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 static struct page *get_huge_zero_page(void) 66 { 67 struct page *zero_page; 68 retry: 69 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 70 return READ_ONCE(huge_zero_page); 71 72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 73 HPAGE_PMD_ORDER); 74 if (!zero_page) { 75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 76 return NULL; 77 } 78 count_vm_event(THP_ZERO_PAGE_ALLOC); 79 preempt_disable(); 80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 81 preempt_enable(); 82 __free_pages(zero_page, compound_order(zero_page)); 83 goto retry; 84 } 85 86 /* We take additional reference here. It will be put back by shrinker */ 87 atomic_set(&huge_zero_refcount, 2); 88 preempt_enable(); 89 return READ_ONCE(huge_zero_page); 90 } 91 92 static void put_huge_zero_page(void) 93 { 94 /* 95 * Counter should never go to zero here. Only shrinker can put 96 * last reference. 97 */ 98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 99 } 100 101 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 102 { 103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 104 return READ_ONCE(huge_zero_page); 105 106 if (!get_huge_zero_page()) 107 return NULL; 108 109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 110 put_huge_zero_page(); 111 112 return READ_ONCE(huge_zero_page); 113 } 114 115 void mm_put_huge_zero_page(struct mm_struct *mm) 116 { 117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 118 put_huge_zero_page(); 119 } 120 121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 122 struct shrink_control *sc) 123 { 124 /* we can free zero page only if last reference remains */ 125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 126 } 127 128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 129 struct shrink_control *sc) 130 { 131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 132 struct page *zero_page = xchg(&huge_zero_page, NULL); 133 BUG_ON(zero_page == NULL); 134 __free_pages(zero_page, compound_order(zero_page)); 135 return HPAGE_PMD_NR; 136 } 137 138 return 0; 139 } 140 141 static struct shrinker huge_zero_page_shrinker = { 142 .count_objects = shrink_huge_zero_page_count, 143 .scan_objects = shrink_huge_zero_page_scan, 144 .seeks = DEFAULT_SEEKS, 145 }; 146 147 #ifdef CONFIG_SYSFS 148 static ssize_t enabled_show(struct kobject *kobj, 149 struct kobj_attribute *attr, char *buf) 150 { 151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 152 return sprintf(buf, "[always] madvise never\n"); 153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 154 return sprintf(buf, "always [madvise] never\n"); 155 else 156 return sprintf(buf, "always madvise [never]\n"); 157 } 158 159 static ssize_t enabled_store(struct kobject *kobj, 160 struct kobj_attribute *attr, 161 const char *buf, size_t count) 162 { 163 ssize_t ret = count; 164 165 if (!memcmp("always", buf, 166 min(sizeof("always")-1, count))) { 167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 169 } else if (!memcmp("madvise", buf, 170 min(sizeof("madvise")-1, count))) { 171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 173 } else if (!memcmp("never", buf, 174 min(sizeof("never")-1, count))) { 175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 177 } else 178 ret = -EINVAL; 179 180 if (ret > 0) { 181 int err = start_stop_khugepaged(); 182 if (err) 183 ret = err; 184 } 185 return ret; 186 } 187 static struct kobj_attribute enabled_attr = 188 __ATTR(enabled, 0644, enabled_show, enabled_store); 189 190 ssize_t single_hugepage_flag_show(struct kobject *kobj, 191 struct kobj_attribute *attr, char *buf, 192 enum transparent_hugepage_flag flag) 193 { 194 return sprintf(buf, "%d\n", 195 !!test_bit(flag, &transparent_hugepage_flags)); 196 } 197 198 ssize_t single_hugepage_flag_store(struct kobject *kobj, 199 struct kobj_attribute *attr, 200 const char *buf, size_t count, 201 enum transparent_hugepage_flag flag) 202 { 203 unsigned long value; 204 int ret; 205 206 ret = kstrtoul(buf, 10, &value); 207 if (ret < 0) 208 return ret; 209 if (value > 1) 210 return -EINVAL; 211 212 if (value) 213 set_bit(flag, &transparent_hugepage_flags); 214 else 215 clear_bit(flag, &transparent_hugepage_flags); 216 217 return count; 218 } 219 220 static ssize_t defrag_show(struct kobject *kobj, 221 struct kobj_attribute *attr, char *buf) 222 { 223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 224 return sprintf(buf, "[always] defer defer+madvise madvise never\n"); 225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 226 return sprintf(buf, "always [defer] defer+madvise madvise never\n"); 227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 228 return sprintf(buf, "always defer [defer+madvise] madvise never\n"); 229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 230 return sprintf(buf, "always defer defer+madvise [madvise] never\n"); 231 return sprintf(buf, "always defer defer+madvise madvise [never]\n"); 232 } 233 234 static ssize_t defrag_store(struct kobject *kobj, 235 struct kobj_attribute *attr, 236 const char *buf, size_t count) 237 { 238 if (!memcmp("always", buf, 239 min(sizeof("always")-1, count))) { 240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 244 } else if (!memcmp("defer+madvise", buf, 245 min(sizeof("defer+madvise")-1, count))) { 246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 250 } else if (!memcmp("defer", buf, 251 min(sizeof("defer")-1, count))) { 252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 256 } else if (!memcmp("madvise", buf, 257 min(sizeof("madvise")-1, count))) { 258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 262 } else if (!memcmp("never", buf, 263 min(sizeof("never")-1, count))) { 264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 268 } else 269 return -EINVAL; 270 271 return count; 272 } 273 static struct kobj_attribute defrag_attr = 274 __ATTR(defrag, 0644, defrag_show, defrag_store); 275 276 static ssize_t use_zero_page_show(struct kobject *kobj, 277 struct kobj_attribute *attr, char *buf) 278 { 279 return single_hugepage_flag_show(kobj, attr, buf, 280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 281 } 282 static ssize_t use_zero_page_store(struct kobject *kobj, 283 struct kobj_attribute *attr, const char *buf, size_t count) 284 { 285 return single_hugepage_flag_store(kobj, attr, buf, count, 286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 287 } 288 static struct kobj_attribute use_zero_page_attr = 289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 290 291 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 292 struct kobj_attribute *attr, char *buf) 293 { 294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); 295 } 296 static struct kobj_attribute hpage_pmd_size_attr = 297 __ATTR_RO(hpage_pmd_size); 298 299 #ifdef CONFIG_DEBUG_VM 300 static ssize_t debug_cow_show(struct kobject *kobj, 301 struct kobj_attribute *attr, char *buf) 302 { 303 return single_hugepage_flag_show(kobj, attr, buf, 304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 305 } 306 static ssize_t debug_cow_store(struct kobject *kobj, 307 struct kobj_attribute *attr, 308 const char *buf, size_t count) 309 { 310 return single_hugepage_flag_store(kobj, attr, buf, count, 311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 312 } 313 static struct kobj_attribute debug_cow_attr = 314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 315 #endif /* CONFIG_DEBUG_VM */ 316 317 static struct attribute *hugepage_attr[] = { 318 &enabled_attr.attr, 319 &defrag_attr.attr, 320 &use_zero_page_attr.attr, 321 &hpage_pmd_size_attr.attr, 322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 323 &shmem_enabled_attr.attr, 324 #endif 325 #ifdef CONFIG_DEBUG_VM 326 &debug_cow_attr.attr, 327 #endif 328 NULL, 329 }; 330 331 static const struct attribute_group hugepage_attr_group = { 332 .attrs = hugepage_attr, 333 }; 334 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 336 { 337 int err; 338 339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 340 if (unlikely(!*hugepage_kobj)) { 341 pr_err("failed to create transparent hugepage kobject\n"); 342 return -ENOMEM; 343 } 344 345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 346 if (err) { 347 pr_err("failed to register transparent hugepage group\n"); 348 goto delete_obj; 349 } 350 351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 352 if (err) { 353 pr_err("failed to register transparent hugepage group\n"); 354 goto remove_hp_group; 355 } 356 357 return 0; 358 359 remove_hp_group: 360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 361 delete_obj: 362 kobject_put(*hugepage_kobj); 363 return err; 364 } 365 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 367 { 368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 370 kobject_put(hugepage_kobj); 371 } 372 #else 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 374 { 375 return 0; 376 } 377 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 379 { 380 } 381 #endif /* CONFIG_SYSFS */ 382 383 static int __init hugepage_init(void) 384 { 385 int err; 386 struct kobject *hugepage_kobj; 387 388 if (!has_transparent_hugepage()) { 389 transparent_hugepage_flags = 0; 390 return -EINVAL; 391 } 392 393 /* 394 * hugepages can't be allocated by the buddy allocator 395 */ 396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 397 /* 398 * we use page->mapping and page->index in second tail page 399 * as list_head: assuming THP order >= 2 400 */ 401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 402 403 err = hugepage_init_sysfs(&hugepage_kobj); 404 if (err) 405 goto err_sysfs; 406 407 err = khugepaged_init(); 408 if (err) 409 goto err_slab; 410 411 err = register_shrinker(&huge_zero_page_shrinker); 412 if (err) 413 goto err_hzp_shrinker; 414 err = register_shrinker(&deferred_split_shrinker); 415 if (err) 416 goto err_split_shrinker; 417 418 /* 419 * By default disable transparent hugepages on smaller systems, 420 * where the extra memory used could hurt more than TLB overhead 421 * is likely to save. The admin can still enable it through /sys. 422 */ 423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { 424 transparent_hugepage_flags = 0; 425 return 0; 426 } 427 428 err = start_stop_khugepaged(); 429 if (err) 430 goto err_khugepaged; 431 432 return 0; 433 err_khugepaged: 434 unregister_shrinker(&deferred_split_shrinker); 435 err_split_shrinker: 436 unregister_shrinker(&huge_zero_page_shrinker); 437 err_hzp_shrinker: 438 khugepaged_destroy(); 439 err_slab: 440 hugepage_exit_sysfs(hugepage_kobj); 441 err_sysfs: 442 return err; 443 } 444 subsys_initcall(hugepage_init); 445 446 static int __init setup_transparent_hugepage(char *str) 447 { 448 int ret = 0; 449 if (!str) 450 goto out; 451 if (!strcmp(str, "always")) { 452 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 453 &transparent_hugepage_flags); 454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 455 &transparent_hugepage_flags); 456 ret = 1; 457 } else if (!strcmp(str, "madvise")) { 458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 459 &transparent_hugepage_flags); 460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 461 &transparent_hugepage_flags); 462 ret = 1; 463 } else if (!strcmp(str, "never")) { 464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 465 &transparent_hugepage_flags); 466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 467 &transparent_hugepage_flags); 468 ret = 1; 469 } 470 out: 471 if (!ret) 472 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 473 return ret; 474 } 475 __setup("transparent_hugepage=", setup_transparent_hugepage); 476 477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 478 { 479 if (likely(vma->vm_flags & VM_WRITE)) 480 pmd = pmd_mkwrite(pmd); 481 return pmd; 482 } 483 484 static inline struct list_head *page_deferred_list(struct page *page) 485 { 486 /* 487 * ->lru in the tail pages is occupied by compound_head. 488 * Let's use ->mapping + ->index in the second tail page as list_head. 489 */ 490 return (struct list_head *)&page[2].mapping; 491 } 492 493 void prep_transhuge_page(struct page *page) 494 { 495 /* 496 * we use page->mapping and page->indexlru in second tail page 497 * as list_head: assuming THP order >= 2 498 */ 499 500 INIT_LIST_HEAD(page_deferred_list(page)); 501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 502 } 503 504 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len, 505 loff_t off, unsigned long flags, unsigned long size) 506 { 507 unsigned long addr; 508 loff_t off_end = off + len; 509 loff_t off_align = round_up(off, size); 510 unsigned long len_pad; 511 512 if (off_end <= off_align || (off_end - off_align) < size) 513 return 0; 514 515 len_pad = len + size; 516 if (len_pad < len || (off + len_pad) < off) 517 return 0; 518 519 addr = current->mm->get_unmapped_area(filp, 0, len_pad, 520 off >> PAGE_SHIFT, flags); 521 if (IS_ERR_VALUE(addr)) 522 return 0; 523 524 addr += (off - addr) & (size - 1); 525 return addr; 526 } 527 528 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 529 unsigned long len, unsigned long pgoff, unsigned long flags) 530 { 531 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 532 533 if (addr) 534 goto out; 535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 536 goto out; 537 538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE); 539 if (addr) 540 return addr; 541 542 out: 543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 544 } 545 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 546 547 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page, 548 gfp_t gfp) 549 { 550 struct vm_area_struct *vma = vmf->vma; 551 struct mem_cgroup *memcg; 552 pgtable_t pgtable; 553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 554 int ret = 0; 555 556 VM_BUG_ON_PAGE(!PageCompound(page), page); 557 558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { 559 put_page(page); 560 count_vm_event(THP_FAULT_FALLBACK); 561 return VM_FAULT_FALLBACK; 562 } 563 564 pgtable = pte_alloc_one(vma->vm_mm, haddr); 565 if (unlikely(!pgtable)) { 566 ret = VM_FAULT_OOM; 567 goto release; 568 } 569 570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 571 /* 572 * The memory barrier inside __SetPageUptodate makes sure that 573 * clear_huge_page writes become visible before the set_pmd_at() 574 * write. 575 */ 576 __SetPageUptodate(page); 577 578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 579 if (unlikely(!pmd_none(*vmf->pmd))) { 580 goto unlock_release; 581 } else { 582 pmd_t entry; 583 584 ret = check_stable_address_space(vma->vm_mm); 585 if (ret) 586 goto unlock_release; 587 588 /* Deliver the page fault to userland */ 589 if (userfaultfd_missing(vma)) { 590 int ret; 591 592 spin_unlock(vmf->ptl); 593 mem_cgroup_cancel_charge(page, memcg, true); 594 put_page(page); 595 pte_free(vma->vm_mm, pgtable); 596 ret = handle_userfault(vmf, VM_UFFD_MISSING); 597 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 598 return ret; 599 } 600 601 entry = mk_huge_pmd(page, vma->vm_page_prot); 602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 603 page_add_new_anon_rmap(page, vma, haddr, true); 604 mem_cgroup_commit_charge(page, memcg, false, true); 605 lru_cache_add_active_or_unevictable(page, vma); 606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 609 mm_inc_nr_ptes(vma->vm_mm); 610 spin_unlock(vmf->ptl); 611 count_vm_event(THP_FAULT_ALLOC); 612 } 613 614 return 0; 615 unlock_release: 616 spin_unlock(vmf->ptl); 617 release: 618 if (pgtable) 619 pte_free(vma->vm_mm, pgtable); 620 mem_cgroup_cancel_charge(page, memcg, true); 621 put_page(page); 622 return ret; 623 624 } 625 626 /* 627 * always: directly stall for all thp allocations 628 * defer: wake kswapd and fail if not immediately available 629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 630 * fail if not immediately available 631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 632 * available 633 * never: never stall for any thp allocation 634 */ 635 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 636 { 637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 638 639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 645 __GFP_KSWAPD_RECLAIM); 646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 648 0); 649 return GFP_TRANSHUGE_LIGHT; 650 } 651 652 /* Caller must hold page table lock. */ 653 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 655 struct page *zero_page) 656 { 657 pmd_t entry; 658 if (!pmd_none(*pmd)) 659 return false; 660 entry = mk_pmd(zero_page, vma->vm_page_prot); 661 entry = pmd_mkhuge(entry); 662 if (pgtable) 663 pgtable_trans_huge_deposit(mm, pmd, pgtable); 664 set_pmd_at(mm, haddr, pmd, entry); 665 mm_inc_nr_ptes(mm); 666 return true; 667 } 668 669 int do_huge_pmd_anonymous_page(struct vm_fault *vmf) 670 { 671 struct vm_area_struct *vma = vmf->vma; 672 gfp_t gfp; 673 struct page *page; 674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 675 676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 677 return VM_FAULT_FALLBACK; 678 if (unlikely(anon_vma_prepare(vma))) 679 return VM_FAULT_OOM; 680 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 681 return VM_FAULT_OOM; 682 if (!(vmf->flags & FAULT_FLAG_WRITE) && 683 !mm_forbids_zeropage(vma->vm_mm) && 684 transparent_hugepage_use_zero_page()) { 685 pgtable_t pgtable; 686 struct page *zero_page; 687 bool set; 688 int ret; 689 pgtable = pte_alloc_one(vma->vm_mm, haddr); 690 if (unlikely(!pgtable)) 691 return VM_FAULT_OOM; 692 zero_page = mm_get_huge_zero_page(vma->vm_mm); 693 if (unlikely(!zero_page)) { 694 pte_free(vma->vm_mm, pgtable); 695 count_vm_event(THP_FAULT_FALLBACK); 696 return VM_FAULT_FALLBACK; 697 } 698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 699 ret = 0; 700 set = false; 701 if (pmd_none(*vmf->pmd)) { 702 ret = check_stable_address_space(vma->vm_mm); 703 if (ret) { 704 spin_unlock(vmf->ptl); 705 } else if (userfaultfd_missing(vma)) { 706 spin_unlock(vmf->ptl); 707 ret = handle_userfault(vmf, VM_UFFD_MISSING); 708 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 709 } else { 710 set_huge_zero_page(pgtable, vma->vm_mm, vma, 711 haddr, vmf->pmd, zero_page); 712 spin_unlock(vmf->ptl); 713 set = true; 714 } 715 } else 716 spin_unlock(vmf->ptl); 717 if (!set) 718 pte_free(vma->vm_mm, pgtable); 719 return ret; 720 } 721 gfp = alloc_hugepage_direct_gfpmask(vma); 722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 723 if (unlikely(!page)) { 724 count_vm_event(THP_FAULT_FALLBACK); 725 return VM_FAULT_FALLBACK; 726 } 727 prep_transhuge_page(page); 728 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 729 } 730 731 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 733 pgtable_t pgtable) 734 { 735 struct mm_struct *mm = vma->vm_mm; 736 pmd_t entry; 737 spinlock_t *ptl; 738 739 ptl = pmd_lock(mm, pmd); 740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 741 if (pfn_t_devmap(pfn)) 742 entry = pmd_mkdevmap(entry); 743 if (write) { 744 entry = pmd_mkyoung(pmd_mkdirty(entry)); 745 entry = maybe_pmd_mkwrite(entry, vma); 746 } 747 748 if (pgtable) { 749 pgtable_trans_huge_deposit(mm, pmd, pgtable); 750 mm_inc_nr_ptes(mm); 751 } 752 753 set_pmd_at(mm, addr, pmd, entry); 754 update_mmu_cache_pmd(vma, addr, pmd); 755 spin_unlock(ptl); 756 } 757 758 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 759 pmd_t *pmd, pfn_t pfn, bool write) 760 { 761 pgprot_t pgprot = vma->vm_page_prot; 762 pgtable_t pgtable = NULL; 763 /* 764 * If we had pmd_special, we could avoid all these restrictions, 765 * but we need to be consistent with PTEs and architectures that 766 * can't support a 'special' bit. 767 */ 768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 770 (VM_PFNMAP|VM_MIXEDMAP)); 771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 772 BUG_ON(!pfn_t_devmap(pfn)); 773 774 if (addr < vma->vm_start || addr >= vma->vm_end) 775 return VM_FAULT_SIGBUS; 776 777 if (arch_needs_pgtable_deposit()) { 778 pgtable = pte_alloc_one(vma->vm_mm, addr); 779 if (!pgtable) 780 return VM_FAULT_OOM; 781 } 782 783 track_pfn_insert(vma, &pgprot, pfn); 784 785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 786 return VM_FAULT_NOPAGE; 787 } 788 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 789 790 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 791 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 792 { 793 if (likely(vma->vm_flags & VM_WRITE)) 794 pud = pud_mkwrite(pud); 795 return pud; 796 } 797 798 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 800 { 801 struct mm_struct *mm = vma->vm_mm; 802 pud_t entry; 803 spinlock_t *ptl; 804 805 ptl = pud_lock(mm, pud); 806 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 807 if (pfn_t_devmap(pfn)) 808 entry = pud_mkdevmap(entry); 809 if (write) { 810 entry = pud_mkyoung(pud_mkdirty(entry)); 811 entry = maybe_pud_mkwrite(entry, vma); 812 } 813 set_pud_at(mm, addr, pud, entry); 814 update_mmu_cache_pud(vma, addr, pud); 815 spin_unlock(ptl); 816 } 817 818 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 819 pud_t *pud, pfn_t pfn, bool write) 820 { 821 pgprot_t pgprot = vma->vm_page_prot; 822 /* 823 * If we had pud_special, we could avoid all these restrictions, 824 * but we need to be consistent with PTEs and architectures that 825 * can't support a 'special' bit. 826 */ 827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 829 (VM_PFNMAP|VM_MIXEDMAP)); 830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 831 BUG_ON(!pfn_t_devmap(pfn)); 832 833 if (addr < vma->vm_start || addr >= vma->vm_end) 834 return VM_FAULT_SIGBUS; 835 836 track_pfn_insert(vma, &pgprot, pfn); 837 838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 839 return VM_FAULT_NOPAGE; 840 } 841 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 842 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 843 844 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 845 pmd_t *pmd, int flags) 846 { 847 pmd_t _pmd; 848 849 _pmd = pmd_mkyoung(*pmd); 850 if (flags & FOLL_WRITE) 851 _pmd = pmd_mkdirty(_pmd); 852 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 853 pmd, _pmd, flags & FOLL_WRITE)) 854 update_mmu_cache_pmd(vma, addr, pmd); 855 } 856 857 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 858 pmd_t *pmd, int flags) 859 { 860 unsigned long pfn = pmd_pfn(*pmd); 861 struct mm_struct *mm = vma->vm_mm; 862 struct dev_pagemap *pgmap; 863 struct page *page; 864 865 assert_spin_locked(pmd_lockptr(mm, pmd)); 866 867 /* 868 * When we COW a devmap PMD entry, we split it into PTEs, so we should 869 * not be in this function with `flags & FOLL_COW` set. 870 */ 871 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 872 873 if (!pmd_access_permitted(*pmd, flags & FOLL_WRITE)) 874 return NULL; 875 876 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 877 /* pass */; 878 else 879 return NULL; 880 881 if (flags & FOLL_TOUCH) 882 touch_pmd(vma, addr, pmd, flags); 883 884 /* 885 * device mapped pages can only be returned if the 886 * caller will manage the page reference count. 887 */ 888 if (!(flags & FOLL_GET)) 889 return ERR_PTR(-EEXIST); 890 891 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 892 pgmap = get_dev_pagemap(pfn, NULL); 893 if (!pgmap) 894 return ERR_PTR(-EFAULT); 895 page = pfn_to_page(pfn); 896 get_page(page); 897 put_dev_pagemap(pgmap); 898 899 return page; 900 } 901 902 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 903 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 904 struct vm_area_struct *vma) 905 { 906 spinlock_t *dst_ptl, *src_ptl; 907 struct page *src_page; 908 pmd_t pmd; 909 pgtable_t pgtable = NULL; 910 int ret = -ENOMEM; 911 912 /* Skip if can be re-fill on fault */ 913 if (!vma_is_anonymous(vma)) 914 return 0; 915 916 pgtable = pte_alloc_one(dst_mm, addr); 917 if (unlikely(!pgtable)) 918 goto out; 919 920 dst_ptl = pmd_lock(dst_mm, dst_pmd); 921 src_ptl = pmd_lockptr(src_mm, src_pmd); 922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 923 924 ret = -EAGAIN; 925 pmd = *src_pmd; 926 927 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 928 if (unlikely(is_swap_pmd(pmd))) { 929 swp_entry_t entry = pmd_to_swp_entry(pmd); 930 931 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 932 if (is_write_migration_entry(entry)) { 933 make_migration_entry_read(&entry); 934 pmd = swp_entry_to_pmd(entry); 935 if (pmd_swp_soft_dirty(*src_pmd)) 936 pmd = pmd_swp_mksoft_dirty(pmd); 937 set_pmd_at(src_mm, addr, src_pmd, pmd); 938 } 939 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 940 mm_inc_nr_ptes(dst_mm); 941 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 942 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 943 ret = 0; 944 goto out_unlock; 945 } 946 #endif 947 948 if (unlikely(!pmd_trans_huge(pmd))) { 949 pte_free(dst_mm, pgtable); 950 goto out_unlock; 951 } 952 /* 953 * When page table lock is held, the huge zero pmd should not be 954 * under splitting since we don't split the page itself, only pmd to 955 * a page table. 956 */ 957 if (is_huge_zero_pmd(pmd)) { 958 struct page *zero_page; 959 /* 960 * get_huge_zero_page() will never allocate a new page here, 961 * since we already have a zero page to copy. It just takes a 962 * reference. 963 */ 964 zero_page = mm_get_huge_zero_page(dst_mm); 965 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 966 zero_page); 967 ret = 0; 968 goto out_unlock; 969 } 970 971 src_page = pmd_page(pmd); 972 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 973 get_page(src_page); 974 page_dup_rmap(src_page, true); 975 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 976 mm_inc_nr_ptes(dst_mm); 977 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 978 979 pmdp_set_wrprotect(src_mm, addr, src_pmd); 980 pmd = pmd_mkold(pmd_wrprotect(pmd)); 981 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 982 983 ret = 0; 984 out_unlock: 985 spin_unlock(src_ptl); 986 spin_unlock(dst_ptl); 987 out: 988 return ret; 989 } 990 991 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 992 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 993 pud_t *pud, int flags) 994 { 995 pud_t _pud; 996 997 _pud = pud_mkyoung(*pud); 998 if (flags & FOLL_WRITE) 999 _pud = pud_mkdirty(_pud); 1000 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1001 pud, _pud, flags & FOLL_WRITE)) 1002 update_mmu_cache_pud(vma, addr, pud); 1003 } 1004 1005 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1006 pud_t *pud, int flags) 1007 { 1008 unsigned long pfn = pud_pfn(*pud); 1009 struct mm_struct *mm = vma->vm_mm; 1010 struct dev_pagemap *pgmap; 1011 struct page *page; 1012 1013 assert_spin_locked(pud_lockptr(mm, pud)); 1014 1015 if (!pud_access_permitted(*pud, flags & FOLL_WRITE)) 1016 return NULL; 1017 1018 if (pud_present(*pud) && pud_devmap(*pud)) 1019 /* pass */; 1020 else 1021 return NULL; 1022 1023 if (flags & FOLL_TOUCH) 1024 touch_pud(vma, addr, pud, flags); 1025 1026 /* 1027 * device mapped pages can only be returned if the 1028 * caller will manage the page reference count. 1029 */ 1030 if (!(flags & FOLL_GET)) 1031 return ERR_PTR(-EEXIST); 1032 1033 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1034 pgmap = get_dev_pagemap(pfn, NULL); 1035 if (!pgmap) 1036 return ERR_PTR(-EFAULT); 1037 page = pfn_to_page(pfn); 1038 get_page(page); 1039 put_dev_pagemap(pgmap); 1040 1041 return page; 1042 } 1043 1044 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1045 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1046 struct vm_area_struct *vma) 1047 { 1048 spinlock_t *dst_ptl, *src_ptl; 1049 pud_t pud; 1050 int ret; 1051 1052 dst_ptl = pud_lock(dst_mm, dst_pud); 1053 src_ptl = pud_lockptr(src_mm, src_pud); 1054 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1055 1056 ret = -EAGAIN; 1057 pud = *src_pud; 1058 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1059 goto out_unlock; 1060 1061 /* 1062 * When page table lock is held, the huge zero pud should not be 1063 * under splitting since we don't split the page itself, only pud to 1064 * a page table. 1065 */ 1066 if (is_huge_zero_pud(pud)) { 1067 /* No huge zero pud yet */ 1068 } 1069 1070 pudp_set_wrprotect(src_mm, addr, src_pud); 1071 pud = pud_mkold(pud_wrprotect(pud)); 1072 set_pud_at(dst_mm, addr, dst_pud, pud); 1073 1074 ret = 0; 1075 out_unlock: 1076 spin_unlock(src_ptl); 1077 spin_unlock(dst_ptl); 1078 return ret; 1079 } 1080 1081 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1082 { 1083 pud_t entry; 1084 unsigned long haddr; 1085 bool write = vmf->flags & FAULT_FLAG_WRITE; 1086 1087 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1088 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1089 goto unlock; 1090 1091 entry = pud_mkyoung(orig_pud); 1092 if (write) 1093 entry = pud_mkdirty(entry); 1094 haddr = vmf->address & HPAGE_PUD_MASK; 1095 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1096 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1097 1098 unlock: 1099 spin_unlock(vmf->ptl); 1100 } 1101 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1102 1103 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1104 { 1105 pmd_t entry; 1106 unsigned long haddr; 1107 bool write = vmf->flags & FAULT_FLAG_WRITE; 1108 1109 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1110 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1111 goto unlock; 1112 1113 entry = pmd_mkyoung(orig_pmd); 1114 if (write) 1115 entry = pmd_mkdirty(entry); 1116 haddr = vmf->address & HPAGE_PMD_MASK; 1117 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1118 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1119 1120 unlock: 1121 spin_unlock(vmf->ptl); 1122 } 1123 1124 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd, 1125 struct page *page) 1126 { 1127 struct vm_area_struct *vma = vmf->vma; 1128 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1129 struct mem_cgroup *memcg; 1130 pgtable_t pgtable; 1131 pmd_t _pmd; 1132 int ret = 0, i; 1133 struct page **pages; 1134 unsigned long mmun_start; /* For mmu_notifiers */ 1135 unsigned long mmun_end; /* For mmu_notifiers */ 1136 1137 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 1138 GFP_KERNEL); 1139 if (unlikely(!pages)) { 1140 ret |= VM_FAULT_OOM; 1141 goto out; 1142 } 1143 1144 for (i = 0; i < HPAGE_PMD_NR; i++) { 1145 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 1146 vmf->address, page_to_nid(page)); 1147 if (unlikely(!pages[i] || 1148 mem_cgroup_try_charge(pages[i], vma->vm_mm, 1149 GFP_KERNEL, &memcg, false))) { 1150 if (pages[i]) 1151 put_page(pages[i]); 1152 while (--i >= 0) { 1153 memcg = (void *)page_private(pages[i]); 1154 set_page_private(pages[i], 0); 1155 mem_cgroup_cancel_charge(pages[i], memcg, 1156 false); 1157 put_page(pages[i]); 1158 } 1159 kfree(pages); 1160 ret |= VM_FAULT_OOM; 1161 goto out; 1162 } 1163 set_page_private(pages[i], (unsigned long)memcg); 1164 } 1165 1166 for (i = 0; i < HPAGE_PMD_NR; i++) { 1167 copy_user_highpage(pages[i], page + i, 1168 haddr + PAGE_SIZE * i, vma); 1169 __SetPageUptodate(pages[i]); 1170 cond_resched(); 1171 } 1172 1173 mmun_start = haddr; 1174 mmun_end = haddr + HPAGE_PMD_SIZE; 1175 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1176 1177 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1178 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1179 goto out_free_pages; 1180 VM_BUG_ON_PAGE(!PageHead(page), page); 1181 1182 /* 1183 * Leave pmd empty until pte is filled note we must notify here as 1184 * concurrent CPU thread might write to new page before the call to 1185 * mmu_notifier_invalidate_range_end() happens which can lead to a 1186 * device seeing memory write in different order than CPU. 1187 * 1188 * See Documentation/vm/mmu_notifier.txt 1189 */ 1190 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1191 1192 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1193 pmd_populate(vma->vm_mm, &_pmd, pgtable); 1194 1195 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1196 pte_t entry; 1197 entry = mk_pte(pages[i], vma->vm_page_prot); 1198 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1199 memcg = (void *)page_private(pages[i]); 1200 set_page_private(pages[i], 0); 1201 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1202 mem_cgroup_commit_charge(pages[i], memcg, false, false); 1203 lru_cache_add_active_or_unevictable(pages[i], vma); 1204 vmf->pte = pte_offset_map(&_pmd, haddr); 1205 VM_BUG_ON(!pte_none(*vmf->pte)); 1206 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 1207 pte_unmap(vmf->pte); 1208 } 1209 kfree(pages); 1210 1211 smp_wmb(); /* make pte visible before pmd */ 1212 pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1213 page_remove_rmap(page, true); 1214 spin_unlock(vmf->ptl); 1215 1216 /* 1217 * No need to double call mmu_notifier->invalidate_range() callback as 1218 * the above pmdp_huge_clear_flush_notify() did already call it. 1219 */ 1220 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1221 mmun_end); 1222 1223 ret |= VM_FAULT_WRITE; 1224 put_page(page); 1225 1226 out: 1227 return ret; 1228 1229 out_free_pages: 1230 spin_unlock(vmf->ptl); 1231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1232 for (i = 0; i < HPAGE_PMD_NR; i++) { 1233 memcg = (void *)page_private(pages[i]); 1234 set_page_private(pages[i], 0); 1235 mem_cgroup_cancel_charge(pages[i], memcg, false); 1236 put_page(pages[i]); 1237 } 1238 kfree(pages); 1239 goto out; 1240 } 1241 1242 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1243 { 1244 struct vm_area_struct *vma = vmf->vma; 1245 struct page *page = NULL, *new_page; 1246 struct mem_cgroup *memcg; 1247 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1248 unsigned long mmun_start; /* For mmu_notifiers */ 1249 unsigned long mmun_end; /* For mmu_notifiers */ 1250 gfp_t huge_gfp; /* for allocation and charge */ 1251 int ret = 0; 1252 1253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1254 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1255 if (is_huge_zero_pmd(orig_pmd)) 1256 goto alloc; 1257 spin_lock(vmf->ptl); 1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1259 goto out_unlock; 1260 1261 page = pmd_page(orig_pmd); 1262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1263 /* 1264 * We can only reuse the page if nobody else maps the huge page or it's 1265 * part. 1266 */ 1267 if (!trylock_page(page)) { 1268 get_page(page); 1269 spin_unlock(vmf->ptl); 1270 lock_page(page); 1271 spin_lock(vmf->ptl); 1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1273 unlock_page(page); 1274 put_page(page); 1275 goto out_unlock; 1276 } 1277 put_page(page); 1278 } 1279 if (reuse_swap_page(page, NULL)) { 1280 pmd_t entry; 1281 entry = pmd_mkyoung(orig_pmd); 1282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1285 ret |= VM_FAULT_WRITE; 1286 unlock_page(page); 1287 goto out_unlock; 1288 } 1289 unlock_page(page); 1290 get_page(page); 1291 spin_unlock(vmf->ptl); 1292 alloc: 1293 if (transparent_hugepage_enabled(vma) && 1294 !transparent_hugepage_debug_cow()) { 1295 huge_gfp = alloc_hugepage_direct_gfpmask(vma); 1296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1297 } else 1298 new_page = NULL; 1299 1300 if (likely(new_page)) { 1301 prep_transhuge_page(new_page); 1302 } else { 1303 if (!page) { 1304 split_huge_pmd(vma, vmf->pmd, vmf->address); 1305 ret |= VM_FAULT_FALLBACK; 1306 } else { 1307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 1308 if (ret & VM_FAULT_OOM) { 1309 split_huge_pmd(vma, vmf->pmd, vmf->address); 1310 ret |= VM_FAULT_FALLBACK; 1311 } 1312 put_page(page); 1313 } 1314 count_vm_event(THP_FAULT_FALLBACK); 1315 goto out; 1316 } 1317 1318 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, 1319 huge_gfp, &memcg, true))) { 1320 put_page(new_page); 1321 split_huge_pmd(vma, vmf->pmd, vmf->address); 1322 if (page) 1323 put_page(page); 1324 ret |= VM_FAULT_FALLBACK; 1325 count_vm_event(THP_FAULT_FALLBACK); 1326 goto out; 1327 } 1328 1329 count_vm_event(THP_FAULT_ALLOC); 1330 1331 if (!page) 1332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 1333 else 1334 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 1335 __SetPageUptodate(new_page); 1336 1337 mmun_start = haddr; 1338 mmun_end = haddr + HPAGE_PMD_SIZE; 1339 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1340 1341 spin_lock(vmf->ptl); 1342 if (page) 1343 put_page(page); 1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1345 spin_unlock(vmf->ptl); 1346 mem_cgroup_cancel_charge(new_page, memcg, true); 1347 put_page(new_page); 1348 goto out_mn; 1349 } else { 1350 pmd_t entry; 1351 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1353 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1354 page_add_new_anon_rmap(new_page, vma, haddr, true); 1355 mem_cgroup_commit_charge(new_page, memcg, false, true); 1356 lru_cache_add_active_or_unevictable(new_page, vma); 1357 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 1358 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1359 if (!page) { 1360 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1361 } else { 1362 VM_BUG_ON_PAGE(!PageHead(page), page); 1363 page_remove_rmap(page, true); 1364 put_page(page); 1365 } 1366 ret |= VM_FAULT_WRITE; 1367 } 1368 spin_unlock(vmf->ptl); 1369 out_mn: 1370 /* 1371 * No need to double call mmu_notifier->invalidate_range() callback as 1372 * the above pmdp_huge_clear_flush_notify() did already call it. 1373 */ 1374 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1375 mmun_end); 1376 out: 1377 return ret; 1378 out_unlock: 1379 spin_unlock(vmf->ptl); 1380 return ret; 1381 } 1382 1383 /* 1384 * FOLL_FORCE can write to even unwritable pmd's, but only 1385 * after we've gone through a COW cycle and they are dirty. 1386 */ 1387 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1388 { 1389 return pmd_access_permitted(pmd, WRITE) || 1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 1391 } 1392 1393 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1394 unsigned long addr, 1395 pmd_t *pmd, 1396 unsigned int flags) 1397 { 1398 struct mm_struct *mm = vma->vm_mm; 1399 struct page *page = NULL; 1400 1401 assert_spin_locked(pmd_lockptr(mm, pmd)); 1402 1403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1404 goto out; 1405 1406 /* Avoid dumping huge zero page */ 1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1408 return ERR_PTR(-EFAULT); 1409 1410 /* Full NUMA hinting faults to serialise migration in fault paths */ 1411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1412 goto out; 1413 1414 page = pmd_page(*pmd); 1415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1416 if (flags & FOLL_TOUCH) 1417 touch_pmd(vma, addr, pmd, flags); 1418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1419 /* 1420 * We don't mlock() pte-mapped THPs. This way we can avoid 1421 * leaking mlocked pages into non-VM_LOCKED VMAs. 1422 * 1423 * For anon THP: 1424 * 1425 * In most cases the pmd is the only mapping of the page as we 1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1427 * writable private mappings in populate_vma_page_range(). 1428 * 1429 * The only scenario when we have the page shared here is if we 1430 * mlocking read-only mapping shared over fork(). We skip 1431 * mlocking such pages. 1432 * 1433 * For file THP: 1434 * 1435 * We can expect PageDoubleMap() to be stable under page lock: 1436 * for file pages we set it in page_add_file_rmap(), which 1437 * requires page to be locked. 1438 */ 1439 1440 if (PageAnon(page) && compound_mapcount(page) != 1) 1441 goto skip_mlock; 1442 if (PageDoubleMap(page) || !page->mapping) 1443 goto skip_mlock; 1444 if (!trylock_page(page)) 1445 goto skip_mlock; 1446 lru_add_drain(); 1447 if (page->mapping && !PageDoubleMap(page)) 1448 mlock_vma_page(page); 1449 unlock_page(page); 1450 } 1451 skip_mlock: 1452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1454 if (flags & FOLL_GET) 1455 get_page(page); 1456 1457 out: 1458 return page; 1459 } 1460 1461 /* NUMA hinting page fault entry point for trans huge pmds */ 1462 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1463 { 1464 struct vm_area_struct *vma = vmf->vma; 1465 struct anon_vma *anon_vma = NULL; 1466 struct page *page; 1467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1468 int page_nid = -1, this_nid = numa_node_id(); 1469 int target_nid, last_cpupid = -1; 1470 bool page_locked; 1471 bool migrated = false; 1472 bool was_writable; 1473 int flags = 0; 1474 1475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1476 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1477 goto out_unlock; 1478 1479 /* 1480 * If there are potential migrations, wait for completion and retry 1481 * without disrupting NUMA hinting information. Do not relock and 1482 * check_same as the page may no longer be mapped. 1483 */ 1484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1485 page = pmd_page(*vmf->pmd); 1486 if (!get_page_unless_zero(page)) 1487 goto out_unlock; 1488 spin_unlock(vmf->ptl); 1489 wait_on_page_locked(page); 1490 put_page(page); 1491 goto out; 1492 } 1493 1494 page = pmd_page(pmd); 1495 BUG_ON(is_huge_zero_page(page)); 1496 page_nid = page_to_nid(page); 1497 last_cpupid = page_cpupid_last(page); 1498 count_vm_numa_event(NUMA_HINT_FAULTS); 1499 if (page_nid == this_nid) { 1500 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1501 flags |= TNF_FAULT_LOCAL; 1502 } 1503 1504 /* See similar comment in do_numa_page for explanation */ 1505 if (!pmd_savedwrite(pmd)) 1506 flags |= TNF_NO_GROUP; 1507 1508 /* 1509 * Acquire the page lock to serialise THP migrations but avoid dropping 1510 * page_table_lock if at all possible 1511 */ 1512 page_locked = trylock_page(page); 1513 target_nid = mpol_misplaced(page, vma, haddr); 1514 if (target_nid == -1) { 1515 /* If the page was locked, there are no parallel migrations */ 1516 if (page_locked) 1517 goto clear_pmdnuma; 1518 } 1519 1520 /* Migration could have started since the pmd_trans_migrating check */ 1521 if (!page_locked) { 1522 page_nid = -1; 1523 if (!get_page_unless_zero(page)) 1524 goto out_unlock; 1525 spin_unlock(vmf->ptl); 1526 wait_on_page_locked(page); 1527 put_page(page); 1528 goto out; 1529 } 1530 1531 /* 1532 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1533 * to serialises splits 1534 */ 1535 get_page(page); 1536 spin_unlock(vmf->ptl); 1537 anon_vma = page_lock_anon_vma_read(page); 1538 1539 /* Confirm the PMD did not change while page_table_lock was released */ 1540 spin_lock(vmf->ptl); 1541 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1542 unlock_page(page); 1543 put_page(page); 1544 page_nid = -1; 1545 goto out_unlock; 1546 } 1547 1548 /* Bail if we fail to protect against THP splits for any reason */ 1549 if (unlikely(!anon_vma)) { 1550 put_page(page); 1551 page_nid = -1; 1552 goto clear_pmdnuma; 1553 } 1554 1555 /* 1556 * Since we took the NUMA fault, we must have observed the !accessible 1557 * bit. Make sure all other CPUs agree with that, to avoid them 1558 * modifying the page we're about to migrate. 1559 * 1560 * Must be done under PTL such that we'll observe the relevant 1561 * inc_tlb_flush_pending(). 1562 * 1563 * We are not sure a pending tlb flush here is for a huge page 1564 * mapping or not. Hence use the tlb range variant 1565 */ 1566 if (mm_tlb_flush_pending(vma->vm_mm)) 1567 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1568 1569 /* 1570 * Migrate the THP to the requested node, returns with page unlocked 1571 * and access rights restored. 1572 */ 1573 spin_unlock(vmf->ptl); 1574 1575 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1576 vmf->pmd, pmd, vmf->address, page, target_nid); 1577 if (migrated) { 1578 flags |= TNF_MIGRATED; 1579 page_nid = target_nid; 1580 } else 1581 flags |= TNF_MIGRATE_FAIL; 1582 1583 goto out; 1584 clear_pmdnuma: 1585 BUG_ON(!PageLocked(page)); 1586 was_writable = pmd_savedwrite(pmd); 1587 pmd = pmd_modify(pmd, vma->vm_page_prot); 1588 pmd = pmd_mkyoung(pmd); 1589 if (was_writable) 1590 pmd = pmd_mkwrite(pmd); 1591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1593 unlock_page(page); 1594 out_unlock: 1595 spin_unlock(vmf->ptl); 1596 1597 out: 1598 if (anon_vma) 1599 page_unlock_anon_vma_read(anon_vma); 1600 1601 if (page_nid != -1) 1602 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1603 flags); 1604 1605 return 0; 1606 } 1607 1608 /* 1609 * Return true if we do MADV_FREE successfully on entire pmd page. 1610 * Otherwise, return false. 1611 */ 1612 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1613 pmd_t *pmd, unsigned long addr, unsigned long next) 1614 { 1615 spinlock_t *ptl; 1616 pmd_t orig_pmd; 1617 struct page *page; 1618 struct mm_struct *mm = tlb->mm; 1619 bool ret = false; 1620 1621 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1622 1623 ptl = pmd_trans_huge_lock(pmd, vma); 1624 if (!ptl) 1625 goto out_unlocked; 1626 1627 orig_pmd = *pmd; 1628 if (is_huge_zero_pmd(orig_pmd)) 1629 goto out; 1630 1631 if (unlikely(!pmd_present(orig_pmd))) { 1632 VM_BUG_ON(thp_migration_supported() && 1633 !is_pmd_migration_entry(orig_pmd)); 1634 goto out; 1635 } 1636 1637 page = pmd_page(orig_pmd); 1638 /* 1639 * If other processes are mapping this page, we couldn't discard 1640 * the page unless they all do MADV_FREE so let's skip the page. 1641 */ 1642 if (page_mapcount(page) != 1) 1643 goto out; 1644 1645 if (!trylock_page(page)) 1646 goto out; 1647 1648 /* 1649 * If user want to discard part-pages of THP, split it so MADV_FREE 1650 * will deactivate only them. 1651 */ 1652 if (next - addr != HPAGE_PMD_SIZE) { 1653 get_page(page); 1654 spin_unlock(ptl); 1655 split_huge_page(page); 1656 unlock_page(page); 1657 put_page(page); 1658 goto out_unlocked; 1659 } 1660 1661 if (PageDirty(page)) 1662 ClearPageDirty(page); 1663 unlock_page(page); 1664 1665 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1666 pmdp_invalidate(vma, addr, pmd); 1667 orig_pmd = pmd_mkold(orig_pmd); 1668 orig_pmd = pmd_mkclean(orig_pmd); 1669 1670 set_pmd_at(mm, addr, pmd, orig_pmd); 1671 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1672 } 1673 1674 mark_page_lazyfree(page); 1675 ret = true; 1676 out: 1677 spin_unlock(ptl); 1678 out_unlocked: 1679 return ret; 1680 } 1681 1682 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1683 { 1684 pgtable_t pgtable; 1685 1686 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1687 pte_free(mm, pgtable); 1688 mm_dec_nr_ptes(mm); 1689 } 1690 1691 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1692 pmd_t *pmd, unsigned long addr) 1693 { 1694 pmd_t orig_pmd; 1695 spinlock_t *ptl; 1696 1697 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1698 1699 ptl = __pmd_trans_huge_lock(pmd, vma); 1700 if (!ptl) 1701 return 0; 1702 /* 1703 * For architectures like ppc64 we look at deposited pgtable 1704 * when calling pmdp_huge_get_and_clear. So do the 1705 * pgtable_trans_huge_withdraw after finishing pmdp related 1706 * operations. 1707 */ 1708 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1709 tlb->fullmm); 1710 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1711 if (vma_is_dax(vma)) { 1712 if (arch_needs_pgtable_deposit()) 1713 zap_deposited_table(tlb->mm, pmd); 1714 spin_unlock(ptl); 1715 if (is_huge_zero_pmd(orig_pmd)) 1716 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1717 } else if (is_huge_zero_pmd(orig_pmd)) { 1718 zap_deposited_table(tlb->mm, pmd); 1719 spin_unlock(ptl); 1720 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1721 } else { 1722 struct page *page = NULL; 1723 int flush_needed = 1; 1724 1725 if (pmd_present(orig_pmd)) { 1726 page = pmd_page(orig_pmd); 1727 page_remove_rmap(page, true); 1728 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1729 VM_BUG_ON_PAGE(!PageHead(page), page); 1730 } else if (thp_migration_supported()) { 1731 swp_entry_t entry; 1732 1733 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1734 entry = pmd_to_swp_entry(orig_pmd); 1735 page = pfn_to_page(swp_offset(entry)); 1736 flush_needed = 0; 1737 } else 1738 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1739 1740 if (PageAnon(page)) { 1741 zap_deposited_table(tlb->mm, pmd); 1742 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1743 } else { 1744 if (arch_needs_pgtable_deposit()) 1745 zap_deposited_table(tlb->mm, pmd); 1746 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); 1747 } 1748 1749 spin_unlock(ptl); 1750 if (flush_needed) 1751 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1752 } 1753 return 1; 1754 } 1755 1756 #ifndef pmd_move_must_withdraw 1757 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1758 spinlock_t *old_pmd_ptl, 1759 struct vm_area_struct *vma) 1760 { 1761 /* 1762 * With split pmd lock we also need to move preallocated 1763 * PTE page table if new_pmd is on different PMD page table. 1764 * 1765 * We also don't deposit and withdraw tables for file pages. 1766 */ 1767 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1768 } 1769 #endif 1770 1771 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1772 { 1773 #ifdef CONFIG_MEM_SOFT_DIRTY 1774 if (unlikely(is_pmd_migration_entry(pmd))) 1775 pmd = pmd_swp_mksoft_dirty(pmd); 1776 else if (pmd_present(pmd)) 1777 pmd = pmd_mksoft_dirty(pmd); 1778 #endif 1779 return pmd; 1780 } 1781 1782 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1783 unsigned long new_addr, unsigned long old_end, 1784 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush) 1785 { 1786 spinlock_t *old_ptl, *new_ptl; 1787 pmd_t pmd; 1788 struct mm_struct *mm = vma->vm_mm; 1789 bool force_flush = false; 1790 1791 if ((old_addr & ~HPAGE_PMD_MASK) || 1792 (new_addr & ~HPAGE_PMD_MASK) || 1793 old_end - old_addr < HPAGE_PMD_SIZE) 1794 return false; 1795 1796 /* 1797 * The destination pmd shouldn't be established, free_pgtables() 1798 * should have release it. 1799 */ 1800 if (WARN_ON(!pmd_none(*new_pmd))) { 1801 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1802 return false; 1803 } 1804 1805 /* 1806 * We don't have to worry about the ordering of src and dst 1807 * ptlocks because exclusive mmap_sem prevents deadlock. 1808 */ 1809 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1810 if (old_ptl) { 1811 new_ptl = pmd_lockptr(mm, new_pmd); 1812 if (new_ptl != old_ptl) 1813 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1814 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1815 if (pmd_present(pmd) && pmd_dirty(pmd)) 1816 force_flush = true; 1817 VM_BUG_ON(!pmd_none(*new_pmd)); 1818 1819 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1820 pgtable_t pgtable; 1821 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1822 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1823 } 1824 pmd = move_soft_dirty_pmd(pmd); 1825 set_pmd_at(mm, new_addr, new_pmd, pmd); 1826 if (new_ptl != old_ptl) 1827 spin_unlock(new_ptl); 1828 if (force_flush) 1829 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1830 else 1831 *need_flush = true; 1832 spin_unlock(old_ptl); 1833 return true; 1834 } 1835 return false; 1836 } 1837 1838 /* 1839 * Returns 1840 * - 0 if PMD could not be locked 1841 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1842 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1843 */ 1844 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1845 unsigned long addr, pgprot_t newprot, int prot_numa) 1846 { 1847 struct mm_struct *mm = vma->vm_mm; 1848 spinlock_t *ptl; 1849 pmd_t entry; 1850 bool preserve_write; 1851 int ret; 1852 1853 ptl = __pmd_trans_huge_lock(pmd, vma); 1854 if (!ptl) 1855 return 0; 1856 1857 preserve_write = prot_numa && pmd_write(*pmd); 1858 ret = 1; 1859 1860 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1861 if (is_swap_pmd(*pmd)) { 1862 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1863 1864 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1865 if (is_write_migration_entry(entry)) { 1866 pmd_t newpmd; 1867 /* 1868 * A protection check is difficult so 1869 * just be safe and disable write 1870 */ 1871 make_migration_entry_read(&entry); 1872 newpmd = swp_entry_to_pmd(entry); 1873 if (pmd_swp_soft_dirty(*pmd)) 1874 newpmd = pmd_swp_mksoft_dirty(newpmd); 1875 set_pmd_at(mm, addr, pmd, newpmd); 1876 } 1877 goto unlock; 1878 } 1879 #endif 1880 1881 /* 1882 * Avoid trapping faults against the zero page. The read-only 1883 * data is likely to be read-cached on the local CPU and 1884 * local/remote hits to the zero page are not interesting. 1885 */ 1886 if (prot_numa && is_huge_zero_pmd(*pmd)) 1887 goto unlock; 1888 1889 if (prot_numa && pmd_protnone(*pmd)) 1890 goto unlock; 1891 1892 /* 1893 * In case prot_numa, we are under down_read(mmap_sem). It's critical 1894 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1895 * which is also under down_read(mmap_sem): 1896 * 1897 * CPU0: CPU1: 1898 * change_huge_pmd(prot_numa=1) 1899 * pmdp_huge_get_and_clear_notify() 1900 * madvise_dontneed() 1901 * zap_pmd_range() 1902 * pmd_trans_huge(*pmd) == 0 (without ptl) 1903 * // skip the pmd 1904 * set_pmd_at(); 1905 * // pmd is re-established 1906 * 1907 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1908 * which may break userspace. 1909 * 1910 * pmdp_invalidate() is required to make sure we don't miss 1911 * dirty/young flags set by hardware. 1912 */ 1913 entry = *pmd; 1914 pmdp_invalidate(vma, addr, pmd); 1915 1916 /* 1917 * Recover dirty/young flags. It relies on pmdp_invalidate to not 1918 * corrupt them. 1919 */ 1920 if (pmd_dirty(*pmd)) 1921 entry = pmd_mkdirty(entry); 1922 if (pmd_young(*pmd)) 1923 entry = pmd_mkyoung(entry); 1924 1925 entry = pmd_modify(entry, newprot); 1926 if (preserve_write) 1927 entry = pmd_mk_savedwrite(entry); 1928 ret = HPAGE_PMD_NR; 1929 set_pmd_at(mm, addr, pmd, entry); 1930 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1931 unlock: 1932 spin_unlock(ptl); 1933 return ret; 1934 } 1935 1936 /* 1937 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1938 * 1939 * Note that if it returns page table lock pointer, this routine returns without 1940 * unlocking page table lock. So callers must unlock it. 1941 */ 1942 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1943 { 1944 spinlock_t *ptl; 1945 ptl = pmd_lock(vma->vm_mm, pmd); 1946 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1947 pmd_devmap(*pmd))) 1948 return ptl; 1949 spin_unlock(ptl); 1950 return NULL; 1951 } 1952 1953 /* 1954 * Returns true if a given pud maps a thp, false otherwise. 1955 * 1956 * Note that if it returns true, this routine returns without unlocking page 1957 * table lock. So callers must unlock it. 1958 */ 1959 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1960 { 1961 spinlock_t *ptl; 1962 1963 ptl = pud_lock(vma->vm_mm, pud); 1964 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1965 return ptl; 1966 spin_unlock(ptl); 1967 return NULL; 1968 } 1969 1970 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1971 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1972 pud_t *pud, unsigned long addr) 1973 { 1974 pud_t orig_pud; 1975 spinlock_t *ptl; 1976 1977 ptl = __pud_trans_huge_lock(pud, vma); 1978 if (!ptl) 1979 return 0; 1980 /* 1981 * For architectures like ppc64 we look at deposited pgtable 1982 * when calling pudp_huge_get_and_clear. So do the 1983 * pgtable_trans_huge_withdraw after finishing pudp related 1984 * operations. 1985 */ 1986 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 1987 tlb->fullmm); 1988 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1989 if (vma_is_dax(vma)) { 1990 spin_unlock(ptl); 1991 /* No zero page support yet */ 1992 } else { 1993 /* No support for anonymous PUD pages yet */ 1994 BUG(); 1995 } 1996 return 1; 1997 } 1998 1999 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2000 unsigned long haddr) 2001 { 2002 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2003 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2004 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2005 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2006 2007 count_vm_event(THP_SPLIT_PUD); 2008 2009 pudp_huge_clear_flush_notify(vma, haddr, pud); 2010 } 2011 2012 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2013 unsigned long address) 2014 { 2015 spinlock_t *ptl; 2016 struct mm_struct *mm = vma->vm_mm; 2017 unsigned long haddr = address & HPAGE_PUD_MASK; 2018 2019 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); 2020 ptl = pud_lock(mm, pud); 2021 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2022 goto out; 2023 __split_huge_pud_locked(vma, pud, haddr); 2024 2025 out: 2026 spin_unlock(ptl); 2027 /* 2028 * No need to double call mmu_notifier->invalidate_range() callback as 2029 * the above pudp_huge_clear_flush_notify() did already call it. 2030 */ 2031 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2032 HPAGE_PUD_SIZE); 2033 } 2034 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2035 2036 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2037 unsigned long haddr, pmd_t *pmd) 2038 { 2039 struct mm_struct *mm = vma->vm_mm; 2040 pgtable_t pgtable; 2041 pmd_t _pmd; 2042 int i; 2043 2044 /* 2045 * Leave pmd empty until pte is filled note that it is fine to delay 2046 * notification until mmu_notifier_invalidate_range_end() as we are 2047 * replacing a zero pmd write protected page with a zero pte write 2048 * protected page. 2049 * 2050 * See Documentation/vm/mmu_notifier.txt 2051 */ 2052 pmdp_huge_clear_flush(vma, haddr, pmd); 2053 2054 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2055 pmd_populate(mm, &_pmd, pgtable); 2056 2057 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2058 pte_t *pte, entry; 2059 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2060 entry = pte_mkspecial(entry); 2061 pte = pte_offset_map(&_pmd, haddr); 2062 VM_BUG_ON(!pte_none(*pte)); 2063 set_pte_at(mm, haddr, pte, entry); 2064 pte_unmap(pte); 2065 } 2066 smp_wmb(); /* make pte visible before pmd */ 2067 pmd_populate(mm, pmd, pgtable); 2068 } 2069 2070 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2071 unsigned long haddr, bool freeze) 2072 { 2073 struct mm_struct *mm = vma->vm_mm; 2074 struct page *page; 2075 pgtable_t pgtable; 2076 pmd_t _pmd; 2077 bool young, write, dirty, soft_dirty, pmd_migration = false; 2078 unsigned long addr; 2079 int i; 2080 2081 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2082 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2083 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2084 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2085 && !pmd_devmap(*pmd)); 2086 2087 count_vm_event(THP_SPLIT_PMD); 2088 2089 if (!vma_is_anonymous(vma)) { 2090 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2091 /* 2092 * We are going to unmap this huge page. So 2093 * just go ahead and zap it 2094 */ 2095 if (arch_needs_pgtable_deposit()) 2096 zap_deposited_table(mm, pmd); 2097 if (vma_is_dax(vma)) 2098 return; 2099 page = pmd_page(_pmd); 2100 if (!PageReferenced(page) && pmd_young(_pmd)) 2101 SetPageReferenced(page); 2102 page_remove_rmap(page, true); 2103 put_page(page); 2104 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); 2105 return; 2106 } else if (is_huge_zero_pmd(*pmd)) { 2107 /* 2108 * FIXME: Do we want to invalidate secondary mmu by calling 2109 * mmu_notifier_invalidate_range() see comments below inside 2110 * __split_huge_pmd() ? 2111 * 2112 * We are going from a zero huge page write protected to zero 2113 * small page also write protected so it does not seems useful 2114 * to invalidate secondary mmu at this time. 2115 */ 2116 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2117 } 2118 2119 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2120 pmd_migration = is_pmd_migration_entry(*pmd); 2121 if (pmd_migration) { 2122 swp_entry_t entry; 2123 2124 entry = pmd_to_swp_entry(*pmd); 2125 page = pfn_to_page(swp_offset(entry)); 2126 } else 2127 #endif 2128 page = pmd_page(*pmd); 2129 VM_BUG_ON_PAGE(!page_count(page), page); 2130 page_ref_add(page, HPAGE_PMD_NR - 1); 2131 write = pmd_write(*pmd); 2132 young = pmd_young(*pmd); 2133 dirty = pmd_dirty(*pmd); 2134 soft_dirty = pmd_soft_dirty(*pmd); 2135 2136 pmdp_huge_split_prepare(vma, haddr, pmd); 2137 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2138 pmd_populate(mm, &_pmd, pgtable); 2139 2140 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2141 pte_t entry, *pte; 2142 /* 2143 * Note that NUMA hinting access restrictions are not 2144 * transferred to avoid any possibility of altering 2145 * permissions across VMAs. 2146 */ 2147 if (freeze || pmd_migration) { 2148 swp_entry_t swp_entry; 2149 swp_entry = make_migration_entry(page + i, write); 2150 entry = swp_entry_to_pte(swp_entry); 2151 if (soft_dirty) 2152 entry = pte_swp_mksoft_dirty(entry); 2153 } else { 2154 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2155 entry = maybe_mkwrite(entry, vma); 2156 if (!write) 2157 entry = pte_wrprotect(entry); 2158 if (!young) 2159 entry = pte_mkold(entry); 2160 if (soft_dirty) 2161 entry = pte_mksoft_dirty(entry); 2162 } 2163 if (dirty) 2164 SetPageDirty(page + i); 2165 pte = pte_offset_map(&_pmd, addr); 2166 BUG_ON(!pte_none(*pte)); 2167 set_pte_at(mm, addr, pte, entry); 2168 atomic_inc(&page[i]._mapcount); 2169 pte_unmap(pte); 2170 } 2171 2172 /* 2173 * Set PG_double_map before dropping compound_mapcount to avoid 2174 * false-negative page_mapped(). 2175 */ 2176 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2177 for (i = 0; i < HPAGE_PMD_NR; i++) 2178 atomic_inc(&page[i]._mapcount); 2179 } 2180 2181 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2182 /* Last compound_mapcount is gone. */ 2183 __dec_node_page_state(page, NR_ANON_THPS); 2184 if (TestClearPageDoubleMap(page)) { 2185 /* No need in mapcount reference anymore */ 2186 for (i = 0; i < HPAGE_PMD_NR; i++) 2187 atomic_dec(&page[i]._mapcount); 2188 } 2189 } 2190 2191 smp_wmb(); /* make pte visible before pmd */ 2192 /* 2193 * Up to this point the pmd is present and huge and userland has the 2194 * whole access to the hugepage during the split (which happens in 2195 * place). If we overwrite the pmd with the not-huge version pointing 2196 * to the pte here (which of course we could if all CPUs were bug 2197 * free), userland could trigger a small page size TLB miss on the 2198 * small sized TLB while the hugepage TLB entry is still established in 2199 * the huge TLB. Some CPU doesn't like that. 2200 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2201 * 383 on page 93. Intel should be safe but is also warns that it's 2202 * only safe if the permission and cache attributes of the two entries 2203 * loaded in the two TLB is identical (which should be the case here). 2204 * But it is generally safer to never allow small and huge TLB entries 2205 * for the same virtual address to be loaded simultaneously. So instead 2206 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2207 * current pmd notpresent (atomically because here the pmd_trans_huge 2208 * and pmd_trans_splitting must remain set at all times on the pmd 2209 * until the split is complete for this pmd), then we flush the SMP TLB 2210 * and finally we write the non-huge version of the pmd entry with 2211 * pmd_populate. 2212 */ 2213 pmdp_invalidate(vma, haddr, pmd); 2214 pmd_populate(mm, pmd, pgtable); 2215 2216 if (freeze) { 2217 for (i = 0; i < HPAGE_PMD_NR; i++) { 2218 page_remove_rmap(page + i, false); 2219 put_page(page + i); 2220 } 2221 } 2222 } 2223 2224 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2225 unsigned long address, bool freeze, struct page *page) 2226 { 2227 spinlock_t *ptl; 2228 struct mm_struct *mm = vma->vm_mm; 2229 unsigned long haddr = address & HPAGE_PMD_MASK; 2230 2231 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 2232 ptl = pmd_lock(mm, pmd); 2233 2234 /* 2235 * If caller asks to setup a migration entries, we need a page to check 2236 * pmd against. Otherwise we can end up replacing wrong page. 2237 */ 2238 VM_BUG_ON(freeze && !page); 2239 if (page && page != pmd_page(*pmd)) 2240 goto out; 2241 2242 if (pmd_trans_huge(*pmd)) { 2243 page = pmd_page(*pmd); 2244 if (PageMlocked(page)) 2245 clear_page_mlock(page); 2246 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2247 goto out; 2248 __split_huge_pmd_locked(vma, pmd, haddr, freeze); 2249 out: 2250 spin_unlock(ptl); 2251 /* 2252 * No need to double call mmu_notifier->invalidate_range() callback. 2253 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2254 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2255 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2256 * fault will trigger a flush_notify before pointing to a new page 2257 * (it is fine if the secondary mmu keeps pointing to the old zero 2258 * page in the meantime) 2259 * 3) Split a huge pmd into pte pointing to the same page. No need 2260 * to invalidate secondary tlb entry they are all still valid. 2261 * any further changes to individual pte will notify. So no need 2262 * to call mmu_notifier->invalidate_range() 2263 */ 2264 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2265 HPAGE_PMD_SIZE); 2266 } 2267 2268 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2269 bool freeze, struct page *page) 2270 { 2271 pgd_t *pgd; 2272 p4d_t *p4d; 2273 pud_t *pud; 2274 pmd_t *pmd; 2275 2276 pgd = pgd_offset(vma->vm_mm, address); 2277 if (!pgd_present(*pgd)) 2278 return; 2279 2280 p4d = p4d_offset(pgd, address); 2281 if (!p4d_present(*p4d)) 2282 return; 2283 2284 pud = pud_offset(p4d, address); 2285 if (!pud_present(*pud)) 2286 return; 2287 2288 pmd = pmd_offset(pud, address); 2289 2290 __split_huge_pmd(vma, pmd, address, freeze, page); 2291 } 2292 2293 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2294 unsigned long start, 2295 unsigned long end, 2296 long adjust_next) 2297 { 2298 /* 2299 * If the new start address isn't hpage aligned and it could 2300 * previously contain an hugepage: check if we need to split 2301 * an huge pmd. 2302 */ 2303 if (start & ~HPAGE_PMD_MASK && 2304 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2305 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2306 split_huge_pmd_address(vma, start, false, NULL); 2307 2308 /* 2309 * If the new end address isn't hpage aligned and it could 2310 * previously contain an hugepage: check if we need to split 2311 * an huge pmd. 2312 */ 2313 if (end & ~HPAGE_PMD_MASK && 2314 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2315 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2316 split_huge_pmd_address(vma, end, false, NULL); 2317 2318 /* 2319 * If we're also updating the vma->vm_next->vm_start, if the new 2320 * vm_next->vm_start isn't page aligned and it could previously 2321 * contain an hugepage: check if we need to split an huge pmd. 2322 */ 2323 if (adjust_next > 0) { 2324 struct vm_area_struct *next = vma->vm_next; 2325 unsigned long nstart = next->vm_start; 2326 nstart += adjust_next << PAGE_SHIFT; 2327 if (nstart & ~HPAGE_PMD_MASK && 2328 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2329 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2330 split_huge_pmd_address(next, nstart, false, NULL); 2331 } 2332 } 2333 2334 static void freeze_page(struct page *page) 2335 { 2336 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2337 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2338 bool unmap_success; 2339 2340 VM_BUG_ON_PAGE(!PageHead(page), page); 2341 2342 if (PageAnon(page)) 2343 ttu_flags |= TTU_SPLIT_FREEZE; 2344 2345 unmap_success = try_to_unmap(page, ttu_flags); 2346 VM_BUG_ON_PAGE(!unmap_success, page); 2347 } 2348 2349 static void unfreeze_page(struct page *page) 2350 { 2351 int i; 2352 if (PageTransHuge(page)) { 2353 remove_migration_ptes(page, page, true); 2354 } else { 2355 for (i = 0; i < HPAGE_PMD_NR; i++) 2356 remove_migration_ptes(page + i, page + i, true); 2357 } 2358 } 2359 2360 static void __split_huge_page_tail(struct page *head, int tail, 2361 struct lruvec *lruvec, struct list_head *list) 2362 { 2363 struct page *page_tail = head + tail; 2364 2365 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2366 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); 2367 2368 /* 2369 * tail_page->_refcount is zero and not changing from under us. But 2370 * get_page_unless_zero() may be running from under us on the 2371 * tail_page. If we used atomic_set() below instead of atomic_inc() or 2372 * atomic_add(), we would then run atomic_set() concurrently with 2373 * get_page_unless_zero(), and atomic_set() is implemented in C not 2374 * using locked ops. spin_unlock on x86 sometime uses locked ops 2375 * because of PPro errata 66, 92, so unless somebody can guarantee 2376 * atomic_set() here would be safe on all archs (and not only on x86), 2377 * it's safer to use atomic_inc()/atomic_add(). 2378 */ 2379 if (PageAnon(head) && !PageSwapCache(head)) { 2380 page_ref_inc(page_tail); 2381 } else { 2382 /* Additional pin to radix tree */ 2383 page_ref_add(page_tail, 2); 2384 } 2385 2386 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2387 page_tail->flags |= (head->flags & 2388 ((1L << PG_referenced) | 2389 (1L << PG_swapbacked) | 2390 (1L << PG_swapcache) | 2391 (1L << PG_mlocked) | 2392 (1L << PG_uptodate) | 2393 (1L << PG_active) | 2394 (1L << PG_locked) | 2395 (1L << PG_unevictable) | 2396 (1L << PG_dirty))); 2397 2398 /* 2399 * After clearing PageTail the gup refcount can be released. 2400 * Page flags also must be visible before we make the page non-compound. 2401 */ 2402 smp_wmb(); 2403 2404 clear_compound_head(page_tail); 2405 2406 if (page_is_young(head)) 2407 set_page_young(page_tail); 2408 if (page_is_idle(head)) 2409 set_page_idle(page_tail); 2410 2411 /* ->mapping in first tail page is compound_mapcount */ 2412 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2413 page_tail); 2414 page_tail->mapping = head->mapping; 2415 2416 page_tail->index = head->index + tail; 2417 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2418 lru_add_page_tail(head, page_tail, lruvec, list); 2419 } 2420 2421 static void __split_huge_page(struct page *page, struct list_head *list, 2422 unsigned long flags) 2423 { 2424 struct page *head = compound_head(page); 2425 struct zone *zone = page_zone(head); 2426 struct lruvec *lruvec; 2427 pgoff_t end = -1; 2428 int i; 2429 2430 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2431 2432 /* complete memcg works before add pages to LRU */ 2433 mem_cgroup_split_huge_fixup(head); 2434 2435 if (!PageAnon(page)) 2436 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); 2437 2438 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2439 __split_huge_page_tail(head, i, lruvec, list); 2440 /* Some pages can be beyond i_size: drop them from page cache */ 2441 if (head[i].index >= end) { 2442 __ClearPageDirty(head + i); 2443 __delete_from_page_cache(head + i, NULL); 2444 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2445 shmem_uncharge(head->mapping->host, 1); 2446 put_page(head + i); 2447 } 2448 } 2449 2450 ClearPageCompound(head); 2451 /* See comment in __split_huge_page_tail() */ 2452 if (PageAnon(head)) { 2453 /* Additional pin to radix tree of swap cache */ 2454 if (PageSwapCache(head)) 2455 page_ref_add(head, 2); 2456 else 2457 page_ref_inc(head); 2458 } else { 2459 /* Additional pin to radix tree */ 2460 page_ref_add(head, 2); 2461 spin_unlock(&head->mapping->tree_lock); 2462 } 2463 2464 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2465 2466 unfreeze_page(head); 2467 2468 for (i = 0; i < HPAGE_PMD_NR; i++) { 2469 struct page *subpage = head + i; 2470 if (subpage == page) 2471 continue; 2472 unlock_page(subpage); 2473 2474 /* 2475 * Subpages may be freed if there wasn't any mapping 2476 * like if add_to_swap() is running on a lru page that 2477 * had its mapping zapped. And freeing these pages 2478 * requires taking the lru_lock so we do the put_page 2479 * of the tail pages after the split is complete. 2480 */ 2481 put_page(subpage); 2482 } 2483 } 2484 2485 int total_mapcount(struct page *page) 2486 { 2487 int i, compound, ret; 2488 2489 VM_BUG_ON_PAGE(PageTail(page), page); 2490 2491 if (likely(!PageCompound(page))) 2492 return atomic_read(&page->_mapcount) + 1; 2493 2494 compound = compound_mapcount(page); 2495 if (PageHuge(page)) 2496 return compound; 2497 ret = compound; 2498 for (i = 0; i < HPAGE_PMD_NR; i++) 2499 ret += atomic_read(&page[i]._mapcount) + 1; 2500 /* File pages has compound_mapcount included in _mapcount */ 2501 if (!PageAnon(page)) 2502 return ret - compound * HPAGE_PMD_NR; 2503 if (PageDoubleMap(page)) 2504 ret -= HPAGE_PMD_NR; 2505 return ret; 2506 } 2507 2508 /* 2509 * This calculates accurately how many mappings a transparent hugepage 2510 * has (unlike page_mapcount() which isn't fully accurate). This full 2511 * accuracy is primarily needed to know if copy-on-write faults can 2512 * reuse the page and change the mapping to read-write instead of 2513 * copying them. At the same time this returns the total_mapcount too. 2514 * 2515 * The function returns the highest mapcount any one of the subpages 2516 * has. If the return value is one, even if different processes are 2517 * mapping different subpages of the transparent hugepage, they can 2518 * all reuse it, because each process is reusing a different subpage. 2519 * 2520 * The total_mapcount is instead counting all virtual mappings of the 2521 * subpages. If the total_mapcount is equal to "one", it tells the 2522 * caller all mappings belong to the same "mm" and in turn the 2523 * anon_vma of the transparent hugepage can become the vma->anon_vma 2524 * local one as no other process may be mapping any of the subpages. 2525 * 2526 * It would be more accurate to replace page_mapcount() with 2527 * page_trans_huge_mapcount(), however we only use 2528 * page_trans_huge_mapcount() in the copy-on-write faults where we 2529 * need full accuracy to avoid breaking page pinning, because 2530 * page_trans_huge_mapcount() is slower than page_mapcount(). 2531 */ 2532 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2533 { 2534 int i, ret, _total_mapcount, mapcount; 2535 2536 /* hugetlbfs shouldn't call it */ 2537 VM_BUG_ON_PAGE(PageHuge(page), page); 2538 2539 if (likely(!PageTransCompound(page))) { 2540 mapcount = atomic_read(&page->_mapcount) + 1; 2541 if (total_mapcount) 2542 *total_mapcount = mapcount; 2543 return mapcount; 2544 } 2545 2546 page = compound_head(page); 2547 2548 _total_mapcount = ret = 0; 2549 for (i = 0; i < HPAGE_PMD_NR; i++) { 2550 mapcount = atomic_read(&page[i]._mapcount) + 1; 2551 ret = max(ret, mapcount); 2552 _total_mapcount += mapcount; 2553 } 2554 if (PageDoubleMap(page)) { 2555 ret -= 1; 2556 _total_mapcount -= HPAGE_PMD_NR; 2557 } 2558 mapcount = compound_mapcount(page); 2559 ret += mapcount; 2560 _total_mapcount += mapcount; 2561 if (total_mapcount) 2562 *total_mapcount = _total_mapcount; 2563 return ret; 2564 } 2565 2566 /* Racy check whether the huge page can be split */ 2567 bool can_split_huge_page(struct page *page, int *pextra_pins) 2568 { 2569 int extra_pins; 2570 2571 /* Additional pins from radix tree */ 2572 if (PageAnon(page)) 2573 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2574 else 2575 extra_pins = HPAGE_PMD_NR; 2576 if (pextra_pins) 2577 *pextra_pins = extra_pins; 2578 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2579 } 2580 2581 /* 2582 * This function splits huge page into normal pages. @page can point to any 2583 * subpage of huge page to split. Split doesn't change the position of @page. 2584 * 2585 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2586 * The huge page must be locked. 2587 * 2588 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2589 * 2590 * Both head page and tail pages will inherit mapping, flags, and so on from 2591 * the hugepage. 2592 * 2593 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2594 * they are not mapped. 2595 * 2596 * Returns 0 if the hugepage is split successfully. 2597 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2598 * us. 2599 */ 2600 int split_huge_page_to_list(struct page *page, struct list_head *list) 2601 { 2602 struct page *head = compound_head(page); 2603 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2604 struct anon_vma *anon_vma = NULL; 2605 struct address_space *mapping = NULL; 2606 int count, mapcount, extra_pins, ret; 2607 bool mlocked; 2608 unsigned long flags; 2609 2610 VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2611 VM_BUG_ON_PAGE(!PageLocked(page), page); 2612 VM_BUG_ON_PAGE(!PageCompound(page), page); 2613 2614 if (PageWriteback(page)) 2615 return -EBUSY; 2616 2617 if (PageAnon(head)) { 2618 /* 2619 * The caller does not necessarily hold an mmap_sem that would 2620 * prevent the anon_vma disappearing so we first we take a 2621 * reference to it and then lock the anon_vma for write. This 2622 * is similar to page_lock_anon_vma_read except the write lock 2623 * is taken to serialise against parallel split or collapse 2624 * operations. 2625 */ 2626 anon_vma = page_get_anon_vma(head); 2627 if (!anon_vma) { 2628 ret = -EBUSY; 2629 goto out; 2630 } 2631 mapping = NULL; 2632 anon_vma_lock_write(anon_vma); 2633 } else { 2634 mapping = head->mapping; 2635 2636 /* Truncated ? */ 2637 if (!mapping) { 2638 ret = -EBUSY; 2639 goto out; 2640 } 2641 2642 anon_vma = NULL; 2643 i_mmap_lock_read(mapping); 2644 } 2645 2646 /* 2647 * Racy check if we can split the page, before freeze_page() will 2648 * split PMDs 2649 */ 2650 if (!can_split_huge_page(head, &extra_pins)) { 2651 ret = -EBUSY; 2652 goto out_unlock; 2653 } 2654 2655 mlocked = PageMlocked(page); 2656 freeze_page(head); 2657 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2658 2659 /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2660 if (mlocked) 2661 lru_add_drain(); 2662 2663 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2664 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2665 2666 if (mapping) { 2667 void **pslot; 2668 2669 spin_lock(&mapping->tree_lock); 2670 pslot = radix_tree_lookup_slot(&mapping->page_tree, 2671 page_index(head)); 2672 /* 2673 * Check if the head page is present in radix tree. 2674 * We assume all tail are present too, if head is there. 2675 */ 2676 if (radix_tree_deref_slot_protected(pslot, 2677 &mapping->tree_lock) != head) 2678 goto fail; 2679 } 2680 2681 /* Prevent deferred_split_scan() touching ->_refcount */ 2682 spin_lock(&pgdata->split_queue_lock); 2683 count = page_count(head); 2684 mapcount = total_mapcount(head); 2685 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2686 if (!list_empty(page_deferred_list(head))) { 2687 pgdata->split_queue_len--; 2688 list_del(page_deferred_list(head)); 2689 } 2690 if (mapping) 2691 __dec_node_page_state(page, NR_SHMEM_THPS); 2692 spin_unlock(&pgdata->split_queue_lock); 2693 __split_huge_page(page, list, flags); 2694 if (PageSwapCache(head)) { 2695 swp_entry_t entry = { .val = page_private(head) }; 2696 2697 ret = split_swap_cluster(entry); 2698 } else 2699 ret = 0; 2700 } else { 2701 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2702 pr_alert("total_mapcount: %u, page_count(): %u\n", 2703 mapcount, count); 2704 if (PageTail(page)) 2705 dump_page(head, NULL); 2706 dump_page(page, "total_mapcount(head) > 0"); 2707 BUG(); 2708 } 2709 spin_unlock(&pgdata->split_queue_lock); 2710 fail: if (mapping) 2711 spin_unlock(&mapping->tree_lock); 2712 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2713 unfreeze_page(head); 2714 ret = -EBUSY; 2715 } 2716 2717 out_unlock: 2718 if (anon_vma) { 2719 anon_vma_unlock_write(anon_vma); 2720 put_anon_vma(anon_vma); 2721 } 2722 if (mapping) 2723 i_mmap_unlock_read(mapping); 2724 out: 2725 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2726 return ret; 2727 } 2728 2729 void free_transhuge_page(struct page *page) 2730 { 2731 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2732 unsigned long flags; 2733 2734 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2735 if (!list_empty(page_deferred_list(page))) { 2736 pgdata->split_queue_len--; 2737 list_del(page_deferred_list(page)); 2738 } 2739 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2740 free_compound_page(page); 2741 } 2742 2743 void deferred_split_huge_page(struct page *page) 2744 { 2745 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2746 unsigned long flags; 2747 2748 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2749 2750 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2751 if (list_empty(page_deferred_list(page))) { 2752 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2753 list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2754 pgdata->split_queue_len++; 2755 } 2756 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2757 } 2758 2759 static unsigned long deferred_split_count(struct shrinker *shrink, 2760 struct shrink_control *sc) 2761 { 2762 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2763 return READ_ONCE(pgdata->split_queue_len); 2764 } 2765 2766 static unsigned long deferred_split_scan(struct shrinker *shrink, 2767 struct shrink_control *sc) 2768 { 2769 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2770 unsigned long flags; 2771 LIST_HEAD(list), *pos, *next; 2772 struct page *page; 2773 int split = 0; 2774 2775 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2776 /* Take pin on all head pages to avoid freeing them under us */ 2777 list_for_each_safe(pos, next, &pgdata->split_queue) { 2778 page = list_entry((void *)pos, struct page, mapping); 2779 page = compound_head(page); 2780 if (get_page_unless_zero(page)) { 2781 list_move(page_deferred_list(page), &list); 2782 } else { 2783 /* We lost race with put_compound_page() */ 2784 list_del_init(page_deferred_list(page)); 2785 pgdata->split_queue_len--; 2786 } 2787 if (!--sc->nr_to_scan) 2788 break; 2789 } 2790 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2791 2792 list_for_each_safe(pos, next, &list) { 2793 page = list_entry((void *)pos, struct page, mapping); 2794 lock_page(page); 2795 /* split_huge_page() removes page from list on success */ 2796 if (!split_huge_page(page)) 2797 split++; 2798 unlock_page(page); 2799 put_page(page); 2800 } 2801 2802 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2803 list_splice_tail(&list, &pgdata->split_queue); 2804 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2805 2806 /* 2807 * Stop shrinker if we didn't split any page, but the queue is empty. 2808 * This can happen if pages were freed under us. 2809 */ 2810 if (!split && list_empty(&pgdata->split_queue)) 2811 return SHRINK_STOP; 2812 return split; 2813 } 2814 2815 static struct shrinker deferred_split_shrinker = { 2816 .count_objects = deferred_split_count, 2817 .scan_objects = deferred_split_scan, 2818 .seeks = DEFAULT_SEEKS, 2819 .flags = SHRINKER_NUMA_AWARE, 2820 }; 2821 2822 #ifdef CONFIG_DEBUG_FS 2823 static int split_huge_pages_set(void *data, u64 val) 2824 { 2825 struct zone *zone; 2826 struct page *page; 2827 unsigned long pfn, max_zone_pfn; 2828 unsigned long total = 0, split = 0; 2829 2830 if (val != 1) 2831 return -EINVAL; 2832 2833 for_each_populated_zone(zone) { 2834 max_zone_pfn = zone_end_pfn(zone); 2835 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2836 if (!pfn_valid(pfn)) 2837 continue; 2838 2839 page = pfn_to_page(pfn); 2840 if (!get_page_unless_zero(page)) 2841 continue; 2842 2843 if (zone != page_zone(page)) 2844 goto next; 2845 2846 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2847 goto next; 2848 2849 total++; 2850 lock_page(page); 2851 if (!split_huge_page(page)) 2852 split++; 2853 unlock_page(page); 2854 next: 2855 put_page(page); 2856 } 2857 } 2858 2859 pr_info("%lu of %lu THP split\n", split, total); 2860 2861 return 0; 2862 } 2863 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2864 "%llu\n"); 2865 2866 static int __init split_huge_pages_debugfs(void) 2867 { 2868 void *ret; 2869 2870 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2871 &split_huge_pages_fops); 2872 if (!ret) 2873 pr_warn("Failed to create split_huge_pages in debugfs"); 2874 return 0; 2875 } 2876 late_initcall(split_huge_pages_debugfs); 2877 #endif 2878 2879 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2880 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2881 struct page *page) 2882 { 2883 struct vm_area_struct *vma = pvmw->vma; 2884 struct mm_struct *mm = vma->vm_mm; 2885 unsigned long address = pvmw->address; 2886 pmd_t pmdval; 2887 swp_entry_t entry; 2888 pmd_t pmdswp; 2889 2890 if (!(pvmw->pmd && !pvmw->pte)) 2891 return; 2892 2893 mmu_notifier_invalidate_range_start(mm, address, 2894 address + HPAGE_PMD_SIZE); 2895 2896 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2897 pmdval = *pvmw->pmd; 2898 pmdp_invalidate(vma, address, pvmw->pmd); 2899 if (pmd_dirty(pmdval)) 2900 set_page_dirty(page); 2901 entry = make_migration_entry(page, pmd_write(pmdval)); 2902 pmdswp = swp_entry_to_pmd(entry); 2903 if (pmd_soft_dirty(pmdval)) 2904 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2905 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2906 page_remove_rmap(page, true); 2907 put_page(page); 2908 2909 mmu_notifier_invalidate_range_end(mm, address, 2910 address + HPAGE_PMD_SIZE); 2911 } 2912 2913 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2914 { 2915 struct vm_area_struct *vma = pvmw->vma; 2916 struct mm_struct *mm = vma->vm_mm; 2917 unsigned long address = pvmw->address; 2918 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2919 pmd_t pmde; 2920 swp_entry_t entry; 2921 2922 if (!(pvmw->pmd && !pvmw->pte)) 2923 return; 2924 2925 entry = pmd_to_swp_entry(*pvmw->pmd); 2926 get_page(new); 2927 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2928 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2929 pmde = pmd_mksoft_dirty(pmde); 2930 if (is_write_migration_entry(entry)) 2931 pmde = maybe_pmd_mkwrite(pmde, vma); 2932 2933 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2934 page_add_anon_rmap(new, vma, mmun_start, true); 2935 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2936 if (vma->vm_flags & VM_LOCKED) 2937 mlock_vma_page(new); 2938 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2939 } 2940 #endif 2941