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(v: &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, v: &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(zone: page_zone(page)); |
158 | ret = dissolve_free_huge_page(page); |
159 | if (!ret) |
160 | ret = take_page_off_buddy(page); |
161 | zone_pcp_enable(zone: 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(page: 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(page: 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(page: 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(page: 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: *pgd)) |
390 | return 0; |
391 | p4d = p4d_offset(pgd, address); |
392 | if (!p4d_present(p4d: *p4d)) |
393 | return 0; |
394 | pud = pud_offset(p4d, address); |
395 | if (!pud_present(pud: *pud)) |
396 | return 0; |
397 | if (pud_devmap(pud: *pud)) |
398 | return PUD_SHIFT; |
399 | pmd = pmd_offset(pud, address); |
400 | if (!pmd_present(pmd: *pmd)) |
401 | return 0; |
402 | if (pmd_devmap(pmd: *pmd)) |
403 | return PMD_SHIFT; |
404 | pte = pte_offset_map(pmd, addr: address); |
405 | if (!pte) |
406 | return 0; |
407 | ptent = ptep_get(ptep: pte); |
408 | if (pte_present(a: ptent) && pte_devmap(a: 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(size: 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(page: p)) { |
445 | if (fsdax_pgoff != FSDAX_INVALID_PGOFF) |
446 | tk->addr = vma_pgoff_address(pgoff: fsdax_pgoff, nr_pages: 1, vma); |
447 | tk->size_shift = dev_pagemap_mapping_shift(vma, address: 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(objp: tk); |
466 | return; |
467 | } |
468 | |
469 | get_task_struct(t: tsk); |
470 | tk->tsk = tsk; |
471 | list_add_tail(new: &tk->nd, head: 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, ksm_addr: 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 | p: tk->tsk, type: 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(entry: &tk->nd); |
540 | put_task_struct(t: tk->tsk); |
541 | kfree(objp: 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 page *page, struct list_head *to_kill, |
599 | int force_early) |
600 | { |
601 | struct folio *folio = page_folio(page); |
602 | struct vm_area_struct *vma; |
603 | struct task_struct *tsk; |
604 | struct anon_vma *av; |
605 | pgoff_t pgoff; |
606 | |
607 | av = folio_lock_anon_vma_read(folio, NULL); |
608 | if (av == NULL) /* Not actually mapped anymore */ |
609 | return; |
610 | |
611 | pgoff = page_to_pgoff(page); |
612 | rcu_read_lock(); |
613 | for_each_process(tsk) { |
614 | struct anon_vma_chain *vmac; |
615 | struct task_struct *t = task_early_kill(tsk, force_early); |
616 | |
617 | if (!t) |
618 | continue; |
619 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
620 | pgoff, pgoff) { |
621 | vma = vmac->vma; |
622 | if (vma->vm_mm != t->mm) |
623 | continue; |
624 | if (!page_mapped_in_vma(page, vma)) |
625 | continue; |
626 | add_to_kill_anon_file(tsk: t, p: page, vma, to_kill); |
627 | } |
628 | } |
629 | rcu_read_unlock(); |
630 | anon_vma_unlock_read(anon_vma: av); |
631 | } |
632 | |
633 | /* |
634 | * Collect processes when the error hit a file mapped page. |
635 | */ |
636 | static void collect_procs_file(struct page *page, struct list_head *to_kill, |
637 | int force_early) |
638 | { |
639 | struct vm_area_struct *vma; |
640 | struct task_struct *tsk; |
641 | struct address_space *mapping = page->mapping; |
642 | pgoff_t pgoff; |
643 | |
644 | i_mmap_lock_read(mapping); |
645 | rcu_read_lock(); |
646 | pgoff = page_to_pgoff(page); |
647 | for_each_process(tsk) { |
648 | struct task_struct *t = task_early_kill(tsk, force_early); |
649 | |
650 | if (!t) |
651 | continue; |
652 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
653 | pgoff) { |
654 | /* |
655 | * Send early kill signal to tasks where a vma covers |
656 | * the page but the corrupted page is not necessarily |
657 | * mapped in its pte. |
658 | * Assume applications who requested early kill want |
659 | * to be informed of all such data corruptions. |
660 | */ |
661 | if (vma->vm_mm == t->mm) |
662 | add_to_kill_anon_file(tsk: t, p: page, vma, to_kill); |
663 | } |
664 | } |
665 | rcu_read_unlock(); |
666 | i_mmap_unlock_read(mapping); |
667 | } |
668 | |
669 | #ifdef CONFIG_FS_DAX |
670 | static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p, |
671 | struct vm_area_struct *vma, |
672 | struct list_head *to_kill, pgoff_t pgoff) |
673 | { |
674 | __add_to_kill(tsk, p, vma, to_kill, ksm_addr: 0, fsdax_pgoff: pgoff); |
675 | } |
676 | |
677 | /* |
678 | * Collect processes when the error hit a fsdax page. |
679 | */ |
680 | static void collect_procs_fsdax(struct page *page, |
681 | struct address_space *mapping, pgoff_t pgoff, |
682 | struct list_head *to_kill) |
683 | { |
684 | struct vm_area_struct *vma; |
685 | struct task_struct *tsk; |
686 | |
687 | i_mmap_lock_read(mapping); |
688 | rcu_read_lock(); |
689 | for_each_process(tsk) { |
690 | struct task_struct *t = task_early_kill(tsk, force_early: true); |
691 | |
692 | if (!t) |
693 | continue; |
694 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { |
695 | if (vma->vm_mm == t->mm) |
696 | add_to_kill_fsdax(tsk: t, p: page, vma, to_kill, pgoff); |
697 | } |
698 | } |
699 | rcu_read_unlock(); |
700 | i_mmap_unlock_read(mapping); |
701 | } |
702 | #endif /* CONFIG_FS_DAX */ |
703 | |
704 | /* |
705 | * Collect the processes who have the corrupted page mapped to kill. |
706 | */ |
707 | static void collect_procs(struct page *page, struct list_head *tokill, |
708 | int force_early) |
709 | { |
710 | if (!page->mapping) |
711 | return; |
712 | if (unlikely(PageKsm(page))) |
713 | collect_procs_ksm(page, to_kill: tokill, force_early); |
714 | else if (PageAnon(page)) |
715 | collect_procs_anon(page, to_kill: tokill, force_early); |
716 | else |
717 | collect_procs_file(page, to_kill: tokill, force_early); |
718 | } |
719 | |
720 | struct hwpoison_walk { |
721 | struct to_kill tk; |
722 | unsigned long pfn; |
723 | int flags; |
724 | }; |
725 | |
726 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) |
727 | { |
728 | tk->addr = addr; |
729 | tk->size_shift = shift; |
730 | } |
731 | |
732 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, |
733 | unsigned long poisoned_pfn, struct to_kill *tk) |
734 | { |
735 | unsigned long pfn = 0; |
736 | |
737 | if (pte_present(a: pte)) { |
738 | pfn = pte_pfn(pte); |
739 | } else { |
740 | swp_entry_t swp = pte_to_swp_entry(pte); |
741 | |
742 | if (is_hwpoison_entry(entry: swp)) |
743 | pfn = swp_offset_pfn(entry: swp); |
744 | } |
745 | |
746 | if (!pfn || pfn != poisoned_pfn) |
747 | return 0; |
748 | |
749 | set_to_kill(tk, addr, shift); |
750 | return 1; |
751 | } |
752 | |
753 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
754 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, |
755 | struct hwpoison_walk *hwp) |
756 | { |
757 | pmd_t pmd = *pmdp; |
758 | unsigned long pfn; |
759 | unsigned long hwpoison_vaddr; |
760 | |
761 | if (!pmd_present(pmd)) |
762 | return 0; |
763 | pfn = pmd_pfn(pmd); |
764 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { |
765 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); |
766 | set_to_kill(tk: &hwp->tk, addr: hwpoison_vaddr, PAGE_SHIFT); |
767 | return 1; |
768 | } |
769 | return 0; |
770 | } |
771 | #else |
772 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, |
773 | struct hwpoison_walk *hwp) |
774 | { |
775 | return 0; |
776 | } |
777 | #endif |
778 | |
779 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, |
780 | unsigned long end, struct mm_walk *walk) |
781 | { |
782 | struct hwpoison_walk *hwp = walk->private; |
783 | int ret = 0; |
784 | pte_t *ptep, *mapped_pte; |
785 | spinlock_t *ptl; |
786 | |
787 | ptl = pmd_trans_huge_lock(pmd: pmdp, vma: walk->vma); |
788 | if (ptl) { |
789 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); |
790 | spin_unlock(lock: ptl); |
791 | goto out; |
792 | } |
793 | |
794 | mapped_pte = ptep = pte_offset_map_lock(mm: walk->vma->vm_mm, pmd: pmdp, |
795 | addr, ptlp: &ptl); |
796 | if (!ptep) |
797 | goto out; |
798 | |
799 | for (; addr != end; ptep++, addr += PAGE_SIZE) { |
800 | ret = check_hwpoisoned_entry(pte: ptep_get(ptep), addr, PAGE_SHIFT, |
801 | poisoned_pfn: hwp->pfn, tk: &hwp->tk); |
802 | if (ret == 1) |
803 | break; |
804 | } |
805 | pte_unmap_unlock(mapped_pte, ptl); |
806 | out: |
807 | cond_resched(); |
808 | return ret; |
809 | } |
810 | |
811 | #ifdef CONFIG_HUGETLB_PAGE |
812 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, |
813 | unsigned long addr, unsigned long end, |
814 | struct mm_walk *walk) |
815 | { |
816 | struct hwpoison_walk *hwp = walk->private; |
817 | pte_t pte = huge_ptep_get(ptep); |
818 | struct hstate *h = hstate_vma(vma: walk->vma); |
819 | |
820 | return check_hwpoisoned_entry(pte, addr, shift: huge_page_shift(h), |
821 | poisoned_pfn: hwp->pfn, tk: &hwp->tk); |
822 | } |
823 | #else |
824 | #define hwpoison_hugetlb_range NULL |
825 | #endif |
826 | |
827 | static const struct mm_walk_ops hwpoison_walk_ops = { |
828 | .pmd_entry = hwpoison_pte_range, |
829 | .hugetlb_entry = hwpoison_hugetlb_range, |
830 | .walk_lock = PGWALK_RDLOCK, |
831 | }; |
832 | |
833 | /* |
834 | * Sends SIGBUS to the current process with error info. |
835 | * |
836 | * This function is intended to handle "Action Required" MCEs on already |
837 | * hardware poisoned pages. They could happen, for example, when |
838 | * memory_failure() failed to unmap the error page at the first call, or |
839 | * when multiple local machine checks happened on different CPUs. |
840 | * |
841 | * MCE handler currently has no easy access to the error virtual address, |
842 | * so this function walks page table to find it. The returned virtual address |
843 | * is proper in most cases, but it could be wrong when the application |
844 | * process has multiple entries mapping the error page. |
845 | */ |
846 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, |
847 | int flags) |
848 | { |
849 | int ret; |
850 | struct hwpoison_walk priv = { |
851 | .pfn = pfn, |
852 | }; |
853 | priv.tk.tsk = p; |
854 | |
855 | if (!p->mm) |
856 | return -EFAULT; |
857 | |
858 | mmap_read_lock(mm: p->mm); |
859 | ret = walk_page_range(mm: p->mm, start: 0, TASK_SIZE, ops: &hwpoison_walk_ops, |
860 | private: (void *)&priv); |
861 | if (ret == 1 && priv.tk.addr) |
862 | kill_proc(tk: &priv.tk, pfn, flags); |
863 | else |
864 | ret = 0; |
865 | mmap_read_unlock(mm: p->mm); |
866 | return ret > 0 ? -EHWPOISON : -EFAULT; |
867 | } |
868 | |
869 | static const char *action_name[] = { |
870 | [MF_IGNORED] = "Ignored" , |
871 | [MF_FAILED] = "Failed" , |
872 | [MF_DELAYED] = "Delayed" , |
873 | [MF_RECOVERED] = "Recovered" , |
874 | }; |
875 | |
876 | static const char * const action_page_types[] = { |
877 | [MF_MSG_KERNEL] = "reserved kernel page" , |
878 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page" , |
879 | [MF_MSG_SLAB] = "kernel slab page" , |
880 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking" , |
881 | [MF_MSG_HUGE] = "huge page" , |
882 | [MF_MSG_FREE_HUGE] = "free huge page" , |
883 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page" , |
884 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page" , |
885 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page" , |
886 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page" , |
887 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page" , |
888 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page" , |
889 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page" , |
890 | [MF_MSG_DIRTY_LRU] = "dirty LRU page" , |
891 | [MF_MSG_CLEAN_LRU] = "clean LRU page" , |
892 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page" , |
893 | [MF_MSG_BUDDY] = "free buddy page" , |
894 | [MF_MSG_DAX] = "dax page" , |
895 | [MF_MSG_UNSPLIT_THP] = "unsplit thp" , |
896 | [MF_MSG_UNKNOWN] = "unknown page" , |
897 | }; |
898 | |
899 | /* |
900 | * XXX: It is possible that a page is isolated from LRU cache, |
901 | * and then kept in swap cache or failed to remove from page cache. |
902 | * The page count will stop it from being freed by unpoison. |
903 | * Stress tests should be aware of this memory leak problem. |
904 | */ |
905 | static int delete_from_lru_cache(struct page *p) |
906 | { |
907 | if (isolate_lru_page(page: p)) { |
908 | /* |
909 | * Clear sensible page flags, so that the buddy system won't |
910 | * complain when the page is unpoison-and-freed. |
911 | */ |
912 | ClearPageActive(page: p); |
913 | ClearPageUnevictable(page: p); |
914 | |
915 | /* |
916 | * Poisoned page might never drop its ref count to 0 so we have |
917 | * to uncharge it manually from its memcg. |
918 | */ |
919 | mem_cgroup_uncharge(page_folio(p)); |
920 | |
921 | /* |
922 | * drop the page count elevated by isolate_lru_page() |
923 | */ |
924 | put_page(page: p); |
925 | return 0; |
926 | } |
927 | return -EIO; |
928 | } |
929 | |
930 | static int truncate_error_page(struct page *p, unsigned long pfn, |
931 | struct address_space *mapping) |
932 | { |
933 | int ret = MF_FAILED; |
934 | |
935 | if (mapping->a_ops->error_remove_page) { |
936 | struct folio *folio = page_folio(p); |
937 | int err = mapping->a_ops->error_remove_page(mapping, p); |
938 | |
939 | if (err != 0) |
940 | pr_info("%#lx: Failed to punch page: %d\n" , pfn, err); |
941 | else if (!filemap_release_folio(folio, GFP_NOIO)) |
942 | pr_info("%#lx: failed to release buffers\n" , pfn); |
943 | else |
944 | ret = MF_RECOVERED; |
945 | } else { |
946 | /* |
947 | * If the file system doesn't support it just invalidate |
948 | * This fails on dirty or anything with private pages |
949 | */ |
950 | if (invalidate_inode_page(page: p)) |
951 | ret = MF_RECOVERED; |
952 | else |
953 | pr_info("%#lx: Failed to invalidate\n" , pfn); |
954 | } |
955 | |
956 | return ret; |
957 | } |
958 | |
959 | struct page_state { |
960 | unsigned long mask; |
961 | unsigned long res; |
962 | enum mf_action_page_type type; |
963 | |
964 | /* Callback ->action() has to unlock the relevant page inside it. */ |
965 | int (*action)(struct page_state *ps, struct page *p); |
966 | }; |
967 | |
968 | /* |
969 | * Return true if page is still referenced by others, otherwise return |
970 | * false. |
971 | * |
972 | * The extra_pins is true when one extra refcount is expected. |
973 | */ |
974 | static bool (struct page_state *ps, struct page *p, |
975 | bool ) |
976 | { |
977 | int count = page_count(page: p) - 1; |
978 | |
979 | if (extra_pins) |
980 | count -= 1; |
981 | |
982 | if (count > 0) { |
983 | pr_err("%#lx: %s still referenced by %d users\n" , |
984 | page_to_pfn(p), action_page_types[ps->type], count); |
985 | return true; |
986 | } |
987 | |
988 | return false; |
989 | } |
990 | |
991 | /* |
992 | * Error hit kernel page. |
993 | * Do nothing, try to be lucky and not touch this instead. For a few cases we |
994 | * could be more sophisticated. |
995 | */ |
996 | static int me_kernel(struct page_state *ps, struct page *p) |
997 | { |
998 | unlock_page(page: p); |
999 | return MF_IGNORED; |
1000 | } |
1001 | |
1002 | /* |
1003 | * Page in unknown state. Do nothing. |
1004 | */ |
1005 | static int me_unknown(struct page_state *ps, struct page *p) |
1006 | { |
1007 | pr_err("%#lx: Unknown page state\n" , page_to_pfn(p)); |
1008 | unlock_page(page: p); |
1009 | return MF_FAILED; |
1010 | } |
1011 | |
1012 | /* |
1013 | * Clean (or cleaned) page cache page. |
1014 | */ |
1015 | static int me_pagecache_clean(struct page_state *ps, struct page *p) |
1016 | { |
1017 | int ret; |
1018 | struct address_space *mapping; |
1019 | bool ; |
1020 | |
1021 | delete_from_lru_cache(p); |
1022 | |
1023 | /* |
1024 | * For anonymous pages we're done the only reference left |
1025 | * should be the one m_f() holds. |
1026 | */ |
1027 | if (PageAnon(page: p)) { |
1028 | ret = MF_RECOVERED; |
1029 | goto out; |
1030 | } |
1031 | |
1032 | /* |
1033 | * Now truncate the page in the page cache. This is really |
1034 | * more like a "temporary hole punch" |
1035 | * Don't do this for block devices when someone else |
1036 | * has a reference, because it could be file system metadata |
1037 | * and that's not safe to truncate. |
1038 | */ |
1039 | mapping = page_mapping(p); |
1040 | if (!mapping) { |
1041 | /* |
1042 | * Page has been teared down in the meanwhile |
1043 | */ |
1044 | ret = MF_FAILED; |
1045 | goto out; |
1046 | } |
1047 | |
1048 | /* |
1049 | * The shmem page is kept in page cache instead of truncating |
1050 | * so is expected to have an extra refcount after error-handling. |
1051 | */ |
1052 | extra_pins = shmem_mapping(mapping); |
1053 | |
1054 | /* |
1055 | * Truncation is a bit tricky. Enable it per file system for now. |
1056 | * |
1057 | * Open: to take i_rwsem or not for this? Right now we don't. |
1058 | */ |
1059 | ret = truncate_error_page(p, page_to_pfn(p), mapping); |
1060 | if (has_extra_refcount(ps, p, extra_pins)) |
1061 | ret = MF_FAILED; |
1062 | |
1063 | out: |
1064 | unlock_page(page: p); |
1065 | |
1066 | return ret; |
1067 | } |
1068 | |
1069 | /* |
1070 | * Dirty pagecache page |
1071 | * Issues: when the error hit a hole page the error is not properly |
1072 | * propagated. |
1073 | */ |
1074 | static int me_pagecache_dirty(struct page_state *ps, struct page *p) |
1075 | { |
1076 | struct address_space *mapping = page_mapping(p); |
1077 | |
1078 | SetPageError(p); |
1079 | /* TBD: print more information about the file. */ |
1080 | if (mapping) { |
1081 | /* |
1082 | * IO error will be reported by write(), fsync(), etc. |
1083 | * who check the mapping. |
1084 | * This way the application knows that something went |
1085 | * wrong with its dirty file data. |
1086 | * |
1087 | * There's one open issue: |
1088 | * |
1089 | * The EIO will be only reported on the next IO |
1090 | * operation and then cleared through the IO map. |
1091 | * Normally Linux has two mechanisms to pass IO error |
1092 | * first through the AS_EIO flag in the address space |
1093 | * and then through the PageError flag in the page. |
1094 | * Since we drop pages on memory failure handling the |
1095 | * only mechanism open to use is through AS_AIO. |
1096 | * |
1097 | * This has the disadvantage that it gets cleared on |
1098 | * the first operation that returns an error, while |
1099 | * the PageError bit is more sticky and only cleared |
1100 | * when the page is reread or dropped. If an |
1101 | * application assumes it will always get error on |
1102 | * fsync, but does other operations on the fd before |
1103 | * and the page is dropped between then the error |
1104 | * will not be properly reported. |
1105 | * |
1106 | * This can already happen even without hwpoisoned |
1107 | * pages: first on metadata IO errors (which only |
1108 | * report through AS_EIO) or when the page is dropped |
1109 | * at the wrong time. |
1110 | * |
1111 | * So right now we assume that the application DTRT on |
1112 | * the first EIO, but we're not worse than other parts |
1113 | * of the kernel. |
1114 | */ |
1115 | mapping_set_error(mapping, error: -EIO); |
1116 | } |
1117 | |
1118 | return me_pagecache_clean(ps, p); |
1119 | } |
1120 | |
1121 | /* |
1122 | * Clean and dirty swap cache. |
1123 | * |
1124 | * Dirty swap cache page is tricky to handle. The page could live both in page |
1125 | * cache and swap cache(ie. page is freshly swapped in). So it could be |
1126 | * referenced concurrently by 2 types of PTEs: |
1127 | * normal PTEs and swap PTEs. We try to handle them consistently by calling |
1128 | * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs, |
1129 | * and then |
1130 | * - clear dirty bit to prevent IO |
1131 | * - remove from LRU |
1132 | * - but keep in the swap cache, so that when we return to it on |
1133 | * a later page fault, we know the application is accessing |
1134 | * corrupted data and shall be killed (we installed simple |
1135 | * interception code in do_swap_page to catch it). |
1136 | * |
1137 | * Clean swap cache pages can be directly isolated. A later page fault will |
1138 | * bring in the known good data from disk. |
1139 | */ |
1140 | static int me_swapcache_dirty(struct page_state *ps, struct page *p) |
1141 | { |
1142 | int ret; |
1143 | bool = false; |
1144 | |
1145 | ClearPageDirty(page: p); |
1146 | /* Trigger EIO in shmem: */ |
1147 | ClearPageUptodate(page: p); |
1148 | |
1149 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
1150 | unlock_page(page: p); |
1151 | |
1152 | if (ret == MF_DELAYED) |
1153 | extra_pins = true; |
1154 | |
1155 | if (has_extra_refcount(ps, p, extra_pins)) |
1156 | ret = MF_FAILED; |
1157 | |
1158 | return ret; |
1159 | } |
1160 | |
1161 | static int me_swapcache_clean(struct page_state *ps, struct page *p) |
1162 | { |
1163 | struct folio *folio = page_folio(p); |
1164 | int ret; |
1165 | |
1166 | delete_from_swap_cache(folio); |
1167 | |
1168 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
1169 | folio_unlock(folio); |
1170 | |
1171 | if (has_extra_refcount(ps, p, extra_pins: false)) |
1172 | ret = MF_FAILED; |
1173 | |
1174 | return ret; |
1175 | } |
1176 | |
1177 | /* |
1178 | * Huge pages. Needs work. |
1179 | * Issues: |
1180 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
1181 | * To narrow down kill region to one page, we need to break up pmd. |
1182 | */ |
1183 | static int me_huge_page(struct page_state *ps, struct page *p) |
1184 | { |
1185 | int res; |
1186 | struct page *hpage = compound_head(p); |
1187 | struct address_space *mapping; |
1188 | bool = false; |
1189 | |
1190 | mapping = page_mapping(hpage); |
1191 | if (mapping) { |
1192 | res = truncate_error_page(p: hpage, page_to_pfn(p), mapping); |
1193 | /* The page is kept in page cache. */ |
1194 | extra_pins = true; |
1195 | unlock_page(page: hpage); |
1196 | } else { |
1197 | unlock_page(page: hpage); |
1198 | /* |
1199 | * migration entry prevents later access on error hugepage, |
1200 | * so we can free and dissolve it into buddy to save healthy |
1201 | * subpages. |
1202 | */ |
1203 | put_page(page: hpage); |
1204 | if (__page_handle_poison(page: p) >= 0) { |
1205 | page_ref_inc(page: p); |
1206 | res = MF_RECOVERED; |
1207 | } else { |
1208 | res = MF_FAILED; |
1209 | } |
1210 | } |
1211 | |
1212 | if (has_extra_refcount(ps, p, extra_pins)) |
1213 | res = MF_FAILED; |
1214 | |
1215 | return res; |
1216 | } |
1217 | |
1218 | /* |
1219 | * Various page states we can handle. |
1220 | * |
1221 | * A page state is defined by its current page->flags bits. |
1222 | * The table matches them in order and calls the right handler. |
1223 | * |
1224 | * This is quite tricky because we can access page at any time |
1225 | * in its live cycle, so all accesses have to be extremely careful. |
1226 | * |
1227 | * This is not complete. More states could be added. |
1228 | * For any missing state don't attempt recovery. |
1229 | */ |
1230 | |
1231 | #define dirty (1UL << PG_dirty) |
1232 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
1233 | #define unevict (1UL << PG_unevictable) |
1234 | #define mlock (1UL << PG_mlocked) |
1235 | #define lru (1UL << PG_lru) |
1236 | #define head (1UL << PG_head) |
1237 | #define slab (1UL << PG_slab) |
1238 | #define reserved (1UL << PG_reserved) |
1239 | |
1240 | static struct page_state error_states[] = { |
1241 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
1242 | /* |
1243 | * free pages are specially detected outside this table: |
1244 | * PG_buddy pages only make a small fraction of all free pages. |
1245 | */ |
1246 | |
1247 | /* |
1248 | * Could in theory check if slab page is free or if we can drop |
1249 | * currently unused objects without touching them. But just |
1250 | * treat it as standard kernel for now. |
1251 | */ |
1252 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
1253 | |
1254 | { head, head, MF_MSG_HUGE, me_huge_page }, |
1255 | |
1256 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1257 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, |
1258 | |
1259 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1260 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, |
1261 | |
1262 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1263 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, |
1264 | |
1265 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1266 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, |
1267 | |
1268 | /* |
1269 | * Catchall entry: must be at end. |
1270 | */ |
1271 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
1272 | }; |
1273 | |
1274 | #undef dirty |
1275 | #undef sc |
1276 | #undef unevict |
1277 | #undef mlock |
1278 | #undef lru |
1279 | #undef head |
1280 | #undef slab |
1281 | #undef reserved |
1282 | |
1283 | static void update_per_node_mf_stats(unsigned long pfn, |
1284 | enum mf_result result) |
1285 | { |
1286 | int nid = MAX_NUMNODES; |
1287 | struct memory_failure_stats *mf_stats = NULL; |
1288 | |
1289 | nid = pfn_to_nid(pfn); |
1290 | if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) { |
1291 | WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d" , pfn, nid); |
1292 | return; |
1293 | } |
1294 | |
1295 | mf_stats = &NODE_DATA(nid)->mf_stats; |
1296 | switch (result) { |
1297 | case MF_IGNORED: |
1298 | ++mf_stats->ignored; |
1299 | break; |
1300 | case MF_FAILED: |
1301 | ++mf_stats->failed; |
1302 | break; |
1303 | case MF_DELAYED: |
1304 | ++mf_stats->delayed; |
1305 | break; |
1306 | case MF_RECOVERED: |
1307 | ++mf_stats->recovered; |
1308 | break; |
1309 | default: |
1310 | WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled" , result); |
1311 | break; |
1312 | } |
1313 | ++mf_stats->total; |
1314 | } |
1315 | |
1316 | /* |
1317 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of |
1318 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). |
1319 | */ |
1320 | static int action_result(unsigned long pfn, enum mf_action_page_type type, |
1321 | enum mf_result result) |
1322 | { |
1323 | trace_memory_failure_event(pfn, type, result); |
1324 | |
1325 | num_poisoned_pages_inc(pfn); |
1326 | |
1327 | update_per_node_mf_stats(pfn, result); |
1328 | |
1329 | pr_err("%#lx: recovery action for %s: %s\n" , |
1330 | pfn, action_page_types[type], action_name[result]); |
1331 | |
1332 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
1333 | } |
1334 | |
1335 | static int page_action(struct page_state *ps, struct page *p, |
1336 | unsigned long pfn) |
1337 | { |
1338 | int result; |
1339 | |
1340 | /* page p should be unlocked after returning from ps->action(). */ |
1341 | result = ps->action(ps, p); |
1342 | |
1343 | /* Could do more checks here if page looks ok */ |
1344 | /* |
1345 | * Could adjust zone counters here to correct for the missing page. |
1346 | */ |
1347 | |
1348 | return action_result(pfn, type: ps->type, result); |
1349 | } |
1350 | |
1351 | static inline bool PageHWPoisonTakenOff(struct page *page) |
1352 | { |
1353 | return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; |
1354 | } |
1355 | |
1356 | void SetPageHWPoisonTakenOff(struct page *page) |
1357 | { |
1358 | set_page_private(page, MAGIC_HWPOISON); |
1359 | } |
1360 | |
1361 | void ClearPageHWPoisonTakenOff(struct page *page) |
1362 | { |
1363 | if (PageHWPoison(page)) |
1364 | set_page_private(page, private: 0); |
1365 | } |
1366 | |
1367 | /* |
1368 | * Return true if a page type of a given page is supported by hwpoison |
1369 | * mechanism (while handling could fail), otherwise false. This function |
1370 | * does not return true for hugetlb or device memory pages, so it's assumed |
1371 | * to be called only in the context where we never have such pages. |
1372 | */ |
1373 | static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) |
1374 | { |
1375 | /* Soft offline could migrate non-LRU movable pages */ |
1376 | if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) |
1377 | return true; |
1378 | |
1379 | return PageLRU(page) || is_free_buddy_page(page); |
1380 | } |
1381 | |
1382 | static int __get_hwpoison_page(struct page *page, unsigned long flags) |
1383 | { |
1384 | struct folio *folio = page_folio(page); |
1385 | int ret = 0; |
1386 | bool hugetlb = false; |
1387 | |
1388 | ret = get_hwpoison_hugetlb_folio(folio, hugetlb: &hugetlb, unpoison: false); |
1389 | if (hugetlb) { |
1390 | /* Make sure hugetlb demotion did not happen from under us. */ |
1391 | if (folio == page_folio(page)) |
1392 | return ret; |
1393 | if (ret > 0) { |
1394 | folio_put(folio); |
1395 | folio = page_folio(page); |
1396 | } |
1397 | } |
1398 | |
1399 | /* |
1400 | * This check prevents from calling folio_try_get() for any |
1401 | * unsupported type of folio in order to reduce the risk of unexpected |
1402 | * races caused by taking a folio refcount. |
1403 | */ |
1404 | if (!HWPoisonHandlable(page: &folio->page, flags)) |
1405 | return -EBUSY; |
1406 | |
1407 | if (folio_try_get(folio)) { |
1408 | if (folio == page_folio(page)) |
1409 | return 1; |
1410 | |
1411 | pr_info("%#lx cannot catch tail\n" , page_to_pfn(page)); |
1412 | folio_put(folio); |
1413 | } |
1414 | |
1415 | return 0; |
1416 | } |
1417 | |
1418 | static int get_any_page(struct page *p, unsigned long flags) |
1419 | { |
1420 | int ret = 0, pass = 0; |
1421 | bool count_increased = false; |
1422 | |
1423 | if (flags & MF_COUNT_INCREASED) |
1424 | count_increased = true; |
1425 | |
1426 | try_again: |
1427 | if (!count_increased) { |
1428 | ret = __get_hwpoison_page(page: p, flags); |
1429 | if (!ret) { |
1430 | if (page_count(page: p)) { |
1431 | /* We raced with an allocation, retry. */ |
1432 | if (pass++ < 3) |
1433 | goto try_again; |
1434 | ret = -EBUSY; |
1435 | } else if (!PageHuge(page: p) && !is_free_buddy_page(page: p)) { |
1436 | /* We raced with put_page, retry. */ |
1437 | if (pass++ < 3) |
1438 | goto try_again; |
1439 | ret = -EIO; |
1440 | } |
1441 | goto out; |
1442 | } else if (ret == -EBUSY) { |
1443 | /* |
1444 | * We raced with (possibly temporary) unhandlable |
1445 | * page, retry. |
1446 | */ |
1447 | if (pass++ < 3) { |
1448 | shake_page(p); |
1449 | goto try_again; |
1450 | } |
1451 | ret = -EIO; |
1452 | goto out; |
1453 | } |
1454 | } |
1455 | |
1456 | if (PageHuge(page: p) || HWPoisonHandlable(page: p, flags)) { |
1457 | ret = 1; |
1458 | } else { |
1459 | /* |
1460 | * A page we cannot handle. Check whether we can turn |
1461 | * it into something we can handle. |
1462 | */ |
1463 | if (pass++ < 3) { |
1464 | put_page(page: p); |
1465 | shake_page(p); |
1466 | count_increased = false; |
1467 | goto try_again; |
1468 | } |
1469 | put_page(page: p); |
1470 | ret = -EIO; |
1471 | } |
1472 | out: |
1473 | if (ret == -EIO) |
1474 | pr_err("%#lx: unhandlable page.\n" , page_to_pfn(p)); |
1475 | |
1476 | return ret; |
1477 | } |
1478 | |
1479 | static int __get_unpoison_page(struct page *page) |
1480 | { |
1481 | struct folio *folio = page_folio(page); |
1482 | int ret = 0; |
1483 | bool hugetlb = false; |
1484 | |
1485 | ret = get_hwpoison_hugetlb_folio(folio, hugetlb: &hugetlb, unpoison: true); |
1486 | if (hugetlb) { |
1487 | /* Make sure hugetlb demotion did not happen from under us. */ |
1488 | if (folio == page_folio(page)) |
1489 | return ret; |
1490 | if (ret > 0) |
1491 | folio_put(folio); |
1492 | } |
1493 | |
1494 | /* |
1495 | * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, |
1496 | * but also isolated from buddy freelist, so need to identify the |
1497 | * state and have to cancel both operations to unpoison. |
1498 | */ |
1499 | if (PageHWPoisonTakenOff(page)) |
1500 | return -EHWPOISON; |
1501 | |
1502 | return get_page_unless_zero(page) ? 1 : 0; |
1503 | } |
1504 | |
1505 | /** |
1506 | * get_hwpoison_page() - Get refcount for memory error handling |
1507 | * @p: Raw error page (hit by memory error) |
1508 | * @flags: Flags controlling behavior of error handling |
1509 | * |
1510 | * get_hwpoison_page() takes a page refcount of an error page to handle memory |
1511 | * error on it, after checking that the error page is in a well-defined state |
1512 | * (defined as a page-type we can successfully handle the memory error on it, |
1513 | * such as LRU page and hugetlb page). |
1514 | * |
1515 | * Memory error handling could be triggered at any time on any type of page, |
1516 | * so it's prone to race with typical memory management lifecycle (like |
1517 | * allocation and free). So to avoid such races, get_hwpoison_page() takes |
1518 | * extra care for the error page's state (as done in __get_hwpoison_page()), |
1519 | * and has some retry logic in get_any_page(). |
1520 | * |
1521 | * When called from unpoison_memory(), the caller should already ensure that |
1522 | * the given page has PG_hwpoison. So it's never reused for other page |
1523 | * allocations, and __get_unpoison_page() never races with them. |
1524 | * |
1525 | * Return: 0 on failure, |
1526 | * 1 on success for in-use pages in a well-defined state, |
1527 | * -EIO for pages on which we can not handle memory errors, |
1528 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle |
1529 | * operations like allocation and free, |
1530 | * -EHWPOISON when the page is hwpoisoned and taken off from buddy. |
1531 | */ |
1532 | static int get_hwpoison_page(struct page *p, unsigned long flags) |
1533 | { |
1534 | int ret; |
1535 | |
1536 | zone_pcp_disable(zone: page_zone(page: p)); |
1537 | if (flags & MF_UNPOISON) |
1538 | ret = __get_unpoison_page(page: p); |
1539 | else |
1540 | ret = get_any_page(p, flags); |
1541 | zone_pcp_enable(zone: page_zone(page: p)); |
1542 | |
1543 | return ret; |
1544 | } |
1545 | |
1546 | /* |
1547 | * Do all that is necessary to remove user space mappings. Unmap |
1548 | * the pages and send SIGBUS to the processes if the data was dirty. |
1549 | */ |
1550 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
1551 | int flags, struct page *hpage) |
1552 | { |
1553 | struct folio *folio = page_folio(hpage); |
1554 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON; |
1555 | struct address_space *mapping; |
1556 | LIST_HEAD(tokill); |
1557 | bool unmap_success; |
1558 | int forcekill; |
1559 | bool mlocked = PageMlocked(page: hpage); |
1560 | |
1561 | /* |
1562 | * Here we are interested only in user-mapped pages, so skip any |
1563 | * other types of pages. |
1564 | */ |
1565 | if (PageReserved(page: p) || PageSlab(page: p) || PageTable(page: p) || PageOffline(page: p)) |
1566 | return true; |
1567 | if (!(PageLRU(page: hpage) || PageHuge(page: p))) |
1568 | return true; |
1569 | |
1570 | /* |
1571 | * This check implies we don't kill processes if their pages |
1572 | * are in the swap cache early. Those are always late kills. |
1573 | */ |
1574 | if (!page_mapped(page: hpage)) |
1575 | return true; |
1576 | |
1577 | if (PageSwapCache(page: p)) { |
1578 | pr_err("%#lx: keeping poisoned page in swap cache\n" , pfn); |
1579 | ttu &= ~TTU_HWPOISON; |
1580 | } |
1581 | |
1582 | /* |
1583 | * Propagate the dirty bit from PTEs to struct page first, because we |
1584 | * need this to decide if we should kill or just drop the page. |
1585 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1586 | * be called inside page lock (it's recommended but not enforced). |
1587 | */ |
1588 | mapping = page_mapping(hpage); |
1589 | if (!(flags & MF_MUST_KILL) && !PageDirty(page: hpage) && mapping && |
1590 | mapping_can_writeback(mapping)) { |
1591 | if (page_mkclean(page: hpage)) { |
1592 | SetPageDirty(hpage); |
1593 | } else { |
1594 | ttu &= ~TTU_HWPOISON; |
1595 | pr_info("%#lx: corrupted page was clean: dropped without side effects\n" , |
1596 | pfn); |
1597 | } |
1598 | } |
1599 | |
1600 | /* |
1601 | * First collect all the processes that have the page |
1602 | * mapped in dirty form. This has to be done before try_to_unmap, |
1603 | * because ttu takes the rmap data structures down. |
1604 | */ |
1605 | collect_procs(page: hpage, tokill: &tokill, force_early: flags & MF_ACTION_REQUIRED); |
1606 | |
1607 | if (PageHuge(page: hpage) && !PageAnon(page: hpage)) { |
1608 | /* |
1609 | * For hugetlb pages in shared mappings, try_to_unmap |
1610 | * could potentially call huge_pmd_unshare. Because of |
1611 | * this, take semaphore in write mode here and set |
1612 | * TTU_RMAP_LOCKED to indicate we have taken the lock |
1613 | * at this higher level. |
1614 | */ |
1615 | mapping = hugetlb_page_mapping_lock_write(hpage); |
1616 | if (mapping) { |
1617 | try_to_unmap(folio, flags: ttu|TTU_RMAP_LOCKED); |
1618 | i_mmap_unlock_write(mapping); |
1619 | } else |
1620 | pr_info("%#lx: could not lock mapping for mapped huge page\n" , pfn); |
1621 | } else { |
1622 | try_to_unmap(folio, flags: ttu); |
1623 | } |
1624 | |
1625 | unmap_success = !page_mapped(page: hpage); |
1626 | if (!unmap_success) |
1627 | pr_err("%#lx: failed to unmap page (mapcount=%d)\n" , |
1628 | pfn, page_mapcount(hpage)); |
1629 | |
1630 | /* |
1631 | * try_to_unmap() might put mlocked page in lru cache, so call |
1632 | * shake_page() again to ensure that it's flushed. |
1633 | */ |
1634 | if (mlocked) |
1635 | shake_page(hpage); |
1636 | |
1637 | /* |
1638 | * Now that the dirty bit has been propagated to the |
1639 | * struct page and all unmaps done we can decide if |
1640 | * killing is needed or not. Only kill when the page |
1641 | * was dirty or the process is not restartable, |
1642 | * otherwise the tokill list is merely |
1643 | * freed. When there was a problem unmapping earlier |
1644 | * use a more force-full uncatchable kill to prevent |
1645 | * any accesses to the poisoned memory. |
1646 | */ |
1647 | forcekill = PageDirty(page: hpage) || (flags & MF_MUST_KILL) || |
1648 | !unmap_success; |
1649 | kill_procs(to_kill: &tokill, forcekill, fail: !unmap_success, pfn, flags); |
1650 | |
1651 | return unmap_success; |
1652 | } |
1653 | |
1654 | static int identify_page_state(unsigned long pfn, struct page *p, |
1655 | unsigned long page_flags) |
1656 | { |
1657 | struct page_state *ps; |
1658 | |
1659 | /* |
1660 | * The first check uses the current page flags which may not have any |
1661 | * relevant information. The second check with the saved page flags is |
1662 | * carried out only if the first check can't determine the page status. |
1663 | */ |
1664 | for (ps = error_states;; ps++) |
1665 | if ((p->flags & ps->mask) == ps->res) |
1666 | break; |
1667 | |
1668 | page_flags |= (p->flags & (1UL << PG_dirty)); |
1669 | |
1670 | if (!ps->mask) |
1671 | for (ps = error_states;; ps++) |
1672 | if ((page_flags & ps->mask) == ps->res) |
1673 | break; |
1674 | return page_action(ps, p, pfn); |
1675 | } |
1676 | |
1677 | static int try_to_split_thp_page(struct page *page) |
1678 | { |
1679 | int ret; |
1680 | |
1681 | lock_page(page); |
1682 | ret = split_huge_page(page); |
1683 | unlock_page(page); |
1684 | |
1685 | if (unlikely(ret)) |
1686 | put_page(page); |
1687 | |
1688 | return ret; |
1689 | } |
1690 | |
1691 | static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn, |
1692 | struct address_space *mapping, pgoff_t index, int flags) |
1693 | { |
1694 | struct to_kill *tk; |
1695 | unsigned long size = 0; |
1696 | |
1697 | list_for_each_entry(tk, to_kill, nd) |
1698 | if (tk->size_shift) |
1699 | size = max(size, 1UL << tk->size_shift); |
1700 | |
1701 | if (size) { |
1702 | /* |
1703 | * Unmap the largest mapping to avoid breaking up device-dax |
1704 | * mappings which are constant size. The actual size of the |
1705 | * mapping being torn down is communicated in siginfo, see |
1706 | * kill_proc() |
1707 | */ |
1708 | loff_t start = (index << PAGE_SHIFT) & ~(size - 1); |
1709 | |
1710 | unmap_mapping_range(mapping, holebegin: start, holelen: size, even_cows: 0); |
1711 | } |
1712 | |
1713 | kill_procs(to_kill, forcekill: flags & MF_MUST_KILL, fail: false, pfn, flags); |
1714 | } |
1715 | |
1716 | /* |
1717 | * Only dev_pagemap pages get here, such as fsdax when the filesystem |
1718 | * either do not claim or fails to claim a hwpoison event, or devdax. |
1719 | * The fsdax pages are initialized per base page, and the devdax pages |
1720 | * could be initialized either as base pages, or as compound pages with |
1721 | * vmemmap optimization enabled. Devdax is simplistic in its dealing with |
1722 | * hwpoison, such that, if a subpage of a compound page is poisoned, |
1723 | * simply mark the compound head page is by far sufficient. |
1724 | */ |
1725 | static int mf_generic_kill_procs(unsigned long long pfn, int flags, |
1726 | struct dev_pagemap *pgmap) |
1727 | { |
1728 | struct folio *folio = pfn_folio(pfn); |
1729 | LIST_HEAD(to_kill); |
1730 | dax_entry_t cookie; |
1731 | int rc = 0; |
1732 | |
1733 | /* |
1734 | * Prevent the inode from being freed while we are interrogating |
1735 | * the address_space, typically this would be handled by |
1736 | * lock_page(), but dax pages do not use the page lock. This |
1737 | * also prevents changes to the mapping of this pfn until |
1738 | * poison signaling is complete. |
1739 | */ |
1740 | cookie = dax_lock_folio(folio); |
1741 | if (!cookie) |
1742 | return -EBUSY; |
1743 | |
1744 | if (hwpoison_filter(&folio->page)) { |
1745 | rc = -EOPNOTSUPP; |
1746 | goto unlock; |
1747 | } |
1748 | |
1749 | switch (pgmap->type) { |
1750 | case MEMORY_DEVICE_PRIVATE: |
1751 | case MEMORY_DEVICE_COHERENT: |
1752 | /* |
1753 | * TODO: Handle device pages which may need coordination |
1754 | * with device-side memory. |
1755 | */ |
1756 | rc = -ENXIO; |
1757 | goto unlock; |
1758 | default: |
1759 | break; |
1760 | } |
1761 | |
1762 | /* |
1763 | * Use this flag as an indication that the dax page has been |
1764 | * remapped UC to prevent speculative consumption of poison. |
1765 | */ |
1766 | SetPageHWPoison(&folio->page); |
1767 | |
1768 | /* |
1769 | * Unlike System-RAM there is no possibility to swap in a |
1770 | * different physical page at a given virtual address, so all |
1771 | * userspace consumption of ZONE_DEVICE memory necessitates |
1772 | * SIGBUS (i.e. MF_MUST_KILL) |
1773 | */ |
1774 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; |
1775 | collect_procs(page: &folio->page, tokill: &to_kill, force_early: true); |
1776 | |
1777 | unmap_and_kill(to_kill: &to_kill, pfn, mapping: folio->mapping, index: folio->index, flags); |
1778 | unlock: |
1779 | dax_unlock_folio(folio, cookie); |
1780 | return rc; |
1781 | } |
1782 | |
1783 | #ifdef CONFIG_FS_DAX |
1784 | /** |
1785 | * mf_dax_kill_procs - Collect and kill processes who are using this file range |
1786 | * @mapping: address_space of the file in use |
1787 | * @index: start pgoff of the range within the file |
1788 | * @count: length of the range, in unit of PAGE_SIZE |
1789 | * @mf_flags: memory failure flags |
1790 | */ |
1791 | int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index, |
1792 | unsigned long count, int mf_flags) |
1793 | { |
1794 | LIST_HEAD(to_kill); |
1795 | dax_entry_t cookie; |
1796 | struct page *page; |
1797 | size_t end = index + count; |
1798 | |
1799 | mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; |
1800 | |
1801 | for (; index < end; index++) { |
1802 | page = NULL; |
1803 | cookie = dax_lock_mapping_entry(mapping, index, page: &page); |
1804 | if (!cookie) |
1805 | return -EBUSY; |
1806 | if (!page) |
1807 | goto unlock; |
1808 | |
1809 | SetPageHWPoison(page); |
1810 | |
1811 | collect_procs_fsdax(page, mapping, pgoff: index, to_kill: &to_kill); |
1812 | unmap_and_kill(to_kill: &to_kill, page_to_pfn(page), mapping, |
1813 | index, flags: mf_flags); |
1814 | unlock: |
1815 | dax_unlock_mapping_entry(mapping, index, cookie); |
1816 | } |
1817 | return 0; |
1818 | } |
1819 | EXPORT_SYMBOL_GPL(mf_dax_kill_procs); |
1820 | #endif /* CONFIG_FS_DAX */ |
1821 | |
1822 | #ifdef CONFIG_HUGETLB_PAGE |
1823 | |
1824 | /* |
1825 | * Struct raw_hwp_page represents information about "raw error page", |
1826 | * constructing singly linked list from ->_hugetlb_hwpoison field of folio. |
1827 | */ |
1828 | struct raw_hwp_page { |
1829 | struct llist_node node; |
1830 | struct page *page; |
1831 | }; |
1832 | |
1833 | static inline struct llist_head *raw_hwp_list_head(struct folio *folio) |
1834 | { |
1835 | return (struct llist_head *)&folio->_hugetlb_hwpoison; |
1836 | } |
1837 | |
1838 | bool is_raw_hwpoison_page_in_hugepage(struct page *page) |
1839 | { |
1840 | struct llist_head *raw_hwp_head; |
1841 | struct raw_hwp_page *p; |
1842 | struct folio *folio = page_folio(page); |
1843 | bool ret = false; |
1844 | |
1845 | if (!folio_test_hwpoison(folio)) |
1846 | return false; |
1847 | |
1848 | if (!folio_test_hugetlb(folio)) |
1849 | return PageHWPoison(page); |
1850 | |
1851 | /* |
1852 | * When RawHwpUnreliable is set, kernel lost track of which subpages |
1853 | * are HWPOISON. So return as if ALL subpages are HWPOISONed. |
1854 | */ |
1855 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
1856 | return true; |
1857 | |
1858 | mutex_lock(&mf_mutex); |
1859 | |
1860 | raw_hwp_head = raw_hwp_list_head(folio); |
1861 | llist_for_each_entry(p, raw_hwp_head->first, node) { |
1862 | if (page == p->page) { |
1863 | ret = true; |
1864 | break; |
1865 | } |
1866 | } |
1867 | |
1868 | mutex_unlock(lock: &mf_mutex); |
1869 | |
1870 | return ret; |
1871 | } |
1872 | |
1873 | static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag) |
1874 | { |
1875 | struct llist_node *head; |
1876 | struct raw_hwp_page *p, *next; |
1877 | unsigned long count = 0; |
1878 | |
1879 | head = llist_del_all(head: raw_hwp_list_head(folio)); |
1880 | llist_for_each_entry_safe(p, next, head, node) { |
1881 | if (move_flag) |
1882 | SetPageHWPoison(p->page); |
1883 | else |
1884 | num_poisoned_pages_sub(page_to_pfn(p->page), i: 1); |
1885 | kfree(objp: p); |
1886 | count++; |
1887 | } |
1888 | return count; |
1889 | } |
1890 | |
1891 | static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page) |
1892 | { |
1893 | struct llist_head *head; |
1894 | struct raw_hwp_page *raw_hwp; |
1895 | struct raw_hwp_page *p, *next; |
1896 | int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0; |
1897 | |
1898 | /* |
1899 | * Once the hwpoison hugepage has lost reliable raw error info, |
1900 | * there is little meaning to keep additional error info precisely, |
1901 | * so skip to add additional raw error info. |
1902 | */ |
1903 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
1904 | return -EHWPOISON; |
1905 | head = raw_hwp_list_head(folio); |
1906 | llist_for_each_entry_safe(p, next, head->first, node) { |
1907 | if (p->page == page) |
1908 | return -EHWPOISON; |
1909 | } |
1910 | |
1911 | raw_hwp = kmalloc(size: sizeof(struct raw_hwp_page), GFP_ATOMIC); |
1912 | if (raw_hwp) { |
1913 | raw_hwp->page = page; |
1914 | llist_add(new: &raw_hwp->node, head); |
1915 | /* the first error event will be counted in action_result(). */ |
1916 | if (ret) |
1917 | num_poisoned_pages_inc(page_to_pfn(page)); |
1918 | } else { |
1919 | /* |
1920 | * Failed to save raw error info. We no longer trace all |
1921 | * hwpoisoned subpages, and we need refuse to free/dissolve |
1922 | * this hwpoisoned hugepage. |
1923 | */ |
1924 | folio_set_hugetlb_raw_hwp_unreliable(folio); |
1925 | /* |
1926 | * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not |
1927 | * used any more, so free it. |
1928 | */ |
1929 | __folio_free_raw_hwp(folio, move_flag: false); |
1930 | } |
1931 | return ret; |
1932 | } |
1933 | |
1934 | static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag) |
1935 | { |
1936 | /* |
1937 | * hugetlb_vmemmap_optimized hugepages can't be freed because struct |
1938 | * pages for tail pages are required but they don't exist. |
1939 | */ |
1940 | if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio)) |
1941 | return 0; |
1942 | |
1943 | /* |
1944 | * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by |
1945 | * definition. |
1946 | */ |
1947 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
1948 | return 0; |
1949 | |
1950 | return __folio_free_raw_hwp(folio, move_flag); |
1951 | } |
1952 | |
1953 | void folio_clear_hugetlb_hwpoison(struct folio *folio) |
1954 | { |
1955 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
1956 | return; |
1957 | if (folio_test_hugetlb_vmemmap_optimized(folio)) |
1958 | return; |
1959 | folio_clear_hwpoison(folio); |
1960 | folio_free_raw_hwp(folio, move_flag: true); |
1961 | } |
1962 | |
1963 | /* |
1964 | * Called from hugetlb code with hugetlb_lock held. |
1965 | * |
1966 | * Return values: |
1967 | * 0 - free hugepage |
1968 | * 1 - in-use hugepage |
1969 | * 2 - not a hugepage |
1970 | * -EBUSY - the hugepage is busy (try to retry) |
1971 | * -EHWPOISON - the hugepage is already hwpoisoned |
1972 | */ |
1973 | int __get_huge_page_for_hwpoison(unsigned long pfn, int flags, |
1974 | bool *migratable_cleared) |
1975 | { |
1976 | struct page *page = pfn_to_page(pfn); |
1977 | struct folio *folio = page_folio(page); |
1978 | int ret = 2; /* fallback to normal page handling */ |
1979 | bool count_increased = false; |
1980 | |
1981 | if (!folio_test_hugetlb(folio)) |
1982 | goto out; |
1983 | |
1984 | if (flags & MF_COUNT_INCREASED) { |
1985 | ret = 1; |
1986 | count_increased = true; |
1987 | } else if (folio_test_hugetlb_freed(folio)) { |
1988 | ret = 0; |
1989 | } else if (folio_test_hugetlb_migratable(folio)) { |
1990 | ret = folio_try_get(folio); |
1991 | if (ret) |
1992 | count_increased = true; |
1993 | } else { |
1994 | ret = -EBUSY; |
1995 | if (!(flags & MF_NO_RETRY)) |
1996 | goto out; |
1997 | } |
1998 | |
1999 | if (folio_set_hugetlb_hwpoison(folio, page)) { |
2000 | ret = -EHWPOISON; |
2001 | goto out; |
2002 | } |
2003 | |
2004 | /* |
2005 | * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them |
2006 | * from being migrated by memory hotremove. |
2007 | */ |
2008 | if (count_increased && folio_test_hugetlb_migratable(folio)) { |
2009 | folio_clear_hugetlb_migratable(folio); |
2010 | *migratable_cleared = true; |
2011 | } |
2012 | |
2013 | return ret; |
2014 | out: |
2015 | if (count_increased) |
2016 | folio_put(folio); |
2017 | return ret; |
2018 | } |
2019 | |
2020 | /* |
2021 | * Taking refcount of hugetlb pages needs extra care about race conditions |
2022 | * with basic operations like hugepage allocation/free/demotion. |
2023 | * So some of prechecks for hwpoison (pinning, and testing/setting |
2024 | * PageHWPoison) should be done in single hugetlb_lock range. |
2025 | */ |
2026 | static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) |
2027 | { |
2028 | int res; |
2029 | struct page *p = pfn_to_page(pfn); |
2030 | struct folio *folio; |
2031 | unsigned long page_flags; |
2032 | bool migratable_cleared = false; |
2033 | |
2034 | *hugetlb = 1; |
2035 | retry: |
2036 | res = get_huge_page_for_hwpoison(pfn, flags, migratable_cleared: &migratable_cleared); |
2037 | if (res == 2) { /* fallback to normal page handling */ |
2038 | *hugetlb = 0; |
2039 | return 0; |
2040 | } else if (res == -EHWPOISON) { |
2041 | pr_err("%#lx: already hardware poisoned\n" , pfn); |
2042 | if (flags & MF_ACTION_REQUIRED) { |
2043 | folio = page_folio(p); |
2044 | res = kill_accessing_process(current, pfn: folio_pfn(folio), flags); |
2045 | } |
2046 | return res; |
2047 | } else if (res == -EBUSY) { |
2048 | if (!(flags & MF_NO_RETRY)) { |
2049 | flags |= MF_NO_RETRY; |
2050 | goto retry; |
2051 | } |
2052 | return action_result(pfn, type: MF_MSG_UNKNOWN, result: MF_IGNORED); |
2053 | } |
2054 | |
2055 | folio = page_folio(p); |
2056 | folio_lock(folio); |
2057 | |
2058 | if (hwpoison_filter(p)) { |
2059 | folio_clear_hugetlb_hwpoison(folio); |
2060 | if (migratable_cleared) |
2061 | folio_set_hugetlb_migratable(folio); |
2062 | folio_unlock(folio); |
2063 | if (res == 1) |
2064 | folio_put(folio); |
2065 | return -EOPNOTSUPP; |
2066 | } |
2067 | |
2068 | /* |
2069 | * Handling free hugepage. The possible race with hugepage allocation |
2070 | * or demotion can be prevented by PageHWPoison flag. |
2071 | */ |
2072 | if (res == 0) { |
2073 | folio_unlock(folio); |
2074 | if (__page_handle_poison(page: p) >= 0) { |
2075 | page_ref_inc(page: p); |
2076 | res = MF_RECOVERED; |
2077 | } else { |
2078 | res = MF_FAILED; |
2079 | } |
2080 | return action_result(pfn, type: MF_MSG_FREE_HUGE, result: res); |
2081 | } |
2082 | |
2083 | page_flags = folio->flags; |
2084 | |
2085 | if (!hwpoison_user_mappings(p, pfn, flags, hpage: &folio->page)) { |
2086 | folio_unlock(folio); |
2087 | return action_result(pfn, type: MF_MSG_UNMAP_FAILED, result: MF_IGNORED); |
2088 | } |
2089 | |
2090 | return identify_page_state(pfn, p, page_flags); |
2091 | } |
2092 | |
2093 | #else |
2094 | static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) |
2095 | { |
2096 | return 0; |
2097 | } |
2098 | |
2099 | static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag) |
2100 | { |
2101 | return 0; |
2102 | } |
2103 | #endif /* CONFIG_HUGETLB_PAGE */ |
2104 | |
2105 | /* Drop the extra refcount in case we come from madvise() */ |
2106 | static void put_ref_page(unsigned long pfn, int flags) |
2107 | { |
2108 | struct page *page; |
2109 | |
2110 | if (!(flags & MF_COUNT_INCREASED)) |
2111 | return; |
2112 | |
2113 | page = pfn_to_page(pfn); |
2114 | if (page) |
2115 | put_page(page); |
2116 | } |
2117 | |
2118 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
2119 | struct dev_pagemap *pgmap) |
2120 | { |
2121 | int rc = -ENXIO; |
2122 | |
2123 | /* device metadata space is not recoverable */ |
2124 | if (!pgmap_pfn_valid(pgmap, pfn)) |
2125 | goto out; |
2126 | |
2127 | /* |
2128 | * Call driver's implementation to handle the memory failure, otherwise |
2129 | * fall back to generic handler. |
2130 | */ |
2131 | if (pgmap_has_memory_failure(pgmap)) { |
2132 | rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags); |
2133 | /* |
2134 | * Fall back to generic handler too if operation is not |
2135 | * supported inside the driver/device/filesystem. |
2136 | */ |
2137 | if (rc != -EOPNOTSUPP) |
2138 | goto out; |
2139 | } |
2140 | |
2141 | rc = mf_generic_kill_procs(pfn, flags, pgmap); |
2142 | out: |
2143 | /* drop pgmap ref acquired in caller */ |
2144 | put_dev_pagemap(pgmap); |
2145 | if (rc != -EOPNOTSUPP) |
2146 | action_result(pfn, type: MF_MSG_DAX, result: rc ? MF_FAILED : MF_RECOVERED); |
2147 | return rc; |
2148 | } |
2149 | |
2150 | /** |
2151 | * memory_failure - Handle memory failure of a page. |
2152 | * @pfn: Page Number of the corrupted page |
2153 | * @flags: fine tune action taken |
2154 | * |
2155 | * This function is called by the low level machine check code |
2156 | * of an architecture when it detects hardware memory corruption |
2157 | * of a page. It tries its best to recover, which includes |
2158 | * dropping pages, killing processes etc. |
2159 | * |
2160 | * The function is primarily of use for corruptions that |
2161 | * happen outside the current execution context (e.g. when |
2162 | * detected by a background scrubber) |
2163 | * |
2164 | * Must run in process context (e.g. a work queue) with interrupts |
2165 | * enabled and no spinlocks held. |
2166 | * |
2167 | * Return: 0 for successfully handled the memory error, |
2168 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event, |
2169 | * < 0(except -EOPNOTSUPP) on failure. |
2170 | */ |
2171 | int memory_failure(unsigned long pfn, int flags) |
2172 | { |
2173 | struct page *p; |
2174 | struct page *hpage; |
2175 | struct dev_pagemap *pgmap; |
2176 | int res = 0; |
2177 | unsigned long page_flags; |
2178 | bool retry = true; |
2179 | int hugetlb = 0; |
2180 | |
2181 | if (!sysctl_memory_failure_recovery) |
2182 | panic(fmt: "Memory failure on page %lx" , pfn); |
2183 | |
2184 | mutex_lock(&mf_mutex); |
2185 | |
2186 | if (!(flags & MF_SW_SIMULATED)) |
2187 | hw_memory_failure = true; |
2188 | |
2189 | p = pfn_to_online_page(pfn); |
2190 | if (!p) { |
2191 | res = arch_memory_failure(pfn, flags); |
2192 | if (res == 0) |
2193 | goto unlock_mutex; |
2194 | |
2195 | if (pfn_valid(pfn)) { |
2196 | pgmap = get_dev_pagemap(pfn, NULL); |
2197 | put_ref_page(pfn, flags); |
2198 | if (pgmap) { |
2199 | res = memory_failure_dev_pagemap(pfn, flags, |
2200 | pgmap); |
2201 | goto unlock_mutex; |
2202 | } |
2203 | } |
2204 | pr_err("%#lx: memory outside kernel control\n" , pfn); |
2205 | res = -ENXIO; |
2206 | goto unlock_mutex; |
2207 | } |
2208 | |
2209 | try_again: |
2210 | res = try_memory_failure_hugetlb(pfn, flags, hugetlb: &hugetlb); |
2211 | if (hugetlb) |
2212 | goto unlock_mutex; |
2213 | |
2214 | if (TestSetPageHWPoison(page: p)) { |
2215 | pr_err("%#lx: already hardware poisoned\n" , pfn); |
2216 | res = -EHWPOISON; |
2217 | if (flags & MF_ACTION_REQUIRED) |
2218 | res = kill_accessing_process(current, pfn, flags); |
2219 | if (flags & MF_COUNT_INCREASED) |
2220 | put_page(page: p); |
2221 | goto unlock_mutex; |
2222 | } |
2223 | |
2224 | /* |
2225 | * We need/can do nothing about count=0 pages. |
2226 | * 1) it's a free page, and therefore in safe hand: |
2227 | * check_new_page() will be the gate keeper. |
2228 | * 2) it's part of a non-compound high order page. |
2229 | * Implies some kernel user: cannot stop them from |
2230 | * R/W the page; let's pray that the page has been |
2231 | * used and will be freed some time later. |
2232 | * In fact it's dangerous to directly bump up page count from 0, |
2233 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
2234 | */ |
2235 | if (!(flags & MF_COUNT_INCREASED)) { |
2236 | res = get_hwpoison_page(p, flags); |
2237 | if (!res) { |
2238 | if (is_free_buddy_page(page: p)) { |
2239 | if (take_page_off_buddy(page: p)) { |
2240 | page_ref_inc(page: p); |
2241 | res = MF_RECOVERED; |
2242 | } else { |
2243 | /* We lost the race, try again */ |
2244 | if (retry) { |
2245 | ClearPageHWPoison(page: p); |
2246 | retry = false; |
2247 | goto try_again; |
2248 | } |
2249 | res = MF_FAILED; |
2250 | } |
2251 | res = action_result(pfn, type: MF_MSG_BUDDY, result: res); |
2252 | } else { |
2253 | res = action_result(pfn, type: MF_MSG_KERNEL_HIGH_ORDER, result: MF_IGNORED); |
2254 | } |
2255 | goto unlock_mutex; |
2256 | } else if (res < 0) { |
2257 | res = action_result(pfn, type: MF_MSG_UNKNOWN, result: MF_IGNORED); |
2258 | goto unlock_mutex; |
2259 | } |
2260 | } |
2261 | |
2262 | hpage = compound_head(p); |
2263 | if (PageTransHuge(page: hpage)) { |
2264 | /* |
2265 | * The flag must be set after the refcount is bumped |
2266 | * otherwise it may race with THP split. |
2267 | * And the flag can't be set in get_hwpoison_page() since |
2268 | * it is called by soft offline too and it is just called |
2269 | * for !MF_COUNT_INCREASED. So here seems to be the best |
2270 | * place. |
2271 | * |
2272 | * Don't need care about the above error handling paths for |
2273 | * get_hwpoison_page() since they handle either free page |
2274 | * or unhandlable page. The refcount is bumped iff the |
2275 | * page is a valid handlable page. |
2276 | */ |
2277 | SetPageHasHWPoisoned(hpage); |
2278 | if (try_to_split_thp_page(page: p) < 0) { |
2279 | res = action_result(pfn, type: MF_MSG_UNSPLIT_THP, result: MF_IGNORED); |
2280 | goto unlock_mutex; |
2281 | } |
2282 | VM_BUG_ON_PAGE(!page_count(p), p); |
2283 | } |
2284 | |
2285 | /* |
2286 | * We ignore non-LRU pages for good reasons. |
2287 | * - PG_locked is only well defined for LRU pages and a few others |
2288 | * - to avoid races with __SetPageLocked() |
2289 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
2290 | * The check (unnecessarily) ignores LRU pages being isolated and |
2291 | * walked by the page reclaim code, however that's not a big loss. |
2292 | */ |
2293 | shake_page(p); |
2294 | |
2295 | lock_page(page: p); |
2296 | |
2297 | /* |
2298 | * We're only intended to deal with the non-Compound page here. |
2299 | * However, the page could have changed compound pages due to |
2300 | * race window. If this happens, we could try again to hopefully |
2301 | * handle the page next round. |
2302 | */ |
2303 | if (PageCompound(page: p)) { |
2304 | if (retry) { |
2305 | ClearPageHWPoison(page: p); |
2306 | unlock_page(page: p); |
2307 | put_page(page: p); |
2308 | flags &= ~MF_COUNT_INCREASED; |
2309 | retry = false; |
2310 | goto try_again; |
2311 | } |
2312 | res = action_result(pfn, type: MF_MSG_DIFFERENT_COMPOUND, result: MF_IGNORED); |
2313 | goto unlock_page; |
2314 | } |
2315 | |
2316 | /* |
2317 | * We use page flags to determine what action should be taken, but |
2318 | * the flags can be modified by the error containment action. One |
2319 | * example is an mlocked page, where PG_mlocked is cleared by |
2320 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status |
2321 | * correctly, we save a copy of the page flags at this time. |
2322 | */ |
2323 | page_flags = p->flags; |
2324 | |
2325 | if (hwpoison_filter(p)) { |
2326 | ClearPageHWPoison(page: p); |
2327 | unlock_page(page: p); |
2328 | put_page(page: p); |
2329 | res = -EOPNOTSUPP; |
2330 | goto unlock_mutex; |
2331 | } |
2332 | |
2333 | /* |
2334 | * __munlock_folio() may clear a writeback page's LRU flag without |
2335 | * page_lock. We need wait writeback completion for this page or it |
2336 | * may trigger vfs BUG while evict inode. |
2337 | */ |
2338 | if (!PageLRU(page: p) && !PageWriteback(page: p)) |
2339 | goto identify_page_state; |
2340 | |
2341 | /* |
2342 | * It's very difficult to mess with pages currently under IO |
2343 | * and in many cases impossible, so we just avoid it here. |
2344 | */ |
2345 | wait_on_page_writeback(page: p); |
2346 | |
2347 | /* |
2348 | * Now take care of user space mappings. |
2349 | * Abort on fail: __filemap_remove_folio() assumes unmapped page. |
2350 | */ |
2351 | if (!hwpoison_user_mappings(p, pfn, flags, hpage: p)) { |
2352 | res = action_result(pfn, type: MF_MSG_UNMAP_FAILED, result: MF_IGNORED); |
2353 | goto unlock_page; |
2354 | } |
2355 | |
2356 | /* |
2357 | * Torn down by someone else? |
2358 | */ |
2359 | if (PageLRU(page: p) && !PageSwapCache(page: p) && p->mapping == NULL) { |
2360 | res = action_result(pfn, type: MF_MSG_TRUNCATED_LRU, result: MF_IGNORED); |
2361 | goto unlock_page; |
2362 | } |
2363 | |
2364 | identify_page_state: |
2365 | res = identify_page_state(pfn, p, page_flags); |
2366 | mutex_unlock(lock: &mf_mutex); |
2367 | return res; |
2368 | unlock_page: |
2369 | unlock_page(page: p); |
2370 | unlock_mutex: |
2371 | mutex_unlock(lock: &mf_mutex); |
2372 | return res; |
2373 | } |
2374 | EXPORT_SYMBOL_GPL(memory_failure); |
2375 | |
2376 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
2377 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) |
2378 | |
2379 | struct memory_failure_entry { |
2380 | unsigned long pfn; |
2381 | int flags; |
2382 | }; |
2383 | |
2384 | struct memory_failure_cpu { |
2385 | DECLARE_KFIFO(fifo, struct memory_failure_entry, |
2386 | MEMORY_FAILURE_FIFO_SIZE); |
2387 | spinlock_t lock; |
2388 | struct work_struct work; |
2389 | }; |
2390 | |
2391 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); |
2392 | |
2393 | /** |
2394 | * memory_failure_queue - Schedule handling memory failure of a page. |
2395 | * @pfn: Page Number of the corrupted page |
2396 | * @flags: Flags for memory failure handling |
2397 | * |
2398 | * This function is called by the low level hardware error handler |
2399 | * when it detects hardware memory corruption of a page. It schedules |
2400 | * the recovering of error page, including dropping pages, killing |
2401 | * processes etc. |
2402 | * |
2403 | * The function is primarily of use for corruptions that |
2404 | * happen outside the current execution context (e.g. when |
2405 | * detected by a background scrubber) |
2406 | * |
2407 | * Can run in IRQ context. |
2408 | */ |
2409 | void memory_failure_queue(unsigned long pfn, int flags) |
2410 | { |
2411 | struct memory_failure_cpu *mf_cpu; |
2412 | unsigned long proc_flags; |
2413 | struct memory_failure_entry entry = { |
2414 | .pfn = pfn, |
2415 | .flags = flags, |
2416 | }; |
2417 | |
2418 | mf_cpu = &get_cpu_var(memory_failure_cpu); |
2419 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); |
2420 | if (kfifo_put(&mf_cpu->fifo, entry)) |
2421 | schedule_work_on(smp_processor_id(), work: &mf_cpu->work); |
2422 | else |
2423 | pr_err("buffer overflow when queuing memory failure at %#lx\n" , |
2424 | pfn); |
2425 | spin_unlock_irqrestore(lock: &mf_cpu->lock, flags: proc_flags); |
2426 | put_cpu_var(memory_failure_cpu); |
2427 | } |
2428 | EXPORT_SYMBOL_GPL(memory_failure_queue); |
2429 | |
2430 | static void memory_failure_work_func(struct work_struct *work) |
2431 | { |
2432 | struct memory_failure_cpu *mf_cpu; |
2433 | struct memory_failure_entry entry = { 0, }; |
2434 | unsigned long proc_flags; |
2435 | int gotten; |
2436 | |
2437 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
2438 | for (;;) { |
2439 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); |
2440 | gotten = kfifo_get(&mf_cpu->fifo, &entry); |
2441 | spin_unlock_irqrestore(lock: &mf_cpu->lock, flags: proc_flags); |
2442 | if (!gotten) |
2443 | break; |
2444 | if (entry.flags & MF_SOFT_OFFLINE) |
2445 | soft_offline_page(pfn: entry.pfn, flags: entry.flags); |
2446 | else |
2447 | memory_failure(entry.pfn, entry.flags); |
2448 | } |
2449 | } |
2450 | |
2451 | /* |
2452 | * Process memory_failure work queued on the specified CPU. |
2453 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. |
2454 | */ |
2455 | void memory_failure_queue_kick(int cpu) |
2456 | { |
2457 | struct memory_failure_cpu *mf_cpu; |
2458 | |
2459 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); |
2460 | cancel_work_sync(work: &mf_cpu->work); |
2461 | memory_failure_work_func(work: &mf_cpu->work); |
2462 | } |
2463 | |
2464 | static int __init memory_failure_init(void) |
2465 | { |
2466 | struct memory_failure_cpu *mf_cpu; |
2467 | int cpu; |
2468 | |
2469 | for_each_possible_cpu(cpu) { |
2470 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); |
2471 | spin_lock_init(&mf_cpu->lock); |
2472 | INIT_KFIFO(mf_cpu->fifo); |
2473 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); |
2474 | } |
2475 | |
2476 | register_sysctl_init("vm" , memory_failure_table); |
2477 | |
2478 | return 0; |
2479 | } |
2480 | core_initcall(memory_failure_init); |
2481 | |
2482 | #undef pr_fmt |
2483 | #define pr_fmt(fmt) "" fmt |
2484 | #define unpoison_pr_info(fmt, pfn, rs) \ |
2485 | ({ \ |
2486 | if (__ratelimit(rs)) \ |
2487 | pr_info(fmt, pfn); \ |
2488 | }) |
2489 | |
2490 | /** |
2491 | * unpoison_memory - Unpoison a previously poisoned page |
2492 | * @pfn: Page number of the to be unpoisoned page |
2493 | * |
2494 | * Software-unpoison a page that has been poisoned by |
2495 | * memory_failure() earlier. |
2496 | * |
2497 | * This is only done on the software-level, so it only works |
2498 | * for linux injected failures, not real hardware failures |
2499 | * |
2500 | * Returns 0 for success, otherwise -errno. |
2501 | */ |
2502 | int unpoison_memory(unsigned long pfn) |
2503 | { |
2504 | struct folio *folio; |
2505 | struct page *p; |
2506 | int ret = -EBUSY, ghp; |
2507 | unsigned long count = 1; |
2508 | bool huge = false; |
2509 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
2510 | DEFAULT_RATELIMIT_BURST); |
2511 | |
2512 | if (!pfn_valid(pfn)) |
2513 | return -ENXIO; |
2514 | |
2515 | p = pfn_to_page(pfn); |
2516 | folio = page_folio(p); |
2517 | |
2518 | mutex_lock(&mf_mutex); |
2519 | |
2520 | if (hw_memory_failure) { |
2521 | unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n" , |
2522 | pfn, &unpoison_rs); |
2523 | ret = -EOPNOTSUPP; |
2524 | goto unlock_mutex; |
2525 | } |
2526 | |
2527 | if (!PageHWPoison(page: p)) { |
2528 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n" , |
2529 | pfn, &unpoison_rs); |
2530 | goto unlock_mutex; |
2531 | } |
2532 | |
2533 | if (folio_ref_count(folio) > 1) { |
2534 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n" , |
2535 | pfn, &unpoison_rs); |
2536 | goto unlock_mutex; |
2537 | } |
2538 | |
2539 | if (folio_test_slab(folio) || PageTable(page: &folio->page) || |
2540 | folio_test_reserved(folio) || PageOffline(page: &folio->page)) |
2541 | goto unlock_mutex; |
2542 | |
2543 | /* |
2544 | * Note that folio->_mapcount is overloaded in SLAB, so the simple test |
2545 | * in folio_mapped() has to be done after folio_test_slab() is checked. |
2546 | */ |
2547 | if (folio_mapped(folio)) { |
2548 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n" , |
2549 | pfn, &unpoison_rs); |
2550 | goto unlock_mutex; |
2551 | } |
2552 | |
2553 | if (folio_mapping(folio)) { |
2554 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n" , |
2555 | pfn, &unpoison_rs); |
2556 | goto unlock_mutex; |
2557 | } |
2558 | |
2559 | ghp = get_hwpoison_page(p, flags: MF_UNPOISON); |
2560 | if (!ghp) { |
2561 | if (PageHuge(page: p)) { |
2562 | huge = true; |
2563 | count = folio_free_raw_hwp(folio, move_flag: false); |
2564 | if (count == 0) |
2565 | goto unlock_mutex; |
2566 | } |
2567 | ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY; |
2568 | } else if (ghp < 0) { |
2569 | if (ghp == -EHWPOISON) { |
2570 | ret = put_page_back_buddy(page: p) ? 0 : -EBUSY; |
2571 | } else { |
2572 | ret = ghp; |
2573 | unpoison_pr_info("Unpoison: failed to grab page %#lx\n" , |
2574 | pfn, &unpoison_rs); |
2575 | } |
2576 | } else { |
2577 | if (PageHuge(page: p)) { |
2578 | huge = true; |
2579 | count = folio_free_raw_hwp(folio, move_flag: false); |
2580 | if (count == 0) { |
2581 | folio_put(folio); |
2582 | goto unlock_mutex; |
2583 | } |
2584 | } |
2585 | |
2586 | folio_put(folio); |
2587 | if (TestClearPageHWPoison(page: p)) { |
2588 | folio_put(folio); |
2589 | ret = 0; |
2590 | } |
2591 | } |
2592 | |
2593 | unlock_mutex: |
2594 | mutex_unlock(lock: &mf_mutex); |
2595 | if (!ret) { |
2596 | if (!huge) |
2597 | num_poisoned_pages_sub(pfn, i: 1); |
2598 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n" , |
2599 | page_to_pfn(p), &unpoison_rs); |
2600 | } |
2601 | return ret; |
2602 | } |
2603 | EXPORT_SYMBOL(unpoison_memory); |
2604 | |
2605 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
2606 | { |
2607 | bool isolated = false; |
2608 | |
2609 | if (PageHuge(page)) { |
2610 | isolated = isolate_hugetlb(page_folio(page), list: pagelist); |
2611 | } else { |
2612 | bool lru = !__PageMovable(page); |
2613 | |
2614 | if (lru) |
2615 | isolated = isolate_lru_page(page); |
2616 | else |
2617 | isolated = isolate_movable_page(page, |
2618 | ISOLATE_UNEVICTABLE); |
2619 | |
2620 | if (isolated) { |
2621 | list_add(new: &page->lru, head: pagelist); |
2622 | if (lru) |
2623 | inc_node_page_state(page, NR_ISOLATED_ANON + |
2624 | page_is_file_lru(page)); |
2625 | } |
2626 | } |
2627 | |
2628 | /* |
2629 | * If we succeed to isolate the page, we grabbed another refcount on |
2630 | * the page, so we can safely drop the one we got from get_any_page(). |
2631 | * If we failed to isolate the page, it means that we cannot go further |
2632 | * and we will return an error, so drop the reference we got from |
2633 | * get_any_page() as well. |
2634 | */ |
2635 | put_page(page); |
2636 | return isolated; |
2637 | } |
2638 | |
2639 | /* |
2640 | * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages. |
2641 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. |
2642 | * If the page is mapped, it migrates the contents over. |
2643 | */ |
2644 | static int soft_offline_in_use_page(struct page *page) |
2645 | { |
2646 | long ret = 0; |
2647 | unsigned long pfn = page_to_pfn(page); |
2648 | struct page *hpage = compound_head(page); |
2649 | char const *msg_page[] = {"page" , "hugepage" }; |
2650 | bool huge = PageHuge(page); |
2651 | LIST_HEAD(pagelist); |
2652 | struct migration_target_control mtc = { |
2653 | .nid = NUMA_NO_NODE, |
2654 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, |
2655 | }; |
2656 | |
2657 | if (!huge && PageTransHuge(page: hpage)) { |
2658 | if (try_to_split_thp_page(page)) { |
2659 | pr_info("soft offline: %#lx: thp split failed\n" , pfn); |
2660 | return -EBUSY; |
2661 | } |
2662 | hpage = page; |
2663 | } |
2664 | |
2665 | lock_page(page); |
2666 | if (!huge) |
2667 | wait_on_page_writeback(page); |
2668 | if (PageHWPoison(page)) { |
2669 | unlock_page(page); |
2670 | put_page(page); |
2671 | pr_info("soft offline: %#lx page already poisoned\n" , pfn); |
2672 | return 0; |
2673 | } |
2674 | |
2675 | if (!huge && PageLRU(page) && !PageSwapCache(page)) |
2676 | /* |
2677 | * Try to invalidate first. This should work for |
2678 | * non dirty unmapped page cache pages. |
2679 | */ |
2680 | ret = invalidate_inode_page(page); |
2681 | unlock_page(page); |
2682 | |
2683 | if (ret) { |
2684 | pr_info("soft_offline: %#lx: invalidated\n" , pfn); |
2685 | page_handle_poison(page, hugepage_or_freepage: false, release: true); |
2686 | return 0; |
2687 | } |
2688 | |
2689 | if (isolate_page(page: hpage, pagelist: &pagelist)) { |
2690 | ret = migrate_pages(l: &pagelist, new: alloc_migration_target, NULL, |
2691 | private: (unsigned long)&mtc, mode: MIGRATE_SYNC, reason: MR_MEMORY_FAILURE, NULL); |
2692 | if (!ret) { |
2693 | bool release = !huge; |
2694 | |
2695 | if (!page_handle_poison(page, hugepage_or_freepage: huge, release)) |
2696 | ret = -EBUSY; |
2697 | } else { |
2698 | if (!list_empty(head: &pagelist)) |
2699 | putback_movable_pages(l: &pagelist); |
2700 | |
2701 | pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n" , |
2702 | pfn, msg_page[huge], ret, &page->flags); |
2703 | if (ret > 0) |
2704 | ret = -EBUSY; |
2705 | } |
2706 | } else { |
2707 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n" , |
2708 | pfn, msg_page[huge], page_count(page), &page->flags); |
2709 | ret = -EBUSY; |
2710 | } |
2711 | return ret; |
2712 | } |
2713 | |
2714 | /** |
2715 | * soft_offline_page - Soft offline a page. |
2716 | * @pfn: pfn to soft-offline |
2717 | * @flags: flags. Same as memory_failure(). |
2718 | * |
2719 | * Returns 0 on success |
2720 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event |
2721 | * < 0 otherwise negated errno. |
2722 | * |
2723 | * Soft offline a page, by migration or invalidation, |
2724 | * without killing anything. This is for the case when |
2725 | * a page is not corrupted yet (so it's still valid to access), |
2726 | * but has had a number of corrected errors and is better taken |
2727 | * out. |
2728 | * |
2729 | * The actual policy on when to do that is maintained by |
2730 | * user space. |
2731 | * |
2732 | * This should never impact any application or cause data loss, |
2733 | * however it might take some time. |
2734 | * |
2735 | * This is not a 100% solution for all memory, but tries to be |
2736 | * ``good enough'' for the majority of memory. |
2737 | */ |
2738 | int soft_offline_page(unsigned long pfn, int flags) |
2739 | { |
2740 | int ret; |
2741 | bool try_again = true; |
2742 | struct page *page; |
2743 | |
2744 | if (!pfn_valid(pfn)) { |
2745 | WARN_ON_ONCE(flags & MF_COUNT_INCREASED); |
2746 | return -ENXIO; |
2747 | } |
2748 | |
2749 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2750 | page = pfn_to_online_page(pfn); |
2751 | if (!page) { |
2752 | put_ref_page(pfn, flags); |
2753 | return -EIO; |
2754 | } |
2755 | |
2756 | mutex_lock(&mf_mutex); |
2757 | |
2758 | if (PageHWPoison(page)) { |
2759 | pr_info("%s: %#lx page already poisoned\n" , __func__, pfn); |
2760 | put_ref_page(pfn, flags); |
2761 | mutex_unlock(lock: &mf_mutex); |
2762 | return 0; |
2763 | } |
2764 | |
2765 | retry: |
2766 | get_online_mems(); |
2767 | ret = get_hwpoison_page(p: page, flags: flags | MF_SOFT_OFFLINE); |
2768 | put_online_mems(); |
2769 | |
2770 | if (hwpoison_filter(page)) { |
2771 | if (ret > 0) |
2772 | put_page(page); |
2773 | |
2774 | mutex_unlock(lock: &mf_mutex); |
2775 | return -EOPNOTSUPP; |
2776 | } |
2777 | |
2778 | if (ret > 0) { |
2779 | ret = soft_offline_in_use_page(page); |
2780 | } else if (ret == 0) { |
2781 | if (!page_handle_poison(page, hugepage_or_freepage: true, release: false)) { |
2782 | if (try_again) { |
2783 | try_again = false; |
2784 | flags &= ~MF_COUNT_INCREASED; |
2785 | goto retry; |
2786 | } |
2787 | ret = -EBUSY; |
2788 | } |
2789 | } |
2790 | |
2791 | mutex_unlock(lock: &mf_mutex); |
2792 | |
2793 | return ret; |
2794 | } |
2795 | |