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 vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 545 struct page *page, 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 vm_fault_t ret = 0; 552 553 VM_BUG_ON_PAGE(!PageCompound(page), page); 554 555 if (mem_cgroup_try_charge_delay(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 vm_fault_t ret2; 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 ret2 = handle_userfault(vmf, VM_UFFD_MISSING); 594 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK); 595 return ret2; 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, unsigned long addr) 633 { 634 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); 635 gfp_t this_node = 0; 636 637 #ifdef CONFIG_NUMA 638 struct mempolicy *pol; 639 /* 640 * __GFP_THISNODE is used only when __GFP_DIRECT_RECLAIM is not 641 * specified, to express a general desire to stay on the current 642 * node for optimistic allocation attempts. If the defrag mode 643 * and/or madvise hint requires the direct reclaim then we prefer 644 * to fallback to other node rather than node reclaim because that 645 * can lead to excessive reclaim even though there is free memory 646 * on other nodes. We expect that NUMA preferences are specified 647 * by memory policies. 648 */ 649 pol = get_vma_policy(vma, addr); 650 if (pol->mode != MPOL_BIND) 651 this_node = __GFP_THISNODE; 652 mpol_cond_put(pol); 653 #endif 654 655 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 656 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 657 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 658 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM | this_node; 659 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 660 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 661 __GFP_KSWAPD_RECLAIM | this_node); 662 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 663 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : 664 this_node); 665 return GFP_TRANSHUGE_LIGHT | this_node; 666 } 667 668 /* Caller must hold page table lock. */ 669 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 670 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 671 struct page *zero_page) 672 { 673 pmd_t entry; 674 if (!pmd_none(*pmd)) 675 return false; 676 entry = mk_pmd(zero_page, vma->vm_page_prot); 677 entry = pmd_mkhuge(entry); 678 if (pgtable) 679 pgtable_trans_huge_deposit(mm, pmd, pgtable); 680 set_pmd_at(mm, haddr, pmd, entry); 681 mm_inc_nr_ptes(mm); 682 return true; 683 } 684 685 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 686 { 687 struct vm_area_struct *vma = vmf->vma; 688 gfp_t gfp; 689 struct page *page; 690 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 691 692 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 693 return VM_FAULT_FALLBACK; 694 if (unlikely(anon_vma_prepare(vma))) 695 return VM_FAULT_OOM; 696 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 697 return VM_FAULT_OOM; 698 if (!(vmf->flags & FAULT_FLAG_WRITE) && 699 !mm_forbids_zeropage(vma->vm_mm) && 700 transparent_hugepage_use_zero_page()) { 701 pgtable_t pgtable; 702 struct page *zero_page; 703 bool set; 704 vm_fault_t ret; 705 pgtable = pte_alloc_one(vma->vm_mm, haddr); 706 if (unlikely(!pgtable)) 707 return VM_FAULT_OOM; 708 zero_page = mm_get_huge_zero_page(vma->vm_mm); 709 if (unlikely(!zero_page)) { 710 pte_free(vma->vm_mm, pgtable); 711 count_vm_event(THP_FAULT_FALLBACK); 712 return VM_FAULT_FALLBACK; 713 } 714 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 715 ret = 0; 716 set = false; 717 if (pmd_none(*vmf->pmd)) { 718 ret = check_stable_address_space(vma->vm_mm); 719 if (ret) { 720 spin_unlock(vmf->ptl); 721 } else if (userfaultfd_missing(vma)) { 722 spin_unlock(vmf->ptl); 723 ret = handle_userfault(vmf, VM_UFFD_MISSING); 724 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 725 } else { 726 set_huge_zero_page(pgtable, vma->vm_mm, vma, 727 haddr, vmf->pmd, zero_page); 728 spin_unlock(vmf->ptl); 729 set = true; 730 } 731 } else 732 spin_unlock(vmf->ptl); 733 if (!set) 734 pte_free(vma->vm_mm, pgtable); 735 return ret; 736 } 737 gfp = alloc_hugepage_direct_gfpmask(vma, haddr); 738 page = alloc_pages_vma(gfp, HPAGE_PMD_ORDER, vma, haddr, numa_node_id()); 739 if (unlikely(!page)) { 740 count_vm_event(THP_FAULT_FALLBACK); 741 return VM_FAULT_FALLBACK; 742 } 743 prep_transhuge_page(page); 744 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 745 } 746 747 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 748 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 749 pgtable_t pgtable) 750 { 751 struct mm_struct *mm = vma->vm_mm; 752 pmd_t entry; 753 spinlock_t *ptl; 754 755 ptl = pmd_lock(mm, pmd); 756 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 757 if (pfn_t_devmap(pfn)) 758 entry = pmd_mkdevmap(entry); 759 if (write) { 760 entry = pmd_mkyoung(pmd_mkdirty(entry)); 761 entry = maybe_pmd_mkwrite(entry, vma); 762 } 763 764 if (pgtable) { 765 pgtable_trans_huge_deposit(mm, pmd, pgtable); 766 mm_inc_nr_ptes(mm); 767 } 768 769 set_pmd_at(mm, addr, pmd, entry); 770 update_mmu_cache_pmd(vma, addr, pmd); 771 spin_unlock(ptl); 772 } 773 774 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 775 pmd_t *pmd, pfn_t pfn, bool write) 776 { 777 pgprot_t pgprot = vma->vm_page_prot; 778 pgtable_t pgtable = NULL; 779 /* 780 * If we had pmd_special, we could avoid all these restrictions, 781 * but we need to be consistent with PTEs and architectures that 782 * can't support a 'special' bit. 783 */ 784 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 785 !pfn_t_devmap(pfn)); 786 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 787 (VM_PFNMAP|VM_MIXEDMAP)); 788 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 789 790 if (addr < vma->vm_start || addr >= vma->vm_end) 791 return VM_FAULT_SIGBUS; 792 793 if (arch_needs_pgtable_deposit()) { 794 pgtable = pte_alloc_one(vma->vm_mm, addr); 795 if (!pgtable) 796 return VM_FAULT_OOM; 797 } 798 799 track_pfn_insert(vma, &pgprot, pfn); 800 801 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 802 return VM_FAULT_NOPAGE; 803 } 804 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 805 806 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 807 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 808 { 809 if (likely(vma->vm_flags & VM_WRITE)) 810 pud = pud_mkwrite(pud); 811 return pud; 812 } 813 814 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 815 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 816 { 817 struct mm_struct *mm = vma->vm_mm; 818 pud_t entry; 819 spinlock_t *ptl; 820 821 ptl = pud_lock(mm, pud); 822 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 823 if (pfn_t_devmap(pfn)) 824 entry = pud_mkdevmap(entry); 825 if (write) { 826 entry = pud_mkyoung(pud_mkdirty(entry)); 827 entry = maybe_pud_mkwrite(entry, vma); 828 } 829 set_pud_at(mm, addr, pud, entry); 830 update_mmu_cache_pud(vma, addr, pud); 831 spin_unlock(ptl); 832 } 833 834 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 835 pud_t *pud, pfn_t pfn, bool write) 836 { 837 pgprot_t pgprot = vma->vm_page_prot; 838 /* 839 * If we had pud_special, we could avoid all these restrictions, 840 * but we need to be consistent with PTEs and architectures that 841 * can't support a 'special' bit. 842 */ 843 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 844 !pfn_t_devmap(pfn)); 845 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 846 (VM_PFNMAP|VM_MIXEDMAP)); 847 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 848 849 if (addr < vma->vm_start || addr >= vma->vm_end) 850 return VM_FAULT_SIGBUS; 851 852 track_pfn_insert(vma, &pgprot, pfn); 853 854 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 855 return VM_FAULT_NOPAGE; 856 } 857 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 858 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 859 860 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 861 pmd_t *pmd, int flags) 862 { 863 pmd_t _pmd; 864 865 _pmd = pmd_mkyoung(*pmd); 866 if (flags & FOLL_WRITE) 867 _pmd = pmd_mkdirty(_pmd); 868 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 869 pmd, _pmd, flags & FOLL_WRITE)) 870 update_mmu_cache_pmd(vma, addr, pmd); 871 } 872 873 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 874 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 875 { 876 unsigned long pfn = pmd_pfn(*pmd); 877 struct mm_struct *mm = vma->vm_mm; 878 struct page *page; 879 880 assert_spin_locked(pmd_lockptr(mm, pmd)); 881 882 /* 883 * When we COW a devmap PMD entry, we split it into PTEs, so we should 884 * not be in this function with `flags & FOLL_COW` set. 885 */ 886 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 887 888 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 889 return NULL; 890 891 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 892 /* pass */; 893 else 894 return NULL; 895 896 if (flags & FOLL_TOUCH) 897 touch_pmd(vma, addr, pmd, flags); 898 899 /* 900 * device mapped pages can only be returned if the 901 * caller will manage the page reference count. 902 */ 903 if (!(flags & FOLL_GET)) 904 return ERR_PTR(-EEXIST); 905 906 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 907 *pgmap = get_dev_pagemap(pfn, *pgmap); 908 if (!*pgmap) 909 return ERR_PTR(-EFAULT); 910 page = pfn_to_page(pfn); 911 get_page(page); 912 913 return page; 914 } 915 916 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 917 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 918 struct vm_area_struct *vma) 919 { 920 spinlock_t *dst_ptl, *src_ptl; 921 struct page *src_page; 922 pmd_t pmd; 923 pgtable_t pgtable = NULL; 924 int ret = -ENOMEM; 925 926 /* Skip if can be re-fill on fault */ 927 if (!vma_is_anonymous(vma)) 928 return 0; 929 930 pgtable = pte_alloc_one(dst_mm, addr); 931 if (unlikely(!pgtable)) 932 goto out; 933 934 dst_ptl = pmd_lock(dst_mm, dst_pmd); 935 src_ptl = pmd_lockptr(src_mm, src_pmd); 936 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 937 938 ret = -EAGAIN; 939 pmd = *src_pmd; 940 941 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 942 if (unlikely(is_swap_pmd(pmd))) { 943 swp_entry_t entry = pmd_to_swp_entry(pmd); 944 945 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 946 if (is_write_migration_entry(entry)) { 947 make_migration_entry_read(&entry); 948 pmd = swp_entry_to_pmd(entry); 949 if (pmd_swp_soft_dirty(*src_pmd)) 950 pmd = pmd_swp_mksoft_dirty(pmd); 951 set_pmd_at(src_mm, addr, src_pmd, pmd); 952 } 953 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 954 mm_inc_nr_ptes(dst_mm); 955 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 956 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 957 ret = 0; 958 goto out_unlock; 959 } 960 #endif 961 962 if (unlikely(!pmd_trans_huge(pmd))) { 963 pte_free(dst_mm, pgtable); 964 goto out_unlock; 965 } 966 /* 967 * When page table lock is held, the huge zero pmd should not be 968 * under splitting since we don't split the page itself, only pmd to 969 * a page table. 970 */ 971 if (is_huge_zero_pmd(pmd)) { 972 struct page *zero_page; 973 /* 974 * get_huge_zero_page() will never allocate a new page here, 975 * since we already have a zero page to copy. It just takes a 976 * reference. 977 */ 978 zero_page = mm_get_huge_zero_page(dst_mm); 979 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 980 zero_page); 981 ret = 0; 982 goto out_unlock; 983 } 984 985 src_page = pmd_page(pmd); 986 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 987 get_page(src_page); 988 page_dup_rmap(src_page, true); 989 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 990 mm_inc_nr_ptes(dst_mm); 991 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 992 993 pmdp_set_wrprotect(src_mm, addr, src_pmd); 994 pmd = pmd_mkold(pmd_wrprotect(pmd)); 995 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 996 997 ret = 0; 998 out_unlock: 999 spin_unlock(src_ptl); 1000 spin_unlock(dst_ptl); 1001 out: 1002 return ret; 1003 } 1004 1005 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1006 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 1007 pud_t *pud, int flags) 1008 { 1009 pud_t _pud; 1010 1011 _pud = pud_mkyoung(*pud); 1012 if (flags & FOLL_WRITE) 1013 _pud = pud_mkdirty(_pud); 1014 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1015 pud, _pud, flags & FOLL_WRITE)) 1016 update_mmu_cache_pud(vma, addr, pud); 1017 } 1018 1019 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1020 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1021 { 1022 unsigned long pfn = pud_pfn(*pud); 1023 struct mm_struct *mm = vma->vm_mm; 1024 struct page *page; 1025 1026 assert_spin_locked(pud_lockptr(mm, pud)); 1027 1028 if (flags & FOLL_WRITE && !pud_write(*pud)) 1029 return NULL; 1030 1031 if (pud_present(*pud) && pud_devmap(*pud)) 1032 /* pass */; 1033 else 1034 return NULL; 1035 1036 if (flags & FOLL_TOUCH) 1037 touch_pud(vma, addr, pud, flags); 1038 1039 /* 1040 * device mapped pages can only be returned if the 1041 * caller will manage the page reference count. 1042 */ 1043 if (!(flags & FOLL_GET)) 1044 return ERR_PTR(-EEXIST); 1045 1046 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1047 *pgmap = get_dev_pagemap(pfn, *pgmap); 1048 if (!*pgmap) 1049 return ERR_PTR(-EFAULT); 1050 page = pfn_to_page(pfn); 1051 get_page(page); 1052 1053 return page; 1054 } 1055 1056 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1057 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1058 struct vm_area_struct *vma) 1059 { 1060 spinlock_t *dst_ptl, *src_ptl; 1061 pud_t pud; 1062 int ret; 1063 1064 dst_ptl = pud_lock(dst_mm, dst_pud); 1065 src_ptl = pud_lockptr(src_mm, src_pud); 1066 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1067 1068 ret = -EAGAIN; 1069 pud = *src_pud; 1070 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1071 goto out_unlock; 1072 1073 /* 1074 * When page table lock is held, the huge zero pud should not be 1075 * under splitting since we don't split the page itself, only pud to 1076 * a page table. 1077 */ 1078 if (is_huge_zero_pud(pud)) { 1079 /* No huge zero pud yet */ 1080 } 1081 1082 pudp_set_wrprotect(src_mm, addr, src_pud); 1083 pud = pud_mkold(pud_wrprotect(pud)); 1084 set_pud_at(dst_mm, addr, dst_pud, pud); 1085 1086 ret = 0; 1087 out_unlock: 1088 spin_unlock(src_ptl); 1089 spin_unlock(dst_ptl); 1090 return ret; 1091 } 1092 1093 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1094 { 1095 pud_t entry; 1096 unsigned long haddr; 1097 bool write = vmf->flags & FAULT_FLAG_WRITE; 1098 1099 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1100 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1101 goto unlock; 1102 1103 entry = pud_mkyoung(orig_pud); 1104 if (write) 1105 entry = pud_mkdirty(entry); 1106 haddr = vmf->address & HPAGE_PUD_MASK; 1107 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1108 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1109 1110 unlock: 1111 spin_unlock(vmf->ptl); 1112 } 1113 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1114 1115 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1116 { 1117 pmd_t entry; 1118 unsigned long haddr; 1119 bool write = vmf->flags & FAULT_FLAG_WRITE; 1120 1121 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1122 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1123 goto unlock; 1124 1125 entry = pmd_mkyoung(orig_pmd); 1126 if (write) 1127 entry = pmd_mkdirty(entry); 1128 haddr = vmf->address & HPAGE_PMD_MASK; 1129 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1130 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1131 1132 unlock: 1133 spin_unlock(vmf->ptl); 1134 } 1135 1136 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, 1137 pmd_t orig_pmd, struct page *page) 1138 { 1139 struct vm_area_struct *vma = vmf->vma; 1140 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1141 struct mem_cgroup *memcg; 1142 pgtable_t pgtable; 1143 pmd_t _pmd; 1144 int i; 1145 vm_fault_t ret = 0; 1146 struct page **pages; 1147 unsigned long mmun_start; /* For mmu_notifiers */ 1148 unsigned long mmun_end; /* For mmu_notifiers */ 1149 1150 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *), 1151 GFP_KERNEL); 1152 if (unlikely(!pages)) { 1153 ret |= VM_FAULT_OOM; 1154 goto out; 1155 } 1156 1157 for (i = 0; i < HPAGE_PMD_NR; i++) { 1158 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 1159 vmf->address, page_to_nid(page)); 1160 if (unlikely(!pages[i] || 1161 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm, 1162 GFP_KERNEL, &memcg, false))) { 1163 if (pages[i]) 1164 put_page(pages[i]); 1165 while (--i >= 0) { 1166 memcg = (void *)page_private(pages[i]); 1167 set_page_private(pages[i], 0); 1168 mem_cgroup_cancel_charge(pages[i], memcg, 1169 false); 1170 put_page(pages[i]); 1171 } 1172 kfree(pages); 1173 ret |= VM_FAULT_OOM; 1174 goto out; 1175 } 1176 set_page_private(pages[i], (unsigned long)memcg); 1177 } 1178 1179 for (i = 0; i < HPAGE_PMD_NR; i++) { 1180 copy_user_highpage(pages[i], page + i, 1181 haddr + PAGE_SIZE * i, vma); 1182 __SetPageUptodate(pages[i]); 1183 cond_resched(); 1184 } 1185 1186 mmun_start = haddr; 1187 mmun_end = haddr + HPAGE_PMD_SIZE; 1188 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1189 1190 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1191 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1192 goto out_free_pages; 1193 VM_BUG_ON_PAGE(!PageHead(page), page); 1194 1195 /* 1196 * Leave pmd empty until pte is filled note we must notify here as 1197 * concurrent CPU thread might write to new page before the call to 1198 * mmu_notifier_invalidate_range_end() happens which can lead to a 1199 * device seeing memory write in different order than CPU. 1200 * 1201 * See Documentation/vm/mmu_notifier.rst 1202 */ 1203 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1204 1205 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1206 pmd_populate(vma->vm_mm, &_pmd, pgtable); 1207 1208 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1209 pte_t entry; 1210 entry = mk_pte(pages[i], vma->vm_page_prot); 1211 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1212 memcg = (void *)page_private(pages[i]); 1213 set_page_private(pages[i], 0); 1214 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1215 mem_cgroup_commit_charge(pages[i], memcg, false, false); 1216 lru_cache_add_active_or_unevictable(pages[i], vma); 1217 vmf->pte = pte_offset_map(&_pmd, haddr); 1218 VM_BUG_ON(!pte_none(*vmf->pte)); 1219 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 1220 pte_unmap(vmf->pte); 1221 } 1222 kfree(pages); 1223 1224 smp_wmb(); /* make pte visible before pmd */ 1225 pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1226 page_remove_rmap(page, true); 1227 spin_unlock(vmf->ptl); 1228 1229 /* 1230 * No need to double call mmu_notifier->invalidate_range() callback as 1231 * the above pmdp_huge_clear_flush_notify() did already call it. 1232 */ 1233 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1234 mmun_end); 1235 1236 ret |= VM_FAULT_WRITE; 1237 put_page(page); 1238 1239 out: 1240 return ret; 1241 1242 out_free_pages: 1243 spin_unlock(vmf->ptl); 1244 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1245 for (i = 0; i < HPAGE_PMD_NR; i++) { 1246 memcg = (void *)page_private(pages[i]); 1247 set_page_private(pages[i], 0); 1248 mem_cgroup_cancel_charge(pages[i], memcg, false); 1249 put_page(pages[i]); 1250 } 1251 kfree(pages); 1252 goto out; 1253 } 1254 1255 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1256 { 1257 struct vm_area_struct *vma = vmf->vma; 1258 struct page *page = NULL, *new_page; 1259 struct mem_cgroup *memcg; 1260 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1261 unsigned long mmun_start; /* For mmu_notifiers */ 1262 unsigned long mmun_end; /* For mmu_notifiers */ 1263 gfp_t huge_gfp; /* for allocation and charge */ 1264 vm_fault_t ret = 0; 1265 1266 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1267 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1268 if (is_huge_zero_pmd(orig_pmd)) 1269 goto alloc; 1270 spin_lock(vmf->ptl); 1271 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1272 goto out_unlock; 1273 1274 page = pmd_page(orig_pmd); 1275 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1276 /* 1277 * We can only reuse the page if nobody else maps the huge page or it's 1278 * part. 1279 */ 1280 if (!trylock_page(page)) { 1281 get_page(page); 1282 spin_unlock(vmf->ptl); 1283 lock_page(page); 1284 spin_lock(vmf->ptl); 1285 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1286 unlock_page(page); 1287 put_page(page); 1288 goto out_unlock; 1289 } 1290 put_page(page); 1291 } 1292 if (reuse_swap_page(page, NULL)) { 1293 pmd_t entry; 1294 entry = pmd_mkyoung(orig_pmd); 1295 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1296 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1297 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1298 ret |= VM_FAULT_WRITE; 1299 unlock_page(page); 1300 goto out_unlock; 1301 } 1302 unlock_page(page); 1303 get_page(page); 1304 spin_unlock(vmf->ptl); 1305 alloc: 1306 if (transparent_hugepage_enabled(vma) && 1307 !transparent_hugepage_debug_cow()) { 1308 huge_gfp = alloc_hugepage_direct_gfpmask(vma, haddr); 1309 new_page = alloc_pages_vma(huge_gfp, HPAGE_PMD_ORDER, vma, 1310 haddr, numa_node_id()); 1311 } else 1312 new_page = NULL; 1313 1314 if (likely(new_page)) { 1315 prep_transhuge_page(new_page); 1316 } else { 1317 if (!page) { 1318 split_huge_pmd(vma, vmf->pmd, vmf->address); 1319 ret |= VM_FAULT_FALLBACK; 1320 } else { 1321 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 1322 if (ret & VM_FAULT_OOM) { 1323 split_huge_pmd(vma, vmf->pmd, vmf->address); 1324 ret |= VM_FAULT_FALLBACK; 1325 } 1326 put_page(page); 1327 } 1328 count_vm_event(THP_FAULT_FALLBACK); 1329 goto out; 1330 } 1331 1332 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm, 1333 huge_gfp, &memcg, true))) { 1334 put_page(new_page); 1335 split_huge_pmd(vma, vmf->pmd, vmf->address); 1336 if (page) 1337 put_page(page); 1338 ret |= VM_FAULT_FALLBACK; 1339 count_vm_event(THP_FAULT_FALLBACK); 1340 goto out; 1341 } 1342 1343 count_vm_event(THP_FAULT_ALLOC); 1344 1345 if (!page) 1346 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 1347 else 1348 copy_user_huge_page(new_page, page, vmf->address, 1349 vma, HPAGE_PMD_NR); 1350 __SetPageUptodate(new_page); 1351 1352 mmun_start = haddr; 1353 mmun_end = haddr + HPAGE_PMD_SIZE; 1354 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1355 1356 spin_lock(vmf->ptl); 1357 if (page) 1358 put_page(page); 1359 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1360 spin_unlock(vmf->ptl); 1361 mem_cgroup_cancel_charge(new_page, memcg, true); 1362 put_page(new_page); 1363 goto out_mn; 1364 } else { 1365 pmd_t entry; 1366 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1367 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1368 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1369 page_add_new_anon_rmap(new_page, vma, haddr, true); 1370 mem_cgroup_commit_charge(new_page, memcg, false, true); 1371 lru_cache_add_active_or_unevictable(new_page, vma); 1372 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 1373 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1374 if (!page) { 1375 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1376 } else { 1377 VM_BUG_ON_PAGE(!PageHead(page), page); 1378 page_remove_rmap(page, true); 1379 put_page(page); 1380 } 1381 ret |= VM_FAULT_WRITE; 1382 } 1383 spin_unlock(vmf->ptl); 1384 out_mn: 1385 /* 1386 * No need to double call mmu_notifier->invalidate_range() callback as 1387 * the above pmdp_huge_clear_flush_notify() did already call it. 1388 */ 1389 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1390 mmun_end); 1391 out: 1392 return ret; 1393 out_unlock: 1394 spin_unlock(vmf->ptl); 1395 return ret; 1396 } 1397 1398 /* 1399 * FOLL_FORCE can write to even unwritable pmd's, but only 1400 * after we've gone through a COW cycle and they are dirty. 1401 */ 1402 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1403 { 1404 return pmd_write(pmd) || 1405 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 1406 } 1407 1408 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1409 unsigned long addr, 1410 pmd_t *pmd, 1411 unsigned int flags) 1412 { 1413 struct mm_struct *mm = vma->vm_mm; 1414 struct page *page = NULL; 1415 1416 assert_spin_locked(pmd_lockptr(mm, pmd)); 1417 1418 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1419 goto out; 1420 1421 /* Avoid dumping huge zero page */ 1422 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1423 return ERR_PTR(-EFAULT); 1424 1425 /* Full NUMA hinting faults to serialise migration in fault paths */ 1426 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1427 goto out; 1428 1429 page = pmd_page(*pmd); 1430 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1431 if (flags & FOLL_TOUCH) 1432 touch_pmd(vma, addr, pmd, flags); 1433 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1434 /* 1435 * We don't mlock() pte-mapped THPs. This way we can avoid 1436 * leaking mlocked pages into non-VM_LOCKED VMAs. 1437 * 1438 * For anon THP: 1439 * 1440 * In most cases the pmd is the only mapping of the page as we 1441 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1442 * writable private mappings in populate_vma_page_range(). 1443 * 1444 * The only scenario when we have the page shared here is if we 1445 * mlocking read-only mapping shared over fork(). We skip 1446 * mlocking such pages. 1447 * 1448 * For file THP: 1449 * 1450 * We can expect PageDoubleMap() to be stable under page lock: 1451 * for file pages we set it in page_add_file_rmap(), which 1452 * requires page to be locked. 1453 */ 1454 1455 if (PageAnon(page) && compound_mapcount(page) != 1) 1456 goto skip_mlock; 1457 if (PageDoubleMap(page) || !page->mapping) 1458 goto skip_mlock; 1459 if (!trylock_page(page)) 1460 goto skip_mlock; 1461 lru_add_drain(); 1462 if (page->mapping && !PageDoubleMap(page)) 1463 mlock_vma_page(page); 1464 unlock_page(page); 1465 } 1466 skip_mlock: 1467 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1468 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1469 if (flags & FOLL_GET) 1470 get_page(page); 1471 1472 out: 1473 return page; 1474 } 1475 1476 /* NUMA hinting page fault entry point for trans huge pmds */ 1477 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1478 { 1479 struct vm_area_struct *vma = vmf->vma; 1480 struct anon_vma *anon_vma = NULL; 1481 struct page *page; 1482 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1483 int page_nid = -1, this_nid = numa_node_id(); 1484 int target_nid, last_cpupid = -1; 1485 bool page_locked; 1486 bool migrated = false; 1487 bool was_writable; 1488 int flags = 0; 1489 1490 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1491 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1492 goto out_unlock; 1493 1494 /* 1495 * If there are potential migrations, wait for completion and retry 1496 * without disrupting NUMA hinting information. Do not relock and 1497 * check_same as the page may no longer be mapped. 1498 */ 1499 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1500 page = pmd_page(*vmf->pmd); 1501 if (!get_page_unless_zero(page)) 1502 goto out_unlock; 1503 spin_unlock(vmf->ptl); 1504 wait_on_page_locked(page); 1505 put_page(page); 1506 goto out; 1507 } 1508 1509 page = pmd_page(pmd); 1510 BUG_ON(is_huge_zero_page(page)); 1511 page_nid = page_to_nid(page); 1512 last_cpupid = page_cpupid_last(page); 1513 count_vm_numa_event(NUMA_HINT_FAULTS); 1514 if (page_nid == this_nid) { 1515 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1516 flags |= TNF_FAULT_LOCAL; 1517 } 1518 1519 /* See similar comment in do_numa_page for explanation */ 1520 if (!pmd_savedwrite(pmd)) 1521 flags |= TNF_NO_GROUP; 1522 1523 /* 1524 * Acquire the page lock to serialise THP migrations but avoid dropping 1525 * page_table_lock if at all possible 1526 */ 1527 page_locked = trylock_page(page); 1528 target_nid = mpol_misplaced(page, vma, haddr); 1529 if (target_nid == -1) { 1530 /* If the page was locked, there are no parallel migrations */ 1531 if (page_locked) 1532 goto clear_pmdnuma; 1533 } 1534 1535 /* Migration could have started since the pmd_trans_migrating check */ 1536 if (!page_locked) { 1537 page_nid = -1; 1538 if (!get_page_unless_zero(page)) 1539 goto out_unlock; 1540 spin_unlock(vmf->ptl); 1541 wait_on_page_locked(page); 1542 put_page(page); 1543 goto out; 1544 } 1545 1546 /* 1547 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1548 * to serialises splits 1549 */ 1550 get_page(page); 1551 spin_unlock(vmf->ptl); 1552 anon_vma = page_lock_anon_vma_read(page); 1553 1554 /* Confirm the PMD did not change while page_table_lock was released */ 1555 spin_lock(vmf->ptl); 1556 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1557 unlock_page(page); 1558 put_page(page); 1559 page_nid = -1; 1560 goto out_unlock; 1561 } 1562 1563 /* Bail if we fail to protect against THP splits for any reason */ 1564 if (unlikely(!anon_vma)) { 1565 put_page(page); 1566 page_nid = -1; 1567 goto clear_pmdnuma; 1568 } 1569 1570 /* 1571 * Since we took the NUMA fault, we must have observed the !accessible 1572 * bit. Make sure all other CPUs agree with that, to avoid them 1573 * modifying the page we're about to migrate. 1574 * 1575 * Must be done under PTL such that we'll observe the relevant 1576 * inc_tlb_flush_pending(). 1577 * 1578 * We are not sure a pending tlb flush here is for a huge page 1579 * mapping or not. Hence use the tlb range variant 1580 */ 1581 if (mm_tlb_flush_pending(vma->vm_mm)) { 1582 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1583 /* 1584 * change_huge_pmd() released the pmd lock before 1585 * invalidating the secondary MMUs sharing the primary 1586 * MMU pagetables (with ->invalidate_range()). The 1587 * mmu_notifier_invalidate_range_end() (which 1588 * internally calls ->invalidate_range()) in 1589 * change_pmd_range() will run after us, so we can't 1590 * rely on it here and we need an explicit invalidate. 1591 */ 1592 mmu_notifier_invalidate_range(vma->vm_mm, haddr, 1593 haddr + HPAGE_PMD_SIZE); 1594 } 1595 1596 /* 1597 * Migrate the THP to the requested node, returns with page unlocked 1598 * and access rights restored. 1599 */ 1600 spin_unlock(vmf->ptl); 1601 1602 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1603 vmf->pmd, pmd, vmf->address, page, target_nid); 1604 if (migrated) { 1605 flags |= TNF_MIGRATED; 1606 page_nid = target_nid; 1607 } else 1608 flags |= TNF_MIGRATE_FAIL; 1609 1610 goto out; 1611 clear_pmdnuma: 1612 BUG_ON(!PageLocked(page)); 1613 was_writable = pmd_savedwrite(pmd); 1614 pmd = pmd_modify(pmd, vma->vm_page_prot); 1615 pmd = pmd_mkyoung(pmd); 1616 if (was_writable) 1617 pmd = pmd_mkwrite(pmd); 1618 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1619 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1620 unlock_page(page); 1621 out_unlock: 1622 spin_unlock(vmf->ptl); 1623 1624 out: 1625 if (anon_vma) 1626 page_unlock_anon_vma_read(anon_vma); 1627 1628 if (page_nid != -1) 1629 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1630 flags); 1631 1632 return 0; 1633 } 1634 1635 /* 1636 * Return true if we do MADV_FREE successfully on entire pmd page. 1637 * Otherwise, return false. 1638 */ 1639 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1640 pmd_t *pmd, unsigned long addr, unsigned long next) 1641 { 1642 spinlock_t *ptl; 1643 pmd_t orig_pmd; 1644 struct page *page; 1645 struct mm_struct *mm = tlb->mm; 1646 bool ret = false; 1647 1648 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1649 1650 ptl = pmd_trans_huge_lock(pmd, vma); 1651 if (!ptl) 1652 goto out_unlocked; 1653 1654 orig_pmd = *pmd; 1655 if (is_huge_zero_pmd(orig_pmd)) 1656 goto out; 1657 1658 if (unlikely(!pmd_present(orig_pmd))) { 1659 VM_BUG_ON(thp_migration_supported() && 1660 !is_pmd_migration_entry(orig_pmd)); 1661 goto out; 1662 } 1663 1664 page = pmd_page(orig_pmd); 1665 /* 1666 * If other processes are mapping this page, we couldn't discard 1667 * the page unless they all do MADV_FREE so let's skip the page. 1668 */ 1669 if (page_mapcount(page) != 1) 1670 goto out; 1671 1672 if (!trylock_page(page)) 1673 goto out; 1674 1675 /* 1676 * If user want to discard part-pages of THP, split it so MADV_FREE 1677 * will deactivate only them. 1678 */ 1679 if (next - addr != HPAGE_PMD_SIZE) { 1680 get_page(page); 1681 spin_unlock(ptl); 1682 split_huge_page(page); 1683 unlock_page(page); 1684 put_page(page); 1685 goto out_unlocked; 1686 } 1687 1688 if (PageDirty(page)) 1689 ClearPageDirty(page); 1690 unlock_page(page); 1691 1692 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1693 pmdp_invalidate(vma, addr, pmd); 1694 orig_pmd = pmd_mkold(orig_pmd); 1695 orig_pmd = pmd_mkclean(orig_pmd); 1696 1697 set_pmd_at(mm, addr, pmd, orig_pmd); 1698 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1699 } 1700 1701 mark_page_lazyfree(page); 1702 ret = true; 1703 out: 1704 spin_unlock(ptl); 1705 out_unlocked: 1706 return ret; 1707 } 1708 1709 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1710 { 1711 pgtable_t pgtable; 1712 1713 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1714 pte_free(mm, pgtable); 1715 mm_dec_nr_ptes(mm); 1716 } 1717 1718 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1719 pmd_t *pmd, unsigned long addr) 1720 { 1721 pmd_t orig_pmd; 1722 spinlock_t *ptl; 1723 1724 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1725 1726 ptl = __pmd_trans_huge_lock(pmd, vma); 1727 if (!ptl) 1728 return 0; 1729 /* 1730 * For architectures like ppc64 we look at deposited pgtable 1731 * when calling pmdp_huge_get_and_clear. So do the 1732 * pgtable_trans_huge_withdraw after finishing pmdp related 1733 * operations. 1734 */ 1735 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1736 tlb->fullmm); 1737 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1738 if (vma_is_dax(vma)) { 1739 if (arch_needs_pgtable_deposit()) 1740 zap_deposited_table(tlb->mm, pmd); 1741 spin_unlock(ptl); 1742 if (is_huge_zero_pmd(orig_pmd)) 1743 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1744 } else if (is_huge_zero_pmd(orig_pmd)) { 1745 zap_deposited_table(tlb->mm, pmd); 1746 spin_unlock(ptl); 1747 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1748 } else { 1749 struct page *page = NULL; 1750 int flush_needed = 1; 1751 1752 if (pmd_present(orig_pmd)) { 1753 page = pmd_page(orig_pmd); 1754 page_remove_rmap(page, true); 1755 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1756 VM_BUG_ON_PAGE(!PageHead(page), page); 1757 } else if (thp_migration_supported()) { 1758 swp_entry_t entry; 1759 1760 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1761 entry = pmd_to_swp_entry(orig_pmd); 1762 page = pfn_to_page(swp_offset(entry)); 1763 flush_needed = 0; 1764 } else 1765 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1766 1767 if (PageAnon(page)) { 1768 zap_deposited_table(tlb->mm, pmd); 1769 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1770 } else { 1771 if (arch_needs_pgtable_deposit()) 1772 zap_deposited_table(tlb->mm, pmd); 1773 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1774 } 1775 1776 spin_unlock(ptl); 1777 if (flush_needed) 1778 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1779 } 1780 return 1; 1781 } 1782 1783 #ifndef pmd_move_must_withdraw 1784 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1785 spinlock_t *old_pmd_ptl, 1786 struct vm_area_struct *vma) 1787 { 1788 /* 1789 * With split pmd lock we also need to move preallocated 1790 * PTE page table if new_pmd is on different PMD page table. 1791 * 1792 * We also don't deposit and withdraw tables for file pages. 1793 */ 1794 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1795 } 1796 #endif 1797 1798 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1799 { 1800 #ifdef CONFIG_MEM_SOFT_DIRTY 1801 if (unlikely(is_pmd_migration_entry(pmd))) 1802 pmd = pmd_swp_mksoft_dirty(pmd); 1803 else if (pmd_present(pmd)) 1804 pmd = pmd_mksoft_dirty(pmd); 1805 #endif 1806 return pmd; 1807 } 1808 1809 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1810 unsigned long new_addr, unsigned long old_end, 1811 pmd_t *old_pmd, pmd_t *new_pmd) 1812 { 1813 spinlock_t *old_ptl, *new_ptl; 1814 pmd_t pmd; 1815 struct mm_struct *mm = vma->vm_mm; 1816 bool force_flush = false; 1817 1818 if ((old_addr & ~HPAGE_PMD_MASK) || 1819 (new_addr & ~HPAGE_PMD_MASK) || 1820 old_end - old_addr < HPAGE_PMD_SIZE) 1821 return false; 1822 1823 /* 1824 * The destination pmd shouldn't be established, free_pgtables() 1825 * should have release it. 1826 */ 1827 if (WARN_ON(!pmd_none(*new_pmd))) { 1828 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1829 return false; 1830 } 1831 1832 /* 1833 * We don't have to worry about the ordering of src and dst 1834 * ptlocks because exclusive mmap_sem prevents deadlock. 1835 */ 1836 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1837 if (old_ptl) { 1838 new_ptl = pmd_lockptr(mm, new_pmd); 1839 if (new_ptl != old_ptl) 1840 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1841 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1842 if (pmd_present(pmd)) 1843 force_flush = true; 1844 VM_BUG_ON(!pmd_none(*new_pmd)); 1845 1846 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1847 pgtable_t pgtable; 1848 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1849 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1850 } 1851 pmd = move_soft_dirty_pmd(pmd); 1852 set_pmd_at(mm, new_addr, new_pmd, pmd); 1853 if (force_flush) 1854 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1855 if (new_ptl != old_ptl) 1856 spin_unlock(new_ptl); 1857 spin_unlock(old_ptl); 1858 return true; 1859 } 1860 return false; 1861 } 1862 1863 /* 1864 * Returns 1865 * - 0 if PMD could not be locked 1866 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1867 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1868 */ 1869 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1870 unsigned long addr, pgprot_t newprot, int prot_numa) 1871 { 1872 struct mm_struct *mm = vma->vm_mm; 1873 spinlock_t *ptl; 1874 pmd_t entry; 1875 bool preserve_write; 1876 int ret; 1877 1878 ptl = __pmd_trans_huge_lock(pmd, vma); 1879 if (!ptl) 1880 return 0; 1881 1882 preserve_write = prot_numa && pmd_write(*pmd); 1883 ret = 1; 1884 1885 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1886 if (is_swap_pmd(*pmd)) { 1887 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1888 1889 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1890 if (is_write_migration_entry(entry)) { 1891 pmd_t newpmd; 1892 /* 1893 * A protection check is difficult so 1894 * just be safe and disable write 1895 */ 1896 make_migration_entry_read(&entry); 1897 newpmd = swp_entry_to_pmd(entry); 1898 if (pmd_swp_soft_dirty(*pmd)) 1899 newpmd = pmd_swp_mksoft_dirty(newpmd); 1900 set_pmd_at(mm, addr, pmd, newpmd); 1901 } 1902 goto unlock; 1903 } 1904 #endif 1905 1906 /* 1907 * Avoid trapping faults against the zero page. The read-only 1908 * data is likely to be read-cached on the local CPU and 1909 * local/remote hits to the zero page are not interesting. 1910 */ 1911 if (prot_numa && is_huge_zero_pmd(*pmd)) 1912 goto unlock; 1913 1914 if (prot_numa && pmd_protnone(*pmd)) 1915 goto unlock; 1916 1917 /* 1918 * In case prot_numa, we are under down_read(mmap_sem). It's critical 1919 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1920 * which is also under down_read(mmap_sem): 1921 * 1922 * CPU0: CPU1: 1923 * change_huge_pmd(prot_numa=1) 1924 * pmdp_huge_get_and_clear_notify() 1925 * madvise_dontneed() 1926 * zap_pmd_range() 1927 * pmd_trans_huge(*pmd) == 0 (without ptl) 1928 * // skip the pmd 1929 * set_pmd_at(); 1930 * // pmd is re-established 1931 * 1932 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1933 * which may break userspace. 1934 * 1935 * pmdp_invalidate() is required to make sure we don't miss 1936 * dirty/young flags set by hardware. 1937 */ 1938 entry = pmdp_invalidate(vma, addr, pmd); 1939 1940 entry = pmd_modify(entry, newprot); 1941 if (preserve_write) 1942 entry = pmd_mk_savedwrite(entry); 1943 ret = HPAGE_PMD_NR; 1944 set_pmd_at(mm, addr, pmd, entry); 1945 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1946 unlock: 1947 spin_unlock(ptl); 1948 return ret; 1949 } 1950 1951 /* 1952 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1953 * 1954 * Note that if it returns page table lock pointer, this routine returns without 1955 * unlocking page table lock. So callers must unlock it. 1956 */ 1957 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1958 { 1959 spinlock_t *ptl; 1960 ptl = pmd_lock(vma->vm_mm, pmd); 1961 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1962 pmd_devmap(*pmd))) 1963 return ptl; 1964 spin_unlock(ptl); 1965 return NULL; 1966 } 1967 1968 /* 1969 * Returns true if a given pud maps a thp, false otherwise. 1970 * 1971 * Note that if it returns true, this routine returns without unlocking page 1972 * table lock. So callers must unlock it. 1973 */ 1974 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1975 { 1976 spinlock_t *ptl; 1977 1978 ptl = pud_lock(vma->vm_mm, pud); 1979 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1980 return ptl; 1981 spin_unlock(ptl); 1982 return NULL; 1983 } 1984 1985 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1986 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1987 pud_t *pud, unsigned long addr) 1988 { 1989 pud_t orig_pud; 1990 spinlock_t *ptl; 1991 1992 ptl = __pud_trans_huge_lock(pud, vma); 1993 if (!ptl) 1994 return 0; 1995 /* 1996 * For architectures like ppc64 we look at deposited pgtable 1997 * when calling pudp_huge_get_and_clear. So do the 1998 * pgtable_trans_huge_withdraw after finishing pudp related 1999 * operations. 2000 */ 2001 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 2002 tlb->fullmm); 2003 tlb_remove_pud_tlb_entry(tlb, pud, addr); 2004 if (vma_is_dax(vma)) { 2005 spin_unlock(ptl); 2006 /* No zero page support yet */ 2007 } else { 2008 /* No support for anonymous PUD pages yet */ 2009 BUG(); 2010 } 2011 return 1; 2012 } 2013 2014 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2015 unsigned long haddr) 2016 { 2017 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2018 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2019 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2020 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2021 2022 count_vm_event(THP_SPLIT_PUD); 2023 2024 pudp_huge_clear_flush_notify(vma, haddr, pud); 2025 } 2026 2027 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2028 unsigned long address) 2029 { 2030 spinlock_t *ptl; 2031 struct mm_struct *mm = vma->vm_mm; 2032 unsigned long haddr = address & HPAGE_PUD_MASK; 2033 2034 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); 2035 ptl = pud_lock(mm, pud); 2036 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2037 goto out; 2038 __split_huge_pud_locked(vma, pud, haddr); 2039 2040 out: 2041 spin_unlock(ptl); 2042 /* 2043 * No need to double call mmu_notifier->invalidate_range() callback as 2044 * the above pudp_huge_clear_flush_notify() did already call it. 2045 */ 2046 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2047 HPAGE_PUD_SIZE); 2048 } 2049 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2050 2051 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2052 unsigned long haddr, pmd_t *pmd) 2053 { 2054 struct mm_struct *mm = vma->vm_mm; 2055 pgtable_t pgtable; 2056 pmd_t _pmd; 2057 int i; 2058 2059 /* 2060 * Leave pmd empty until pte is filled note that it is fine to delay 2061 * notification until mmu_notifier_invalidate_range_end() as we are 2062 * replacing a zero pmd write protected page with a zero pte write 2063 * protected page. 2064 * 2065 * See Documentation/vm/mmu_notifier.rst 2066 */ 2067 pmdp_huge_clear_flush(vma, haddr, pmd); 2068 2069 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2070 pmd_populate(mm, &_pmd, pgtable); 2071 2072 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2073 pte_t *pte, entry; 2074 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2075 entry = pte_mkspecial(entry); 2076 pte = pte_offset_map(&_pmd, haddr); 2077 VM_BUG_ON(!pte_none(*pte)); 2078 set_pte_at(mm, haddr, pte, entry); 2079 pte_unmap(pte); 2080 } 2081 smp_wmb(); /* make pte visible before pmd */ 2082 pmd_populate(mm, pmd, pgtable); 2083 } 2084 2085 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2086 unsigned long haddr, bool freeze) 2087 { 2088 struct mm_struct *mm = vma->vm_mm; 2089 struct page *page; 2090 pgtable_t pgtable; 2091 pmd_t old_pmd, _pmd; 2092 bool young, write, soft_dirty, pmd_migration = false; 2093 unsigned long addr; 2094 int i; 2095 2096 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2097 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2098 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2099 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2100 && !pmd_devmap(*pmd)); 2101 2102 count_vm_event(THP_SPLIT_PMD); 2103 2104 if (!vma_is_anonymous(vma)) { 2105 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2106 /* 2107 * We are going to unmap this huge page. So 2108 * just go ahead and zap it 2109 */ 2110 if (arch_needs_pgtable_deposit()) 2111 zap_deposited_table(mm, pmd); 2112 if (vma_is_dax(vma)) 2113 return; 2114 page = pmd_page(_pmd); 2115 if (!PageDirty(page) && pmd_dirty(_pmd)) 2116 set_page_dirty(page); 2117 if (!PageReferenced(page) && pmd_young(_pmd)) 2118 SetPageReferenced(page); 2119 page_remove_rmap(page, true); 2120 put_page(page); 2121 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2122 return; 2123 } else if (is_huge_zero_pmd(*pmd)) { 2124 /* 2125 * FIXME: Do we want to invalidate secondary mmu by calling 2126 * mmu_notifier_invalidate_range() see comments below inside 2127 * __split_huge_pmd() ? 2128 * 2129 * We are going from a zero huge page write protected to zero 2130 * small page also write protected so it does not seems useful 2131 * to invalidate secondary mmu at this time. 2132 */ 2133 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2134 } 2135 2136 /* 2137 * Up to this point the pmd is present and huge and userland has the 2138 * whole access to the hugepage during the split (which happens in 2139 * place). If we overwrite the pmd with the not-huge version pointing 2140 * to the pte here (which of course we could if all CPUs were bug 2141 * free), userland could trigger a small page size TLB miss on the 2142 * small sized TLB while the hugepage TLB entry is still established in 2143 * the huge TLB. Some CPU doesn't like that. 2144 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2145 * 383 on page 93. Intel should be safe but is also warns that it's 2146 * only safe if the permission and cache attributes of the two entries 2147 * loaded in the two TLB is identical (which should be the case here). 2148 * But it is generally safer to never allow small and huge TLB entries 2149 * for the same virtual address to be loaded simultaneously. So instead 2150 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2151 * current pmd notpresent (atomically because here the pmd_trans_huge 2152 * must remain set at all times on the pmd until the split is complete 2153 * for this pmd), then we flush the SMP TLB and finally we write the 2154 * non-huge version of the pmd entry with pmd_populate. 2155 */ 2156 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2157 2158 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2159 pmd_migration = is_pmd_migration_entry(old_pmd); 2160 if (pmd_migration) { 2161 swp_entry_t entry; 2162 2163 entry = pmd_to_swp_entry(old_pmd); 2164 page = pfn_to_page(swp_offset(entry)); 2165 } else 2166 #endif 2167 page = pmd_page(old_pmd); 2168 VM_BUG_ON_PAGE(!page_count(page), page); 2169 page_ref_add(page, HPAGE_PMD_NR - 1); 2170 if (pmd_dirty(old_pmd)) 2171 SetPageDirty(page); 2172 write = pmd_write(old_pmd); 2173 young = pmd_young(old_pmd); 2174 soft_dirty = pmd_soft_dirty(old_pmd); 2175 2176 /* 2177 * Withdraw the table only after we mark the pmd entry invalid. 2178 * This's critical for some architectures (Power). 2179 */ 2180 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2181 pmd_populate(mm, &_pmd, pgtable); 2182 2183 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2184 pte_t entry, *pte; 2185 /* 2186 * Note that NUMA hinting access restrictions are not 2187 * transferred to avoid any possibility of altering 2188 * permissions across VMAs. 2189 */ 2190 if (freeze || pmd_migration) { 2191 swp_entry_t swp_entry; 2192 swp_entry = make_migration_entry(page + i, write); 2193 entry = swp_entry_to_pte(swp_entry); 2194 if (soft_dirty) 2195 entry = pte_swp_mksoft_dirty(entry); 2196 } else { 2197 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2198 entry = maybe_mkwrite(entry, vma); 2199 if (!write) 2200 entry = pte_wrprotect(entry); 2201 if (!young) 2202 entry = pte_mkold(entry); 2203 if (soft_dirty) 2204 entry = pte_mksoft_dirty(entry); 2205 } 2206 pte = pte_offset_map(&_pmd, addr); 2207 BUG_ON(!pte_none(*pte)); 2208 set_pte_at(mm, addr, pte, entry); 2209 atomic_inc(&page[i]._mapcount); 2210 pte_unmap(pte); 2211 } 2212 2213 /* 2214 * Set PG_double_map before dropping compound_mapcount to avoid 2215 * false-negative page_mapped(). 2216 */ 2217 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2218 for (i = 0; i < HPAGE_PMD_NR; i++) 2219 atomic_inc(&page[i]._mapcount); 2220 } 2221 2222 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2223 /* Last compound_mapcount is gone. */ 2224 __dec_node_page_state(page, NR_ANON_THPS); 2225 if (TestClearPageDoubleMap(page)) { 2226 /* No need in mapcount reference anymore */ 2227 for (i = 0; i < HPAGE_PMD_NR; i++) 2228 atomic_dec(&page[i]._mapcount); 2229 } 2230 } 2231 2232 smp_wmb(); /* make pte visible before pmd */ 2233 pmd_populate(mm, pmd, pgtable); 2234 2235 if (freeze) { 2236 for (i = 0; i < HPAGE_PMD_NR; i++) { 2237 page_remove_rmap(page + i, false); 2238 put_page(page + i); 2239 } 2240 } 2241 } 2242 2243 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2244 unsigned long address, bool freeze, struct page *page) 2245 { 2246 spinlock_t *ptl; 2247 struct mm_struct *mm = vma->vm_mm; 2248 unsigned long haddr = address & HPAGE_PMD_MASK; 2249 2250 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 2251 ptl = pmd_lock(mm, pmd); 2252 2253 /* 2254 * If caller asks to setup a migration entries, we need a page to check 2255 * pmd against. Otherwise we can end up replacing wrong page. 2256 */ 2257 VM_BUG_ON(freeze && !page); 2258 if (page && page != pmd_page(*pmd)) 2259 goto out; 2260 2261 if (pmd_trans_huge(*pmd)) { 2262 page = pmd_page(*pmd); 2263 if (PageMlocked(page)) 2264 clear_page_mlock(page); 2265 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2266 goto out; 2267 __split_huge_pmd_locked(vma, pmd, haddr, freeze); 2268 out: 2269 spin_unlock(ptl); 2270 /* 2271 * No need to double call mmu_notifier->invalidate_range() callback. 2272 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2273 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2274 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2275 * fault will trigger a flush_notify before pointing to a new page 2276 * (it is fine if the secondary mmu keeps pointing to the old zero 2277 * page in the meantime) 2278 * 3) Split a huge pmd into pte pointing to the same page. No need 2279 * to invalidate secondary tlb entry they are all still valid. 2280 * any further changes to individual pte will notify. So no need 2281 * to call mmu_notifier->invalidate_range() 2282 */ 2283 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2284 HPAGE_PMD_SIZE); 2285 } 2286 2287 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2288 bool freeze, struct page *page) 2289 { 2290 pgd_t *pgd; 2291 p4d_t *p4d; 2292 pud_t *pud; 2293 pmd_t *pmd; 2294 2295 pgd = pgd_offset(vma->vm_mm, address); 2296 if (!pgd_present(*pgd)) 2297 return; 2298 2299 p4d = p4d_offset(pgd, address); 2300 if (!p4d_present(*p4d)) 2301 return; 2302 2303 pud = pud_offset(p4d, address); 2304 if (!pud_present(*pud)) 2305 return; 2306 2307 pmd = pmd_offset(pud, address); 2308 2309 __split_huge_pmd(vma, pmd, address, freeze, page); 2310 } 2311 2312 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2313 unsigned long start, 2314 unsigned long end, 2315 long adjust_next) 2316 { 2317 /* 2318 * If the new start address isn't hpage aligned and it could 2319 * previously contain an hugepage: check if we need to split 2320 * an huge pmd. 2321 */ 2322 if (start & ~HPAGE_PMD_MASK && 2323 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2324 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2325 split_huge_pmd_address(vma, start, false, NULL); 2326 2327 /* 2328 * If the new end address isn't hpage aligned and it could 2329 * previously contain an hugepage: check if we need to split 2330 * an huge pmd. 2331 */ 2332 if (end & ~HPAGE_PMD_MASK && 2333 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2334 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2335 split_huge_pmd_address(vma, end, false, NULL); 2336 2337 /* 2338 * If we're also updating the vma->vm_next->vm_start, if the new 2339 * vm_next->vm_start isn't page aligned and it could previously 2340 * contain an hugepage: check if we need to split an huge pmd. 2341 */ 2342 if (adjust_next > 0) { 2343 struct vm_area_struct *next = vma->vm_next; 2344 unsigned long nstart = next->vm_start; 2345 nstart += adjust_next << PAGE_SHIFT; 2346 if (nstart & ~HPAGE_PMD_MASK && 2347 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2348 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2349 split_huge_pmd_address(next, nstart, false, NULL); 2350 } 2351 } 2352 2353 static void freeze_page(struct page *page) 2354 { 2355 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2356 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2357 bool unmap_success; 2358 2359 VM_BUG_ON_PAGE(!PageHead(page), page); 2360 2361 if (PageAnon(page)) 2362 ttu_flags |= TTU_SPLIT_FREEZE; 2363 2364 unmap_success = try_to_unmap(page, ttu_flags); 2365 VM_BUG_ON_PAGE(!unmap_success, page); 2366 } 2367 2368 static void unfreeze_page(struct page *page) 2369 { 2370 int i; 2371 if (PageTransHuge(page)) { 2372 remove_migration_ptes(page, page, true); 2373 } else { 2374 for (i = 0; i < HPAGE_PMD_NR; i++) 2375 remove_migration_ptes(page + i, page + i, true); 2376 } 2377 } 2378 2379 static void __split_huge_page_tail(struct page *head, int tail, 2380 struct lruvec *lruvec, struct list_head *list) 2381 { 2382 struct page *page_tail = head + tail; 2383 2384 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2385 2386 /* 2387 * Clone page flags before unfreezing refcount. 2388 * 2389 * After successful get_page_unless_zero() might follow flags change, 2390 * for exmaple lock_page() which set PG_waiters. 2391 */ 2392 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2393 page_tail->flags |= (head->flags & 2394 ((1L << PG_referenced) | 2395 (1L << PG_swapbacked) | 2396 (1L << PG_swapcache) | 2397 (1L << PG_mlocked) | 2398 (1L << PG_uptodate) | 2399 (1L << PG_active) | 2400 (1L << PG_workingset) | 2401 (1L << PG_locked) | 2402 (1L << PG_unevictable) | 2403 (1L << PG_dirty))); 2404 2405 /* Page flags must be visible before we make the page non-compound. */ 2406 smp_wmb(); 2407 2408 /* 2409 * Clear PageTail before unfreezing page refcount. 2410 * 2411 * After successful get_page_unless_zero() might follow put_page() 2412 * which needs correct compound_head(). 2413 */ 2414 clear_compound_head(page_tail); 2415 2416 /* Finally unfreeze refcount. Additional reference from page cache. */ 2417 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2418 PageSwapCache(head))); 2419 2420 if (page_is_young(head)) 2421 set_page_young(page_tail); 2422 if (page_is_idle(head)) 2423 set_page_idle(page_tail); 2424 2425 /* ->mapping in first tail page is compound_mapcount */ 2426 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2427 page_tail); 2428 page_tail->mapping = head->mapping; 2429 2430 page_tail->index = head->index + tail; 2431 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2432 2433 /* 2434 * always add to the tail because some iterators expect new 2435 * pages to show after the currently processed elements - e.g. 2436 * migrate_pages 2437 */ 2438 lru_add_page_tail(head, page_tail, lruvec, list); 2439 } 2440 2441 static void __split_huge_page(struct page *page, struct list_head *list, 2442 unsigned long flags) 2443 { 2444 struct page *head = compound_head(page); 2445 struct zone *zone = page_zone(head); 2446 struct lruvec *lruvec; 2447 pgoff_t end = -1; 2448 int i; 2449 2450 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2451 2452 /* complete memcg works before add pages to LRU */ 2453 mem_cgroup_split_huge_fixup(head); 2454 2455 if (!PageAnon(page)) 2456 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); 2457 2458 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2459 __split_huge_page_tail(head, i, lruvec, list); 2460 /* Some pages can be beyond i_size: drop them from page cache */ 2461 if (head[i].index >= end) { 2462 ClearPageDirty(head + i); 2463 __delete_from_page_cache(head + i, NULL); 2464 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2465 shmem_uncharge(head->mapping->host, 1); 2466 put_page(head + i); 2467 } 2468 } 2469 2470 ClearPageCompound(head); 2471 /* See comment in __split_huge_page_tail() */ 2472 if (PageAnon(head)) { 2473 /* Additional pin to swap cache */ 2474 if (PageSwapCache(head)) 2475 page_ref_add(head, 2); 2476 else 2477 page_ref_inc(head); 2478 } else { 2479 /* Additional pin to page cache */ 2480 page_ref_add(head, 2); 2481 xa_unlock(&head->mapping->i_pages); 2482 } 2483 2484 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2485 2486 unfreeze_page(head); 2487 2488 for (i = 0; i < HPAGE_PMD_NR; i++) { 2489 struct page *subpage = head + i; 2490 if (subpage == page) 2491 continue; 2492 unlock_page(subpage); 2493 2494 /* 2495 * Subpages may be freed if there wasn't any mapping 2496 * like if add_to_swap() is running on a lru page that 2497 * had its mapping zapped. And freeing these pages 2498 * requires taking the lru_lock so we do the put_page 2499 * of the tail pages after the split is complete. 2500 */ 2501 put_page(subpage); 2502 } 2503 } 2504 2505 int total_mapcount(struct page *page) 2506 { 2507 int i, compound, ret; 2508 2509 VM_BUG_ON_PAGE(PageTail(page), page); 2510 2511 if (likely(!PageCompound(page))) 2512 return atomic_read(&page->_mapcount) + 1; 2513 2514 compound = compound_mapcount(page); 2515 if (PageHuge(page)) 2516 return compound; 2517 ret = compound; 2518 for (i = 0; i < HPAGE_PMD_NR; i++) 2519 ret += atomic_read(&page[i]._mapcount) + 1; 2520 /* File pages has compound_mapcount included in _mapcount */ 2521 if (!PageAnon(page)) 2522 return ret - compound * HPAGE_PMD_NR; 2523 if (PageDoubleMap(page)) 2524 ret -= HPAGE_PMD_NR; 2525 return ret; 2526 } 2527 2528 /* 2529 * This calculates accurately how many mappings a transparent hugepage 2530 * has (unlike page_mapcount() which isn't fully accurate). This full 2531 * accuracy is primarily needed to know if copy-on-write faults can 2532 * reuse the page and change the mapping to read-write instead of 2533 * copying them. At the same time this returns the total_mapcount too. 2534 * 2535 * The function returns the highest mapcount any one of the subpages 2536 * has. If the return value is one, even if different processes are 2537 * mapping different subpages of the transparent hugepage, they can 2538 * all reuse it, because each process is reusing a different subpage. 2539 * 2540 * The total_mapcount is instead counting all virtual mappings of the 2541 * subpages. If the total_mapcount is equal to "one", it tells the 2542 * caller all mappings belong to the same "mm" and in turn the 2543 * anon_vma of the transparent hugepage can become the vma->anon_vma 2544 * local one as no other process may be mapping any of the subpages. 2545 * 2546 * It would be more accurate to replace page_mapcount() with 2547 * page_trans_huge_mapcount(), however we only use 2548 * page_trans_huge_mapcount() in the copy-on-write faults where we 2549 * need full accuracy to avoid breaking page pinning, because 2550 * page_trans_huge_mapcount() is slower than page_mapcount(). 2551 */ 2552 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2553 { 2554 int i, ret, _total_mapcount, mapcount; 2555 2556 /* hugetlbfs shouldn't call it */ 2557 VM_BUG_ON_PAGE(PageHuge(page), page); 2558 2559 if (likely(!PageTransCompound(page))) { 2560 mapcount = atomic_read(&page->_mapcount) + 1; 2561 if (total_mapcount) 2562 *total_mapcount = mapcount; 2563 return mapcount; 2564 } 2565 2566 page = compound_head(page); 2567 2568 _total_mapcount = ret = 0; 2569 for (i = 0; i < HPAGE_PMD_NR; i++) { 2570 mapcount = atomic_read(&page[i]._mapcount) + 1; 2571 ret = max(ret, mapcount); 2572 _total_mapcount += mapcount; 2573 } 2574 if (PageDoubleMap(page)) { 2575 ret -= 1; 2576 _total_mapcount -= HPAGE_PMD_NR; 2577 } 2578 mapcount = compound_mapcount(page); 2579 ret += mapcount; 2580 _total_mapcount += mapcount; 2581 if (total_mapcount) 2582 *total_mapcount = _total_mapcount; 2583 return ret; 2584 } 2585 2586 /* Racy check whether the huge page can be split */ 2587 bool can_split_huge_page(struct page *page, int *pextra_pins) 2588 { 2589 int extra_pins; 2590 2591 /* Additional pins from page cache */ 2592 if (PageAnon(page)) 2593 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2594 else 2595 extra_pins = HPAGE_PMD_NR; 2596 if (pextra_pins) 2597 *pextra_pins = extra_pins; 2598 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2599 } 2600 2601 /* 2602 * This function splits huge page into normal pages. @page can point to any 2603 * subpage of huge page to split. Split doesn't change the position of @page. 2604 * 2605 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2606 * The huge page must be locked. 2607 * 2608 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2609 * 2610 * Both head page and tail pages will inherit mapping, flags, and so on from 2611 * the hugepage. 2612 * 2613 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2614 * they are not mapped. 2615 * 2616 * Returns 0 if the hugepage is split successfully. 2617 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2618 * us. 2619 */ 2620 int split_huge_page_to_list(struct page *page, struct list_head *list) 2621 { 2622 struct page *head = compound_head(page); 2623 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2624 struct anon_vma *anon_vma = NULL; 2625 struct address_space *mapping = NULL; 2626 int count, mapcount, extra_pins, ret; 2627 bool mlocked; 2628 unsigned long flags; 2629 2630 VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2631 VM_BUG_ON_PAGE(!PageLocked(page), page); 2632 VM_BUG_ON_PAGE(!PageCompound(page), page); 2633 2634 if (PageWriteback(page)) 2635 return -EBUSY; 2636 2637 if (PageAnon(head)) { 2638 /* 2639 * The caller does not necessarily hold an mmap_sem that would 2640 * prevent the anon_vma disappearing so we first we take a 2641 * reference to it and then lock the anon_vma for write. This 2642 * is similar to page_lock_anon_vma_read except the write lock 2643 * is taken to serialise against parallel split or collapse 2644 * operations. 2645 */ 2646 anon_vma = page_get_anon_vma(head); 2647 if (!anon_vma) { 2648 ret = -EBUSY; 2649 goto out; 2650 } 2651 mapping = NULL; 2652 anon_vma_lock_write(anon_vma); 2653 } else { 2654 mapping = head->mapping; 2655 2656 /* Truncated ? */ 2657 if (!mapping) { 2658 ret = -EBUSY; 2659 goto out; 2660 } 2661 2662 anon_vma = NULL; 2663 i_mmap_lock_read(mapping); 2664 } 2665 2666 /* 2667 * Racy check if we can split the page, before freeze_page() will 2668 * split PMDs 2669 */ 2670 if (!can_split_huge_page(head, &extra_pins)) { 2671 ret = -EBUSY; 2672 goto out_unlock; 2673 } 2674 2675 mlocked = PageMlocked(page); 2676 freeze_page(head); 2677 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2678 2679 /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2680 if (mlocked) 2681 lru_add_drain(); 2682 2683 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2684 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2685 2686 if (mapping) { 2687 XA_STATE(xas, &mapping->i_pages, page_index(head)); 2688 2689 /* 2690 * Check if the head page is present in page cache. 2691 * We assume all tail are present too, if head is there. 2692 */ 2693 xa_lock(&mapping->i_pages); 2694 if (xas_load(&xas) != head) 2695 goto fail; 2696 } 2697 2698 /* Prevent deferred_split_scan() touching ->_refcount */ 2699 spin_lock(&pgdata->split_queue_lock); 2700 count = page_count(head); 2701 mapcount = total_mapcount(head); 2702 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2703 if (!list_empty(page_deferred_list(head))) { 2704 pgdata->split_queue_len--; 2705 list_del(page_deferred_list(head)); 2706 } 2707 if (mapping) 2708 __dec_node_page_state(page, NR_SHMEM_THPS); 2709 spin_unlock(&pgdata->split_queue_lock); 2710 __split_huge_page(page, list, flags); 2711 if (PageSwapCache(head)) { 2712 swp_entry_t entry = { .val = page_private(head) }; 2713 2714 ret = split_swap_cluster(entry); 2715 } else 2716 ret = 0; 2717 } else { 2718 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2719 pr_alert("total_mapcount: %u, page_count(): %u\n", 2720 mapcount, count); 2721 if (PageTail(page)) 2722 dump_page(head, NULL); 2723 dump_page(page, "total_mapcount(head) > 0"); 2724 BUG(); 2725 } 2726 spin_unlock(&pgdata->split_queue_lock); 2727 fail: if (mapping) 2728 xa_unlock(&mapping->i_pages); 2729 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2730 unfreeze_page(head); 2731 ret = -EBUSY; 2732 } 2733 2734 out_unlock: 2735 if (anon_vma) { 2736 anon_vma_unlock_write(anon_vma); 2737 put_anon_vma(anon_vma); 2738 } 2739 if (mapping) 2740 i_mmap_unlock_read(mapping); 2741 out: 2742 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2743 return ret; 2744 } 2745 2746 void free_transhuge_page(struct page *page) 2747 { 2748 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2749 unsigned long flags; 2750 2751 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2752 if (!list_empty(page_deferred_list(page))) { 2753 pgdata->split_queue_len--; 2754 list_del(page_deferred_list(page)); 2755 } 2756 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2757 free_compound_page(page); 2758 } 2759 2760 void deferred_split_huge_page(struct page *page) 2761 { 2762 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2763 unsigned long flags; 2764 2765 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2766 2767 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2768 if (list_empty(page_deferred_list(page))) { 2769 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2770 list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2771 pgdata->split_queue_len++; 2772 } 2773 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2774 } 2775 2776 static unsigned long deferred_split_count(struct shrinker *shrink, 2777 struct shrink_control *sc) 2778 { 2779 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2780 return READ_ONCE(pgdata->split_queue_len); 2781 } 2782 2783 static unsigned long deferred_split_scan(struct shrinker *shrink, 2784 struct shrink_control *sc) 2785 { 2786 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2787 unsigned long flags; 2788 LIST_HEAD(list), *pos, *next; 2789 struct page *page; 2790 int split = 0; 2791 2792 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2793 /* Take pin on all head pages to avoid freeing them under us */ 2794 list_for_each_safe(pos, next, &pgdata->split_queue) { 2795 page = list_entry((void *)pos, struct page, mapping); 2796 page = compound_head(page); 2797 if (get_page_unless_zero(page)) { 2798 list_move(page_deferred_list(page), &list); 2799 } else { 2800 /* We lost race with put_compound_page() */ 2801 list_del_init(page_deferred_list(page)); 2802 pgdata->split_queue_len--; 2803 } 2804 if (!--sc->nr_to_scan) 2805 break; 2806 } 2807 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2808 2809 list_for_each_safe(pos, next, &list) { 2810 page = list_entry((void *)pos, struct page, mapping); 2811 if (!trylock_page(page)) 2812 goto next; 2813 /* split_huge_page() removes page from list on success */ 2814 if (!split_huge_page(page)) 2815 split++; 2816 unlock_page(page); 2817 next: 2818 put_page(page); 2819 } 2820 2821 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2822 list_splice_tail(&list, &pgdata->split_queue); 2823 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2824 2825 /* 2826 * Stop shrinker if we didn't split any page, but the queue is empty. 2827 * This can happen if pages were freed under us. 2828 */ 2829 if (!split && list_empty(&pgdata->split_queue)) 2830 return SHRINK_STOP; 2831 return split; 2832 } 2833 2834 static struct shrinker deferred_split_shrinker = { 2835 .count_objects = deferred_split_count, 2836 .scan_objects = deferred_split_scan, 2837 .seeks = DEFAULT_SEEKS, 2838 .flags = SHRINKER_NUMA_AWARE, 2839 }; 2840 2841 #ifdef CONFIG_DEBUG_FS 2842 static int split_huge_pages_set(void *data, u64 val) 2843 { 2844 struct zone *zone; 2845 struct page *page; 2846 unsigned long pfn, max_zone_pfn; 2847 unsigned long total = 0, split = 0; 2848 2849 if (val != 1) 2850 return -EINVAL; 2851 2852 for_each_populated_zone(zone) { 2853 max_zone_pfn = zone_end_pfn(zone); 2854 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2855 if (!pfn_valid(pfn)) 2856 continue; 2857 2858 page = pfn_to_page(pfn); 2859 if (!get_page_unless_zero(page)) 2860 continue; 2861 2862 if (zone != page_zone(page)) 2863 goto next; 2864 2865 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2866 goto next; 2867 2868 total++; 2869 lock_page(page); 2870 if (!split_huge_page(page)) 2871 split++; 2872 unlock_page(page); 2873 next: 2874 put_page(page); 2875 } 2876 } 2877 2878 pr_info("%lu of %lu THP split\n", split, total); 2879 2880 return 0; 2881 } 2882 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2883 "%llu\n"); 2884 2885 static int __init split_huge_pages_debugfs(void) 2886 { 2887 void *ret; 2888 2889 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2890 &split_huge_pages_fops); 2891 if (!ret) 2892 pr_warn("Failed to create split_huge_pages in debugfs"); 2893 return 0; 2894 } 2895 late_initcall(split_huge_pages_debugfs); 2896 #endif 2897 2898 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2899 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2900 struct page *page) 2901 { 2902 struct vm_area_struct *vma = pvmw->vma; 2903 struct mm_struct *mm = vma->vm_mm; 2904 unsigned long address = pvmw->address; 2905 pmd_t pmdval; 2906 swp_entry_t entry; 2907 pmd_t pmdswp; 2908 2909 if (!(pvmw->pmd && !pvmw->pte)) 2910 return; 2911 2912 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2913 pmdval = *pvmw->pmd; 2914 pmdp_invalidate(vma, address, pvmw->pmd); 2915 if (pmd_dirty(pmdval)) 2916 set_page_dirty(page); 2917 entry = make_migration_entry(page, pmd_write(pmdval)); 2918 pmdswp = swp_entry_to_pmd(entry); 2919 if (pmd_soft_dirty(pmdval)) 2920 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2921 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2922 page_remove_rmap(page, true); 2923 put_page(page); 2924 } 2925 2926 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2927 { 2928 struct vm_area_struct *vma = pvmw->vma; 2929 struct mm_struct *mm = vma->vm_mm; 2930 unsigned long address = pvmw->address; 2931 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2932 pmd_t pmde; 2933 swp_entry_t entry; 2934 2935 if (!(pvmw->pmd && !pvmw->pte)) 2936 return; 2937 2938 entry = pmd_to_swp_entry(*pvmw->pmd); 2939 get_page(new); 2940 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2941 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2942 pmde = pmd_mksoft_dirty(pmde); 2943 if (is_write_migration_entry(entry)) 2944 pmde = maybe_pmd_mkwrite(pmde, vma); 2945 2946 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2947 if (PageAnon(new)) 2948 page_add_anon_rmap(new, vma, mmun_start, true); 2949 else 2950 page_add_file_rmap(new, true); 2951 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2952 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) 2953 mlock_vma_page(new); 2954 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2955 } 2956 #endif 2957