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 that avoid 43 * to risk increase the memory footprint of applications without a guaranteed 44 * benefit. When transparent hugepage support is enabled, is for all mappings, 45 * and khugepaged scans all mappings. 46 * Defrag is invoked by khugepaged hugepage allocations and by page faults 47 * for all hugepage allocations. 48 */ 49 unsigned long transparent_hugepage_flags __read_mostly = 50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 51 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 52 #endif 53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 55 #endif 56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 59 60 static struct shrinker deferred_split_shrinker; 61 62 static atomic_t huge_zero_refcount; 63 struct page *huge_zero_page __read_mostly; 64 65 static struct page *get_huge_zero_page(void) 66 { 67 struct page *zero_page; 68 retry: 69 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 70 return READ_ONCE(huge_zero_page); 71 72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 73 HPAGE_PMD_ORDER); 74 if (!zero_page) { 75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 76 return NULL; 77 } 78 count_vm_event(THP_ZERO_PAGE_ALLOC); 79 preempt_disable(); 80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 81 preempt_enable(); 82 __free_pages(zero_page, compound_order(zero_page)); 83 goto retry; 84 } 85 86 /* We take additional reference here. It will be put back by shrinker */ 87 atomic_set(&huge_zero_refcount, 2); 88 preempt_enable(); 89 return READ_ONCE(huge_zero_page); 90 } 91 92 static void put_huge_zero_page(void) 93 { 94 /* 95 * Counter should never go to zero here. Only shrinker can put 96 * last reference. 97 */ 98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 99 } 100 101 struct page *mm_get_huge_zero_page(struct mm_struct *mm) 102 { 103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 104 return READ_ONCE(huge_zero_page); 105 106 if (!get_huge_zero_page()) 107 return NULL; 108 109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 110 put_huge_zero_page(); 111 112 return READ_ONCE(huge_zero_page); 113 } 114 115 void mm_put_huge_zero_page(struct mm_struct *mm) 116 { 117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 118 put_huge_zero_page(); 119 } 120 121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 122 struct shrink_control *sc) 123 { 124 /* we can free zero page only if last reference remains */ 125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 126 } 127 128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 129 struct shrink_control *sc) 130 { 131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 132 struct page *zero_page = xchg(&huge_zero_page, NULL); 133 BUG_ON(zero_page == NULL); 134 __free_pages(zero_page, compound_order(zero_page)); 135 return HPAGE_PMD_NR; 136 } 137 138 return 0; 139 } 140 141 static struct shrinker huge_zero_page_shrinker = { 142 .count_objects = shrink_huge_zero_page_count, 143 .scan_objects = shrink_huge_zero_page_scan, 144 .seeks = DEFAULT_SEEKS, 145 }; 146 147 #ifdef CONFIG_SYSFS 148 static ssize_t enabled_show(struct kobject *kobj, 149 struct kobj_attribute *attr, char *buf) 150 { 151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 152 return sprintf(buf, "[always] madvise never\n"); 153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) 154 return sprintf(buf, "always [madvise] never\n"); 155 else 156 return sprintf(buf, "always madvise [never]\n"); 157 } 158 159 static ssize_t enabled_store(struct kobject *kobj, 160 struct kobj_attribute *attr, 161 const char *buf, size_t count) 162 { 163 ssize_t ret = count; 164 165 if (!memcmp("always", buf, 166 min(sizeof("always")-1, count))) { 167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 169 } else if (!memcmp("madvise", buf, 170 min(sizeof("madvise")-1, count))) { 171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 173 } else if (!memcmp("never", buf, 174 min(sizeof("never")-1, count))) { 175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 177 } else 178 ret = -EINVAL; 179 180 if (ret > 0) { 181 int err = start_stop_khugepaged(); 182 if (err) 183 ret = err; 184 } 185 return ret; 186 } 187 static struct kobj_attribute enabled_attr = 188 __ATTR(enabled, 0644, enabled_show, enabled_store); 189 190 ssize_t single_hugepage_flag_show(struct kobject *kobj, 191 struct kobj_attribute *attr, char *buf, 192 enum transparent_hugepage_flag flag) 193 { 194 return sprintf(buf, "%d\n", 195 !!test_bit(flag, &transparent_hugepage_flags)); 196 } 197 198 ssize_t single_hugepage_flag_store(struct kobject *kobj, 199 struct kobj_attribute *attr, 200 const char *buf, size_t count, 201 enum transparent_hugepage_flag flag) 202 { 203 unsigned long value; 204 int ret; 205 206 ret = kstrtoul(buf, 10, &value); 207 if (ret < 0) 208 return ret; 209 if (value > 1) 210 return -EINVAL; 211 212 if (value) 213 set_bit(flag, &transparent_hugepage_flags); 214 else 215 clear_bit(flag, &transparent_hugepage_flags); 216 217 return count; 218 } 219 220 static ssize_t defrag_show(struct kobject *kobj, 221 struct kobj_attribute *attr, char *buf) 222 { 223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 224 return sprintf(buf, "[always] defer defer+madvise madvise never\n"); 225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 226 return sprintf(buf, "always [defer] defer+madvise madvise never\n"); 227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 228 return sprintf(buf, "always defer [defer+madvise] madvise never\n"); 229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 230 return sprintf(buf, "always defer defer+madvise [madvise] never\n"); 231 return sprintf(buf, "always defer defer+madvise madvise [never]\n"); 232 } 233 234 static ssize_t defrag_store(struct kobject *kobj, 235 struct kobj_attribute *attr, 236 const char *buf, size_t count) 237 { 238 if (!memcmp("always", buf, 239 min(sizeof("always")-1, count))) { 240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 244 } else if (!memcmp("defer+madvise", buf, 245 min(sizeof("defer+madvise")-1, count))) { 246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 250 } else if (!memcmp("defer", buf, 251 min(sizeof("defer")-1, count))) { 252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 256 } else if (!memcmp("madvise", buf, 257 min(sizeof("madvise")-1, count))) { 258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 262 } else if (!memcmp("never", buf, 263 min(sizeof("never")-1, count))) { 264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 268 } else 269 return -EINVAL; 270 271 return count; 272 } 273 static struct kobj_attribute defrag_attr = 274 __ATTR(defrag, 0644, defrag_show, defrag_store); 275 276 static ssize_t use_zero_page_show(struct kobject *kobj, 277 struct kobj_attribute *attr, char *buf) 278 { 279 return single_hugepage_flag_show(kobj, attr, buf, 280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 281 } 282 static ssize_t use_zero_page_store(struct kobject *kobj, 283 struct kobj_attribute *attr, const char *buf, size_t count) 284 { 285 return single_hugepage_flag_store(kobj, attr, buf, count, 286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 287 } 288 static struct kobj_attribute use_zero_page_attr = 289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 290 291 static ssize_t hpage_pmd_size_show(struct kobject *kobj, 292 struct kobj_attribute *attr, char *buf) 293 { 294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); 295 } 296 static struct kobj_attribute hpage_pmd_size_attr = 297 __ATTR_RO(hpage_pmd_size); 298 299 #ifdef CONFIG_DEBUG_VM 300 static ssize_t debug_cow_show(struct kobject *kobj, 301 struct kobj_attribute *attr, char *buf) 302 { 303 return single_hugepage_flag_show(kobj, attr, buf, 304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 305 } 306 static ssize_t debug_cow_store(struct kobject *kobj, 307 struct kobj_attribute *attr, 308 const char *buf, size_t count) 309 { 310 return single_hugepage_flag_store(kobj, attr, buf, count, 311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 312 } 313 static struct kobj_attribute debug_cow_attr = 314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 315 #endif /* CONFIG_DEBUG_VM */ 316 317 static struct attribute *hugepage_attr[] = { 318 &enabled_attr.attr, 319 &defrag_attr.attr, 320 &use_zero_page_attr.attr, 321 &hpage_pmd_size_attr.attr, 322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 323 &shmem_enabled_attr.attr, 324 #endif 325 #ifdef CONFIG_DEBUG_VM 326 &debug_cow_attr.attr, 327 #endif 328 NULL, 329 }; 330 331 static const struct attribute_group hugepage_attr_group = { 332 .attrs = hugepage_attr, 333 }; 334 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 336 { 337 int err; 338 339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 340 if (unlikely(!*hugepage_kobj)) { 341 pr_err("failed to create transparent hugepage kobject\n"); 342 return -ENOMEM; 343 } 344 345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 346 if (err) { 347 pr_err("failed to register transparent hugepage group\n"); 348 goto delete_obj; 349 } 350 351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 352 if (err) { 353 pr_err("failed to register transparent hugepage group\n"); 354 goto remove_hp_group; 355 } 356 357 return 0; 358 359 remove_hp_group: 360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 361 delete_obj: 362 kobject_put(*hugepage_kobj); 363 return err; 364 } 365 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 367 { 368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 370 kobject_put(hugepage_kobj); 371 } 372 #else 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 374 { 375 return 0; 376 } 377 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 379 { 380 } 381 #endif /* CONFIG_SYSFS */ 382 383 static int __init hugepage_init(void) 384 { 385 int err; 386 struct kobject *hugepage_kobj; 387 388 if (!has_transparent_hugepage()) { 389 transparent_hugepage_flags = 0; 390 return -EINVAL; 391 } 392 393 /* 394 * hugepages can't be allocated by the buddy allocator 395 */ 396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); 397 /* 398 * we use page->mapping and page->index in second tail page 399 * as list_head: assuming THP order >= 2 400 */ 401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 402 403 err = hugepage_init_sysfs(&hugepage_kobj); 404 if (err) 405 goto err_sysfs; 406 407 err = khugepaged_init(); 408 if (err) 409 goto err_slab; 410 411 err = register_shrinker(&huge_zero_page_shrinker); 412 if (err) 413 goto err_hzp_shrinker; 414 err = register_shrinker(&deferred_split_shrinker); 415 if (err) 416 goto err_split_shrinker; 417 418 /* 419 * By default disable transparent hugepages on smaller systems, 420 * where the extra memory used could hurt more than TLB overhead 421 * is likely to save. The admin can still enable it through /sys. 422 */ 423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { 424 transparent_hugepage_flags = 0; 425 return 0; 426 } 427 428 err = start_stop_khugepaged(); 429 if (err) 430 goto err_khugepaged; 431 432 return 0; 433 err_khugepaged: 434 unregister_shrinker(&deferred_split_shrinker); 435 err_split_shrinker: 436 unregister_shrinker(&huge_zero_page_shrinker); 437 err_hzp_shrinker: 438 khugepaged_destroy(); 439 err_slab: 440 hugepage_exit_sysfs(hugepage_kobj); 441 err_sysfs: 442 return err; 443 } 444 subsys_initcall(hugepage_init); 445 446 static int __init setup_transparent_hugepage(char *str) 447 { 448 int ret = 0; 449 if (!str) 450 goto out; 451 if (!strcmp(str, "always")) { 452 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 453 &transparent_hugepage_flags); 454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 455 &transparent_hugepage_flags); 456 ret = 1; 457 } else if (!strcmp(str, "madvise")) { 458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 459 &transparent_hugepage_flags); 460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 461 &transparent_hugepage_flags); 462 ret = 1; 463 } else if (!strcmp(str, "never")) { 464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 465 &transparent_hugepage_flags); 466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 467 &transparent_hugepage_flags); 468 ret = 1; 469 } 470 out: 471 if (!ret) 472 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 473 return ret; 474 } 475 __setup("transparent_hugepage=", setup_transparent_hugepage); 476 477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 478 { 479 if (likely(vma->vm_flags & VM_WRITE)) 480 pmd = pmd_mkwrite(pmd); 481 return pmd; 482 } 483 484 static inline struct list_head *page_deferred_list(struct page *page) 485 { 486 /* 487 * ->lru in the tail pages is occupied by compound_head. 488 * Let's use ->mapping + ->index in the second tail page as list_head. 489 */ 490 return (struct list_head *)&page[2].mapping; 491 } 492 493 void prep_transhuge_page(struct page *page) 494 { 495 /* 496 * we use page->mapping and page->indexlru in second tail page 497 * as list_head: assuming THP order >= 2 498 */ 499 500 INIT_LIST_HEAD(page_deferred_list(page)); 501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); 502 } 503 504 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len, 505 loff_t off, unsigned long flags, unsigned long size) 506 { 507 unsigned long addr; 508 loff_t off_end = off + len; 509 loff_t off_align = round_up(off, size); 510 unsigned long len_pad; 511 512 if (off_end <= off_align || (off_end - off_align) < size) 513 return 0; 514 515 len_pad = len + size; 516 if (len_pad < len || (off + len_pad) < off) 517 return 0; 518 519 addr = current->mm->get_unmapped_area(filp, 0, len_pad, 520 off >> PAGE_SHIFT, flags); 521 if (IS_ERR_VALUE(addr)) 522 return 0; 523 524 addr += (off - addr) & (size - 1); 525 return addr; 526 } 527 528 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 529 unsigned long len, unsigned long pgoff, unsigned long flags) 530 { 531 loff_t off = (loff_t)pgoff << PAGE_SHIFT; 532 533 if (addr) 534 goto out; 535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) 536 goto out; 537 538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE); 539 if (addr) 540 return addr; 541 542 out: 543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 544 } 545 EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 546 547 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page, 548 gfp_t gfp) 549 { 550 struct vm_area_struct *vma = vmf->vma; 551 struct mem_cgroup *memcg; 552 pgtable_t pgtable; 553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 554 int ret = 0; 555 556 VM_BUG_ON_PAGE(!PageCompound(page), page); 557 558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { 559 put_page(page); 560 count_vm_event(THP_FAULT_FALLBACK); 561 return VM_FAULT_FALLBACK; 562 } 563 564 pgtable = pte_alloc_one(vma->vm_mm, haddr); 565 if (unlikely(!pgtable)) { 566 ret = VM_FAULT_OOM; 567 goto release; 568 } 569 570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 571 /* 572 * The memory barrier inside __SetPageUptodate makes sure that 573 * clear_huge_page writes become visible before the set_pmd_at() 574 * write. 575 */ 576 __SetPageUptodate(page); 577 578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 579 if (unlikely(!pmd_none(*vmf->pmd))) { 580 goto unlock_release; 581 } else { 582 pmd_t entry; 583 584 ret = check_stable_address_space(vma->vm_mm); 585 if (ret) 586 goto unlock_release; 587 588 /* Deliver the page fault to userland */ 589 if (userfaultfd_missing(vma)) { 590 int ret; 591 592 spin_unlock(vmf->ptl); 593 mem_cgroup_cancel_charge(page, memcg, true); 594 put_page(page); 595 pte_free(vma->vm_mm, pgtable); 596 ret = handle_userfault(vmf, VM_UFFD_MISSING); 597 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 598 return ret; 599 } 600 601 entry = mk_huge_pmd(page, vma->vm_page_prot); 602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 603 page_add_new_anon_rmap(page, vma, haddr, true); 604 mem_cgroup_commit_charge(page, memcg, false, true); 605 lru_cache_add_active_or_unevictable(page, vma); 606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 609 atomic_long_inc(&vma->vm_mm->nr_ptes); 610 spin_unlock(vmf->ptl); 611 count_vm_event(THP_FAULT_ALLOC); 612 } 613 614 return 0; 615 unlock_release: 616 spin_unlock(vmf->ptl); 617 release: 618 if (pgtable) 619 pte_free(vma->vm_mm, pgtable); 620 mem_cgroup_cancel_charge(page, memcg, true); 621 put_page(page); 622 return ret; 623 624 } 625 626 /* 627 * always: directly stall for all thp allocations 628 * defer: wake kswapd and fail if not immediately available 629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 630 * fail if not immediately available 631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 632 * available 633 * never: never stall for any thp allocation 634 */ 635 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) 636 { 637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 638 639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 645 __GFP_KSWAPD_RECLAIM); 646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 648 0); 649 return GFP_TRANSHUGE_LIGHT; 650 } 651 652 /* Caller must hold page table lock. */ 653 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 655 struct page *zero_page) 656 { 657 pmd_t entry; 658 if (!pmd_none(*pmd)) 659 return false; 660 entry = mk_pmd(zero_page, vma->vm_page_prot); 661 entry = pmd_mkhuge(entry); 662 if (pgtable) 663 pgtable_trans_huge_deposit(mm, pmd, pgtable); 664 set_pmd_at(mm, haddr, pmd, entry); 665 atomic_long_inc(&mm->nr_ptes); 666 return true; 667 } 668 669 int do_huge_pmd_anonymous_page(struct vm_fault *vmf) 670 { 671 struct vm_area_struct *vma = vmf->vma; 672 gfp_t gfp; 673 struct page *page; 674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 675 676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 677 return VM_FAULT_FALLBACK; 678 if (unlikely(anon_vma_prepare(vma))) 679 return VM_FAULT_OOM; 680 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 681 return VM_FAULT_OOM; 682 if (!(vmf->flags & FAULT_FLAG_WRITE) && 683 !mm_forbids_zeropage(vma->vm_mm) && 684 transparent_hugepage_use_zero_page()) { 685 pgtable_t pgtable; 686 struct page *zero_page; 687 bool set; 688 int ret; 689 pgtable = pte_alloc_one(vma->vm_mm, haddr); 690 if (unlikely(!pgtable)) 691 return VM_FAULT_OOM; 692 zero_page = mm_get_huge_zero_page(vma->vm_mm); 693 if (unlikely(!zero_page)) { 694 pte_free(vma->vm_mm, pgtable); 695 count_vm_event(THP_FAULT_FALLBACK); 696 return VM_FAULT_FALLBACK; 697 } 698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 699 ret = 0; 700 set = false; 701 if (pmd_none(*vmf->pmd)) { 702 ret = check_stable_address_space(vma->vm_mm); 703 if (ret) { 704 spin_unlock(vmf->ptl); 705 } else if (userfaultfd_missing(vma)) { 706 spin_unlock(vmf->ptl); 707 ret = handle_userfault(vmf, VM_UFFD_MISSING); 708 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 709 } else { 710 set_huge_zero_page(pgtable, vma->vm_mm, vma, 711 haddr, vmf->pmd, zero_page); 712 spin_unlock(vmf->ptl); 713 set = true; 714 } 715 } else 716 spin_unlock(vmf->ptl); 717 if (!set) 718 pte_free(vma->vm_mm, pgtable); 719 return ret; 720 } 721 gfp = alloc_hugepage_direct_gfpmask(vma); 722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 723 if (unlikely(!page)) { 724 count_vm_event(THP_FAULT_FALLBACK); 725 return VM_FAULT_FALLBACK; 726 } 727 prep_transhuge_page(page); 728 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 729 } 730 731 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 733 pgtable_t pgtable) 734 { 735 struct mm_struct *mm = vma->vm_mm; 736 pmd_t entry; 737 spinlock_t *ptl; 738 739 ptl = pmd_lock(mm, pmd); 740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 741 if (pfn_t_devmap(pfn)) 742 entry = pmd_mkdevmap(entry); 743 if (write) { 744 entry = pmd_mkyoung(pmd_mkdirty(entry)); 745 entry = maybe_pmd_mkwrite(entry, vma); 746 } 747 748 if (pgtable) { 749 pgtable_trans_huge_deposit(mm, pmd, pgtable); 750 atomic_long_inc(&mm->nr_ptes); 751 } 752 753 set_pmd_at(mm, addr, pmd, entry); 754 update_mmu_cache_pmd(vma, addr, pmd); 755 spin_unlock(ptl); 756 } 757 758 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 759 pmd_t *pmd, pfn_t pfn, bool write) 760 { 761 pgprot_t pgprot = vma->vm_page_prot; 762 pgtable_t pgtable = NULL; 763 /* 764 * If we had pmd_special, we could avoid all these restrictions, 765 * but we need to be consistent with PTEs and architectures that 766 * can't support a 'special' bit. 767 */ 768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 770 (VM_PFNMAP|VM_MIXEDMAP)); 771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 772 BUG_ON(!pfn_t_devmap(pfn)); 773 774 if (addr < vma->vm_start || addr >= vma->vm_end) 775 return VM_FAULT_SIGBUS; 776 777 if (arch_needs_pgtable_deposit()) { 778 pgtable = pte_alloc_one(vma->vm_mm, addr); 779 if (!pgtable) 780 return VM_FAULT_OOM; 781 } 782 783 track_pfn_insert(vma, &pgprot, pfn); 784 785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 786 return VM_FAULT_NOPAGE; 787 } 788 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 789 790 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 791 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 792 { 793 if (likely(vma->vm_flags & VM_WRITE)) 794 pud = pud_mkwrite(pud); 795 return pud; 796 } 797 798 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 800 { 801 struct mm_struct *mm = vma->vm_mm; 802 pud_t entry; 803 spinlock_t *ptl; 804 805 ptl = pud_lock(mm, pud); 806 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 807 if (pfn_t_devmap(pfn)) 808 entry = pud_mkdevmap(entry); 809 if (write) { 810 entry = pud_mkyoung(pud_mkdirty(entry)); 811 entry = maybe_pud_mkwrite(entry, vma); 812 } 813 set_pud_at(mm, addr, pud, entry); 814 update_mmu_cache_pud(vma, addr, pud); 815 spin_unlock(ptl); 816 } 817 818 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 819 pud_t *pud, pfn_t pfn, bool write) 820 { 821 pgprot_t pgprot = vma->vm_page_prot; 822 /* 823 * If we had pud_special, we could avoid all these restrictions, 824 * but we need to be consistent with PTEs and architectures that 825 * can't support a 'special' bit. 826 */ 827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 829 (VM_PFNMAP|VM_MIXEDMAP)); 830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 831 BUG_ON(!pfn_t_devmap(pfn)); 832 833 if (addr < vma->vm_start || addr >= vma->vm_end) 834 return VM_FAULT_SIGBUS; 835 836 track_pfn_insert(vma, &pgprot, pfn); 837 838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 839 return VM_FAULT_NOPAGE; 840 } 841 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 842 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 843 844 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 845 pmd_t *pmd) 846 { 847 pmd_t _pmd; 848 849 /* 850 * We should set the dirty bit only for FOLL_WRITE but for now 851 * the dirty bit in the pmd is meaningless. And if the dirty 852 * bit will become meaningful and we'll only set it with 853 * FOLL_WRITE, an atomic set_bit will be required on the pmd to 854 * set the young bit, instead of the current set_pmd_at. 855 */ 856 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 857 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 858 pmd, _pmd, 1)) 859 update_mmu_cache_pmd(vma, addr, pmd); 860 } 861 862 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 863 pmd_t *pmd, int flags) 864 { 865 unsigned long pfn = pmd_pfn(*pmd); 866 struct mm_struct *mm = vma->vm_mm; 867 struct dev_pagemap *pgmap; 868 struct page *page; 869 870 assert_spin_locked(pmd_lockptr(mm, pmd)); 871 872 /* 873 * When we COW a devmap PMD entry, we split it into PTEs, so we should 874 * not be in this function with `flags & FOLL_COW` set. 875 */ 876 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 877 878 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 879 return NULL; 880 881 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 882 /* pass */; 883 else 884 return NULL; 885 886 if (flags & FOLL_TOUCH) 887 touch_pmd(vma, addr, pmd); 888 889 /* 890 * device mapped pages can only be returned if the 891 * caller will manage the page reference count. 892 */ 893 if (!(flags & FOLL_GET)) 894 return ERR_PTR(-EEXIST); 895 896 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 897 pgmap = get_dev_pagemap(pfn, NULL); 898 if (!pgmap) 899 return ERR_PTR(-EFAULT); 900 page = pfn_to_page(pfn); 901 get_page(page); 902 put_dev_pagemap(pgmap); 903 904 return page; 905 } 906 907 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 908 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 909 struct vm_area_struct *vma) 910 { 911 spinlock_t *dst_ptl, *src_ptl; 912 struct page *src_page; 913 pmd_t pmd; 914 pgtable_t pgtable = NULL; 915 int ret = -ENOMEM; 916 917 /* Skip if can be re-fill on fault */ 918 if (!vma_is_anonymous(vma)) 919 return 0; 920 921 pgtable = pte_alloc_one(dst_mm, addr); 922 if (unlikely(!pgtable)) 923 goto out; 924 925 dst_ptl = pmd_lock(dst_mm, dst_pmd); 926 src_ptl = pmd_lockptr(src_mm, src_pmd); 927 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 928 929 ret = -EAGAIN; 930 pmd = *src_pmd; 931 932 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 933 if (unlikely(is_swap_pmd(pmd))) { 934 swp_entry_t entry = pmd_to_swp_entry(pmd); 935 936 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 937 if (is_write_migration_entry(entry)) { 938 make_migration_entry_read(&entry); 939 pmd = swp_entry_to_pmd(entry); 940 if (pmd_swp_soft_dirty(*src_pmd)) 941 pmd = pmd_swp_mksoft_dirty(pmd); 942 set_pmd_at(src_mm, addr, src_pmd, pmd); 943 } 944 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 945 atomic_long_inc(&dst_mm->nr_ptes); 946 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 947 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 948 ret = 0; 949 goto out_unlock; 950 } 951 #endif 952 953 if (unlikely(!pmd_trans_huge(pmd))) { 954 pte_free(dst_mm, pgtable); 955 goto out_unlock; 956 } 957 /* 958 * When page table lock is held, the huge zero pmd should not be 959 * under splitting since we don't split the page itself, only pmd to 960 * a page table. 961 */ 962 if (is_huge_zero_pmd(pmd)) { 963 struct page *zero_page; 964 /* 965 * get_huge_zero_page() will never allocate a new page here, 966 * since we already have a zero page to copy. It just takes a 967 * reference. 968 */ 969 zero_page = mm_get_huge_zero_page(dst_mm); 970 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 971 zero_page); 972 ret = 0; 973 goto out_unlock; 974 } 975 976 src_page = pmd_page(pmd); 977 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 978 get_page(src_page); 979 page_dup_rmap(src_page, true); 980 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 981 atomic_long_inc(&dst_mm->nr_ptes); 982 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 983 984 pmdp_set_wrprotect(src_mm, addr, src_pmd); 985 pmd = pmd_mkold(pmd_wrprotect(pmd)); 986 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 987 988 ret = 0; 989 out_unlock: 990 spin_unlock(src_ptl); 991 spin_unlock(dst_ptl); 992 out: 993 return ret; 994 } 995 996 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 997 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 998 pud_t *pud) 999 { 1000 pud_t _pud; 1001 1002 /* 1003 * We should set the dirty bit only for FOLL_WRITE but for now 1004 * the dirty bit in the pud is meaningless. And if the dirty 1005 * bit will become meaningful and we'll only set it with 1006 * FOLL_WRITE, an atomic set_bit will be required on the pud to 1007 * set the young bit, instead of the current set_pud_at. 1008 */ 1009 _pud = pud_mkyoung(pud_mkdirty(*pud)); 1010 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1011 pud, _pud, 1)) 1012 update_mmu_cache_pud(vma, addr, pud); 1013 } 1014 1015 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1016 pud_t *pud, int flags) 1017 { 1018 unsigned long pfn = pud_pfn(*pud); 1019 struct mm_struct *mm = vma->vm_mm; 1020 struct dev_pagemap *pgmap; 1021 struct page *page; 1022 1023 assert_spin_locked(pud_lockptr(mm, pud)); 1024 1025 if (flags & FOLL_WRITE && !pud_write(*pud)) 1026 return NULL; 1027 1028 if (pud_present(*pud) && pud_devmap(*pud)) 1029 /* pass */; 1030 else 1031 return NULL; 1032 1033 if (flags & FOLL_TOUCH) 1034 touch_pud(vma, addr, pud); 1035 1036 /* 1037 * device mapped pages can only be returned if the 1038 * caller will manage the page reference count. 1039 */ 1040 if (!(flags & FOLL_GET)) 1041 return ERR_PTR(-EEXIST); 1042 1043 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1044 pgmap = get_dev_pagemap(pfn, NULL); 1045 if (!pgmap) 1046 return ERR_PTR(-EFAULT); 1047 page = pfn_to_page(pfn); 1048 get_page(page); 1049 put_dev_pagemap(pgmap); 1050 1051 return page; 1052 } 1053 1054 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1055 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1056 struct vm_area_struct *vma) 1057 { 1058 spinlock_t *dst_ptl, *src_ptl; 1059 pud_t pud; 1060 int ret; 1061 1062 dst_ptl = pud_lock(dst_mm, dst_pud); 1063 src_ptl = pud_lockptr(src_mm, src_pud); 1064 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1065 1066 ret = -EAGAIN; 1067 pud = *src_pud; 1068 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1069 goto out_unlock; 1070 1071 /* 1072 * When page table lock is held, the huge zero pud should not be 1073 * under splitting since we don't split the page itself, only pud to 1074 * a page table. 1075 */ 1076 if (is_huge_zero_pud(pud)) { 1077 /* No huge zero pud yet */ 1078 } 1079 1080 pudp_set_wrprotect(src_mm, addr, src_pud); 1081 pud = pud_mkold(pud_wrprotect(pud)); 1082 set_pud_at(dst_mm, addr, dst_pud, pud); 1083 1084 ret = 0; 1085 out_unlock: 1086 spin_unlock(src_ptl); 1087 spin_unlock(dst_ptl); 1088 return ret; 1089 } 1090 1091 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1092 { 1093 pud_t entry; 1094 unsigned long haddr; 1095 bool write = vmf->flags & FAULT_FLAG_WRITE; 1096 1097 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1098 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1099 goto unlock; 1100 1101 entry = pud_mkyoung(orig_pud); 1102 if (write) 1103 entry = pud_mkdirty(entry); 1104 haddr = vmf->address & HPAGE_PUD_MASK; 1105 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1106 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1107 1108 unlock: 1109 spin_unlock(vmf->ptl); 1110 } 1111 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1112 1113 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1114 { 1115 pmd_t entry; 1116 unsigned long haddr; 1117 bool write = vmf->flags & FAULT_FLAG_WRITE; 1118 1119 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1120 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1121 goto unlock; 1122 1123 entry = pmd_mkyoung(orig_pmd); 1124 if (write) 1125 entry = pmd_mkdirty(entry); 1126 haddr = vmf->address & HPAGE_PMD_MASK; 1127 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1128 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1129 1130 unlock: 1131 spin_unlock(vmf->ptl); 1132 } 1133 1134 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd, 1135 struct page *page) 1136 { 1137 struct vm_area_struct *vma = vmf->vma; 1138 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1139 struct mem_cgroup *memcg; 1140 pgtable_t pgtable; 1141 pmd_t _pmd; 1142 int ret = 0, i; 1143 struct page **pages; 1144 unsigned long mmun_start; /* For mmu_notifiers */ 1145 unsigned long mmun_end; /* For mmu_notifiers */ 1146 1147 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 1148 GFP_KERNEL); 1149 if (unlikely(!pages)) { 1150 ret |= VM_FAULT_OOM; 1151 goto out; 1152 } 1153 1154 for (i = 0; i < HPAGE_PMD_NR; i++) { 1155 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 1156 vmf->address, page_to_nid(page)); 1157 if (unlikely(!pages[i] || 1158 mem_cgroup_try_charge(pages[i], vma->vm_mm, 1159 GFP_KERNEL, &memcg, false))) { 1160 if (pages[i]) 1161 put_page(pages[i]); 1162 while (--i >= 0) { 1163 memcg = (void *)page_private(pages[i]); 1164 set_page_private(pages[i], 0); 1165 mem_cgroup_cancel_charge(pages[i], memcg, 1166 false); 1167 put_page(pages[i]); 1168 } 1169 kfree(pages); 1170 ret |= VM_FAULT_OOM; 1171 goto out; 1172 } 1173 set_page_private(pages[i], (unsigned long)memcg); 1174 } 1175 1176 for (i = 0; i < HPAGE_PMD_NR; i++) { 1177 copy_user_highpage(pages[i], page + i, 1178 haddr + PAGE_SIZE * i, vma); 1179 __SetPageUptodate(pages[i]); 1180 cond_resched(); 1181 } 1182 1183 mmun_start = haddr; 1184 mmun_end = haddr + HPAGE_PMD_SIZE; 1185 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1186 1187 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1188 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1189 goto out_free_pages; 1190 VM_BUG_ON_PAGE(!PageHead(page), page); 1191 1192 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1193 /* leave pmd empty until pte is filled */ 1194 1195 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1196 pmd_populate(vma->vm_mm, &_pmd, pgtable); 1197 1198 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1199 pte_t entry; 1200 entry = mk_pte(pages[i], vma->vm_page_prot); 1201 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1202 memcg = (void *)page_private(pages[i]); 1203 set_page_private(pages[i], 0); 1204 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1205 mem_cgroup_commit_charge(pages[i], memcg, false, false); 1206 lru_cache_add_active_or_unevictable(pages[i], vma); 1207 vmf->pte = pte_offset_map(&_pmd, haddr); 1208 VM_BUG_ON(!pte_none(*vmf->pte)); 1209 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 1210 pte_unmap(vmf->pte); 1211 } 1212 kfree(pages); 1213 1214 smp_wmb(); /* make pte visible before pmd */ 1215 pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1216 page_remove_rmap(page, true); 1217 spin_unlock(vmf->ptl); 1218 1219 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1220 1221 ret |= VM_FAULT_WRITE; 1222 put_page(page); 1223 1224 out: 1225 return ret; 1226 1227 out_free_pages: 1228 spin_unlock(vmf->ptl); 1229 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1230 for (i = 0; i < HPAGE_PMD_NR; i++) { 1231 memcg = (void *)page_private(pages[i]); 1232 set_page_private(pages[i], 0); 1233 mem_cgroup_cancel_charge(pages[i], memcg, false); 1234 put_page(pages[i]); 1235 } 1236 kfree(pages); 1237 goto out; 1238 } 1239 1240 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1241 { 1242 struct vm_area_struct *vma = vmf->vma; 1243 struct page *page = NULL, *new_page; 1244 struct mem_cgroup *memcg; 1245 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1246 unsigned long mmun_start; /* For mmu_notifiers */ 1247 unsigned long mmun_end; /* For mmu_notifiers */ 1248 gfp_t huge_gfp; /* for allocation and charge */ 1249 int ret = 0; 1250 1251 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1252 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1253 if (is_huge_zero_pmd(orig_pmd)) 1254 goto alloc; 1255 spin_lock(vmf->ptl); 1256 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1257 goto out_unlock; 1258 1259 page = pmd_page(orig_pmd); 1260 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1261 /* 1262 * We can only reuse the page if nobody else maps the huge page or it's 1263 * part. 1264 */ 1265 if (!trylock_page(page)) { 1266 get_page(page); 1267 spin_unlock(vmf->ptl); 1268 lock_page(page); 1269 spin_lock(vmf->ptl); 1270 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1271 unlock_page(page); 1272 put_page(page); 1273 goto out_unlock; 1274 } 1275 put_page(page); 1276 } 1277 if (reuse_swap_page(page, NULL)) { 1278 pmd_t entry; 1279 entry = pmd_mkyoung(orig_pmd); 1280 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1281 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1282 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1283 ret |= VM_FAULT_WRITE; 1284 unlock_page(page); 1285 goto out_unlock; 1286 } 1287 unlock_page(page); 1288 get_page(page); 1289 spin_unlock(vmf->ptl); 1290 alloc: 1291 if (transparent_hugepage_enabled(vma) && 1292 !transparent_hugepage_debug_cow()) { 1293 huge_gfp = alloc_hugepage_direct_gfpmask(vma); 1294 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1295 } else 1296 new_page = NULL; 1297 1298 if (likely(new_page)) { 1299 prep_transhuge_page(new_page); 1300 } else { 1301 if (!page) { 1302 split_huge_pmd(vma, vmf->pmd, vmf->address); 1303 ret |= VM_FAULT_FALLBACK; 1304 } else { 1305 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 1306 if (ret & VM_FAULT_OOM) { 1307 split_huge_pmd(vma, vmf->pmd, vmf->address); 1308 ret |= VM_FAULT_FALLBACK; 1309 } 1310 put_page(page); 1311 } 1312 count_vm_event(THP_FAULT_FALLBACK); 1313 goto out; 1314 } 1315 1316 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, 1317 huge_gfp, &memcg, true))) { 1318 put_page(new_page); 1319 split_huge_pmd(vma, vmf->pmd, vmf->address); 1320 if (page) 1321 put_page(page); 1322 ret |= VM_FAULT_FALLBACK; 1323 count_vm_event(THP_FAULT_FALLBACK); 1324 goto out; 1325 } 1326 1327 count_vm_event(THP_FAULT_ALLOC); 1328 1329 if (!page) 1330 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 1331 else 1332 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 1333 __SetPageUptodate(new_page); 1334 1335 mmun_start = haddr; 1336 mmun_end = haddr + HPAGE_PMD_SIZE; 1337 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1338 1339 spin_lock(vmf->ptl); 1340 if (page) 1341 put_page(page); 1342 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1343 spin_unlock(vmf->ptl); 1344 mem_cgroup_cancel_charge(new_page, memcg, true); 1345 put_page(new_page); 1346 goto out_mn; 1347 } else { 1348 pmd_t entry; 1349 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1350 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1351 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1352 page_add_new_anon_rmap(new_page, vma, haddr, true); 1353 mem_cgroup_commit_charge(new_page, memcg, false, true); 1354 lru_cache_add_active_or_unevictable(new_page, vma); 1355 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 1356 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1357 if (!page) { 1358 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1359 } else { 1360 VM_BUG_ON_PAGE(!PageHead(page), page); 1361 page_remove_rmap(page, true); 1362 put_page(page); 1363 } 1364 ret |= VM_FAULT_WRITE; 1365 } 1366 spin_unlock(vmf->ptl); 1367 out_mn: 1368 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1369 out: 1370 return ret; 1371 out_unlock: 1372 spin_unlock(vmf->ptl); 1373 return ret; 1374 } 1375 1376 /* 1377 * FOLL_FORCE can write to even unwritable pmd's, but only 1378 * after we've gone through a COW cycle and they are dirty. 1379 */ 1380 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1381 { 1382 return pmd_write(pmd) || 1383 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 1384 } 1385 1386 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1387 unsigned long addr, 1388 pmd_t *pmd, 1389 unsigned int flags) 1390 { 1391 struct mm_struct *mm = vma->vm_mm; 1392 struct page *page = NULL; 1393 1394 assert_spin_locked(pmd_lockptr(mm, pmd)); 1395 1396 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1397 goto out; 1398 1399 /* Avoid dumping huge zero page */ 1400 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1401 return ERR_PTR(-EFAULT); 1402 1403 /* Full NUMA hinting faults to serialise migration in fault paths */ 1404 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1405 goto out; 1406 1407 page = pmd_page(*pmd); 1408 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1409 if (flags & FOLL_TOUCH) 1410 touch_pmd(vma, addr, pmd); 1411 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1412 /* 1413 * We don't mlock() pte-mapped THPs. This way we can avoid 1414 * leaking mlocked pages into non-VM_LOCKED VMAs. 1415 * 1416 * For anon THP: 1417 * 1418 * In most cases the pmd is the only mapping of the page as we 1419 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1420 * writable private mappings in populate_vma_page_range(). 1421 * 1422 * The only scenario when we have the page shared here is if we 1423 * mlocking read-only mapping shared over fork(). We skip 1424 * mlocking such pages. 1425 * 1426 * For file THP: 1427 * 1428 * We can expect PageDoubleMap() to be stable under page lock: 1429 * for file pages we set it in page_add_file_rmap(), which 1430 * requires page to be locked. 1431 */ 1432 1433 if (PageAnon(page) && compound_mapcount(page) != 1) 1434 goto skip_mlock; 1435 if (PageDoubleMap(page) || !page->mapping) 1436 goto skip_mlock; 1437 if (!trylock_page(page)) 1438 goto skip_mlock; 1439 lru_add_drain(); 1440 if (page->mapping && !PageDoubleMap(page)) 1441 mlock_vma_page(page); 1442 unlock_page(page); 1443 } 1444 skip_mlock: 1445 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1446 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1447 if (flags & FOLL_GET) 1448 get_page(page); 1449 1450 out: 1451 return page; 1452 } 1453 1454 /* NUMA hinting page fault entry point for trans huge pmds */ 1455 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1456 { 1457 struct vm_area_struct *vma = vmf->vma; 1458 struct anon_vma *anon_vma = NULL; 1459 struct page *page; 1460 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1461 int page_nid = -1, this_nid = numa_node_id(); 1462 int target_nid, last_cpupid = -1; 1463 bool page_locked; 1464 bool migrated = false; 1465 bool was_writable; 1466 int flags = 0; 1467 1468 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1469 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1470 goto out_unlock; 1471 1472 /* 1473 * If there are potential migrations, wait for completion and retry 1474 * without disrupting NUMA hinting information. Do not relock and 1475 * check_same as the page may no longer be mapped. 1476 */ 1477 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1478 page = pmd_page(*vmf->pmd); 1479 if (!get_page_unless_zero(page)) 1480 goto out_unlock; 1481 spin_unlock(vmf->ptl); 1482 wait_on_page_locked(page); 1483 put_page(page); 1484 goto out; 1485 } 1486 1487 page = pmd_page(pmd); 1488 BUG_ON(is_huge_zero_page(page)); 1489 page_nid = page_to_nid(page); 1490 last_cpupid = page_cpupid_last(page); 1491 count_vm_numa_event(NUMA_HINT_FAULTS); 1492 if (page_nid == this_nid) { 1493 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1494 flags |= TNF_FAULT_LOCAL; 1495 } 1496 1497 /* See similar comment in do_numa_page for explanation */ 1498 if (!pmd_savedwrite(pmd)) 1499 flags |= TNF_NO_GROUP; 1500 1501 /* 1502 * Acquire the page lock to serialise THP migrations but avoid dropping 1503 * page_table_lock if at all possible 1504 */ 1505 page_locked = trylock_page(page); 1506 target_nid = mpol_misplaced(page, vma, haddr); 1507 if (target_nid == -1) { 1508 /* If the page was locked, there are no parallel migrations */ 1509 if (page_locked) 1510 goto clear_pmdnuma; 1511 } 1512 1513 /* Migration could have started since the pmd_trans_migrating check */ 1514 if (!page_locked) { 1515 page_nid = -1; 1516 if (!get_page_unless_zero(page)) 1517 goto out_unlock; 1518 spin_unlock(vmf->ptl); 1519 wait_on_page_locked(page); 1520 put_page(page); 1521 goto out; 1522 } 1523 1524 /* 1525 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1526 * to serialises splits 1527 */ 1528 get_page(page); 1529 spin_unlock(vmf->ptl); 1530 anon_vma = page_lock_anon_vma_read(page); 1531 1532 /* Confirm the PMD did not change while page_table_lock was released */ 1533 spin_lock(vmf->ptl); 1534 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1535 unlock_page(page); 1536 put_page(page); 1537 page_nid = -1; 1538 goto out_unlock; 1539 } 1540 1541 /* Bail if we fail to protect against THP splits for any reason */ 1542 if (unlikely(!anon_vma)) { 1543 put_page(page); 1544 page_nid = -1; 1545 goto clear_pmdnuma; 1546 } 1547 1548 /* 1549 * Since we took the NUMA fault, we must have observed the !accessible 1550 * bit. Make sure all other CPUs agree with that, to avoid them 1551 * modifying the page we're about to migrate. 1552 * 1553 * Must be done under PTL such that we'll observe the relevant 1554 * inc_tlb_flush_pending(). 1555 * 1556 * We are not sure a pending tlb flush here is for a huge page 1557 * mapping or not. Hence use the tlb range variant 1558 */ 1559 if (mm_tlb_flush_pending(vma->vm_mm)) 1560 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1561 1562 /* 1563 * Migrate the THP to the requested node, returns with page unlocked 1564 * and access rights restored. 1565 */ 1566 spin_unlock(vmf->ptl); 1567 1568 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1569 vmf->pmd, pmd, vmf->address, page, target_nid); 1570 if (migrated) { 1571 flags |= TNF_MIGRATED; 1572 page_nid = target_nid; 1573 } else 1574 flags |= TNF_MIGRATE_FAIL; 1575 1576 goto out; 1577 clear_pmdnuma: 1578 BUG_ON(!PageLocked(page)); 1579 was_writable = pmd_savedwrite(pmd); 1580 pmd = pmd_modify(pmd, vma->vm_page_prot); 1581 pmd = pmd_mkyoung(pmd); 1582 if (was_writable) 1583 pmd = pmd_mkwrite(pmd); 1584 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1585 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1586 unlock_page(page); 1587 out_unlock: 1588 spin_unlock(vmf->ptl); 1589 1590 out: 1591 if (anon_vma) 1592 page_unlock_anon_vma_read(anon_vma); 1593 1594 if (page_nid != -1) 1595 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1596 flags); 1597 1598 return 0; 1599 } 1600 1601 /* 1602 * Return true if we do MADV_FREE successfully on entire pmd page. 1603 * Otherwise, return false. 1604 */ 1605 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1606 pmd_t *pmd, unsigned long addr, unsigned long next) 1607 { 1608 spinlock_t *ptl; 1609 pmd_t orig_pmd; 1610 struct page *page; 1611 struct mm_struct *mm = tlb->mm; 1612 bool ret = false; 1613 1614 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1615 1616 ptl = pmd_trans_huge_lock(pmd, vma); 1617 if (!ptl) 1618 goto out_unlocked; 1619 1620 orig_pmd = *pmd; 1621 if (is_huge_zero_pmd(orig_pmd)) 1622 goto out; 1623 1624 if (unlikely(!pmd_present(orig_pmd))) { 1625 VM_BUG_ON(thp_migration_supported() && 1626 !is_pmd_migration_entry(orig_pmd)); 1627 goto out; 1628 } 1629 1630 page = pmd_page(orig_pmd); 1631 /* 1632 * If other processes are mapping this page, we couldn't discard 1633 * the page unless they all do MADV_FREE so let's skip the page. 1634 */ 1635 if (page_mapcount(page) != 1) 1636 goto out; 1637 1638 if (!trylock_page(page)) 1639 goto out; 1640 1641 /* 1642 * If user want to discard part-pages of THP, split it so MADV_FREE 1643 * will deactivate only them. 1644 */ 1645 if (next - addr != HPAGE_PMD_SIZE) { 1646 get_page(page); 1647 spin_unlock(ptl); 1648 split_huge_page(page); 1649 unlock_page(page); 1650 put_page(page); 1651 goto out_unlocked; 1652 } 1653 1654 if (PageDirty(page)) 1655 ClearPageDirty(page); 1656 unlock_page(page); 1657 1658 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1659 pmdp_invalidate(vma, addr, pmd); 1660 orig_pmd = pmd_mkold(orig_pmd); 1661 orig_pmd = pmd_mkclean(orig_pmd); 1662 1663 set_pmd_at(mm, addr, pmd, orig_pmd); 1664 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1665 } 1666 1667 mark_page_lazyfree(page); 1668 ret = true; 1669 out: 1670 spin_unlock(ptl); 1671 out_unlocked: 1672 return ret; 1673 } 1674 1675 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1676 { 1677 pgtable_t pgtable; 1678 1679 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1680 pte_free(mm, pgtable); 1681 atomic_long_dec(&mm->nr_ptes); 1682 } 1683 1684 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1685 pmd_t *pmd, unsigned long addr) 1686 { 1687 pmd_t orig_pmd; 1688 spinlock_t *ptl; 1689 1690 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1691 1692 ptl = __pmd_trans_huge_lock(pmd, vma); 1693 if (!ptl) 1694 return 0; 1695 /* 1696 * For architectures like ppc64 we look at deposited pgtable 1697 * when calling pmdp_huge_get_and_clear. So do the 1698 * pgtable_trans_huge_withdraw after finishing pmdp related 1699 * operations. 1700 */ 1701 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1702 tlb->fullmm); 1703 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1704 if (vma_is_dax(vma)) { 1705 if (arch_needs_pgtable_deposit()) 1706 zap_deposited_table(tlb->mm, pmd); 1707 spin_unlock(ptl); 1708 if (is_huge_zero_pmd(orig_pmd)) 1709 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1710 } else if (is_huge_zero_pmd(orig_pmd)) { 1711 zap_deposited_table(tlb->mm, pmd); 1712 spin_unlock(ptl); 1713 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1714 } else { 1715 struct page *page = NULL; 1716 int flush_needed = 1; 1717 1718 if (pmd_present(orig_pmd)) { 1719 page = pmd_page(orig_pmd); 1720 page_remove_rmap(page, true); 1721 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1722 VM_BUG_ON_PAGE(!PageHead(page), page); 1723 } else if (thp_migration_supported()) { 1724 swp_entry_t entry; 1725 1726 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1727 entry = pmd_to_swp_entry(orig_pmd); 1728 page = pfn_to_page(swp_offset(entry)); 1729 flush_needed = 0; 1730 } else 1731 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1732 1733 if (PageAnon(page)) { 1734 zap_deposited_table(tlb->mm, pmd); 1735 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1736 } else { 1737 if (arch_needs_pgtable_deposit()) 1738 zap_deposited_table(tlb->mm, pmd); 1739 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); 1740 } 1741 1742 spin_unlock(ptl); 1743 if (flush_needed) 1744 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1745 } 1746 return 1; 1747 } 1748 1749 #ifndef pmd_move_must_withdraw 1750 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1751 spinlock_t *old_pmd_ptl, 1752 struct vm_area_struct *vma) 1753 { 1754 /* 1755 * With split pmd lock we also need to move preallocated 1756 * PTE page table if new_pmd is on different PMD page table. 1757 * 1758 * We also don't deposit and withdraw tables for file pages. 1759 */ 1760 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1761 } 1762 #endif 1763 1764 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1765 { 1766 #ifdef CONFIG_MEM_SOFT_DIRTY 1767 if (unlikely(is_pmd_migration_entry(pmd))) 1768 pmd = pmd_swp_mksoft_dirty(pmd); 1769 else if (pmd_present(pmd)) 1770 pmd = pmd_mksoft_dirty(pmd); 1771 #endif 1772 return pmd; 1773 } 1774 1775 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1776 unsigned long new_addr, unsigned long old_end, 1777 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush) 1778 { 1779 spinlock_t *old_ptl, *new_ptl; 1780 pmd_t pmd; 1781 struct mm_struct *mm = vma->vm_mm; 1782 bool force_flush = false; 1783 1784 if ((old_addr & ~HPAGE_PMD_MASK) || 1785 (new_addr & ~HPAGE_PMD_MASK) || 1786 old_end - old_addr < HPAGE_PMD_SIZE) 1787 return false; 1788 1789 /* 1790 * The destination pmd shouldn't be established, free_pgtables() 1791 * should have release it. 1792 */ 1793 if (WARN_ON(!pmd_none(*new_pmd))) { 1794 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1795 return false; 1796 } 1797 1798 /* 1799 * We don't have to worry about the ordering of src and dst 1800 * ptlocks because exclusive mmap_sem prevents deadlock. 1801 */ 1802 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1803 if (old_ptl) { 1804 new_ptl = pmd_lockptr(mm, new_pmd); 1805 if (new_ptl != old_ptl) 1806 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1807 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1808 if (pmd_present(pmd) && pmd_dirty(pmd)) 1809 force_flush = true; 1810 VM_BUG_ON(!pmd_none(*new_pmd)); 1811 1812 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1813 pgtable_t pgtable; 1814 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1815 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1816 } 1817 pmd = move_soft_dirty_pmd(pmd); 1818 set_pmd_at(mm, new_addr, new_pmd, pmd); 1819 if (new_ptl != old_ptl) 1820 spin_unlock(new_ptl); 1821 if (force_flush) 1822 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1823 else 1824 *need_flush = true; 1825 spin_unlock(old_ptl); 1826 return true; 1827 } 1828 return false; 1829 } 1830 1831 /* 1832 * Returns 1833 * - 0 if PMD could not be locked 1834 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1835 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1836 */ 1837 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1838 unsigned long addr, pgprot_t newprot, int prot_numa) 1839 { 1840 struct mm_struct *mm = vma->vm_mm; 1841 spinlock_t *ptl; 1842 pmd_t entry; 1843 bool preserve_write; 1844 int ret; 1845 1846 ptl = __pmd_trans_huge_lock(pmd, vma); 1847 if (!ptl) 1848 return 0; 1849 1850 preserve_write = prot_numa && pmd_write(*pmd); 1851 ret = 1; 1852 1853 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1854 if (is_swap_pmd(*pmd)) { 1855 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1856 1857 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1858 if (is_write_migration_entry(entry)) { 1859 pmd_t newpmd; 1860 /* 1861 * A protection check is difficult so 1862 * just be safe and disable write 1863 */ 1864 make_migration_entry_read(&entry); 1865 newpmd = swp_entry_to_pmd(entry); 1866 if (pmd_swp_soft_dirty(*pmd)) 1867 newpmd = pmd_swp_mksoft_dirty(newpmd); 1868 set_pmd_at(mm, addr, pmd, newpmd); 1869 } 1870 goto unlock; 1871 } 1872 #endif 1873 1874 /* 1875 * Avoid trapping faults against the zero page. The read-only 1876 * data is likely to be read-cached on the local CPU and 1877 * local/remote hits to the zero page are not interesting. 1878 */ 1879 if (prot_numa && is_huge_zero_pmd(*pmd)) 1880 goto unlock; 1881 1882 if (prot_numa && pmd_protnone(*pmd)) 1883 goto unlock; 1884 1885 /* 1886 * In case prot_numa, we are under down_read(mmap_sem). It's critical 1887 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1888 * which is also under down_read(mmap_sem): 1889 * 1890 * CPU0: CPU1: 1891 * change_huge_pmd(prot_numa=1) 1892 * pmdp_huge_get_and_clear_notify() 1893 * madvise_dontneed() 1894 * zap_pmd_range() 1895 * pmd_trans_huge(*pmd) == 0 (without ptl) 1896 * // skip the pmd 1897 * set_pmd_at(); 1898 * // pmd is re-established 1899 * 1900 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1901 * which may break userspace. 1902 * 1903 * pmdp_invalidate() is required to make sure we don't miss 1904 * dirty/young flags set by hardware. 1905 */ 1906 entry = *pmd; 1907 pmdp_invalidate(vma, addr, pmd); 1908 1909 /* 1910 * Recover dirty/young flags. It relies on pmdp_invalidate to not 1911 * corrupt them. 1912 */ 1913 if (pmd_dirty(*pmd)) 1914 entry = pmd_mkdirty(entry); 1915 if (pmd_young(*pmd)) 1916 entry = pmd_mkyoung(entry); 1917 1918 entry = pmd_modify(entry, newprot); 1919 if (preserve_write) 1920 entry = pmd_mk_savedwrite(entry); 1921 ret = HPAGE_PMD_NR; 1922 set_pmd_at(mm, addr, pmd, entry); 1923 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1924 unlock: 1925 spin_unlock(ptl); 1926 return ret; 1927 } 1928 1929 /* 1930 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1931 * 1932 * Note that if it returns page table lock pointer, this routine returns without 1933 * unlocking page table lock. So callers must unlock it. 1934 */ 1935 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1936 { 1937 spinlock_t *ptl; 1938 ptl = pmd_lock(vma->vm_mm, pmd); 1939 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1940 pmd_devmap(*pmd))) 1941 return ptl; 1942 spin_unlock(ptl); 1943 return NULL; 1944 } 1945 1946 /* 1947 * Returns true if a given pud maps a thp, false otherwise. 1948 * 1949 * Note that if it returns true, this routine returns without unlocking page 1950 * table lock. So callers must unlock it. 1951 */ 1952 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1953 { 1954 spinlock_t *ptl; 1955 1956 ptl = pud_lock(vma->vm_mm, pud); 1957 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1958 return ptl; 1959 spin_unlock(ptl); 1960 return NULL; 1961 } 1962 1963 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1964 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1965 pud_t *pud, unsigned long addr) 1966 { 1967 pud_t orig_pud; 1968 spinlock_t *ptl; 1969 1970 ptl = __pud_trans_huge_lock(pud, vma); 1971 if (!ptl) 1972 return 0; 1973 /* 1974 * For architectures like ppc64 we look at deposited pgtable 1975 * when calling pudp_huge_get_and_clear. So do the 1976 * pgtable_trans_huge_withdraw after finishing pudp related 1977 * operations. 1978 */ 1979 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 1980 tlb->fullmm); 1981 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1982 if (vma_is_dax(vma)) { 1983 spin_unlock(ptl); 1984 /* No zero page support yet */ 1985 } else { 1986 /* No support for anonymous PUD pages yet */ 1987 BUG(); 1988 } 1989 return 1; 1990 } 1991 1992 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 1993 unsigned long haddr) 1994 { 1995 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 1996 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 1997 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 1998 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 1999 2000 count_vm_event(THP_SPLIT_PUD); 2001 2002 pudp_huge_clear_flush_notify(vma, haddr, pud); 2003 } 2004 2005 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2006 unsigned long address) 2007 { 2008 spinlock_t *ptl; 2009 struct mm_struct *mm = vma->vm_mm; 2010 unsigned long haddr = address & HPAGE_PUD_MASK; 2011 2012 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); 2013 ptl = pud_lock(mm, pud); 2014 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2015 goto out; 2016 __split_huge_pud_locked(vma, pud, haddr); 2017 2018 out: 2019 spin_unlock(ptl); 2020 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE); 2021 } 2022 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2023 2024 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2025 unsigned long haddr, pmd_t *pmd) 2026 { 2027 struct mm_struct *mm = vma->vm_mm; 2028 pgtable_t pgtable; 2029 pmd_t _pmd; 2030 int i; 2031 2032 /* leave pmd empty until pte is filled */ 2033 pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2034 2035 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2036 pmd_populate(mm, &_pmd, pgtable); 2037 2038 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2039 pte_t *pte, entry; 2040 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2041 entry = pte_mkspecial(entry); 2042 pte = pte_offset_map(&_pmd, haddr); 2043 VM_BUG_ON(!pte_none(*pte)); 2044 set_pte_at(mm, haddr, pte, entry); 2045 pte_unmap(pte); 2046 } 2047 smp_wmb(); /* make pte visible before pmd */ 2048 pmd_populate(mm, pmd, pgtable); 2049 } 2050 2051 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2052 unsigned long haddr, bool freeze) 2053 { 2054 struct mm_struct *mm = vma->vm_mm; 2055 struct page *page; 2056 pgtable_t pgtable; 2057 pmd_t _pmd; 2058 bool young, write, dirty, soft_dirty, pmd_migration = false; 2059 unsigned long addr; 2060 int i; 2061 2062 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2063 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2064 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2065 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2066 && !pmd_devmap(*pmd)); 2067 2068 count_vm_event(THP_SPLIT_PMD); 2069 2070 if (!vma_is_anonymous(vma)) { 2071 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2072 /* 2073 * We are going to unmap this huge page. So 2074 * just go ahead and zap it 2075 */ 2076 if (arch_needs_pgtable_deposit()) 2077 zap_deposited_table(mm, pmd); 2078 if (vma_is_dax(vma)) 2079 return; 2080 page = pmd_page(_pmd); 2081 if (!PageReferenced(page) && pmd_young(_pmd)) 2082 SetPageReferenced(page); 2083 page_remove_rmap(page, true); 2084 put_page(page); 2085 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); 2086 return; 2087 } else if (is_huge_zero_pmd(*pmd)) { 2088 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2089 } 2090 2091 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2092 pmd_migration = is_pmd_migration_entry(*pmd); 2093 if (pmd_migration) { 2094 swp_entry_t entry; 2095 2096 entry = pmd_to_swp_entry(*pmd); 2097 page = pfn_to_page(swp_offset(entry)); 2098 } else 2099 #endif 2100 page = pmd_page(*pmd); 2101 VM_BUG_ON_PAGE(!page_count(page), page); 2102 page_ref_add(page, HPAGE_PMD_NR - 1); 2103 write = pmd_write(*pmd); 2104 young = pmd_young(*pmd); 2105 dirty = pmd_dirty(*pmd); 2106 soft_dirty = pmd_soft_dirty(*pmd); 2107 2108 pmdp_huge_split_prepare(vma, haddr, pmd); 2109 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2110 pmd_populate(mm, &_pmd, pgtable); 2111 2112 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2113 pte_t entry, *pte; 2114 /* 2115 * Note that NUMA hinting access restrictions are not 2116 * transferred to avoid any possibility of altering 2117 * permissions across VMAs. 2118 */ 2119 if (freeze || pmd_migration) { 2120 swp_entry_t swp_entry; 2121 swp_entry = make_migration_entry(page + i, write); 2122 entry = swp_entry_to_pte(swp_entry); 2123 if (soft_dirty) 2124 entry = pte_swp_mksoft_dirty(entry); 2125 } else { 2126 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2127 entry = maybe_mkwrite(entry, vma); 2128 if (!write) 2129 entry = pte_wrprotect(entry); 2130 if (!young) 2131 entry = pte_mkold(entry); 2132 if (soft_dirty) 2133 entry = pte_mksoft_dirty(entry); 2134 } 2135 if (dirty) 2136 SetPageDirty(page + i); 2137 pte = pte_offset_map(&_pmd, addr); 2138 BUG_ON(!pte_none(*pte)); 2139 set_pte_at(mm, addr, pte, entry); 2140 atomic_inc(&page[i]._mapcount); 2141 pte_unmap(pte); 2142 } 2143 2144 /* 2145 * Set PG_double_map before dropping compound_mapcount to avoid 2146 * false-negative page_mapped(). 2147 */ 2148 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2149 for (i = 0; i < HPAGE_PMD_NR; i++) 2150 atomic_inc(&page[i]._mapcount); 2151 } 2152 2153 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2154 /* Last compound_mapcount is gone. */ 2155 __dec_node_page_state(page, NR_ANON_THPS); 2156 if (TestClearPageDoubleMap(page)) { 2157 /* No need in mapcount reference anymore */ 2158 for (i = 0; i < HPAGE_PMD_NR; i++) 2159 atomic_dec(&page[i]._mapcount); 2160 } 2161 } 2162 2163 smp_wmb(); /* make pte visible before pmd */ 2164 /* 2165 * Up to this point the pmd is present and huge and userland has the 2166 * whole access to the hugepage during the split (which happens in 2167 * place). If we overwrite the pmd with the not-huge version pointing 2168 * to the pte here (which of course we could if all CPUs were bug 2169 * free), userland could trigger a small page size TLB miss on the 2170 * small sized TLB while the hugepage TLB entry is still established in 2171 * the huge TLB. Some CPU doesn't like that. 2172 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2173 * 383 on page 93. Intel should be safe but is also warns that it's 2174 * only safe if the permission and cache attributes of the two entries 2175 * loaded in the two TLB is identical (which should be the case here). 2176 * But it is generally safer to never allow small and huge TLB entries 2177 * for the same virtual address to be loaded simultaneously. So instead 2178 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2179 * current pmd notpresent (atomically because here the pmd_trans_huge 2180 * and pmd_trans_splitting must remain set at all times on the pmd 2181 * until the split is complete for this pmd), then we flush the SMP TLB 2182 * and finally we write the non-huge version of the pmd entry with 2183 * pmd_populate. 2184 */ 2185 pmdp_invalidate(vma, haddr, pmd); 2186 pmd_populate(mm, pmd, pgtable); 2187 2188 if (freeze) { 2189 for (i = 0; i < HPAGE_PMD_NR; i++) { 2190 page_remove_rmap(page + i, false); 2191 put_page(page + i); 2192 } 2193 } 2194 } 2195 2196 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2197 unsigned long address, bool freeze, struct page *page) 2198 { 2199 spinlock_t *ptl; 2200 struct mm_struct *mm = vma->vm_mm; 2201 unsigned long haddr = address & HPAGE_PMD_MASK; 2202 2203 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 2204 ptl = pmd_lock(mm, pmd); 2205 2206 /* 2207 * If caller asks to setup a migration entries, we need a page to check 2208 * pmd against. Otherwise we can end up replacing wrong page. 2209 */ 2210 VM_BUG_ON(freeze && !page); 2211 if (page && page != pmd_page(*pmd)) 2212 goto out; 2213 2214 if (pmd_trans_huge(*pmd)) { 2215 page = pmd_page(*pmd); 2216 if (PageMlocked(page)) 2217 clear_page_mlock(page); 2218 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2219 goto out; 2220 __split_huge_pmd_locked(vma, pmd, haddr, freeze); 2221 out: 2222 spin_unlock(ptl); 2223 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); 2224 } 2225 2226 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2227 bool freeze, struct page *page) 2228 { 2229 pgd_t *pgd; 2230 p4d_t *p4d; 2231 pud_t *pud; 2232 pmd_t *pmd; 2233 2234 pgd = pgd_offset(vma->vm_mm, address); 2235 if (!pgd_present(*pgd)) 2236 return; 2237 2238 p4d = p4d_offset(pgd, address); 2239 if (!p4d_present(*p4d)) 2240 return; 2241 2242 pud = pud_offset(p4d, address); 2243 if (!pud_present(*pud)) 2244 return; 2245 2246 pmd = pmd_offset(pud, address); 2247 2248 __split_huge_pmd(vma, pmd, address, freeze, page); 2249 } 2250 2251 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2252 unsigned long start, 2253 unsigned long end, 2254 long adjust_next) 2255 { 2256 /* 2257 * If the new start address isn't hpage aligned and it could 2258 * previously contain an hugepage: check if we need to split 2259 * an huge pmd. 2260 */ 2261 if (start & ~HPAGE_PMD_MASK && 2262 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2263 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2264 split_huge_pmd_address(vma, start, false, NULL); 2265 2266 /* 2267 * If the new end address isn't hpage aligned and it could 2268 * previously contain an hugepage: check if we need to split 2269 * an huge pmd. 2270 */ 2271 if (end & ~HPAGE_PMD_MASK && 2272 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2273 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2274 split_huge_pmd_address(vma, end, false, NULL); 2275 2276 /* 2277 * If we're also updating the vma->vm_next->vm_start, if the new 2278 * vm_next->vm_start isn't page aligned and it could previously 2279 * contain an hugepage: check if we need to split an huge pmd. 2280 */ 2281 if (adjust_next > 0) { 2282 struct vm_area_struct *next = vma->vm_next; 2283 unsigned long nstart = next->vm_start; 2284 nstart += adjust_next << PAGE_SHIFT; 2285 if (nstart & ~HPAGE_PMD_MASK && 2286 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2287 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2288 split_huge_pmd_address(next, nstart, false, NULL); 2289 } 2290 } 2291 2292 static void freeze_page(struct page *page) 2293 { 2294 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2295 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2296 bool unmap_success; 2297 2298 VM_BUG_ON_PAGE(!PageHead(page), page); 2299 2300 if (PageAnon(page)) 2301 ttu_flags |= TTU_SPLIT_FREEZE; 2302 2303 unmap_success = try_to_unmap(page, ttu_flags); 2304 VM_BUG_ON_PAGE(!unmap_success, page); 2305 } 2306 2307 static void unfreeze_page(struct page *page) 2308 { 2309 int i; 2310 if (PageTransHuge(page)) { 2311 remove_migration_ptes(page, page, true); 2312 } else { 2313 for (i = 0; i < HPAGE_PMD_NR; i++) 2314 remove_migration_ptes(page + i, page + i, true); 2315 } 2316 } 2317 2318 static void __split_huge_page_tail(struct page *head, int tail, 2319 struct lruvec *lruvec, struct list_head *list) 2320 { 2321 struct page *page_tail = head + tail; 2322 2323 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2324 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); 2325 2326 /* 2327 * tail_page->_refcount is zero and not changing from under us. But 2328 * get_page_unless_zero() may be running from under us on the 2329 * tail_page. If we used atomic_set() below instead of atomic_inc() or 2330 * atomic_add(), we would then run atomic_set() concurrently with 2331 * get_page_unless_zero(), and atomic_set() is implemented in C not 2332 * using locked ops. spin_unlock on x86 sometime uses locked ops 2333 * because of PPro errata 66, 92, so unless somebody can guarantee 2334 * atomic_set() here would be safe on all archs (and not only on x86), 2335 * it's safer to use atomic_inc()/atomic_add(). 2336 */ 2337 if (PageAnon(head) && !PageSwapCache(head)) { 2338 page_ref_inc(page_tail); 2339 } else { 2340 /* Additional pin to radix tree */ 2341 page_ref_add(page_tail, 2); 2342 } 2343 2344 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2345 page_tail->flags |= (head->flags & 2346 ((1L << PG_referenced) | 2347 (1L << PG_swapbacked) | 2348 (1L << PG_swapcache) | 2349 (1L << PG_mlocked) | 2350 (1L << PG_uptodate) | 2351 (1L << PG_active) | 2352 (1L << PG_locked) | 2353 (1L << PG_unevictable) | 2354 (1L << PG_dirty))); 2355 2356 /* 2357 * After clearing PageTail the gup refcount can be released. 2358 * Page flags also must be visible before we make the page non-compound. 2359 */ 2360 smp_wmb(); 2361 2362 clear_compound_head(page_tail); 2363 2364 if (page_is_young(head)) 2365 set_page_young(page_tail); 2366 if (page_is_idle(head)) 2367 set_page_idle(page_tail); 2368 2369 /* ->mapping in first tail page is compound_mapcount */ 2370 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2371 page_tail); 2372 page_tail->mapping = head->mapping; 2373 2374 page_tail->index = head->index + tail; 2375 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2376 lru_add_page_tail(head, page_tail, lruvec, list); 2377 } 2378 2379 static void __split_huge_page(struct page *page, struct list_head *list, 2380 unsigned long flags) 2381 { 2382 struct page *head = compound_head(page); 2383 struct zone *zone = page_zone(head); 2384 struct lruvec *lruvec; 2385 pgoff_t end = -1; 2386 int i; 2387 2388 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2389 2390 /* complete memcg works before add pages to LRU */ 2391 mem_cgroup_split_huge_fixup(head); 2392 2393 if (!PageAnon(page)) 2394 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); 2395 2396 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2397 __split_huge_page_tail(head, i, lruvec, list); 2398 /* Some pages can be beyond i_size: drop them from page cache */ 2399 if (head[i].index >= end) { 2400 __ClearPageDirty(head + i); 2401 __delete_from_page_cache(head + i, NULL); 2402 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2403 shmem_uncharge(head->mapping->host, 1); 2404 put_page(head + i); 2405 } 2406 } 2407 2408 ClearPageCompound(head); 2409 /* See comment in __split_huge_page_tail() */ 2410 if (PageAnon(head)) { 2411 /* Additional pin to radix tree of swap cache */ 2412 if (PageSwapCache(head)) 2413 page_ref_add(head, 2); 2414 else 2415 page_ref_inc(head); 2416 } else { 2417 /* Additional pin to radix tree */ 2418 page_ref_add(head, 2); 2419 spin_unlock(&head->mapping->tree_lock); 2420 } 2421 2422 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2423 2424 unfreeze_page(head); 2425 2426 for (i = 0; i < HPAGE_PMD_NR; i++) { 2427 struct page *subpage = head + i; 2428 if (subpage == page) 2429 continue; 2430 unlock_page(subpage); 2431 2432 /* 2433 * Subpages may be freed if there wasn't any mapping 2434 * like if add_to_swap() is running on a lru page that 2435 * had its mapping zapped. And freeing these pages 2436 * requires taking the lru_lock so we do the put_page 2437 * of the tail pages after the split is complete. 2438 */ 2439 put_page(subpage); 2440 } 2441 } 2442 2443 int total_mapcount(struct page *page) 2444 { 2445 int i, compound, ret; 2446 2447 VM_BUG_ON_PAGE(PageTail(page), page); 2448 2449 if (likely(!PageCompound(page))) 2450 return atomic_read(&page->_mapcount) + 1; 2451 2452 compound = compound_mapcount(page); 2453 if (PageHuge(page)) 2454 return compound; 2455 ret = compound; 2456 for (i = 0; i < HPAGE_PMD_NR; i++) 2457 ret += atomic_read(&page[i]._mapcount) + 1; 2458 /* File pages has compound_mapcount included in _mapcount */ 2459 if (!PageAnon(page)) 2460 return ret - compound * HPAGE_PMD_NR; 2461 if (PageDoubleMap(page)) 2462 ret -= HPAGE_PMD_NR; 2463 return ret; 2464 } 2465 2466 /* 2467 * This calculates accurately how many mappings a transparent hugepage 2468 * has (unlike page_mapcount() which isn't fully accurate). This full 2469 * accuracy is primarily needed to know if copy-on-write faults can 2470 * reuse the page and change the mapping to read-write instead of 2471 * copying them. At the same time this returns the total_mapcount too. 2472 * 2473 * The function returns the highest mapcount any one of the subpages 2474 * has. If the return value is one, even if different processes are 2475 * mapping different subpages of the transparent hugepage, they can 2476 * all reuse it, because each process is reusing a different subpage. 2477 * 2478 * The total_mapcount is instead counting all virtual mappings of the 2479 * subpages. If the total_mapcount is equal to "one", it tells the 2480 * caller all mappings belong to the same "mm" and in turn the 2481 * anon_vma of the transparent hugepage can become the vma->anon_vma 2482 * local one as no other process may be mapping any of the subpages. 2483 * 2484 * It would be more accurate to replace page_mapcount() with 2485 * page_trans_huge_mapcount(), however we only use 2486 * page_trans_huge_mapcount() in the copy-on-write faults where we 2487 * need full accuracy to avoid breaking page pinning, because 2488 * page_trans_huge_mapcount() is slower than page_mapcount(). 2489 */ 2490 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2491 { 2492 int i, ret, _total_mapcount, mapcount; 2493 2494 /* hugetlbfs shouldn't call it */ 2495 VM_BUG_ON_PAGE(PageHuge(page), page); 2496 2497 if (likely(!PageTransCompound(page))) { 2498 mapcount = atomic_read(&page->_mapcount) + 1; 2499 if (total_mapcount) 2500 *total_mapcount = mapcount; 2501 return mapcount; 2502 } 2503 2504 page = compound_head(page); 2505 2506 _total_mapcount = ret = 0; 2507 for (i = 0; i < HPAGE_PMD_NR; i++) { 2508 mapcount = atomic_read(&page[i]._mapcount) + 1; 2509 ret = max(ret, mapcount); 2510 _total_mapcount += mapcount; 2511 } 2512 if (PageDoubleMap(page)) { 2513 ret -= 1; 2514 _total_mapcount -= HPAGE_PMD_NR; 2515 } 2516 mapcount = compound_mapcount(page); 2517 ret += mapcount; 2518 _total_mapcount += mapcount; 2519 if (total_mapcount) 2520 *total_mapcount = _total_mapcount; 2521 return ret; 2522 } 2523 2524 /* Racy check whether the huge page can be split */ 2525 bool can_split_huge_page(struct page *page, int *pextra_pins) 2526 { 2527 int extra_pins; 2528 2529 /* Additional pins from radix tree */ 2530 if (PageAnon(page)) 2531 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2532 else 2533 extra_pins = HPAGE_PMD_NR; 2534 if (pextra_pins) 2535 *pextra_pins = extra_pins; 2536 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2537 } 2538 2539 /* 2540 * This function splits huge page into normal pages. @page can point to any 2541 * subpage of huge page to split. Split doesn't change the position of @page. 2542 * 2543 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2544 * The huge page must be locked. 2545 * 2546 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2547 * 2548 * Both head page and tail pages will inherit mapping, flags, and so on from 2549 * the hugepage. 2550 * 2551 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2552 * they are not mapped. 2553 * 2554 * Returns 0 if the hugepage is split successfully. 2555 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2556 * us. 2557 */ 2558 int split_huge_page_to_list(struct page *page, struct list_head *list) 2559 { 2560 struct page *head = compound_head(page); 2561 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2562 struct anon_vma *anon_vma = NULL; 2563 struct address_space *mapping = NULL; 2564 int count, mapcount, extra_pins, ret; 2565 bool mlocked; 2566 unsigned long flags; 2567 2568 VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2569 VM_BUG_ON_PAGE(!PageLocked(page), page); 2570 VM_BUG_ON_PAGE(!PageCompound(page), page); 2571 2572 if (PageWriteback(page)) 2573 return -EBUSY; 2574 2575 if (PageAnon(head)) { 2576 /* 2577 * The caller does not necessarily hold an mmap_sem that would 2578 * prevent the anon_vma disappearing so we first we take a 2579 * reference to it and then lock the anon_vma for write. This 2580 * is similar to page_lock_anon_vma_read except the write lock 2581 * is taken to serialise against parallel split or collapse 2582 * operations. 2583 */ 2584 anon_vma = page_get_anon_vma(head); 2585 if (!anon_vma) { 2586 ret = -EBUSY; 2587 goto out; 2588 } 2589 mapping = NULL; 2590 anon_vma_lock_write(anon_vma); 2591 } else { 2592 mapping = head->mapping; 2593 2594 /* Truncated ? */ 2595 if (!mapping) { 2596 ret = -EBUSY; 2597 goto out; 2598 } 2599 2600 anon_vma = NULL; 2601 i_mmap_lock_read(mapping); 2602 } 2603 2604 /* 2605 * Racy check if we can split the page, before freeze_page() will 2606 * split PMDs 2607 */ 2608 if (!can_split_huge_page(head, &extra_pins)) { 2609 ret = -EBUSY; 2610 goto out_unlock; 2611 } 2612 2613 mlocked = PageMlocked(page); 2614 freeze_page(head); 2615 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2616 2617 /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2618 if (mlocked) 2619 lru_add_drain(); 2620 2621 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2622 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2623 2624 if (mapping) { 2625 void **pslot; 2626 2627 spin_lock(&mapping->tree_lock); 2628 pslot = radix_tree_lookup_slot(&mapping->page_tree, 2629 page_index(head)); 2630 /* 2631 * Check if the head page is present in radix tree. 2632 * We assume all tail are present too, if head is there. 2633 */ 2634 if (radix_tree_deref_slot_protected(pslot, 2635 &mapping->tree_lock) != head) 2636 goto fail; 2637 } 2638 2639 /* Prevent deferred_split_scan() touching ->_refcount */ 2640 spin_lock(&pgdata->split_queue_lock); 2641 count = page_count(head); 2642 mapcount = total_mapcount(head); 2643 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2644 if (!list_empty(page_deferred_list(head))) { 2645 pgdata->split_queue_len--; 2646 list_del(page_deferred_list(head)); 2647 } 2648 if (mapping) 2649 __dec_node_page_state(page, NR_SHMEM_THPS); 2650 spin_unlock(&pgdata->split_queue_lock); 2651 __split_huge_page(page, list, flags); 2652 if (PageSwapCache(head)) { 2653 swp_entry_t entry = { .val = page_private(head) }; 2654 2655 ret = split_swap_cluster(entry); 2656 } else 2657 ret = 0; 2658 } else { 2659 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2660 pr_alert("total_mapcount: %u, page_count(): %u\n", 2661 mapcount, count); 2662 if (PageTail(page)) 2663 dump_page(head, NULL); 2664 dump_page(page, "total_mapcount(head) > 0"); 2665 BUG(); 2666 } 2667 spin_unlock(&pgdata->split_queue_lock); 2668 fail: if (mapping) 2669 spin_unlock(&mapping->tree_lock); 2670 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2671 unfreeze_page(head); 2672 ret = -EBUSY; 2673 } 2674 2675 out_unlock: 2676 if (anon_vma) { 2677 anon_vma_unlock_write(anon_vma); 2678 put_anon_vma(anon_vma); 2679 } 2680 if (mapping) 2681 i_mmap_unlock_read(mapping); 2682 out: 2683 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2684 return ret; 2685 } 2686 2687 void free_transhuge_page(struct page *page) 2688 { 2689 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2690 unsigned long flags; 2691 2692 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2693 if (!list_empty(page_deferred_list(page))) { 2694 pgdata->split_queue_len--; 2695 list_del(page_deferred_list(page)); 2696 } 2697 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2698 free_compound_page(page); 2699 } 2700 2701 void deferred_split_huge_page(struct page *page) 2702 { 2703 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2704 unsigned long flags; 2705 2706 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2707 2708 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2709 if (list_empty(page_deferred_list(page))) { 2710 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2711 list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2712 pgdata->split_queue_len++; 2713 } 2714 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2715 } 2716 2717 static unsigned long deferred_split_count(struct shrinker *shrink, 2718 struct shrink_control *sc) 2719 { 2720 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2721 return ACCESS_ONCE(pgdata->split_queue_len); 2722 } 2723 2724 static unsigned long deferred_split_scan(struct shrinker *shrink, 2725 struct shrink_control *sc) 2726 { 2727 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2728 unsigned long flags; 2729 LIST_HEAD(list), *pos, *next; 2730 struct page *page; 2731 int split = 0; 2732 2733 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2734 /* Take pin on all head pages to avoid freeing them under us */ 2735 list_for_each_safe(pos, next, &pgdata->split_queue) { 2736 page = list_entry((void *)pos, struct page, mapping); 2737 page = compound_head(page); 2738 if (get_page_unless_zero(page)) { 2739 list_move(page_deferred_list(page), &list); 2740 } else { 2741 /* We lost race with put_compound_page() */ 2742 list_del_init(page_deferred_list(page)); 2743 pgdata->split_queue_len--; 2744 } 2745 if (!--sc->nr_to_scan) 2746 break; 2747 } 2748 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2749 2750 list_for_each_safe(pos, next, &list) { 2751 page = list_entry((void *)pos, struct page, mapping); 2752 lock_page(page); 2753 /* split_huge_page() removes page from list on success */ 2754 if (!split_huge_page(page)) 2755 split++; 2756 unlock_page(page); 2757 put_page(page); 2758 } 2759 2760 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2761 list_splice_tail(&list, &pgdata->split_queue); 2762 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2763 2764 /* 2765 * Stop shrinker if we didn't split any page, but the queue is empty. 2766 * This can happen if pages were freed under us. 2767 */ 2768 if (!split && list_empty(&pgdata->split_queue)) 2769 return SHRINK_STOP; 2770 return split; 2771 } 2772 2773 static struct shrinker deferred_split_shrinker = { 2774 .count_objects = deferred_split_count, 2775 .scan_objects = deferred_split_scan, 2776 .seeks = DEFAULT_SEEKS, 2777 .flags = SHRINKER_NUMA_AWARE, 2778 }; 2779 2780 #ifdef CONFIG_DEBUG_FS 2781 static int split_huge_pages_set(void *data, u64 val) 2782 { 2783 struct zone *zone; 2784 struct page *page; 2785 unsigned long pfn, max_zone_pfn; 2786 unsigned long total = 0, split = 0; 2787 2788 if (val != 1) 2789 return -EINVAL; 2790 2791 for_each_populated_zone(zone) { 2792 max_zone_pfn = zone_end_pfn(zone); 2793 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2794 if (!pfn_valid(pfn)) 2795 continue; 2796 2797 page = pfn_to_page(pfn); 2798 if (!get_page_unless_zero(page)) 2799 continue; 2800 2801 if (zone != page_zone(page)) 2802 goto next; 2803 2804 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2805 goto next; 2806 2807 total++; 2808 lock_page(page); 2809 if (!split_huge_page(page)) 2810 split++; 2811 unlock_page(page); 2812 next: 2813 put_page(page); 2814 } 2815 } 2816 2817 pr_info("%lu of %lu THP split\n", split, total); 2818 2819 return 0; 2820 } 2821 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2822 "%llu\n"); 2823 2824 static int __init split_huge_pages_debugfs(void) 2825 { 2826 void *ret; 2827 2828 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2829 &split_huge_pages_fops); 2830 if (!ret) 2831 pr_warn("Failed to create split_huge_pages in debugfs"); 2832 return 0; 2833 } 2834 late_initcall(split_huge_pages_debugfs); 2835 #endif 2836 2837 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2838 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2839 struct page *page) 2840 { 2841 struct vm_area_struct *vma = pvmw->vma; 2842 struct mm_struct *mm = vma->vm_mm; 2843 unsigned long address = pvmw->address; 2844 pmd_t pmdval; 2845 swp_entry_t entry; 2846 pmd_t pmdswp; 2847 2848 if (!(pvmw->pmd && !pvmw->pte)) 2849 return; 2850 2851 mmu_notifier_invalidate_range_start(mm, address, 2852 address + HPAGE_PMD_SIZE); 2853 2854 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2855 pmdval = *pvmw->pmd; 2856 pmdp_invalidate(vma, address, pvmw->pmd); 2857 if (pmd_dirty(pmdval)) 2858 set_page_dirty(page); 2859 entry = make_migration_entry(page, pmd_write(pmdval)); 2860 pmdswp = swp_entry_to_pmd(entry); 2861 if (pmd_soft_dirty(pmdval)) 2862 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2863 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2864 page_remove_rmap(page, true); 2865 put_page(page); 2866 2867 mmu_notifier_invalidate_range_end(mm, address, 2868 address + HPAGE_PMD_SIZE); 2869 } 2870 2871 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2872 { 2873 struct vm_area_struct *vma = pvmw->vma; 2874 struct mm_struct *mm = vma->vm_mm; 2875 unsigned long address = pvmw->address; 2876 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2877 pmd_t pmde; 2878 swp_entry_t entry; 2879 2880 if (!(pvmw->pmd && !pvmw->pte)) 2881 return; 2882 2883 entry = pmd_to_swp_entry(*pvmw->pmd); 2884 get_page(new); 2885 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2886 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2887 pmde = pmd_mksoft_dirty(pmde); 2888 if (is_write_migration_entry(entry)) 2889 pmde = maybe_pmd_mkwrite(pmde, vma); 2890 2891 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2892 page_add_anon_rmap(new, vma, mmun_start, true); 2893 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2894 if (vma->vm_flags & VM_LOCKED) 2895 mlock_vma_page(new); 2896 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2897 } 2898 #endif 2899