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