1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2008, 2009 Intel Corporation 4 * Authors: Andi Kleen, Fengguang Wu 5 * 6 * High level machine check handler. Handles pages reported by the 7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache 8 * failure. 9 * 10 * In addition there is a "soft offline" entry point that allows stop using 11 * not-yet-corrupted-by-suspicious pages without killing anything. 12 * 13 * Handles page cache pages in various states. The tricky part 14 * here is that we can access any page asynchronously in respect to 15 * other VM users, because memory failures could happen anytime and 16 * anywhere. This could violate some of their assumptions. This is why 17 * this code has to be extremely careful. Generally it tries to use 18 * normal locking rules, as in get the standard locks, even if that means 19 * the error handling takes potentially a long time. 20 * 21 * It can be very tempting to add handling for obscure cases here. 22 * In general any code for handling new cases should only be added iff: 23 * - You know how to test it. 24 * - You have a test that can be added to mce-test 25 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ 26 * - The case actually shows up as a frequent (top 10) page state in 27 * tools/mm/page-types when running a real workload. 28 * 29 * There are several operations here with exponential complexity because 30 * of unsuitable VM data structures. For example the operation to map back 31 * from RMAP chains to processes has to walk the complete process list and 32 * has non linear complexity with the number. But since memory corruptions 33 * are rare we hope to get away with this. This avoids impacting the core 34 * VM. 35 */ 36 37 #define pr_fmt(fmt) "Memory failure: " fmt 38 39 #include <linux/kernel.h> 40 #include <linux/mm.h> 41 #include <linux/page-flags.h> 42 #include <linux/sched/signal.h> 43 #include <linux/sched/task.h> 44 #include <linux/dax.h> 45 #include <linux/ksm.h> 46 #include <linux/rmap.h> 47 #include <linux/export.h> 48 #include <linux/pagemap.h> 49 #include <linux/swap.h> 50 #include <linux/backing-dev.h> 51 #include <linux/migrate.h> 52 #include <linux/slab.h> 53 #include <linux/swapops.h> 54 #include <linux/hugetlb.h> 55 #include <linux/memory_hotplug.h> 56 #include <linux/mm_inline.h> 57 #include <linux/memremap.h> 58 #include <linux/kfifo.h> 59 #include <linux/ratelimit.h> 60 #include <linux/pagewalk.h> 61 #include <linux/shmem_fs.h> 62 #include <linux/sysctl.h> 63 #include "swap.h" 64 #include "internal.h" 65 #include "ras/ras_event.h" 66 67 static int sysctl_memory_failure_early_kill __read_mostly; 68 69 static int sysctl_memory_failure_recovery __read_mostly = 1; 70 71 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); 72 73 static bool hw_memory_failure __read_mostly = false; 74 75 static DEFINE_MUTEX(mf_mutex); 76 77 void num_poisoned_pages_inc(unsigned long pfn) 78 { 79 atomic_long_inc(&num_poisoned_pages); 80 memblk_nr_poison_inc(pfn); 81 } 82 83 void num_poisoned_pages_sub(unsigned long pfn, long i) 84 { 85 atomic_long_sub(i, &num_poisoned_pages); 86 if (pfn != -1UL) 87 memblk_nr_poison_sub(pfn, i); 88 } 89 90 /** 91 * MF_ATTR_RO - Create sysfs entry for each memory failure statistics. 92 * @_name: name of the file in the per NUMA sysfs directory. 93 */ 94 #define MF_ATTR_RO(_name) \ 95 static ssize_t _name##_show(struct device *dev, \ 96 struct device_attribute *attr, \ 97 char *buf) \ 98 { \ 99 struct memory_failure_stats *mf_stats = \ 100 &NODE_DATA(dev->id)->mf_stats; \ 101 return sprintf(buf, "%lu\n", mf_stats->_name); \ 102 } \ 103 static DEVICE_ATTR_RO(_name) 104 105 MF_ATTR_RO(total); 106 MF_ATTR_RO(ignored); 107 MF_ATTR_RO(failed); 108 MF_ATTR_RO(delayed); 109 MF_ATTR_RO(recovered); 110 111 static struct attribute *memory_failure_attr[] = { 112 &dev_attr_total.attr, 113 &dev_attr_ignored.attr, 114 &dev_attr_failed.attr, 115 &dev_attr_delayed.attr, 116 &dev_attr_recovered.attr, 117 NULL, 118 }; 119 120 const struct attribute_group memory_failure_attr_group = { 121 .name = "memory_failure", 122 .attrs = memory_failure_attr, 123 }; 124 125 static struct ctl_table memory_failure_table[] = { 126 { 127 .procname = "memory_failure_early_kill", 128 .data = &sysctl_memory_failure_early_kill, 129 .maxlen = sizeof(sysctl_memory_failure_early_kill), 130 .mode = 0644, 131 .proc_handler = proc_dointvec_minmax, 132 .extra1 = SYSCTL_ZERO, 133 .extra2 = SYSCTL_ONE, 134 }, 135 { 136 .procname = "memory_failure_recovery", 137 .data = &sysctl_memory_failure_recovery, 138 .maxlen = sizeof(sysctl_memory_failure_recovery), 139 .mode = 0644, 140 .proc_handler = proc_dointvec_minmax, 141 .extra1 = SYSCTL_ZERO, 142 .extra2 = SYSCTL_ONE, 143 }, 144 { } 145 }; 146 147 /* 148 * Return values: 149 * 1: the page is dissolved (if needed) and taken off from buddy, 150 * 0: the page is dissolved (if needed) and not taken off from buddy, 151 * < 0: failed to dissolve. 152 */ 153 static int __page_handle_poison(struct page *page) 154 { 155 int ret; 156 157 zone_pcp_disable(page_zone(page)); 158 ret = dissolve_free_huge_page(page); 159 if (!ret) 160 ret = take_page_off_buddy(page); 161 zone_pcp_enable(page_zone(page)); 162 163 return ret; 164 } 165 166 static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) 167 { 168 if (hugepage_or_freepage) { 169 /* 170 * Doing this check for free pages is also fine since dissolve_free_huge_page 171 * returns 0 for non-hugetlb pages as well. 172 */ 173 if (__page_handle_poison(page) <= 0) 174 /* 175 * We could fail to take off the target page from buddy 176 * for example due to racy page allocation, but that's 177 * acceptable because soft-offlined page is not broken 178 * and if someone really want to use it, they should 179 * take it. 180 */ 181 return false; 182 } 183 184 SetPageHWPoison(page); 185 if (release) 186 put_page(page); 187 page_ref_inc(page); 188 num_poisoned_pages_inc(page_to_pfn(page)); 189 190 return true; 191 } 192 193 #if IS_ENABLED(CONFIG_HWPOISON_INJECT) 194 195 u32 hwpoison_filter_enable = 0; 196 u32 hwpoison_filter_dev_major = ~0U; 197 u32 hwpoison_filter_dev_minor = ~0U; 198 u64 hwpoison_filter_flags_mask; 199 u64 hwpoison_filter_flags_value; 200 EXPORT_SYMBOL_GPL(hwpoison_filter_enable); 201 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); 202 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); 203 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); 204 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); 205 206 static int hwpoison_filter_dev(struct page *p) 207 { 208 struct address_space *mapping; 209 dev_t dev; 210 211 if (hwpoison_filter_dev_major == ~0U && 212 hwpoison_filter_dev_minor == ~0U) 213 return 0; 214 215 mapping = page_mapping(p); 216 if (mapping == NULL || mapping->host == NULL) 217 return -EINVAL; 218 219 dev = mapping->host->i_sb->s_dev; 220 if (hwpoison_filter_dev_major != ~0U && 221 hwpoison_filter_dev_major != MAJOR(dev)) 222 return -EINVAL; 223 if (hwpoison_filter_dev_minor != ~0U && 224 hwpoison_filter_dev_minor != MINOR(dev)) 225 return -EINVAL; 226 227 return 0; 228 } 229 230 static int hwpoison_filter_flags(struct page *p) 231 { 232 if (!hwpoison_filter_flags_mask) 233 return 0; 234 235 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == 236 hwpoison_filter_flags_value) 237 return 0; 238 else 239 return -EINVAL; 240 } 241 242 /* 243 * This allows stress tests to limit test scope to a collection of tasks 244 * by putting them under some memcg. This prevents killing unrelated/important 245 * processes such as /sbin/init. Note that the target task may share clean 246 * pages with init (eg. libc text), which is harmless. If the target task 247 * share _dirty_ pages with another task B, the test scheme must make sure B 248 * is also included in the memcg. At last, due to race conditions this filter 249 * can only guarantee that the page either belongs to the memcg tasks, or is 250 * a freed page. 251 */ 252 #ifdef CONFIG_MEMCG 253 u64 hwpoison_filter_memcg; 254 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); 255 static int hwpoison_filter_task(struct page *p) 256 { 257 if (!hwpoison_filter_memcg) 258 return 0; 259 260 if (page_cgroup_ino(p) != hwpoison_filter_memcg) 261 return -EINVAL; 262 263 return 0; 264 } 265 #else 266 static int hwpoison_filter_task(struct page *p) { return 0; } 267 #endif 268 269 int hwpoison_filter(struct page *p) 270 { 271 if (!hwpoison_filter_enable) 272 return 0; 273 274 if (hwpoison_filter_dev(p)) 275 return -EINVAL; 276 277 if (hwpoison_filter_flags(p)) 278 return -EINVAL; 279 280 if (hwpoison_filter_task(p)) 281 return -EINVAL; 282 283 return 0; 284 } 285 #else 286 int hwpoison_filter(struct page *p) 287 { 288 return 0; 289 } 290 #endif 291 292 EXPORT_SYMBOL_GPL(hwpoison_filter); 293 294 /* 295 * Kill all processes that have a poisoned page mapped and then isolate 296 * the page. 297 * 298 * General strategy: 299 * Find all processes having the page mapped and kill them. 300 * But we keep a page reference around so that the page is not 301 * actually freed yet. 302 * Then stash the page away 303 * 304 * There's no convenient way to get back to mapped processes 305 * from the VMAs. So do a brute-force search over all 306 * running processes. 307 * 308 * Remember that machine checks are not common (or rather 309 * if they are common you have other problems), so this shouldn't 310 * be a performance issue. 311 * 312 * Also there are some races possible while we get from the 313 * error detection to actually handle it. 314 */ 315 316 struct to_kill { 317 struct list_head nd; 318 struct task_struct *tsk; 319 unsigned long addr; 320 short size_shift; 321 }; 322 323 /* 324 * Send all the processes who have the page mapped a signal. 325 * ``action optional'' if they are not immediately affected by the error 326 * ``action required'' if error happened in current execution context 327 */ 328 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) 329 { 330 struct task_struct *t = tk->tsk; 331 short addr_lsb = tk->size_shift; 332 int ret = 0; 333 334 pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", 335 pfn, t->comm, t->pid); 336 337 if ((flags & MF_ACTION_REQUIRED) && (t == current)) 338 ret = force_sig_mceerr(BUS_MCEERR_AR, 339 (void __user *)tk->addr, addr_lsb); 340 else 341 /* 342 * Signal other processes sharing the page if they have 343 * PF_MCE_EARLY set. 344 * Don't use force here, it's convenient if the signal 345 * can be temporarily blocked. 346 * This could cause a loop when the user sets SIGBUS 347 * to SIG_IGN, but hopefully no one will do that? 348 */ 349 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, 350 addr_lsb, t); 351 if (ret < 0) 352 pr_info("Error sending signal to %s:%d: %d\n", 353 t->comm, t->pid, ret); 354 return ret; 355 } 356 357 /* 358 * Unknown page type encountered. Try to check whether it can turn PageLRU by 359 * lru_add_drain_all. 360 */ 361 void shake_page(struct page *p) 362 { 363 if (PageHuge(p)) 364 return; 365 /* 366 * TODO: Could shrink slab caches here if a lightweight range-based 367 * shrinker will be available. 368 */ 369 if (PageSlab(p)) 370 return; 371 372 lru_add_drain_all(); 373 } 374 EXPORT_SYMBOL_GPL(shake_page); 375 376 static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma, 377 unsigned long address) 378 { 379 unsigned long ret = 0; 380 pgd_t *pgd; 381 p4d_t *p4d; 382 pud_t *pud; 383 pmd_t *pmd; 384 pte_t *pte; 385 pte_t ptent; 386 387 VM_BUG_ON_VMA(address == -EFAULT, vma); 388 pgd = pgd_offset(vma->vm_mm, address); 389 if (!pgd_present(*pgd)) 390 return 0; 391 p4d = p4d_offset(pgd, address); 392 if (!p4d_present(*p4d)) 393 return 0; 394 pud = pud_offset(p4d, address); 395 if (!pud_present(*pud)) 396 return 0; 397 if (pud_devmap(*pud)) 398 return PUD_SHIFT; 399 pmd = pmd_offset(pud, address); 400 if (!pmd_present(*pmd)) 401 return 0; 402 if (pmd_devmap(*pmd)) 403 return PMD_SHIFT; 404 pte = pte_offset_map(pmd, address); 405 if (!pte) 406 return 0; 407 ptent = ptep_get(pte); 408 if (pte_present(ptent) && pte_devmap(ptent)) 409 ret = PAGE_SHIFT; 410 pte_unmap(pte); 411 return ret; 412 } 413 414 /* 415 * Failure handling: if we can't find or can't kill a process there's 416 * not much we can do. We just print a message and ignore otherwise. 417 */ 418 419 #define FSDAX_INVALID_PGOFF ULONG_MAX 420 421 /* 422 * Schedule a process for later kill. 423 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. 424 * 425 * Note: @fsdax_pgoff is used only when @p is a fsdax page and a 426 * filesystem with a memory failure handler has claimed the 427 * memory_failure event. In all other cases, page->index and 428 * page->mapping are sufficient for mapping the page back to its 429 * corresponding user virtual address. 430 */ 431 static void __add_to_kill(struct task_struct *tsk, struct page *p, 432 struct vm_area_struct *vma, struct list_head *to_kill, 433 unsigned long ksm_addr, pgoff_t fsdax_pgoff) 434 { 435 struct to_kill *tk; 436 437 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); 438 if (!tk) { 439 pr_err("Out of memory while machine check handling\n"); 440 return; 441 } 442 443 tk->addr = ksm_addr ? ksm_addr : page_address_in_vma(p, vma); 444 if (is_zone_device_page(p)) { 445 if (fsdax_pgoff != FSDAX_INVALID_PGOFF) 446 tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma); 447 tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr); 448 } else 449 tk->size_shift = page_shift(compound_head(p)); 450 451 /* 452 * Send SIGKILL if "tk->addr == -EFAULT". Also, as 453 * "tk->size_shift" is always non-zero for !is_zone_device_page(), 454 * so "tk->size_shift == 0" effectively checks no mapping on 455 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times 456 * to a process' address space, it's possible not all N VMAs 457 * contain mappings for the page, but at least one VMA does. 458 * Only deliver SIGBUS with payload derived from the VMA that 459 * has a mapping for the page. 460 */ 461 if (tk->addr == -EFAULT) { 462 pr_info("Unable to find user space address %lx in %s\n", 463 page_to_pfn(p), tsk->comm); 464 } else if (tk->size_shift == 0) { 465 kfree(tk); 466 return; 467 } 468 469 get_task_struct(tsk); 470 tk->tsk = tsk; 471 list_add_tail(&tk->nd, to_kill); 472 } 473 474 static void add_to_kill_anon_file(struct task_struct *tsk, struct page *p, 475 struct vm_area_struct *vma, 476 struct list_head *to_kill) 477 { 478 __add_to_kill(tsk, p, vma, to_kill, 0, FSDAX_INVALID_PGOFF); 479 } 480 481 #ifdef CONFIG_KSM 482 static bool task_in_to_kill_list(struct list_head *to_kill, 483 struct task_struct *tsk) 484 { 485 struct to_kill *tk, *next; 486 487 list_for_each_entry_safe(tk, next, to_kill, nd) { 488 if (tk->tsk == tsk) 489 return true; 490 } 491 492 return false; 493 } 494 void add_to_kill_ksm(struct task_struct *tsk, struct page *p, 495 struct vm_area_struct *vma, struct list_head *to_kill, 496 unsigned long ksm_addr) 497 { 498 if (!task_in_to_kill_list(to_kill, tsk)) 499 __add_to_kill(tsk, p, vma, to_kill, ksm_addr, FSDAX_INVALID_PGOFF); 500 } 501 #endif 502 /* 503 * Kill the processes that have been collected earlier. 504 * 505 * Only do anything when FORCEKILL is set, otherwise just free the 506 * list (this is used for clean pages which do not need killing) 507 * Also when FAIL is set do a force kill because something went 508 * wrong earlier. 509 */ 510 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, 511 unsigned long pfn, int flags) 512 { 513 struct to_kill *tk, *next; 514 515 list_for_each_entry_safe(tk, next, to_kill, nd) { 516 if (forcekill) { 517 /* 518 * In case something went wrong with munmapping 519 * make sure the process doesn't catch the 520 * signal and then access the memory. Just kill it. 521 */ 522 if (fail || tk->addr == -EFAULT) { 523 pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", 524 pfn, tk->tsk->comm, tk->tsk->pid); 525 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, 526 tk->tsk, PIDTYPE_PID); 527 } 528 529 /* 530 * In theory the process could have mapped 531 * something else on the address in-between. We could 532 * check for that, but we need to tell the 533 * process anyways. 534 */ 535 else if (kill_proc(tk, pfn, flags) < 0) 536 pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n", 537 pfn, tk->tsk->comm, tk->tsk->pid); 538 } 539 list_del(&tk->nd); 540 put_task_struct(tk->tsk); 541 kfree(tk); 542 } 543 } 544 545 /* 546 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) 547 * on behalf of the thread group. Return task_struct of the (first found) 548 * dedicated thread if found, and return NULL otherwise. 549 * 550 * We already hold rcu lock in the caller, so we don't have to call 551 * rcu_read_lock/unlock() in this function. 552 */ 553 static struct task_struct *find_early_kill_thread(struct task_struct *tsk) 554 { 555 struct task_struct *t; 556 557 for_each_thread(tsk, t) { 558 if (t->flags & PF_MCE_PROCESS) { 559 if (t->flags & PF_MCE_EARLY) 560 return t; 561 } else { 562 if (sysctl_memory_failure_early_kill) 563 return t; 564 } 565 } 566 return NULL; 567 } 568 569 /* 570 * Determine whether a given process is "early kill" process which expects 571 * to be signaled when some page under the process is hwpoisoned. 572 * Return task_struct of the dedicated thread (main thread unless explicitly 573 * specified) if the process is "early kill" and otherwise returns NULL. 574 * 575 * Note that the above is true for Action Optional case. For Action Required 576 * case, it's only meaningful to the current thread which need to be signaled 577 * with SIGBUS, this error is Action Optional for other non current 578 * processes sharing the same error page,if the process is "early kill", the 579 * task_struct of the dedicated thread will also be returned. 580 */ 581 struct task_struct *task_early_kill(struct task_struct *tsk, int force_early) 582 { 583 if (!tsk->mm) 584 return NULL; 585 /* 586 * Comparing ->mm here because current task might represent 587 * a subthread, while tsk always points to the main thread. 588 */ 589 if (force_early && tsk->mm == current->mm) 590 return current; 591 592 return find_early_kill_thread(tsk); 593 } 594 595 /* 596 * Collect processes when the error hit an anonymous page. 597 */ 598 static void collect_procs_anon(struct folio *folio, struct page *page, 599 struct list_head *to_kill, int force_early) 600 { 601 struct vm_area_struct *vma; 602 struct task_struct *tsk; 603 struct anon_vma *av; 604 pgoff_t pgoff; 605 606 av = folio_lock_anon_vma_read(folio, NULL); 607 if (av == NULL) /* Not actually mapped anymore */ 608 return; 609 610 pgoff = page_to_pgoff(page); 611 rcu_read_lock(); 612 for_each_process(tsk) { 613 struct anon_vma_chain *vmac; 614 struct task_struct *t = task_early_kill(tsk, force_early); 615 616 if (!t) 617 continue; 618 anon_vma_interval_tree_foreach(vmac, &av->rb_root, 619 pgoff, pgoff) { 620 vma = vmac->vma; 621 if (vma->vm_mm != t->mm) 622 continue; 623 if (!page_mapped_in_vma(page, vma)) 624 continue; 625 add_to_kill_anon_file(t, page, vma, to_kill); 626 } 627 } 628 rcu_read_unlock(); 629 anon_vma_unlock_read(av); 630 } 631 632 /* 633 * Collect processes when the error hit a file mapped page. 634 */ 635 static void collect_procs_file(struct folio *folio, struct page *page, 636 struct list_head *to_kill, int force_early) 637 { 638 struct vm_area_struct *vma; 639 struct task_struct *tsk; 640 struct address_space *mapping = folio->mapping; 641 pgoff_t pgoff; 642 643 i_mmap_lock_read(mapping); 644 rcu_read_lock(); 645 pgoff = page_to_pgoff(page); 646 for_each_process(tsk) { 647 struct task_struct *t = task_early_kill(tsk, force_early); 648 649 if (!t) 650 continue; 651 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, 652 pgoff) { 653 /* 654 * Send early kill signal to tasks where a vma covers 655 * the page but the corrupted page is not necessarily 656 * mapped in its pte. 657 * Assume applications who requested early kill want 658 * to be informed of all such data corruptions. 659 */ 660 if (vma->vm_mm == t->mm) 661 add_to_kill_anon_file(t, page, vma, to_kill); 662 } 663 } 664 rcu_read_unlock(); 665 i_mmap_unlock_read(mapping); 666 } 667 668 #ifdef CONFIG_FS_DAX 669 static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p, 670 struct vm_area_struct *vma, 671 struct list_head *to_kill, pgoff_t pgoff) 672 { 673 __add_to_kill(tsk, p, vma, to_kill, 0, pgoff); 674 } 675 676 /* 677 * Collect processes when the error hit a fsdax page. 678 */ 679 static void collect_procs_fsdax(struct page *page, 680 struct address_space *mapping, pgoff_t pgoff, 681 struct list_head *to_kill, bool pre_remove) 682 { 683 struct vm_area_struct *vma; 684 struct task_struct *tsk; 685 686 i_mmap_lock_read(mapping); 687 rcu_read_lock(); 688 for_each_process(tsk) { 689 struct task_struct *t = tsk; 690 691 /* 692 * Search for all tasks while MF_MEM_PRE_REMOVE is set, because 693 * the current may not be the one accessing the fsdax page. 694 * Otherwise, search for the current task. 695 */ 696 if (!pre_remove) 697 t = task_early_kill(tsk, true); 698 if (!t) 699 continue; 700 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 701 if (vma->vm_mm == t->mm) 702 add_to_kill_fsdax(t, page, vma, to_kill, pgoff); 703 } 704 } 705 rcu_read_unlock(); 706 i_mmap_unlock_read(mapping); 707 } 708 #endif /* CONFIG_FS_DAX */ 709 710 /* 711 * Collect the processes who have the corrupted page mapped to kill. 712 */ 713 static void collect_procs(struct folio *folio, struct page *page, 714 struct list_head *tokill, int force_early) 715 { 716 if (!folio->mapping) 717 return; 718 if (unlikely(PageKsm(page))) 719 collect_procs_ksm(page, tokill, force_early); 720 else if (PageAnon(page)) 721 collect_procs_anon(folio, page, tokill, force_early); 722 else 723 collect_procs_file(folio, page, tokill, force_early); 724 } 725 726 struct hwpoison_walk { 727 struct to_kill tk; 728 unsigned long pfn; 729 int flags; 730 }; 731 732 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) 733 { 734 tk->addr = addr; 735 tk->size_shift = shift; 736 } 737 738 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, 739 unsigned long poisoned_pfn, struct to_kill *tk) 740 { 741 unsigned long pfn = 0; 742 743 if (pte_present(pte)) { 744 pfn = pte_pfn(pte); 745 } else { 746 swp_entry_t swp = pte_to_swp_entry(pte); 747 748 if (is_hwpoison_entry(swp)) 749 pfn = swp_offset_pfn(swp); 750 } 751 752 if (!pfn || pfn != poisoned_pfn) 753 return 0; 754 755 set_to_kill(tk, addr, shift); 756 return 1; 757 } 758 759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 760 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, 761 struct hwpoison_walk *hwp) 762 { 763 pmd_t pmd = *pmdp; 764 unsigned long pfn; 765 unsigned long hwpoison_vaddr; 766 767 if (!pmd_present(pmd)) 768 return 0; 769 pfn = pmd_pfn(pmd); 770 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { 771 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); 772 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); 773 return 1; 774 } 775 return 0; 776 } 777 #else 778 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, 779 struct hwpoison_walk *hwp) 780 { 781 return 0; 782 } 783 #endif 784 785 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, 786 unsigned long end, struct mm_walk *walk) 787 { 788 struct hwpoison_walk *hwp = walk->private; 789 int ret = 0; 790 pte_t *ptep, *mapped_pte; 791 spinlock_t *ptl; 792 793 ptl = pmd_trans_huge_lock(pmdp, walk->vma); 794 if (ptl) { 795 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); 796 spin_unlock(ptl); 797 goto out; 798 } 799 800 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, 801 addr, &ptl); 802 if (!ptep) 803 goto out; 804 805 for (; addr != end; ptep++, addr += PAGE_SIZE) { 806 ret = check_hwpoisoned_entry(ptep_get(ptep), addr, PAGE_SHIFT, 807 hwp->pfn, &hwp->tk); 808 if (ret == 1) 809 break; 810 } 811 pte_unmap_unlock(mapped_pte, ptl); 812 out: 813 cond_resched(); 814 return ret; 815 } 816 817 #ifdef CONFIG_HUGETLB_PAGE 818 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, 819 unsigned long addr, unsigned long end, 820 struct mm_walk *walk) 821 { 822 struct hwpoison_walk *hwp = walk->private; 823 pte_t pte = huge_ptep_get(ptep); 824 struct hstate *h = hstate_vma(walk->vma); 825 826 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), 827 hwp->pfn, &hwp->tk); 828 } 829 #else 830 #define hwpoison_hugetlb_range NULL 831 #endif 832 833 static const struct mm_walk_ops hwpoison_walk_ops = { 834 .pmd_entry = hwpoison_pte_range, 835 .hugetlb_entry = hwpoison_hugetlb_range, 836 .walk_lock = PGWALK_RDLOCK, 837 }; 838 839 /* 840 * Sends SIGBUS to the current process with error info. 841 * 842 * This function is intended to handle "Action Required" MCEs on already 843 * hardware poisoned pages. They could happen, for example, when 844 * memory_failure() failed to unmap the error page at the first call, or 845 * when multiple local machine checks happened on different CPUs. 846 * 847 * MCE handler currently has no easy access to the error virtual address, 848 * so this function walks page table to find it. The returned virtual address 849 * is proper in most cases, but it could be wrong when the application 850 * process has multiple entries mapping the error page. 851 */ 852 static int kill_accessing_process(struct task_struct *p, unsigned long pfn, 853 int flags) 854 { 855 int ret; 856 struct hwpoison_walk priv = { 857 .pfn = pfn, 858 }; 859 priv.tk.tsk = p; 860 861 if (!p->mm) 862 return -EFAULT; 863 864 mmap_read_lock(p->mm); 865 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwpoison_walk_ops, 866 (void *)&priv); 867 if (ret == 1 && priv.tk.addr) 868 kill_proc(&priv.tk, pfn, flags); 869 else 870 ret = 0; 871 mmap_read_unlock(p->mm); 872 return ret > 0 ? -EHWPOISON : -EFAULT; 873 } 874 875 static const char *action_name[] = { 876 [MF_IGNORED] = "Ignored", 877 [MF_FAILED] = "Failed", 878 [MF_DELAYED] = "Delayed", 879 [MF_RECOVERED] = "Recovered", 880 }; 881 882 static const char * const action_page_types[] = { 883 [MF_MSG_KERNEL] = "reserved kernel page", 884 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", 885 [MF_MSG_SLAB] = "kernel slab page", 886 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", 887 [MF_MSG_HUGE] = "huge page", 888 [MF_MSG_FREE_HUGE] = "free huge page", 889 [MF_MSG_UNMAP_FAILED] = "unmapping failed page", 890 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", 891 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", 892 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", 893 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", 894 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", 895 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", 896 [MF_MSG_DIRTY_LRU] = "dirty LRU page", 897 [MF_MSG_CLEAN_LRU] = "clean LRU page", 898 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", 899 [MF_MSG_BUDDY] = "free buddy page", 900 [MF_MSG_DAX] = "dax page", 901 [MF_MSG_UNSPLIT_THP] = "unsplit thp", 902 [MF_MSG_UNKNOWN] = "unknown page", 903 }; 904 905 /* 906 * XXX: It is possible that a page is isolated from LRU cache, 907 * and then kept in swap cache or failed to remove from page cache. 908 * The page count will stop it from being freed by unpoison. 909 * Stress tests should be aware of this memory leak problem. 910 */ 911 static int delete_from_lru_cache(struct folio *folio) 912 { 913 if (folio_isolate_lru(folio)) { 914 /* 915 * Clear sensible page flags, so that the buddy system won't 916 * complain when the folio is unpoison-and-freed. 917 */ 918 folio_clear_active(folio); 919 folio_clear_unevictable(folio); 920 921 /* 922 * Poisoned page might never drop its ref count to 0 so we have 923 * to uncharge it manually from its memcg. 924 */ 925 mem_cgroup_uncharge(folio); 926 927 /* 928 * drop the refcount elevated by folio_isolate_lru() 929 */ 930 folio_put(folio); 931 return 0; 932 } 933 return -EIO; 934 } 935 936 static int truncate_error_folio(struct folio *folio, unsigned long pfn, 937 struct address_space *mapping) 938 { 939 int ret = MF_FAILED; 940 941 if (mapping->a_ops->error_remove_folio) { 942 int err = mapping->a_ops->error_remove_folio(mapping, folio); 943 944 if (err != 0) 945 pr_info("%#lx: Failed to punch page: %d\n", pfn, err); 946 else if (!filemap_release_folio(folio, GFP_NOIO)) 947 pr_info("%#lx: failed to release buffers\n", pfn); 948 else 949 ret = MF_RECOVERED; 950 } else { 951 /* 952 * If the file system doesn't support it just invalidate 953 * This fails on dirty or anything with private pages 954 */ 955 if (mapping_evict_folio(mapping, folio)) 956 ret = MF_RECOVERED; 957 else 958 pr_info("%#lx: Failed to invalidate\n", pfn); 959 } 960 961 return ret; 962 } 963 964 struct page_state { 965 unsigned long mask; 966 unsigned long res; 967 enum mf_action_page_type type; 968 969 /* Callback ->action() has to unlock the relevant page inside it. */ 970 int (*action)(struct page_state *ps, struct page *p); 971 }; 972 973 /* 974 * Return true if page is still referenced by others, otherwise return 975 * false. 976 * 977 * The extra_pins is true when one extra refcount is expected. 978 */ 979 static bool has_extra_refcount(struct page_state *ps, struct page *p, 980 bool extra_pins) 981 { 982 int count = page_count(p) - 1; 983 984 if (extra_pins) 985 count -= folio_nr_pages(page_folio(p)); 986 987 if (count > 0) { 988 pr_err("%#lx: %s still referenced by %d users\n", 989 page_to_pfn(p), action_page_types[ps->type], count); 990 return true; 991 } 992 993 return false; 994 } 995 996 /* 997 * Error hit kernel page. 998 * Do nothing, try to be lucky and not touch this instead. For a few cases we 999 * could be more sophisticated. 1000 */ 1001 static int me_kernel(struct page_state *ps, struct page *p) 1002 { 1003 unlock_page(p); 1004 return MF_IGNORED; 1005 } 1006 1007 /* 1008 * Page in unknown state. Do nothing. 1009 */ 1010 static int me_unknown(struct page_state *ps, struct page *p) 1011 { 1012 pr_err("%#lx: Unknown page state\n", page_to_pfn(p)); 1013 unlock_page(p); 1014 return MF_FAILED; 1015 } 1016 1017 /* 1018 * Clean (or cleaned) page cache page. 1019 */ 1020 static int me_pagecache_clean(struct page_state *ps, struct page *p) 1021 { 1022 struct folio *folio = page_folio(p); 1023 int ret; 1024 struct address_space *mapping; 1025 bool extra_pins; 1026 1027 delete_from_lru_cache(folio); 1028 1029 /* 1030 * For anonymous folios the only reference left 1031 * should be the one m_f() holds. 1032 */ 1033 if (folio_test_anon(folio)) { 1034 ret = MF_RECOVERED; 1035 goto out; 1036 } 1037 1038 /* 1039 * Now truncate the page in the page cache. This is really 1040 * more like a "temporary hole punch" 1041 * Don't do this for block devices when someone else 1042 * has a reference, because it could be file system metadata 1043 * and that's not safe to truncate. 1044 */ 1045 mapping = folio_mapping(folio); 1046 if (!mapping) { 1047 /* Folio has been torn down in the meantime */ 1048 ret = MF_FAILED; 1049 goto out; 1050 } 1051 1052 /* 1053 * The shmem page is kept in page cache instead of truncating 1054 * so is expected to have an extra refcount after error-handling. 1055 */ 1056 extra_pins = shmem_mapping(mapping); 1057 1058 /* 1059 * Truncation is a bit tricky. Enable it per file system for now. 1060 * 1061 * Open: to take i_rwsem or not for this? Right now we don't. 1062 */ 1063 ret = truncate_error_folio(folio, page_to_pfn(p), mapping); 1064 if (has_extra_refcount(ps, p, extra_pins)) 1065 ret = MF_FAILED; 1066 1067 out: 1068 folio_unlock(folio); 1069 1070 return ret; 1071 } 1072 1073 /* 1074 * Dirty pagecache page 1075 * Issues: when the error hit a hole page the error is not properly 1076 * propagated. 1077 */ 1078 static int me_pagecache_dirty(struct page_state *ps, struct page *p) 1079 { 1080 struct address_space *mapping = page_mapping(p); 1081 1082 SetPageError(p); 1083 /* TBD: print more information about the file. */ 1084 if (mapping) { 1085 /* 1086 * IO error will be reported by write(), fsync(), etc. 1087 * who check the mapping. 1088 * This way the application knows that something went 1089 * wrong with its dirty file data. 1090 * 1091 * There's one open issue: 1092 * 1093 * The EIO will be only reported on the next IO 1094 * operation and then cleared through the IO map. 1095 * Normally Linux has two mechanisms to pass IO error 1096 * first through the AS_EIO flag in the address space 1097 * and then through the PageError flag in the page. 1098 * Since we drop pages on memory failure handling the 1099 * only mechanism open to use is through AS_AIO. 1100 * 1101 * This has the disadvantage that it gets cleared on 1102 * the first operation that returns an error, while 1103 * the PageError bit is more sticky and only cleared 1104 * when the page is reread or dropped. If an 1105 * application assumes it will always get error on 1106 * fsync, but does other operations on the fd before 1107 * and the page is dropped between then the error 1108 * will not be properly reported. 1109 * 1110 * This can already happen even without hwpoisoned 1111 * pages: first on metadata IO errors (which only 1112 * report through AS_EIO) or when the page is dropped 1113 * at the wrong time. 1114 * 1115 * So right now we assume that the application DTRT on 1116 * the first EIO, but we're not worse than other parts 1117 * of the kernel. 1118 */ 1119 mapping_set_error(mapping, -EIO); 1120 } 1121 1122 return me_pagecache_clean(ps, p); 1123 } 1124 1125 /* 1126 * Clean and dirty swap cache. 1127 * 1128 * Dirty swap cache page is tricky to handle. The page could live both in page 1129 * cache and swap cache(ie. page is freshly swapped in). So it could be 1130 * referenced concurrently by 2 types of PTEs: 1131 * normal PTEs and swap PTEs. We try to handle them consistently by calling 1132 * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs, 1133 * and then 1134 * - clear dirty bit to prevent IO 1135 * - remove from LRU 1136 * - but keep in the swap cache, so that when we return to it on 1137 * a later page fault, we know the application is accessing 1138 * corrupted data and shall be killed (we installed simple 1139 * interception code in do_swap_page to catch it). 1140 * 1141 * Clean swap cache pages can be directly isolated. A later page fault will 1142 * bring in the known good data from disk. 1143 */ 1144 static int me_swapcache_dirty(struct page_state *ps, struct page *p) 1145 { 1146 struct folio *folio = page_folio(p); 1147 int ret; 1148 bool extra_pins = false; 1149 1150 folio_clear_dirty(folio); 1151 /* Trigger EIO in shmem: */ 1152 folio_clear_uptodate(folio); 1153 1154 ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_DELAYED; 1155 folio_unlock(folio); 1156 1157 if (ret == MF_DELAYED) 1158 extra_pins = true; 1159 1160 if (has_extra_refcount(ps, p, extra_pins)) 1161 ret = MF_FAILED; 1162 1163 return ret; 1164 } 1165 1166 static int me_swapcache_clean(struct page_state *ps, struct page *p) 1167 { 1168 struct folio *folio = page_folio(p); 1169 int ret; 1170 1171 delete_from_swap_cache(folio); 1172 1173 ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_RECOVERED; 1174 folio_unlock(folio); 1175 1176 if (has_extra_refcount(ps, p, false)) 1177 ret = MF_FAILED; 1178 1179 return ret; 1180 } 1181 1182 /* 1183 * Huge pages. Needs work. 1184 * Issues: 1185 * - Error on hugepage is contained in hugepage unit (not in raw page unit.) 1186 * To narrow down kill region to one page, we need to break up pmd. 1187 */ 1188 static int me_huge_page(struct page_state *ps, struct page *p) 1189 { 1190 struct folio *folio = page_folio(p); 1191 int res; 1192 struct address_space *mapping; 1193 bool extra_pins = false; 1194 1195 mapping = folio_mapping(folio); 1196 if (mapping) { 1197 res = truncate_error_folio(folio, page_to_pfn(p), mapping); 1198 /* The page is kept in page cache. */ 1199 extra_pins = true; 1200 folio_unlock(folio); 1201 } else { 1202 folio_unlock(folio); 1203 /* 1204 * migration entry prevents later access on error hugepage, 1205 * so we can free and dissolve it into buddy to save healthy 1206 * subpages. 1207 */ 1208 folio_put(folio); 1209 if (__page_handle_poison(p) >= 0) { 1210 page_ref_inc(p); 1211 res = MF_RECOVERED; 1212 } else { 1213 res = MF_FAILED; 1214 } 1215 } 1216 1217 if (has_extra_refcount(ps, p, extra_pins)) 1218 res = MF_FAILED; 1219 1220 return res; 1221 } 1222 1223 /* 1224 * Various page states we can handle. 1225 * 1226 * A page state is defined by its current page->flags bits. 1227 * The table matches them in order and calls the right handler. 1228 * 1229 * This is quite tricky because we can access page at any time 1230 * in its live cycle, so all accesses have to be extremely careful. 1231 * 1232 * This is not complete. More states could be added. 1233 * For any missing state don't attempt recovery. 1234 */ 1235 1236 #define dirty (1UL << PG_dirty) 1237 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) 1238 #define unevict (1UL << PG_unevictable) 1239 #define mlock (1UL << PG_mlocked) 1240 #define lru (1UL << PG_lru) 1241 #define head (1UL << PG_head) 1242 #define slab (1UL << PG_slab) 1243 #define reserved (1UL << PG_reserved) 1244 1245 static struct page_state error_states[] = { 1246 { reserved, reserved, MF_MSG_KERNEL, me_kernel }, 1247 /* 1248 * free pages are specially detected outside this table: 1249 * PG_buddy pages only make a small fraction of all free pages. 1250 */ 1251 1252 /* 1253 * Could in theory check if slab page is free or if we can drop 1254 * currently unused objects without touching them. But just 1255 * treat it as standard kernel for now. 1256 */ 1257 { slab, slab, MF_MSG_SLAB, me_kernel }, 1258 1259 { head, head, MF_MSG_HUGE, me_huge_page }, 1260 1261 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, 1262 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, 1263 1264 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, 1265 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, 1266 1267 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, 1268 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, 1269 1270 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, 1271 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, 1272 1273 /* 1274 * Catchall entry: must be at end. 1275 */ 1276 { 0, 0, MF_MSG_UNKNOWN, me_unknown }, 1277 }; 1278 1279 #undef dirty 1280 #undef sc 1281 #undef unevict 1282 #undef mlock 1283 #undef lru 1284 #undef head 1285 #undef slab 1286 #undef reserved 1287 1288 static void update_per_node_mf_stats(unsigned long pfn, 1289 enum mf_result result) 1290 { 1291 int nid = MAX_NUMNODES; 1292 struct memory_failure_stats *mf_stats = NULL; 1293 1294 nid = pfn_to_nid(pfn); 1295 if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) { 1296 WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d", pfn, nid); 1297 return; 1298 } 1299 1300 mf_stats = &NODE_DATA(nid)->mf_stats; 1301 switch (result) { 1302 case MF_IGNORED: 1303 ++mf_stats->ignored; 1304 break; 1305 case MF_FAILED: 1306 ++mf_stats->failed; 1307 break; 1308 case MF_DELAYED: 1309 ++mf_stats->delayed; 1310 break; 1311 case MF_RECOVERED: 1312 ++mf_stats->recovered; 1313 break; 1314 default: 1315 WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled", result); 1316 break; 1317 } 1318 ++mf_stats->total; 1319 } 1320 1321 /* 1322 * "Dirty/Clean" indication is not 100% accurate due to the possibility of 1323 * setting PG_dirty outside page lock. See also comment above set_page_dirty(). 1324 */ 1325 static int action_result(unsigned long pfn, enum mf_action_page_type type, 1326 enum mf_result result) 1327 { 1328 trace_memory_failure_event(pfn, type, result); 1329 1330 num_poisoned_pages_inc(pfn); 1331 1332 update_per_node_mf_stats(pfn, result); 1333 1334 pr_err("%#lx: recovery action for %s: %s\n", 1335 pfn, action_page_types[type], action_name[result]); 1336 1337 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; 1338 } 1339 1340 static int page_action(struct page_state *ps, struct page *p, 1341 unsigned long pfn) 1342 { 1343 int result; 1344 1345 /* page p should be unlocked after returning from ps->action(). */ 1346 result = ps->action(ps, p); 1347 1348 /* Could do more checks here if page looks ok */ 1349 /* 1350 * Could adjust zone counters here to correct for the missing page. 1351 */ 1352 1353 return action_result(pfn, ps->type, result); 1354 } 1355 1356 static inline bool PageHWPoisonTakenOff(struct page *page) 1357 { 1358 return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; 1359 } 1360 1361 void SetPageHWPoisonTakenOff(struct page *page) 1362 { 1363 set_page_private(page, MAGIC_HWPOISON); 1364 } 1365 1366 void ClearPageHWPoisonTakenOff(struct page *page) 1367 { 1368 if (PageHWPoison(page)) 1369 set_page_private(page, 0); 1370 } 1371 1372 /* 1373 * Return true if a page type of a given page is supported by hwpoison 1374 * mechanism (while handling could fail), otherwise false. This function 1375 * does not return true for hugetlb or device memory pages, so it's assumed 1376 * to be called only in the context where we never have such pages. 1377 */ 1378 static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) 1379 { 1380 /* Soft offline could migrate non-LRU movable pages */ 1381 if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) 1382 return true; 1383 1384 return PageLRU(page) || is_free_buddy_page(page); 1385 } 1386 1387 static int __get_hwpoison_page(struct page *page, unsigned long flags) 1388 { 1389 struct folio *folio = page_folio(page); 1390 int ret = 0; 1391 bool hugetlb = false; 1392 1393 ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false); 1394 if (hugetlb) { 1395 /* Make sure hugetlb demotion did not happen from under us. */ 1396 if (folio == page_folio(page)) 1397 return ret; 1398 if (ret > 0) { 1399 folio_put(folio); 1400 folio = page_folio(page); 1401 } 1402 } 1403 1404 /* 1405 * This check prevents from calling folio_try_get() for any 1406 * unsupported type of folio in order to reduce the risk of unexpected 1407 * races caused by taking a folio refcount. 1408 */ 1409 if (!HWPoisonHandlable(&folio->page, flags)) 1410 return -EBUSY; 1411 1412 if (folio_try_get(folio)) { 1413 if (folio == page_folio(page)) 1414 return 1; 1415 1416 pr_info("%#lx cannot catch tail\n", page_to_pfn(page)); 1417 folio_put(folio); 1418 } 1419 1420 return 0; 1421 } 1422 1423 static int get_any_page(struct page *p, unsigned long flags) 1424 { 1425 int ret = 0, pass = 0; 1426 bool count_increased = false; 1427 1428 if (flags & MF_COUNT_INCREASED) 1429 count_increased = true; 1430 1431 try_again: 1432 if (!count_increased) { 1433 ret = __get_hwpoison_page(p, flags); 1434 if (!ret) { 1435 if (page_count(p)) { 1436 /* We raced with an allocation, retry. */ 1437 if (pass++ < 3) 1438 goto try_again; 1439 ret = -EBUSY; 1440 } else if (!PageHuge(p) && !is_free_buddy_page(p)) { 1441 /* We raced with put_page, retry. */ 1442 if (pass++ < 3) 1443 goto try_again; 1444 ret = -EIO; 1445 } 1446 goto out; 1447 } else if (ret == -EBUSY) { 1448 /* 1449 * We raced with (possibly temporary) unhandlable 1450 * page, retry. 1451 */ 1452 if (pass++ < 3) { 1453 shake_page(p); 1454 goto try_again; 1455 } 1456 ret = -EIO; 1457 goto out; 1458 } 1459 } 1460 1461 if (PageHuge(p) || HWPoisonHandlable(p, flags)) { 1462 ret = 1; 1463 } else { 1464 /* 1465 * A page we cannot handle. Check whether we can turn 1466 * it into something we can handle. 1467 */ 1468 if (pass++ < 3) { 1469 put_page(p); 1470 shake_page(p); 1471 count_increased = false; 1472 goto try_again; 1473 } 1474 put_page(p); 1475 ret = -EIO; 1476 } 1477 out: 1478 if (ret == -EIO) 1479 pr_err("%#lx: unhandlable page.\n", page_to_pfn(p)); 1480 1481 return ret; 1482 } 1483 1484 static int __get_unpoison_page(struct page *page) 1485 { 1486 struct folio *folio = page_folio(page); 1487 int ret = 0; 1488 bool hugetlb = false; 1489 1490 ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, true); 1491 if (hugetlb) { 1492 /* Make sure hugetlb demotion did not happen from under us. */ 1493 if (folio == page_folio(page)) 1494 return ret; 1495 if (ret > 0) 1496 folio_put(folio); 1497 } 1498 1499 /* 1500 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, 1501 * but also isolated from buddy freelist, so need to identify the 1502 * state and have to cancel both operations to unpoison. 1503 */ 1504 if (PageHWPoisonTakenOff(page)) 1505 return -EHWPOISON; 1506 1507 return get_page_unless_zero(page) ? 1 : 0; 1508 } 1509 1510 /** 1511 * get_hwpoison_page() - Get refcount for memory error handling 1512 * @p: Raw error page (hit by memory error) 1513 * @flags: Flags controlling behavior of error handling 1514 * 1515 * get_hwpoison_page() takes a page refcount of an error page to handle memory 1516 * error on it, after checking that the error page is in a well-defined state 1517 * (defined as a page-type we can successfully handle the memory error on it, 1518 * such as LRU page and hugetlb page). 1519 * 1520 * Memory error handling could be triggered at any time on any type of page, 1521 * so it's prone to race with typical memory management lifecycle (like 1522 * allocation and free). So to avoid such races, get_hwpoison_page() takes 1523 * extra care for the error page's state (as done in __get_hwpoison_page()), 1524 * and has some retry logic in get_any_page(). 1525 * 1526 * When called from unpoison_memory(), the caller should already ensure that 1527 * the given page has PG_hwpoison. So it's never reused for other page 1528 * allocations, and __get_unpoison_page() never races with them. 1529 * 1530 * Return: 0 on failure, 1531 * 1 on success for in-use pages in a well-defined state, 1532 * -EIO for pages on which we can not handle memory errors, 1533 * -EBUSY when get_hwpoison_page() has raced with page lifecycle 1534 * operations like allocation and free, 1535 * -EHWPOISON when the page is hwpoisoned and taken off from buddy. 1536 */ 1537 static int get_hwpoison_page(struct page *p, unsigned long flags) 1538 { 1539 int ret; 1540 1541 zone_pcp_disable(page_zone(p)); 1542 if (flags & MF_UNPOISON) 1543 ret = __get_unpoison_page(p); 1544 else 1545 ret = get_any_page(p, flags); 1546 zone_pcp_enable(page_zone(p)); 1547 1548 return ret; 1549 } 1550 1551 /* 1552 * Do all that is necessary to remove user space mappings. Unmap 1553 * the pages and send SIGBUS to the processes if the data was dirty. 1554 */ 1555 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, 1556 int flags, struct page *hpage) 1557 { 1558 struct folio *folio = page_folio(hpage); 1559 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON; 1560 struct address_space *mapping; 1561 LIST_HEAD(tokill); 1562 bool unmap_success; 1563 int forcekill; 1564 bool mlocked = PageMlocked(hpage); 1565 1566 /* 1567 * Here we are interested only in user-mapped pages, so skip any 1568 * other types of pages. 1569 */ 1570 if (PageReserved(p) || PageSlab(p) || PageTable(p) || PageOffline(p)) 1571 return true; 1572 if (!(PageLRU(hpage) || PageHuge(p))) 1573 return true; 1574 1575 /* 1576 * This check implies we don't kill processes if their pages 1577 * are in the swap cache early. Those are always late kills. 1578 */ 1579 if (!page_mapped(p)) 1580 return true; 1581 1582 if (PageSwapCache(p)) { 1583 pr_err("%#lx: keeping poisoned page in swap cache\n", pfn); 1584 ttu &= ~TTU_HWPOISON; 1585 } 1586 1587 /* 1588 * Propagate the dirty bit from PTEs to struct page first, because we 1589 * need this to decide if we should kill or just drop the page. 1590 * XXX: the dirty test could be racy: set_page_dirty() may not always 1591 * be called inside page lock (it's recommended but not enforced). 1592 */ 1593 mapping = page_mapping(hpage); 1594 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && 1595 mapping_can_writeback(mapping)) { 1596 if (page_mkclean(hpage)) { 1597 SetPageDirty(hpage); 1598 } else { 1599 ttu &= ~TTU_HWPOISON; 1600 pr_info("%#lx: corrupted page was clean: dropped without side effects\n", 1601 pfn); 1602 } 1603 } 1604 1605 /* 1606 * First collect all the processes that have the page 1607 * mapped in dirty form. This has to be done before try_to_unmap, 1608 * because ttu takes the rmap data structures down. 1609 */ 1610 collect_procs(folio, p, &tokill, flags & MF_ACTION_REQUIRED); 1611 1612 if (PageHuge(hpage) && !PageAnon(hpage)) { 1613 /* 1614 * For hugetlb pages in shared mappings, try_to_unmap 1615 * could potentially call huge_pmd_unshare. Because of 1616 * this, take semaphore in write mode here and set 1617 * TTU_RMAP_LOCKED to indicate we have taken the lock 1618 * at this higher level. 1619 */ 1620 mapping = hugetlb_page_mapping_lock_write(hpage); 1621 if (mapping) { 1622 try_to_unmap(folio, ttu|TTU_RMAP_LOCKED); 1623 i_mmap_unlock_write(mapping); 1624 } else 1625 pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn); 1626 } else { 1627 try_to_unmap(folio, ttu); 1628 } 1629 1630 unmap_success = !page_mapped(p); 1631 if (!unmap_success) 1632 pr_err("%#lx: failed to unmap page (mapcount=%d)\n", 1633 pfn, page_mapcount(p)); 1634 1635 /* 1636 * try_to_unmap() might put mlocked page in lru cache, so call 1637 * shake_page() again to ensure that it's flushed. 1638 */ 1639 if (mlocked) 1640 shake_page(hpage); 1641 1642 /* 1643 * Now that the dirty bit has been propagated to the 1644 * struct page and all unmaps done we can decide if 1645 * killing is needed or not. Only kill when the page 1646 * was dirty or the process is not restartable, 1647 * otherwise the tokill list is merely 1648 * freed. When there was a problem unmapping earlier 1649 * use a more force-full uncatchable kill to prevent 1650 * any accesses to the poisoned memory. 1651 */ 1652 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) || 1653 !unmap_success; 1654 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); 1655 1656 return unmap_success; 1657 } 1658 1659 static int identify_page_state(unsigned long pfn, struct page *p, 1660 unsigned long page_flags) 1661 { 1662 struct page_state *ps; 1663 1664 /* 1665 * The first check uses the current page flags which may not have any 1666 * relevant information. The second check with the saved page flags is 1667 * carried out only if the first check can't determine the page status. 1668 */ 1669 for (ps = error_states;; ps++) 1670 if ((p->flags & ps->mask) == ps->res) 1671 break; 1672 1673 page_flags |= (p->flags & (1UL << PG_dirty)); 1674 1675 if (!ps->mask) 1676 for (ps = error_states;; ps++) 1677 if ((page_flags & ps->mask) == ps->res) 1678 break; 1679 return page_action(ps, p, pfn); 1680 } 1681 1682 static int try_to_split_thp_page(struct page *page) 1683 { 1684 int ret; 1685 1686 lock_page(page); 1687 ret = split_huge_page(page); 1688 unlock_page(page); 1689 1690 if (unlikely(ret)) 1691 put_page(page); 1692 1693 return ret; 1694 } 1695 1696 static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn, 1697 struct address_space *mapping, pgoff_t index, int flags) 1698 { 1699 struct to_kill *tk; 1700 unsigned long size = 0; 1701 1702 list_for_each_entry(tk, to_kill, nd) 1703 if (tk->size_shift) 1704 size = max(size, 1UL << tk->size_shift); 1705 1706 if (size) { 1707 /* 1708 * Unmap the largest mapping to avoid breaking up device-dax 1709 * mappings which are constant size. The actual size of the 1710 * mapping being torn down is communicated in siginfo, see 1711 * kill_proc() 1712 */ 1713 loff_t start = ((loff_t)index << PAGE_SHIFT) & ~(size - 1); 1714 1715 unmap_mapping_range(mapping, start, size, 0); 1716 } 1717 1718 kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags); 1719 } 1720 1721 /* 1722 * Only dev_pagemap pages get here, such as fsdax when the filesystem 1723 * either do not claim or fails to claim a hwpoison event, or devdax. 1724 * The fsdax pages are initialized per base page, and the devdax pages 1725 * could be initialized either as base pages, or as compound pages with 1726 * vmemmap optimization enabled. Devdax is simplistic in its dealing with 1727 * hwpoison, such that, if a subpage of a compound page is poisoned, 1728 * simply mark the compound head page is by far sufficient. 1729 */ 1730 static int mf_generic_kill_procs(unsigned long long pfn, int flags, 1731 struct dev_pagemap *pgmap) 1732 { 1733 struct folio *folio = pfn_folio(pfn); 1734 LIST_HEAD(to_kill); 1735 dax_entry_t cookie; 1736 int rc = 0; 1737 1738 /* 1739 * Prevent the inode from being freed while we are interrogating 1740 * the address_space, typically this would be handled by 1741 * lock_page(), but dax pages do not use the page lock. This 1742 * also prevents changes to the mapping of this pfn until 1743 * poison signaling is complete. 1744 */ 1745 cookie = dax_lock_folio(folio); 1746 if (!cookie) 1747 return -EBUSY; 1748 1749 if (hwpoison_filter(&folio->page)) { 1750 rc = -EOPNOTSUPP; 1751 goto unlock; 1752 } 1753 1754 switch (pgmap->type) { 1755 case MEMORY_DEVICE_PRIVATE: 1756 case MEMORY_DEVICE_COHERENT: 1757 /* 1758 * TODO: Handle device pages which may need coordination 1759 * with device-side memory. 1760 */ 1761 rc = -ENXIO; 1762 goto unlock; 1763 default: 1764 break; 1765 } 1766 1767 /* 1768 * Use this flag as an indication that the dax page has been 1769 * remapped UC to prevent speculative consumption of poison. 1770 */ 1771 SetPageHWPoison(&folio->page); 1772 1773 /* 1774 * Unlike System-RAM there is no possibility to swap in a 1775 * different physical page at a given virtual address, so all 1776 * userspace consumption of ZONE_DEVICE memory necessitates 1777 * SIGBUS (i.e. MF_MUST_KILL) 1778 */ 1779 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; 1780 collect_procs(folio, &folio->page, &to_kill, true); 1781 1782 unmap_and_kill(&to_kill, pfn, folio->mapping, folio->index, flags); 1783 unlock: 1784 dax_unlock_folio(folio, cookie); 1785 return rc; 1786 } 1787 1788 #ifdef CONFIG_FS_DAX 1789 /** 1790 * mf_dax_kill_procs - Collect and kill processes who are using this file range 1791 * @mapping: address_space of the file in use 1792 * @index: start pgoff of the range within the file 1793 * @count: length of the range, in unit of PAGE_SIZE 1794 * @mf_flags: memory failure flags 1795 */ 1796 int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index, 1797 unsigned long count, int mf_flags) 1798 { 1799 LIST_HEAD(to_kill); 1800 dax_entry_t cookie; 1801 struct page *page; 1802 size_t end = index + count; 1803 bool pre_remove = mf_flags & MF_MEM_PRE_REMOVE; 1804 1805 mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; 1806 1807 for (; index < end; index++) { 1808 page = NULL; 1809 cookie = dax_lock_mapping_entry(mapping, index, &page); 1810 if (!cookie) 1811 return -EBUSY; 1812 if (!page) 1813 goto unlock; 1814 1815 if (!pre_remove) 1816 SetPageHWPoison(page); 1817 1818 /* 1819 * The pre_remove case is revoking access, the memory is still 1820 * good and could theoretically be put back into service. 1821 */ 1822 collect_procs_fsdax(page, mapping, index, &to_kill, pre_remove); 1823 unmap_and_kill(&to_kill, page_to_pfn(page), mapping, 1824 index, mf_flags); 1825 unlock: 1826 dax_unlock_mapping_entry(mapping, index, cookie); 1827 } 1828 return 0; 1829 } 1830 EXPORT_SYMBOL_GPL(mf_dax_kill_procs); 1831 #endif /* CONFIG_FS_DAX */ 1832 1833 #ifdef CONFIG_HUGETLB_PAGE 1834 1835 /* 1836 * Struct raw_hwp_page represents information about "raw error page", 1837 * constructing singly linked list from ->_hugetlb_hwpoison field of folio. 1838 */ 1839 struct raw_hwp_page { 1840 struct llist_node node; 1841 struct page *page; 1842 }; 1843 1844 static inline struct llist_head *raw_hwp_list_head(struct folio *folio) 1845 { 1846 return (struct llist_head *)&folio->_hugetlb_hwpoison; 1847 } 1848 1849 bool is_raw_hwpoison_page_in_hugepage(struct page *page) 1850 { 1851 struct llist_head *raw_hwp_head; 1852 struct raw_hwp_page *p; 1853 struct folio *folio = page_folio(page); 1854 bool ret = false; 1855 1856 if (!folio_test_hwpoison(folio)) 1857 return false; 1858 1859 if (!folio_test_hugetlb(folio)) 1860 return PageHWPoison(page); 1861 1862 /* 1863 * When RawHwpUnreliable is set, kernel lost track of which subpages 1864 * are HWPOISON. So return as if ALL subpages are HWPOISONed. 1865 */ 1866 if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1867 return true; 1868 1869 mutex_lock(&mf_mutex); 1870 1871 raw_hwp_head = raw_hwp_list_head(folio); 1872 llist_for_each_entry(p, raw_hwp_head->first, node) { 1873 if (page == p->page) { 1874 ret = true; 1875 break; 1876 } 1877 } 1878 1879 mutex_unlock(&mf_mutex); 1880 1881 return ret; 1882 } 1883 1884 static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag) 1885 { 1886 struct llist_node *head; 1887 struct raw_hwp_page *p, *next; 1888 unsigned long count = 0; 1889 1890 head = llist_del_all(raw_hwp_list_head(folio)); 1891 llist_for_each_entry_safe(p, next, head, node) { 1892 if (move_flag) 1893 SetPageHWPoison(p->page); 1894 else 1895 num_poisoned_pages_sub(page_to_pfn(p->page), 1); 1896 kfree(p); 1897 count++; 1898 } 1899 return count; 1900 } 1901 1902 static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page) 1903 { 1904 struct llist_head *head; 1905 struct raw_hwp_page *raw_hwp; 1906 struct raw_hwp_page *p, *next; 1907 int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0; 1908 1909 /* 1910 * Once the hwpoison hugepage has lost reliable raw error info, 1911 * there is little meaning to keep additional error info precisely, 1912 * so skip to add additional raw error info. 1913 */ 1914 if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1915 return -EHWPOISON; 1916 head = raw_hwp_list_head(folio); 1917 llist_for_each_entry_safe(p, next, head->first, node) { 1918 if (p->page == page) 1919 return -EHWPOISON; 1920 } 1921 1922 raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC); 1923 if (raw_hwp) { 1924 raw_hwp->page = page; 1925 llist_add(&raw_hwp->node, head); 1926 /* the first error event will be counted in action_result(). */ 1927 if (ret) 1928 num_poisoned_pages_inc(page_to_pfn(page)); 1929 } else { 1930 /* 1931 * Failed to save raw error info. We no longer trace all 1932 * hwpoisoned subpages, and we need refuse to free/dissolve 1933 * this hwpoisoned hugepage. 1934 */ 1935 folio_set_hugetlb_raw_hwp_unreliable(folio); 1936 /* 1937 * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not 1938 * used any more, so free it. 1939 */ 1940 __folio_free_raw_hwp(folio, false); 1941 } 1942 return ret; 1943 } 1944 1945 static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag) 1946 { 1947 /* 1948 * hugetlb_vmemmap_optimized hugepages can't be freed because struct 1949 * pages for tail pages are required but they don't exist. 1950 */ 1951 if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio)) 1952 return 0; 1953 1954 /* 1955 * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by 1956 * definition. 1957 */ 1958 if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1959 return 0; 1960 1961 return __folio_free_raw_hwp(folio, move_flag); 1962 } 1963 1964 void folio_clear_hugetlb_hwpoison(struct folio *folio) 1965 { 1966 if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1967 return; 1968 if (folio_test_hugetlb_vmemmap_optimized(folio)) 1969 return; 1970 folio_clear_hwpoison(folio); 1971 folio_free_raw_hwp(folio, true); 1972 } 1973 1974 /* 1975 * Called from hugetlb code with hugetlb_lock held. 1976 * 1977 * Return values: 1978 * 0 - free hugepage 1979 * 1 - in-use hugepage 1980 * 2 - not a hugepage 1981 * -EBUSY - the hugepage is busy (try to retry) 1982 * -EHWPOISON - the hugepage is already hwpoisoned 1983 */ 1984 int __get_huge_page_for_hwpoison(unsigned long pfn, int flags, 1985 bool *migratable_cleared) 1986 { 1987 struct page *page = pfn_to_page(pfn); 1988 struct folio *folio = page_folio(page); 1989 int ret = 2; /* fallback to normal page handling */ 1990 bool count_increased = false; 1991 1992 if (!folio_test_hugetlb(folio)) 1993 goto out; 1994 1995 if (flags & MF_COUNT_INCREASED) { 1996 ret = 1; 1997 count_increased = true; 1998 } else if (folio_test_hugetlb_freed(folio)) { 1999 ret = 0; 2000 } else if (folio_test_hugetlb_migratable(folio)) { 2001 ret = folio_try_get(folio); 2002 if (ret) 2003 count_increased = true; 2004 } else { 2005 ret = -EBUSY; 2006 if (!(flags & MF_NO_RETRY)) 2007 goto out; 2008 } 2009 2010 if (folio_set_hugetlb_hwpoison(folio, page)) { 2011 ret = -EHWPOISON; 2012 goto out; 2013 } 2014 2015 /* 2016 * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them 2017 * from being migrated by memory hotremove. 2018 */ 2019 if (count_increased && folio_test_hugetlb_migratable(folio)) { 2020 folio_clear_hugetlb_migratable(folio); 2021 *migratable_cleared = true; 2022 } 2023 2024 return ret; 2025 out: 2026 if (count_increased) 2027 folio_put(folio); 2028 return ret; 2029 } 2030 2031 /* 2032 * Taking refcount of hugetlb pages needs extra care about race conditions 2033 * with basic operations like hugepage allocation/free/demotion. 2034 * So some of prechecks for hwpoison (pinning, and testing/setting 2035 * PageHWPoison) should be done in single hugetlb_lock range. 2036 */ 2037 static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) 2038 { 2039 int res; 2040 struct page *p = pfn_to_page(pfn); 2041 struct folio *folio; 2042 unsigned long page_flags; 2043 bool migratable_cleared = false; 2044 2045 *hugetlb = 1; 2046 retry: 2047 res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared); 2048 if (res == 2) { /* fallback to normal page handling */ 2049 *hugetlb = 0; 2050 return 0; 2051 } else if (res == -EHWPOISON) { 2052 pr_err("%#lx: already hardware poisoned\n", pfn); 2053 if (flags & MF_ACTION_REQUIRED) { 2054 folio = page_folio(p); 2055 res = kill_accessing_process(current, folio_pfn(folio), flags); 2056 } 2057 return res; 2058 } else if (res == -EBUSY) { 2059 if (!(flags & MF_NO_RETRY)) { 2060 flags |= MF_NO_RETRY; 2061 goto retry; 2062 } 2063 return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); 2064 } 2065 2066 folio = page_folio(p); 2067 folio_lock(folio); 2068 2069 if (hwpoison_filter(p)) { 2070 folio_clear_hugetlb_hwpoison(folio); 2071 if (migratable_cleared) 2072 folio_set_hugetlb_migratable(folio); 2073 folio_unlock(folio); 2074 if (res == 1) 2075 folio_put(folio); 2076 return -EOPNOTSUPP; 2077 } 2078 2079 /* 2080 * Handling free hugepage. The possible race with hugepage allocation 2081 * or demotion can be prevented by PageHWPoison flag. 2082 */ 2083 if (res == 0) { 2084 folio_unlock(folio); 2085 if (__page_handle_poison(p) >= 0) { 2086 page_ref_inc(p); 2087 res = MF_RECOVERED; 2088 } else { 2089 res = MF_FAILED; 2090 } 2091 return action_result(pfn, MF_MSG_FREE_HUGE, res); 2092 } 2093 2094 page_flags = folio->flags; 2095 2096 if (!hwpoison_user_mappings(p, pfn, flags, &folio->page)) { 2097 folio_unlock(folio); 2098 return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); 2099 } 2100 2101 return identify_page_state(pfn, p, page_flags); 2102 } 2103 2104 #else 2105 static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) 2106 { 2107 return 0; 2108 } 2109 2110 static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag) 2111 { 2112 return 0; 2113 } 2114 #endif /* CONFIG_HUGETLB_PAGE */ 2115 2116 /* Drop the extra refcount in case we come from madvise() */ 2117 static void put_ref_page(unsigned long pfn, int flags) 2118 { 2119 struct page *page; 2120 2121 if (!(flags & MF_COUNT_INCREASED)) 2122 return; 2123 2124 page = pfn_to_page(pfn); 2125 if (page) 2126 put_page(page); 2127 } 2128 2129 static int memory_failure_dev_pagemap(unsigned long pfn, int flags, 2130 struct dev_pagemap *pgmap) 2131 { 2132 int rc = -ENXIO; 2133 2134 /* device metadata space is not recoverable */ 2135 if (!pgmap_pfn_valid(pgmap, pfn)) 2136 goto out; 2137 2138 /* 2139 * Call driver's implementation to handle the memory failure, otherwise 2140 * fall back to generic handler. 2141 */ 2142 if (pgmap_has_memory_failure(pgmap)) { 2143 rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags); 2144 /* 2145 * Fall back to generic handler too if operation is not 2146 * supported inside the driver/device/filesystem. 2147 */ 2148 if (rc != -EOPNOTSUPP) 2149 goto out; 2150 } 2151 2152 rc = mf_generic_kill_procs(pfn, flags, pgmap); 2153 out: 2154 /* drop pgmap ref acquired in caller */ 2155 put_dev_pagemap(pgmap); 2156 if (rc != -EOPNOTSUPP) 2157 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); 2158 return rc; 2159 } 2160 2161 /** 2162 * memory_failure - Handle memory failure of a page. 2163 * @pfn: Page Number of the corrupted page 2164 * @flags: fine tune action taken 2165 * 2166 * This function is called by the low level machine check code 2167 * of an architecture when it detects hardware memory corruption 2168 * of a page. It tries its best to recover, which includes 2169 * dropping pages, killing processes etc. 2170 * 2171 * The function is primarily of use for corruptions that 2172 * happen outside the current execution context (e.g. when 2173 * detected by a background scrubber) 2174 * 2175 * Must run in process context (e.g. a work queue) with interrupts 2176 * enabled and no spinlocks held. 2177 * 2178 * Return: 0 for successfully handled the memory error, 2179 * -EOPNOTSUPP for hwpoison_filter() filtered the error event, 2180 * < 0(except -EOPNOTSUPP) on failure. 2181 */ 2182 int memory_failure(unsigned long pfn, int flags) 2183 { 2184 struct page *p; 2185 struct page *hpage; 2186 struct dev_pagemap *pgmap; 2187 int res = 0; 2188 unsigned long page_flags; 2189 bool retry = true; 2190 int hugetlb = 0; 2191 2192 if (!sysctl_memory_failure_recovery) 2193 panic("Memory failure on page %lx", pfn); 2194 2195 mutex_lock(&mf_mutex); 2196 2197 if (!(flags & MF_SW_SIMULATED)) 2198 hw_memory_failure = true; 2199 2200 p = pfn_to_online_page(pfn); 2201 if (!p) { 2202 res = arch_memory_failure(pfn, flags); 2203 if (res == 0) 2204 goto unlock_mutex; 2205 2206 if (pfn_valid(pfn)) { 2207 pgmap = get_dev_pagemap(pfn, NULL); 2208 put_ref_page(pfn, flags); 2209 if (pgmap) { 2210 res = memory_failure_dev_pagemap(pfn, flags, 2211 pgmap); 2212 goto unlock_mutex; 2213 } 2214 } 2215 pr_err("%#lx: memory outside kernel control\n", pfn); 2216 res = -ENXIO; 2217 goto unlock_mutex; 2218 } 2219 2220 try_again: 2221 res = try_memory_failure_hugetlb(pfn, flags, &hugetlb); 2222 if (hugetlb) 2223 goto unlock_mutex; 2224 2225 if (TestSetPageHWPoison(p)) { 2226 pr_err("%#lx: already hardware poisoned\n", pfn); 2227 res = -EHWPOISON; 2228 if (flags & MF_ACTION_REQUIRED) 2229 res = kill_accessing_process(current, pfn, flags); 2230 if (flags & MF_COUNT_INCREASED) 2231 put_page(p); 2232 goto unlock_mutex; 2233 } 2234 2235 /* 2236 * We need/can do nothing about count=0 pages. 2237 * 1) it's a free page, and therefore in safe hand: 2238 * check_new_page() will be the gate keeper. 2239 * 2) it's part of a non-compound high order page. 2240 * Implies some kernel user: cannot stop them from 2241 * R/W the page; let's pray that the page has been 2242 * used and will be freed some time later. 2243 * In fact it's dangerous to directly bump up page count from 0, 2244 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. 2245 */ 2246 if (!(flags & MF_COUNT_INCREASED)) { 2247 res = get_hwpoison_page(p, flags); 2248 if (!res) { 2249 if (is_free_buddy_page(p)) { 2250 if (take_page_off_buddy(p)) { 2251 page_ref_inc(p); 2252 res = MF_RECOVERED; 2253 } else { 2254 /* We lost the race, try again */ 2255 if (retry) { 2256 ClearPageHWPoison(p); 2257 retry = false; 2258 goto try_again; 2259 } 2260 res = MF_FAILED; 2261 } 2262 res = action_result(pfn, MF_MSG_BUDDY, res); 2263 } else { 2264 res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); 2265 } 2266 goto unlock_mutex; 2267 } else if (res < 0) { 2268 res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); 2269 goto unlock_mutex; 2270 } 2271 } 2272 2273 hpage = compound_head(p); 2274 if (PageTransHuge(hpage)) { 2275 /* 2276 * The flag must be set after the refcount is bumped 2277 * otherwise it may race with THP split. 2278 * And the flag can't be set in get_hwpoison_page() since 2279 * it is called by soft offline too and it is just called 2280 * for !MF_COUNT_INCREASED. So here seems to be the best 2281 * place. 2282 * 2283 * Don't need care about the above error handling paths for 2284 * get_hwpoison_page() since they handle either free page 2285 * or unhandlable page. The refcount is bumped iff the 2286 * page is a valid handlable page. 2287 */ 2288 SetPageHasHWPoisoned(hpage); 2289 if (try_to_split_thp_page(p) < 0) { 2290 res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); 2291 goto unlock_mutex; 2292 } 2293 VM_BUG_ON_PAGE(!page_count(p), p); 2294 } 2295 2296 /* 2297 * We ignore non-LRU pages for good reasons. 2298 * - PG_locked is only well defined for LRU pages and a few others 2299 * - to avoid races with __SetPageLocked() 2300 * - to avoid races with __SetPageSlab*() (and more non-atomic ops) 2301 * The check (unnecessarily) ignores LRU pages being isolated and 2302 * walked by the page reclaim code, however that's not a big loss. 2303 */ 2304 shake_page(p); 2305 2306 lock_page(p); 2307 2308 /* 2309 * We're only intended to deal with the non-Compound page here. 2310 * However, the page could have changed compound pages due to 2311 * race window. If this happens, we could try again to hopefully 2312 * handle the page next round. 2313 */ 2314 if (PageCompound(p)) { 2315 if (retry) { 2316 ClearPageHWPoison(p); 2317 unlock_page(p); 2318 put_page(p); 2319 flags &= ~MF_COUNT_INCREASED; 2320 retry = false; 2321 goto try_again; 2322 } 2323 res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); 2324 goto unlock_page; 2325 } 2326 2327 /* 2328 * We use page flags to determine what action should be taken, but 2329 * the flags can be modified by the error containment action. One 2330 * example is an mlocked page, where PG_mlocked is cleared by 2331 * folio_remove_rmap_*() in try_to_unmap_one(). So to determine page 2332 * status correctly, we save a copy of the page flags at this time. 2333 */ 2334 page_flags = p->flags; 2335 2336 if (hwpoison_filter(p)) { 2337 ClearPageHWPoison(p); 2338 unlock_page(p); 2339 put_page(p); 2340 res = -EOPNOTSUPP; 2341 goto unlock_mutex; 2342 } 2343 2344 /* 2345 * __munlock_folio() may clear a writeback page's LRU flag without 2346 * page_lock. We need wait writeback completion for this page or it 2347 * may trigger vfs BUG while evict inode. 2348 */ 2349 if (!PageLRU(p) && !PageWriteback(p)) 2350 goto identify_page_state; 2351 2352 /* 2353 * It's very difficult to mess with pages currently under IO 2354 * and in many cases impossible, so we just avoid it here. 2355 */ 2356 wait_on_page_writeback(p); 2357 2358 /* 2359 * Now take care of user space mappings. 2360 * Abort on fail: __filemap_remove_folio() assumes unmapped page. 2361 */ 2362 if (!hwpoison_user_mappings(p, pfn, flags, p)) { 2363 res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); 2364 goto unlock_page; 2365 } 2366 2367 /* 2368 * Torn down by someone else? 2369 */ 2370 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { 2371 res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); 2372 goto unlock_page; 2373 } 2374 2375 identify_page_state: 2376 res = identify_page_state(pfn, p, page_flags); 2377 mutex_unlock(&mf_mutex); 2378 return res; 2379 unlock_page: 2380 unlock_page(p); 2381 unlock_mutex: 2382 mutex_unlock(&mf_mutex); 2383 return res; 2384 } 2385 EXPORT_SYMBOL_GPL(memory_failure); 2386 2387 #define MEMORY_FAILURE_FIFO_ORDER 4 2388 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) 2389 2390 struct memory_failure_entry { 2391 unsigned long pfn; 2392 int flags; 2393 }; 2394 2395 struct memory_failure_cpu { 2396 DECLARE_KFIFO(fifo, struct memory_failure_entry, 2397 MEMORY_FAILURE_FIFO_SIZE); 2398 spinlock_t lock; 2399 struct work_struct work; 2400 }; 2401 2402 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); 2403 2404 /** 2405 * memory_failure_queue - Schedule handling memory failure of a page. 2406 * @pfn: Page Number of the corrupted page 2407 * @flags: Flags for memory failure handling 2408 * 2409 * This function is called by the low level hardware error handler 2410 * when it detects hardware memory corruption of a page. It schedules 2411 * the recovering of error page, including dropping pages, killing 2412 * processes etc. 2413 * 2414 * The function is primarily of use for corruptions that 2415 * happen outside the current execution context (e.g. when 2416 * detected by a background scrubber) 2417 * 2418 * Can run in IRQ context. 2419 */ 2420 void memory_failure_queue(unsigned long pfn, int flags) 2421 { 2422 struct memory_failure_cpu *mf_cpu; 2423 unsigned long proc_flags; 2424 struct memory_failure_entry entry = { 2425 .pfn = pfn, 2426 .flags = flags, 2427 }; 2428 2429 mf_cpu = &get_cpu_var(memory_failure_cpu); 2430 spin_lock_irqsave(&mf_cpu->lock, proc_flags); 2431 if (kfifo_put(&mf_cpu->fifo, entry)) 2432 schedule_work_on(smp_processor_id(), &mf_cpu->work); 2433 else 2434 pr_err("buffer overflow when queuing memory failure at %#lx\n", 2435 pfn); 2436 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); 2437 put_cpu_var(memory_failure_cpu); 2438 } 2439 EXPORT_SYMBOL_GPL(memory_failure_queue); 2440 2441 static void memory_failure_work_func(struct work_struct *work) 2442 { 2443 struct memory_failure_cpu *mf_cpu; 2444 struct memory_failure_entry entry = { 0, }; 2445 unsigned long proc_flags; 2446 int gotten; 2447 2448 mf_cpu = container_of(work, struct memory_failure_cpu, work); 2449 for (;;) { 2450 spin_lock_irqsave(&mf_cpu->lock, proc_flags); 2451 gotten = kfifo_get(&mf_cpu->fifo, &entry); 2452 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); 2453 if (!gotten) 2454 break; 2455 if (entry.flags & MF_SOFT_OFFLINE) 2456 soft_offline_page(entry.pfn, entry.flags); 2457 else 2458 memory_failure(entry.pfn, entry.flags); 2459 } 2460 } 2461 2462 /* 2463 * Process memory_failure work queued on the specified CPU. 2464 * Used to avoid return-to-userspace racing with the memory_failure workqueue. 2465 */ 2466 void memory_failure_queue_kick(int cpu) 2467 { 2468 struct memory_failure_cpu *mf_cpu; 2469 2470 mf_cpu = &per_cpu(memory_failure_cpu, cpu); 2471 cancel_work_sync(&mf_cpu->work); 2472 memory_failure_work_func(&mf_cpu->work); 2473 } 2474 2475 static int __init memory_failure_init(void) 2476 { 2477 struct memory_failure_cpu *mf_cpu; 2478 int cpu; 2479 2480 for_each_possible_cpu(cpu) { 2481 mf_cpu = &per_cpu(memory_failure_cpu, cpu); 2482 spin_lock_init(&mf_cpu->lock); 2483 INIT_KFIFO(mf_cpu->fifo); 2484 INIT_WORK(&mf_cpu->work, memory_failure_work_func); 2485 } 2486 2487 register_sysctl_init("vm", memory_failure_table); 2488 2489 return 0; 2490 } 2491 core_initcall(memory_failure_init); 2492 2493 #undef pr_fmt 2494 #define pr_fmt(fmt) "" fmt 2495 #define unpoison_pr_info(fmt, pfn, rs) \ 2496 ({ \ 2497 if (__ratelimit(rs)) \ 2498 pr_info(fmt, pfn); \ 2499 }) 2500 2501 /** 2502 * unpoison_memory - Unpoison a previously poisoned page 2503 * @pfn: Page number of the to be unpoisoned page 2504 * 2505 * Software-unpoison a page that has been poisoned by 2506 * memory_failure() earlier. 2507 * 2508 * This is only done on the software-level, so it only works 2509 * for linux injected failures, not real hardware failures 2510 * 2511 * Returns 0 for success, otherwise -errno. 2512 */ 2513 int unpoison_memory(unsigned long pfn) 2514 { 2515 struct folio *folio; 2516 struct page *p; 2517 int ret = -EBUSY, ghp; 2518 unsigned long count = 1; 2519 bool huge = false; 2520 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, 2521 DEFAULT_RATELIMIT_BURST); 2522 2523 if (!pfn_valid(pfn)) 2524 return -ENXIO; 2525 2526 p = pfn_to_page(pfn); 2527 folio = page_folio(p); 2528 2529 mutex_lock(&mf_mutex); 2530 2531 if (hw_memory_failure) { 2532 unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n", 2533 pfn, &unpoison_rs); 2534 ret = -EOPNOTSUPP; 2535 goto unlock_mutex; 2536 } 2537 2538 if (!PageHWPoison(p)) { 2539 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", 2540 pfn, &unpoison_rs); 2541 goto unlock_mutex; 2542 } 2543 2544 if (folio_ref_count(folio) > 1) { 2545 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", 2546 pfn, &unpoison_rs); 2547 goto unlock_mutex; 2548 } 2549 2550 if (folio_test_slab(folio) || PageTable(&folio->page) || 2551 folio_test_reserved(folio) || PageOffline(&folio->page)) 2552 goto unlock_mutex; 2553 2554 /* 2555 * Note that folio->_mapcount is overloaded in SLAB, so the simple test 2556 * in folio_mapped() has to be done after folio_test_slab() is checked. 2557 */ 2558 if (folio_mapped(folio)) { 2559 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", 2560 pfn, &unpoison_rs); 2561 goto unlock_mutex; 2562 } 2563 2564 if (folio_mapping(folio)) { 2565 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", 2566 pfn, &unpoison_rs); 2567 goto unlock_mutex; 2568 } 2569 2570 ghp = get_hwpoison_page(p, MF_UNPOISON); 2571 if (!ghp) { 2572 if (PageHuge(p)) { 2573 huge = true; 2574 count = folio_free_raw_hwp(folio, false); 2575 if (count == 0) 2576 goto unlock_mutex; 2577 } 2578 ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY; 2579 } else if (ghp < 0) { 2580 if (ghp == -EHWPOISON) { 2581 ret = put_page_back_buddy(p) ? 0 : -EBUSY; 2582 } else { 2583 ret = ghp; 2584 unpoison_pr_info("Unpoison: failed to grab page %#lx\n", 2585 pfn, &unpoison_rs); 2586 } 2587 } else { 2588 if (PageHuge(p)) { 2589 huge = true; 2590 count = folio_free_raw_hwp(folio, false); 2591 if (count == 0) { 2592 folio_put(folio); 2593 goto unlock_mutex; 2594 } 2595 } 2596 2597 folio_put(folio); 2598 if (TestClearPageHWPoison(p)) { 2599 folio_put(folio); 2600 ret = 0; 2601 } 2602 } 2603 2604 unlock_mutex: 2605 mutex_unlock(&mf_mutex); 2606 if (!ret) { 2607 if (!huge) 2608 num_poisoned_pages_sub(pfn, 1); 2609 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", 2610 page_to_pfn(p), &unpoison_rs); 2611 } 2612 return ret; 2613 } 2614 EXPORT_SYMBOL(unpoison_memory); 2615 2616 static bool mf_isolate_folio(struct folio *folio, struct list_head *pagelist) 2617 { 2618 bool isolated = false; 2619 2620 if (folio_test_hugetlb(folio)) { 2621 isolated = isolate_hugetlb(folio, pagelist); 2622 } else { 2623 bool lru = !__folio_test_movable(folio); 2624 2625 if (lru) 2626 isolated = folio_isolate_lru(folio); 2627 else 2628 isolated = isolate_movable_page(&folio->page, 2629 ISOLATE_UNEVICTABLE); 2630 2631 if (isolated) { 2632 list_add(&folio->lru, pagelist); 2633 if (lru) 2634 node_stat_add_folio(folio, NR_ISOLATED_ANON + 2635 folio_is_file_lru(folio)); 2636 } 2637 } 2638 2639 /* 2640 * If we succeed to isolate the folio, we grabbed another refcount on 2641 * the folio, so we can safely drop the one we got from get_any_page(). 2642 * If we failed to isolate the folio, it means that we cannot go further 2643 * and we will return an error, so drop the reference we got from 2644 * get_any_page() as well. 2645 */ 2646 folio_put(folio); 2647 return isolated; 2648 } 2649 2650 /* 2651 * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages. 2652 * If the page is a non-dirty unmapped page-cache page, it simply invalidates. 2653 * If the page is mapped, it migrates the contents over. 2654 */ 2655 static int soft_offline_in_use_page(struct page *page) 2656 { 2657 long ret = 0; 2658 unsigned long pfn = page_to_pfn(page); 2659 struct folio *folio = page_folio(page); 2660 char const *msg_page[] = {"page", "hugepage"}; 2661 bool huge = folio_test_hugetlb(folio); 2662 LIST_HEAD(pagelist); 2663 struct migration_target_control mtc = { 2664 .nid = NUMA_NO_NODE, 2665 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, 2666 }; 2667 2668 if (!huge && folio_test_large(folio)) { 2669 if (try_to_split_thp_page(page)) { 2670 pr_info("soft offline: %#lx: thp split failed\n", pfn); 2671 return -EBUSY; 2672 } 2673 folio = page_folio(page); 2674 } 2675 2676 folio_lock(folio); 2677 if (!huge) 2678 folio_wait_writeback(folio); 2679 if (PageHWPoison(page)) { 2680 folio_unlock(folio); 2681 folio_put(folio); 2682 pr_info("soft offline: %#lx page already poisoned\n", pfn); 2683 return 0; 2684 } 2685 2686 if (!huge && folio_test_lru(folio) && !folio_test_swapcache(folio)) 2687 /* 2688 * Try to invalidate first. This should work for 2689 * non dirty unmapped page cache pages. 2690 */ 2691 ret = mapping_evict_folio(folio_mapping(folio), folio); 2692 folio_unlock(folio); 2693 2694 if (ret) { 2695 pr_info("soft_offline: %#lx: invalidated\n", pfn); 2696 page_handle_poison(page, false, true); 2697 return 0; 2698 } 2699 2700 if (mf_isolate_folio(folio, &pagelist)) { 2701 ret = migrate_pages(&pagelist, alloc_migration_target, NULL, 2702 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL); 2703 if (!ret) { 2704 bool release = !huge; 2705 2706 if (!page_handle_poison(page, huge, release)) 2707 ret = -EBUSY; 2708 } else { 2709 if (!list_empty(&pagelist)) 2710 putback_movable_pages(&pagelist); 2711 2712 pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n", 2713 pfn, msg_page[huge], ret, &page->flags); 2714 if (ret > 0) 2715 ret = -EBUSY; 2716 } 2717 } else { 2718 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n", 2719 pfn, msg_page[huge], page_count(page), &page->flags); 2720 ret = -EBUSY; 2721 } 2722 return ret; 2723 } 2724 2725 /** 2726 * soft_offline_page - Soft offline a page. 2727 * @pfn: pfn to soft-offline 2728 * @flags: flags. Same as memory_failure(). 2729 * 2730 * Returns 0 on success 2731 * -EOPNOTSUPP for hwpoison_filter() filtered the error event 2732 * < 0 otherwise negated errno. 2733 * 2734 * Soft offline a page, by migration or invalidation, 2735 * without killing anything. This is for the case when 2736 * a page is not corrupted yet (so it's still valid to access), 2737 * but has had a number of corrected errors and is better taken 2738 * out. 2739 * 2740 * The actual policy on when to do that is maintained by 2741 * user space. 2742 * 2743 * This should never impact any application or cause data loss, 2744 * however it might take some time. 2745 * 2746 * This is not a 100% solution for all memory, but tries to be 2747 * ``good enough'' for the majority of memory. 2748 */ 2749 int soft_offline_page(unsigned long pfn, int flags) 2750 { 2751 int ret; 2752 bool try_again = true; 2753 struct page *page; 2754 2755 if (!pfn_valid(pfn)) { 2756 WARN_ON_ONCE(flags & MF_COUNT_INCREASED); 2757 return -ENXIO; 2758 } 2759 2760 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ 2761 page = pfn_to_online_page(pfn); 2762 if (!page) { 2763 put_ref_page(pfn, flags); 2764 return -EIO; 2765 } 2766 2767 mutex_lock(&mf_mutex); 2768 2769 if (PageHWPoison(page)) { 2770 pr_info("%s: %#lx page already poisoned\n", __func__, pfn); 2771 put_ref_page(pfn, flags); 2772 mutex_unlock(&mf_mutex); 2773 return 0; 2774 } 2775 2776 retry: 2777 get_online_mems(); 2778 ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE); 2779 put_online_mems(); 2780 2781 if (hwpoison_filter(page)) { 2782 if (ret > 0) 2783 put_page(page); 2784 2785 mutex_unlock(&mf_mutex); 2786 return -EOPNOTSUPP; 2787 } 2788 2789 if (ret > 0) { 2790 ret = soft_offline_in_use_page(page); 2791 } else if (ret == 0) { 2792 if (!page_handle_poison(page, true, false)) { 2793 if (try_again) { 2794 try_again = false; 2795 flags &= ~MF_COUNT_INCREASED; 2796 goto retry; 2797 } 2798 ret = -EBUSY; 2799 } 2800 } 2801 2802 mutex_unlock(&mf_mutex); 2803 2804 return ret; 2805 } 2806