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/coredump.h> 11 #include <linux/sched/numa_balancing.h> 12 #include <linux/highmem.h> 13 #include <linux/hugetlb.h> 14 #include <linux/mmu_notifier.h> 15 #include <linux/rmap.h> 16 #include <linux/swap.h> 17 #include <linux/shrinker.h> 18 #include <linux/mm_inline.h> 19 #include <linux/swapops.h> 20 #include <linux/dax.h> 21 #include <linux/khugepaged.h> 22 #include <linux/freezer.h> 23 #include <linux/pfn_t.h> 24 #include <linux/mman.h> 25 #include <linux/memremap.h> 26 #include <linux/pagemap.h> 27 #include <linux/debugfs.h> 28 #include <linux/migrate.h> 29 #include <linux/hashtable.h> 30 #include <linux/userfaultfd_k.h> 31 #include <linux/page_idle.h> 32 #include <linux/shmem_fs.h> 33 #include <linux/oom.h> 34 #include <linux/numa.h> 35 #include <linux/page_owner.h> 36 37 #include <asm/tlb.h> 38 #include <asm/pgalloc.h> 39 #include "internal.h" 40 41 /* 42 * By default, transparent hugepage support is disabled in order to avoid 43 * risking an increased memory footprint for applications that are not 44 * guaranteed to benefit from it. When transparent hugepage support is 45 * enabled, it is for all mappings, and khugepaged scans all mappings. 46 * Defrag is invoked by khugepaged hugepage allocations and by page faults 47 * for all hugepage allocations. 48 */ 49 unsigned long transparent_hugepage_flags __read_mostly = 50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 51 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 52 #endif 53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 55 #endif 56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 59 60 static struct shrinker deferred_split_shrinker; 61 62 static atomic_t huge_zero_refcount; 63 struct page *huge_zero_page __read_mostly; 64 65 bool transparent_hugepage_enabled(struct vm_area_struct *vma) 66 { 67 /* The addr is used to check if the vma size fits */ 68 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE; 69 70 if (!transhuge_vma_suitable(vma, addr)) 71 return false; 72 if (vma_is_anonymous(vma)) 73 return __transparent_hugepage_enabled(vma); 74 if (vma_is_shmem(vma)) 75 return shmem_huge_enabled(vma); 76 77 return false; 78 } 79 80 static struct page *get_huge_zero_page(void) 81 { 82 struct page *zero_page; 83 retry: 84 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 85 return READ_ONCE(huge_zero_page); 86 87 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 88 HPAGE_PMD_ORDER); 89 if (!zero_page) { 90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 91 return NULL; 92 } 93 count_vm_event(THP_ZERO_PAGE_ALLOC); 94 preempt_disable(); 95 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 96 preempt_enable(); 97 __free_pages(zero_page, compound_order(zero_page)); 98 goto retry; 99 } 100 101 /* We take additional reference here. It will be put back by shrinker */ 102 atomic_set(&huge_zero_refcount, 2); 103 preempt_enable(); 104 return READ_ONCE(huge_zero_page); 105 } 106 107 static void put_huge_zero_page(void) 108 { 109 /* 110 * Counter should never go to zero here. Only shrinker can put 111 * last reference. 112 */ 113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 114 } 115 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 117 { 118 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 119 return READ_ONCE(huge_zero_page); 120 121 if (!get_huge_zero_page()) 122 return NULL; 123 124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 125 put_huge_zero_page(); 126 127 return READ_ONCE(huge_zero_page); 128 } 129 130 void mm_put_huge_zero_page(struct mm_struct *mm) 131 { 132 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 133 put_huge_zero_page(); 134 } 135 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 137 struct shrink_control *sc) 138 { 139 /* we can free zero page only if last reference remains */ 140 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 141 } 142 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 144 struct shrink_control *sc) 145 { 146 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 147 struct page *zero_page = xchg(&huge_zero_page, NULL); 148 BUG_ON(zero_page == NULL); 149 __free_pages(zero_page, compound_order(zero_page)); 150 return HPAGE_PMD_NR; 151 } 152 153 return 0; 154 } 155 156 static struct shrinker huge_zero_page_shrinker = { 157 .count_objects = shrink_huge_zero_page_count, 158 .scan_objects = shrink_huge_zero_page_scan, 159 .seeks = DEFAULT_SEEKS, 160 }; 161 162 #ifdef CONFIG_SYSFS 163 static ssize_t enabled_show(struct kobject *kobj, 164 struct kobj_attribute *attr, char *buf) 165 { 166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 167 return sprintf(buf, "[always] madvise never\n"); 168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 169 return sprintf(buf, "always [madvise] never\n"); 170 else 171 return sprintf(buf, "always madvise [never]\n"); 172 } 173 174 static ssize_t enabled_store(struct kobject *kobj, 175 struct kobj_attribute *attr, 176 const char *buf, size_t count) 177 { 178 ssize_t ret = count; 179 180 if (sysfs_streq(buf, "always")) { 181 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 182 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 183 } else if (sysfs_streq(buf, "madvise")) { 184 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 185 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 186 } else if (sysfs_streq(buf, "never")) { 187 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 188 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 189 } else 190 ret = -EINVAL; 191 192 if (ret > 0) { 193 int err = start_stop_khugepaged(); 194 if (err) 195 ret = err; 196 } 197 return ret; 198 } 199 static struct kobj_attribute enabled_attr = 200 __ATTR(enabled, 0644, enabled_show, enabled_store); 201 202 ssize_t single_hugepage_flag_show(struct kobject *kobj, 203 struct kobj_attribute *attr, char *buf, 204 enum transparent_hugepage_flag flag) 205 { 206 return sprintf(buf, "%d\n", 207 !!test_bit(flag, &transparent_hugepage_flags)); 208 } 209 210 ssize_t single_hugepage_flag_store(struct kobject *kobj, 211 struct kobj_attribute *attr, 212 const char *buf, size_t count, 213 enum transparent_hugepage_flag flag) 214 { 215 unsigned long value; 216 int ret; 217 218 ret = kstrtoul(buf, 10, &value); 219 if (ret < 0) 220 return ret; 221 if (value > 1) 222 return -EINVAL; 223 224 if (value) 225 set_bit(flag, &transparent_hugepage_flags); 226 else 227 clear_bit(flag, &transparent_hugepage_flags); 228 229 return count; 230 } 231 232 static ssize_t defrag_show(struct kobject *kobj, 233 struct kobj_attribute *attr, char *buf) 234 { 235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 236 return sprintf(buf, "[always] defer defer+madvise madvise never\n"); 237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 238 return sprintf(buf, "always [defer] defer+madvise madvise never\n"); 239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 240 return sprintf(buf, "always defer [defer+madvise] madvise never\n"); 241 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 242 return sprintf(buf, "always defer defer+madvise [madvise] never\n"); 243 return sprintf(buf, "always defer defer+madvise madvise [never]\n"); 244 } 245 246 static ssize_t defrag_store(struct kobject *kobj, 247 struct kobj_attribute *attr, 248 const char *buf, size_t count) 249 { 250 if (sysfs_streq(buf, "always")) { 251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 255 } else if (sysfs_streq(buf, "defer+madvise")) { 256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 259 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 260 } else if (sysfs_streq(buf, "defer")) { 261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 265 } else if (sysfs_streq(buf, "madvise")) { 266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 269 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 270 } else if (sysfs_streq(buf, "never")) { 271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 272 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 275 } else 276 return -EINVAL; 277 278 return count; 279 } 280 static struct kobj_attribute defrag_attr = 281 __ATTR(defrag, 0644, defrag_show, defrag_store); 282 283 static ssize_t use_zero_page_show(struct kobject *kobj, 284 struct kobj_attribute *attr, char *buf) 285 { 286 return single_hugepage_flag_show(kobj, attr, buf, 287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 288 } 289 static ssize_t use_zero_page_store(struct kobject *kobj, 290 struct kobj_attribute *attr, const char *buf, size_t count) 291 { 292 return single_hugepage_flag_store(kobj, attr, buf, count, 293 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 294 } 295 static struct kobj_attribute use_zero_page_attr = 296 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 297 298 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 299 struct kobj_attribute *attr, char *buf) 300 { 301 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); 302 } 303 static struct kobj_attribute hpage_pmd_size_attr = 304 __ATTR_RO(hpage_pmd_size); 305 306 #ifdef CONFIG_DEBUG_VM 307 static ssize_t debug_cow_show(struct kobject *kobj, 308 struct kobj_attribute *attr, char *buf) 309 { 310 return single_hugepage_flag_show(kobj, attr, buf, 311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 312 } 313 static ssize_t debug_cow_store(struct kobject *kobj, 314 struct kobj_attribute *attr, 315 const char *buf, size_t count) 316 { 317 return single_hugepage_flag_store(kobj, attr, buf, count, 318 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 319 } 320 static struct kobj_attribute debug_cow_attr = 321 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 322 #endif /* CONFIG_DEBUG_VM */ 323 324 static struct attribute *hugepage_attr[] = { 325 &enabled_attr.attr, 326 &defrag_attr.attr, 327 &use_zero_page_attr.attr, 328 &hpage_pmd_size_attr.attr, 329 #ifdef CONFIG_SHMEM 330 &shmem_enabled_attr.attr, 331 #endif 332 #ifdef CONFIG_DEBUG_VM 333 &debug_cow_attr.attr, 334 #endif 335 NULL, 336 }; 337 338 static const struct attribute_group hugepage_attr_group = { 339 .attrs = hugepage_attr, 340 }; 341 342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 343 { 344 int err; 345 346 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 347 if (unlikely(!*hugepage_kobj)) { 348 pr_err("failed to create transparent hugepage kobject\n"); 349 return -ENOMEM; 350 } 351 352 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 353 if (err) { 354 pr_err("failed to register transparent hugepage group\n"); 355 goto delete_obj; 356 } 357 358 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 359 if (err) { 360 pr_err("failed to register transparent hugepage group\n"); 361 goto remove_hp_group; 362 } 363 364 return 0; 365 366 remove_hp_group: 367 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 368 delete_obj: 369 kobject_put(*hugepage_kobj); 370 return err; 371 } 372 373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 374 { 375 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 376 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 377 kobject_put(hugepage_kobj); 378 } 379 #else 380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 381 { 382 return 0; 383 } 384 385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 386 { 387 } 388 #endif /* CONFIG_SYSFS */ 389 390 static int __init hugepage_init(void) 391 { 392 int err; 393 struct kobject *hugepage_kobj; 394 395 if (!has_transparent_hugepage()) { 396 transparent_hugepage_flags = 0; 397 return -EINVAL; 398 } 399 400 /* 401 * hugepages can't be allocated by the buddy allocator 402 */ 403 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 404 /* 405 * we use page->mapping and page->index in second tail page 406 * as list_head: assuming THP order >= 2 407 */ 408 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 409 410 err = hugepage_init_sysfs(&hugepage_kobj); 411 if (err) 412 goto err_sysfs; 413 414 err = khugepaged_init(); 415 if (err) 416 goto err_slab; 417 418 err = register_shrinker(&huge_zero_page_shrinker); 419 if (err) 420 goto err_hzp_shrinker; 421 err = register_shrinker(&deferred_split_shrinker); 422 if (err) 423 goto err_split_shrinker; 424 425 /* 426 * By default disable transparent hugepages on smaller systems, 427 * where the extra memory used could hurt more than TLB overhead 428 * is likely to save. The admin can still enable it through /sys. 429 */ 430 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 431 transparent_hugepage_flags = 0; 432 return 0; 433 } 434 435 err = start_stop_khugepaged(); 436 if (err) 437 goto err_khugepaged; 438 439 return 0; 440 err_khugepaged: 441 unregister_shrinker(&deferred_split_shrinker); 442 err_split_shrinker: 443 unregister_shrinker(&huge_zero_page_shrinker); 444 err_hzp_shrinker: 445 khugepaged_destroy(); 446 err_slab: 447 hugepage_exit_sysfs(hugepage_kobj); 448 err_sysfs: 449 return err; 450 } 451 subsys_initcall(hugepage_init); 452 453 static int __init setup_transparent_hugepage(char *str) 454 { 455 int ret = 0; 456 if (!str) 457 goto out; 458 if (!strcmp(str, "always")) { 459 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 460 &transparent_hugepage_flags); 461 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 462 &transparent_hugepage_flags); 463 ret = 1; 464 } else if (!strcmp(str, "madvise")) { 465 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 466 &transparent_hugepage_flags); 467 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 468 &transparent_hugepage_flags); 469 ret = 1; 470 } else if (!strcmp(str, "never")) { 471 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 472 &transparent_hugepage_flags); 473 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 474 &transparent_hugepage_flags); 475 ret = 1; 476 } 477 out: 478 if (!ret) 479 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 480 return ret; 481 } 482 __setup("transparent_hugepage=", setup_transparent_hugepage); 483 484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 485 { 486 if (likely(vma->vm_flags & VM_WRITE)) 487 pmd = pmd_mkwrite(pmd); 488 return pmd; 489 } 490 491 #ifdef CONFIG_MEMCG 492 static inline struct deferred_split *get_deferred_split_queue(struct page *page) 493 { 494 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup; 495 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 496 497 if (memcg) 498 return &memcg->deferred_split_queue; 499 else 500 return &pgdat->deferred_split_queue; 501 } 502 #else 503 static inline struct deferred_split *get_deferred_split_queue(struct page *page) 504 { 505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); 506 507 return &pgdat->deferred_split_queue; 508 } 509 #endif 510 511 void prep_transhuge_page(struct page *page) 512 { 513 /* 514 * we use page->mapping and page->indexlru in second tail page 515 * as list_head: assuming THP order >= 2 516 */ 517 518 INIT_LIST_HEAD(page_deferred_list(page)); 519 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 520 } 521 522 bool is_transparent_hugepage(struct page *page) 523 { 524 if (!PageCompound(page)) 525 return false; 526 527 page = compound_head(page); 528 return is_huge_zero_page(page) || 529 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; 530 } 531 EXPORT_SYMBOL_GPL(is_transparent_hugepage); 532 533 static unsigned long __thp_get_unmapped_area(struct file *filp, 534 unsigned long addr, unsigned long len, 535 loff_t off, unsigned long flags, unsigned long size) 536 { 537 loff_t off_end = off + len; 538 loff_t off_align = round_up(off, size); 539 unsigned long len_pad, ret; 540 541 if (off_end <= off_align || (off_end - off_align) < size) 542 return 0; 543 544 len_pad = len + size; 545 if (len_pad < len || (off + len_pad) < off) 546 return 0; 547 548 ret = current->mm->get_unmapped_area(filp, addr, len_pad, 549 off >> PAGE_SHIFT, flags); 550 551 /* 552 * The failure might be due to length padding. The caller will retry 553 * without the padding. 554 */ 555 if (IS_ERR_VALUE(ret)) 556 return 0; 557 558 /* 559 * Do not try to align to THP boundary if allocation at the address 560 * hint succeeds. 561 */ 562 if (ret == addr) 563 return addr; 564 565 ret += (off - ret) & (size - 1); 566 return ret; 567 } 568 569 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 570 unsigned long len, unsigned long pgoff, unsigned long flags) 571 { 572 unsigned long ret; 573 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 574 575 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 576 goto out; 577 578 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 579 if (ret) 580 return ret; 581 out: 582 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 583 } 584 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 585 586 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 587 struct page *page, gfp_t gfp) 588 { 589 struct vm_area_struct *vma = vmf->vma; 590 pgtable_t pgtable; 591 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 592 vm_fault_t ret = 0; 593 594 VM_BUG_ON_PAGE(!PageCompound(page), page); 595 596 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) { 597 put_page(page); 598 count_vm_event(THP_FAULT_FALLBACK); 599 count_vm_event(THP_FAULT_FALLBACK_CHARGE); 600 return VM_FAULT_FALLBACK; 601 } 602 cgroup_throttle_swaprate(page, gfp); 603 604 pgtable = pte_alloc_one(vma->vm_mm); 605 if (unlikely(!pgtable)) { 606 ret = VM_FAULT_OOM; 607 goto release; 608 } 609 610 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 611 /* 612 * The memory barrier inside __SetPageUptodate makes sure that 613 * clear_huge_page writes become visible before the set_pmd_at() 614 * write. 615 */ 616 __SetPageUptodate(page); 617 618 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 619 if (unlikely(!pmd_none(*vmf->pmd))) { 620 goto unlock_release; 621 } else { 622 pmd_t entry; 623 624 ret = check_stable_address_space(vma->vm_mm); 625 if (ret) 626 goto unlock_release; 627 628 /* Deliver the page fault to userland */ 629 if (userfaultfd_missing(vma)) { 630 vm_fault_t ret2; 631 632 spin_unlock(vmf->ptl); 633 put_page(page); 634 pte_free(vma->vm_mm, pgtable); 635 ret2 = handle_userfault(vmf, VM_UFFD_MISSING); 636 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK); 637 return ret2; 638 } 639 640 entry = mk_huge_pmd(page, vma->vm_page_prot); 641 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 642 page_add_new_anon_rmap(page, vma, haddr, true); 643 lru_cache_add_active_or_unevictable(page, vma); 644 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 645 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 646 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 647 mm_inc_nr_ptes(vma->vm_mm); 648 spin_unlock(vmf->ptl); 649 count_vm_event(THP_FAULT_ALLOC); 650 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 651 } 652 653 return 0; 654 unlock_release: 655 spin_unlock(vmf->ptl); 656 release: 657 if (pgtable) 658 pte_free(vma->vm_mm, pgtable); 659 put_page(page); 660 return ret; 661 662 } 663 664 /* 665 * always: directly stall for all thp allocations 666 * defer: wake kswapd and fail if not immediately available 667 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 668 * fail if not immediately available 669 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 670 * available 671 * never: never stall for any thp allocation 672 */ 673 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 674 { 675 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 676 677 /* Always do synchronous compaction */ 678 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 679 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 680 681 /* Kick kcompactd and fail quickly */ 682 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 683 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 684 685 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 686 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 687 return GFP_TRANSHUGE_LIGHT | 688 (vma_madvised ? __GFP_DIRECT_RECLAIM : 689 __GFP_KSWAPD_RECLAIM); 690 691 /* Only do synchronous compaction if madvised */ 692 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 693 return GFP_TRANSHUGE_LIGHT | 694 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 695 696 return GFP_TRANSHUGE_LIGHT; 697 } 698 699 /* Caller must hold page table lock. */ 700 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 701 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 702 struct page *zero_page) 703 { 704 pmd_t entry; 705 if (!pmd_none(*pmd)) 706 return false; 707 entry = mk_pmd(zero_page, vma->vm_page_prot); 708 entry = pmd_mkhuge(entry); 709 if (pgtable) 710 pgtable_trans_huge_deposit(mm, pmd, pgtable); 711 set_pmd_at(mm, haddr, pmd, entry); 712 mm_inc_nr_ptes(mm); 713 return true; 714 } 715 716 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 717 { 718 struct vm_area_struct *vma = vmf->vma; 719 gfp_t gfp; 720 struct page *page; 721 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 722 723 if (!transhuge_vma_suitable(vma, haddr)) 724 return VM_FAULT_FALLBACK; 725 if (unlikely(anon_vma_prepare(vma))) 726 return VM_FAULT_OOM; 727 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 728 return VM_FAULT_OOM; 729 if (!(vmf->flags & FAULT_FLAG_WRITE) && 730 !mm_forbids_zeropage(vma->vm_mm) && 731 transparent_hugepage_use_zero_page()) { 732 pgtable_t pgtable; 733 struct page *zero_page; 734 bool set; 735 vm_fault_t ret; 736 pgtable = pte_alloc_one(vma->vm_mm); 737 if (unlikely(!pgtable)) 738 return VM_FAULT_OOM; 739 zero_page = mm_get_huge_zero_page(vma->vm_mm); 740 if (unlikely(!zero_page)) { 741 pte_free(vma->vm_mm, pgtable); 742 count_vm_event(THP_FAULT_FALLBACK); 743 return VM_FAULT_FALLBACK; 744 } 745 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 746 ret = 0; 747 set = false; 748 if (pmd_none(*vmf->pmd)) { 749 ret = check_stable_address_space(vma->vm_mm); 750 if (ret) { 751 spin_unlock(vmf->ptl); 752 } else if (userfaultfd_missing(vma)) { 753 spin_unlock(vmf->ptl); 754 ret = handle_userfault(vmf, VM_UFFD_MISSING); 755 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 756 } else { 757 set_huge_zero_page(pgtable, vma->vm_mm, vma, 758 haddr, vmf->pmd, zero_page); 759 spin_unlock(vmf->ptl); 760 set = true; 761 } 762 } else 763 spin_unlock(vmf->ptl); 764 if (!set) 765 pte_free(vma->vm_mm, pgtable); 766 return ret; 767 } 768 gfp = alloc_hugepage_direct_gfpmask(vma); 769 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 770 if (unlikely(!page)) { 771 count_vm_event(THP_FAULT_FALLBACK); 772 return VM_FAULT_FALLBACK; 773 } 774 prep_transhuge_page(page); 775 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 776 } 777 778 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 779 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 780 pgtable_t pgtable) 781 { 782 struct mm_struct *mm = vma->vm_mm; 783 pmd_t entry; 784 spinlock_t *ptl; 785 786 ptl = pmd_lock(mm, pmd); 787 if (!pmd_none(*pmd)) { 788 if (write) { 789 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 790 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 791 goto out_unlock; 792 } 793 entry = pmd_mkyoung(*pmd); 794 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 795 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 796 update_mmu_cache_pmd(vma, addr, pmd); 797 } 798 799 goto out_unlock; 800 } 801 802 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 803 if (pfn_t_devmap(pfn)) 804 entry = pmd_mkdevmap(entry); 805 if (write) { 806 entry = pmd_mkyoung(pmd_mkdirty(entry)); 807 entry = maybe_pmd_mkwrite(entry, vma); 808 } 809 810 if (pgtable) { 811 pgtable_trans_huge_deposit(mm, pmd, pgtable); 812 mm_inc_nr_ptes(mm); 813 pgtable = NULL; 814 } 815 816 set_pmd_at(mm, addr, pmd, entry); 817 update_mmu_cache_pmd(vma, addr, pmd); 818 819 out_unlock: 820 spin_unlock(ptl); 821 if (pgtable) 822 pte_free(mm, pgtable); 823 } 824 825 /** 826 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn 827 * @vmf: Structure describing the fault 828 * @pfn: pfn to insert 829 * @pgprot: page protection to use 830 * @write: whether it's a write fault 831 * 832 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and 833 * also consult the vmf_insert_mixed_prot() documentation when 834 * @pgprot != @vmf->vma->vm_page_prot. 835 * 836 * Return: vm_fault_t value. 837 */ 838 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, 839 pgprot_t pgprot, bool write) 840 { 841 unsigned long addr = vmf->address & PMD_MASK; 842 struct vm_area_struct *vma = vmf->vma; 843 pgtable_t pgtable = NULL; 844 845 /* 846 * If we had pmd_special, we could avoid all these restrictions, 847 * but we need to be consistent with PTEs and architectures that 848 * can't support a 'special' bit. 849 */ 850 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 851 !pfn_t_devmap(pfn)); 852 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 853 (VM_PFNMAP|VM_MIXEDMAP)); 854 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 855 856 if (addr < vma->vm_start || addr >= vma->vm_end) 857 return VM_FAULT_SIGBUS; 858 859 if (arch_needs_pgtable_deposit()) { 860 pgtable = pte_alloc_one(vma->vm_mm); 861 if (!pgtable) 862 return VM_FAULT_OOM; 863 } 864 865 track_pfn_insert(vma, &pgprot, pfn); 866 867 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 868 return VM_FAULT_NOPAGE; 869 } 870 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); 871 872 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 873 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 874 { 875 if (likely(vma->vm_flags & VM_WRITE)) 876 pud = pud_mkwrite(pud); 877 return pud; 878 } 879 880 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 881 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 882 { 883 struct mm_struct *mm = vma->vm_mm; 884 pud_t entry; 885 spinlock_t *ptl; 886 887 ptl = pud_lock(mm, pud); 888 if (!pud_none(*pud)) { 889 if (write) { 890 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 891 WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 892 goto out_unlock; 893 } 894 entry = pud_mkyoung(*pud); 895 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 896 if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 897 update_mmu_cache_pud(vma, addr, pud); 898 } 899 goto out_unlock; 900 } 901 902 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 903 if (pfn_t_devmap(pfn)) 904 entry = pud_mkdevmap(entry); 905 if (write) { 906 entry = pud_mkyoung(pud_mkdirty(entry)); 907 entry = maybe_pud_mkwrite(entry, vma); 908 } 909 set_pud_at(mm, addr, pud, entry); 910 update_mmu_cache_pud(vma, addr, pud); 911 912 out_unlock: 913 spin_unlock(ptl); 914 } 915 916 /** 917 * vmf_insert_pfn_pud_prot - insert a pud size pfn 918 * @vmf: Structure describing the fault 919 * @pfn: pfn to insert 920 * @pgprot: page protection to use 921 * @write: whether it's a write fault 922 * 923 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and 924 * also consult the vmf_insert_mixed_prot() documentation when 925 * @pgprot != @vmf->vma->vm_page_prot. 926 * 927 * Return: vm_fault_t value. 928 */ 929 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, 930 pgprot_t pgprot, bool write) 931 { 932 unsigned long addr = vmf->address & PUD_MASK; 933 struct vm_area_struct *vma = vmf->vma; 934 935 /* 936 * If we had pud_special, we could avoid all these restrictions, 937 * but we need to be consistent with PTEs and architectures that 938 * can't support a 'special' bit. 939 */ 940 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 941 !pfn_t_devmap(pfn)); 942 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 943 (VM_PFNMAP|VM_MIXEDMAP)); 944 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 945 946 if (addr < vma->vm_start || addr >= vma->vm_end) 947 return VM_FAULT_SIGBUS; 948 949 track_pfn_insert(vma, &pgprot, pfn); 950 951 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); 952 return VM_FAULT_NOPAGE; 953 } 954 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); 955 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 956 957 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 958 pmd_t *pmd, int flags) 959 { 960 pmd_t _pmd; 961 962 _pmd = pmd_mkyoung(*pmd); 963 if (flags & FOLL_WRITE) 964 _pmd = pmd_mkdirty(_pmd); 965 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 966 pmd, _pmd, flags & FOLL_WRITE)) 967 update_mmu_cache_pmd(vma, addr, pmd); 968 } 969 970 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 971 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 972 { 973 unsigned long pfn = pmd_pfn(*pmd); 974 struct mm_struct *mm = vma->vm_mm; 975 struct page *page; 976 977 assert_spin_locked(pmd_lockptr(mm, pmd)); 978 979 /* 980 * When we COW a devmap PMD entry, we split it into PTEs, so we should 981 * not be in this function with `flags & FOLL_COW` set. 982 */ 983 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 984 985 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 986 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 987 (FOLL_PIN | FOLL_GET))) 988 return NULL; 989 990 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 991 return NULL; 992 993 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 994 /* pass */; 995 else 996 return NULL; 997 998 if (flags & FOLL_TOUCH) 999 touch_pmd(vma, addr, pmd, flags); 1000 1001 /* 1002 * device mapped pages can only be returned if the 1003 * caller will manage the page reference count. 1004 */ 1005 if (!(flags & (FOLL_GET | FOLL_PIN))) 1006 return ERR_PTR(-EEXIST); 1007 1008 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1009 *pgmap = get_dev_pagemap(pfn, *pgmap); 1010 if (!*pgmap) 1011 return ERR_PTR(-EFAULT); 1012 page = pfn_to_page(pfn); 1013 if (!try_grab_page(page, flags)) 1014 page = ERR_PTR(-ENOMEM); 1015 1016 return page; 1017 } 1018 1019 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1020 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1021 struct vm_area_struct *vma) 1022 { 1023 spinlock_t *dst_ptl, *src_ptl; 1024 struct page *src_page; 1025 pmd_t pmd; 1026 pgtable_t pgtable = NULL; 1027 int ret = -ENOMEM; 1028 1029 /* Skip if can be re-fill on fault */ 1030 if (!vma_is_anonymous(vma)) 1031 return 0; 1032 1033 pgtable = pte_alloc_one(dst_mm); 1034 if (unlikely(!pgtable)) 1035 goto out; 1036 1037 dst_ptl = pmd_lock(dst_mm, dst_pmd); 1038 src_ptl = pmd_lockptr(src_mm, src_pmd); 1039 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1040 1041 ret = -EAGAIN; 1042 pmd = *src_pmd; 1043 1044 /* 1045 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA 1046 * does not have the VM_UFFD_WP, which means that the uffd 1047 * fork event is not enabled. 1048 */ 1049 if (!(vma->vm_flags & VM_UFFD_WP)) 1050 pmd = pmd_clear_uffd_wp(pmd); 1051 1052 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1053 if (unlikely(is_swap_pmd(pmd))) { 1054 swp_entry_t entry = pmd_to_swp_entry(pmd); 1055 1056 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 1057 if (is_write_migration_entry(entry)) { 1058 make_migration_entry_read(&entry); 1059 pmd = swp_entry_to_pmd(entry); 1060 if (pmd_swp_soft_dirty(*src_pmd)) 1061 pmd = pmd_swp_mksoft_dirty(pmd); 1062 set_pmd_at(src_mm, addr, src_pmd, pmd); 1063 } 1064 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1065 mm_inc_nr_ptes(dst_mm); 1066 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1067 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1068 ret = 0; 1069 goto out_unlock; 1070 } 1071 #endif 1072 1073 if (unlikely(!pmd_trans_huge(pmd))) { 1074 pte_free(dst_mm, pgtable); 1075 goto out_unlock; 1076 } 1077 /* 1078 * When page table lock is held, the huge zero pmd should not be 1079 * under splitting since we don't split the page itself, only pmd to 1080 * a page table. 1081 */ 1082 if (is_huge_zero_pmd(pmd)) { 1083 struct page *zero_page; 1084 /* 1085 * get_huge_zero_page() will never allocate a new page here, 1086 * since we already have a zero page to copy. It just takes a 1087 * reference. 1088 */ 1089 zero_page = mm_get_huge_zero_page(dst_mm); 1090 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 1091 zero_page); 1092 ret = 0; 1093 goto out_unlock; 1094 } 1095 1096 src_page = pmd_page(pmd); 1097 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 1098 get_page(src_page); 1099 page_dup_rmap(src_page, true); 1100 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1101 mm_inc_nr_ptes(dst_mm); 1102 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 1103 1104 pmdp_set_wrprotect(src_mm, addr, src_pmd); 1105 pmd = pmd_mkold(pmd_wrprotect(pmd)); 1106 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 1107 1108 ret = 0; 1109 out_unlock: 1110 spin_unlock(src_ptl); 1111 spin_unlock(dst_ptl); 1112 out: 1113 return ret; 1114 } 1115 1116 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1117 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1118 pud_t *pud, int flags) 1119 { 1120 pud_t _pud; 1121 1122 _pud = pud_mkyoung(*pud); 1123 if (flags & FOLL_WRITE) 1124 _pud = pud_mkdirty(_pud); 1125 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1126 pud, _pud, flags & FOLL_WRITE)) 1127 update_mmu_cache_pud(vma, addr, pud); 1128 } 1129 1130 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1131 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1132 { 1133 unsigned long pfn = pud_pfn(*pud); 1134 struct mm_struct *mm = vma->vm_mm; 1135 struct page *page; 1136 1137 assert_spin_locked(pud_lockptr(mm, pud)); 1138 1139 if (flags & FOLL_WRITE && !pud_write(*pud)) 1140 return NULL; 1141 1142 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 1143 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 1144 (FOLL_PIN | FOLL_GET))) 1145 return NULL; 1146 1147 if (pud_present(*pud) && pud_devmap(*pud)) 1148 /* pass */; 1149 else 1150 return NULL; 1151 1152 if (flags & FOLL_TOUCH) 1153 touch_pud(vma, addr, pud, flags); 1154 1155 /* 1156 * device mapped pages can only be returned if the 1157 * caller will manage the page reference count. 1158 * 1159 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1160 */ 1161 if (!(flags & (FOLL_GET | FOLL_PIN))) 1162 return ERR_PTR(-EEXIST); 1163 1164 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1165 *pgmap = get_dev_pagemap(pfn, *pgmap); 1166 if (!*pgmap) 1167 return ERR_PTR(-EFAULT); 1168 page = pfn_to_page(pfn); 1169 if (!try_grab_page(page, flags)) 1170 page = ERR_PTR(-ENOMEM); 1171 1172 return page; 1173 } 1174 1175 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1176 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1177 struct vm_area_struct *vma) 1178 { 1179 spinlock_t *dst_ptl, *src_ptl; 1180 pud_t pud; 1181 int ret; 1182 1183 dst_ptl = pud_lock(dst_mm, dst_pud); 1184 src_ptl = pud_lockptr(src_mm, src_pud); 1185 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1186 1187 ret = -EAGAIN; 1188 pud = *src_pud; 1189 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1190 goto out_unlock; 1191 1192 /* 1193 * When page table lock is held, the huge zero pud should not be 1194 * under splitting since we don't split the page itself, only pud to 1195 * a page table. 1196 */ 1197 if (is_huge_zero_pud(pud)) { 1198 /* No huge zero pud yet */ 1199 } 1200 1201 pudp_set_wrprotect(src_mm, addr, src_pud); 1202 pud = pud_mkold(pud_wrprotect(pud)); 1203 set_pud_at(dst_mm, addr, dst_pud, pud); 1204 1205 ret = 0; 1206 out_unlock: 1207 spin_unlock(src_ptl); 1208 spin_unlock(dst_ptl); 1209 return ret; 1210 } 1211 1212 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1213 { 1214 pud_t entry; 1215 unsigned long haddr; 1216 bool write = vmf->flags & FAULT_FLAG_WRITE; 1217 1218 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1219 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1220 goto unlock; 1221 1222 entry = pud_mkyoung(orig_pud); 1223 if (write) 1224 entry = pud_mkdirty(entry); 1225 haddr = vmf->address & HPAGE_PUD_MASK; 1226 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1227 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1228 1229 unlock: 1230 spin_unlock(vmf->ptl); 1231 } 1232 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1233 1234 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1235 { 1236 pmd_t entry; 1237 unsigned long haddr; 1238 bool write = vmf->flags & FAULT_FLAG_WRITE; 1239 1240 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1241 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1242 goto unlock; 1243 1244 entry = pmd_mkyoung(orig_pmd); 1245 if (write) 1246 entry = pmd_mkdirty(entry); 1247 haddr = vmf->address & HPAGE_PMD_MASK; 1248 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1249 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1250 1251 unlock: 1252 spin_unlock(vmf->ptl); 1253 } 1254 1255 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1256 { 1257 struct vm_area_struct *vma = vmf->vma; 1258 struct page *page; 1259 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1260 1261 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1262 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1263 1264 if (is_huge_zero_pmd(orig_pmd)) 1265 goto fallback; 1266 1267 spin_lock(vmf->ptl); 1268 1269 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1270 spin_unlock(vmf->ptl); 1271 return 0; 1272 } 1273 1274 page = pmd_page(orig_pmd); 1275 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1276 1277 /* Lock page for reuse_swap_page() */ 1278 if (!trylock_page(page)) { 1279 get_page(page); 1280 spin_unlock(vmf->ptl); 1281 lock_page(page); 1282 spin_lock(vmf->ptl); 1283 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1284 spin_unlock(vmf->ptl); 1285 unlock_page(page); 1286 put_page(page); 1287 return 0; 1288 } 1289 put_page(page); 1290 } 1291 1292 /* 1293 * We can only reuse the page if nobody else maps the huge page or it's 1294 * part. 1295 */ 1296 if (reuse_swap_page(page, NULL)) { 1297 pmd_t entry; 1298 entry = pmd_mkyoung(orig_pmd); 1299 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1300 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1301 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1302 unlock_page(page); 1303 spin_unlock(vmf->ptl); 1304 return VM_FAULT_WRITE; 1305 } 1306 1307 unlock_page(page); 1308 spin_unlock(vmf->ptl); 1309 fallback: 1310 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 1311 return VM_FAULT_FALLBACK; 1312 } 1313 1314 /* 1315 * FOLL_FORCE or a forced COW break can write even to unwritable pmd's, 1316 * but only after we've gone through a COW cycle and they are dirty. 1317 */ 1318 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1319 { 1320 return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd)); 1321 } 1322 1323 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1324 unsigned long addr, 1325 pmd_t *pmd, 1326 unsigned int flags) 1327 { 1328 struct mm_struct *mm = vma->vm_mm; 1329 struct page *page = NULL; 1330 1331 assert_spin_locked(pmd_lockptr(mm, pmd)); 1332 1333 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1334 goto out; 1335 1336 /* Avoid dumping huge zero page */ 1337 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1338 return ERR_PTR(-EFAULT); 1339 1340 /* Full NUMA hinting faults to serialise migration in fault paths */ 1341 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1342 goto out; 1343 1344 page = pmd_page(*pmd); 1345 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1346 1347 if (!try_grab_page(page, flags)) 1348 return ERR_PTR(-ENOMEM); 1349 1350 if (flags & FOLL_TOUCH) 1351 touch_pmd(vma, addr, pmd, flags); 1352 1353 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1354 /* 1355 * We don't mlock() pte-mapped THPs. This way we can avoid 1356 * leaking mlocked pages into non-VM_LOCKED VMAs. 1357 * 1358 * For anon THP: 1359 * 1360 * In most cases the pmd is the only mapping of the page as we 1361 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1362 * writable private mappings in populate_vma_page_range(). 1363 * 1364 * The only scenario when we have the page shared here is if we 1365 * mlocking read-only mapping shared over fork(). We skip 1366 * mlocking such pages. 1367 * 1368 * For file THP: 1369 * 1370 * We can expect PageDoubleMap() to be stable under page lock: 1371 * for file pages we set it in page_add_file_rmap(), which 1372 * requires page to be locked. 1373 */ 1374 1375 if (PageAnon(page) && compound_mapcount(page) != 1) 1376 goto skip_mlock; 1377 if (PageDoubleMap(page) || !page->mapping) 1378 goto skip_mlock; 1379 if (!trylock_page(page)) 1380 goto skip_mlock; 1381 if (page->mapping && !PageDoubleMap(page)) 1382 mlock_vma_page(page); 1383 unlock_page(page); 1384 } 1385 skip_mlock: 1386 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1387 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1388 1389 out: 1390 return page; 1391 } 1392 1393 /* NUMA hinting page fault entry point for trans huge pmds */ 1394 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1395 { 1396 struct vm_area_struct *vma = vmf->vma; 1397 struct anon_vma *anon_vma = NULL; 1398 struct page *page; 1399 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1400 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id(); 1401 int target_nid, last_cpupid = -1; 1402 bool page_locked; 1403 bool migrated = false; 1404 bool was_writable; 1405 int flags = 0; 1406 1407 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1408 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1409 goto out_unlock; 1410 1411 /* 1412 * If there are potential migrations, wait for completion and retry 1413 * without disrupting NUMA hinting information. Do not relock and 1414 * check_same as the page may no longer be mapped. 1415 */ 1416 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1417 page = pmd_page(*vmf->pmd); 1418 if (!get_page_unless_zero(page)) 1419 goto out_unlock; 1420 spin_unlock(vmf->ptl); 1421 put_and_wait_on_page_locked(page); 1422 goto out; 1423 } 1424 1425 page = pmd_page(pmd); 1426 BUG_ON(is_huge_zero_page(page)); 1427 page_nid = page_to_nid(page); 1428 last_cpupid = page_cpupid_last(page); 1429 count_vm_numa_event(NUMA_HINT_FAULTS); 1430 if (page_nid == this_nid) { 1431 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1432 flags |= TNF_FAULT_LOCAL; 1433 } 1434 1435 /* See similar comment in do_numa_page for explanation */ 1436 if (!pmd_savedwrite(pmd)) 1437 flags |= TNF_NO_GROUP; 1438 1439 /* 1440 * Acquire the page lock to serialise THP migrations but avoid dropping 1441 * page_table_lock if at all possible 1442 */ 1443 page_locked = trylock_page(page); 1444 target_nid = mpol_misplaced(page, vma, haddr); 1445 if (target_nid == NUMA_NO_NODE) { 1446 /* If the page was locked, there are no parallel migrations */ 1447 if (page_locked) 1448 goto clear_pmdnuma; 1449 } 1450 1451 /* Migration could have started since the pmd_trans_migrating check */ 1452 if (!page_locked) { 1453 page_nid = NUMA_NO_NODE; 1454 if (!get_page_unless_zero(page)) 1455 goto out_unlock; 1456 spin_unlock(vmf->ptl); 1457 put_and_wait_on_page_locked(page); 1458 goto out; 1459 } 1460 1461 /* 1462 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1463 * to serialises splits 1464 */ 1465 get_page(page); 1466 spin_unlock(vmf->ptl); 1467 anon_vma = page_lock_anon_vma_read(page); 1468 1469 /* Confirm the PMD did not change while page_table_lock was released */ 1470 spin_lock(vmf->ptl); 1471 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1472 unlock_page(page); 1473 put_page(page); 1474 page_nid = NUMA_NO_NODE; 1475 goto out_unlock; 1476 } 1477 1478 /* Bail if we fail to protect against THP splits for any reason */ 1479 if (unlikely(!anon_vma)) { 1480 put_page(page); 1481 page_nid = NUMA_NO_NODE; 1482 goto clear_pmdnuma; 1483 } 1484 1485 /* 1486 * Since we took the NUMA fault, we must have observed the !accessible 1487 * bit. Make sure all other CPUs agree with that, to avoid them 1488 * modifying the page we're about to migrate. 1489 * 1490 * Must be done under PTL such that we'll observe the relevant 1491 * inc_tlb_flush_pending(). 1492 * 1493 * We are not sure a pending tlb flush here is for a huge page 1494 * mapping or not. Hence use the tlb range variant 1495 */ 1496 if (mm_tlb_flush_pending(vma->vm_mm)) { 1497 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1498 /* 1499 * change_huge_pmd() released the pmd lock before 1500 * invalidating the secondary MMUs sharing the primary 1501 * MMU pagetables (with ->invalidate_range()). The 1502 * mmu_notifier_invalidate_range_end() (which 1503 * internally calls ->invalidate_range()) in 1504 * change_pmd_range() will run after us, so we can't 1505 * rely on it here and we need an explicit invalidate. 1506 */ 1507 mmu_notifier_invalidate_range(vma->vm_mm, haddr, 1508 haddr + HPAGE_PMD_SIZE); 1509 } 1510 1511 /* 1512 * Migrate the THP to the requested node, returns with page unlocked 1513 * and access rights restored. 1514 */ 1515 spin_unlock(vmf->ptl); 1516 1517 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1518 vmf->pmd, pmd, vmf->address, page, target_nid); 1519 if (migrated) { 1520 flags |= TNF_MIGRATED; 1521 page_nid = target_nid; 1522 } else 1523 flags |= TNF_MIGRATE_FAIL; 1524 1525 goto out; 1526 clear_pmdnuma: 1527 BUG_ON(!PageLocked(page)); 1528 was_writable = pmd_savedwrite(pmd); 1529 pmd = pmd_modify(pmd, vma->vm_page_prot); 1530 pmd = pmd_mkyoung(pmd); 1531 if (was_writable) 1532 pmd = pmd_mkwrite(pmd); 1533 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1534 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1535 unlock_page(page); 1536 out_unlock: 1537 spin_unlock(vmf->ptl); 1538 1539 out: 1540 if (anon_vma) 1541 page_unlock_anon_vma_read(anon_vma); 1542 1543 if (page_nid != NUMA_NO_NODE) 1544 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1545 flags); 1546 1547 return 0; 1548 } 1549 1550 /* 1551 * Return true if we do MADV_FREE successfully on entire pmd page. 1552 * Otherwise, return false. 1553 */ 1554 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1555 pmd_t *pmd, unsigned long addr, unsigned long next) 1556 { 1557 spinlock_t *ptl; 1558 pmd_t orig_pmd; 1559 struct page *page; 1560 struct mm_struct *mm = tlb->mm; 1561 bool ret = false; 1562 1563 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1564 1565 ptl = pmd_trans_huge_lock(pmd, vma); 1566 if (!ptl) 1567 goto out_unlocked; 1568 1569 orig_pmd = *pmd; 1570 if (is_huge_zero_pmd(orig_pmd)) 1571 goto out; 1572 1573 if (unlikely(!pmd_present(orig_pmd))) { 1574 VM_BUG_ON(thp_migration_supported() && 1575 !is_pmd_migration_entry(orig_pmd)); 1576 goto out; 1577 } 1578 1579 page = pmd_page(orig_pmd); 1580 /* 1581 * If other processes are mapping this page, we couldn't discard 1582 * the page unless they all do MADV_FREE so let's skip the page. 1583 */ 1584 if (page_mapcount(page) != 1) 1585 goto out; 1586 1587 if (!trylock_page(page)) 1588 goto out; 1589 1590 /* 1591 * If user want to discard part-pages of THP, split it so MADV_FREE 1592 * will deactivate only them. 1593 */ 1594 if (next - addr != HPAGE_PMD_SIZE) { 1595 get_page(page); 1596 spin_unlock(ptl); 1597 split_huge_page(page); 1598 unlock_page(page); 1599 put_page(page); 1600 goto out_unlocked; 1601 } 1602 1603 if (PageDirty(page)) 1604 ClearPageDirty(page); 1605 unlock_page(page); 1606 1607 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1608 pmdp_invalidate(vma, addr, pmd); 1609 orig_pmd = pmd_mkold(orig_pmd); 1610 orig_pmd = pmd_mkclean(orig_pmd); 1611 1612 set_pmd_at(mm, addr, pmd, orig_pmd); 1613 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1614 } 1615 1616 mark_page_lazyfree(page); 1617 ret = true; 1618 out: 1619 spin_unlock(ptl); 1620 out_unlocked: 1621 return ret; 1622 } 1623 1624 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1625 { 1626 pgtable_t pgtable; 1627 1628 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1629 pte_free(mm, pgtable); 1630 mm_dec_nr_ptes(mm); 1631 } 1632 1633 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1634 pmd_t *pmd, unsigned long addr) 1635 { 1636 pmd_t orig_pmd; 1637 spinlock_t *ptl; 1638 1639 tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 1640 1641 ptl = __pmd_trans_huge_lock(pmd, vma); 1642 if (!ptl) 1643 return 0; 1644 /* 1645 * For architectures like ppc64 we look at deposited pgtable 1646 * when calling pmdp_huge_get_and_clear. So do the 1647 * pgtable_trans_huge_withdraw after finishing pmdp related 1648 * operations. 1649 */ 1650 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1651 tlb->fullmm); 1652 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1653 if (vma_is_special_huge(vma)) { 1654 if (arch_needs_pgtable_deposit()) 1655 zap_deposited_table(tlb->mm, pmd); 1656 spin_unlock(ptl); 1657 if (is_huge_zero_pmd(orig_pmd)) 1658 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1659 } else if (is_huge_zero_pmd(orig_pmd)) { 1660 zap_deposited_table(tlb->mm, pmd); 1661 spin_unlock(ptl); 1662 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1663 } else { 1664 struct page *page = NULL; 1665 int flush_needed = 1; 1666 1667 if (pmd_present(orig_pmd)) { 1668 page = pmd_page(orig_pmd); 1669 page_remove_rmap(page, true); 1670 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1671 VM_BUG_ON_PAGE(!PageHead(page), page); 1672 } else if (thp_migration_supported()) { 1673 swp_entry_t entry; 1674 1675 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1676 entry = pmd_to_swp_entry(orig_pmd); 1677 page = pfn_to_page(swp_offset(entry)); 1678 flush_needed = 0; 1679 } else 1680 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1681 1682 if (PageAnon(page)) { 1683 zap_deposited_table(tlb->mm, pmd); 1684 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1685 } else { 1686 if (arch_needs_pgtable_deposit()) 1687 zap_deposited_table(tlb->mm, pmd); 1688 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1689 } 1690 1691 spin_unlock(ptl); 1692 if (flush_needed) 1693 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1694 } 1695 return 1; 1696 } 1697 1698 #ifndef pmd_move_must_withdraw 1699 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1700 spinlock_t *old_pmd_ptl, 1701 struct vm_area_struct *vma) 1702 { 1703 /* 1704 * With split pmd lock we also need to move preallocated 1705 * PTE page table if new_pmd is on different PMD page table. 1706 * 1707 * We also don't deposit and withdraw tables for file pages. 1708 */ 1709 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1710 } 1711 #endif 1712 1713 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1714 { 1715 #ifdef CONFIG_MEM_SOFT_DIRTY 1716 if (unlikely(is_pmd_migration_entry(pmd))) 1717 pmd = pmd_swp_mksoft_dirty(pmd); 1718 else if (pmd_present(pmd)) 1719 pmd = pmd_mksoft_dirty(pmd); 1720 #endif 1721 return pmd; 1722 } 1723 1724 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1725 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 1726 { 1727 spinlock_t *old_ptl, *new_ptl; 1728 pmd_t pmd; 1729 struct mm_struct *mm = vma->vm_mm; 1730 bool force_flush = false; 1731 1732 /* 1733 * The destination pmd shouldn't be established, free_pgtables() 1734 * should have release it. 1735 */ 1736 if (WARN_ON(!pmd_none(*new_pmd))) { 1737 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1738 return false; 1739 } 1740 1741 /* 1742 * We don't have to worry about the ordering of src and dst 1743 * ptlocks because exclusive mmap_lock prevents deadlock. 1744 */ 1745 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1746 if (old_ptl) { 1747 new_ptl = pmd_lockptr(mm, new_pmd); 1748 if (new_ptl != old_ptl) 1749 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1750 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1751 if (pmd_present(pmd)) 1752 force_flush = true; 1753 VM_BUG_ON(!pmd_none(*new_pmd)); 1754 1755 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1756 pgtable_t pgtable; 1757 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1758 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1759 } 1760 pmd = move_soft_dirty_pmd(pmd); 1761 set_pmd_at(mm, new_addr, new_pmd, pmd); 1762 if (force_flush) 1763 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1764 if (new_ptl != old_ptl) 1765 spin_unlock(new_ptl); 1766 spin_unlock(old_ptl); 1767 return true; 1768 } 1769 return false; 1770 } 1771 1772 /* 1773 * Returns 1774 * - 0 if PMD could not be locked 1775 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1776 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1777 */ 1778 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1779 unsigned long addr, pgprot_t newprot, unsigned long cp_flags) 1780 { 1781 struct mm_struct *mm = vma->vm_mm; 1782 spinlock_t *ptl; 1783 pmd_t entry; 1784 bool preserve_write; 1785 int ret; 1786 bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 1787 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 1788 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 1789 1790 ptl = __pmd_trans_huge_lock(pmd, vma); 1791 if (!ptl) 1792 return 0; 1793 1794 preserve_write = prot_numa && pmd_write(*pmd); 1795 ret = 1; 1796 1797 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1798 if (is_swap_pmd(*pmd)) { 1799 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1800 1801 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1802 if (is_write_migration_entry(entry)) { 1803 pmd_t newpmd; 1804 /* 1805 * A protection check is difficult so 1806 * just be safe and disable write 1807 */ 1808 make_migration_entry_read(&entry); 1809 newpmd = swp_entry_to_pmd(entry); 1810 if (pmd_swp_soft_dirty(*pmd)) 1811 newpmd = pmd_swp_mksoft_dirty(newpmd); 1812 set_pmd_at(mm, addr, pmd, newpmd); 1813 } 1814 goto unlock; 1815 } 1816 #endif 1817 1818 /* 1819 * Avoid trapping faults against the zero page. The read-only 1820 * data is likely to be read-cached on the local CPU and 1821 * local/remote hits to the zero page are not interesting. 1822 */ 1823 if (prot_numa && is_huge_zero_pmd(*pmd)) 1824 goto unlock; 1825 1826 if (prot_numa && pmd_protnone(*pmd)) 1827 goto unlock; 1828 1829 /* 1830 * In case prot_numa, we are under mmap_read_lock(mm). It's critical 1831 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1832 * which is also under mmap_read_lock(mm): 1833 * 1834 * CPU0: CPU1: 1835 * change_huge_pmd(prot_numa=1) 1836 * pmdp_huge_get_and_clear_notify() 1837 * madvise_dontneed() 1838 * zap_pmd_range() 1839 * pmd_trans_huge(*pmd) == 0 (without ptl) 1840 * // skip the pmd 1841 * set_pmd_at(); 1842 * // pmd is re-established 1843 * 1844 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1845 * which may break userspace. 1846 * 1847 * pmdp_invalidate() is required to make sure we don't miss 1848 * dirty/young flags set by hardware. 1849 */ 1850 entry = pmdp_invalidate(vma, addr, pmd); 1851 1852 entry = pmd_modify(entry, newprot); 1853 if (preserve_write) 1854 entry = pmd_mk_savedwrite(entry); 1855 if (uffd_wp) { 1856 entry = pmd_wrprotect(entry); 1857 entry = pmd_mkuffd_wp(entry); 1858 } else if (uffd_wp_resolve) { 1859 /* 1860 * Leave the write bit to be handled by PF interrupt 1861 * handler, then things like COW could be properly 1862 * handled. 1863 */ 1864 entry = pmd_clear_uffd_wp(entry); 1865 } 1866 ret = HPAGE_PMD_NR; 1867 set_pmd_at(mm, addr, pmd, entry); 1868 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1869 unlock: 1870 spin_unlock(ptl); 1871 return ret; 1872 } 1873 1874 /* 1875 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1876 * 1877 * Note that if it returns page table lock pointer, this routine returns without 1878 * unlocking page table lock. So callers must unlock it. 1879 */ 1880 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1881 { 1882 spinlock_t *ptl; 1883 ptl = pmd_lock(vma->vm_mm, pmd); 1884 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1885 pmd_devmap(*pmd))) 1886 return ptl; 1887 spin_unlock(ptl); 1888 return NULL; 1889 } 1890 1891 /* 1892 * Returns true if a given pud maps a thp, false otherwise. 1893 * 1894 * Note that if it returns true, this routine returns without unlocking page 1895 * table lock. So callers must unlock it. 1896 */ 1897 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1898 { 1899 spinlock_t *ptl; 1900 1901 ptl = pud_lock(vma->vm_mm, pud); 1902 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1903 return ptl; 1904 spin_unlock(ptl); 1905 return NULL; 1906 } 1907 1908 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1909 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1910 pud_t *pud, unsigned long addr) 1911 { 1912 spinlock_t *ptl; 1913 1914 ptl = __pud_trans_huge_lock(pud, vma); 1915 if (!ptl) 1916 return 0; 1917 /* 1918 * For architectures like ppc64 we look at deposited pgtable 1919 * when calling pudp_huge_get_and_clear. So do the 1920 * pgtable_trans_huge_withdraw after finishing pudp related 1921 * operations. 1922 */ 1923 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); 1924 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1925 if (vma_is_special_huge(vma)) { 1926 spin_unlock(ptl); 1927 /* No zero page support yet */ 1928 } else { 1929 /* No support for anonymous PUD pages yet */ 1930 BUG(); 1931 } 1932 return 1; 1933 } 1934 1935 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1936 unsigned long haddr) 1937 { 1938 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1939 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1940 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1941 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1942 1943 count_vm_event(THP_SPLIT_PUD); 1944 1945 pudp_huge_clear_flush_notify(vma, haddr, pud); 1946 } 1947 1948 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 1949 unsigned long address) 1950 { 1951 spinlock_t *ptl; 1952 struct mmu_notifier_range range; 1953 1954 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1955 address & HPAGE_PUD_MASK, 1956 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 1957 mmu_notifier_invalidate_range_start(&range); 1958 ptl = pud_lock(vma->vm_mm, pud); 1959 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 1960 goto out; 1961 __split_huge_pud_locked(vma, pud, range.start); 1962 1963 out: 1964 spin_unlock(ptl); 1965 /* 1966 * No need to double call mmu_notifier->invalidate_range() callback as 1967 * the above pudp_huge_clear_flush_notify() did already call it. 1968 */ 1969 mmu_notifier_invalidate_range_only_end(&range); 1970 } 1971 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1972 1973 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1974 unsigned long haddr, pmd_t *pmd) 1975 { 1976 struct mm_struct *mm = vma->vm_mm; 1977 pgtable_t pgtable; 1978 pmd_t _pmd; 1979 int i; 1980 1981 /* 1982 * Leave pmd empty until pte is filled note that it is fine to delay 1983 * notification until mmu_notifier_invalidate_range_end() as we are 1984 * replacing a zero pmd write protected page with a zero pte write 1985 * protected page. 1986 * 1987 * See Documentation/vm/mmu_notifier.rst 1988 */ 1989 pmdp_huge_clear_flush(vma, haddr, pmd); 1990 1991 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1992 pmd_populate(mm, &_pmd, pgtable); 1993 1994 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1995 pte_t *pte, entry; 1996 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 1997 entry = pte_mkspecial(entry); 1998 pte = pte_offset_map(&_pmd, haddr); 1999 VM_BUG_ON(!pte_none(*pte)); 2000 set_pte_at(mm, haddr, pte, entry); 2001 pte_unmap(pte); 2002 } 2003 smp_wmb(); /* make pte visible before pmd */ 2004 pmd_populate(mm, pmd, pgtable); 2005 } 2006 2007 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2008 unsigned long haddr, bool freeze) 2009 { 2010 struct mm_struct *mm = vma->vm_mm; 2011 struct page *page; 2012 pgtable_t pgtable; 2013 pmd_t old_pmd, _pmd; 2014 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 2015 unsigned long addr; 2016 int i; 2017 2018 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2019 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2020 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2021 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2022 && !pmd_devmap(*pmd)); 2023 2024 count_vm_event(THP_SPLIT_PMD); 2025 2026 if (!vma_is_anonymous(vma)) { 2027 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2028 /* 2029 * We are going to unmap this huge page. So 2030 * just go ahead and zap it 2031 */ 2032 if (arch_needs_pgtable_deposit()) 2033 zap_deposited_table(mm, pmd); 2034 if (vma_is_special_huge(vma)) 2035 return; 2036 page = pmd_page(_pmd); 2037 if (!PageDirty(page) && pmd_dirty(_pmd)) 2038 set_page_dirty(page); 2039 if (!PageReferenced(page) && pmd_young(_pmd)) 2040 SetPageReferenced(page); 2041 page_remove_rmap(page, true); 2042 put_page(page); 2043 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2044 return; 2045 } else if (is_huge_zero_pmd(*pmd)) { 2046 /* 2047 * FIXME: Do we want to invalidate secondary mmu by calling 2048 * mmu_notifier_invalidate_range() see comments below inside 2049 * __split_huge_pmd() ? 2050 * 2051 * We are going from a zero huge page write protected to zero 2052 * small page also write protected so it does not seems useful 2053 * to invalidate secondary mmu at this time. 2054 */ 2055 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2056 } 2057 2058 /* 2059 * Up to this point the pmd is present and huge and userland has the 2060 * whole access to the hugepage during the split (which happens in 2061 * place). If we overwrite the pmd with the not-huge version pointing 2062 * to the pte here (which of course we could if all CPUs were bug 2063 * free), userland could trigger a small page size TLB miss on the 2064 * small sized TLB while the hugepage TLB entry is still established in 2065 * the huge TLB. Some CPU doesn't like that. 2066 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2067 * 383 on page 93. Intel should be safe but is also warns that it's 2068 * only safe if the permission and cache attributes of the two entries 2069 * loaded in the two TLB is identical (which should be the case here). 2070 * But it is generally safer to never allow small and huge TLB entries 2071 * for the same virtual address to be loaded simultaneously. So instead 2072 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2073 * current pmd notpresent (atomically because here the pmd_trans_huge 2074 * must remain set at all times on the pmd until the split is complete 2075 * for this pmd), then we flush the SMP TLB and finally we write the 2076 * non-huge version of the pmd entry with pmd_populate. 2077 */ 2078 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2079 2080 pmd_migration = is_pmd_migration_entry(old_pmd); 2081 if (unlikely(pmd_migration)) { 2082 swp_entry_t entry; 2083 2084 entry = pmd_to_swp_entry(old_pmd); 2085 page = pfn_to_page(swp_offset(entry)); 2086 write = is_write_migration_entry(entry); 2087 young = false; 2088 soft_dirty = pmd_swp_soft_dirty(old_pmd); 2089 uffd_wp = pmd_swp_uffd_wp(old_pmd); 2090 } else { 2091 page = pmd_page(old_pmd); 2092 if (pmd_dirty(old_pmd)) 2093 SetPageDirty(page); 2094 write = pmd_write(old_pmd); 2095 young = pmd_young(old_pmd); 2096 soft_dirty = pmd_soft_dirty(old_pmd); 2097 uffd_wp = pmd_uffd_wp(old_pmd); 2098 } 2099 VM_BUG_ON_PAGE(!page_count(page), page); 2100 page_ref_add(page, HPAGE_PMD_NR - 1); 2101 2102 /* 2103 * Withdraw the table only after we mark the pmd entry invalid. 2104 * This's critical for some architectures (Power). 2105 */ 2106 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2107 pmd_populate(mm, &_pmd, pgtable); 2108 2109 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2110 pte_t entry, *pte; 2111 /* 2112 * Note that NUMA hinting access restrictions are not 2113 * transferred to avoid any possibility of altering 2114 * permissions across VMAs. 2115 */ 2116 if (freeze || pmd_migration) { 2117 swp_entry_t swp_entry; 2118 swp_entry = make_migration_entry(page + i, write); 2119 entry = swp_entry_to_pte(swp_entry); 2120 if (soft_dirty) 2121 entry = pte_swp_mksoft_dirty(entry); 2122 if (uffd_wp) 2123 entry = pte_swp_mkuffd_wp(entry); 2124 } else { 2125 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2126 entry = maybe_mkwrite(entry, vma); 2127 if (!write) 2128 entry = pte_wrprotect(entry); 2129 if (!young) 2130 entry = pte_mkold(entry); 2131 if (soft_dirty) 2132 entry = pte_mksoft_dirty(entry); 2133 if (uffd_wp) 2134 entry = pte_mkuffd_wp(entry); 2135 } 2136 pte = pte_offset_map(&_pmd, addr); 2137 BUG_ON(!pte_none(*pte)); 2138 set_pte_at(mm, addr, pte, entry); 2139 atomic_inc(&page[i]._mapcount); 2140 pte_unmap(pte); 2141 } 2142 2143 /* 2144 * Set PG_double_map before dropping compound_mapcount to avoid 2145 * false-negative page_mapped(). 2146 */ 2147 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2148 for (i = 0; i < HPAGE_PMD_NR; i++) 2149 atomic_inc(&page[i]._mapcount); 2150 } 2151 2152 lock_page_memcg(page); 2153 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2154 /* Last compound_mapcount is gone. */ 2155 __dec_lruvec_page_state(page, NR_ANON_THPS); 2156 if (TestClearPageDoubleMap(page)) { 2157 /* No need in mapcount reference anymore */ 2158 for (i = 0; i < HPAGE_PMD_NR; i++) 2159 atomic_dec(&page[i]._mapcount); 2160 } 2161 } 2162 unlock_page_memcg(page); 2163 2164 smp_wmb(); /* make pte visible before pmd */ 2165 pmd_populate(mm, pmd, pgtable); 2166 2167 if (freeze) { 2168 for (i = 0; i < HPAGE_PMD_NR; i++) { 2169 page_remove_rmap(page + i, false); 2170 put_page(page + i); 2171 } 2172 } 2173 } 2174 2175 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2176 unsigned long address, bool freeze, struct page *page) 2177 { 2178 spinlock_t *ptl; 2179 struct mmu_notifier_range range; 2180 bool was_locked = false; 2181 pmd_t _pmd; 2182 2183 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 2184 address & HPAGE_PMD_MASK, 2185 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2186 mmu_notifier_invalidate_range_start(&range); 2187 ptl = pmd_lock(vma->vm_mm, pmd); 2188 2189 /* 2190 * If caller asks to setup a migration entries, we need a page to check 2191 * pmd against. Otherwise we can end up replacing wrong page. 2192 */ 2193 VM_BUG_ON(freeze && !page); 2194 if (page) { 2195 VM_WARN_ON_ONCE(!PageLocked(page)); 2196 was_locked = true; 2197 if (page != pmd_page(*pmd)) 2198 goto out; 2199 } 2200 2201 repeat: 2202 if (pmd_trans_huge(*pmd)) { 2203 if (!page) { 2204 page = pmd_page(*pmd); 2205 if (unlikely(!trylock_page(page))) { 2206 get_page(page); 2207 _pmd = *pmd; 2208 spin_unlock(ptl); 2209 lock_page(page); 2210 spin_lock(ptl); 2211 if (unlikely(!pmd_same(*pmd, _pmd))) { 2212 unlock_page(page); 2213 put_page(page); 2214 page = NULL; 2215 goto repeat; 2216 } 2217 put_page(page); 2218 } 2219 } 2220 if (PageMlocked(page)) 2221 clear_page_mlock(page); 2222 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2223 goto out; 2224 __split_huge_pmd_locked(vma, pmd, range.start, freeze); 2225 out: 2226 spin_unlock(ptl); 2227 if (!was_locked && page) 2228 unlock_page(page); 2229 /* 2230 * No need to double call mmu_notifier->invalidate_range() callback. 2231 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2232 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2233 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2234 * fault will trigger a flush_notify before pointing to a new page 2235 * (it is fine if the secondary mmu keeps pointing to the old zero 2236 * page in the meantime) 2237 * 3) Split a huge pmd into pte pointing to the same page. No need 2238 * to invalidate secondary tlb entry they are all still valid. 2239 * any further changes to individual pte will notify. So no need 2240 * to call mmu_notifier->invalidate_range() 2241 */ 2242 mmu_notifier_invalidate_range_only_end(&range); 2243 } 2244 2245 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2246 bool freeze, struct page *page) 2247 { 2248 pgd_t *pgd; 2249 p4d_t *p4d; 2250 pud_t *pud; 2251 pmd_t *pmd; 2252 2253 pgd = pgd_offset(vma->vm_mm, address); 2254 if (!pgd_present(*pgd)) 2255 return; 2256 2257 p4d = p4d_offset(pgd, address); 2258 if (!p4d_present(*p4d)) 2259 return; 2260 2261 pud = pud_offset(p4d, address); 2262 if (!pud_present(*pud)) 2263 return; 2264 2265 pmd = pmd_offset(pud, address); 2266 2267 __split_huge_pmd(vma, pmd, address, freeze, page); 2268 } 2269 2270 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2271 unsigned long start, 2272 unsigned long end, 2273 long adjust_next) 2274 { 2275 /* 2276 * If the new start address isn't hpage aligned and it could 2277 * previously contain an hugepage: check if we need to split 2278 * an huge pmd. 2279 */ 2280 if (start & ~HPAGE_PMD_MASK && 2281 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2282 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2283 split_huge_pmd_address(vma, start, false, NULL); 2284 2285 /* 2286 * If the new end address isn't hpage aligned and it could 2287 * previously contain an hugepage: check if we need to split 2288 * an huge pmd. 2289 */ 2290 if (end & ~HPAGE_PMD_MASK && 2291 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2292 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2293 split_huge_pmd_address(vma, end, false, NULL); 2294 2295 /* 2296 * If we're also updating the vma->vm_next->vm_start, if the new 2297 * vm_next->vm_start isn't page aligned and it could previously 2298 * contain an hugepage: check if we need to split an huge pmd. 2299 */ 2300 if (adjust_next > 0) { 2301 struct vm_area_struct *next = vma->vm_next; 2302 unsigned long nstart = next->vm_start; 2303 nstart += adjust_next << PAGE_SHIFT; 2304 if (nstart & ~HPAGE_PMD_MASK && 2305 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2306 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2307 split_huge_pmd_address(next, nstart, false, NULL); 2308 } 2309 } 2310 2311 static void unmap_page(struct page *page) 2312 { 2313 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2314 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2315 bool unmap_success; 2316 2317 VM_BUG_ON_PAGE(!PageHead(page), page); 2318 2319 if (PageAnon(page)) 2320 ttu_flags |= TTU_SPLIT_FREEZE; 2321 2322 unmap_success = try_to_unmap(page, ttu_flags); 2323 VM_BUG_ON_PAGE(!unmap_success, page); 2324 } 2325 2326 static void remap_page(struct page *page) 2327 { 2328 int i; 2329 if (PageTransHuge(page)) { 2330 remove_migration_ptes(page, page, true); 2331 } else { 2332 for (i = 0; i < HPAGE_PMD_NR; i++) 2333 remove_migration_ptes(page + i, page + i, true); 2334 } 2335 } 2336 2337 static void __split_huge_page_tail(struct page *head, int tail, 2338 struct lruvec *lruvec, struct list_head *list) 2339 { 2340 struct page *page_tail = head + tail; 2341 2342 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2343 2344 /* 2345 * Clone page flags before unfreezing refcount. 2346 * 2347 * After successful get_page_unless_zero() might follow flags change, 2348 * for exmaple lock_page() which set PG_waiters. 2349 */ 2350 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2351 page_tail->flags |= (head->flags & 2352 ((1L << PG_referenced) | 2353 (1L << PG_swapbacked) | 2354 (1L << PG_swapcache) | 2355 (1L << PG_mlocked) | 2356 (1L << PG_uptodate) | 2357 (1L << PG_active) | 2358 (1L << PG_workingset) | 2359 (1L << PG_locked) | 2360 (1L << PG_unevictable) | 2361 (1L << PG_dirty))); 2362 2363 /* ->mapping in first tail page is compound_mapcount */ 2364 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2365 page_tail); 2366 page_tail->mapping = head->mapping; 2367 page_tail->index = head->index + tail; 2368 2369 /* Page flags must be visible before we make the page non-compound. */ 2370 smp_wmb(); 2371 2372 /* 2373 * Clear PageTail before unfreezing page refcount. 2374 * 2375 * After successful get_page_unless_zero() might follow put_page() 2376 * which needs correct compound_head(). 2377 */ 2378 clear_compound_head(page_tail); 2379 2380 /* Finally unfreeze refcount. Additional reference from page cache. */ 2381 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2382 PageSwapCache(head))); 2383 2384 if (page_is_young(head)) 2385 set_page_young(page_tail); 2386 if (page_is_idle(head)) 2387 set_page_idle(page_tail); 2388 2389 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2390 2391 /* 2392 * always add to the tail because some iterators expect new 2393 * pages to show after the currently processed elements - e.g. 2394 * migrate_pages 2395 */ 2396 lru_add_page_tail(head, page_tail, lruvec, list); 2397 } 2398 2399 static void __split_huge_page(struct page *page, struct list_head *list, 2400 pgoff_t end, unsigned long flags) 2401 { 2402 struct page *head = compound_head(page); 2403 pg_data_t *pgdat = page_pgdat(head); 2404 struct lruvec *lruvec; 2405 struct address_space *swap_cache = NULL; 2406 unsigned long offset = 0; 2407 int i; 2408 2409 lruvec = mem_cgroup_page_lruvec(head, pgdat); 2410 2411 /* complete memcg works before add pages to LRU */ 2412 mem_cgroup_split_huge_fixup(head); 2413 2414 if (PageAnon(head) && PageSwapCache(head)) { 2415 swp_entry_t entry = { .val = page_private(head) }; 2416 2417 offset = swp_offset(entry); 2418 swap_cache = swap_address_space(entry); 2419 xa_lock(&swap_cache->i_pages); 2420 } 2421 2422 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2423 __split_huge_page_tail(head, i, lruvec, list); 2424 /* Some pages can be beyond i_size: drop them from page cache */ 2425 if (head[i].index >= end) { 2426 ClearPageDirty(head + i); 2427 __delete_from_page_cache(head + i, NULL); 2428 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2429 shmem_uncharge(head->mapping->host, 1); 2430 put_page(head + i); 2431 } else if (!PageAnon(page)) { 2432 __xa_store(&head->mapping->i_pages, head[i].index, 2433 head + i, 0); 2434 } else if (swap_cache) { 2435 __xa_store(&swap_cache->i_pages, offset + i, 2436 head + i, 0); 2437 } 2438 } 2439 2440 ClearPageCompound(head); 2441 2442 split_page_owner(head, HPAGE_PMD_ORDER); 2443 2444 /* See comment in __split_huge_page_tail() */ 2445 if (PageAnon(head)) { 2446 /* Additional pin to swap cache */ 2447 if (PageSwapCache(head)) { 2448 page_ref_add(head, 2); 2449 xa_unlock(&swap_cache->i_pages); 2450 } else { 2451 page_ref_inc(head); 2452 } 2453 } else { 2454 /* Additional pin to page cache */ 2455 page_ref_add(head, 2); 2456 xa_unlock(&head->mapping->i_pages); 2457 } 2458 2459 spin_unlock_irqrestore(&pgdat->lru_lock, flags); 2460 2461 remap_page(head); 2462 2463 for (i = 0; i < HPAGE_PMD_NR; i++) { 2464 struct page *subpage = head + i; 2465 if (subpage == page) 2466 continue; 2467 unlock_page(subpage); 2468 2469 /* 2470 * Subpages may be freed if there wasn't any mapping 2471 * like if add_to_swap() is running on a lru page that 2472 * had its mapping zapped. And freeing these pages 2473 * requires taking the lru_lock so we do the put_page 2474 * of the tail pages after the split is complete. 2475 */ 2476 put_page(subpage); 2477 } 2478 } 2479 2480 int total_mapcount(struct page *page) 2481 { 2482 int i, compound, ret; 2483 2484 VM_BUG_ON_PAGE(PageTail(page), page); 2485 2486 if (likely(!PageCompound(page))) 2487 return atomic_read(&page->_mapcount) + 1; 2488 2489 compound = compound_mapcount(page); 2490 if (PageHuge(page)) 2491 return compound; 2492 ret = compound; 2493 for (i = 0; i < HPAGE_PMD_NR; i++) 2494 ret += atomic_read(&page[i]._mapcount) + 1; 2495 /* File pages has compound_mapcount included in _mapcount */ 2496 if (!PageAnon(page)) 2497 return ret - compound * HPAGE_PMD_NR; 2498 if (PageDoubleMap(page)) 2499 ret -= HPAGE_PMD_NR; 2500 return ret; 2501 } 2502 2503 /* 2504 * This calculates accurately how many mappings a transparent hugepage 2505 * has (unlike page_mapcount() which isn't fully accurate). This full 2506 * accuracy is primarily needed to know if copy-on-write faults can 2507 * reuse the page and change the mapping to read-write instead of 2508 * copying them. At the same time this returns the total_mapcount too. 2509 * 2510 * The function returns the highest mapcount any one of the subpages 2511 * has. If the return value is one, even if different processes are 2512 * mapping different subpages of the transparent hugepage, they can 2513 * all reuse it, because each process is reusing a different subpage. 2514 * 2515 * The total_mapcount is instead counting all virtual mappings of the 2516 * subpages. If the total_mapcount is equal to "one", it tells the 2517 * caller all mappings belong to the same "mm" and in turn the 2518 * anon_vma of the transparent hugepage can become the vma->anon_vma 2519 * local one as no other process may be mapping any of the subpages. 2520 * 2521 * It would be more accurate to replace page_mapcount() with 2522 * page_trans_huge_mapcount(), however we only use 2523 * page_trans_huge_mapcount() in the copy-on-write faults where we 2524 * need full accuracy to avoid breaking page pinning, because 2525 * page_trans_huge_mapcount() is slower than page_mapcount(). 2526 */ 2527 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2528 { 2529 int i, ret, _total_mapcount, mapcount; 2530 2531 /* hugetlbfs shouldn't call it */ 2532 VM_BUG_ON_PAGE(PageHuge(page), page); 2533 2534 if (likely(!PageTransCompound(page))) { 2535 mapcount = atomic_read(&page->_mapcount) + 1; 2536 if (total_mapcount) 2537 *total_mapcount = mapcount; 2538 return mapcount; 2539 } 2540 2541 page = compound_head(page); 2542 2543 _total_mapcount = ret = 0; 2544 for (i = 0; i < HPAGE_PMD_NR; i++) { 2545 mapcount = atomic_read(&page[i]._mapcount) + 1; 2546 ret = max(ret, mapcount); 2547 _total_mapcount += mapcount; 2548 } 2549 if (PageDoubleMap(page)) { 2550 ret -= 1; 2551 _total_mapcount -= HPAGE_PMD_NR; 2552 } 2553 mapcount = compound_mapcount(page); 2554 ret += mapcount; 2555 _total_mapcount += mapcount; 2556 if (total_mapcount) 2557 *total_mapcount = _total_mapcount; 2558 return ret; 2559 } 2560 2561 /* Racy check whether the huge page can be split */ 2562 bool can_split_huge_page(struct page *page, int *pextra_pins) 2563 { 2564 int extra_pins; 2565 2566 /* Additional pins from page cache */ 2567 if (PageAnon(page)) 2568 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2569 else 2570 extra_pins = HPAGE_PMD_NR; 2571 if (pextra_pins) 2572 *pextra_pins = extra_pins; 2573 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2574 } 2575 2576 /* 2577 * This function splits huge page into normal pages. @page can point to any 2578 * subpage of huge page to split. Split doesn't change the position of @page. 2579 * 2580 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2581 * The huge page must be locked. 2582 * 2583 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2584 * 2585 * Both head page and tail pages will inherit mapping, flags, and so on from 2586 * the hugepage. 2587 * 2588 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2589 * they are not mapped. 2590 * 2591 * Returns 0 if the hugepage is split successfully. 2592 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2593 * us. 2594 */ 2595 int split_huge_page_to_list(struct page *page, struct list_head *list) 2596 { 2597 struct page *head = compound_head(page); 2598 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2599 struct deferred_split *ds_queue = get_deferred_split_queue(head); 2600 struct anon_vma *anon_vma = NULL; 2601 struct address_space *mapping = NULL; 2602 int count, mapcount, extra_pins, ret; 2603 unsigned long flags; 2604 pgoff_t end; 2605 2606 VM_BUG_ON_PAGE(is_huge_zero_page(head), head); 2607 VM_BUG_ON_PAGE(!PageLocked(head), head); 2608 VM_BUG_ON_PAGE(!PageCompound(head), head); 2609 2610 if (PageWriteback(head)) 2611 return -EBUSY; 2612 2613 if (PageAnon(head)) { 2614 /* 2615 * The caller does not necessarily hold an mmap_lock that would 2616 * prevent the anon_vma disappearing so we first we take a 2617 * reference to it and then lock the anon_vma for write. This 2618 * is similar to page_lock_anon_vma_read except the write lock 2619 * is taken to serialise against parallel split or collapse 2620 * operations. 2621 */ 2622 anon_vma = page_get_anon_vma(head); 2623 if (!anon_vma) { 2624 ret = -EBUSY; 2625 goto out; 2626 } 2627 end = -1; 2628 mapping = NULL; 2629 anon_vma_lock_write(anon_vma); 2630 } else { 2631 mapping = head->mapping; 2632 2633 /* Truncated ? */ 2634 if (!mapping) { 2635 ret = -EBUSY; 2636 goto out; 2637 } 2638 2639 anon_vma = NULL; 2640 i_mmap_lock_read(mapping); 2641 2642 /* 2643 *__split_huge_page() may need to trim off pages beyond EOF: 2644 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2645 * which cannot be nested inside the page tree lock. So note 2646 * end now: i_size itself may be changed at any moment, but 2647 * head page lock is good enough to serialize the trimming. 2648 */ 2649 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2650 } 2651 2652 /* 2653 * Racy check if we can split the page, before unmap_page() will 2654 * split PMDs 2655 */ 2656 if (!can_split_huge_page(head, &extra_pins)) { 2657 ret = -EBUSY; 2658 goto out_unlock; 2659 } 2660 2661 unmap_page(head); 2662 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2663 2664 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2665 spin_lock_irqsave(&pgdata->lru_lock, flags); 2666 2667 if (mapping) { 2668 XA_STATE(xas, &mapping->i_pages, page_index(head)); 2669 2670 /* 2671 * Check if the head page is present in page cache. 2672 * We assume all tail are present too, if head is there. 2673 */ 2674 xa_lock(&mapping->i_pages); 2675 if (xas_load(&xas) != head) 2676 goto fail; 2677 } 2678 2679 /* Prevent deferred_split_scan() touching ->_refcount */ 2680 spin_lock(&ds_queue->split_queue_lock); 2681 count = page_count(head); 2682 mapcount = total_mapcount(head); 2683 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2684 if (!list_empty(page_deferred_list(head))) { 2685 ds_queue->split_queue_len--; 2686 list_del(page_deferred_list(head)); 2687 } 2688 spin_unlock(&ds_queue->split_queue_lock); 2689 if (mapping) { 2690 if (PageSwapBacked(head)) 2691 __dec_node_page_state(head, NR_SHMEM_THPS); 2692 else 2693 __dec_node_page_state(head, NR_FILE_THPS); 2694 } 2695 2696 __split_huge_page(page, list, end, flags); 2697 if (PageSwapCache(head)) { 2698 swp_entry_t entry = { .val = page_private(head) }; 2699 2700 ret = split_swap_cluster(entry); 2701 } else 2702 ret = 0; 2703 } else { 2704 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2705 pr_alert("total_mapcount: %u, page_count(): %u\n", 2706 mapcount, count); 2707 if (PageTail(page)) 2708 dump_page(head, NULL); 2709 dump_page(page, "total_mapcount(head) > 0"); 2710 BUG(); 2711 } 2712 spin_unlock(&ds_queue->split_queue_lock); 2713 fail: if (mapping) 2714 xa_unlock(&mapping->i_pages); 2715 spin_unlock_irqrestore(&pgdata->lru_lock, flags); 2716 remap_page(head); 2717 ret = -EBUSY; 2718 } 2719 2720 out_unlock: 2721 if (anon_vma) { 2722 anon_vma_unlock_write(anon_vma); 2723 put_anon_vma(anon_vma); 2724 } 2725 if (mapping) 2726 i_mmap_unlock_read(mapping); 2727 out: 2728 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2729 return ret; 2730 } 2731 2732 void free_transhuge_page(struct page *page) 2733 { 2734 struct deferred_split *ds_queue = get_deferred_split_queue(page); 2735 unsigned long flags; 2736 2737 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2738 if (!list_empty(page_deferred_list(page))) { 2739 ds_queue->split_queue_len--; 2740 list_del(page_deferred_list(page)); 2741 } 2742 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2743 free_compound_page(page); 2744 } 2745 2746 void deferred_split_huge_page(struct page *page) 2747 { 2748 struct deferred_split *ds_queue = get_deferred_split_queue(page); 2749 #ifdef CONFIG_MEMCG 2750 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup; 2751 #endif 2752 unsigned long flags; 2753 2754 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2755 2756 /* 2757 * The try_to_unmap() in page reclaim path might reach here too, 2758 * this may cause a race condition to corrupt deferred split queue. 2759 * And, if page reclaim is already handling the same page, it is 2760 * unnecessary to handle it again in shrinker. 2761 * 2762 * Check PageSwapCache to determine if the page is being 2763 * handled by page reclaim since THP swap would add the page into 2764 * swap cache before calling try_to_unmap(). 2765 */ 2766 if (PageSwapCache(page)) 2767 return; 2768 2769 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2770 if (list_empty(page_deferred_list(page))) { 2771 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2772 list_add_tail(page_deferred_list(page), &ds_queue->split_queue); 2773 ds_queue->split_queue_len++; 2774 #ifdef CONFIG_MEMCG 2775 if (memcg) 2776 memcg_set_shrinker_bit(memcg, page_to_nid(page), 2777 deferred_split_shrinker.id); 2778 #endif 2779 } 2780 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2781 } 2782 2783 static unsigned long deferred_split_count(struct shrinker *shrink, 2784 struct shrink_control *sc) 2785 { 2786 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2787 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2788 2789 #ifdef CONFIG_MEMCG 2790 if (sc->memcg) 2791 ds_queue = &sc->memcg->deferred_split_queue; 2792 #endif 2793 return READ_ONCE(ds_queue->split_queue_len); 2794 } 2795 2796 static unsigned long deferred_split_scan(struct shrinker *shrink, 2797 struct shrink_control *sc) 2798 { 2799 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2800 struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 2801 unsigned long flags; 2802 LIST_HEAD(list), *pos, *next; 2803 struct page *page; 2804 int split = 0; 2805 2806 #ifdef CONFIG_MEMCG 2807 if (sc->memcg) 2808 ds_queue = &sc->memcg->deferred_split_queue; 2809 #endif 2810 2811 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2812 /* Take pin on all head pages to avoid freeing them under us */ 2813 list_for_each_safe(pos, next, &ds_queue->split_queue) { 2814 page = list_entry((void *)pos, struct page, mapping); 2815 page = compound_head(page); 2816 if (get_page_unless_zero(page)) { 2817 list_move(page_deferred_list(page), &list); 2818 } else { 2819 /* We lost race with put_compound_page() */ 2820 list_del_init(page_deferred_list(page)); 2821 ds_queue->split_queue_len--; 2822 } 2823 if (!--sc->nr_to_scan) 2824 break; 2825 } 2826 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2827 2828 list_for_each_safe(pos, next, &list) { 2829 page = list_entry((void *)pos, struct page, mapping); 2830 if (!trylock_page(page)) 2831 goto next; 2832 /* split_huge_page() removes page from list on success */ 2833 if (!split_huge_page(page)) 2834 split++; 2835 unlock_page(page); 2836 next: 2837 put_page(page); 2838 } 2839 2840 spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 2841 list_splice_tail(&list, &ds_queue->split_queue); 2842 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 2843 2844 /* 2845 * Stop shrinker if we didn't split any page, but the queue is empty. 2846 * This can happen if pages were freed under us. 2847 */ 2848 if (!split && list_empty(&ds_queue->split_queue)) 2849 return SHRINK_STOP; 2850 return split; 2851 } 2852 2853 static struct shrinker deferred_split_shrinker = { 2854 .count_objects = deferred_split_count, 2855 .scan_objects = deferred_split_scan, 2856 .seeks = DEFAULT_SEEKS, 2857 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | 2858 SHRINKER_NONSLAB, 2859 }; 2860 2861 #ifdef CONFIG_DEBUG_FS 2862 static int split_huge_pages_set(void *data, u64 val) 2863 { 2864 struct zone *zone; 2865 struct page *page; 2866 unsigned long pfn, max_zone_pfn; 2867 unsigned long total = 0, split = 0; 2868 2869 if (val != 1) 2870 return -EINVAL; 2871 2872 for_each_populated_zone(zone) { 2873 max_zone_pfn = zone_end_pfn(zone); 2874 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2875 if (!pfn_valid(pfn)) 2876 continue; 2877 2878 page = pfn_to_page(pfn); 2879 if (!get_page_unless_zero(page)) 2880 continue; 2881 2882 if (zone != page_zone(page)) 2883 goto next; 2884 2885 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2886 goto next; 2887 2888 total++; 2889 lock_page(page); 2890 if (!split_huge_page(page)) 2891 split++; 2892 unlock_page(page); 2893 next: 2894 put_page(page); 2895 } 2896 } 2897 2898 pr_info("%lu of %lu THP split\n", split, total); 2899 2900 return 0; 2901 } 2902 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2903 "%llu\n"); 2904 2905 static int __init split_huge_pages_debugfs(void) 2906 { 2907 debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2908 &split_huge_pages_fops); 2909 return 0; 2910 } 2911 late_initcall(split_huge_pages_debugfs); 2912 #endif 2913 2914 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2915 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2916 struct page *page) 2917 { 2918 struct vm_area_struct *vma = pvmw->vma; 2919 struct mm_struct *mm = vma->vm_mm; 2920 unsigned long address = pvmw->address; 2921 pmd_t pmdval; 2922 swp_entry_t entry; 2923 pmd_t pmdswp; 2924 2925 if (!(pvmw->pmd && !pvmw->pte)) 2926 return; 2927 2928 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2929 pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 2930 if (pmd_dirty(pmdval)) 2931 set_page_dirty(page); 2932 entry = make_migration_entry(page, pmd_write(pmdval)); 2933 pmdswp = swp_entry_to_pmd(entry); 2934 if (pmd_soft_dirty(pmdval)) 2935 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2936 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2937 page_remove_rmap(page, true); 2938 put_page(page); 2939 } 2940 2941 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2942 { 2943 struct vm_area_struct *vma = pvmw->vma; 2944 struct mm_struct *mm = vma->vm_mm; 2945 unsigned long address = pvmw->address; 2946 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2947 pmd_t pmde; 2948 swp_entry_t entry; 2949 2950 if (!(pvmw->pmd && !pvmw->pte)) 2951 return; 2952 2953 entry = pmd_to_swp_entry(*pvmw->pmd); 2954 get_page(new); 2955 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2956 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2957 pmde = pmd_mksoft_dirty(pmde); 2958 if (is_write_migration_entry(entry)) 2959 pmde = maybe_pmd_mkwrite(pmde, vma); 2960 2961 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2962 if (PageAnon(new)) 2963 page_add_anon_rmap(new, vma, mmun_start, true); 2964 else 2965 page_add_file_rmap(new, true); 2966 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2967 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) 2968 mlock_vma_page(new); 2969 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2970 } 2971 #endif 2972