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