1 /* 2 * Copyright (C) 2009 Red Hat, Inc. 3 * 4 * This work is licensed under the terms of the GNU GPL, version 2. See 5 * the COPYING file in the top-level directory. 6 */ 7 8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9 10 #include <linux/mm.h> 11 #include <linux/sched.h> 12 #include <linux/sched/coredump.h> 13 #include <linux/sched/numa_balancing.h> 14 #include <linux/highmem.h> 15 #include <linux/hugetlb.h> 16 #include <linux/mmu_notifier.h> 17 #include <linux/rmap.h> 18 #include <linux/swap.h> 19 #include <linux/shrinker.h> 20 #include <linux/mm_inline.h> 21 #include <linux/swapops.h> 22 #include <linux/dax.h> 23 #include <linux/khugepaged.h> 24 #include <linux/freezer.h> 25 #include <linux/pfn_t.h> 26 #include <linux/mman.h> 27 #include <linux/memremap.h> 28 #include <linux/pagemap.h> 29 #include <linux/debugfs.h> 30 #include <linux/migrate.h> 31 #include <linux/hashtable.h> 32 #include <linux/userfaultfd_k.h> 33 #include <linux/page_idle.h> 34 #include <linux/shmem_fs.h> 35 #include <linux/oom.h> 36 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 if (vma_is_anonymous(vma)) 68 return __transparent_hugepage_enabled(vma); 69 if (vma_is_shmem(vma) && shmem_huge_enabled(vma)) 70 return __transparent_hugepage_enabled(vma); 71 72 return false; 73 } 74 75 static struct page *get_huge_zero_page(void) 76 { 77 struct page *zero_page; 78 retry: 79 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 80 return READ_ONCE(huge_zero_page); 81 82 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 83 HPAGE_PMD_ORDER); 84 if (!zero_page) { 85 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 86 return NULL; 87 } 88 count_vm_event(THP_ZERO_PAGE_ALLOC); 89 preempt_disable(); 90 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 91 preempt_enable(); 92 __free_pages(zero_page, compound_order(zero_page)); 93 goto retry; 94 } 95 96 /* We take additional reference here. It will be put back by shrinker */ 97 atomic_set(&huge_zero_refcount, 2); 98 preempt_enable(); 99 return READ_ONCE(huge_zero_page); 100 } 101 102 static void put_huge_zero_page(void) 103 { 104 /* 105 * Counter should never go to zero here. Only shrinker can put 106 * last reference. 107 */ 108 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 109 } 110 111 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 112 { 113 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 114 return READ_ONCE(huge_zero_page); 115 116 if (!get_huge_zero_page()) 117 return NULL; 118 119 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 120 put_huge_zero_page(); 121 122 return READ_ONCE(huge_zero_page); 123 } 124 125 void mm_put_huge_zero_page(struct mm_struct *mm) 126 { 127 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 128 put_huge_zero_page(); 129 } 130 131 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 132 struct shrink_control *sc) 133 { 134 /* we can free zero page only if last reference remains */ 135 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 136 } 137 138 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 139 struct shrink_control *sc) 140 { 141 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 142 struct page *zero_page = xchg(&huge_zero_page, NULL); 143 BUG_ON(zero_page == NULL); 144 __free_pages(zero_page, compound_order(zero_page)); 145 return HPAGE_PMD_NR; 146 } 147 148 return 0; 149 } 150 151 static struct shrinker huge_zero_page_shrinker = { 152 .count_objects = shrink_huge_zero_page_count, 153 .scan_objects = shrink_huge_zero_page_scan, 154 .seeks = DEFAULT_SEEKS, 155 }; 156 157 #ifdef CONFIG_SYSFS 158 static ssize_t enabled_show(struct kobject *kobj, 159 struct kobj_attribute *attr, char *buf) 160 { 161 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 162 return sprintf(buf, "[always] madvise never\n"); 163 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 164 return sprintf(buf, "always [madvise] never\n"); 165 else 166 return sprintf(buf, "always madvise [never]\n"); 167 } 168 169 static ssize_t enabled_store(struct kobject *kobj, 170 struct kobj_attribute *attr, 171 const char *buf, size_t count) 172 { 173 ssize_t ret = count; 174 175 if (!memcmp("always", buf, 176 min(sizeof("always")-1, count))) { 177 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 178 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 179 } else if (!memcmp("madvise", buf, 180 min(sizeof("madvise")-1, count))) { 181 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 182 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 183 } else if (!memcmp("never", buf, 184 min(sizeof("never")-1, count))) { 185 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 186 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 187 } else 188 ret = -EINVAL; 189 190 if (ret > 0) { 191 int err = start_stop_khugepaged(); 192 if (err) 193 ret = err; 194 } 195 return ret; 196 } 197 static struct kobj_attribute enabled_attr = 198 __ATTR(enabled, 0644, enabled_show, enabled_store); 199 200 ssize_t single_hugepage_flag_show(struct kobject *kobj, 201 struct kobj_attribute *attr, char *buf, 202 enum transparent_hugepage_flag flag) 203 { 204 return sprintf(buf, "%d\n", 205 !!test_bit(flag, &transparent_hugepage_flags)); 206 } 207 208 ssize_t single_hugepage_flag_store(struct kobject *kobj, 209 struct kobj_attribute *attr, 210 const char *buf, size_t count, 211 enum transparent_hugepage_flag flag) 212 { 213 unsigned long value; 214 int ret; 215 216 ret = kstrtoul(buf, 10, &value); 217 if (ret < 0) 218 return ret; 219 if (value > 1) 220 return -EINVAL; 221 222 if (value) 223 set_bit(flag, &transparent_hugepage_flags); 224 else 225 clear_bit(flag, &transparent_hugepage_flags); 226 227 return count; 228 } 229 230 static ssize_t defrag_show(struct kobject *kobj, 231 struct kobj_attribute *attr, char *buf) 232 { 233 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 234 return sprintf(buf, "[always] defer defer+madvise madvise never\n"); 235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 236 return sprintf(buf, "always [defer] defer+madvise madvise never\n"); 237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 238 return sprintf(buf, "always defer [defer+madvise] madvise never\n"); 239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 240 return sprintf(buf, "always defer defer+madvise [madvise] never\n"); 241 return sprintf(buf, "always defer defer+madvise madvise [never]\n"); 242 } 243 244 static ssize_t defrag_store(struct kobject *kobj, 245 struct kobj_attribute *attr, 246 const char *buf, size_t count) 247 { 248 if (!memcmp("always", buf, 249 min(sizeof("always")-1, count))) { 250 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 253 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 254 } else if (!memcmp("defer+madvise", buf, 255 min(sizeof("defer+madvise")-1, count))) { 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 (!memcmp("defer", buf, 261 min(sizeof("defer")-1, count))) { 262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 265 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 266 } else if (!memcmp("madvise", buf, 267 min(sizeof("madvise")-1, count))) { 268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 269 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 270 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 271 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 272 } else if (!memcmp("never", buf, 273 min(sizeof("never")-1, count))) { 274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 275 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 276 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 277 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 278 } else 279 return -EINVAL; 280 281 return count; 282 } 283 static struct kobj_attribute defrag_attr = 284 __ATTR(defrag, 0644, defrag_show, defrag_store); 285 286 static ssize_t use_zero_page_show(struct kobject *kobj, 287 struct kobj_attribute *attr, char *buf) 288 { 289 return single_hugepage_flag_show(kobj, attr, buf, 290 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 291 } 292 static ssize_t use_zero_page_store(struct kobject *kobj, 293 struct kobj_attribute *attr, const char *buf, size_t count) 294 { 295 return single_hugepage_flag_store(kobj, attr, buf, count, 296 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 297 } 298 static struct kobj_attribute use_zero_page_attr = 299 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 300 301 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 302 struct kobj_attribute *attr, char *buf) 303 { 304 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); 305 } 306 static struct kobj_attribute hpage_pmd_size_attr = 307 __ATTR_RO(hpage_pmd_size); 308 309 #ifdef CONFIG_DEBUG_VM 310 static ssize_t debug_cow_show(struct kobject *kobj, 311 struct kobj_attribute *attr, char *buf) 312 { 313 return single_hugepage_flag_show(kobj, attr, buf, 314 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 315 } 316 static ssize_t debug_cow_store(struct kobject *kobj, 317 struct kobj_attribute *attr, 318 const char *buf, size_t count) 319 { 320 return single_hugepage_flag_store(kobj, attr, buf, count, 321 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 322 } 323 static struct kobj_attribute debug_cow_attr = 324 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 325 #endif /* CONFIG_DEBUG_VM */ 326 327 static struct attribute *hugepage_attr[] = { 328 &enabled_attr.attr, 329 &defrag_attr.attr, 330 &use_zero_page_attr.attr, 331 &hpage_pmd_size_attr.attr, 332 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 333 &shmem_enabled_attr.attr, 334 #endif 335 #ifdef CONFIG_DEBUG_VM 336 &debug_cow_attr.attr, 337 #endif 338 NULL, 339 }; 340 341 static const struct attribute_group hugepage_attr_group = { 342 .attrs = hugepage_attr, 343 }; 344 345 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 346 { 347 int err; 348 349 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 350 if (unlikely(!*hugepage_kobj)) { 351 pr_err("failed to create transparent hugepage kobject\n"); 352 return -ENOMEM; 353 } 354 355 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 356 if (err) { 357 pr_err("failed to register transparent hugepage group\n"); 358 goto delete_obj; 359 } 360 361 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 362 if (err) { 363 pr_err("failed to register transparent hugepage group\n"); 364 goto remove_hp_group; 365 } 366 367 return 0; 368 369 remove_hp_group: 370 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 371 delete_obj: 372 kobject_put(*hugepage_kobj); 373 return err; 374 } 375 376 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 377 { 378 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 379 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 380 kobject_put(hugepage_kobj); 381 } 382 #else 383 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 384 { 385 return 0; 386 } 387 388 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 389 { 390 } 391 #endif /* CONFIG_SYSFS */ 392 393 static int __init hugepage_init(void) 394 { 395 int err; 396 struct kobject *hugepage_kobj; 397 398 if (!has_transparent_hugepage()) { 399 transparent_hugepage_flags = 0; 400 return -EINVAL; 401 } 402 403 /* 404 * hugepages can't be allocated by the buddy allocator 405 */ 406 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 407 /* 408 * we use page->mapping and page->index in second tail page 409 * as list_head: assuming THP order >= 2 410 */ 411 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 412 413 err = hugepage_init_sysfs(&hugepage_kobj); 414 if (err) 415 goto err_sysfs; 416 417 err = khugepaged_init(); 418 if (err) 419 goto err_slab; 420 421 err = register_shrinker(&huge_zero_page_shrinker); 422 if (err) 423 goto err_hzp_shrinker; 424 err = register_shrinker(&deferred_split_shrinker); 425 if (err) 426 goto err_split_shrinker; 427 428 /* 429 * By default disable transparent hugepages on smaller systems, 430 * where the extra memory used could hurt more than TLB overhead 431 * is likely to save. The admin can still enable it through /sys. 432 */ 433 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 434 transparent_hugepage_flags = 0; 435 return 0; 436 } 437 438 err = start_stop_khugepaged(); 439 if (err) 440 goto err_khugepaged; 441 442 return 0; 443 err_khugepaged: 444 unregister_shrinker(&deferred_split_shrinker); 445 err_split_shrinker: 446 unregister_shrinker(&huge_zero_page_shrinker); 447 err_hzp_shrinker: 448 khugepaged_destroy(); 449 err_slab: 450 hugepage_exit_sysfs(hugepage_kobj); 451 err_sysfs: 452 return err; 453 } 454 subsys_initcall(hugepage_init); 455 456 static int __init setup_transparent_hugepage(char *str) 457 { 458 int ret = 0; 459 if (!str) 460 goto out; 461 if (!strcmp(str, "always")) { 462 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 463 &transparent_hugepage_flags); 464 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 465 &transparent_hugepage_flags); 466 ret = 1; 467 } else if (!strcmp(str, "madvise")) { 468 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 469 &transparent_hugepage_flags); 470 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 471 &transparent_hugepage_flags); 472 ret = 1; 473 } else if (!strcmp(str, "never")) { 474 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 475 &transparent_hugepage_flags); 476 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 477 &transparent_hugepage_flags); 478 ret = 1; 479 } 480 out: 481 if (!ret) 482 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 483 return ret; 484 } 485 __setup("transparent_hugepage=", setup_transparent_hugepage); 486 487 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 488 { 489 if (likely(vma->vm_flags & VM_WRITE)) 490 pmd = pmd_mkwrite(pmd); 491 return pmd; 492 } 493 494 static inline struct list_head *page_deferred_list(struct page *page) 495 { 496 /* ->lru in the tail pages is occupied by compound_head. */ 497 return &page[2].deferred_list; 498 } 499 500 void prep_transhuge_page(struct page *page) 501 { 502 /* 503 * we use page->mapping and page->indexlru in second tail page 504 * as list_head: assuming THP order >= 2 505 */ 506 507 INIT_LIST_HEAD(page_deferred_list(page)); 508 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 509 } 510 511 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len, 512 loff_t off, unsigned long flags, unsigned long size) 513 { 514 unsigned long addr; 515 loff_t off_end = off + len; 516 loff_t off_align = round_up(off, size); 517 unsigned long len_pad; 518 519 if (off_end <= off_align || (off_end - off_align) < size) 520 return 0; 521 522 len_pad = len + size; 523 if (len_pad < len || (off + len_pad) < off) 524 return 0; 525 526 addr = current->mm->get_unmapped_area(filp, 0, len_pad, 527 off >> PAGE_SHIFT, flags); 528 if (IS_ERR_VALUE(addr)) 529 return 0; 530 531 addr += (off - addr) & (size - 1); 532 return addr; 533 } 534 535 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 536 unsigned long len, unsigned long pgoff, unsigned long flags) 537 { 538 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 539 540 if (addr) 541 goto out; 542 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 543 goto out; 544 545 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE); 546 if (addr) 547 return addr; 548 549 out: 550 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 551 } 552 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 553 554 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 555 struct page *page, gfp_t gfp) 556 { 557 struct vm_area_struct *vma = vmf->vma; 558 struct mem_cgroup *memcg; 559 pgtable_t pgtable; 560 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 561 vm_fault_t ret = 0; 562 563 VM_BUG_ON_PAGE(!PageCompound(page), page); 564 565 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) { 566 put_page(page); 567 count_vm_event(THP_FAULT_FALLBACK); 568 return VM_FAULT_FALLBACK; 569 } 570 571 pgtable = pte_alloc_one(vma->vm_mm); 572 if (unlikely(!pgtable)) { 573 ret = VM_FAULT_OOM; 574 goto release; 575 } 576 577 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 578 /* 579 * The memory barrier inside __SetPageUptodate makes sure that 580 * clear_huge_page writes become visible before the set_pmd_at() 581 * write. 582 */ 583 __SetPageUptodate(page); 584 585 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 586 if (unlikely(!pmd_none(*vmf->pmd))) { 587 goto unlock_release; 588 } else { 589 pmd_t entry; 590 591 ret = check_stable_address_space(vma->vm_mm); 592 if (ret) 593 goto unlock_release; 594 595 /* Deliver the page fault to userland */ 596 if (userfaultfd_missing(vma)) { 597 vm_fault_t ret2; 598 599 spin_unlock(vmf->ptl); 600 mem_cgroup_cancel_charge(page, memcg, true); 601 put_page(page); 602 pte_free(vma->vm_mm, pgtable); 603 ret2 = handle_userfault(vmf, VM_UFFD_MISSING); 604 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK); 605 return ret2; 606 } 607 608 entry = mk_huge_pmd(page, vma->vm_page_prot); 609 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 610 page_add_new_anon_rmap(page, vma, haddr, true); 611 mem_cgroup_commit_charge(page, memcg, false, true); 612 lru_cache_add_active_or_unevictable(page, vma); 613 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 614 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 615 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 616 mm_inc_nr_ptes(vma->vm_mm); 617 spin_unlock(vmf->ptl); 618 count_vm_event(THP_FAULT_ALLOC); 619 } 620 621 return 0; 622 unlock_release: 623 spin_unlock(vmf->ptl); 624 release: 625 if (pgtable) 626 pte_free(vma->vm_mm, pgtable); 627 mem_cgroup_cancel_charge(page, memcg, true); 628 put_page(page); 629 return ret; 630 631 } 632 633 /* 634 * always: directly stall for all thp allocations 635 * defer: wake kswapd and fail if not immediately available 636 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 637 * fail if not immediately available 638 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 639 * available 640 * never: never stall for any thp allocation 641 */ 642 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 643 { 644 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 645 646 /* Always do synchronous compaction */ 647 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 648 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 649 650 /* Kick kcompactd and fail quickly */ 651 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 652 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 653 654 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 655 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 656 return GFP_TRANSHUGE_LIGHT | 657 (vma_madvised ? __GFP_DIRECT_RECLAIM : 658 __GFP_KSWAPD_RECLAIM); 659 660 /* Only do synchronous compaction if madvised */ 661 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 662 return GFP_TRANSHUGE_LIGHT | 663 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 664 665 return GFP_TRANSHUGE_LIGHT; 666 } 667 668 /* Caller must hold page table lock. */ 669 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 670 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 671 struct page *zero_page) 672 { 673 pmd_t entry; 674 if (!pmd_none(*pmd)) 675 return false; 676 entry = mk_pmd(zero_page, vma->vm_page_prot); 677 entry = pmd_mkhuge(entry); 678 if (pgtable) 679 pgtable_trans_huge_deposit(mm, pmd, pgtable); 680 set_pmd_at(mm, haddr, pmd, entry); 681 mm_inc_nr_ptes(mm); 682 return true; 683 } 684 685 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 686 { 687 struct vm_area_struct *vma = vmf->vma; 688 gfp_t gfp; 689 struct page *page; 690 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 691 692 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 693 return VM_FAULT_FALLBACK; 694 if (unlikely(anon_vma_prepare(vma))) 695 return VM_FAULT_OOM; 696 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 697 return VM_FAULT_OOM; 698 if (!(vmf->flags & FAULT_FLAG_WRITE) && 699 !mm_forbids_zeropage(vma->vm_mm) && 700 transparent_hugepage_use_zero_page()) { 701 pgtable_t pgtable; 702 struct page *zero_page; 703 bool set; 704 vm_fault_t ret; 705 pgtable = pte_alloc_one(vma->vm_mm); 706 if (unlikely(!pgtable)) 707 return VM_FAULT_OOM; 708 zero_page = mm_get_huge_zero_page(vma->vm_mm); 709 if (unlikely(!zero_page)) { 710 pte_free(vma->vm_mm, pgtable); 711 count_vm_event(THP_FAULT_FALLBACK); 712 return VM_FAULT_FALLBACK; 713 } 714 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 715 ret = 0; 716 set = false; 717 if (pmd_none(*vmf->pmd)) { 718 ret = check_stable_address_space(vma->vm_mm); 719 if (ret) { 720 spin_unlock(vmf->ptl); 721 } else if (userfaultfd_missing(vma)) { 722 spin_unlock(vmf->ptl); 723 ret = handle_userfault(vmf, VM_UFFD_MISSING); 724 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 725 } else { 726 set_huge_zero_page(pgtable, vma->vm_mm, vma, 727 haddr, vmf->pmd, zero_page); 728 spin_unlock(vmf->ptl); 729 set = true; 730 } 731 } else 732 spin_unlock(vmf->ptl); 733 if (!set) 734 pte_free(vma->vm_mm, pgtable); 735 return ret; 736 } 737 gfp = alloc_hugepage_direct_gfpmask(vma); 738 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 739 if (unlikely(!page)) { 740 count_vm_event(THP_FAULT_FALLBACK); 741 return VM_FAULT_FALLBACK; 742 } 743 prep_transhuge_page(page); 744 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 745 } 746 747 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 748 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 749 pgtable_t pgtable) 750 { 751 struct mm_struct *mm = vma->vm_mm; 752 pmd_t entry; 753 spinlock_t *ptl; 754 755 ptl = pmd_lock(mm, pmd); 756 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 757 if (pfn_t_devmap(pfn)) 758 entry = pmd_mkdevmap(entry); 759 if (write) { 760 entry = pmd_mkyoung(pmd_mkdirty(entry)); 761 entry = maybe_pmd_mkwrite(entry, vma); 762 } 763 764 if (pgtable) { 765 pgtable_trans_huge_deposit(mm, pmd, pgtable); 766 mm_inc_nr_ptes(mm); 767 } 768 769 set_pmd_at(mm, addr, pmd, entry); 770 update_mmu_cache_pmd(vma, addr, pmd); 771 spin_unlock(ptl); 772 } 773 774 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 775 pmd_t *pmd, pfn_t pfn, bool write) 776 { 777 pgprot_t pgprot = vma->vm_page_prot; 778 pgtable_t pgtable = NULL; 779 /* 780 * If we had pmd_special, we could avoid all these restrictions, 781 * but we need to be consistent with PTEs and architectures that 782 * can't support a 'special' bit. 783 */ 784 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 785 !pfn_t_devmap(pfn)); 786 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 787 (VM_PFNMAP|VM_MIXEDMAP)); 788 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 789 790 if (addr < vma->vm_start || addr >= vma->vm_end) 791 return VM_FAULT_SIGBUS; 792 793 if (arch_needs_pgtable_deposit()) { 794 pgtable = pte_alloc_one(vma->vm_mm); 795 if (!pgtable) 796 return VM_FAULT_OOM; 797 } 798 799 track_pfn_insert(vma, &pgprot, pfn); 800 801 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 802 return VM_FAULT_NOPAGE; 803 } 804 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 805 806 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 807 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 808 { 809 if (likely(vma->vm_flags & VM_WRITE)) 810 pud = pud_mkwrite(pud); 811 return pud; 812 } 813 814 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 815 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 816 { 817 struct mm_struct *mm = vma->vm_mm; 818 pud_t entry; 819 spinlock_t *ptl; 820 821 ptl = pud_lock(mm, pud); 822 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 823 if (pfn_t_devmap(pfn)) 824 entry = pud_mkdevmap(entry); 825 if (write) { 826 entry = pud_mkyoung(pud_mkdirty(entry)); 827 entry = maybe_pud_mkwrite(entry, vma); 828 } 829 set_pud_at(mm, addr, pud, entry); 830 update_mmu_cache_pud(vma, addr, pud); 831 spin_unlock(ptl); 832 } 833 834 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 835 pud_t *pud, pfn_t pfn, bool write) 836 { 837 pgprot_t pgprot = vma->vm_page_prot; 838 /* 839 * If we had pud_special, we could avoid all these restrictions, 840 * but we need to be consistent with PTEs and architectures that 841 * can't support a 'special' bit. 842 */ 843 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 844 !pfn_t_devmap(pfn)); 845 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 846 (VM_PFNMAP|VM_MIXEDMAP)); 847 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 848 849 if (addr < vma->vm_start || addr >= vma->vm_end) 850 return VM_FAULT_SIGBUS; 851 852 track_pfn_insert(vma, &pgprot, pfn); 853 854 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 855 return VM_FAULT_NOPAGE; 856 } 857 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 858 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 859 860 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 861 pmd_t *pmd, int flags) 862 { 863 pmd_t _pmd; 864 865 _pmd = pmd_mkyoung(*pmd); 866 if (flags & FOLL_WRITE) 867 _pmd = pmd_mkdirty(_pmd); 868 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 869 pmd, _pmd, flags & FOLL_WRITE)) 870 update_mmu_cache_pmd(vma, addr, pmd); 871 } 872 873 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 874 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 875 { 876 unsigned long pfn = pmd_pfn(*pmd); 877 struct mm_struct *mm = vma->vm_mm; 878 struct page *page; 879 880 assert_spin_locked(pmd_lockptr(mm, pmd)); 881 882 /* 883 * When we COW a devmap PMD entry, we split it into PTEs, so we should 884 * not be in this function with `flags & FOLL_COW` set. 885 */ 886 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 887 888 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 889 return NULL; 890 891 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 892 /* pass */; 893 else 894 return NULL; 895 896 if (flags & FOLL_TOUCH) 897 touch_pmd(vma, addr, pmd, flags); 898 899 /* 900 * device mapped pages can only be returned if the 901 * caller will manage the page reference count. 902 */ 903 if (!(flags & FOLL_GET)) 904 return ERR_PTR(-EEXIST); 905 906 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 907 *pgmap = get_dev_pagemap(pfn, *pgmap); 908 if (!*pgmap) 909 return ERR_PTR(-EFAULT); 910 page = pfn_to_page(pfn); 911 get_page(page); 912 913 return page; 914 } 915 916 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 917 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 918 struct vm_area_struct *vma) 919 { 920 spinlock_t *dst_ptl, *src_ptl; 921 struct page *src_page; 922 pmd_t pmd; 923 pgtable_t pgtable = NULL; 924 int ret = -ENOMEM; 925 926 /* Skip if can be re-fill on fault */ 927 if (!vma_is_anonymous(vma)) 928 return 0; 929 930 pgtable = pte_alloc_one(dst_mm); 931 if (unlikely(!pgtable)) 932 goto out; 933 934 dst_ptl = pmd_lock(dst_mm, dst_pmd); 935 src_ptl = pmd_lockptr(src_mm, src_pmd); 936 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 937 938 ret = -EAGAIN; 939 pmd = *src_pmd; 940 941 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 942 if (unlikely(is_swap_pmd(pmd))) { 943 swp_entry_t entry = pmd_to_swp_entry(pmd); 944 945 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 946 if (is_write_migration_entry(entry)) { 947 make_migration_entry_read(&entry); 948 pmd = swp_entry_to_pmd(entry); 949 if (pmd_swp_soft_dirty(*src_pmd)) 950 pmd = pmd_swp_mksoft_dirty(pmd); 951 set_pmd_at(src_mm, addr, src_pmd, pmd); 952 } 953 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 954 mm_inc_nr_ptes(dst_mm); 955 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 956 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 957 ret = 0; 958 goto out_unlock; 959 } 960 #endif 961 962 if (unlikely(!pmd_trans_huge(pmd))) { 963 pte_free(dst_mm, pgtable); 964 goto out_unlock; 965 } 966 /* 967 * When page table lock is held, the huge zero pmd should not be 968 * under splitting since we don't split the page itself, only pmd to 969 * a page table. 970 */ 971 if (is_huge_zero_pmd(pmd)) { 972 struct page *zero_page; 973 /* 974 * get_huge_zero_page() will never allocate a new page here, 975 * since we already have a zero page to copy. It just takes a 976 * reference. 977 */ 978 zero_page = mm_get_huge_zero_page(dst_mm); 979 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 980 zero_page); 981 ret = 0; 982 goto out_unlock; 983 } 984 985 src_page = pmd_page(pmd); 986 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 987 get_page(src_page); 988 page_dup_rmap(src_page, true); 989 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 990 mm_inc_nr_ptes(dst_mm); 991 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 992 993 pmdp_set_wrprotect(src_mm, addr, src_pmd); 994 pmd = pmd_mkold(pmd_wrprotect(pmd)); 995 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 996 997 ret = 0; 998 out_unlock: 999 spin_unlock(src_ptl); 1000 spin_unlock(dst_ptl); 1001 out: 1002 return ret; 1003 } 1004 1005 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1006 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1007 pud_t *pud, int flags) 1008 { 1009 pud_t _pud; 1010 1011 _pud = pud_mkyoung(*pud); 1012 if (flags & FOLL_WRITE) 1013 _pud = pud_mkdirty(_pud); 1014 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1015 pud, _pud, flags & FOLL_WRITE)) 1016 update_mmu_cache_pud(vma, addr, pud); 1017 } 1018 1019 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1020 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1021 { 1022 unsigned long pfn = pud_pfn(*pud); 1023 struct mm_struct *mm = vma->vm_mm; 1024 struct page *page; 1025 1026 assert_spin_locked(pud_lockptr(mm, pud)); 1027 1028 if (flags & FOLL_WRITE && !pud_write(*pud)) 1029 return NULL; 1030 1031 if (pud_present(*pud) && pud_devmap(*pud)) 1032 /* pass */; 1033 else 1034 return NULL; 1035 1036 if (flags & FOLL_TOUCH) 1037 touch_pud(vma, addr, pud, flags); 1038 1039 /* 1040 * device mapped pages can only be returned if the 1041 * caller will manage the page reference count. 1042 */ 1043 if (!(flags & FOLL_GET)) 1044 return ERR_PTR(-EEXIST); 1045 1046 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1047 *pgmap = get_dev_pagemap(pfn, *pgmap); 1048 if (!*pgmap) 1049 return ERR_PTR(-EFAULT); 1050 page = pfn_to_page(pfn); 1051 get_page(page); 1052 1053 return page; 1054 } 1055 1056 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1057 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1058 struct vm_area_struct *vma) 1059 { 1060 spinlock_t *dst_ptl, *src_ptl; 1061 pud_t pud; 1062 int ret; 1063 1064 dst_ptl = pud_lock(dst_mm, dst_pud); 1065 src_ptl = pud_lockptr(src_mm, src_pud); 1066 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1067 1068 ret = -EAGAIN; 1069 pud = *src_pud; 1070 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1071 goto out_unlock; 1072 1073 /* 1074 * When page table lock is held, the huge zero pud should not be 1075 * under splitting since we don't split the page itself, only pud to 1076 * a page table. 1077 */ 1078 if (is_huge_zero_pud(pud)) { 1079 /* No huge zero pud yet */ 1080 } 1081 1082 pudp_set_wrprotect(src_mm, addr, src_pud); 1083 pud = pud_mkold(pud_wrprotect(pud)); 1084 set_pud_at(dst_mm, addr, dst_pud, pud); 1085 1086 ret = 0; 1087 out_unlock: 1088 spin_unlock(src_ptl); 1089 spin_unlock(dst_ptl); 1090 return ret; 1091 } 1092 1093 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1094 { 1095 pud_t entry; 1096 unsigned long haddr; 1097 bool write = vmf->flags & FAULT_FLAG_WRITE; 1098 1099 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1100 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1101 goto unlock; 1102 1103 entry = pud_mkyoung(orig_pud); 1104 if (write) 1105 entry = pud_mkdirty(entry); 1106 haddr = vmf->address & HPAGE_PUD_MASK; 1107 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1108 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1109 1110 unlock: 1111 spin_unlock(vmf->ptl); 1112 } 1113 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1114 1115 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1116 { 1117 pmd_t entry; 1118 unsigned long haddr; 1119 bool write = vmf->flags & FAULT_FLAG_WRITE; 1120 1121 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1122 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1123 goto unlock; 1124 1125 entry = pmd_mkyoung(orig_pmd); 1126 if (write) 1127 entry = pmd_mkdirty(entry); 1128 haddr = vmf->address & HPAGE_PMD_MASK; 1129 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1130 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1131 1132 unlock: 1133 spin_unlock(vmf->ptl); 1134 } 1135 1136 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, 1137 pmd_t orig_pmd, struct page *page) 1138 { 1139 struct vm_area_struct *vma = vmf->vma; 1140 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1141 struct mem_cgroup *memcg; 1142 pgtable_t pgtable; 1143 pmd_t _pmd; 1144 int i; 1145 vm_fault_t ret = 0; 1146 struct page **pages; 1147 struct mmu_notifier_range range; 1148 1149 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *), 1150 GFP_KERNEL); 1151 if (unlikely(!pages)) { 1152 ret |= VM_FAULT_OOM; 1153 goto out; 1154 } 1155 1156 for (i = 0; i < HPAGE_PMD_NR; i++) { 1157 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 1158 vmf->address, page_to_nid(page)); 1159 if (unlikely(!pages[i] || 1160 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm, 1161 GFP_KERNEL, &memcg, false))) { 1162 if (pages[i]) 1163 put_page(pages[i]); 1164 while (--i >= 0) { 1165 memcg = (void *)page_private(pages[i]); 1166 set_page_private(pages[i], 0); 1167 mem_cgroup_cancel_charge(pages[i], memcg, 1168 false); 1169 put_page(pages[i]); 1170 } 1171 kfree(pages); 1172 ret |= VM_FAULT_OOM; 1173 goto out; 1174 } 1175 set_page_private(pages[i], (unsigned long)memcg); 1176 } 1177 1178 for (i = 0; i < HPAGE_PMD_NR; i++) { 1179 copy_user_highpage(pages[i], page + i, 1180 haddr + PAGE_SIZE * i, vma); 1181 __SetPageUptodate(pages[i]); 1182 cond_resched(); 1183 } 1184 1185 mmu_notifier_range_init(&range, vma->vm_mm, haddr, 1186 haddr + HPAGE_PMD_SIZE); 1187 mmu_notifier_invalidate_range_start(&range); 1188 1189 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1190 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1191 goto out_free_pages; 1192 VM_BUG_ON_PAGE(!PageHead(page), page); 1193 1194 /* 1195 * Leave pmd empty until pte is filled note we must notify here as 1196 * concurrent CPU thread might write to new page before the call to 1197 * mmu_notifier_invalidate_range_end() happens which can lead to a 1198 * device seeing memory write in different order than CPU. 1199 * 1200 * See Documentation/vm/mmu_notifier.rst 1201 */ 1202 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1203 1204 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1205 pmd_populate(vma->vm_mm, &_pmd, pgtable); 1206 1207 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1208 pte_t entry; 1209 entry = mk_pte(pages[i], vma->vm_page_prot); 1210 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1211 memcg = (void *)page_private(pages[i]); 1212 set_page_private(pages[i], 0); 1213 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1214 mem_cgroup_commit_charge(pages[i], memcg, false, false); 1215 lru_cache_add_active_or_unevictable(pages[i], vma); 1216 vmf->pte = pte_offset_map(&_pmd, haddr); 1217 VM_BUG_ON(!pte_none(*vmf->pte)); 1218 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 1219 pte_unmap(vmf->pte); 1220 } 1221 kfree(pages); 1222 1223 smp_wmb(); /* make pte visible before pmd */ 1224 pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1225 page_remove_rmap(page, true); 1226 spin_unlock(vmf->ptl); 1227 1228 /* 1229 * No need to double call mmu_notifier->invalidate_range() callback as 1230 * the above pmdp_huge_clear_flush_notify() did already call it. 1231 */ 1232 mmu_notifier_invalidate_range_only_end(&range); 1233 1234 ret |= VM_FAULT_WRITE; 1235 put_page(page); 1236 1237 out: 1238 return ret; 1239 1240 out_free_pages: 1241 spin_unlock(vmf->ptl); 1242 mmu_notifier_invalidate_range_end(&range); 1243 for (i = 0; i < HPAGE_PMD_NR; i++) { 1244 memcg = (void *)page_private(pages[i]); 1245 set_page_private(pages[i], 0); 1246 mem_cgroup_cancel_charge(pages[i], memcg, false); 1247 put_page(pages[i]); 1248 } 1249 kfree(pages); 1250 goto out; 1251 } 1252 1253 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1254 { 1255 struct vm_area_struct *vma = vmf->vma; 1256 struct page *page = NULL, *new_page; 1257 struct mem_cgroup *memcg; 1258 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1259 struct mmu_notifier_range range; 1260 gfp_t huge_gfp; /* for allocation and charge */ 1261 vm_fault_t ret = 0; 1262 1263 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1264 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1265 if (is_huge_zero_pmd(orig_pmd)) 1266 goto alloc; 1267 spin_lock(vmf->ptl); 1268 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1269 goto out_unlock; 1270 1271 page = pmd_page(orig_pmd); 1272 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1273 /* 1274 * We can only reuse the page if nobody else maps the huge page or it's 1275 * part. 1276 */ 1277 if (!trylock_page(page)) { 1278 get_page(page); 1279 spin_unlock(vmf->ptl); 1280 lock_page(page); 1281 spin_lock(vmf->ptl); 1282 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1283 unlock_page(page); 1284 put_page(page); 1285 goto out_unlock; 1286 } 1287 put_page(page); 1288 } 1289 if (reuse_swap_page(page, NULL)) { 1290 pmd_t entry; 1291 entry = pmd_mkyoung(orig_pmd); 1292 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1293 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1294 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1295 ret |= VM_FAULT_WRITE; 1296 unlock_page(page); 1297 goto out_unlock; 1298 } 1299 unlock_page(page); 1300 get_page(page); 1301 spin_unlock(vmf->ptl); 1302 alloc: 1303 if (__transparent_hugepage_enabled(vma) && 1304 !transparent_hugepage_debug_cow()) { 1305 huge_gfp = alloc_hugepage_direct_gfpmask(vma); 1306 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1307 } else 1308 new_page = NULL; 1309 1310 if (likely(new_page)) { 1311 prep_transhuge_page(new_page); 1312 } else { 1313 if (!page) { 1314 split_huge_pmd(vma, vmf->pmd, vmf->address); 1315 ret |= VM_FAULT_FALLBACK; 1316 } else { 1317 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 1318 if (ret & VM_FAULT_OOM) { 1319 split_huge_pmd(vma, vmf->pmd, vmf->address); 1320 ret |= VM_FAULT_FALLBACK; 1321 } 1322 put_page(page); 1323 } 1324 count_vm_event(THP_FAULT_FALLBACK); 1325 goto out; 1326 } 1327 1328 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm, 1329 huge_gfp, &memcg, true))) { 1330 put_page(new_page); 1331 split_huge_pmd(vma, vmf->pmd, vmf->address); 1332 if (page) 1333 put_page(page); 1334 ret |= VM_FAULT_FALLBACK; 1335 count_vm_event(THP_FAULT_FALLBACK); 1336 goto out; 1337 } 1338 1339 count_vm_event(THP_FAULT_ALLOC); 1340 1341 if (!page) 1342 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 1343 else 1344 copy_user_huge_page(new_page, page, vmf->address, 1345 vma, HPAGE_PMD_NR); 1346 __SetPageUptodate(new_page); 1347 1348 mmu_notifier_range_init(&range, vma->vm_mm, haddr, 1349 haddr + HPAGE_PMD_SIZE); 1350 mmu_notifier_invalidate_range_start(&range); 1351 1352 spin_lock(vmf->ptl); 1353 if (page) 1354 put_page(page); 1355 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1356 spin_unlock(vmf->ptl); 1357 mem_cgroup_cancel_charge(new_page, memcg, true); 1358 put_page(new_page); 1359 goto out_mn; 1360 } else { 1361 pmd_t entry; 1362 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1363 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1364 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1365 page_add_new_anon_rmap(new_page, vma, haddr, true); 1366 mem_cgroup_commit_charge(new_page, memcg, false, true); 1367 lru_cache_add_active_or_unevictable(new_page, vma); 1368 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 1369 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1370 if (!page) { 1371 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1372 } else { 1373 VM_BUG_ON_PAGE(!PageHead(page), page); 1374 page_remove_rmap(page, true); 1375 put_page(page); 1376 } 1377 ret |= VM_FAULT_WRITE; 1378 } 1379 spin_unlock(vmf->ptl); 1380 out_mn: 1381 /* 1382 * No need to double call mmu_notifier->invalidate_range() callback as 1383 * the above pmdp_huge_clear_flush_notify() did already call it. 1384 */ 1385 mmu_notifier_invalidate_range_only_end(&range); 1386 out: 1387 return ret; 1388 out_unlock: 1389 spin_unlock(vmf->ptl); 1390 return ret; 1391 } 1392 1393 /* 1394 * FOLL_FORCE can write to even unwritable pmd's, but only 1395 * after we've gone through a COW cycle and they are dirty. 1396 */ 1397 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1398 { 1399 return pmd_write(pmd) || 1400 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 1401 } 1402 1403 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1404 unsigned long addr, 1405 pmd_t *pmd, 1406 unsigned int flags) 1407 { 1408 struct mm_struct *mm = vma->vm_mm; 1409 struct page *page = NULL; 1410 1411 assert_spin_locked(pmd_lockptr(mm, pmd)); 1412 1413 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1414 goto out; 1415 1416 /* Avoid dumping huge zero page */ 1417 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1418 return ERR_PTR(-EFAULT); 1419 1420 /* Full NUMA hinting faults to serialise migration in fault paths */ 1421 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1422 goto out; 1423 1424 page = pmd_page(*pmd); 1425 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1426 if (flags & FOLL_TOUCH) 1427 touch_pmd(vma, addr, pmd, flags); 1428 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1429 /* 1430 * We don't mlock() pte-mapped THPs. This way we can avoid 1431 * leaking mlocked pages into non-VM_LOCKED VMAs. 1432 * 1433 * For anon THP: 1434 * 1435 * In most cases the pmd is the only mapping of the page as we 1436 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1437 * writable private mappings in populate_vma_page_range(). 1438 * 1439 * The only scenario when we have the page shared here is if we 1440 * mlocking read-only mapping shared over fork(). We skip 1441 * mlocking such pages. 1442 * 1443 * For file THP: 1444 * 1445 * We can expect PageDoubleMap() to be stable under page lock: 1446 * for file pages we set it in page_add_file_rmap(), which 1447 * requires page to be locked. 1448 */ 1449 1450 if (PageAnon(page) && compound_mapcount(page) != 1) 1451 goto skip_mlock; 1452 if (PageDoubleMap(page) || !page->mapping) 1453 goto skip_mlock; 1454 if (!trylock_page(page)) 1455 goto skip_mlock; 1456 lru_add_drain(); 1457 if (page->mapping && !PageDoubleMap(page)) 1458 mlock_vma_page(page); 1459 unlock_page(page); 1460 } 1461 skip_mlock: 1462 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1463 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1464 if (flags & FOLL_GET) 1465 get_page(page); 1466 1467 out: 1468 return page; 1469 } 1470 1471 /* NUMA hinting page fault entry point for trans huge pmds */ 1472 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1473 { 1474 struct vm_area_struct *vma = vmf->vma; 1475 struct anon_vma *anon_vma = NULL; 1476 struct page *page; 1477 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1478 int page_nid = -1, this_nid = numa_node_id(); 1479 int target_nid, last_cpupid = -1; 1480 bool page_locked; 1481 bool migrated = false; 1482 bool was_writable; 1483 int flags = 0; 1484 1485 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1486 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1487 goto out_unlock; 1488 1489 /* 1490 * If there are potential migrations, wait for completion and retry 1491 * without disrupting NUMA hinting information. Do not relock and 1492 * check_same as the page may no longer be mapped. 1493 */ 1494 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1495 page = pmd_page(*vmf->pmd); 1496 if (!get_page_unless_zero(page)) 1497 goto out_unlock; 1498 spin_unlock(vmf->ptl); 1499 put_and_wait_on_page_locked(page); 1500 goto out; 1501 } 1502 1503 page = pmd_page(pmd); 1504 BUG_ON(is_huge_zero_page(page)); 1505 page_nid = page_to_nid(page); 1506 last_cpupid = page_cpupid_last(page); 1507 count_vm_numa_event(NUMA_HINT_FAULTS); 1508 if (page_nid == this_nid) { 1509 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1510 flags |= TNF_FAULT_LOCAL; 1511 } 1512 1513 /* See similar comment in do_numa_page for explanation */ 1514 if (!pmd_savedwrite(pmd)) 1515 flags |= TNF_NO_GROUP; 1516 1517 /* 1518 * Acquire the page lock to serialise THP migrations but avoid dropping 1519 * page_table_lock if at all possible 1520 */ 1521 page_locked = trylock_page(page); 1522 target_nid = mpol_misplaced(page, vma, haddr); 1523 if (target_nid == -1) { 1524 /* If the page was locked, there are no parallel migrations */ 1525 if (page_locked) 1526 goto clear_pmdnuma; 1527 } 1528 1529 /* Migration could have started since the pmd_trans_migrating check */ 1530 if (!page_locked) { 1531 page_nid = -1; 1532 if (!get_page_unless_zero(page)) 1533 goto out_unlock; 1534 spin_unlock(vmf->ptl); 1535 put_and_wait_on_page_locked(page); 1536 goto out; 1537 } 1538 1539 /* 1540 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1541 * to serialises splits 1542 */ 1543 get_page(page); 1544 spin_unlock(vmf->ptl); 1545 anon_vma = page_lock_anon_vma_read(page); 1546 1547 /* Confirm the PMD did not change while page_table_lock was released */ 1548 spin_lock(vmf->ptl); 1549 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1550 unlock_page(page); 1551 put_page(page); 1552 page_nid = -1; 1553 goto out_unlock; 1554 } 1555 1556 /* Bail if we fail to protect against THP splits for any reason */ 1557 if (unlikely(!anon_vma)) { 1558 put_page(page); 1559 page_nid = -1; 1560 goto clear_pmdnuma; 1561 } 1562 1563 /* 1564 * Since we took the NUMA fault, we must have observed the !accessible 1565 * bit. Make sure all other CPUs agree with that, to avoid them 1566 * modifying the page we're about to migrate. 1567 * 1568 * Must be done under PTL such that we'll observe the relevant 1569 * inc_tlb_flush_pending(). 1570 * 1571 * We are not sure a pending tlb flush here is for a huge page 1572 * mapping or not. Hence use the tlb range variant 1573 */ 1574 if (mm_tlb_flush_pending(vma->vm_mm)) { 1575 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1576 /* 1577 * change_huge_pmd() released the pmd lock before 1578 * invalidating the secondary MMUs sharing the primary 1579 * MMU pagetables (with ->invalidate_range()). The 1580 * mmu_notifier_invalidate_range_end() (which 1581 * internally calls ->invalidate_range()) in 1582 * change_pmd_range() will run after us, so we can't 1583 * rely on it here and we need an explicit invalidate. 1584 */ 1585 mmu_notifier_invalidate_range(vma->vm_mm, haddr, 1586 haddr + HPAGE_PMD_SIZE); 1587 } 1588 1589 /* 1590 * Migrate the THP to the requested node, returns with page unlocked 1591 * and access rights restored. 1592 */ 1593 spin_unlock(vmf->ptl); 1594 1595 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1596 vmf->pmd, pmd, vmf->address, page, target_nid); 1597 if (migrated) { 1598 flags |= TNF_MIGRATED; 1599 page_nid = target_nid; 1600 } else 1601 flags |= TNF_MIGRATE_FAIL; 1602 1603 goto out; 1604 clear_pmdnuma: 1605 BUG_ON(!PageLocked(page)); 1606 was_writable = pmd_savedwrite(pmd); 1607 pmd = pmd_modify(pmd, vma->vm_page_prot); 1608 pmd = pmd_mkyoung(pmd); 1609 if (was_writable) 1610 pmd = pmd_mkwrite(pmd); 1611 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1612 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1613 unlock_page(page); 1614 out_unlock: 1615 spin_unlock(vmf->ptl); 1616 1617 out: 1618 if (anon_vma) 1619 page_unlock_anon_vma_read(anon_vma); 1620 1621 if (page_nid != -1) 1622 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1623 flags); 1624 1625 return 0; 1626 } 1627 1628 /* 1629 * Return true if we do MADV_FREE successfully on entire pmd page. 1630 * Otherwise, return false. 1631 */ 1632 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1633 pmd_t *pmd, unsigned long addr, unsigned long next) 1634 { 1635 spinlock_t *ptl; 1636 pmd_t orig_pmd; 1637 struct page *page; 1638 struct mm_struct *mm = tlb->mm; 1639 bool ret = false; 1640 1641 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1642 1643 ptl = pmd_trans_huge_lock(pmd, vma); 1644 if (!ptl) 1645 goto out_unlocked; 1646 1647 orig_pmd = *pmd; 1648 if (is_huge_zero_pmd(orig_pmd)) 1649 goto out; 1650 1651 if (unlikely(!pmd_present(orig_pmd))) { 1652 VM_BUG_ON(thp_migration_supported() && 1653 !is_pmd_migration_entry(orig_pmd)); 1654 goto out; 1655 } 1656 1657 page = pmd_page(orig_pmd); 1658 /* 1659 * If other processes are mapping this page, we couldn't discard 1660 * the page unless they all do MADV_FREE so let's skip the page. 1661 */ 1662 if (page_mapcount(page) != 1) 1663 goto out; 1664 1665 if (!trylock_page(page)) 1666 goto out; 1667 1668 /* 1669 * If user want to discard part-pages of THP, split it so MADV_FREE 1670 * will deactivate only them. 1671 */ 1672 if (next - addr != HPAGE_PMD_SIZE) { 1673 get_page(page); 1674 spin_unlock(ptl); 1675 split_huge_page(page); 1676 unlock_page(page); 1677 put_page(page); 1678 goto out_unlocked; 1679 } 1680 1681 if (PageDirty(page)) 1682 ClearPageDirty(page); 1683 unlock_page(page); 1684 1685 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1686 pmdp_invalidate(vma, addr, pmd); 1687 orig_pmd = pmd_mkold(orig_pmd); 1688 orig_pmd = pmd_mkclean(orig_pmd); 1689 1690 set_pmd_at(mm, addr, pmd, orig_pmd); 1691 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1692 } 1693 1694 mark_page_lazyfree(page); 1695 ret = true; 1696 out: 1697 spin_unlock(ptl); 1698 out_unlocked: 1699 return ret; 1700 } 1701 1702 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1703 { 1704 pgtable_t pgtable; 1705 1706 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1707 pte_free(mm, pgtable); 1708 mm_dec_nr_ptes(mm); 1709 } 1710 1711 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1712 pmd_t *pmd, unsigned long addr) 1713 { 1714 pmd_t orig_pmd; 1715 spinlock_t *ptl; 1716 1717 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1718 1719 ptl = __pmd_trans_huge_lock(pmd, vma); 1720 if (!ptl) 1721 return 0; 1722 /* 1723 * For architectures like ppc64 we look at deposited pgtable 1724 * when calling pmdp_huge_get_and_clear. So do the 1725 * pgtable_trans_huge_withdraw after finishing pmdp related 1726 * operations. 1727 */ 1728 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1729 tlb->fullmm); 1730 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1731 if (vma_is_dax(vma)) { 1732 if (arch_needs_pgtable_deposit()) 1733 zap_deposited_table(tlb->mm, pmd); 1734 spin_unlock(ptl); 1735 if (is_huge_zero_pmd(orig_pmd)) 1736 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1737 } else if (is_huge_zero_pmd(orig_pmd)) { 1738 zap_deposited_table(tlb->mm, pmd); 1739 spin_unlock(ptl); 1740 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1741 } else { 1742 struct page *page = NULL; 1743 int flush_needed = 1; 1744 1745 if (pmd_present(orig_pmd)) { 1746 page = pmd_page(orig_pmd); 1747 page_remove_rmap(page, true); 1748 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1749 VM_BUG_ON_PAGE(!PageHead(page), page); 1750 } else if (thp_migration_supported()) { 1751 swp_entry_t entry; 1752 1753 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1754 entry = pmd_to_swp_entry(orig_pmd); 1755 page = pfn_to_page(swp_offset(entry)); 1756 flush_needed = 0; 1757 } else 1758 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1759 1760 if (PageAnon(page)) { 1761 zap_deposited_table(tlb->mm, pmd); 1762 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1763 } else { 1764 if (arch_needs_pgtable_deposit()) 1765 zap_deposited_table(tlb->mm, pmd); 1766 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1767 } 1768 1769 spin_unlock(ptl); 1770 if (flush_needed) 1771 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1772 } 1773 return 1; 1774 } 1775 1776 #ifndef pmd_move_must_withdraw 1777 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1778 spinlock_t *old_pmd_ptl, 1779 struct vm_area_struct *vma) 1780 { 1781 /* 1782 * With split pmd lock we also need to move preallocated 1783 * PTE page table if new_pmd is on different PMD page table. 1784 * 1785 * We also don't deposit and withdraw tables for file pages. 1786 */ 1787 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1788 } 1789 #endif 1790 1791 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1792 { 1793 #ifdef CONFIG_MEM_SOFT_DIRTY 1794 if (unlikely(is_pmd_migration_entry(pmd))) 1795 pmd = pmd_swp_mksoft_dirty(pmd); 1796 else if (pmd_present(pmd)) 1797 pmd = pmd_mksoft_dirty(pmd); 1798 #endif 1799 return pmd; 1800 } 1801 1802 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1803 unsigned long new_addr, unsigned long old_end, 1804 pmd_t *old_pmd, pmd_t *new_pmd) 1805 { 1806 spinlock_t *old_ptl, *new_ptl; 1807 pmd_t pmd; 1808 struct mm_struct *mm = vma->vm_mm; 1809 bool force_flush = false; 1810 1811 if ((old_addr & ~HPAGE_PMD_MASK) || 1812 (new_addr & ~HPAGE_PMD_MASK) || 1813 old_end - old_addr < HPAGE_PMD_SIZE) 1814 return false; 1815 1816 /* 1817 * The destination pmd shouldn't be established, free_pgtables() 1818 * should have release it. 1819 */ 1820 if (WARN_ON(!pmd_none(*new_pmd))) { 1821 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1822 return false; 1823 } 1824 1825 /* 1826 * We don't have to worry about the ordering of src and dst 1827 * ptlocks because exclusive mmap_sem prevents deadlock. 1828 */ 1829 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1830 if (old_ptl) { 1831 new_ptl = pmd_lockptr(mm, new_pmd); 1832 if (new_ptl != old_ptl) 1833 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1834 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1835 if (pmd_present(pmd)) 1836 force_flush = true; 1837 VM_BUG_ON(!pmd_none(*new_pmd)); 1838 1839 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1840 pgtable_t pgtable; 1841 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1842 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1843 } 1844 pmd = move_soft_dirty_pmd(pmd); 1845 set_pmd_at(mm, new_addr, new_pmd, pmd); 1846 if (force_flush) 1847 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1848 if (new_ptl != old_ptl) 1849 spin_unlock(new_ptl); 1850 spin_unlock(old_ptl); 1851 return true; 1852 } 1853 return false; 1854 } 1855 1856 /* 1857 * Returns 1858 * - 0 if PMD could not be locked 1859 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1860 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1861 */ 1862 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1863 unsigned long addr, pgprot_t newprot, int prot_numa) 1864 { 1865 struct mm_struct *mm = vma->vm_mm; 1866 spinlock_t *ptl; 1867 pmd_t entry; 1868 bool preserve_write; 1869 int ret; 1870 1871 ptl = __pmd_trans_huge_lock(pmd, vma); 1872 if (!ptl) 1873 return 0; 1874 1875 preserve_write = prot_numa && pmd_write(*pmd); 1876 ret = 1; 1877 1878 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1879 if (is_swap_pmd(*pmd)) { 1880 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1881 1882 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1883 if (is_write_migration_entry(entry)) { 1884 pmd_t newpmd; 1885 /* 1886 * A protection check is difficult so 1887 * just be safe and disable write 1888 */ 1889 make_migration_entry_read(&entry); 1890 newpmd = swp_entry_to_pmd(entry); 1891 if (pmd_swp_soft_dirty(*pmd)) 1892 newpmd = pmd_swp_mksoft_dirty(newpmd); 1893 set_pmd_at(mm, addr, pmd, newpmd); 1894 } 1895 goto unlock; 1896 } 1897 #endif 1898 1899 /* 1900 * Avoid trapping faults against the zero page. The read-only 1901 * data is likely to be read-cached on the local CPU and 1902 * local/remote hits to the zero page are not interesting. 1903 */ 1904 if (prot_numa && is_huge_zero_pmd(*pmd)) 1905 goto unlock; 1906 1907 if (prot_numa && pmd_protnone(*pmd)) 1908 goto unlock; 1909 1910 /* 1911 * In case prot_numa, we are under down_read(mmap_sem). It's critical 1912 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1913 * which is also under down_read(mmap_sem): 1914 * 1915 * CPU0: CPU1: 1916 * change_huge_pmd(prot_numa=1) 1917 * pmdp_huge_get_and_clear_notify() 1918 * madvise_dontneed() 1919 * zap_pmd_range() 1920 * pmd_trans_huge(*pmd) == 0 (without ptl) 1921 * // skip the pmd 1922 * set_pmd_at(); 1923 * // pmd is re-established 1924 * 1925 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1926 * which may break userspace. 1927 * 1928 * pmdp_invalidate() is required to make sure we don't miss 1929 * dirty/young flags set by hardware. 1930 */ 1931 entry = pmdp_invalidate(vma, addr, pmd); 1932 1933 entry = pmd_modify(entry, newprot); 1934 if (preserve_write) 1935 entry = pmd_mk_savedwrite(entry); 1936 ret = HPAGE_PMD_NR; 1937 set_pmd_at(mm, addr, pmd, entry); 1938 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1939 unlock: 1940 spin_unlock(ptl); 1941 return ret; 1942 } 1943 1944 /* 1945 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1946 * 1947 * Note that if it returns page table lock pointer, this routine returns without 1948 * unlocking page table lock. So callers must unlock it. 1949 */ 1950 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1951 { 1952 spinlock_t *ptl; 1953 ptl = pmd_lock(vma->vm_mm, pmd); 1954 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1955 pmd_devmap(*pmd))) 1956 return ptl; 1957 spin_unlock(ptl); 1958 return NULL; 1959 } 1960 1961 /* 1962 * Returns true if a given pud maps a thp, false otherwise. 1963 * 1964 * Note that if it returns true, this routine returns without unlocking page 1965 * table lock. So callers must unlock it. 1966 */ 1967 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1968 { 1969 spinlock_t *ptl; 1970 1971 ptl = pud_lock(vma->vm_mm, pud); 1972 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1973 return ptl; 1974 spin_unlock(ptl); 1975 return NULL; 1976 } 1977 1978 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1979 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1980 pud_t *pud, unsigned long addr) 1981 { 1982 pud_t orig_pud; 1983 spinlock_t *ptl; 1984 1985 ptl = __pud_trans_huge_lock(pud, vma); 1986 if (!ptl) 1987 return 0; 1988 /* 1989 * For architectures like ppc64 we look at deposited pgtable 1990 * when calling pudp_huge_get_and_clear. So do the 1991 * pgtable_trans_huge_withdraw after finishing pudp related 1992 * operations. 1993 */ 1994 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 1995 tlb->fullmm); 1996 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1997 if (vma_is_dax(vma)) { 1998 spin_unlock(ptl); 1999 /* No zero page support yet */ 2000 } else { 2001 /* No support for anonymous PUD pages yet */ 2002 BUG(); 2003 } 2004 return 1; 2005 } 2006 2007 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2008 unsigned long haddr) 2009 { 2010 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2011 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2012 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2013 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2014 2015 count_vm_event(THP_SPLIT_PUD); 2016 2017 pudp_huge_clear_flush_notify(vma, haddr, pud); 2018 } 2019 2020 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2021 unsigned long address) 2022 { 2023 spinlock_t *ptl; 2024 struct mmu_notifier_range range; 2025 2026 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PUD_MASK, 2027 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 2028 mmu_notifier_invalidate_range_start(&range); 2029 ptl = pud_lock(vma->vm_mm, pud); 2030 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2031 goto out; 2032 __split_huge_pud_locked(vma, pud, range.start); 2033 2034 out: 2035 spin_unlock(ptl); 2036 /* 2037 * No need to double call mmu_notifier->invalidate_range() callback as 2038 * the above pudp_huge_clear_flush_notify() did already call it. 2039 */ 2040 mmu_notifier_invalidate_range_only_end(&range); 2041 } 2042 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2043 2044 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2045 unsigned long haddr, pmd_t *pmd) 2046 { 2047 struct mm_struct *mm = vma->vm_mm; 2048 pgtable_t pgtable; 2049 pmd_t _pmd; 2050 int i; 2051 2052 /* 2053 * Leave pmd empty until pte is filled note that it is fine to delay 2054 * notification until mmu_notifier_invalidate_range_end() as we are 2055 * replacing a zero pmd write protected page with a zero pte write 2056 * protected page. 2057 * 2058 * See Documentation/vm/mmu_notifier.rst 2059 */ 2060 pmdp_huge_clear_flush(vma, haddr, pmd); 2061 2062 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2063 pmd_populate(mm, &_pmd, pgtable); 2064 2065 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2066 pte_t *pte, entry; 2067 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2068 entry = pte_mkspecial(entry); 2069 pte = pte_offset_map(&_pmd, haddr); 2070 VM_BUG_ON(!pte_none(*pte)); 2071 set_pte_at(mm, haddr, pte, entry); 2072 pte_unmap(pte); 2073 } 2074 smp_wmb(); /* make pte visible before pmd */ 2075 pmd_populate(mm, pmd, pgtable); 2076 } 2077 2078 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2079 unsigned long haddr, bool freeze) 2080 { 2081 struct mm_struct *mm = vma->vm_mm; 2082 struct page *page; 2083 pgtable_t pgtable; 2084 pmd_t old_pmd, _pmd; 2085 bool young, write, soft_dirty, pmd_migration = false; 2086 unsigned long addr; 2087 int i; 2088 2089 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2090 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2091 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2092 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2093 && !pmd_devmap(*pmd)); 2094 2095 count_vm_event(THP_SPLIT_PMD); 2096 2097 if (!vma_is_anonymous(vma)) { 2098 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2099 /* 2100 * We are going to unmap this huge page. So 2101 * just go ahead and zap it 2102 */ 2103 if (arch_needs_pgtable_deposit()) 2104 zap_deposited_table(mm, pmd); 2105 if (vma_is_dax(vma)) 2106 return; 2107 page = pmd_page(_pmd); 2108 if (!PageDirty(page) && pmd_dirty(_pmd)) 2109 set_page_dirty(page); 2110 if (!PageReferenced(page) && pmd_young(_pmd)) 2111 SetPageReferenced(page); 2112 page_remove_rmap(page, true); 2113 put_page(page); 2114 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2115 return; 2116 } else if (is_huge_zero_pmd(*pmd)) { 2117 /* 2118 * FIXME: Do we want to invalidate secondary mmu by calling 2119 * mmu_notifier_invalidate_range() see comments below inside 2120 * __split_huge_pmd() ? 2121 * 2122 * We are going from a zero huge page write protected to zero 2123 * small page also write protected so it does not seems useful 2124 * to invalidate secondary mmu at this time. 2125 */ 2126 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2127 } 2128 2129 /* 2130 * Up to this point the pmd is present and huge and userland has the 2131 * whole access to the hugepage during the split (which happens in 2132 * place). If we overwrite the pmd with the not-huge version pointing 2133 * to the pte here (which of course we could if all CPUs were bug 2134 * free), userland could trigger a small page size TLB miss on the 2135 * small sized TLB while the hugepage TLB entry is still established in 2136 * the huge TLB. Some CPU doesn't like that. 2137 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2138 * 383 on page 93. Intel should be safe but is also warns that it's 2139 * only safe if the permission and cache attributes of the two entries 2140 * loaded in the two TLB is identical (which should be the case here). 2141 * But it is generally safer to never allow small and huge TLB entries 2142 * for the same virtual address to be loaded simultaneously. So instead 2143 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2144 * current pmd notpresent (atomically because here the pmd_trans_huge 2145 * must remain set at all times on the pmd until the split is complete 2146 * for this pmd), then we flush the SMP TLB and finally we write the 2147 * non-huge version of the pmd entry with pmd_populate. 2148 */ 2149 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2150 2151 pmd_migration = is_pmd_migration_entry(old_pmd); 2152 if (unlikely(pmd_migration)) { 2153 swp_entry_t entry; 2154 2155 entry = pmd_to_swp_entry(old_pmd); 2156 page = pfn_to_page(swp_offset(entry)); 2157 write = is_write_migration_entry(entry); 2158 young = false; 2159 soft_dirty = pmd_swp_soft_dirty(old_pmd); 2160 } else { 2161 page = pmd_page(old_pmd); 2162 if (pmd_dirty(old_pmd)) 2163 SetPageDirty(page); 2164 write = pmd_write(old_pmd); 2165 young = pmd_young(old_pmd); 2166 soft_dirty = pmd_soft_dirty(old_pmd); 2167 } 2168 VM_BUG_ON_PAGE(!page_count(page), page); 2169 page_ref_add(page, HPAGE_PMD_NR - 1); 2170 2171 /* 2172 * Withdraw the table only after we mark the pmd entry invalid. 2173 * This's critical for some architectures (Power). 2174 */ 2175 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2176 pmd_populate(mm, &_pmd, pgtable); 2177 2178 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2179 pte_t entry, *pte; 2180 /* 2181 * Note that NUMA hinting access restrictions are not 2182 * transferred to avoid any possibility of altering 2183 * permissions across VMAs. 2184 */ 2185 if (freeze || pmd_migration) { 2186 swp_entry_t swp_entry; 2187 swp_entry = make_migration_entry(page + i, write); 2188 entry = swp_entry_to_pte(swp_entry); 2189 if (soft_dirty) 2190 entry = pte_swp_mksoft_dirty(entry); 2191 } else { 2192 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2193 entry = maybe_mkwrite(entry, vma); 2194 if (!write) 2195 entry = pte_wrprotect(entry); 2196 if (!young) 2197 entry = pte_mkold(entry); 2198 if (soft_dirty) 2199 entry = pte_mksoft_dirty(entry); 2200 } 2201 pte = pte_offset_map(&_pmd, addr); 2202 BUG_ON(!pte_none(*pte)); 2203 set_pte_at(mm, addr, pte, entry); 2204 atomic_inc(&page[i]._mapcount); 2205 pte_unmap(pte); 2206 } 2207 2208 /* 2209 * Set PG_double_map before dropping compound_mapcount to avoid 2210 * false-negative page_mapped(). 2211 */ 2212 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2213 for (i = 0; i < HPAGE_PMD_NR; i++) 2214 atomic_inc(&page[i]._mapcount); 2215 } 2216 2217 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2218 /* Last compound_mapcount is gone. */ 2219 __dec_node_page_state(page, NR_ANON_THPS); 2220 if (TestClearPageDoubleMap(page)) { 2221 /* No need in mapcount reference anymore */ 2222 for (i = 0; i < HPAGE_PMD_NR; i++) 2223 atomic_dec(&page[i]._mapcount); 2224 } 2225 } 2226 2227 smp_wmb(); /* make pte visible before pmd */ 2228 pmd_populate(mm, pmd, pgtable); 2229 2230 if (freeze) { 2231 for (i = 0; i < HPAGE_PMD_NR; i++) { 2232 page_remove_rmap(page + i, false); 2233 put_page(page + i); 2234 } 2235 } 2236 } 2237 2238 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2239 unsigned long address, bool freeze, struct page *page) 2240 { 2241 spinlock_t *ptl; 2242 struct mmu_notifier_range range; 2243 2244 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PMD_MASK, 2245 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2246 mmu_notifier_invalidate_range_start(&range); 2247 ptl = pmd_lock(vma->vm_mm, pmd); 2248 2249 /* 2250 * If caller asks to setup a migration entries, we need a page to check 2251 * pmd against. Otherwise we can end up replacing wrong page. 2252 */ 2253 VM_BUG_ON(freeze && !page); 2254 if (page && page != pmd_page(*pmd)) 2255 goto out; 2256 2257 if (pmd_trans_huge(*pmd)) { 2258 page = pmd_page(*pmd); 2259 if (PageMlocked(page)) 2260 clear_page_mlock(page); 2261 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2262 goto out; 2263 __split_huge_pmd_locked(vma, pmd, range.start, freeze); 2264 out: 2265 spin_unlock(ptl); 2266 /* 2267 * No need to double call mmu_notifier->invalidate_range() callback. 2268 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2269 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2270 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2271 * fault will trigger a flush_notify before pointing to a new page 2272 * (it is fine if the secondary mmu keeps pointing to the old zero 2273 * page in the meantime) 2274 * 3) Split a huge pmd into pte pointing to the same page. No need 2275 * to invalidate secondary tlb entry they are all still valid. 2276 * any further changes to individual pte will notify. So no need 2277 * to call mmu_notifier->invalidate_range() 2278 */ 2279 mmu_notifier_invalidate_range_only_end(&range); 2280 } 2281 2282 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2283 bool freeze, struct page *page) 2284 { 2285 pgd_t *pgd; 2286 p4d_t *p4d; 2287 pud_t *pud; 2288 pmd_t *pmd; 2289 2290 pgd = pgd_offset(vma->vm_mm, address); 2291 if (!pgd_present(*pgd)) 2292 return; 2293 2294 p4d = p4d_offset(pgd, address); 2295 if (!p4d_present(*p4d)) 2296 return; 2297 2298 pud = pud_offset(p4d, address); 2299 if (!pud_present(*pud)) 2300 return; 2301 2302 pmd = pmd_offset(pud, address); 2303 2304 __split_huge_pmd(vma, pmd, address, freeze, page); 2305 } 2306 2307 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2308 unsigned long start, 2309 unsigned long end, 2310 long adjust_next) 2311 { 2312 /* 2313 * If the new start address isn't hpage aligned and it could 2314 * previously contain an hugepage: check if we need to split 2315 * an huge pmd. 2316 */ 2317 if (start & ~HPAGE_PMD_MASK && 2318 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2319 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2320 split_huge_pmd_address(vma, start, false, NULL); 2321 2322 /* 2323 * If the new end address isn't hpage aligned and it could 2324 * previously contain an hugepage: check if we need to split 2325 * an huge pmd. 2326 */ 2327 if (end & ~HPAGE_PMD_MASK && 2328 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2329 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2330 split_huge_pmd_address(vma, end, false, NULL); 2331 2332 /* 2333 * If we're also updating the vma->vm_next->vm_start, if the new 2334 * vm_next->vm_start isn't page aligned and it could previously 2335 * contain an hugepage: check if we need to split an huge pmd. 2336 */ 2337 if (adjust_next > 0) { 2338 struct vm_area_struct *next = vma->vm_next; 2339 unsigned long nstart = next->vm_start; 2340 nstart += adjust_next << PAGE_SHIFT; 2341 if (nstart & ~HPAGE_PMD_MASK && 2342 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2343 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2344 split_huge_pmd_address(next, nstart, false, NULL); 2345 } 2346 } 2347 2348 static void unmap_page(struct page *page) 2349 { 2350 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2351 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2352 bool unmap_success; 2353 2354 VM_BUG_ON_PAGE(!PageHead(page), page); 2355 2356 if (PageAnon(page)) 2357 ttu_flags |= TTU_SPLIT_FREEZE; 2358 2359 unmap_success = try_to_unmap(page, ttu_flags); 2360 VM_BUG_ON_PAGE(!unmap_success, page); 2361 } 2362 2363 static void remap_page(struct page *page) 2364 { 2365 int i; 2366 if (PageTransHuge(page)) { 2367 remove_migration_ptes(page, page, true); 2368 } else { 2369 for (i = 0; i < HPAGE_PMD_NR; i++) 2370 remove_migration_ptes(page + i, page + i, true); 2371 } 2372 } 2373 2374 static void __split_huge_page_tail(struct page *head, int tail, 2375 struct lruvec *lruvec, struct list_head *list) 2376 { 2377 struct page *page_tail = head + tail; 2378 2379 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2380 2381 /* 2382 * Clone page flags before unfreezing refcount. 2383 * 2384 * After successful get_page_unless_zero() might follow flags change, 2385 * for exmaple lock_page() which set PG_waiters. 2386 */ 2387 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2388 page_tail->flags |= (head->flags & 2389 ((1L << PG_referenced) | 2390 (1L << PG_swapbacked) | 2391 (1L << PG_swapcache) | 2392 (1L << PG_mlocked) | 2393 (1L << PG_uptodate) | 2394 (1L << PG_active) | 2395 (1L << PG_workingset) | 2396 (1L << PG_locked) | 2397 (1L << PG_unevictable) | 2398 (1L << PG_dirty))); 2399 2400 /* ->mapping in first tail page is compound_mapcount */ 2401 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2402 page_tail); 2403 page_tail->mapping = head->mapping; 2404 page_tail->index = head->index + tail; 2405 2406 /* Page flags must be visible before we make the page non-compound. */ 2407 smp_wmb(); 2408 2409 /* 2410 * Clear PageTail before unfreezing page refcount. 2411 * 2412 * After successful get_page_unless_zero() might follow put_page() 2413 * which needs correct compound_head(). 2414 */ 2415 clear_compound_head(page_tail); 2416 2417 /* Finally unfreeze refcount. Additional reference from page cache. */ 2418 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2419 PageSwapCache(head))); 2420 2421 if (page_is_young(head)) 2422 set_page_young(page_tail); 2423 if (page_is_idle(head)) 2424 set_page_idle(page_tail); 2425 2426 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2427 2428 /* 2429 * always add to the tail because some iterators expect new 2430 * pages to show after the currently processed elements - e.g. 2431 * migrate_pages 2432 */ 2433 lru_add_page_tail(head, page_tail, lruvec, list); 2434 } 2435 2436 static void __split_huge_page(struct page *page, struct list_head *list, 2437 pgoff_t end, unsigned long flags) 2438 { 2439 struct page *head = compound_head(page); 2440 struct zone *zone = page_zone(head); 2441 struct lruvec *lruvec; 2442 int i; 2443 2444 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2445 2446 /* complete memcg works before add pages to LRU */ 2447 mem_cgroup_split_huge_fixup(head); 2448 2449 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2450 __split_huge_page_tail(head, i, lruvec, list); 2451 /* Some pages can be beyond i_size: drop them from page cache */ 2452 if (head[i].index >= end) { 2453 ClearPageDirty(head + i); 2454 __delete_from_page_cache(head + i, NULL); 2455 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2456 shmem_uncharge(head->mapping->host, 1); 2457 put_page(head + i); 2458 } 2459 } 2460 2461 ClearPageCompound(head); 2462 /* See comment in __split_huge_page_tail() */ 2463 if (PageAnon(head)) { 2464 /* Additional pin to swap cache */ 2465 if (PageSwapCache(head)) 2466 page_ref_add(head, 2); 2467 else 2468 page_ref_inc(head); 2469 } else { 2470 /* Additional pin to page cache */ 2471 page_ref_add(head, 2); 2472 xa_unlock(&head->mapping->i_pages); 2473 } 2474 2475 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2476 2477 remap_page(head); 2478 2479 for (i = 0; i < HPAGE_PMD_NR; i++) { 2480 struct page *subpage = head + i; 2481 if (subpage == page) 2482 continue; 2483 unlock_page(subpage); 2484 2485 /* 2486 * Subpages may be freed if there wasn't any mapping 2487 * like if add_to_swap() is running on a lru page that 2488 * had its mapping zapped. And freeing these pages 2489 * requires taking the lru_lock so we do the put_page 2490 * of the tail pages after the split is complete. 2491 */ 2492 put_page(subpage); 2493 } 2494 } 2495 2496 int total_mapcount(struct page *page) 2497 { 2498 int i, compound, ret; 2499 2500 VM_BUG_ON_PAGE(PageTail(page), page); 2501 2502 if (likely(!PageCompound(page))) 2503 return atomic_read(&page->_mapcount) + 1; 2504 2505 compound = compound_mapcount(page); 2506 if (PageHuge(page)) 2507 return compound; 2508 ret = compound; 2509 for (i = 0; i < HPAGE_PMD_NR; i++) 2510 ret += atomic_read(&page[i]._mapcount) + 1; 2511 /* File pages has compound_mapcount included in _mapcount */ 2512 if (!PageAnon(page)) 2513 return ret - compound * HPAGE_PMD_NR; 2514 if (PageDoubleMap(page)) 2515 ret -= HPAGE_PMD_NR; 2516 return ret; 2517 } 2518 2519 /* 2520 * This calculates accurately how many mappings a transparent hugepage 2521 * has (unlike page_mapcount() which isn't fully accurate). This full 2522 * accuracy is primarily needed to know if copy-on-write faults can 2523 * reuse the page and change the mapping to read-write instead of 2524 * copying them. At the same time this returns the total_mapcount too. 2525 * 2526 * The function returns the highest mapcount any one of the subpages 2527 * has. If the return value is one, even if different processes are 2528 * mapping different subpages of the transparent hugepage, they can 2529 * all reuse it, because each process is reusing a different subpage. 2530 * 2531 * The total_mapcount is instead counting all virtual mappings of the 2532 * subpages. If the total_mapcount is equal to "one", it tells the 2533 * caller all mappings belong to the same "mm" and in turn the 2534 * anon_vma of the transparent hugepage can become the vma->anon_vma 2535 * local one as no other process may be mapping any of the subpages. 2536 * 2537 * It would be more accurate to replace page_mapcount() with 2538 * page_trans_huge_mapcount(), however we only use 2539 * page_trans_huge_mapcount() in the copy-on-write faults where we 2540 * need full accuracy to avoid breaking page pinning, because 2541 * page_trans_huge_mapcount() is slower than page_mapcount(). 2542 */ 2543 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2544 { 2545 int i, ret, _total_mapcount, mapcount; 2546 2547 /* hugetlbfs shouldn't call it */ 2548 VM_BUG_ON_PAGE(PageHuge(page), page); 2549 2550 if (likely(!PageTransCompound(page))) { 2551 mapcount = atomic_read(&page->_mapcount) + 1; 2552 if (total_mapcount) 2553 *total_mapcount = mapcount; 2554 return mapcount; 2555 } 2556 2557 page = compound_head(page); 2558 2559 _total_mapcount = ret = 0; 2560 for (i = 0; i < HPAGE_PMD_NR; i++) { 2561 mapcount = atomic_read(&page[i]._mapcount) + 1; 2562 ret = max(ret, mapcount); 2563 _total_mapcount += mapcount; 2564 } 2565 if (PageDoubleMap(page)) { 2566 ret -= 1; 2567 _total_mapcount -= HPAGE_PMD_NR; 2568 } 2569 mapcount = compound_mapcount(page); 2570 ret += mapcount; 2571 _total_mapcount += mapcount; 2572 if (total_mapcount) 2573 *total_mapcount = _total_mapcount; 2574 return ret; 2575 } 2576 2577 /* Racy check whether the huge page can be split */ 2578 bool can_split_huge_page(struct page *page, int *pextra_pins) 2579 { 2580 int extra_pins; 2581 2582 /* Additional pins from page cache */ 2583 if (PageAnon(page)) 2584 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2585 else 2586 extra_pins = HPAGE_PMD_NR; 2587 if (pextra_pins) 2588 *pextra_pins = extra_pins; 2589 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2590 } 2591 2592 /* 2593 * This function splits huge page into normal pages. @page can point to any 2594 * subpage of huge page to split. Split doesn't change the position of @page. 2595 * 2596 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2597 * The huge page must be locked. 2598 * 2599 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2600 * 2601 * Both head page and tail pages will inherit mapping, flags, and so on from 2602 * the hugepage. 2603 * 2604 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2605 * they are not mapped. 2606 * 2607 * Returns 0 if the hugepage is split successfully. 2608 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2609 * us. 2610 */ 2611 int split_huge_page_to_list(struct page *page, struct list_head *list) 2612 { 2613 struct page *head = compound_head(page); 2614 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2615 struct anon_vma *anon_vma = NULL; 2616 struct address_space *mapping = NULL; 2617 int count, mapcount, extra_pins, ret; 2618 bool mlocked; 2619 unsigned long flags; 2620 pgoff_t end; 2621 2622 VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2623 VM_BUG_ON_PAGE(!PageLocked(page), page); 2624 VM_BUG_ON_PAGE(!PageCompound(page), page); 2625 2626 if (PageWriteback(page)) 2627 return -EBUSY; 2628 2629 if (PageAnon(head)) { 2630 /* 2631 * The caller does not necessarily hold an mmap_sem that would 2632 * prevent the anon_vma disappearing so we first we take a 2633 * reference to it and then lock the anon_vma for write. This 2634 * is similar to page_lock_anon_vma_read except the write lock 2635 * is taken to serialise against parallel split or collapse 2636 * operations. 2637 */ 2638 anon_vma = page_get_anon_vma(head); 2639 if (!anon_vma) { 2640 ret = -EBUSY; 2641 goto out; 2642 } 2643 end = -1; 2644 mapping = NULL; 2645 anon_vma_lock_write(anon_vma); 2646 } else { 2647 mapping = head->mapping; 2648 2649 /* Truncated ? */ 2650 if (!mapping) { 2651 ret = -EBUSY; 2652 goto out; 2653 } 2654 2655 anon_vma = NULL; 2656 i_mmap_lock_read(mapping); 2657 2658 /* 2659 *__split_huge_page() may need to trim off pages beyond EOF: 2660 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2661 * which cannot be nested inside the page tree lock. So note 2662 * end now: i_size itself may be changed at any moment, but 2663 * head page lock is good enough to serialize the trimming. 2664 */ 2665 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2666 } 2667 2668 /* 2669 * Racy check if we can split the page, before unmap_page() will 2670 * split PMDs 2671 */ 2672 if (!can_split_huge_page(head, &extra_pins)) { 2673 ret = -EBUSY; 2674 goto out_unlock; 2675 } 2676 2677 mlocked = PageMlocked(page); 2678 unmap_page(head); 2679 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2680 2681 /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2682 if (mlocked) 2683 lru_add_drain(); 2684 2685 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2686 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2687 2688 if (mapping) { 2689 XA_STATE(xas, &mapping->i_pages, page_index(head)); 2690 2691 /* 2692 * Check if the head page is present in page cache. 2693 * We assume all tail are present too, if head is there. 2694 */ 2695 xa_lock(&mapping->i_pages); 2696 if (xas_load(&xas) != head) 2697 goto fail; 2698 } 2699 2700 /* Prevent deferred_split_scan() touching ->_refcount */ 2701 spin_lock(&pgdata->split_queue_lock); 2702 count = page_count(head); 2703 mapcount = total_mapcount(head); 2704 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2705 if (!list_empty(page_deferred_list(head))) { 2706 pgdata->split_queue_len--; 2707 list_del(page_deferred_list(head)); 2708 } 2709 if (mapping) 2710 __dec_node_page_state(page, NR_SHMEM_THPS); 2711 spin_unlock(&pgdata->split_queue_lock); 2712 __split_huge_page(page, list, end, flags); 2713 if (PageSwapCache(head)) { 2714 swp_entry_t entry = { .val = page_private(head) }; 2715 2716 ret = split_swap_cluster(entry); 2717 } else 2718 ret = 0; 2719 } else { 2720 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2721 pr_alert("total_mapcount: %u, page_count(): %u\n", 2722 mapcount, count); 2723 if (PageTail(page)) 2724 dump_page(head, NULL); 2725 dump_page(page, "total_mapcount(head) > 0"); 2726 BUG(); 2727 } 2728 spin_unlock(&pgdata->split_queue_lock); 2729 fail: if (mapping) 2730 xa_unlock(&mapping->i_pages); 2731 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2732 remap_page(head); 2733 ret = -EBUSY; 2734 } 2735 2736 out_unlock: 2737 if (anon_vma) { 2738 anon_vma_unlock_write(anon_vma); 2739 put_anon_vma(anon_vma); 2740 } 2741 if (mapping) 2742 i_mmap_unlock_read(mapping); 2743 out: 2744 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2745 return ret; 2746 } 2747 2748 void free_transhuge_page(struct page *page) 2749 { 2750 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2751 unsigned long flags; 2752 2753 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2754 if (!list_empty(page_deferred_list(page))) { 2755 pgdata->split_queue_len--; 2756 list_del(page_deferred_list(page)); 2757 } 2758 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2759 free_compound_page(page); 2760 } 2761 2762 void deferred_split_huge_page(struct page *page) 2763 { 2764 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2765 unsigned long flags; 2766 2767 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2768 2769 spin_lock_irqsave(&pgdata->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), &pgdata->split_queue); 2773 pgdata->split_queue_len++; 2774 } 2775 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2776 } 2777 2778 static unsigned long deferred_split_count(struct shrinker *shrink, 2779 struct shrink_control *sc) 2780 { 2781 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2782 return READ_ONCE(pgdata->split_queue_len); 2783 } 2784 2785 static unsigned long deferred_split_scan(struct shrinker *shrink, 2786 struct shrink_control *sc) 2787 { 2788 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2789 unsigned long flags; 2790 LIST_HEAD(list), *pos, *next; 2791 struct page *page; 2792 int split = 0; 2793 2794 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2795 /* Take pin on all head pages to avoid freeing them under us */ 2796 list_for_each_safe(pos, next, &pgdata->split_queue) { 2797 page = list_entry((void *)pos, struct page, mapping); 2798 page = compound_head(page); 2799 if (get_page_unless_zero(page)) { 2800 list_move(page_deferred_list(page), &list); 2801 } else { 2802 /* We lost race with put_compound_page() */ 2803 list_del_init(page_deferred_list(page)); 2804 pgdata->split_queue_len--; 2805 } 2806 if (!--sc->nr_to_scan) 2807 break; 2808 } 2809 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2810 2811 list_for_each_safe(pos, next, &list) { 2812 page = list_entry((void *)pos, struct page, mapping); 2813 if (!trylock_page(page)) 2814 goto next; 2815 /* split_huge_page() removes page from list on success */ 2816 if (!split_huge_page(page)) 2817 split++; 2818 unlock_page(page); 2819 next: 2820 put_page(page); 2821 } 2822 2823 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2824 list_splice_tail(&list, &pgdata->split_queue); 2825 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2826 2827 /* 2828 * Stop shrinker if we didn't split any page, but the queue is empty. 2829 * This can happen if pages were freed under us. 2830 */ 2831 if (!split && list_empty(&pgdata->split_queue)) 2832 return SHRINK_STOP; 2833 return split; 2834 } 2835 2836 static struct shrinker deferred_split_shrinker = { 2837 .count_objects = deferred_split_count, 2838 .scan_objects = deferred_split_scan, 2839 .seeks = DEFAULT_SEEKS, 2840 .flags = SHRINKER_NUMA_AWARE, 2841 }; 2842 2843 #ifdef CONFIG_DEBUG_FS 2844 static int split_huge_pages_set(void *data, u64 val) 2845 { 2846 struct zone *zone; 2847 struct page *page; 2848 unsigned long pfn, max_zone_pfn; 2849 unsigned long total = 0, split = 0; 2850 2851 if (val != 1) 2852 return -EINVAL; 2853 2854 for_each_populated_zone(zone) { 2855 max_zone_pfn = zone_end_pfn(zone); 2856 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2857 if (!pfn_valid(pfn)) 2858 continue; 2859 2860 page = pfn_to_page(pfn); 2861 if (!get_page_unless_zero(page)) 2862 continue; 2863 2864 if (zone != page_zone(page)) 2865 goto next; 2866 2867 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2868 goto next; 2869 2870 total++; 2871 lock_page(page); 2872 if (!split_huge_page(page)) 2873 split++; 2874 unlock_page(page); 2875 next: 2876 put_page(page); 2877 } 2878 } 2879 2880 pr_info("%lu of %lu THP split\n", split, total); 2881 2882 return 0; 2883 } 2884 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2885 "%llu\n"); 2886 2887 static int __init split_huge_pages_debugfs(void) 2888 { 2889 void *ret; 2890 2891 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2892 &split_huge_pages_fops); 2893 if (!ret) 2894 pr_warn("Failed to create split_huge_pages in debugfs"); 2895 return 0; 2896 } 2897 late_initcall(split_huge_pages_debugfs); 2898 #endif 2899 2900 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2901 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2902 struct page *page) 2903 { 2904 struct vm_area_struct *vma = pvmw->vma; 2905 struct mm_struct *mm = vma->vm_mm; 2906 unsigned long address = pvmw->address; 2907 pmd_t pmdval; 2908 swp_entry_t entry; 2909 pmd_t pmdswp; 2910 2911 if (!(pvmw->pmd && !pvmw->pte)) 2912 return; 2913 2914 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2915 pmdval = *pvmw->pmd; 2916 pmdp_invalidate(vma, address, pvmw->pmd); 2917 if (pmd_dirty(pmdval)) 2918 set_page_dirty(page); 2919 entry = make_migration_entry(page, pmd_write(pmdval)); 2920 pmdswp = swp_entry_to_pmd(entry); 2921 if (pmd_soft_dirty(pmdval)) 2922 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2923 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2924 page_remove_rmap(page, true); 2925 put_page(page); 2926 } 2927 2928 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2929 { 2930 struct vm_area_struct *vma = pvmw->vma; 2931 struct mm_struct *mm = vma->vm_mm; 2932 unsigned long address = pvmw->address; 2933 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2934 pmd_t pmde; 2935 swp_entry_t entry; 2936 2937 if (!(pvmw->pmd && !pvmw->pte)) 2938 return; 2939 2940 entry = pmd_to_swp_entry(*pvmw->pmd); 2941 get_page(new); 2942 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2943 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2944 pmde = pmd_mksoft_dirty(pmde); 2945 if (is_write_migration_entry(entry)) 2946 pmde = maybe_pmd_mkwrite(pmde, vma); 2947 2948 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2949 if (PageAnon(new)) 2950 page_add_anon_rmap(new, vma, mmun_start, true); 2951 else 2952 page_add_file_rmap(new, true); 2953 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2954 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) 2955 mlock_vma_page(new); 2956 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2957 } 2958 #endif 2959