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