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