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 /* Always do synchronous compaction */ 637 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 638 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 639 640 /* Kick kcompactd and fail quickly */ 641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 643 644 /* Synchronous compaction if madvised, otherwise kick kcompactd */ 645 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 646 return GFP_TRANSHUGE_LIGHT | 647 (vma_madvised ? __GFP_DIRECT_RECLAIM : 648 __GFP_KSWAPD_RECLAIM); 649 650 /* Only do synchronous compaction if madvised */ 651 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 652 return GFP_TRANSHUGE_LIGHT | 653 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 654 655 return GFP_TRANSHUGE_LIGHT; 656 } 657 658 /* Caller must hold page table lock. */ 659 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 660 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 661 struct page *zero_page) 662 { 663 pmd_t entry; 664 if (!pmd_none(*pmd)) 665 return false; 666 entry = mk_pmd(zero_page, vma->vm_page_prot); 667 entry = pmd_mkhuge(entry); 668 if (pgtable) 669 pgtable_trans_huge_deposit(mm, pmd, pgtable); 670 set_pmd_at(mm, haddr, pmd, entry); 671 mm_inc_nr_ptes(mm); 672 return true; 673 } 674 675 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 676 { 677 struct vm_area_struct *vma = vmf->vma; 678 gfp_t gfp; 679 struct page *page; 680 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 681 682 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 683 return VM_FAULT_FALLBACK; 684 if (unlikely(anon_vma_prepare(vma))) 685 return VM_FAULT_OOM; 686 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 687 return VM_FAULT_OOM; 688 if (!(vmf->flags & FAULT_FLAG_WRITE) && 689 !mm_forbids_zeropage(vma->vm_mm) && 690 transparent_hugepage_use_zero_page()) { 691 pgtable_t pgtable; 692 struct page *zero_page; 693 bool set; 694 vm_fault_t ret; 695 pgtable = pte_alloc_one(vma->vm_mm, haddr); 696 if (unlikely(!pgtable)) 697 return VM_FAULT_OOM; 698 zero_page = mm_get_huge_zero_page(vma->vm_mm); 699 if (unlikely(!zero_page)) { 700 pte_free(vma->vm_mm, pgtable); 701 count_vm_event(THP_FAULT_FALLBACK); 702 return VM_FAULT_FALLBACK; 703 } 704 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 705 ret = 0; 706 set = false; 707 if (pmd_none(*vmf->pmd)) { 708 ret = check_stable_address_space(vma->vm_mm); 709 if (ret) { 710 spin_unlock(vmf->ptl); 711 } else if (userfaultfd_missing(vma)) { 712 spin_unlock(vmf->ptl); 713 ret = handle_userfault(vmf, VM_UFFD_MISSING); 714 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 715 } else { 716 set_huge_zero_page(pgtable, vma->vm_mm, vma, 717 haddr, vmf->pmd, zero_page); 718 spin_unlock(vmf->ptl); 719 set = true; 720 } 721 } else 722 spin_unlock(vmf->ptl); 723 if (!set) 724 pte_free(vma->vm_mm, pgtable); 725 return ret; 726 } 727 gfp = alloc_hugepage_direct_gfpmask(vma); 728 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 729 if (unlikely(!page)) { 730 count_vm_event(THP_FAULT_FALLBACK); 731 return VM_FAULT_FALLBACK; 732 } 733 prep_transhuge_page(page); 734 return __do_huge_pmd_anonymous_page(vmf, page, gfp); 735 } 736 737 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 738 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 739 pgtable_t pgtable) 740 { 741 struct mm_struct *mm = vma->vm_mm; 742 pmd_t entry; 743 spinlock_t *ptl; 744 745 ptl = pmd_lock(mm, pmd); 746 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 747 if (pfn_t_devmap(pfn)) 748 entry = pmd_mkdevmap(entry); 749 if (write) { 750 entry = pmd_mkyoung(pmd_mkdirty(entry)); 751 entry = maybe_pmd_mkwrite(entry, vma); 752 } 753 754 if (pgtable) { 755 pgtable_trans_huge_deposit(mm, pmd, pgtable); 756 mm_inc_nr_ptes(mm); 757 } 758 759 set_pmd_at(mm, addr, pmd, entry); 760 update_mmu_cache_pmd(vma, addr, pmd); 761 spin_unlock(ptl); 762 } 763 764 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 765 pmd_t *pmd, pfn_t pfn, bool write) 766 { 767 pgprot_t pgprot = vma->vm_page_prot; 768 pgtable_t pgtable = NULL; 769 /* 770 * If we had pmd_special, we could avoid all these restrictions, 771 * but we need to be consistent with PTEs and architectures that 772 * can't support a 'special' bit. 773 */ 774 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 775 !pfn_t_devmap(pfn)); 776 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 777 (VM_PFNMAP|VM_MIXEDMAP)); 778 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 779 780 if (addr < vma->vm_start || addr >= vma->vm_end) 781 return VM_FAULT_SIGBUS; 782 783 if (arch_needs_pgtable_deposit()) { 784 pgtable = pte_alloc_one(vma->vm_mm, addr); 785 if (!pgtable) 786 return VM_FAULT_OOM; 787 } 788 789 track_pfn_insert(vma, &pgprot, pfn); 790 791 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); 792 return VM_FAULT_NOPAGE; 793 } 794 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 795 796 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 797 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 798 { 799 if (likely(vma->vm_flags & VM_WRITE)) 800 pud = pud_mkwrite(pud); 801 return pud; 802 } 803 804 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 805 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) 806 { 807 struct mm_struct *mm = vma->vm_mm; 808 pud_t entry; 809 spinlock_t *ptl; 810 811 ptl = pud_lock(mm, pud); 812 entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 813 if (pfn_t_devmap(pfn)) 814 entry = pud_mkdevmap(entry); 815 if (write) { 816 entry = pud_mkyoung(pud_mkdirty(entry)); 817 entry = maybe_pud_mkwrite(entry, vma); 818 } 819 set_pud_at(mm, addr, pud, entry); 820 update_mmu_cache_pud(vma, addr, pud); 821 spin_unlock(ptl); 822 } 823 824 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 825 pud_t *pud, pfn_t pfn, bool write) 826 { 827 pgprot_t pgprot = vma->vm_page_prot; 828 /* 829 * If we had pud_special, we could avoid all these restrictions, 830 * but we need to be consistent with PTEs and architectures that 831 * can't support a 'special' bit. 832 */ 833 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 834 !pfn_t_devmap(pfn)); 835 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 836 (VM_PFNMAP|VM_MIXEDMAP)); 837 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 838 839 if (addr < vma->vm_start || addr >= vma->vm_end) 840 return VM_FAULT_SIGBUS; 841 842 track_pfn_insert(vma, &pgprot, pfn); 843 844 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); 845 return VM_FAULT_NOPAGE; 846 } 847 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 848 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 849 850 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 851 pmd_t *pmd, int flags) 852 { 853 pmd_t _pmd; 854 855 _pmd = pmd_mkyoung(*pmd); 856 if (flags & FOLL_WRITE) 857 _pmd = pmd_mkdirty(_pmd); 858 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 859 pmd, _pmd, flags & FOLL_WRITE)) 860 update_mmu_cache_pmd(vma, addr, pmd); 861 } 862 863 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 864 pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 865 { 866 unsigned long pfn = pmd_pfn(*pmd); 867 struct mm_struct *mm = vma->vm_mm; 868 struct page *page; 869 870 assert_spin_locked(pmd_lockptr(mm, pmd)); 871 872 /* 873 * When we COW a devmap PMD entry, we split it into PTEs, so we should 874 * not be in this function with `flags & FOLL_COW` set. 875 */ 876 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); 877 878 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 879 return NULL; 880 881 if (pmd_present(*pmd) && pmd_devmap(*pmd)) 882 /* pass */; 883 else 884 return NULL; 885 886 if (flags & FOLL_TOUCH) 887 touch_pmd(vma, addr, pmd, flags); 888 889 /* 890 * device mapped pages can only be returned if the 891 * caller will manage the page reference count. 892 */ 893 if (!(flags & FOLL_GET)) 894 return ERR_PTR(-EEXIST); 895 896 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 897 *pgmap = get_dev_pagemap(pfn, *pgmap); 898 if (!*pgmap) 899 return ERR_PTR(-EFAULT); 900 page = pfn_to_page(pfn); 901 get_page(page); 902 903 return page; 904 } 905 906 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 907 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 908 struct vm_area_struct *vma) 909 { 910 spinlock_t *dst_ptl, *src_ptl; 911 struct page *src_page; 912 pmd_t pmd; 913 pgtable_t pgtable = NULL; 914 int ret = -ENOMEM; 915 916 /* Skip if can be re-fill on fault */ 917 if (!vma_is_anonymous(vma)) 918 return 0; 919 920 pgtable = pte_alloc_one(dst_mm, addr); 921 if (unlikely(!pgtable)) 922 goto out; 923 924 dst_ptl = pmd_lock(dst_mm, dst_pmd); 925 src_ptl = pmd_lockptr(src_mm, src_pmd); 926 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 927 928 ret = -EAGAIN; 929 pmd = *src_pmd; 930 931 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 932 if (unlikely(is_swap_pmd(pmd))) { 933 swp_entry_t entry = pmd_to_swp_entry(pmd); 934 935 VM_BUG_ON(!is_pmd_migration_entry(pmd)); 936 if (is_write_migration_entry(entry)) { 937 make_migration_entry_read(&entry); 938 pmd = swp_entry_to_pmd(entry); 939 if (pmd_swp_soft_dirty(*src_pmd)) 940 pmd = pmd_swp_mksoft_dirty(pmd); 941 set_pmd_at(src_mm, addr, src_pmd, pmd); 942 } 943 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 944 mm_inc_nr_ptes(dst_mm); 945 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 946 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 947 ret = 0; 948 goto out_unlock; 949 } 950 #endif 951 952 if (unlikely(!pmd_trans_huge(pmd))) { 953 pte_free(dst_mm, pgtable); 954 goto out_unlock; 955 } 956 /* 957 * When page table lock is held, the huge zero pmd should not be 958 * under splitting since we don't split the page itself, only pmd to 959 * a page table. 960 */ 961 if (is_huge_zero_pmd(pmd)) { 962 struct page *zero_page; 963 /* 964 * get_huge_zero_page() will never allocate a new page here, 965 * since we already have a zero page to copy. It just takes a 966 * reference. 967 */ 968 zero_page = mm_get_huge_zero_page(dst_mm); 969 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 970 zero_page); 971 ret = 0; 972 goto out_unlock; 973 } 974 975 src_page = pmd_page(pmd); 976 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 977 get_page(src_page); 978 page_dup_rmap(src_page, true); 979 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 980 mm_inc_nr_ptes(dst_mm); 981 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 982 983 pmdp_set_wrprotect(src_mm, addr, src_pmd); 984 pmd = pmd_mkold(pmd_wrprotect(pmd)); 985 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 986 987 ret = 0; 988 out_unlock: 989 spin_unlock(src_ptl); 990 spin_unlock(dst_ptl); 991 out: 992 return ret; 993 } 994 995 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 996 static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 997 pud_t *pud, int flags) 998 { 999 pud_t _pud; 1000 1001 _pud = pud_mkyoung(*pud); 1002 if (flags & FOLL_WRITE) 1003 _pud = pud_mkdirty(_pud); 1004 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 1005 pud, _pud, flags & FOLL_WRITE)) 1006 update_mmu_cache_pud(vma, addr, pud); 1007 } 1008 1009 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1010 pud_t *pud, int flags, struct dev_pagemap **pgmap) 1011 { 1012 unsigned long pfn = pud_pfn(*pud); 1013 struct mm_struct *mm = vma->vm_mm; 1014 struct page *page; 1015 1016 assert_spin_locked(pud_lockptr(mm, pud)); 1017 1018 if (flags & FOLL_WRITE && !pud_write(*pud)) 1019 return NULL; 1020 1021 if (pud_present(*pud) && pud_devmap(*pud)) 1022 /* pass */; 1023 else 1024 return NULL; 1025 1026 if (flags & FOLL_TOUCH) 1027 touch_pud(vma, addr, pud, flags); 1028 1029 /* 1030 * device mapped pages can only be returned if the 1031 * caller will manage the page reference count. 1032 */ 1033 if (!(flags & FOLL_GET)) 1034 return ERR_PTR(-EEXIST); 1035 1036 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1037 *pgmap = get_dev_pagemap(pfn, *pgmap); 1038 if (!*pgmap) 1039 return ERR_PTR(-EFAULT); 1040 page = pfn_to_page(pfn); 1041 get_page(page); 1042 1043 return page; 1044 } 1045 1046 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1047 pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1048 struct vm_area_struct *vma) 1049 { 1050 spinlock_t *dst_ptl, *src_ptl; 1051 pud_t pud; 1052 int ret; 1053 1054 dst_ptl = pud_lock(dst_mm, dst_pud); 1055 src_ptl = pud_lockptr(src_mm, src_pud); 1056 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1057 1058 ret = -EAGAIN; 1059 pud = *src_pud; 1060 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1061 goto out_unlock; 1062 1063 /* 1064 * When page table lock is held, the huge zero pud should not be 1065 * under splitting since we don't split the page itself, only pud to 1066 * a page table. 1067 */ 1068 if (is_huge_zero_pud(pud)) { 1069 /* No huge zero pud yet */ 1070 } 1071 1072 pudp_set_wrprotect(src_mm, addr, src_pud); 1073 pud = pud_mkold(pud_wrprotect(pud)); 1074 set_pud_at(dst_mm, addr, dst_pud, pud); 1075 1076 ret = 0; 1077 out_unlock: 1078 spin_unlock(src_ptl); 1079 spin_unlock(dst_ptl); 1080 return ret; 1081 } 1082 1083 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1084 { 1085 pud_t entry; 1086 unsigned long haddr; 1087 bool write = vmf->flags & FAULT_FLAG_WRITE; 1088 1089 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1090 if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1091 goto unlock; 1092 1093 entry = pud_mkyoung(orig_pud); 1094 if (write) 1095 entry = pud_mkdirty(entry); 1096 haddr = vmf->address & HPAGE_PUD_MASK; 1097 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) 1098 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); 1099 1100 unlock: 1101 spin_unlock(vmf->ptl); 1102 } 1103 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1104 1105 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) 1106 { 1107 pmd_t entry; 1108 unsigned long haddr; 1109 bool write = vmf->flags & FAULT_FLAG_WRITE; 1110 1111 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1112 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1113 goto unlock; 1114 1115 entry = pmd_mkyoung(orig_pmd); 1116 if (write) 1117 entry = pmd_mkdirty(entry); 1118 haddr = vmf->address & HPAGE_PMD_MASK; 1119 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) 1120 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); 1121 1122 unlock: 1123 spin_unlock(vmf->ptl); 1124 } 1125 1126 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, 1127 pmd_t orig_pmd, struct page *page) 1128 { 1129 struct vm_area_struct *vma = vmf->vma; 1130 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1131 struct mem_cgroup *memcg; 1132 pgtable_t pgtable; 1133 pmd_t _pmd; 1134 int i; 1135 vm_fault_t ret = 0; 1136 struct page **pages; 1137 unsigned long mmun_start; /* For mmu_notifiers */ 1138 unsigned long mmun_end; /* For mmu_notifiers */ 1139 1140 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *), 1141 GFP_KERNEL); 1142 if (unlikely(!pages)) { 1143 ret |= VM_FAULT_OOM; 1144 goto out; 1145 } 1146 1147 for (i = 0; i < HPAGE_PMD_NR; i++) { 1148 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, 1149 vmf->address, page_to_nid(page)); 1150 if (unlikely(!pages[i] || 1151 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm, 1152 GFP_KERNEL, &memcg, false))) { 1153 if (pages[i]) 1154 put_page(pages[i]); 1155 while (--i >= 0) { 1156 memcg = (void *)page_private(pages[i]); 1157 set_page_private(pages[i], 0); 1158 mem_cgroup_cancel_charge(pages[i], memcg, 1159 false); 1160 put_page(pages[i]); 1161 } 1162 kfree(pages); 1163 ret |= VM_FAULT_OOM; 1164 goto out; 1165 } 1166 set_page_private(pages[i], (unsigned long)memcg); 1167 } 1168 1169 for (i = 0; i < HPAGE_PMD_NR; i++) { 1170 copy_user_highpage(pages[i], page + i, 1171 haddr + PAGE_SIZE * i, vma); 1172 __SetPageUptodate(pages[i]); 1173 cond_resched(); 1174 } 1175 1176 mmun_start = haddr; 1177 mmun_end = haddr + HPAGE_PMD_SIZE; 1178 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1179 1180 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1181 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1182 goto out_free_pages; 1183 VM_BUG_ON_PAGE(!PageHead(page), page); 1184 1185 /* 1186 * Leave pmd empty until pte is filled note we must notify here as 1187 * concurrent CPU thread might write to new page before the call to 1188 * mmu_notifier_invalidate_range_end() happens which can lead to a 1189 * device seeing memory write in different order than CPU. 1190 * 1191 * See Documentation/vm/mmu_notifier.rst 1192 */ 1193 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1194 1195 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); 1196 pmd_populate(vma->vm_mm, &_pmd, pgtable); 1197 1198 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1199 pte_t entry; 1200 entry = mk_pte(pages[i], vma->vm_page_prot); 1201 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1202 memcg = (void *)page_private(pages[i]); 1203 set_page_private(pages[i], 0); 1204 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); 1205 mem_cgroup_commit_charge(pages[i], memcg, false, false); 1206 lru_cache_add_active_or_unevictable(pages[i], vma); 1207 vmf->pte = pte_offset_map(&_pmd, haddr); 1208 VM_BUG_ON(!pte_none(*vmf->pte)); 1209 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); 1210 pte_unmap(vmf->pte); 1211 } 1212 kfree(pages); 1213 1214 smp_wmb(); /* make pte visible before pmd */ 1215 pmd_populate(vma->vm_mm, vmf->pmd, pgtable); 1216 page_remove_rmap(page, true); 1217 spin_unlock(vmf->ptl); 1218 1219 /* 1220 * No need to double call mmu_notifier->invalidate_range() callback as 1221 * the above pmdp_huge_clear_flush_notify() did already call it. 1222 */ 1223 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1224 mmun_end); 1225 1226 ret |= VM_FAULT_WRITE; 1227 put_page(page); 1228 1229 out: 1230 return ret; 1231 1232 out_free_pages: 1233 spin_unlock(vmf->ptl); 1234 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); 1235 for (i = 0; i < HPAGE_PMD_NR; i++) { 1236 memcg = (void *)page_private(pages[i]); 1237 set_page_private(pages[i], 0); 1238 mem_cgroup_cancel_charge(pages[i], memcg, false); 1239 put_page(pages[i]); 1240 } 1241 kfree(pages); 1242 goto out; 1243 } 1244 1245 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) 1246 { 1247 struct vm_area_struct *vma = vmf->vma; 1248 struct page *page = NULL, *new_page; 1249 struct mem_cgroup *memcg; 1250 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1251 unsigned long mmun_start; /* For mmu_notifiers */ 1252 unsigned long mmun_end; /* For mmu_notifiers */ 1253 gfp_t huge_gfp; /* for allocation and charge */ 1254 vm_fault_t ret = 0; 1255 1256 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 1257 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1258 if (is_huge_zero_pmd(orig_pmd)) 1259 goto alloc; 1260 spin_lock(vmf->ptl); 1261 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) 1262 goto out_unlock; 1263 1264 page = pmd_page(orig_pmd); 1265 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1266 /* 1267 * We can only reuse the page if nobody else maps the huge page or it's 1268 * part. 1269 */ 1270 if (!trylock_page(page)) { 1271 get_page(page); 1272 spin_unlock(vmf->ptl); 1273 lock_page(page); 1274 spin_lock(vmf->ptl); 1275 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1276 unlock_page(page); 1277 put_page(page); 1278 goto out_unlock; 1279 } 1280 put_page(page); 1281 } 1282 if (reuse_swap_page(page, NULL)) { 1283 pmd_t entry; 1284 entry = pmd_mkyoung(orig_pmd); 1285 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1286 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 1287 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1288 ret |= VM_FAULT_WRITE; 1289 unlock_page(page); 1290 goto out_unlock; 1291 } 1292 unlock_page(page); 1293 get_page(page); 1294 spin_unlock(vmf->ptl); 1295 alloc: 1296 if (transparent_hugepage_enabled(vma) && 1297 !transparent_hugepage_debug_cow()) { 1298 huge_gfp = alloc_hugepage_direct_gfpmask(vma); 1299 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1300 } else 1301 new_page = NULL; 1302 1303 if (likely(new_page)) { 1304 prep_transhuge_page(new_page); 1305 } else { 1306 if (!page) { 1307 split_huge_pmd(vma, vmf->pmd, vmf->address); 1308 ret |= VM_FAULT_FALLBACK; 1309 } else { 1310 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); 1311 if (ret & VM_FAULT_OOM) { 1312 split_huge_pmd(vma, vmf->pmd, vmf->address); 1313 ret |= VM_FAULT_FALLBACK; 1314 } 1315 put_page(page); 1316 } 1317 count_vm_event(THP_FAULT_FALLBACK); 1318 goto out; 1319 } 1320 1321 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm, 1322 huge_gfp, &memcg, true))) { 1323 put_page(new_page); 1324 split_huge_pmd(vma, vmf->pmd, vmf->address); 1325 if (page) 1326 put_page(page); 1327 ret |= VM_FAULT_FALLBACK; 1328 count_vm_event(THP_FAULT_FALLBACK); 1329 goto out; 1330 } 1331 1332 count_vm_event(THP_FAULT_ALLOC); 1333 1334 if (!page) 1335 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); 1336 else 1337 copy_user_huge_page(new_page, page, vmf->address, 1338 vma, HPAGE_PMD_NR); 1339 __SetPageUptodate(new_page); 1340 1341 mmun_start = haddr; 1342 mmun_end = haddr + HPAGE_PMD_SIZE; 1343 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); 1344 1345 spin_lock(vmf->ptl); 1346 if (page) 1347 put_page(page); 1348 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 1349 spin_unlock(vmf->ptl); 1350 mem_cgroup_cancel_charge(new_page, memcg, true); 1351 put_page(new_page); 1352 goto out_mn; 1353 } else { 1354 pmd_t entry; 1355 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1356 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1357 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); 1358 page_add_new_anon_rmap(new_page, vma, haddr, true); 1359 mem_cgroup_commit_charge(new_page, memcg, false, true); 1360 lru_cache_add_active_or_unevictable(new_page, vma); 1361 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 1362 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1363 if (!page) { 1364 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1365 } else { 1366 VM_BUG_ON_PAGE(!PageHead(page), page); 1367 page_remove_rmap(page, true); 1368 put_page(page); 1369 } 1370 ret |= VM_FAULT_WRITE; 1371 } 1372 spin_unlock(vmf->ptl); 1373 out_mn: 1374 /* 1375 * No need to double call mmu_notifier->invalidate_range() callback as 1376 * the above pmdp_huge_clear_flush_notify() did already call it. 1377 */ 1378 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, 1379 mmun_end); 1380 out: 1381 return ret; 1382 out_unlock: 1383 spin_unlock(vmf->ptl); 1384 return ret; 1385 } 1386 1387 /* 1388 * FOLL_FORCE can write to even unwritable pmd's, but only 1389 * after we've gone through a COW cycle and they are dirty. 1390 */ 1391 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) 1392 { 1393 return pmd_write(pmd) || 1394 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); 1395 } 1396 1397 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1398 unsigned long addr, 1399 pmd_t *pmd, 1400 unsigned int flags) 1401 { 1402 struct mm_struct *mm = vma->vm_mm; 1403 struct page *page = NULL; 1404 1405 assert_spin_locked(pmd_lockptr(mm, pmd)); 1406 1407 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) 1408 goto out; 1409 1410 /* Avoid dumping huge zero page */ 1411 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1412 return ERR_PTR(-EFAULT); 1413 1414 /* Full NUMA hinting faults to serialise migration in fault paths */ 1415 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1416 goto out; 1417 1418 page = pmd_page(*pmd); 1419 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 1420 if (flags & FOLL_TOUCH) 1421 touch_pmd(vma, addr, pmd, flags); 1422 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1423 /* 1424 * We don't mlock() pte-mapped THPs. This way we can avoid 1425 * leaking mlocked pages into non-VM_LOCKED VMAs. 1426 * 1427 * For anon THP: 1428 * 1429 * In most cases the pmd is the only mapping of the page as we 1430 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for 1431 * writable private mappings in populate_vma_page_range(). 1432 * 1433 * The only scenario when we have the page shared here is if we 1434 * mlocking read-only mapping shared over fork(). We skip 1435 * mlocking such pages. 1436 * 1437 * For file THP: 1438 * 1439 * We can expect PageDoubleMap() to be stable under page lock: 1440 * for file pages we set it in page_add_file_rmap(), which 1441 * requires page to be locked. 1442 */ 1443 1444 if (PageAnon(page) && compound_mapcount(page) != 1) 1445 goto skip_mlock; 1446 if (PageDoubleMap(page) || !page->mapping) 1447 goto skip_mlock; 1448 if (!trylock_page(page)) 1449 goto skip_mlock; 1450 lru_add_drain(); 1451 if (page->mapping && !PageDoubleMap(page)) 1452 mlock_vma_page(page); 1453 unlock_page(page); 1454 } 1455 skip_mlock: 1456 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1457 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 1458 if (flags & FOLL_GET) 1459 get_page(page); 1460 1461 out: 1462 return page; 1463 } 1464 1465 /* NUMA hinting page fault entry point for trans huge pmds */ 1466 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) 1467 { 1468 struct vm_area_struct *vma = vmf->vma; 1469 struct anon_vma *anon_vma = NULL; 1470 struct page *page; 1471 unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1472 int page_nid = -1, this_nid = numa_node_id(); 1473 int target_nid, last_cpupid = -1; 1474 bool page_locked; 1475 bool migrated = false; 1476 bool was_writable; 1477 int flags = 0; 1478 1479 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1480 if (unlikely(!pmd_same(pmd, *vmf->pmd))) 1481 goto out_unlock; 1482 1483 /* 1484 * If there are potential migrations, wait for completion and retry 1485 * without disrupting NUMA hinting information. Do not relock and 1486 * check_same as the page may no longer be mapped. 1487 */ 1488 if (unlikely(pmd_trans_migrating(*vmf->pmd))) { 1489 page = pmd_page(*vmf->pmd); 1490 if (!get_page_unless_zero(page)) 1491 goto out_unlock; 1492 spin_unlock(vmf->ptl); 1493 wait_on_page_locked(page); 1494 put_page(page); 1495 goto out; 1496 } 1497 1498 page = pmd_page(pmd); 1499 BUG_ON(is_huge_zero_page(page)); 1500 page_nid = page_to_nid(page); 1501 last_cpupid = page_cpupid_last(page); 1502 count_vm_numa_event(NUMA_HINT_FAULTS); 1503 if (page_nid == this_nid) { 1504 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1505 flags |= TNF_FAULT_LOCAL; 1506 } 1507 1508 /* See similar comment in do_numa_page for explanation */ 1509 if (!pmd_savedwrite(pmd)) 1510 flags |= TNF_NO_GROUP; 1511 1512 /* 1513 * Acquire the page lock to serialise THP migrations but avoid dropping 1514 * page_table_lock if at all possible 1515 */ 1516 page_locked = trylock_page(page); 1517 target_nid = mpol_misplaced(page, vma, haddr); 1518 if (target_nid == -1) { 1519 /* If the page was locked, there are no parallel migrations */ 1520 if (page_locked) 1521 goto clear_pmdnuma; 1522 } 1523 1524 /* Migration could have started since the pmd_trans_migrating check */ 1525 if (!page_locked) { 1526 page_nid = -1; 1527 if (!get_page_unless_zero(page)) 1528 goto out_unlock; 1529 spin_unlock(vmf->ptl); 1530 wait_on_page_locked(page); 1531 put_page(page); 1532 goto out; 1533 } 1534 1535 /* 1536 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1537 * to serialises splits 1538 */ 1539 get_page(page); 1540 spin_unlock(vmf->ptl); 1541 anon_vma = page_lock_anon_vma_read(page); 1542 1543 /* Confirm the PMD did not change while page_table_lock was released */ 1544 spin_lock(vmf->ptl); 1545 if (unlikely(!pmd_same(pmd, *vmf->pmd))) { 1546 unlock_page(page); 1547 put_page(page); 1548 page_nid = -1; 1549 goto out_unlock; 1550 } 1551 1552 /* Bail if we fail to protect against THP splits for any reason */ 1553 if (unlikely(!anon_vma)) { 1554 put_page(page); 1555 page_nid = -1; 1556 goto clear_pmdnuma; 1557 } 1558 1559 /* 1560 * Since we took the NUMA fault, we must have observed the !accessible 1561 * bit. Make sure all other CPUs agree with that, to avoid them 1562 * modifying the page we're about to migrate. 1563 * 1564 * Must be done under PTL such that we'll observe the relevant 1565 * inc_tlb_flush_pending(). 1566 * 1567 * We are not sure a pending tlb flush here is for a huge page 1568 * mapping or not. Hence use the tlb range variant 1569 */ 1570 if (mm_tlb_flush_pending(vma->vm_mm)) { 1571 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); 1572 /* 1573 * change_huge_pmd() released the pmd lock before 1574 * invalidating the secondary MMUs sharing the primary 1575 * MMU pagetables (with ->invalidate_range()). The 1576 * mmu_notifier_invalidate_range_end() (which 1577 * internally calls ->invalidate_range()) in 1578 * change_pmd_range() will run after us, so we can't 1579 * rely on it here and we need an explicit invalidate. 1580 */ 1581 mmu_notifier_invalidate_range(vma->vm_mm, haddr, 1582 haddr + HPAGE_PMD_SIZE); 1583 } 1584 1585 /* 1586 * Migrate the THP to the requested node, returns with page unlocked 1587 * and access rights restored. 1588 */ 1589 spin_unlock(vmf->ptl); 1590 1591 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, 1592 vmf->pmd, pmd, vmf->address, page, target_nid); 1593 if (migrated) { 1594 flags |= TNF_MIGRATED; 1595 page_nid = target_nid; 1596 } else 1597 flags |= TNF_MIGRATE_FAIL; 1598 1599 goto out; 1600 clear_pmdnuma: 1601 BUG_ON(!PageLocked(page)); 1602 was_writable = pmd_savedwrite(pmd); 1603 pmd = pmd_modify(pmd, vma->vm_page_prot); 1604 pmd = pmd_mkyoung(pmd); 1605 if (was_writable) 1606 pmd = pmd_mkwrite(pmd); 1607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1608 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1609 unlock_page(page); 1610 out_unlock: 1611 spin_unlock(vmf->ptl); 1612 1613 out: 1614 if (anon_vma) 1615 page_unlock_anon_vma_read(anon_vma); 1616 1617 if (page_nid != -1) 1618 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, 1619 flags); 1620 1621 return 0; 1622 } 1623 1624 /* 1625 * Return true if we do MADV_FREE successfully on entire pmd page. 1626 * Otherwise, return false. 1627 */ 1628 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1629 pmd_t *pmd, unsigned long addr, unsigned long next) 1630 { 1631 spinlock_t *ptl; 1632 pmd_t orig_pmd; 1633 struct page *page; 1634 struct mm_struct *mm = tlb->mm; 1635 bool ret = false; 1636 1637 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1638 1639 ptl = pmd_trans_huge_lock(pmd, vma); 1640 if (!ptl) 1641 goto out_unlocked; 1642 1643 orig_pmd = *pmd; 1644 if (is_huge_zero_pmd(orig_pmd)) 1645 goto out; 1646 1647 if (unlikely(!pmd_present(orig_pmd))) { 1648 VM_BUG_ON(thp_migration_supported() && 1649 !is_pmd_migration_entry(orig_pmd)); 1650 goto out; 1651 } 1652 1653 page = pmd_page(orig_pmd); 1654 /* 1655 * If other processes are mapping this page, we couldn't discard 1656 * the page unless they all do MADV_FREE so let's skip the page. 1657 */ 1658 if (page_mapcount(page) != 1) 1659 goto out; 1660 1661 if (!trylock_page(page)) 1662 goto out; 1663 1664 /* 1665 * If user want to discard part-pages of THP, split it so MADV_FREE 1666 * will deactivate only them. 1667 */ 1668 if (next - addr != HPAGE_PMD_SIZE) { 1669 get_page(page); 1670 spin_unlock(ptl); 1671 split_huge_page(page); 1672 unlock_page(page); 1673 put_page(page); 1674 goto out_unlocked; 1675 } 1676 1677 if (PageDirty(page)) 1678 ClearPageDirty(page); 1679 unlock_page(page); 1680 1681 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 1682 pmdp_invalidate(vma, addr, pmd); 1683 orig_pmd = pmd_mkold(orig_pmd); 1684 orig_pmd = pmd_mkclean(orig_pmd); 1685 1686 set_pmd_at(mm, addr, pmd, orig_pmd); 1687 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1688 } 1689 1690 mark_page_lazyfree(page); 1691 ret = true; 1692 out: 1693 spin_unlock(ptl); 1694 out_unlocked: 1695 return ret; 1696 } 1697 1698 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1699 { 1700 pgtable_t pgtable; 1701 1702 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1703 pte_free(mm, pgtable); 1704 mm_dec_nr_ptes(mm); 1705 } 1706 1707 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1708 pmd_t *pmd, unsigned long addr) 1709 { 1710 pmd_t orig_pmd; 1711 spinlock_t *ptl; 1712 1713 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); 1714 1715 ptl = __pmd_trans_huge_lock(pmd, vma); 1716 if (!ptl) 1717 return 0; 1718 /* 1719 * For architectures like ppc64 we look at deposited pgtable 1720 * when calling pmdp_huge_get_and_clear. So do the 1721 * pgtable_trans_huge_withdraw after finishing pmdp related 1722 * operations. 1723 */ 1724 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1725 tlb->fullmm); 1726 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1727 if (vma_is_dax(vma)) { 1728 if (arch_needs_pgtable_deposit()) 1729 zap_deposited_table(tlb->mm, pmd); 1730 spin_unlock(ptl); 1731 if (is_huge_zero_pmd(orig_pmd)) 1732 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1733 } else if (is_huge_zero_pmd(orig_pmd)) { 1734 zap_deposited_table(tlb->mm, pmd); 1735 spin_unlock(ptl); 1736 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); 1737 } else { 1738 struct page *page = NULL; 1739 int flush_needed = 1; 1740 1741 if (pmd_present(orig_pmd)) { 1742 page = pmd_page(orig_pmd); 1743 page_remove_rmap(page, true); 1744 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1745 VM_BUG_ON_PAGE(!PageHead(page), page); 1746 } else if (thp_migration_supported()) { 1747 swp_entry_t entry; 1748 1749 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1750 entry = pmd_to_swp_entry(orig_pmd); 1751 page = pfn_to_page(swp_offset(entry)); 1752 flush_needed = 0; 1753 } else 1754 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1755 1756 if (PageAnon(page)) { 1757 zap_deposited_table(tlb->mm, pmd); 1758 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1759 } else { 1760 if (arch_needs_pgtable_deposit()) 1761 zap_deposited_table(tlb->mm, pmd); 1762 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1763 } 1764 1765 spin_unlock(ptl); 1766 if (flush_needed) 1767 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1768 } 1769 return 1; 1770 } 1771 1772 #ifndef pmd_move_must_withdraw 1773 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 1774 spinlock_t *old_pmd_ptl, 1775 struct vm_area_struct *vma) 1776 { 1777 /* 1778 * With split pmd lock we also need to move preallocated 1779 * PTE page table if new_pmd is on different PMD page table. 1780 * 1781 * We also don't deposit and withdraw tables for file pages. 1782 */ 1783 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 1784 } 1785 #endif 1786 1787 static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1788 { 1789 #ifdef CONFIG_MEM_SOFT_DIRTY 1790 if (unlikely(is_pmd_migration_entry(pmd))) 1791 pmd = pmd_swp_mksoft_dirty(pmd); 1792 else if (pmd_present(pmd)) 1793 pmd = pmd_mksoft_dirty(pmd); 1794 #endif 1795 return pmd; 1796 } 1797 1798 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1799 unsigned long new_addr, unsigned long old_end, 1800 pmd_t *old_pmd, pmd_t *new_pmd) 1801 { 1802 spinlock_t *old_ptl, *new_ptl; 1803 pmd_t pmd; 1804 struct mm_struct *mm = vma->vm_mm; 1805 bool force_flush = false; 1806 1807 if ((old_addr & ~HPAGE_PMD_MASK) || 1808 (new_addr & ~HPAGE_PMD_MASK) || 1809 old_end - old_addr < HPAGE_PMD_SIZE) 1810 return false; 1811 1812 /* 1813 * The destination pmd shouldn't be established, free_pgtables() 1814 * should have release it. 1815 */ 1816 if (WARN_ON(!pmd_none(*new_pmd))) { 1817 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1818 return false; 1819 } 1820 1821 /* 1822 * We don't have to worry about the ordering of src and dst 1823 * ptlocks because exclusive mmap_sem prevents deadlock. 1824 */ 1825 old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1826 if (old_ptl) { 1827 new_ptl = pmd_lockptr(mm, new_pmd); 1828 if (new_ptl != old_ptl) 1829 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1830 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1831 if (pmd_present(pmd)) 1832 force_flush = true; 1833 VM_BUG_ON(!pmd_none(*new_pmd)); 1834 1835 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 1836 pgtable_t pgtable; 1837 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1838 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1839 } 1840 pmd = move_soft_dirty_pmd(pmd); 1841 set_pmd_at(mm, new_addr, new_pmd, pmd); 1842 if (force_flush) 1843 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 1844 if (new_ptl != old_ptl) 1845 spin_unlock(new_ptl); 1846 spin_unlock(old_ptl); 1847 return true; 1848 } 1849 return false; 1850 } 1851 1852 /* 1853 * Returns 1854 * - 0 if PMD could not be locked 1855 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1856 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1857 */ 1858 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1859 unsigned long addr, pgprot_t newprot, int prot_numa) 1860 { 1861 struct mm_struct *mm = vma->vm_mm; 1862 spinlock_t *ptl; 1863 pmd_t entry; 1864 bool preserve_write; 1865 int ret; 1866 1867 ptl = __pmd_trans_huge_lock(pmd, vma); 1868 if (!ptl) 1869 return 0; 1870 1871 preserve_write = prot_numa && pmd_write(*pmd); 1872 ret = 1; 1873 1874 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1875 if (is_swap_pmd(*pmd)) { 1876 swp_entry_t entry = pmd_to_swp_entry(*pmd); 1877 1878 VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 1879 if (is_write_migration_entry(entry)) { 1880 pmd_t newpmd; 1881 /* 1882 * A protection check is difficult so 1883 * just be safe and disable write 1884 */ 1885 make_migration_entry_read(&entry); 1886 newpmd = swp_entry_to_pmd(entry); 1887 if (pmd_swp_soft_dirty(*pmd)) 1888 newpmd = pmd_swp_mksoft_dirty(newpmd); 1889 set_pmd_at(mm, addr, pmd, newpmd); 1890 } 1891 goto unlock; 1892 } 1893 #endif 1894 1895 /* 1896 * Avoid trapping faults against the zero page. The read-only 1897 * data is likely to be read-cached on the local CPU and 1898 * local/remote hits to the zero page are not interesting. 1899 */ 1900 if (prot_numa && is_huge_zero_pmd(*pmd)) 1901 goto unlock; 1902 1903 if (prot_numa && pmd_protnone(*pmd)) 1904 goto unlock; 1905 1906 /* 1907 * In case prot_numa, we are under down_read(mmap_sem). It's critical 1908 * to not clear pmd intermittently to avoid race with MADV_DONTNEED 1909 * which is also under down_read(mmap_sem): 1910 * 1911 * CPU0: CPU1: 1912 * change_huge_pmd(prot_numa=1) 1913 * pmdp_huge_get_and_clear_notify() 1914 * madvise_dontneed() 1915 * zap_pmd_range() 1916 * pmd_trans_huge(*pmd) == 0 (without ptl) 1917 * // skip the pmd 1918 * set_pmd_at(); 1919 * // pmd is re-established 1920 * 1921 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 1922 * which may break userspace. 1923 * 1924 * pmdp_invalidate() is required to make sure we don't miss 1925 * dirty/young flags set by hardware. 1926 */ 1927 entry = pmdp_invalidate(vma, addr, pmd); 1928 1929 entry = pmd_modify(entry, newprot); 1930 if (preserve_write) 1931 entry = pmd_mk_savedwrite(entry); 1932 ret = HPAGE_PMD_NR; 1933 set_pmd_at(mm, addr, pmd, entry); 1934 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); 1935 unlock: 1936 spin_unlock(ptl); 1937 return ret; 1938 } 1939 1940 /* 1941 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 1942 * 1943 * Note that if it returns page table lock pointer, this routine returns without 1944 * unlocking page table lock. So callers must unlock it. 1945 */ 1946 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1947 { 1948 spinlock_t *ptl; 1949 ptl = pmd_lock(vma->vm_mm, pmd); 1950 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 1951 pmd_devmap(*pmd))) 1952 return ptl; 1953 spin_unlock(ptl); 1954 return NULL; 1955 } 1956 1957 /* 1958 * Returns true if a given pud maps a thp, false otherwise. 1959 * 1960 * Note that if it returns true, this routine returns without unlocking page 1961 * table lock. So callers must unlock it. 1962 */ 1963 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 1964 { 1965 spinlock_t *ptl; 1966 1967 ptl = pud_lock(vma->vm_mm, pud); 1968 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 1969 return ptl; 1970 spin_unlock(ptl); 1971 return NULL; 1972 } 1973 1974 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1975 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 1976 pud_t *pud, unsigned long addr) 1977 { 1978 pud_t orig_pud; 1979 spinlock_t *ptl; 1980 1981 ptl = __pud_trans_huge_lock(pud, vma); 1982 if (!ptl) 1983 return 0; 1984 /* 1985 * For architectures like ppc64 we look at deposited pgtable 1986 * when calling pudp_huge_get_and_clear. So do the 1987 * pgtable_trans_huge_withdraw after finishing pudp related 1988 * operations. 1989 */ 1990 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, 1991 tlb->fullmm); 1992 tlb_remove_pud_tlb_entry(tlb, pud, addr); 1993 if (vma_is_dax(vma)) { 1994 spin_unlock(ptl); 1995 /* No zero page support yet */ 1996 } else { 1997 /* No support for anonymous PUD pages yet */ 1998 BUG(); 1999 } 2000 return 1; 2001 } 2002 2003 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2004 unsigned long haddr) 2005 { 2006 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2007 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2008 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2009 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2010 2011 count_vm_event(THP_SPLIT_PUD); 2012 2013 pudp_huge_clear_flush_notify(vma, haddr, pud); 2014 } 2015 2016 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2017 unsigned long address) 2018 { 2019 spinlock_t *ptl; 2020 struct mm_struct *mm = vma->vm_mm; 2021 unsigned long haddr = address & HPAGE_PUD_MASK; 2022 2023 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); 2024 ptl = pud_lock(mm, pud); 2025 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2026 goto out; 2027 __split_huge_pud_locked(vma, pud, haddr); 2028 2029 out: 2030 spin_unlock(ptl); 2031 /* 2032 * No need to double call mmu_notifier->invalidate_range() callback as 2033 * the above pudp_huge_clear_flush_notify() did already call it. 2034 */ 2035 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2036 HPAGE_PUD_SIZE); 2037 } 2038 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2039 2040 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2041 unsigned long haddr, pmd_t *pmd) 2042 { 2043 struct mm_struct *mm = vma->vm_mm; 2044 pgtable_t pgtable; 2045 pmd_t _pmd; 2046 int i; 2047 2048 /* 2049 * Leave pmd empty until pte is filled note that it is fine to delay 2050 * notification until mmu_notifier_invalidate_range_end() as we are 2051 * replacing a zero pmd write protected page with a zero pte write 2052 * protected page. 2053 * 2054 * See Documentation/vm/mmu_notifier.rst 2055 */ 2056 pmdp_huge_clear_flush(vma, haddr, pmd); 2057 2058 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2059 pmd_populate(mm, &_pmd, pgtable); 2060 2061 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2062 pte_t *pte, entry; 2063 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2064 entry = pte_mkspecial(entry); 2065 pte = pte_offset_map(&_pmd, haddr); 2066 VM_BUG_ON(!pte_none(*pte)); 2067 set_pte_at(mm, haddr, pte, entry); 2068 pte_unmap(pte); 2069 } 2070 smp_wmb(); /* make pte visible before pmd */ 2071 pmd_populate(mm, pmd, pgtable); 2072 } 2073 2074 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2075 unsigned long haddr, bool freeze) 2076 { 2077 struct mm_struct *mm = vma->vm_mm; 2078 struct page *page; 2079 pgtable_t pgtable; 2080 pmd_t old_pmd, _pmd; 2081 bool young, write, soft_dirty, pmd_migration = false; 2082 unsigned long addr; 2083 int i; 2084 2085 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2086 VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2087 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 2088 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 2089 && !pmd_devmap(*pmd)); 2090 2091 count_vm_event(THP_SPLIT_PMD); 2092 2093 if (!vma_is_anonymous(vma)) { 2094 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2095 /* 2096 * We are going to unmap this huge page. So 2097 * just go ahead and zap it 2098 */ 2099 if (arch_needs_pgtable_deposit()) 2100 zap_deposited_table(mm, pmd); 2101 if (vma_is_dax(vma)) 2102 return; 2103 page = pmd_page(_pmd); 2104 if (!PageDirty(page) && pmd_dirty(_pmd)) 2105 set_page_dirty(page); 2106 if (!PageReferenced(page) && pmd_young(_pmd)) 2107 SetPageReferenced(page); 2108 page_remove_rmap(page, true); 2109 put_page(page); 2110 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2111 return; 2112 } else if (is_huge_zero_pmd(*pmd)) { 2113 /* 2114 * FIXME: Do we want to invalidate secondary mmu by calling 2115 * mmu_notifier_invalidate_range() see comments below inside 2116 * __split_huge_pmd() ? 2117 * 2118 * We are going from a zero huge page write protected to zero 2119 * small page also write protected so it does not seems useful 2120 * to invalidate secondary mmu at this time. 2121 */ 2122 return __split_huge_zero_page_pmd(vma, haddr, pmd); 2123 } 2124 2125 /* 2126 * Up to this point the pmd is present and huge and userland has the 2127 * whole access to the hugepage during the split (which happens in 2128 * place). If we overwrite the pmd with the not-huge version pointing 2129 * to the pte here (which of course we could if all CPUs were bug 2130 * free), userland could trigger a small page size TLB miss on the 2131 * small sized TLB while the hugepage TLB entry is still established in 2132 * the huge TLB. Some CPU doesn't like that. 2133 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum 2134 * 383 on page 93. Intel should be safe but is also warns that it's 2135 * only safe if the permission and cache attributes of the two entries 2136 * loaded in the two TLB is identical (which should be the case here). 2137 * But it is generally safer to never allow small and huge TLB entries 2138 * for the same virtual address to be loaded simultaneously. So instead 2139 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2140 * current pmd notpresent (atomically because here the pmd_trans_huge 2141 * must remain set at all times on the pmd until the split is complete 2142 * for this pmd), then we flush the SMP TLB and finally we write the 2143 * non-huge version of the pmd entry with pmd_populate. 2144 */ 2145 old_pmd = pmdp_invalidate(vma, haddr, pmd); 2146 2147 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2148 pmd_migration = is_pmd_migration_entry(old_pmd); 2149 if (pmd_migration) { 2150 swp_entry_t entry; 2151 2152 entry = pmd_to_swp_entry(old_pmd); 2153 page = pfn_to_page(swp_offset(entry)); 2154 } else 2155 #endif 2156 page = pmd_page(old_pmd); 2157 VM_BUG_ON_PAGE(!page_count(page), page); 2158 page_ref_add(page, HPAGE_PMD_NR - 1); 2159 if (pmd_dirty(old_pmd)) 2160 SetPageDirty(page); 2161 write = pmd_write(old_pmd); 2162 young = pmd_young(old_pmd); 2163 soft_dirty = pmd_soft_dirty(old_pmd); 2164 2165 /* 2166 * Withdraw the table only after we mark the pmd entry invalid. 2167 * This's critical for some architectures (Power). 2168 */ 2169 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2170 pmd_populate(mm, &_pmd, pgtable); 2171 2172 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2173 pte_t entry, *pte; 2174 /* 2175 * Note that NUMA hinting access restrictions are not 2176 * transferred to avoid any possibility of altering 2177 * permissions across VMAs. 2178 */ 2179 if (freeze || pmd_migration) { 2180 swp_entry_t swp_entry; 2181 swp_entry = make_migration_entry(page + i, write); 2182 entry = swp_entry_to_pte(swp_entry); 2183 if (soft_dirty) 2184 entry = pte_swp_mksoft_dirty(entry); 2185 } else { 2186 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 2187 entry = maybe_mkwrite(entry, vma); 2188 if (!write) 2189 entry = pte_wrprotect(entry); 2190 if (!young) 2191 entry = pte_mkold(entry); 2192 if (soft_dirty) 2193 entry = pte_mksoft_dirty(entry); 2194 } 2195 pte = pte_offset_map(&_pmd, addr); 2196 BUG_ON(!pte_none(*pte)); 2197 set_pte_at(mm, addr, pte, entry); 2198 atomic_inc(&page[i]._mapcount); 2199 pte_unmap(pte); 2200 } 2201 2202 /* 2203 * Set PG_double_map before dropping compound_mapcount to avoid 2204 * false-negative page_mapped(). 2205 */ 2206 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { 2207 for (i = 0; i < HPAGE_PMD_NR; i++) 2208 atomic_inc(&page[i]._mapcount); 2209 } 2210 2211 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { 2212 /* Last compound_mapcount is gone. */ 2213 __dec_node_page_state(page, NR_ANON_THPS); 2214 if (TestClearPageDoubleMap(page)) { 2215 /* No need in mapcount reference anymore */ 2216 for (i = 0; i < HPAGE_PMD_NR; i++) 2217 atomic_dec(&page[i]._mapcount); 2218 } 2219 } 2220 2221 smp_wmb(); /* make pte visible before pmd */ 2222 pmd_populate(mm, pmd, pgtable); 2223 2224 if (freeze) { 2225 for (i = 0; i < HPAGE_PMD_NR; i++) { 2226 page_remove_rmap(page + i, false); 2227 put_page(page + i); 2228 } 2229 } 2230 } 2231 2232 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2233 unsigned long address, bool freeze, struct page *page) 2234 { 2235 spinlock_t *ptl; 2236 struct mm_struct *mm = vma->vm_mm; 2237 unsigned long haddr = address & HPAGE_PMD_MASK; 2238 2239 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); 2240 ptl = pmd_lock(mm, pmd); 2241 2242 /* 2243 * If caller asks to setup a migration entries, we need a page to check 2244 * pmd against. Otherwise we can end up replacing wrong page. 2245 */ 2246 VM_BUG_ON(freeze && !page); 2247 if (page && page != pmd_page(*pmd)) 2248 goto out; 2249 2250 if (pmd_trans_huge(*pmd)) { 2251 page = pmd_page(*pmd); 2252 if (PageMlocked(page)) 2253 clear_page_mlock(page); 2254 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) 2255 goto out; 2256 __split_huge_pmd_locked(vma, pmd, haddr, freeze); 2257 out: 2258 spin_unlock(ptl); 2259 /* 2260 * No need to double call mmu_notifier->invalidate_range() callback. 2261 * They are 3 cases to consider inside __split_huge_pmd_locked(): 2262 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious 2263 * 2) __split_huge_zero_page_pmd() read only zero page and any write 2264 * fault will trigger a flush_notify before pointing to a new page 2265 * (it is fine if the secondary mmu keeps pointing to the old zero 2266 * page in the meantime) 2267 * 3) Split a huge pmd into pte pointing to the same page. No need 2268 * to invalidate secondary tlb entry they are all still valid. 2269 * any further changes to individual pte will notify. So no need 2270 * to call mmu_notifier->invalidate_range() 2271 */ 2272 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + 2273 HPAGE_PMD_SIZE); 2274 } 2275 2276 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2277 bool freeze, struct page *page) 2278 { 2279 pgd_t *pgd; 2280 p4d_t *p4d; 2281 pud_t *pud; 2282 pmd_t *pmd; 2283 2284 pgd = pgd_offset(vma->vm_mm, address); 2285 if (!pgd_present(*pgd)) 2286 return; 2287 2288 p4d = p4d_offset(pgd, address); 2289 if (!p4d_present(*p4d)) 2290 return; 2291 2292 pud = pud_offset(p4d, address); 2293 if (!pud_present(*pud)) 2294 return; 2295 2296 pmd = pmd_offset(pud, address); 2297 2298 __split_huge_pmd(vma, pmd, address, freeze, page); 2299 } 2300 2301 void vma_adjust_trans_huge(struct vm_area_struct *vma, 2302 unsigned long start, 2303 unsigned long end, 2304 long adjust_next) 2305 { 2306 /* 2307 * If the new start address isn't hpage aligned and it could 2308 * previously contain an hugepage: check if we need to split 2309 * an huge pmd. 2310 */ 2311 if (start & ~HPAGE_PMD_MASK && 2312 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2313 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2314 split_huge_pmd_address(vma, start, false, NULL); 2315 2316 /* 2317 * If the new end address isn't hpage aligned and it could 2318 * previously contain an hugepage: check if we need to split 2319 * an huge pmd. 2320 */ 2321 if (end & ~HPAGE_PMD_MASK && 2322 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2323 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2324 split_huge_pmd_address(vma, end, false, NULL); 2325 2326 /* 2327 * If we're also updating the vma->vm_next->vm_start, if the new 2328 * vm_next->vm_start isn't page aligned and it could previously 2329 * contain an hugepage: check if we need to split an huge pmd. 2330 */ 2331 if (adjust_next > 0) { 2332 struct vm_area_struct *next = vma->vm_next; 2333 unsigned long nstart = next->vm_start; 2334 nstart += adjust_next << PAGE_SHIFT; 2335 if (nstart & ~HPAGE_PMD_MASK && 2336 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2337 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2338 split_huge_pmd_address(next, nstart, false, NULL); 2339 } 2340 } 2341 2342 static void unmap_page(struct page *page) 2343 { 2344 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | 2345 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; 2346 bool unmap_success; 2347 2348 VM_BUG_ON_PAGE(!PageHead(page), page); 2349 2350 if (PageAnon(page)) 2351 ttu_flags |= TTU_SPLIT_FREEZE; 2352 2353 unmap_success = try_to_unmap(page, ttu_flags); 2354 VM_BUG_ON_PAGE(!unmap_success, page); 2355 } 2356 2357 static void remap_page(struct page *page) 2358 { 2359 int i; 2360 if (PageTransHuge(page)) { 2361 remove_migration_ptes(page, page, true); 2362 } else { 2363 for (i = 0; i < HPAGE_PMD_NR; i++) 2364 remove_migration_ptes(page + i, page + i, true); 2365 } 2366 } 2367 2368 static void __split_huge_page_tail(struct page *head, int tail, 2369 struct lruvec *lruvec, struct list_head *list) 2370 { 2371 struct page *page_tail = head + tail; 2372 2373 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2374 2375 /* 2376 * Clone page flags before unfreezing refcount. 2377 * 2378 * After successful get_page_unless_zero() might follow flags change, 2379 * for exmaple lock_page() which set PG_waiters. 2380 */ 2381 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2382 page_tail->flags |= (head->flags & 2383 ((1L << PG_referenced) | 2384 (1L << PG_swapbacked) | 2385 (1L << PG_swapcache) | 2386 (1L << PG_mlocked) | 2387 (1L << PG_uptodate) | 2388 (1L << PG_active) | 2389 (1L << PG_workingset) | 2390 (1L << PG_locked) | 2391 (1L << PG_unevictable) | 2392 (1L << PG_dirty))); 2393 2394 /* ->mapping in first tail page is compound_mapcount */ 2395 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2396 page_tail); 2397 page_tail->mapping = head->mapping; 2398 page_tail->index = head->index + tail; 2399 2400 /* Page flags must be visible before we make the page non-compound. */ 2401 smp_wmb(); 2402 2403 /* 2404 * Clear PageTail before unfreezing page refcount. 2405 * 2406 * After successful get_page_unless_zero() might follow put_page() 2407 * which needs correct compound_head(). 2408 */ 2409 clear_compound_head(page_tail); 2410 2411 /* Finally unfreeze refcount. Additional reference from page cache. */ 2412 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || 2413 PageSwapCache(head))); 2414 2415 if (page_is_young(head)) 2416 set_page_young(page_tail); 2417 if (page_is_idle(head)) 2418 set_page_idle(page_tail); 2419 2420 page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); 2421 2422 /* 2423 * always add to the tail because some iterators expect new 2424 * pages to show after the currently processed elements - e.g. 2425 * migrate_pages 2426 */ 2427 lru_add_page_tail(head, page_tail, lruvec, list); 2428 } 2429 2430 static void __split_huge_page(struct page *page, struct list_head *list, 2431 pgoff_t end, unsigned long flags) 2432 { 2433 struct page *head = compound_head(page); 2434 struct zone *zone = page_zone(head); 2435 struct lruvec *lruvec; 2436 int i; 2437 2438 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); 2439 2440 /* complete memcg works before add pages to LRU */ 2441 mem_cgroup_split_huge_fixup(head); 2442 2443 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 2444 __split_huge_page_tail(head, i, lruvec, list); 2445 /* Some pages can be beyond i_size: drop them from page cache */ 2446 if (head[i].index >= end) { 2447 ClearPageDirty(head + i); 2448 __delete_from_page_cache(head + i, NULL); 2449 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) 2450 shmem_uncharge(head->mapping->host, 1); 2451 put_page(head + i); 2452 } 2453 } 2454 2455 ClearPageCompound(head); 2456 /* See comment in __split_huge_page_tail() */ 2457 if (PageAnon(head)) { 2458 /* Additional pin to swap cache */ 2459 if (PageSwapCache(head)) 2460 page_ref_add(head, 2); 2461 else 2462 page_ref_inc(head); 2463 } else { 2464 /* Additional pin to page cache */ 2465 page_ref_add(head, 2); 2466 xa_unlock(&head->mapping->i_pages); 2467 } 2468 2469 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2470 2471 remap_page(head); 2472 2473 for (i = 0; i < HPAGE_PMD_NR; i++) { 2474 struct page *subpage = head + i; 2475 if (subpage == page) 2476 continue; 2477 unlock_page(subpage); 2478 2479 /* 2480 * Subpages may be freed if there wasn't any mapping 2481 * like if add_to_swap() is running on a lru page that 2482 * had its mapping zapped. And freeing these pages 2483 * requires taking the lru_lock so we do the put_page 2484 * of the tail pages after the split is complete. 2485 */ 2486 put_page(subpage); 2487 } 2488 } 2489 2490 int total_mapcount(struct page *page) 2491 { 2492 int i, compound, ret; 2493 2494 VM_BUG_ON_PAGE(PageTail(page), page); 2495 2496 if (likely(!PageCompound(page))) 2497 return atomic_read(&page->_mapcount) + 1; 2498 2499 compound = compound_mapcount(page); 2500 if (PageHuge(page)) 2501 return compound; 2502 ret = compound; 2503 for (i = 0; i < HPAGE_PMD_NR; i++) 2504 ret += atomic_read(&page[i]._mapcount) + 1; 2505 /* File pages has compound_mapcount included in _mapcount */ 2506 if (!PageAnon(page)) 2507 return ret - compound * HPAGE_PMD_NR; 2508 if (PageDoubleMap(page)) 2509 ret -= HPAGE_PMD_NR; 2510 return ret; 2511 } 2512 2513 /* 2514 * This calculates accurately how many mappings a transparent hugepage 2515 * has (unlike page_mapcount() which isn't fully accurate). This full 2516 * accuracy is primarily needed to know if copy-on-write faults can 2517 * reuse the page and change the mapping to read-write instead of 2518 * copying them. At the same time this returns the total_mapcount too. 2519 * 2520 * The function returns the highest mapcount any one of the subpages 2521 * has. If the return value is one, even if different processes are 2522 * mapping different subpages of the transparent hugepage, they can 2523 * all reuse it, because each process is reusing a different subpage. 2524 * 2525 * The total_mapcount is instead counting all virtual mappings of the 2526 * subpages. If the total_mapcount is equal to "one", it tells the 2527 * caller all mappings belong to the same "mm" and in turn the 2528 * anon_vma of the transparent hugepage can become the vma->anon_vma 2529 * local one as no other process may be mapping any of the subpages. 2530 * 2531 * It would be more accurate to replace page_mapcount() with 2532 * page_trans_huge_mapcount(), however we only use 2533 * page_trans_huge_mapcount() in the copy-on-write faults where we 2534 * need full accuracy to avoid breaking page pinning, because 2535 * page_trans_huge_mapcount() is slower than page_mapcount(). 2536 */ 2537 int page_trans_huge_mapcount(struct page *page, int *total_mapcount) 2538 { 2539 int i, ret, _total_mapcount, mapcount; 2540 2541 /* hugetlbfs shouldn't call it */ 2542 VM_BUG_ON_PAGE(PageHuge(page), page); 2543 2544 if (likely(!PageTransCompound(page))) { 2545 mapcount = atomic_read(&page->_mapcount) + 1; 2546 if (total_mapcount) 2547 *total_mapcount = mapcount; 2548 return mapcount; 2549 } 2550 2551 page = compound_head(page); 2552 2553 _total_mapcount = ret = 0; 2554 for (i = 0; i < HPAGE_PMD_NR; i++) { 2555 mapcount = atomic_read(&page[i]._mapcount) + 1; 2556 ret = max(ret, mapcount); 2557 _total_mapcount += mapcount; 2558 } 2559 if (PageDoubleMap(page)) { 2560 ret -= 1; 2561 _total_mapcount -= HPAGE_PMD_NR; 2562 } 2563 mapcount = compound_mapcount(page); 2564 ret += mapcount; 2565 _total_mapcount += mapcount; 2566 if (total_mapcount) 2567 *total_mapcount = _total_mapcount; 2568 return ret; 2569 } 2570 2571 /* Racy check whether the huge page can be split */ 2572 bool can_split_huge_page(struct page *page, int *pextra_pins) 2573 { 2574 int extra_pins; 2575 2576 /* Additional pins from page cache */ 2577 if (PageAnon(page)) 2578 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; 2579 else 2580 extra_pins = HPAGE_PMD_NR; 2581 if (pextra_pins) 2582 *pextra_pins = extra_pins; 2583 return total_mapcount(page) == page_count(page) - extra_pins - 1; 2584 } 2585 2586 /* 2587 * This function splits huge page into normal pages. @page can point to any 2588 * subpage of huge page to split. Split doesn't change the position of @page. 2589 * 2590 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2591 * The huge page must be locked. 2592 * 2593 * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2594 * 2595 * Both head page and tail pages will inherit mapping, flags, and so on from 2596 * the hugepage. 2597 * 2598 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2599 * they are not mapped. 2600 * 2601 * Returns 0 if the hugepage is split successfully. 2602 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2603 * us. 2604 */ 2605 int split_huge_page_to_list(struct page *page, struct list_head *list) 2606 { 2607 struct page *head = compound_head(page); 2608 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); 2609 struct anon_vma *anon_vma = NULL; 2610 struct address_space *mapping = NULL; 2611 int count, mapcount, extra_pins, ret; 2612 bool mlocked; 2613 unsigned long flags; 2614 pgoff_t end; 2615 2616 VM_BUG_ON_PAGE(is_huge_zero_page(page), page); 2617 VM_BUG_ON_PAGE(!PageLocked(page), page); 2618 VM_BUG_ON_PAGE(!PageCompound(page), page); 2619 2620 if (PageWriteback(page)) 2621 return -EBUSY; 2622 2623 if (PageAnon(head)) { 2624 /* 2625 * The caller does not necessarily hold an mmap_sem that would 2626 * prevent the anon_vma disappearing so we first we take a 2627 * reference to it and then lock the anon_vma for write. This 2628 * is similar to page_lock_anon_vma_read except the write lock 2629 * is taken to serialise against parallel split or collapse 2630 * operations. 2631 */ 2632 anon_vma = page_get_anon_vma(head); 2633 if (!anon_vma) { 2634 ret = -EBUSY; 2635 goto out; 2636 } 2637 end = -1; 2638 mapping = NULL; 2639 anon_vma_lock_write(anon_vma); 2640 } else { 2641 mapping = head->mapping; 2642 2643 /* Truncated ? */ 2644 if (!mapping) { 2645 ret = -EBUSY; 2646 goto out; 2647 } 2648 2649 anon_vma = NULL; 2650 i_mmap_lock_read(mapping); 2651 2652 /* 2653 *__split_huge_page() may need to trim off pages beyond EOF: 2654 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 2655 * which cannot be nested inside the page tree lock. So note 2656 * end now: i_size itself may be changed at any moment, but 2657 * head page lock is good enough to serialize the trimming. 2658 */ 2659 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 2660 } 2661 2662 /* 2663 * Racy check if we can split the page, before unmap_page() will 2664 * split PMDs 2665 */ 2666 if (!can_split_huge_page(head, &extra_pins)) { 2667 ret = -EBUSY; 2668 goto out_unlock; 2669 } 2670 2671 mlocked = PageMlocked(page); 2672 unmap_page(head); 2673 VM_BUG_ON_PAGE(compound_mapcount(head), head); 2674 2675 /* Make sure the page is not on per-CPU pagevec as it takes pin */ 2676 if (mlocked) 2677 lru_add_drain(); 2678 2679 /* prevent PageLRU to go away from under us, and freeze lru stats */ 2680 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); 2681 2682 if (mapping) { 2683 XA_STATE(xas, &mapping->i_pages, page_index(head)); 2684 2685 /* 2686 * Check if the head page is present in page cache. 2687 * We assume all tail are present too, if head is there. 2688 */ 2689 xa_lock(&mapping->i_pages); 2690 if (xas_load(&xas) != head) 2691 goto fail; 2692 } 2693 2694 /* Prevent deferred_split_scan() touching ->_refcount */ 2695 spin_lock(&pgdata->split_queue_lock); 2696 count = page_count(head); 2697 mapcount = total_mapcount(head); 2698 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { 2699 if (!list_empty(page_deferred_list(head))) { 2700 pgdata->split_queue_len--; 2701 list_del(page_deferred_list(head)); 2702 } 2703 if (mapping) 2704 __dec_node_page_state(page, NR_SHMEM_THPS); 2705 spin_unlock(&pgdata->split_queue_lock); 2706 __split_huge_page(page, list, end, flags); 2707 if (PageSwapCache(head)) { 2708 swp_entry_t entry = { .val = page_private(head) }; 2709 2710 ret = split_swap_cluster(entry); 2711 } else 2712 ret = 0; 2713 } else { 2714 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { 2715 pr_alert("total_mapcount: %u, page_count(): %u\n", 2716 mapcount, count); 2717 if (PageTail(page)) 2718 dump_page(head, NULL); 2719 dump_page(page, "total_mapcount(head) > 0"); 2720 BUG(); 2721 } 2722 spin_unlock(&pgdata->split_queue_lock); 2723 fail: if (mapping) 2724 xa_unlock(&mapping->i_pages); 2725 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); 2726 remap_page(head); 2727 ret = -EBUSY; 2728 } 2729 2730 out_unlock: 2731 if (anon_vma) { 2732 anon_vma_unlock_write(anon_vma); 2733 put_anon_vma(anon_vma); 2734 } 2735 if (mapping) 2736 i_mmap_unlock_read(mapping); 2737 out: 2738 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 2739 return ret; 2740 } 2741 2742 void free_transhuge_page(struct page *page) 2743 { 2744 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2745 unsigned long flags; 2746 2747 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2748 if (!list_empty(page_deferred_list(page))) { 2749 pgdata->split_queue_len--; 2750 list_del(page_deferred_list(page)); 2751 } 2752 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2753 free_compound_page(page); 2754 } 2755 2756 void deferred_split_huge_page(struct page *page) 2757 { 2758 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); 2759 unsigned long flags; 2760 2761 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 2762 2763 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2764 if (list_empty(page_deferred_list(page))) { 2765 count_vm_event(THP_DEFERRED_SPLIT_PAGE); 2766 list_add_tail(page_deferred_list(page), &pgdata->split_queue); 2767 pgdata->split_queue_len++; 2768 } 2769 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2770 } 2771 2772 static unsigned long deferred_split_count(struct shrinker *shrink, 2773 struct shrink_control *sc) 2774 { 2775 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2776 return READ_ONCE(pgdata->split_queue_len); 2777 } 2778 2779 static unsigned long deferred_split_scan(struct shrinker *shrink, 2780 struct shrink_control *sc) 2781 { 2782 struct pglist_data *pgdata = NODE_DATA(sc->nid); 2783 unsigned long flags; 2784 LIST_HEAD(list), *pos, *next; 2785 struct page *page; 2786 int split = 0; 2787 2788 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2789 /* Take pin on all head pages to avoid freeing them under us */ 2790 list_for_each_safe(pos, next, &pgdata->split_queue) { 2791 page = list_entry((void *)pos, struct page, mapping); 2792 page = compound_head(page); 2793 if (get_page_unless_zero(page)) { 2794 list_move(page_deferred_list(page), &list); 2795 } else { 2796 /* We lost race with put_compound_page() */ 2797 list_del_init(page_deferred_list(page)); 2798 pgdata->split_queue_len--; 2799 } 2800 if (!--sc->nr_to_scan) 2801 break; 2802 } 2803 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2804 2805 list_for_each_safe(pos, next, &list) { 2806 page = list_entry((void *)pos, struct page, mapping); 2807 if (!trylock_page(page)) 2808 goto next; 2809 /* split_huge_page() removes page from list on success */ 2810 if (!split_huge_page(page)) 2811 split++; 2812 unlock_page(page); 2813 next: 2814 put_page(page); 2815 } 2816 2817 spin_lock_irqsave(&pgdata->split_queue_lock, flags); 2818 list_splice_tail(&list, &pgdata->split_queue); 2819 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); 2820 2821 /* 2822 * Stop shrinker if we didn't split any page, but the queue is empty. 2823 * This can happen if pages were freed under us. 2824 */ 2825 if (!split && list_empty(&pgdata->split_queue)) 2826 return SHRINK_STOP; 2827 return split; 2828 } 2829 2830 static struct shrinker deferred_split_shrinker = { 2831 .count_objects = deferred_split_count, 2832 .scan_objects = deferred_split_scan, 2833 .seeks = DEFAULT_SEEKS, 2834 .flags = SHRINKER_NUMA_AWARE, 2835 }; 2836 2837 #ifdef CONFIG_DEBUG_FS 2838 static int split_huge_pages_set(void *data, u64 val) 2839 { 2840 struct zone *zone; 2841 struct page *page; 2842 unsigned long pfn, max_zone_pfn; 2843 unsigned long total = 0, split = 0; 2844 2845 if (val != 1) 2846 return -EINVAL; 2847 2848 for_each_populated_zone(zone) { 2849 max_zone_pfn = zone_end_pfn(zone); 2850 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 2851 if (!pfn_valid(pfn)) 2852 continue; 2853 2854 page = pfn_to_page(pfn); 2855 if (!get_page_unless_zero(page)) 2856 continue; 2857 2858 if (zone != page_zone(page)) 2859 goto next; 2860 2861 if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) 2862 goto next; 2863 2864 total++; 2865 lock_page(page); 2866 if (!split_huge_page(page)) 2867 split++; 2868 unlock_page(page); 2869 next: 2870 put_page(page); 2871 } 2872 } 2873 2874 pr_info("%lu of %lu THP split\n", split, total); 2875 2876 return 0; 2877 } 2878 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, 2879 "%llu\n"); 2880 2881 static int __init split_huge_pages_debugfs(void) 2882 { 2883 void *ret; 2884 2885 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 2886 &split_huge_pages_fops); 2887 if (!ret) 2888 pr_warn("Failed to create split_huge_pages in debugfs"); 2889 return 0; 2890 } 2891 late_initcall(split_huge_pages_debugfs); 2892 #endif 2893 2894 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 2895 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 2896 struct page *page) 2897 { 2898 struct vm_area_struct *vma = pvmw->vma; 2899 struct mm_struct *mm = vma->vm_mm; 2900 unsigned long address = pvmw->address; 2901 pmd_t pmdval; 2902 swp_entry_t entry; 2903 pmd_t pmdswp; 2904 2905 if (!(pvmw->pmd && !pvmw->pte)) 2906 return; 2907 2908 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 2909 pmdval = *pvmw->pmd; 2910 pmdp_invalidate(vma, address, pvmw->pmd); 2911 if (pmd_dirty(pmdval)) 2912 set_page_dirty(page); 2913 entry = make_migration_entry(page, pmd_write(pmdval)); 2914 pmdswp = swp_entry_to_pmd(entry); 2915 if (pmd_soft_dirty(pmdval)) 2916 pmdswp = pmd_swp_mksoft_dirty(pmdswp); 2917 set_pmd_at(mm, address, pvmw->pmd, pmdswp); 2918 page_remove_rmap(page, true); 2919 put_page(page); 2920 } 2921 2922 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 2923 { 2924 struct vm_area_struct *vma = pvmw->vma; 2925 struct mm_struct *mm = vma->vm_mm; 2926 unsigned long address = pvmw->address; 2927 unsigned long mmun_start = address & HPAGE_PMD_MASK; 2928 pmd_t pmde; 2929 swp_entry_t entry; 2930 2931 if (!(pvmw->pmd && !pvmw->pte)) 2932 return; 2933 2934 entry = pmd_to_swp_entry(*pvmw->pmd); 2935 get_page(new); 2936 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); 2937 if (pmd_swp_soft_dirty(*pvmw->pmd)) 2938 pmde = pmd_mksoft_dirty(pmde); 2939 if (is_write_migration_entry(entry)) 2940 pmde = maybe_pmd_mkwrite(pmde, vma); 2941 2942 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); 2943 if (PageAnon(new)) 2944 page_add_anon_rmap(new, vma, mmun_start, true); 2945 else 2946 page_add_file_rmap(new, true); 2947 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); 2948 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) 2949 mlock_vma_page(new); 2950 update_mmu_cache_pmd(vma, address, pvmw->pmd); 2951 } 2952 #endif 2953