 #endif 1717 return pmd; 1718 } 1719 1720 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1721 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 1722 { 1723 spinlock_t *old_ptl, *new_ptl; 1724 pmd_t pmd; 1725 struct mm_struct *mm = vma->vm_mm; 1726 bool force_flush = false; 1727 1728 /* 1729 * The destination pmd shouldn't be established, free_pgtables() 1730 * should have release it. 1731 */ 1732 if (WARN_ON(!pmd_none(*new_pmd))) { 1733 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1734 return false; 1735 } 1736 1737 /* 1738 * We don't have to worry about the ordering of src and dst 1739 * ptlocks because exclusive mmap_lock prevents deadlock. 1740 */ 1741 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1742 if (old_ptl) { 1743 new_ptl = pmd_lockptr(mm, new_pmd); 1744 if (new_ptl != old_ptl) 1745 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1746 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1747 if (pmd_present(pmd)) 1748 force_flush = true; 1749 VM_BUG_ON(!pmd_none(*new_pmd)); 1750 1751 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1752 pgtable_t pgtable; 1753 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1754 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1755 } 1756 pmd = move_soft_dirty_pmd(pmd); 1757 set_pmd_at(mm, new_addr, new_pmd, pmd); 1758 if (force_flush) 1759 flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1760 if (new_ptl != old_ptl) 1761 spin_unlock(new_ptl); 1762 spin_unlock(old_ptl); 1763 return true; 1764 } 1765 return false; 1766 } 1767 1768 /* 1769 * Returns 1770 * - 0 if PMD could not be locked 1771 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 1772 * or if prot_numa but THP migration is not supported 1773 * - HPAGE_PMD_NR if protections changed and TLB flush necessary 1774 */ 1775 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1776 pmd_t *pmd, unsigned long addr, pgprot_t newprot, 1777 unsigned long cp_flags) 1778 { 1779 struct mm_struct *mm = vma->vm_mm; 1780 spinlock_t *ptl; 1781 pmd_t oldpmd, entry; 1782 bool preserve_write; 1783 int ret; 1784 bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 1785 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 1786 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 1787 1788 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1789 1790 if (prot_numa && !thp_migration_supported()) 1791 return 1; 1792 1793 ptl = __pmd_trans_huge_lock(pmd, vma); 1794 if (!ptl) 1795 return 0; 1796 1797 preserve_write = prot_numa && pmd_write(*pmd); 1798 ret = 1; 1799 1800 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1801 if (is_swap_pmd(*pmd)) { 1802 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1803 struct page *page = pfn_swap_entry_to_page(entry); 1804 1805 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1806 if (is_writable_migration_entry(entry)) { 1807 pmd_t newpmd; 1808 /* 1809 * A protection check is difficult so 1810 * just be safe and disable write 1811 */ 1812 if (PageAnon(page)) 1813 entry = make_readable_exclusive_migration_entry(swp_offset(entry)); 1814 else 1815 entry = make_readable_migration_entry(swp_offset(entry)); 1816 newpmd = swp_entry_to_pmd(entry); 1817 if (pmd_swp_soft_dirty(*pmd)) 1818 newpmd = pmd_swp_mksoft_dirty(newpmd); 1819 if (pmd_swp_uffd_wp(*pmd)) 1820 newpmd = pmd_swp_mkuffd_wp(newpmd); 1821 set_pmd_at(mm, addr, pmd, newpmd); 1822 } 1823 goto unlock; 1824 } 1825 #endif 1826 1827 if (prot_numa) { 1828 struct page *page; 1829 bool toptier; 1830 /* 1831 * Avoid trapping faults against the zero page. The read-only 1832 * data is likely to be read-cached on the local CPU and 1833 * local/remote hits to the zero page are not interesting. 1834 */ 1835 if (is_huge_zero_pmd(*pmd)) 1836 goto unlock; 1837 1838 if (pmd_protnone(*pmd)) 1839 goto unlock; 1840 1841 page = pmd_page(*pmd); 1842 toptier = node_is_toptier(page_to_nid(page)); 1843 /* 1844 * Skip scanning top tier node if normal numa 1845 * balancing is disabled 1846 */ 1847 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 1848 toptier) 1849 goto unlock; 1850 1851 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && 1852 !toptier) 1853 xchg_page_access_time(page, jiffies_to_msecs(jiffies)); 1854 } 1855 /* 1856 * In case prot_numa, we are under mmap_read_lock(mm). It's critical 1857 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1858 * which is also under mmap_read_lock(mm): 1859 * 1860 * CPU0: CPU1: 1861 * change_huge_pmd(prot_numa=1) 1862 * pmdp_huge_get_and_clear_notify() 1863 * madvise_dontneed() 1864 * zap_pmd_range() 1865 * pmd_trans_huge(*pmd) == 0 (without ptl) 1866 * // skip the pmd 1867 * set_pmd_at(); 1868 * // pmd is re-established 1869 * 1870 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1871 * which may break userspace. 1872 * 1873 * pmdp_invalidate_ad() is required to make sure we don't miss 1874 * dirty/young flags set by hardware. 1875 */ 1876 oldpmd = pmdp_invalidate_ad(vma, addr, pmd); 1877 1878 entry = pmd_modify(oldpmd, newprot); 1879 if (preserve_write) 1880 entry = pmd_mk_savedwrite(entry); 1881 if (uffd_wp) { 1882 entry = pmd_wrprotect(entry); 1883 entry = pmd_mkuffd_wp(entry); 1884 } else if (uffd_wp_resolve) { 1885 /* 1886 * Leave the write bit to be handled by PF interrupt 1887 * handler, then things like COW could be properly 1888 * handled. 1889 */ 1890 entry = pmd_clear_uffd_wp(entry); 1891 } 1892 ret = HPAGE_PMD_NR; 1893 set_pmd_at(mm, addr, pmd, entry); 1894 1895 if (huge_pmd_needs_flush(oldpmd, entry)) 1896 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE); 1897 1898 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1899 unlock: 1900 spin_unlock(ptl); 1901 return ret; 1902 } 1903 1904 /* 1905 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1906 * 1907 * Note that if it returns page table lock pointer, this routine returns without 1908 * unlocking page table lock. So callers must unlock it. 1909 */ 1910 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1911 { 1912 spinlock_t *ptl; 1913 ptl = pmd_lock(vma->vm_mm, pmd); 1914 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1915 pmd_devmap(*pmd))) 1916 return ptl; 1917 spin_unlock(ptl); 1918 return NULL; 1919 } 1920 1921 /* 1922 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise. 1923 * 1924 * Note that if it returns page table lock pointer, this routine returns without 1925 * unlocking page table lock. So callers must unlock it. 1926 */ 1927 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1928 { 1929 spinlock_t *ptl; 1930 1931 ptl = pud_lock(vma->vm_mm, pud); 1932 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1933 return ptl; 1934 spin_unlock(ptl); 1935 return NULL; 1936 } 1937 1938 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1939 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1940 pud_t *pud, unsigned long addr) 1941 { 1942 spinlock_t *ptl; 1943 1944 ptl = __pud_trans_huge_lock(pud, vma); 1945 if (!ptl) 1946 return 0; 1947 1948 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); 1949 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1950 if (vma_is_special_huge(vma)) { 1951 spin_unlock(ptl); 1952 /* No zero page support yet */ 1953 } else { 1954 /* No support for anonymous PUD pages yet */ 1955 BUG(); 1956 } 1957 return 1; 1958 } 1959 1960 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1961 unsigned long haddr) 1962 { 1963 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1964 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1965 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1966 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1967 1968 count_vm_event(THP_SPLIT_PUD); 1969 1970 pudp_huge_clear_flush_notify(vma, haddr, pud); 1971 } 1972 1973 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 1974 unsigned long address) 1975 { 1976 spinlock_t *ptl; 1977 struct mmu_notifier_range range; 1978 1979 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1980 address & HPAGE_PUD_MASK, 1981 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 1982 mmu_notifier_invalidate_range_start(&range); 1983 ptl = pud_lock(vma->vm_mm, pud); 1984 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 1985 goto out; 1986 __split_huge_pud_locked(vma, pud, range.start); 1987 1988 out: 1989 spin_unlock(ptl); 1990 /* 1991 * No need to double call mmu_notifier->invalidate_range() callback as 1992 * the above pudp_huge_clear_flush_notify() did already call it. 1993 */ 1994 mmu_notifier_invalidate_range_only_end(&range); 1995 } 1996 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1997 1998 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1999 unsigned long haddr, pmd_t *pmd) 2000 { 2001 struct mm_struct *mm = vma->vm_mm; 2002 pgtable_t pgtable; 2003 pmd_t _pmd; 2004 int i; 2005 2006 /* 2007 * Leave pmd empty until pte is filled note that it is fine to delay 2008 * notification until mmu_notifier_invalidate_range_end() as we are 2009 * replacing a zero pmd write protected page with a zero pte write 2010 * protected page. 2011 * 2012 * See Documentation/mm/mmu_notifier.rst 2013 */ 2014 pmdp_huge_clear_flush(vma, haddr, pmd); 2015 2016 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2017 pmd_populate(mm, &_pmd, pgtable); 2018 2019 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2020 pte_t *pte, entry; 2021 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2022 entry = pte_mkspecial(entry); 2023 pte = pte_offset_map(&_pmd, haddr); 2024 VM_BUG_ON(!pte_none(*pte)); 2025 set_pte_at(mm, haddr, pte, entry); 2026 pte_unmap(pte); 2027 } 2028 smp_wmb(); /* make pte visible before pmd */ 2029 pmd_populate(mm, pmd, pgtable); 2030 } 2031 2032 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2033 unsigned long haddr, bool freeze) 2034 { 2035 struct mm_struct *mm = vma->vm_mm; 2036 struct page *page; 2037 pgtable_t pgtable; 2038 pmd_t old_pmd, _pmd; 2039 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 2040 bool anon_exclusive = false, dirty = false; 2041 unsigned long addr; 2042 int i; 2043 2044 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2045 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2046 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2047 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2048 && !pmd_devmap(*pmd)); 2049 2050 count_vm_event(THP_SPLIT_PMD); 2051 2052 if (!vma_is_anonymous(vma)) { 2053 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2054 /* 2055 * We are going to unmap this huge page. So 2056 * just go ahead and zap it 2057 */ 2058 if (arch_needs_pgtable_deposit()) 2059 zap_deposited_table(mm, pmd); 2060 if (vma_is_special_huge(vma)) 2061 return; 2062 if (unlikely(is_pmd_migration_entry(old_pmd))) { 2063 swp_entry_t entry; 2064 2065 entry = pmd_to_swp_entry(old_pmd); 2066 page = pfn_swap_entry_to_page(entry); 2067 } else { 2068 page = pmd_page(old_pmd); 2069 if (!PageDirty(page) && pmd_dirty(old_pmd)) 2070 set_page_dirty(page); 2071 if (!PageReferenced(page) && pmd_young(old_pmd)) 2072 SetPageReferenced(page); 2073 page_remove_rmap(page, vma, true); 2074 put_page(page); 2075 } 2076 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2077 return; 2078 } 2079 2080 if (is_huge_zero_pmd(*pmd)) { 2081 /* 2082 * FIXME: Do we want to invalidate secondary mmu by calling 2083 * mmu_notifier_invalidate_range() see comments below inside 2084 * __split_huge_pmd() ? 2085 * 2086 * We are going from a zero huge page write protected to zero 2087 * small page also write protected so it does not seems useful 2088 * to invalidate secondary mmu at this time. 2089 */ 2090 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2091 } 2092 2093 /* 2094 * Up to this point the pmd is present and huge and userland has the 2095 * whole access to the hugepage during the split (which happens in 2096 * place). If we overwrite the pmd with the not-huge version pointing 2097 * to the pte here (which of course we could if all CPUs were bug 2098 * free), userland could trigger a small page size TLB miss on the 2099 * small sized TLB while the hugepage TLB entry is still established in 2100 * the huge TLB. Some CPU doesn't like that. 2101 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 2102 * 383 on page 105. Intel should be safe but is also warns that it's 2103 * only safe if the permission and cache attributes of the two entries 2104 * loaded in the two TLB is identical (which should be the case here). 2105 * But it is generally safer to never allow small and huge TLB entries 2106 * for the same virtual address to be loaded simultaneously. So instead 2107 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2108 * current pmd notpresent (atomically because here the pmd_trans_huge 2109 * must remain set at all times on the pmd until the split is complete 2110 * for this pmd), then we flush the SMP TLB and finally we write the 2111 * non-huge version of the pmd entry with pmd_populate. 2112 */ 2113 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2114 2115 pmd_migration = is_pmd_migration_entry(old_pmd); 2116 if (unlikely(pmd_migration)) { 2117 swp_entry_t entry; 2118 2119 entry = pmd_to_swp_entry(old_pmd); 2120 page = pfn_swap_entry_to_page(entry); 2121 write = is_writable_migration_entry(entry); 2122 if (PageAnon(page)) 2123 anon_exclusive = is_readable_exclusive_migration_entry(entry); 2124 young = is_migration_entry_young(entry); 2125 dirty = is_migration_entry_dirty(entry); 2126 soft_dirty = pmd_swp_soft_dirty(old_pmd); 2127 uffd_wp = pmd_swp_uffd_wp(old_pmd); 2128 } else { 2129 page = pmd_page(old_pmd); 2130 if (pmd_dirty(old_pmd)) { 2131 dirty = true; 2132 SetPageDirty(page); 2133 } 2134 write = pmd_write(old_pmd); 2135 young = pmd_young(old_pmd); 2136 soft_dirty = pmd_soft_dirty(old_pmd); 2137 uffd_wp = pmd_uffd_wp(old_pmd); 2138 2139 VM_BUG_ON_PAGE(!page_count(page), page); 2140 page_ref_add(page, HPAGE_PMD_NR - 1); 2141 2142 /* 2143 * Without "freeze", we'll simply split the PMD, propagating the 2144 * PageAnonExclusive() flag for each PTE by setting it for 2145 * each subpage -- no need to (temporarily) clear. 2146 * 2147 * With "freeze" we want to replace mapped pages by 2148 * migration entries right away. This is only possible if we 2149 * managed to clear PageAnonExclusive() -- see 2150 * set_pmd_migration_entry(). 2151 * 2152 * In case we cannot clear PageAnonExclusive(), split the PMD 2153 * only and let try_to_migrate_one() fail later. 2154 * 2155 * See page_try_share_anon_rmap(): invalidate PMD first. 2156 */ 2157 anon_exclusive = PageAnon(page) && PageAnonExclusive(page); 2158 if (freeze && anon_exclusive && page_try_share_anon_rmap(page)) 2159 freeze = false; 2160 } 2161 2162 /* 2163 * Withdraw the table only after we mark the pmd entry invalid. 2164 * This's critical for some architectures (Power). 2165 */ 2166 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2167 pmd_populate(mm, &_pmd, pgtable); 2168 2169 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2170 pte_t entry, *pte; 2171 /* 2172 * Note that NUMA hinting access restrictions are not 2173 * transferred to avoid any possibility of altering 2174 * permissions across VMAs. 2175 */ 2176 if (freeze || pmd_migration) { 2177 swp_entry_t swp_entry; 2178 if (write) 2179 swp_entry = make_writable_migration_entry( 2180 page_to_pfn(page + i)); 2181 else if (anon_exclusive) 2182 swp_entry = make_readable_exclusive_migration_entry( 2183 page_to_pfn(page + i)); 2184 else 2185 swp_entry = make_readable_migration_entry( 2186 page_to_pfn(page + i)); 2187 if (young) 2188 swp_entry = make_migration_entry_young(swp_entry); 2189 if (dirty) 2190 swp_entry = make_migration_entry_dirty(swp_entry); 2191 entry = swp_entry_to_pte(swp_entry); 2192 if (soft_dirty) 2193 entry = pte_swp_mksoft_dirty(entry); 2194 if (uffd_wp) 2195 entry = pte_swp_mkuffd_wp(entry); 2196 } else { 2197 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2198 entry = maybe_mkwrite(entry, vma); 2199 if (anon_exclusive) 2200 SetPageAnonExclusive(page + i); 2201 if (!write) 2202 entry = pte_wrprotect(entry); 2203 if (!young) 2204 entry = pte_mkold(entry); 2205 /* NOTE: this may set soft-dirty too on some archs */ 2206 if (dirty) 2207 entry = pte_mkdirty(entry); 2208 if (soft_dirty) 2209 entry = pte_mksoft_dirty(entry); 2210 if (uffd_wp) 2211 entry = pte_mkuffd_wp(entry); 2212 } 2213 pte = pte_offset_map(&_pmd, addr); 2214 BUG_ON(!pte_none(*pte)); 2215 set_pte_at(mm, addr, pte, entry); 2216 if (!pmd_migration) 2217 atomic_inc(&page[i]._mapcount); 2218 pte_unmap(pte); 2219 } 2220 2221 if (!pmd_migration) { 2222 /* 2223 * Set PG_double_map before dropping compound_mapcount to avoid 2224 * false-negative page_mapped(). 2225 */ 2226 if (compound_mapcount(page) > 1 && 2227 !TestSetPageDoubleMap(page)) { 2228 for (i = 0; i < HPAGE_PMD_NR; i++) 2229 atomic_inc(&page[i]._mapcount); 2230 } 2231 2232 lock_page_memcg(page); 2233 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2234 /* Last compound_mapcount is gone. */ 2235 __mod_lruvec_page_state(page, NR_ANON_THPS, 2236 -HPAGE_PMD_NR); 2237 if (TestClearPageDoubleMap(page)) { 2238 /* No need in mapcount reference anymore */ 2239 for (i = 0; i < HPAGE_PMD_NR; i++) 2240 atomic_dec(&page[i]._mapcount); 2241 } 2242 } 2243 unlock_page_memcg(page); 2244 2245 /* Above is effectively page_remove_rmap(page, vma, true) */ 2246 munlock_vma_page(page, vma, true); 2247 } 2248 2249 smp_wmb(); /* make pte visible before pmd */ 2250 pmd_populate(mm, pmd, pgtable); 2251 2252 if (freeze) { 2253 for (i = 0; i < HPAGE_PMD_NR; i++) { 2254 page_remove_rmap(page + i, vma, false); 2255 put_page(page + i); 2256 } 2257 } 2258 } 2259 2260 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2261 unsigned long address, bool freeze, struct folio *folio) 2262 { 2263 spinlock_t *ptl; 2264 struct mmu_notifier_range range; 2265 2266 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 2267 address & HPAGE_PMD_MASK, 2268 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2269 mmu_notifier_invalidate_range_start(&range); 2270 ptl = pmd_lock(vma->vm_mm, pmd); 2271 2272 /* 2273 * If caller asks to setup a migration entry, we need a folio to check 2274 * pmd against. Otherwise we can end up replacing wrong folio. 2275 */ 2276 VM_BUG_ON(freeze && !folio); 2277 VM_WARN_ON_ONCE(folio && !folio_test_locked(folio)); 2278 2279 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || 2280 is_pmd_migration_entry(*pmd)) { 2281 /* 2282 * It's safe to call pmd_page when folio is set because it's 2283 * guaranteed that pmd is present. 2284 */ 2285 if (folio && folio != page_folio(pmd_page(*pmd))) 2286 goto out; 2287 __split_huge_pmd_locked(vma, pmd, range.start, freeze); 2288 } 2289 2290 out: 2291 spin_unlock(ptl); 2292 /* 2293 * No need to double call mmu_notifier->invalidate_range() callback. 2294 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2295 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2296 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2297 * fault will trigger a flush_notify before pointing to a new page 2298 * (it is fine if the secondary mmu keeps pointing to the old zero 2299 * page in the meantime) 2300 * 3) Split a huge pmd into pte pointing to the same page. No need 2301 * to invalidate secondary tlb entry they are all still valid. 2302 * any further changes to individual pte will notify. So no need 2303 * to call mmu_notifier->invalidate_range() 2304 */ 2305 mmu_notifier_invalidate_range_only_end(&range); 2306 } 2307 2308 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2309 bool freeze, struct folio *folio) 2310 { 2311 pmd_t *pmd = mm_find_pmd(vma->vm_mm, address); 2312 2313 if (!pmd) 2314 return; 2315 2316 __split_huge_pmd(vma, pmd, address, freeze, folio); 2317 } 2318 2319 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 2320 { 2321 /* 2322 * If the new address isn't hpage aligned and it could previously 2323 * contain an hugepage: check if we need to split an huge pmd. 2324 */ 2325 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 2326 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 2327 ALIGN(address, HPAGE_PMD_SIZE))) 2328 split_huge_pmd_address(vma, address, false, NULL); 2329 } 2330 2331 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2332 unsigned long start, 2333 unsigned long end, 2334 long adjust_next) 2335 { 2336 /* Check if we need to split start first. */ 2337 split_huge_pmd_if_needed(vma, start); 2338 2339 /* Check if we need to split end next. */ 2340 split_huge_pmd_if_needed(vma, end); 2341 2342 /* 2343 * If we're also updating the vma->vm_next->vm_start, 2344 * check if we need to split it. 2345 */ 2346 if (adjust_next > 0) { 2347 struct vm_area_struct *next = vma->vm_next; 2348 unsigned long nstart = next->vm_start; 2349 nstart += adjust_next; 2350 split_huge_pmd_if_needed(next, nstart); 2351 } 2352 } 2353 2354 static void unmap_page(struct page *page) 2355 { 2356 struct folio *folio = page_folio(page); 2357 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | 2358 TTU_SYNC; 2359 2360 VM_BUG_ON_PAGE(!PageHead(page), page); 2361 2362 /* 2363 * Anon pages need migration entries to preserve them, but file 2364 * pages can simply be left unmapped, then faulted back on demand. 2365 * If that is ever changed (perhaps for mlock), update remap_page(). 2366 */ 2367 if (folio_test_anon(folio)) 2368 try_to_migrate(folio, ttu_flags); 2369 else 2370 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK); 2371 } 2372 2373 static void remap_page(struct folio *folio, unsigned long nr) 2374 { 2375 int i = 0; 2376 2377 /* If unmap_page() uses try_to_migrate() on file, remove this check */ 2378 if (!folio_test_anon(folio)) 2379 return; 2380 for (;;) { 2381 remove_migration_ptes(folio, folio, true); 2382 i += folio_nr_pages(folio); 2383 if (i >= nr) 2384 break; 2385 folio = folio_next(folio); 2386 } 2387 } 2388 2389 static void lru_add_page_tail(struct page *head, struct page *tail, 2390 struct lruvec *lruvec, struct list_head *list) 2391 { 2392 VM_BUG_ON_PAGE(!PageHead(head), head); 2393 VM_BUG_ON_PAGE(PageCompound(tail), head); 2394 VM_BUG_ON_PAGE(PageLRU(tail), head); 2395 lockdep_assert_held(&lruvec->lru_lock); 2396 2397 if (list) { 2398 /* page reclaim is reclaiming a huge page */ 2399 VM_WARN_ON(PageLRU(head)); 2400 get_page(tail); 2401 list_add_tail(&tail->lru, list); 2402 } else { 2403 /* head is still on lru (and we have it frozen) */ 2404 VM_WARN_ON(!PageLRU(head)); 2405 if (PageUnevictable(tail)) 2406 tail->mlock_count = 0; 2407 else 2408 list_add_tail(&tail->lru, &head->lru); 2409 SetPageLRU(tail); 2410 } 2411 } 2412 2413 static void __split_huge_page_tail(struct page *head, int tail, 2414 struct lruvec *lruvec, struct list_head *list) 2415 { 2416 struct page *page_tail = head + tail; 2417 2418 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2419 2420 /* 2421 * Clone page flags before unfreezing refcount. 2422 * 2423 * After successful get_page_unless_zero() might follow flags change, 2424 * for example lock_page() which set PG_waiters. 2425 * 2426 * Note that for mapped sub-pages of an anonymous THP, 2427 * PG_anon_exclusive has been cleared in unmap_page() and is stored in 2428 * the migration entry instead from where remap_page() will restore it. 2429 * We can still have PG_anon_exclusive set on effectively unmapped and 2430 * unreferenced sub-pages of an anonymous THP: we can simply drop 2431 * PG_anon_exclusive (-> PG_mappedtodisk) for these here. 2432 */ 2433 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2434 page_tail->flags |= (head->flags & 2435 ((1L << PG_referenced) | 2436 (1L << PG_swapbacked) | 2437 (1L << PG_swapcache) | 2438 (1L << PG_mlocked) | 2439 (1L << PG_uptodate) | 2440 (1L << PG_active) | 2441 (1L << PG_workingset) | 2442 (1L << PG_locked) | 2443 (1L << PG_unevictable) | 2444 #ifdef CONFIG_64BIT 2445 (1L << PG_arch_2) | 2446 #endif 2447 (1L << PG_dirty))); 2448 2449 /* ->mapping in first tail page is compound_mapcount */ 2450 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2451 page_tail); 2452 page_tail->mapping = head->mapping; 2453 page_tail->index = head->index + tail; 2454 page_tail->private = 0; 2455 2456 /* Page flags must be visible before we make the page non-compound. */ 2457 smp_wmb(); 2458 2459 /* 2460 * Clear PageTail before unfreezing page refcount. 2461 * 2462 * After successful get_page_unless_zero() might follow put_page() 2463 * which needs correct compound_head(). 2464 */ 2465 clear_compound_head(page_tail); 2466 2467 /* Finally unfreeze refcount. Additional reference from page cache. */ 2468 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2469 PageSwapCache(head))); 2470 2471 if (page_is_young(head)) 2472 set_page_young(page_tail); 2473 if (page_is_idle(head)) 2474 set_page_idle(page_tail); 2475 2476 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2477 2478 /* 2479 * always add to the tail because some iterators expect new 2480 * pages to show after the currently processed elements - e.g. 2481 * migrate_pages 2482 */ 2483 lru_add_page_tail(head, page_tail, lruvec, list); 2484 } 2485 2486 static void __split_huge_page(struct page *page, struct list_head *list, 2487 pgoff_t end) 2488 { 2489 struct folio *folio = page_folio(page); 2490 struct page *head = &folio->page; 2491 struct lruvec *lruvec; 2492 struct address_space *swap_cache = NULL; 2493 unsigned long offset = 0; 2494 unsigned int nr = thp_nr_pages(head); 2495 int i; 2496 2497 /* complete memcg works before add pages to LRU */ 2498 split_page_memcg(head, nr); 2499 2500 if (PageAnon(head) && PageSwapCache(head)) { 2501 swp_entry_t entry = { .val = page_private(head) }; 2502 2503 offset = swp_offset(entry); 2504 swap_cache = swap_address_space(entry); 2505 xa_lock(&swap_cache->i_pages); 2506 } 2507 2508 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 2509 lruvec = folio_lruvec_lock(folio); 2510 2511 ClearPageHasHWPoisoned(head); 2512 2513 for (i = nr - 1; i >= 1; i--) { 2514 __split_huge_page_tail(head, i, lruvec, list); 2515 /* Some pages can be beyond EOF: drop them from page cache */ 2516 if (head[i].index >= end) { 2517 struct folio *tail = page_folio(head + i); 2518 2519 if (shmem_mapping(head->mapping)) 2520 shmem_uncharge(head->mapping->host, 1); 2521 else if (folio_test_clear_dirty(tail)) 2522 folio_account_cleaned(tail, 2523 inode_to_wb(folio->mapping->host)); 2524 __filemap_remove_folio(tail, NULL); 2525 folio_put(tail); 2526 } else if (!PageAnon(page)) { 2527 __xa_store(&head->mapping->i_pages, head[i].index, 2528 head + i, 0); 2529 } else if (swap_cache) { 2530 __xa_store(&swap_cache->i_pages, offset + i, 2531 head + i, 0); 2532 } 2533 } 2534 2535 ClearPageCompound(head); 2536 unlock_page_lruvec(lruvec); 2537 /* Caller disabled irqs, so they are still disabled here */ 2538 2539 split_page_owner(head, nr); 2540 2541 /* See comment in __split_huge_page_tail() */ 2542 if (PageAnon(head)) { 2543 /* Additional pin to swap cache */ 2544 if (PageSwapCache(head)) { 2545 page_ref_add(head, 2); 2546 xa_unlock(&swap_cache->i_pages); 2547 } else { 2548 page_ref_inc(head); 2549 } 2550 } else { 2551 /* Additional pin to page cache */ 2552 page_ref_add(head, 2); 2553 xa_unlock(&head->mapping->i_pages); 2554 } 2555 local_irq_enable(); 2556 2557 remap_page(folio, nr); 2558 2559 if (PageSwapCache(head)) { 2560 swp_entry_t entry = { .val = page_private(head) }; 2561 2562 split_swap_cluster(entry); 2563 } 2564 2565 for (i = 0; i < nr; i++) { 2566 struct page *subpage = head + i; 2567 if (subpage == page) 2568 continue; 2569 unlock_page(subpage); 2570 2571 /* 2572 * Subpages may be freed if there wasn't any mapping 2573 * like if add_to_swap() is running on a lru page that 2574 * had its mapping zapped. And freeing these pages 2575 * requires taking the lru_lock so we do the put_page 2576 * of the tail pages after the split is complete. 2577 */ 2578 free_page_and_swap_cache(subpage); 2579 } 2580 } 2581 2582 /* Racy check whether the huge page can be split */ 2583 bool can_split_folio(struct folio *folio, int *pextra_pins) 2584 { 2585 int extra_pins; 2586 2587 /* Additional pins from page cache */ 2588 if (folio_test_anon(folio)) 2589 extra_pins = folio_test_swapcache(folio) ? 2590 folio_nr_pages(folio) : 0; 2591 else 2592 extra_pins = folio_nr_pages(folio); 2593 if (pextra_pins) 2594 *pextra_pins = extra_pins; 2595 return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1; 2596 } 2597 2598 /* 2599 * This function splits huge page into normal pages. @page can point to any 2600 * subpage of huge page to split. Split doesn't change the position of @page. 2601 * 2602 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2603 * The huge page must be locked. 2604 * 2605 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2606 * 2607 * Both head page and tail pages will inherit mapping, flags, and so on from 2608 * the hugepage. 2609 * 2610 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2611 * they are not mapped. 2612 * 2613 * Returns 0 if the hugepage is split successfully. 2614 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2615 * us. 2616 */ 2617 int split_huge_page_to_list(struct page *page, struct list_head *list) 2618 { 2619 struct folio *folio = page_folio(page); 2620 struct page *head = &folio->page; 2621 struct deferred_split *ds_queue = get_deferred_split_queue(head); 2622 XA_STATE(xas, &head->mapping->i_pages, head->index); 2623 struct anon_vma *anon_vma = NULL; 2624 struct address_space *mapping = NULL; 2625 int extra_pins, ret; 2626 pgoff_t end; 2627 bool is_hzp; 2628 2629 VM_BUG_ON_PAGE(!PageLocked(head), head); 2630 VM_BUG_ON_PAGE(!PageCompound(head), head); 2631 2632 is_hzp = is_huge_zero_page(head); 2633 VM_WARN_ON_ONCE_PAGE(is_hzp, head); 2634 if (is_hzp) 2635 return -EBUSY; 2636 2637 if (PageWriteback(head)) 2638 return -EBUSY; 2639 2640 if (PageAnon(head)) { 2641 /* 2642 * The caller does not necessarily hold an mmap_lock that would 2643 * prevent the anon_vma disappearing so we first we take a 2644 * reference to it and then lock the anon_vma for write. This 2645 * is similar to folio_lock_anon_vma_read except the write lock 2646 * is taken to serialise against parallel split or collapse 2647 * operations. 2648 */ 2649 anon_vma = page_get_anon_vma(head); 2650 if (!anon_vma) { 2651 ret = -EBUSY; 2652 goto out; 2653 } 2654 end = -1; 2655 mapping = NULL; 2656 anon_vma_lock_write(anon_vma); 2657 } else { 2658 gfp_t gfp; 2659 2660 mapping = head->mapping; 2661 2662 /* Truncated ? */ 2663 if (!mapping) { 2664 ret = -EBUSY; 2665 goto out; 2666 } 2667 2668 gfp = current_gfp_context(mapping_gfp_mask(mapping) & 2669 GFP_RECLAIM_MASK); 2670 2671 if (folio_test_private(folio) && 2672 !filemap_release_folio(folio, gfp)) { 2673 ret = -EBUSY; 2674 goto out; 2675 } 2676 2677 xas_split_alloc(&xas, head, compound_order(head), gfp); 2678 if (xas_error(&xas)) { 2679 ret = xas_error(&xas); 2680 goto out; 2681 } 2682 2683 anon_vma = NULL; 2684 i_mmap_lock_read(mapping); 2685 2686 /* 2687 *__split_huge_page() may need to trim off pages beyond EOF: 2688 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2689 * which cannot be nested inside the page tree lock. So note 2690 * end now: i_size itself may be changed at any moment, but 2691 * head page lock is good enough to serialize the trimming. 2692 */ 2693 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2694 if (shmem_mapping(mapping)) 2695 end = shmem_fallocend(mapping->host, end); 2696 } 2697 2698 /* 2699 * Racy check if we can split the page, before unmap_page() will 2700 * split PMDs 2701 */ 2702 if (!can_split_folio(folio, &extra_pins)) { 2703 ret = -EBUSY; 2704 goto out_unlock; 2705 } 2706 2707 unmap_page(head); 2708 2709 /* block interrupt reentry in xa_lock and spinlock */ 2710 local_irq_disable(); 2711 if (mapping) { 2712 /* 2713 * Check if the head page is present in page cache. 2714 * We assume all tail are present too, if head is there. 2715 */ 2716 xas_lock(&xas); 2717 xas_reset(&xas); 2718 if (xas_load(&xas) != head) 2719 goto fail; 2720 } 2721 2722 /* Prevent deferred_split_scan() touching ->_refcount */ 2723 spin_lock(&ds_queue->split_queue_lock); 2724 if (page_ref_freeze(head, 1 + extra_pins)) { 2725 if (!list_empty(page_deferred_list(head))) { 2726 ds_queue->split_queue_len--; 2727 list_del(page_deferred_list(head)); 2728 } 2729 spin_unlock(&ds_queue->split_queue_lock); 2730 if (mapping) { 2731 int nr = thp_nr_pages(head); 2732 2733 xas_split(&xas, head, thp_order(head)); 2734 if (PageSwapBacked(head)) { 2735 __mod_lruvec_page_state(head, NR_SHMEM_THPS, 2736 -nr); 2737 } else { 2738 __mod_lruvec_page_state(head, NR_FILE_THPS, 2739 -nr); 2740 filemap_nr_thps_dec(mapping); 2741 } 2742 } 2743 2744 __split_huge_page(page, list, end); 2745 ret = 0; 2746 } else { 2747 spin_unlock(&ds_queue->split_queue_lock); 2748 fail: 2749 if (mapping) 2750 xas_unlock(&xas); 2751 local_irq_enable(); 2752 remap_page(folio, folio_nr_pages(folio)); 2753 ret = -EBUSY; 2754 } 2755 2756 out_unlock: 2757 if (anon_vma) { 2758 anon_vma_unlock_write(anon_vma); 2759 put_anon_vma(anon_vma); 2760 } 2761 if (mapping) 2762 i_mmap_unlock_read(mapping); 2763 out: 2764 xas_destroy(&xas); 2765 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2766 return ret; 2767 } 2768 2769 void free_transhuge_page(struct page *page) 2770 { 2771 struct deferred_split *ds_queue = get_deferred_split_queue(page); 2772 unsigned long flags; 2773 2774 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2775 if (!list_empty(page_deferred_list(page))) { 2776 ds_queue->split_queue_len--; 2777 list_del(page_deferred_list(page)); 2778 } 2779 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2780 free_compound_page(page); 2781 } 2782 2783 void deferred_split_huge_page(struct page *page) 2784 { 2785 struct deferred_split *ds_queue = get_deferred_split_queue(page); 2786 #ifdef CONFIG_MEMCG 2787 struct mem_cgroup *memcg = page_memcg(compound_head(page)); 2788 #endif 2789 unsigned long flags; 2790 2791 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2792 2793 /* 2794 * The try_to_unmap() in page reclaim path might reach here too, 2795 * this may cause a race condition to corrupt deferred split queue. 2796 * And, if page reclaim is already handling the same page, it is 2797 * unnecessary to handle it again in shrinker. 2798 * 2799 * Check PageSwapCache to determine if the page is being 2800 * handled by page reclaim since THP swap would add the page into 2801 * swap cache before calling try_to_unmap(). 2802 */ 2803 if (PageSwapCache(page)) 2804 return; 2805 2806 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2807 if (list_empty(page_deferred_list(page))) { 2808 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2809 list_add_tail(page_deferred_list(page), &ds_queue->split_queue); 2810 ds_queue->split_queue_len++; 2811 #ifdef CONFIG_MEMCG 2812 if (memcg) 2813 set_shrinker_bit(memcg, page_to_nid(page), 2814 deferred_split_shrinker.id); 2815 #endif 2816 } 2817 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2818 } 2819 2820 static unsigned long deferred_split_count(struct shrinker *shrink, 2821 struct shrink_control *sc) 2822 { 2823 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2824 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2825 2826 #ifdef CONFIG_MEMCG 2827 if (sc->memcg) 2828 ds_queue = &sc->memcg->deferred_split_queue; 2829 #endif 2830 return READ_ONCE(ds_queue->split_queue_len); 2831 } 2832 2833 static unsigned long deferred_split_scan(struct shrinker *shrink, 2834 struct shrink_control *sc) 2835 { 2836 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2837 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2838 unsigned long flags; 2839 LIST_HEAD(list), *pos, *next; 2840 struct page *page; 2841 int split = 0; 2842 2843 #ifdef CONFIG_MEMCG 2844 if (sc->memcg) 2845 ds_queue = &sc->memcg->deferred_split_queue; 2846 #endif 2847 2848 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2849 /* Take pin on all head pages to avoid freeing them under us */ 2850 list_for_each_safe(pos, next, &ds_queue->split_queue) { 2851 page = list_entry((void *)pos, struct page, deferred_list); 2852 page = compound_head(page); 2853 if (get_page_unless_zero(page)) { 2854 list_move(page_deferred_list(page), &list); 2855 } else { 2856 /* We lost race with put_compound_page() */ 2857 list_del_init(page_deferred_list(page)); 2858 ds_queue->split_queue_len--; 2859 } 2860 if (!--sc->nr_to_scan) 2861 break; 2862 } 2863 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2864 2865 list_for_each_safe(pos, next, &list) { 2866 page = list_entry((void *)pos, struct page, deferred_list); 2867 if (!trylock_page(page)) 2868 goto next; 2869 /* split_huge_page() removes page from list on success */ 2870 if (!split_huge_page(page)) 2871 split++; 2872 unlock_page(page); 2873 next: 2874 put_page(page); 2875 } 2876 2877 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2878 list_splice_tail(&list, &ds_queue->split_queue); 2879 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2880 2881 /* 2882 * Stop shrinker if we didn't split any page, but the queue is empty. 2883 * This can happen if pages were freed under us. 2884 */ 2885 if (!split && list_empty(&ds_queue->split_queue)) 2886 return SHRINK_STOP; 2887 return split; 2888 } 2889 2890 static struct shrinker deferred_split_shrinker = { 2891 .count_objects = deferred_split_count, 2892 .scan_objects = deferred_split_scan, 2893 .seeks = DEFAULT_SEEKS, 2894 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | 2895 SHRINKER_NONSLAB, 2896 }; 2897 2898 #ifdef CONFIG_DEBUG_FS 2899 static void split_huge_pages_all(void) 2900 { 2901 struct zone *zone; 2902 struct page *page; 2903 unsigned long pfn, max_zone_pfn; 2904 unsigned long total = 0, split = 0; 2905 2906 pr_debug("Split all THPs\n"); 2907 for_each_zone(zone) { 2908 if (!managed_zone(zone)) 2909 continue; 2910 max_zone_pfn = zone_end_pfn(zone); 2911 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2912 int nr_pages; 2913 2914 page = pfn_to_online_page(pfn); 2915 if (!page || !get_page_unless_zero(page)) 2916 continue; 2917 2918 if (zone != page_zone(page)) 2919 goto next; 2920 2921 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2922 goto next; 2923 2924 total++; 2925 lock_page(page); 2926 nr_pages = thp_nr_pages(page); 2927 if (!split_huge_page(page)) 2928 split++; 2929 pfn += nr_pages - 1; 2930 unlock_page(page); 2931 next: 2932 put_page(page); 2933 cond_resched(); 2934 } 2935 } 2936 2937 pr_debug("%lu of %lu THP split\n", split, total); 2938 } 2939 2940 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 2941 { 2942 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 2943 is_vm_hugetlb_page(vma); 2944 } 2945 2946 static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 2947 unsigned long vaddr_end) 2948 { 2949 int ret = 0; 2950 struct task_struct *task; 2951 struct mm_struct *mm; 2952 unsigned long total = 0, split = 0; 2953 unsigned long addr; 2954 2955 vaddr_start &= PAGE_MASK; 2956 vaddr_end &= PAGE_MASK; 2957 2958 /* Find the task_struct from pid */ 2959 rcu_read_lock(); 2960 task = find_task_by_vpid(pid); 2961 if (!task) { 2962 rcu_read_unlock(); 2963 ret = -ESRCH; 2964 goto out; 2965 } 2966 get_task_struct(task); 2967 rcu_read_unlock(); 2968 2969 /* Find the mm_struct */ 2970 mm = get_task_mm(task); 2971 put_task_struct(task); 2972 2973 if (!mm) { 2974 ret = -EINVAL; 2975 goto out; 2976 } 2977 2978 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 2979 pid, vaddr_start, vaddr_end); 2980 2981 mmap_read_lock(mm); 2982 /* 2983 * always increase addr by PAGE_SIZE, since we could have a PTE page 2984 * table filled with PTE-mapped THPs, each of which is distinct. 2985 */ 2986 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 2987 struct vm_area_struct *vma = vma_lookup(mm, addr); 2988 struct page *page; 2989 2990 if (!vma) 2991 break; 2992 2993 /* skip special VMA and hugetlb VMA */ 2994 if (vma_not_suitable_for_thp_split(vma)) { 2995 addr = vma->vm_end; 2996 continue; 2997 } 2998 2999 /* FOLL_DUMP to ignore special (like zero) pages */ 3000 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 3001 3002 if (IS_ERR_OR_NULL(page) || is_zone_device_page(page)) 3003 continue; 3004 3005 if (!is_transparent_hugepage(page)) 3006 goto next; 3007 3008 total++; 3009 if (!can_split_folio(page_folio(page), NULL)) 3010 goto next; 3011 3012 if (!trylock_page(page)) 3013 goto next; 3014 3015 if (!split_huge_page(page)) 3016 split++; 3017 3018 unlock_page(page); 3019 next: 3020 put_page(page); 3021 cond_resched(); 3022 } 3023 mmap_read_unlock(mm); 3024 mmput(mm); 3025 3026 pr_debug("%lu of %lu THP split\n", split, total); 3027 3028 out: 3029 return ret; 3030 } 3031 3032 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 3033 pgoff_t off_end) 3034 { 3035 struct filename *file; 3036 struct file *candidate; 3037 struct address_space *mapping; 3038 int ret = -EINVAL; 3039 pgoff_t index; 3040 int nr_pages = 1; 3041 unsigned long total = 0, split = 0; 3042 3043 file = getname_kernel(file_path); 3044 if (IS_ERR(file)) 3045 return ret; 3046 3047 candidate = file_open_name(file, O_RDONLY, 0); 3048 if (IS_ERR(candidate)) 3049 goto out; 3050 3051 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 3052 file_path, off_start, off_end); 3053 3054 mapping = candidate->f_mapping; 3055 3056 for (index = off_start; index < off_end; index += nr_pages) { 3057 struct page *fpage = pagecache_get_page(mapping, index, 3058 FGP_ENTRY | FGP_HEAD, 0); 3059 3060 nr_pages = 1; 3061 if (xa_is_value(fpage) || !fpage) 3062 continue; 3063 3064 if (!is_transparent_hugepage(fpage)) 3065 goto next; 3066 3067 total++; 3068 nr_pages = thp_nr_pages(fpage); 3069 3070 if (!trylock_page(fpage)) 3071 goto next; 3072 3073 if (!split_huge_page(fpage)) 3074 split++; 3075 3076 unlock_page(fpage); 3077 next: 3078 put_page(fpage); 3079 cond_resched(); 3080 } 3081 3082 filp_close(candidate, NULL); 3083 ret = 0; 3084 3085 pr_debug("%lu of %lu file-backed THP split\n", split, total); 3086 out: 3087 putname(file); 3088 return ret; 3089 } 3090 3091 #define MAX_INPUT_BUF_SZ 255 3092 3093 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 3094 size_t count, loff_t *ppops) 3095 { 3096 static DEFINE_MUTEX(split_debug_mutex); 3097 ssize_t ret; 3098 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */ 3099 char input_buf[MAX_INPUT_BUF_SZ]; 3100 int pid; 3101 unsigned long vaddr_start, vaddr_end; 3102 3103 ret = mutex_lock_interruptible(&split_debug_mutex); 3104 if (ret) 3105 return ret; 3106 3107 ret = -EFAULT; 3108 3109 memset(input_buf, 0, MAX_INPUT_BUF_SZ); 3110 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 3111 goto out; 3112 3113 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 3114 3115 if (input_buf[0] == '/') { 3116 char *tok; 3117 char *buf = input_buf; 3118 char file_path[MAX_INPUT_BUF_SZ]; 3119 pgoff_t off_start = 0, off_end = 0; 3120 size_t input_len = strlen(input_buf); 3121 3122 tok = strsep(&buf, ","); 3123 if (tok) { 3124 strcpy(file_path, tok); 3125 } else { 3126 ret = -EINVAL; 3127 goto out; 3128 } 3129 3130 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end); 3131 if (ret != 2) { 3132 ret = -EINVAL; 3133 goto out; 3134 } 3135 ret = split_huge_pages_in_file(file_path, off_start, off_end); 3136 if (!ret) 3137 ret = input_len; 3138 3139 goto out; 3140 } 3141 3142 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end); 3143 if (ret == 1 && pid == 1) { 3144 split_huge_pages_all(); 3145 ret = strlen(input_buf); 3146 goto out; 3147 } else if (ret != 3) { 3148 ret = -EINVAL; 3149 goto out; 3150 } 3151 3152 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end); 3153 if (!ret) 3154 ret = strlen(input_buf); 3155 out: 3156 mutex_unlock(&split_debug_mutex); 3157 return ret; 3158 3159 } 3160 3161 static const struct file_operations split_huge_pages_fops = { 3162 .owner = THIS_MODULE, 3163 .write = split_huge_pages_write, 3164 .llseek = no_llseek, 3165 }; 3166 3167 static int __init split_huge_pages_debugfs(void) 3168 { 3169 debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 3170 &split_huge_pages_fops); 3171 return 0; 3172 } 3173 late_initcall(split_huge_pages_debugfs); 3174 #endif 3175 3176 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 3177 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 3178 struct page *page) 3179 { 3180 struct vm_area_struct *vma = pvmw->vma; 3181 struct mm_struct *mm = vma->vm_mm; 3182 unsigned long address = pvmw->address; 3183 bool anon_exclusive; 3184 pmd_t pmdval; 3185 swp_entry_t entry; 3186 pmd_t pmdswp; 3187 3188 if (!(pvmw->pmd && !pvmw->pte)) 3189 return 0; 3190 3191 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 3192 pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 3193 3194 /* See page_try_share_anon_rmap(): invalidate PMD first. */ 3195 anon_exclusive = PageAnon(page) && PageAnonExclusive(page); 3196 if (anon_exclusive && page_try_share_anon_rmap(page)) { 3197 set_pmd_at(mm, address, pvmw->pmd, pmdval); 3198 return -EBUSY; 3199 } 3200 3201 if (pmd_dirty(pmdval)) 3202 set_page_dirty(page); 3203 if (pmd_write(pmdval)) 3204 entry = make_writable_migration_entry(page_to_pfn(page)); 3205 else if (anon_exclusive) 3206 entry = make_readable_exclusive_migration_entry(page_to_pfn(page)); 3207 else 3208 entry = make_readable_migration_entry(page_to_pfn(page)); 3209 if (pmd_young(pmdval)) 3210 entry = make_migration_entry_young(entry); 3211 if (pmd_dirty(pmdval)) 3212 entry = make_migration_entry_dirty(entry); 3213 pmdswp = swp_entry_to_pmd(entry); 3214 if (pmd_soft_dirty(pmdval)) 3215 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 3216 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 3217 page_remove_rmap(page, vma, true); 3218 put_page(page); 3219 trace_set_migration_pmd(address, pmd_val(pmdswp)); 3220 3221 return 0; 3222 } 3223 3224 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 3225 { 3226 struct vm_area_struct *vma = pvmw->vma; 3227 struct mm_struct *mm = vma->vm_mm; 3228 unsigned long address = pvmw->address; 3229 unsigned long haddr = address & HPAGE_PMD_MASK; 3230 pmd_t pmde; 3231 swp_entry_t entry; 3232 3233 if (!(pvmw->pmd && !pvmw->pte)) 3234 return; 3235 3236 entry = pmd_to_swp_entry(*pvmw->pmd); 3237 get_page(new); 3238 pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)); 3239 if (pmd_swp_soft_dirty(*pvmw->pmd)) 3240 pmde = pmd_mksoft_dirty(pmde); 3241 if (is_writable_migration_entry(entry)) 3242 pmde = maybe_pmd_mkwrite(pmde, vma); 3243 if (pmd_swp_uffd_wp(*pvmw->pmd)) 3244 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde)); 3245 if (!is_migration_entry_young(entry)) 3246 pmde = pmd_mkold(pmde); 3247 /* NOTE: this may contain setting soft-dirty on some archs */ 3248 if (PageDirty(new) && is_migration_entry_dirty(entry)) 3249 pmde = pmd_mkdirty(pmde); 3250 3251 if (PageAnon(new)) { 3252 rmap_t rmap_flags = RMAP_COMPOUND; 3253 3254 if (!is_readable_migration_entry(entry)) 3255 rmap_flags |= RMAP_EXCLUSIVE; 3256 3257 page_add_anon_rmap(new, vma, haddr, rmap_flags); 3258 } else { 3259 page_add_file_rmap(new, vma, true); 3260 } 3261 VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new)); 3262 set_pmd_at(mm, haddr, pvmw->pmd, pmde); 3263 3264 /* No need to invalidate - it was non-present before */ 3265 update_mmu_cache_pmd(vma, address, pvmw->pmd); 3266 trace_remove_migration_pmd(address, pmd_val(pmde)); 3267 } 3268 #endif 3269